extern/whisper.cpp
1
0
Fork 1
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-08-09 21:38:44 +02:00
Code Issues Packages Projects Releases Wiki Activity

ggml-cpu: enable IBM NNPA Vector Intrinsics (llama/14317)

Browse Source 
* ggml-cpu: add nnpa compile flag

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>
(cherry picked from commit 4a9f60c201573128f73a65999b3e5cc497fae5c1)

* ggml-cpu: add fp16->fp32 nnpa first

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>
(cherry picked from commit 8d4a7987f9c1887f716be96250f2caeee0253929)

* ggml-cpu: add fp32->fp16

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>
(cherry picked from commit 0ff0d6516247a41d2ade42b42cf0d676a4dd1627)

* ggml-cpu: better variable names

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>
(cherry picked from commit 2f58bbcbb89c183340e252362b2a40651f573f1f)

* docs: update s390x docs

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>
(cherry picked from commit 01b929491b50071a5d0572235dcf5a449da70aa7)

* ggml-cpu: add debugging prints to see if dlf16 is correct

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: fix print vs printf

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: fix float placeholder

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: ensure fp16 and fp32 load and stores are called

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: fp16 load ensured to hit

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: remove sigint from fp16 store

for some reason, the function is not getting a hit when debugged with
    gdb. we will need to investigate further

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: activate nnpa for ggml_cpu_fp16_to_fp32

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: nnpa activate ggml_cpu_fp16_to_fp32 for 8 elements

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: nnpa switch to vec_xst test

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: switch to vec_xst for 4 element loops also

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: rework noop

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: remove noop, general code cleanup

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: clarify variable naming

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: activate nnpa for ggml_cpu_fp32_to_fp16

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: add breakpoint for debugging

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: test fix for conversion failure

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: disable fp32->fp16 nnpa conversions for now

there are some conversion failures in nnpa that requires the eyes of an
ibm stsm. will create a separate pr to introduce the fp32->fp16 change.

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: switch to elif macro

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: reattempt fp32->fp16

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: fix typo

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: reattempt fp32->fp16

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: fix compiler types

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: change to typedef vector types

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: add 4 element loops for fp32->fp16

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: clarified vector naming

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: bring back fp32->fp16 store nnpa

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: activate nnpa fp32->fp16 or fp16->fp32 compute

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: add nnpa macro check in ggml-impl

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: add missing __func__

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: diagnose why __NNPA__ macro is not being defined

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: import vecintrin.h to fix compiler errors

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: update macro tests

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: move s390x typedef to own header file

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* Revert "ggml-cpu: move s390x typedef to own header file"

This reverts commit 157f856c34589566151630e294563a420702db39.

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: switch to importing ggml-cpu-impl instead

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: fix macro declaration

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: test more macros

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: add debug prints

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: bruteforce macro definitions

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: move macro definitions

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: add ggml-impl.h to cmakelists

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: switch to private macros

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: move s390x typedef to own header file

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>
(cherry picked from commit 157f856c34589566151630e294563a420702db39)

* ggml-cpu: move things around

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: bring back compile macros

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: switch to quotes for import

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: add compiler error macro

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: add s390x detection in ggml-src

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: bring back compile definitions

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: undo cmakelists work

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* Revert "ggml-cpu: move s390x typedef to own header file"

This reverts commit 18d79e1a30b39d9aaa0bd58400c5cf2c32135c9a.

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: remove typedefs.h

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: remove typedef from cmakelists

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: add ggml-impl.h future notes

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: add todo comment for future reference

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: clarify naming of dlf16

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: remove unnecessary target compile definitions

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: move nnpa fp16->fp32 and fp32->fp16 to simd-mappings

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml: refactor fp32->fp16 and fp16->fp32 simd to ggml-cpu

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* docs: update broken huggingface link for s390x

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: fix duplicate func names during compile

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* Revert "ggml-cpu: fix duplicate func names during compile"

This reverts commit fbb733451f27677063b914d4f6c9a9841d45b38d.

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* Revert "ggml: refactor fp32->fp16 and fp16->fp32 simd to ggml-cpu"

This reverts commit bd288e8fa52b5244f65cee21cb61062f1a9e0ca5.

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml: refactor fp16<->fp32 simd to ggml-cpu

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: fix missing simd-mappings.h import in quants.c

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: fix missing simd-mappings.h within repack

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: fix amx mmq missing simd-mappings.h

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: attempt at fixing loongarch failing build

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: move nnpa together with other fp16<->fp32 simd

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: fix wrong refactor of ggml-base

ref: https://github.com/ggml-org/llama.cpp/pull/14317#discussion_r2164176555

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml: remove dependency on ggml-cpu from ggml-base

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: rename all fp16<->fp32 macros to prefix with ggml_cpu

ref: https://github.com/ggml-org/llama.cpp/pull/14317#discussion_r2164449406

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: remove mistaken fallback macro

fallback logic was already implemented but i was too sleepy to realise

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml: move ggml_table_f32_f16 to ggml-cpu

ref: https://github.com/ggml-org/llama.cpp/pull/14317#discussion_r2164775006

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: move ggml_table_f32_f16 back to ggml-base due to ci failures

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* Revert "ggml-cpu: move ggml_table_f32_f16 back to ggml-base due to ci failures"

This reverts commit 32a3533564bdb7902cefb9c89b1c9e956a81ce29.

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* Revert "ggml: move ggml_table_f32_f16 to ggml-cpu"

This reverts commit 9e40d984ad27d7b60392fb2b7548885201864fe4.

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml: move ggml_table_f32_f16 to ggml-cpu

ref: https://github.com/ggml-org/llama.cpp/pull/14317#discussion_r2164775006

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>
(cherry picked from commit 9e40d984ad27d7b60392fb2b7548885201864fe4)

* ggml: move ggml_table_f32_f16 to ggml-cpu.c

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: extern c ggml_table_f32_f16 + chore docs

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: dedup ggml_table_f32_f16 from simd-mappings.h

we rely on the variable declaration in ggml-cpu.c instead

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* Revert "ggml-cpu: dedup ggml_table_f32_f16 from simd-mappings.h"

This reverts commit f71b21d2f74f5e03ec0c2b4fefd3cbf395aecf16.

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* ggml-cpu: bring back ggml_table_f32_f16

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* Revert "ggml-cpu: bring back ggml_table_f32_f16"

This reverts commit 2dce119178bed5ef5c8398c4230ddd14fef80e49.

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>

* fix ggml time initialization

* fix f32_f16 table init

* remove extra line

---------

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>
Co-authored-by: slaren <slarengh@gmail.com>
This commit is contained in:
Aaron Teo
2025-06-26 05:49:04 +08:00
committed by Georgi Gerganov
parent 99764f5767
commit 16aa7d151d
27 changed files with 972 additions and 850 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
ggml/CMakeLists.txt
View File

@ -131,6 +131,7 @@ option(GGML_RVV "ggml: enable rvv" ON)
option(GGML_RV_ZFH "ggml: enable riscv zfh" OFF) option(GGML_RV_ZFH "ggml: enable riscv zfh" OFF)
option(GGML_XTHEADVECTOR "ggml: enable xtheadvector" OFF) option(GGML_XTHEADVECTOR "ggml: enable xtheadvector" OFF)
option(GGML_VXE "ggml: enable vxe" ON) option(GGML_VXE "ggml: enable vxe" ON)
option(GGML_NNPA "ggml: enable nnpa" ON)
option(GGML_CPU_ALL_VARIANTS "ggml: build all variants of the CPU backend (requires GGML_BACKEND_DL)" OFF) option(GGML_CPU_ALL_VARIANTS "ggml: build all variants of the CPU backend (requires GGML_BACKEND_DL)" OFF)
set(GGML_CPU_ARM_ARCH "" CACHE STRING "ggml: CPU architecture for ARM") set(GGML_CPU_ARM_ARCH "" CACHE STRING "ggml: CPU architecture for ARM")

1
ggml/include/ggml-cpu.h
View File

@ -101,6 +101,7 @@ extern "C" {
GGML_BACKEND_API int ggml_cpu_has_riscv_v (void); GGML_BACKEND_API int ggml_cpu_has_riscv_v (void);
GGML_BACKEND_API int ggml_cpu_has_vsx (void); GGML_BACKEND_API int ggml_cpu_has_vsx (void);
GGML_BACKEND_API int ggml_cpu_has_vxe (void); GGML_BACKEND_API int ggml_cpu_has_vxe (void);
GGML_BACKEND_API int ggml_cpu_has_nnpa (void);
GGML_BACKEND_API int ggml_cpu_has_wasm_simd (void); GGML_BACKEND_API int ggml_cpu_has_wasm_simd (void);
GGML_BACKEND_API int ggml_cpu_has_llamafile (void); GGML_BACKEND_API int ggml_cpu_has_llamafile (void);

8
ggml/src/ggml-cpu/CMakeLists.txt
View File

@ -448,6 +448,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
# TODO: Separation to determine activation of VX/VXE/VXE2 # TODO: Separation to determine activation of VX/VXE/VXE2
if (${S390X_M} MATCHES "8561|8562") if (${S390X_M} MATCHES "8561|8562")
set(GGML_NNPA OFF)
message(STATUS "z15 target") message(STATUS "z15 target")
list(APPEND ARCH_FLAGS -march=z15) list(APPEND ARCH_FLAGS -march=z15)
elseif (${S390X_M} MATCHES "3931") elseif (${S390X_M} MATCHES "3931")
@ -464,7 +465,14 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
endif() endif()
if (GGML_VXE) if (GGML_VXE)
message(STATUS "VX/VXE/VXE2 enabled")
list(APPEND ARCH_FLAGS -mvx -mzvector) list(APPEND ARCH_FLAGS -mvx -mzvector)
list(APPEND ARCH_DEFINITIONS GGML_VXE)
endif()
if (GGML_NNPA)
message(STATUS "NNPA enabled")
list(APPEND ARCH_DEFINITIONS GGML_NNPA)
endif() endif()
elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "wasm") elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "wasm")
message(STATUS "Wasm detected") message(STATUS "Wasm detected")

19
ggml/src/ggml-cpu/amx/mmq.cpp
View File

@ -8,6 +8,7 @@
#include "mmq.h" #include "mmq.h"
#include "ggml-impl.h" #include "ggml-impl.h"
#include "ggml-cpu-impl.h" #include "ggml-cpu-impl.h"
#include "simd-mappings.h"
#include "quants.h" #include "quants.h"
#include "ggml-quants.h" #include "ggml-quants.h"
#include <algorithm> #include <algorithm>
@ -453,7 +454,7 @@ void quantize_row_q8_K_vnni(const float * RESTRICT x, void * RESTRICT vy, int64_
// Quantize these floats // Quantize these floats
const float iscale = 127.f / amax; const float iscale = 127.f / amax;
y[i].d = GGML_FP32_TO_FP16(1 / iscale); y[i].d = GGML_CPU_FP32_TO_FP16(1 / iscale);
const float id = ( amax != 0.0f ) ? iscale : 0.f; const float id = ( amax != 0.0f ) ? iscale : 0.f;
const __m512 vscale = _mm512_set1_ps(id); const __m512 vscale = _mm512_set1_ps(id);
@ -1090,7 +1091,7 @@ struct acc_C<block_q8_0, block_q4_0, is_acc> {
const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset))); const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset)));
for (int m = 0; m < nr; ++m) { for (int m = 0; m < nr; ++m) {
const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d)); const __m512 vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[m * lda].d));
const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N)); const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));
__m512 vsum; __m512 vsum;
@ -1113,8 +1114,8 @@ struct acc_C<block_q8_1, block_q4_1, is_acc> {
const __m512 vm0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset + TILE_N * sizeof(ggml_half)))); const __m512 vm0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset + TILE_N * sizeof(ggml_half))));
for (int m = 0; m < nr; ++m) { for (int m = 0; m < nr; ++m) {
const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d)); const __m512 vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[m * lda].d));
const __m512 vs1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].s)); const __m512 vs1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[m * lda].s));
const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N)); const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));
__m512 vsum; __m512 vsum;
@ -1137,7 +1138,7 @@ struct acc_C<block_q8_0, block_q8_0, is_acc> {
const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset))); const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset)));
for (int m = 0; m < nr; ++m) { for (int m = 0; m < nr; ++m) {
const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d)); const __m512 vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[m * lda].d));
const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N)); const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));
__m512 vsum; __m512 vsum;
@ -1437,7 +1438,7 @@ struct tinygemm_kernel_vnni<block_q8_0, block_q4_0, float, BLOCK_M, BLOCK_N, BLO
va[k] = _mm512_set1_epi32(a_ptr[k]); va[k] = _mm512_set1_epi32(a_ptr[k]);
vcomp = _mm512_dpbusd_epi32(vcomp, off, va[k]); vcomp = _mm512_dpbusd_epi32(vcomp, off, va[k]);
} }
vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d)); vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[0 * KB + i].d));
} }
// load b // load b
@ -1498,8 +1499,8 @@ struct tinygemm_kernel_vnni<block_q8_1, block_q4_1, float, 1, BLOCK_N, BLOCK_K>
for (int k = 0; k < 8; ++k) { for (int k = 0; k < 8; ++k) {
va[k] = _mm512_set1_epi32(a_ptr[k]); va[k] = _mm512_set1_epi32(a_ptr[k]);
} }
vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d)); vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[0 * KB + i].d));
vs1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].s)); vs1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[0 * KB + i].s));
} }
// load b // load b
@ -1571,7 +1572,7 @@ struct tinygemm_kernel_vnni<block_q8_0, block_q8_0, float, BLOCK_M, BLOCK_N, BLO
va[k] = _mm512_set1_epi32(a_ptr[k]); va[k] = _mm512_set1_epi32(a_ptr[k]);
va[k] = _mm512_add_epi8(va[k], off); va[k] = _mm512_add_epi8(va[k], off);
} }
vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d)); vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[0 * KB + i].d));
} }
// load b // load b

217
ggml/src/ggml-cpu/arch/arm/quants.c
View File

@ -3,6 +3,7 @@
#include "ggml-quants.h" #include "ggml-quants.h"
#include "ggml-impl.h" #include "ggml-impl.h"
#include "ggml-cpu.h" #include "ggml-cpu.h"
#include "simd-mappings.h"
#include "../../quants.h" #include "../../quants.h"
#include "../../ggml-cpu-impl.h" #include "../../ggml-cpu-impl.h"
@ -62,7 +63,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
const float d = amax / ((1 << 7) - 1); const float d = amax / ((1 << 7) - 1);
const float id = d ? 1.0f/d : 0.0f; const float id = d ? 1.0f/d : 0.0f;
y[i].d = GGML_FP32_TO_FP16(d); y[i].d = GGML_CPU_FP32_TO_FP16(d);
for (int j = 0; j < 8; j++) { for (int j = 0; j < 8; j++) {
const float32x4_t v = vmulq_n_f32(srcv[j], id); const float32x4_t v = vmulq_n_f32(srcv[j], id);
@ -104,7 +105,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
const float d = amax / ((1 << 7) - 1); const float d = amax / ((1 << 7) - 1);
const float id = d ? 1.0f/d : 0.0f; const float id = d ? 1.0f/d : 0.0f;
y[i].d = GGML_FP32_TO_FP16(d); y[i].d = GGML_CPU_FP32_TO_FP16(d);
int32x4_t accv = vdupq_n_s32(0); int32x4_t accv = vdupq_n_s32(0);
@ -120,7 +121,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
accv = vaddq_s32(accv, vi); accv = vaddq_s32(accv, vi);
} }
y[i].s = GGML_FP32_TO_FP16(d * vaddvq_s32(accv)); y[i].s = GGML_CPU_FP32_TO_FP16(d * vaddvq_s32(accv));
} }
#else #else
GGML_UNUSED(nb); GGML_UNUSED(nb);
@ -194,10 +195,10 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16); const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
float32_t _scale[4] = { float32_t _scale[4] = {
GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d), GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y0->d),
GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d), GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y1->d),
GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d), GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y0->d),
GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d) GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y1->d)
}; };
float32x4_t scale = vld1q_f32(_scale); float32x4_t scale = vld1q_f32(_scale);
@ -274,10 +275,10 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
// dot product // dot product
sumv0 = svmla_n_f32_x(ph4, sumv0, svcvt_f32_s32_x(ph4, svadd_x(ph4, sumv0 = svmla_n_f32_x(ph4, sumv0, svcvt_f32_s32_x(ph4, svadd_x(ph4,
svdot_s32(svdup_n_s32(0), qx0ls, qy0l), svdot_s32(svdup_n_s32(0), qx0ls, qy0l),
svdot_s32(svdup_n_s32(0), qx0hs, qy0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); svdot_s32(svdup_n_s32(0), qx0hs, qy0h))), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
sumv1 = svmla_n_f32_x(ph4, sumv1, svcvt_f32_s32_x(ph4, svadd_x(ph4, sumv1 = svmla_n_f32_x(ph4, sumv1, svcvt_f32_s32_x(ph4, svadd_x(ph4,
svdot_s32(svdup_n_s32(0), qx1ls, qy1l), svdot_s32(svdup_n_s32(0), qx1ls, qy1l),
svdot_s32(svdup_n_s32(0), qx1hs, qy1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); svdot_s32(svdup_n_s32(0), qx1hs, qy1h))), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
} }
sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1)); sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
@ -313,9 +314,9 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
// dot product // dot product
sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(), sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(),
svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(), sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(),
svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
} }
sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1)); sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
@ -354,9 +355,9 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
// dot product // dot product
sumv0 = svmla_n_f32_x(ph32, sumv0, svcvt_f32_s32_x(ph32, sumv0 = svmla_n_f32_x(ph32, sumv0, svcvt_f32_s32_x(ph32,
svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
sumv1 = svmla_n_f32_x(ph32, sumv1, svcvt_f32_s32_x(ph32, sumv1 = svmla_n_f32_x(ph32, sumv1, svcvt_f32_s32_x(ph32,
svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
} }
sumf = svaddv_f32(ph32, svadd_f32_x(ph32, sumv0, sumv1)); sumf = svaddv_f32(ph32, svadd_f32_x(ph32, sumv0, sumv1));
@ -404,8 +405,8 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
const int32x4_t p_0 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0l), v0_0hs, v1_0h); const int32x4_t p_0 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0l), v0_0hs, v1_0h);
const int32x4_t p_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1l), v0_1hs, v1_1h); const int32x4_t p_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1l), v0_1hs, v1_1h);
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
} }
sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
@ -423,7 +424,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
} }
int sumi = sumi0 + sumi1; int sumi = sumi0 + sumi1;
sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d); sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
} }
*s = sumf; *s = sumf;
@ -464,10 +465,10 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
const block_q8_1 * GGML_RESTRICT b_y1 = &vy1[i]; const block_q8_1 * GGML_RESTRICT b_y1 = &vy1[i];
float32_t summs_t[4] = { float32_t summs_t[4] = {
GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y0->s), GGML_CPU_FP16_TO_FP32(b_x0->m) * GGML_CPU_FP16_TO_FP32(b_y0->s),
GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y0->s), GGML_CPU_FP16_TO_FP32(b_x1->m) * GGML_CPU_FP16_TO_FP32(b_y0->s),
GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y1->s), GGML_CPU_FP16_TO_FP32(b_x0->m) * GGML_CPU_FP16_TO_FP32(b_y1->s),
GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y1->s) GGML_CPU_FP16_TO_FP32(b_x1->m) * GGML_CPU_FP16_TO_FP32(b_y1->s)
}; };
summs0 = vaddq_f32(summs0, vld1q_f32(summs_t)); summs0 = vaddq_f32(summs0, vld1q_f32(summs_t));
@ -490,10 +491,10 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
// mmla into int32x4_t // mmla into int32x4_t
float32_t _scale[4] = { float32_t _scale[4] = {
GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d), GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y0->d),
GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d), GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y1->d),
GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d), GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y0->d),
GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d) GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y1->d)
}; };
float32x4_t scale = vld1q_f32(_scale); float32x4_t scale = vld1q_f32(_scale);
@ -539,7 +540,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
const block_q8_1 * GGML_RESTRICT y0 = &y[ib + 0]; const block_q8_1 * GGML_RESTRICT y0 = &y[ib + 0];
const block_q8_1 * GGML_RESTRICT y1 = &y[ib + 1]; const block_q8_1 * GGML_RESTRICT y1 = &y[ib + 1];
summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s) + GGML_FP16_TO_FP32(x1->m) * GGML_FP16_TO_FP32(y1->s); summs += GGML_CPU_FP16_TO_FP32(x0->m) * GGML_CPU_FP16_TO_FP32(y0->s) + GGML_CPU_FP16_TO_FP32(x1->m) * GGML_CPU_FP16_TO_FP32(y1->s);
const uint8x16_t m4b = vdupq_n_u8(0x0F); const uint8x16_t m4b = vdupq_n_u8(0x0F);
@ -562,8 +563,8 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
const int32x4_t p_0 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0l), v0_0h, v1_0h); const int32x4_t p_0 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0l), v0_0h, v1_0h);
const int32x4_t p_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1l), v0_1h, v1_1h); const int32x4_t p_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1l), v0_1h, v1_1h);
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
} }
sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs; sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs;
@ -582,7 +583,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
} }
int sumi = sumi0 + sumi1; int sumi = sumi0 + sumi1;
sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
} }
*s = sumf; *s = sumf;
@ -666,10 +667,10 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32( sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l), ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32( sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l), ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
} }
sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
@ -694,7 +695,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
} }
int sumi = sumi0 + sumi1; int sumi = sumi0 + sumi1;
sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi; sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
} }
*s = sumf; *s = sumf;
@ -739,8 +740,8 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
const uint8x16_t m4b = vdupq_n_u8(0x0F); const uint8x16_t m4b = vdupq_n_u8(0x0F);
summs0 += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s); summs0 += GGML_CPU_FP16_TO_FP32(x0->m) * GGML_CPU_FP16_TO_FP32(y0->s);
summs1 += GGML_FP16_TO_FP32(x1->m) * GGML_FP16_TO_FP32(y1->s); summs1 += GGML_CPU_FP16_TO_FP32(x1->m) * GGML_CPU_FP16_TO_FP32(y1->s);
// extract the 5th bit via lookup table ((b) << 4) // extract the 5th bit via lookup table ((b) << 4)
memcpy(&qh0, x0->qh, sizeof(qh0)); memcpy(&qh0, x0->qh, sizeof(qh0));
@ -784,10 +785,10 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32( sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l), ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32( sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l), ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
} }
sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs0 + summs1; sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs0 + summs1;
@ -812,7 +813,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
} }
int sumi = sumi0 + sumi1; int sumi = sumi0 + sumi1;
sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
} }
*s = sumf; *s = sumf;
@ -864,10 +865,10 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16); const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
float32_t _scale[4] = { float32_t _scale[4] = {
GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d), GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y0->d),
GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d), GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y1->d),
GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d), GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y0->d),
GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d) GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y1->d)
}; };
float32x4_t scale = vld1q_f32(_scale); float32x4_t scale = vld1q_f32(_scale);
@ -934,10 +935,10 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumv0 = svmla_n_f32_x(pl16, sumv0, svcvt_f32_s32_x(pl16, svadd_x(pl16, sumv0 = svmla_n_f32_x(pl16, sumv0, svcvt_f32_s32_x(pl16, svadd_x(pl16,
svdot_s32(svdup_n_s32(0), qx0_0, qy0_0), svdot_s32(svdup_n_s32(0), qx0_0, qy0_0),
svdot_s32(svdup_n_s32(0), qx0_1, qy0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); svdot_s32(svdup_n_s32(0), qx0_1, qy0_1))), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
sumv1 = svmla_n_f32_x(pl16, sumv1, svcvt_f32_s32_x(pl16, svadd_x(pl16, sumv1 = svmla_n_f32_x(pl16, sumv1, svcvt_f32_s32_x(pl16, svadd_x(pl16,
svdot_s32(svdup_n_s32(0), qx1_0, qy1_0), svdot_s32(svdup_n_s32(0), qx1_0, qy1_0),
svdot_s32(svdup_n_s32(0), qx1_1, qy1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); svdot_s32(svdup_n_s32(0), qx1_1, qy1_1))), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
} }
sumf = svaddv_f32(pl16, svadd_f32_x(pl16, sumv0, sumv1)); sumf = svaddv_f32(pl16, svadd_f32_x(pl16, sumv0, sumv1));
@ -960,9 +961,9 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs); const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs);
sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(), sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(),
svdot_s32(svdup_n_s32(0), qx0, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); svdot_s32(svdup_n_s32(0), qx0, qy0)), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(), sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(),
svdot_s32(svdup_n_s32(0), qx1, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); svdot_s32(svdup_n_s32(0), qx1, qy1)), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
} }
sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1)); sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
@ -1002,8 +1003,8 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
qy_64 = svadd_s8_x(svptrue_b8(), qy_32, qy_64); qy_64 = svadd_s8_x(svptrue_b8(), qy_32, qy_64);
// scale creation // scale creation
const float32_t deq1 = GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d); const float32_t deq1 = GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d);
const float32_t deq2 = GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d); const float32_t deq2 = GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d);
// duplicate deq1 in first half of vector and deq2 in second half of vector // duplicate deq1 in first half of vector and deq2 in second half of vector
const svfloat32_t temp = svdup_f32_m(svdup_f32_z(ph8, deq1), pl8, deq2); const svfloat32_t temp = svdup_f32_m(svdup_f32_z(ph8, deq1), pl8, deq2);
@ -1043,11 +1044,11 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32( sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
ggml_vdotq_s32(vdupq_n_s32(0), x0_0, y0_0), ggml_vdotq_s32(vdupq_n_s32(0), x0_0, y0_0),
ggml_vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); ggml_vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32( sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
ggml_vdotq_s32(vdupq_n_s32(0), x1_0, y1_0), ggml_vdotq_s32(vdupq_n_s32(0), x1_0, y1_0),
ggml_vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); ggml_vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
} }
sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
@ -1059,7 +1060,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumi += x[ib].qs[j]*y[ib].qs[j]; sumi += x[ib].qs[j]*y[ib].qs[j];
} }
sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)); sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
} }
*s = sumf; *s = sumf;
@ -1217,7 +1218,7 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
const int16x8_t ysum0 = vld1q_s16(y[i].bsums); const int16x8_t ysum0 = vld1q_s16(y[i].bsums);
const int16x8_t ysum1 = vld1q_s16(y[i].bsums + 8); const int16x8_t ysum1 = vld1q_s16(y[i].bsums + 8);
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
#if defined(__ARM_FEATURE_DOTPROD) #if defined(__ARM_FEATURE_DOTPROD)
sumi0 = vaddq_s32(sumi0, sumi1); sumi0 = vaddq_s32(sumi0, sumi1);
@ -1269,7 +1270,7 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
} }
} }
sumf += (float) sum * (GGML_FP16_TO_FP32(x[i].d) * y[i].d); sumf += (float) sum * (GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d);
} }
*s = sumf; *s = sumf;
@ -1362,7 +1363,7 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
const int16x8_t ysum0 = vld1q_s16(y[i].bsums); const int16x8_t ysum0 = vld1q_s16(y[i].bsums);
const int16x8_t ysum1 = vld1q_s16(y[i].bsums + 8); const int16x8_t ysum1 = vld1q_s16(y[i].bsums + 8);
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
#if defined(__ARM_FEATURE_DOTPROD) #if defined(__ARM_FEATURE_DOTPROD)
sumi0 = vaddq_s32(sumi0, sumi1); sumi0 = vaddq_s32(sumi0, sumi1);
@ -1393,7 +1394,7 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
} }
} }
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
sumf += (float) sumi * d; sumf += (float) sumi * d;
} }
@ -1425,9 +1426,9 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
switch (vector_length) { switch (vector_length) {
case 128: case 128:
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
svfloat32_t d_broad = svdup_n_f32((float32_t)d); svfloat32_t d_broad = svdup_n_f32((float32_t)d);
const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
svfloat32_t dmin_broad = svdup_n_f32((float32_t)dmin); svfloat32_t dmin_broad = svdup_n_f32((float32_t)dmin);
const uint8_t * GGML_RESTRICT q2 = x[i].qs; const uint8_t * GGML_RESTRICT q2 = x[i].qs;
@ -1570,9 +1571,9 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
case 256: case 256:
case 512: case 512:
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
svfloat32_t d_broad = svdup_n_f32((float32_t)d); svfloat32_t d_broad = svdup_n_f32((float32_t)d);
const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
svfloat32_t dmin_broad = svdup_n_f32((float32_t)dmin); svfloat32_t dmin_broad = svdup_n_f32((float32_t)dmin);
const uint8_t * GGML_RESTRICT q2 = x[i].qs; const uint8_t * GGML_RESTRICT q2 = x[i].qs;
@ -1671,8 +1672,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
float sum = 0; float sum = 0;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
const uint8_t * GGML_RESTRICT q2 = x[i].qs; const uint8_t * GGML_RESTRICT q2 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
@ -1742,8 +1743,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
summs += y[i].bsums[j] * (sc[j] >> 4); summs += y[i].bsums[j] * (sc[j] >> 4);
} }
const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
int isum = 0; int isum = 0;
int is = 0; int is = 0;
@ -1805,7 +1806,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const uint8_t * GGML_RESTRICT q3_sv = x[i].qs; const uint8_t * GGML_RESTRICT q3_sv = x[i].qs;
const uint8_t * GGML_RESTRICT qh_sv = x[i].hmask; const uint8_t * GGML_RESTRICT qh_sv = x[i].hmask;
@ -1981,7 +1982,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const uint8_t * GGML_RESTRICT q3 = x[i].qs; const uint8_t * GGML_RESTRICT q3 = x[i].qs;
const uint8_t * GGML_RESTRICT qh = x[i].hmask; const uint8_t * GGML_RESTRICT qh = x[i].hmask;
@ -2112,7 +2113,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];
@ -2258,18 +2259,18 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
bias[3] = vaddvq_s32(vaddq_s32(vmull_s16(vget_low_s16(y1_sums), vget_low_s16(x1_mins)), bias[3] = vaddvq_s32(vaddq_s32(vmull_s16(vget_low_s16(y1_sums), vget_low_s16(x1_mins)),
vmull_s16(vget_high_s16(y1_sums), vget_high_s16(x1_mins)))); vmull_s16(vget_high_s16(y1_sums), vget_high_s16(x1_mins))));
const float32x4_t dmins = { const float32x4_t dmins = {
GGML_FP16_TO_FP32(x0->dmin) * y0->d, GGML_CPU_FP16_TO_FP32(x0->dmin) * y0->d,
GGML_FP16_TO_FP32(x0->dmin) * y1->d, GGML_CPU_FP16_TO_FP32(x0->dmin) * y1->d,
GGML_FP16_TO_FP32(x1->dmin) * y0->d, GGML_CPU_FP16_TO_FP32(x1->dmin) * y0->d,
GGML_FP16_TO_FP32(x1->dmin) * y1->d, GGML_CPU_FP16_TO_FP32(x1->dmin) * y1->d,
}; };
vfsum = vmlsq_f32(vfsum, vcvtq_f32_s32(vld1q_s32(bias)), dmins); vfsum = vmlsq_f32(vfsum, vcvtq_f32_s32(vld1q_s32(bias)), dmins);
const float32x4_t superblock_scale = { const float32x4_t superblock_scale = {
GGML_FP16_TO_FP32(x0->d) * y0->d, GGML_CPU_FP16_TO_FP32(x0->d) * y0->d,
GGML_FP16_TO_FP32(x0->d) * y1->d, GGML_CPU_FP16_TO_FP32(x0->d) * y1->d,
GGML_FP16_TO_FP32(x1->d) * y0->d, GGML_CPU_FP16_TO_FP32(x1->d) * y0->d,
GGML_FP16_TO_FP32(x1->d) * y1->d, GGML_CPU_FP16_TO_FP32(x1->d) * y1->d,
}; };
vfsum = vmlaq_f32(vfsum, vcvtq_f32_s32(visum), superblock_scale); vfsum = vmlaq_f32(vfsum, vcvtq_f32_s32(visum), superblock_scale);
} }
@ -2289,8 +2290,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
float sumf = 0; float sumf = 0;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8)); const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));
@ -2377,8 +2378,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8)); const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));
@ -2478,9 +2479,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi; sumf -= dmin * sumi;
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];
@ -2520,8 +2521,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8)); const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));
@ -2630,9 +2631,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi; sumf -= dmin * sumi;
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];
@ -2827,10 +2828,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
const int32x4_t vibias = vmulq_n_s32(vld1q_s32(bias), 32); const int32x4_t vibias = vmulq_n_s32(vld1q_s32(bias), 32);
const float32x4_t superblock_scale = { const float32x4_t superblock_scale = {
GGML_FP16_TO_FP32(x0->d) * y0->d, GGML_CPU_FP16_TO_FP32(x0->d) * y0->d,
GGML_FP16_TO_FP32(x0->d) * y1->d, GGML_CPU_FP16_TO_FP32(x0->d) * y1->d,
GGML_FP16_TO_FP32(x1->d) * y0->d, GGML_CPU_FP16_TO_FP32(x1->d) * y0->d,
GGML_FP16_TO_FP32(x1->d) * y1->d, GGML_CPU_FP16_TO_FP32(x1->d) * y1->d,
}; };
visum = vsubq_s32(visum, vibias); visum = vsubq_s32(visum, vibias);
@ -2858,7 +2859,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
svuint8_t q6h_1, q6h_2, q6h_3, q6h_4; svuint8_t q6h_1, q6h_2, q6h_3, q6h_4;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d_all = GGML_FP16_TO_FP32(x[i].d); const float d_all = GGML_CPU_FP16_TO_FP32(x[i].d);
const uint8_t * GGML_RESTRICT q6 = x[i].ql; const uint8_t * GGML_RESTRICT q6 = x[i].ql;
const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint8_t * GGML_RESTRICT qh = x[i].qh;
@ -3011,7 +3012,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d_all = GGML_FP16_TO_FP32(x[i].d); const float d_all = GGML_CPU_FP16_TO_FP32(x[i].d);
const uint8_t * GGML_RESTRICT q6 = x[i].ql; const uint8_t * GGML_RESTRICT q6 = x[i].ql;
const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint8_t * GGML_RESTRICT qh = x[i].qh;
@ -3128,7 +3129,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];
@ -3199,7 +3200,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
float sumf = 0; float sumf = 0;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint16_t * GGML_RESTRICT q2 = x[i].qs; const uint16_t * GGML_RESTRICT q2 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
float sumf1 = 0, sumf2 = 0; float sumf1 = 0, sumf2 = 0;
@ -3234,7 +3235,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
float sumf = 0.f; float sumf = 0.f;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint16_t * GGML_RESTRICT q2 = x[i].qs; const uint16_t * GGML_RESTRICT q2 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
int32_t bsum = 0; int32_t bsum = 0;
@ -3284,7 +3285,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
float sumf = 0; float sumf = 0;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint16_t * GGML_RESTRICT q2 = x[i].qs; const uint16_t * GGML_RESTRICT q2 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
const uint8x8_t scales8 = vld1_u8(x[i].scales); const uint8x8_t scales8 = vld1_u8(x[i].scales);
@ -3329,7 +3330,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
float sumf = 0.f; float sumf = 0.f;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint16_t * GGML_RESTRICT q2 = x[i].qs; const uint16_t * GGML_RESTRICT q2 = x[i].qs;
const uint8_t * GGML_RESTRICT sc = x[i].scales; const uint8_t * GGML_RESTRICT sc = x[i].scales;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
@ -3398,7 +3399,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
float sumf = 0; float sumf = 0;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * GGML_RESTRICT qs = x[i].qs; const uint8_t * GGML_RESTRICT qs = x[i].qs;
const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint8_t * GGML_RESTRICT qh = x[i].qh;
@ -3458,7 +3459,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
float sumf = 0; float sumf = 0;
for (int i = 0; i < nb; i++) { for (int i = 0; i < nb; i++) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const int8_t * q8 = y[i].qs; const int8_t * q8 = y[i].qs;
const uint8_t * qs = x[i].qs; const uint8_t * qs = x[i].qs;
const uint8_t * qh = x[i].qh; const uint8_t * qh = x[i].qh;
@ -3521,7 +3522,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
float sumf = 0; float sumf = 0;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * GGML_RESTRICT q3 = x[i].qs; const uint8_t * GGML_RESTRICT q3 = x[i].qs;
const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4; const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
@ -3557,7 +3558,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
float sumf = 0.f; float sumf = 0.f;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * GGML_RESTRICT q3 = x[i].qs; const uint8_t * GGML_RESTRICT q3 = x[i].qs;
const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4; const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
@ -3630,7 +3631,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
float sumf = 0; float sumf = 0;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * GGML_RESTRICT qs = x[i].qs; const uint8_t * GGML_RESTRICT qs = x[i].qs;
const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint8_t * GGML_RESTRICT qh = x[i].qh;
const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs; const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs;
@ -3691,7 +3692,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
float sumf = 0.f; float sumf = 0.f;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * GGML_RESTRICT qs = x[i].qs; const uint8_t * GGML_RESTRICT qs = x[i].qs;
const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint8_t * GGML_RESTRICT qh = x[i].qh;
const uint8_t * GGML_RESTRICT signs = x[i].signs; const uint8_t * GGML_RESTRICT signs = x[i].signs;
@ -3786,7 +3787,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
} }
sumf += y[i].d * GGML_FP16_TO_FP32(x[i].d) * (sumi1 + sumi2 + IQ1S_DELTA * sumi3); sumf += y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d) * (sumi1 + sumi2 + IQ1S_DELTA * sumi3);
} }
*s = sumf; *s = sumf;
@ -3817,7 +3818,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
qs += 4; qs += 4;
} }
sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1); sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
} }
*s = sumf; *s = sumf;
@ -3905,7 +3906,7 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
} }
sumf += y[i].d * GGML_FP16_TO_FP32(scale.f16) * (vaddvq_s32(sumi1) + IQ1M_DELTA * vaddvq_s32(sumi2)); sumf += y[i].d * GGML_CPU_FP16_TO_FP32(scale.f16) * (vaddvq_s32(sumi1) + IQ1M_DELTA * vaddvq_s32(sumi2));
} }
*s = sumf; *s = sumf;
@ -3952,7 +3953,7 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
qh += 2; qh += 2;
} }
sumf += GGML_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2); sumf += GGML_CPU_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
} }
*s = sumf; *s = sumf;
@ -4003,13 +4004,13 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
prod_2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[2], q8b.val[2]), q4b.val[3], q8b.val[3]); prod_2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[2], q8b.val[2]), q4b.val[3], q8b.val[3]);
sumf += sumf +=
GGML_FP16_TO_FP32(x[ib+0].d) * GGML_FP16_TO_FP32(y[ib + 0].d) * vaddvq_s32(prod_1) + GGML_CPU_FP16_TO_FP32(x[ib+0].d) * GGML_CPU_FP16_TO_FP32(y[ib + 0].d) * vaddvq_s32(prod_1) +
GGML_FP16_TO_FP32(x[ib+1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) * vaddvq_s32(prod_2); GGML_CPU_FP16_TO_FP32(x[ib+1].d) * GGML_CPU_FP16_TO_FP32(y[ib + 1].d) * vaddvq_s32(prod_2);
} }
#endif #endif
for (; ib < nb; ++ib) { for (; ib < nb; ++ib) {
const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d); const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
int sumi1 = 0, sumi2 = 0; int sumi1 = 0, sumi2 = 0;
for (int j = 0; j < QK4_NL/2; ++j) { for (int j = 0; j < QK4_NL/2; ++j) {
sumi1 += y[ib].qs[j+ 0] * kvalues_iq4nl[x[ib].qs[j] & 0xf]; sumi1 += y[ib].qs[j+ 0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
@ -4071,7 +4072,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
} }
sumf += GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2); sumf += GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2);
} }
*s = sumf; *s = sumf;
@ -4079,7 +4080,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
#else #else
float sumf = 0; float sumf = 0;
for (int ibl = 0; ibl < nb; ++ibl) { for (int ibl = 0; ibl < nb; ++ibl) {
const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d; const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
uint16_t h = x[ibl].scales_h; uint16_t h = x[ibl].scales_h;
const uint8_t * qs = x[ibl].qs; const uint8_t * qs = x[ibl].qs;
const int8_t * q8 = y[ibl].qs; const int8_t * q8 = y[ibl].qs;

25
ggml/src/ggml-cpu/arch/arm/repack.cpp
View File

@ -6,6 +6,7 @@
#include "ggml-impl.h" #include "ggml-impl.h"
#include "ggml-cpu.h" #include "ggml-cpu.h"
#include "ggml-cpu-impl.h" #include "ggml-cpu-impl.h"
#include "simd-mappings.h"
#include "traits.h" #include "traits.h"
#include <cmath> #include <cmath>
@ -51,7 +52,7 @@ void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTR
const float d = amax / ((1 << 7) - 1); const float d = amax / ((1 << 7) - 1);
id[row_iter] = d ? 1.0f / d : 0.0f; id[row_iter] = d ? 1.0f / d : 0.0f;
y[i].d[row_iter] = GGML_FP32_TO_FP16(d); y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
} }
for (int j = 0; j < 8; j++) { for (int j = 0; j < 8; j++) {
@ -102,7 +103,7 @@ void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTR
const float d = amax / ((1 << 7) - 1); const float d = amax / ((1 << 7) - 1);
id[row_iter] = d ? 1.0f / d : 0.0f; id[row_iter] = d ? 1.0f / d : 0.0f;
y[i].d[row_iter] = GGML_FP32_TO_FP16(d); y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
} }
for (int j = 0; j < QK8_0 * 4; j++) { for (int j = 0; j < QK8_0 * 4; j++) {
@ -145,7 +146,7 @@ void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR
const float d = amax / ((1 << 7) - 1); const float d = amax / ((1 << 7) - 1);
id[row_iter] = d ? 1.0f / d : 0.0f; id[row_iter] = d ? 1.0f / d : 0.0f;
y[i].d[row_iter] = GGML_FP32_TO_FP16(d); y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
} }
for (int j = 0; j < 4; j++) { for (int j = 0; j < 4; j++) {
@ -221,7 +222,7 @@ void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR
const float d = amax / ((1 << 7) - 1); const float d = amax / ((1 << 7) - 1);
id[row_iter] = d ? 1.0f / d : 0.0f; id[row_iter] = d ? 1.0f / d : 0.0f;
y[i].d[row_iter] = GGML_FP32_TO_FP16(d); y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
} }
for (int j = 0; j < QK8_0 * 4; j++) { for (int j = 0; j < QK8_0 * 4; j++) {
@ -311,7 +312,7 @@ void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
} }
sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
} }
} }
} }
@ -399,7 +400,7 @@ void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
} }
sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
} }
} }
} }
@ -514,7 +515,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
} }
sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
} }
} }
} }
@ -608,7 +609,7 @@ void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])); sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
} }
sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
} }
} }
} }
@ -1117,7 +1118,7 @@ void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
(v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
} }
sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]); sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
} }
} }
} }
@ -1570,7 +1571,7 @@ void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
(v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
} }
sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]); sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
} }
} }
} }
@ -2039,7 +2040,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
(v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
} }
sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]); sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
} }
} }
} }
@ -2147,7 +2148,7 @@ void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
(v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])); (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4]));
} }
sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]); sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
} }
} }
} }

105
ggml/src/ggml-cpu/arch/loongarch/quants.c
View File

@ -3,6 +3,7 @@
#include "ggml-quants.h" #include "ggml-quants.h"
#include "ggml-impl.h" #include "ggml-impl.h"
#include "ggml-cpu.h" #include "ggml-cpu.h"
#include "simd-mappings.h"
#include "../../quants.h" #include "../../quants.h"
#include "../../ggml-cpu-impl.h" #include "../../ggml-cpu-impl.h"
@ -474,7 +475,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
// Quantize these floats // Quantize these floats
const float d = max_scalar / 127.f; const float d = max_scalar / 127.f;
y[i].d = GGML_FP32_TO_FP16(d); y[i].d = GGML_CPU_FP32_TO_FP16(d);
const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f; const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f;
const __m256 mul = (__m256)__lasx_xvreplfr2vr_s( id ); const __m256 mul = (__m256)__lasx_xvreplfr2vr_s( id );
@ -548,7 +549,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
// Quantize these floats // Quantize these floats
const float d = max_scalar / 127.f; const float d = max_scalar / 127.f;
y[i].d = GGML_FP32_TO_FP16(d); y[i].d = GGML_CPU_FP32_TO_FP16(d);
const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f; const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f;
const __m256 mul = __lasx_xvreplfr2vr_s( id ); const __m256 mul = __lasx_xvreplfr2vr_s( id );
@ -576,7 +577,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
// Compute the sum of the quants and set y[i].s // Compute the sum of the quants and set y[i].s
const __m128i s0 = __lsx_vadd_w(__lsx_vadd_w(ni0, ni1), __lsx_vadd_w(ni2, ni3)); const __m128i s0 = __lsx_vadd_w(__lsx_vadd_w(ni0, ni1), __lsx_vadd_w(ni2, ni3));
const __m128i s1 = __lsx_vadd_w(__lsx_vadd_w(ni4, ni5), __lsx_vadd_w(ni6, ni7)); const __m128i s1 = __lsx_vadd_w(__lsx_vadd_w(ni4, ni5), __lsx_vadd_w(ni6, ni7));
y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_4(__lsx_vadd_w(s0, s1))); y[i].s = GGML_CPU_FP32_TO_FP16(d * hsum_i32_4(__lsx_vadd_w(s0, s1)));
// Convert int32 to int16 // Convert int32 to int16
ni0 = lsx_packs_w( ni0, ni1 ); ni0 = lsx_packs_w( ni0, ni1 );
@ -667,7 +668,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
// Main loop // Main loop
for (; ib < nb; ++ib) { for (; ib < nb; ++ib) {
/* Compute combined scale for the block */ /* Compute combined scale for the block */
const __m256 d = __lasx_xvreplfr2vr_s( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) ); const __m256 d = __lasx_xvreplfr2vr_s( GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d) );
__m256i qx = bytes_from_nibbles_32(x[ib].qs); __m256i qx = bytes_from_nibbles_32(x[ib].qs);
@ -699,7 +700,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (; ib + 1 < nb; ib += 2) { for (; ib + 1 < nb; ib += 2) {
// Compute combined scale for the block 0 and 1 // Compute combined scale for the block 0 and 1
const __m128 d_0_1 = (__m128)__lsx_vreplgr2vr_w( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) ); const __m128 d_0_1 = (__m128)__lsx_vreplgr2vr_w( GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d) );
const __m128i tmp_0_1 = __lsx_vld((const __m128i *)x[ib].qs, 0); const __m128i tmp_0_1 = __lsx_vld((const __m128i *)x[ib].qs, 0);
@ -717,7 +718,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
//_mm_prefetch(&y[ib] + 2 * sizeof(block_q8_0), _MM_HINT_T0); //_mm_prefetch(&y[ib] + 2 * sizeof(block_q8_0), _MM_HINT_T0);
// Compute combined scale for the block 2 and 3 // Compute combined scale for the block 2 and 3
const __m128 d_2_3 = (__m128)__lsx_vreplgr2vr_w( GGML_FP16_TO_FP32(x[ib + 1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) ); const __m128 d_2_3 = (__m128)__lsx_vreplgr2vr_w( GGML_CPU_FP16_TO_FP32(x[ib + 1].d) * GGML_CPU_FP16_TO_FP32(y[ib + 1].d) );
const __m128i tmp_2_3 = __lsx_vld((const __m128i *)x[ib + 1].qs, 0); const __m128i tmp_2_3 = __lsx_vld((const __m128i *)x[ib + 1].qs, 0);
@ -766,7 +767,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
} }
int sumi = sumi0 + sumi1; int sumi = sumi0 + sumi1;
sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d); sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
} }
*s = sumf; *s = sumf;
@ -797,10 +798,10 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
// Main loop // Main loop
for (; ib < nb; ++ib) { for (; ib < nb; ++ib) {
const float d0 = GGML_FP16_TO_FP32(x[ib].d); const float d0 = GGML_CPU_FP16_TO_FP32(x[ib].d);
const float d1 = GGML_FP16_TO_FP32(y[ib].d); const float d1 = GGML_CPU_FP16_TO_FP32(y[ib].d);
summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s); summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);
const __m256 d0v = __lasx_xvreplfr2vr_s( d0 ); const __m256 d0v = __lasx_xvreplfr2vr_s( d0 );
const __m256 d1v = __lasx_xvreplfr2vr_s( d1 ); const __m256 d1v = __lasx_xvreplfr2vr_s( d1 );
@ -834,7 +835,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
} }
int sumi = sumi0 + sumi1; int sumi = sumi0 + sumi1;
sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
} }
*s = sumf; *s = sumf;
@ -865,7 +866,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
// Main loop // Main loop
for (; ib < nb; ++ib) { for (; ib < nb; ++ib) {
/* Compute combined scale for the block */ /* Compute combined scale for the block */
const __m256 d = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)); //FIXME const __m256 d = __lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d)); //FIXME
__m256i qx = bytes_from_nibbles_32(x[ib].qs); __m256i qx = bytes_from_nibbles_32(x[ib].qs);
__m256i bxhi = bytes_from_bits_32(x[ib].qh); __m256i bxhi = bytes_from_bits_32(x[ib].qh);
@ -902,7 +903,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
} }
int sumi = sumi0 + sumi1; int sumi = sumi0 + sumi1;
sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi; sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
} }
*s = sumf; *s = sumf;
@ -934,16 +935,16 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
// Main loop // Main loop
for (; ib < nb; ++ib) { for (; ib < nb; ++ib) {
const __m256 dx = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ib].d)); const __m256 dx = __lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(x[ib].d));
summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s); summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);
__m256i qx = bytes_from_nibbles_32(x[ib].qs); __m256i qx = bytes_from_nibbles_32(x[ib].qs);
__m256i bxhi = bytes_from_bits_32(x[ib].qh); __m256i bxhi = bytes_from_bits_32(x[ib].qh);
bxhi = __lasx_xvand_v(bxhi, __lasx_xvreplgr2vr_b(0x10)); bxhi = __lasx_xvand_v(bxhi, __lasx_xvreplgr2vr_b(0x10));
qx = __lasx_xvor_v(qx, bxhi); qx = __lasx_xvor_v(qx, bxhi);
const __m256 dy = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[ib].d)); const __m256 dy = __lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(y[ib].d));
const __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0); const __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);
const __m256 q = mul_sum_us8_pairs_float(qx, qy); const __m256 q = mul_sum_us8_pairs_float(qx, qy);
@ -973,7 +974,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
} }
int sumi = sumi0 + sumi1; int sumi = sumi0 + sumi1;
sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
} }
*s = sumf; *s = sumf;
@ -1003,7 +1004,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
// Main loop // Main loop
for (; ib < nb; ++ib) { for (; ib < nb; ++ib) {
// Compute combined scale for the block // Compute combined scale for the block
const __m256 d = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)); const __m256 d = __lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
__m256i qx = __lasx_xvld((const __m256i *)x[ib].qs, 0); __m256i qx = __lasx_xvld((const __m256i *)x[ib].qs, 0);
__m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0); __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);
@ -1023,7 +1024,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumi += x[ib].qs[j]*y[ib].qs[j]; sumi += x[ib].qs[j]*y[ib].qs[j];
} }
sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)); sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
} }
*s = sumf; *s = sumf;
@ -1047,8 +1048,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
const uint8_t * GGML_RESTRICT q2 = x[i].qs; const uint8_t * GGML_RESTRICT q2 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
@ -1116,8 +1117,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
summs += y[i].bsums[j] * (sc[j] >> 4); summs += y[i].bsums[j] * (sc[j] >> 4);
} }
const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
int isum = 0; int isum = 0;
int is = 0; int is = 0;
@ -1170,7 +1171,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const uint8_t * GGML_RESTRICT q3 = x[i].qs; const uint8_t * GGML_RESTRICT q3 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
// Set up scales // Set up scales
@ -1294,7 +1295,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];
@ -1330,8 +1331,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
memcpy(utmp, x[i].scales, 12); memcpy(utmp, x[i].scales, 12);
utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4); utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
@ -1438,9 +1439,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi; sumf -= dmin * sumi;
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];
@ -1477,8 +1478,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
const uint8_t * GGML_RESTRICT q5 = x[i].qs; const uint8_t * GGML_RESTRICT q5 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
memcpy(utmp, x[i].scales, 12); memcpy(utmp, x[i].scales, 12);
utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4); utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
@ -1593,9 +1594,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi; sumf -= dmin * sumi;
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];
@ -1624,7 +1625,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const uint8_t * GGML_RESTRICT q4 = x[i].ql; const uint8_t * GGML_RESTRICT q4 = x[i].ql;
const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint8_t * GGML_RESTRICT qh = x[i].qh;
@ -1713,7 +1714,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];
@ -1780,7 +1781,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
__m256 accumf = (__m256)__lasx_xvldi(0); __m256 accumf = (__m256)__lasx_xvldi(0);
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint16_t * GGML_RESTRICT q2 = x[i].qs; const uint16_t * GGML_RESTRICT q2 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
__m256i sumi1 = __lasx_xvldi(0); __m256i sumi1 = __lasx_xvldi(0);
@ -1820,7 +1821,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
float sumf = 0.f; float sumf = 0.f;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint16_t * GGML_RESTRICT q2 = x[i].qs; const uint16_t * GGML_RESTRICT q2 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
int32_t bsum = 0; int32_t bsum = 0;
@ -1895,7 +1896,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
__m256 accumf = (__m256)__lasx_xvldi(0); __m256 accumf = (__m256)__lasx_xvldi(0);
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint16_t * GGML_RESTRICT q2 = x[i].qs; const uint16_t * GGML_RESTRICT q2 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
@ -1980,7 +1981,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
float sumf = 0.f; float sumf = 0.f;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint16_t * GGML_RESTRICT q2 = x[i].qs; const uint16_t * GGML_RESTRICT q2 = x[i].qs;
const uint8_t * GGML_RESTRICT sc = x[i].scales; const uint8_t * GGML_RESTRICT sc = x[i].scales;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
@ -2049,7 +2050,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
__m256 accumf = (__m256)__lasx_xvldi(0); __m256 accumf = (__m256)__lasx_xvldi(0);
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * GGML_RESTRICT qs = x[i].qs; const uint8_t * GGML_RESTRICT qs = x[i].qs;
const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint8_t * GGML_RESTRICT qh = x[i].qh;
const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].qs + QK_K/8); const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].qs + QK_K/8);
@ -2108,7 +2109,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
float sumf = 0; float sumf = 0;
for (int i = 0; i < nb; i++) { for (int i = 0; i < nb; i++) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const int8_t * q8 = y[i].qs; const int8_t * q8 = y[i].qs;
const uint8_t * qs = x[i].qs; const uint8_t * qs = x[i].qs;
const uint8_t * qh = x[i].qh; const uint8_t * qh = x[i].qh;
@ -2168,7 +2169,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
__m256 accumf = (__m256)__lasx_xvldi(0); __m256 accumf = (__m256)__lasx_xvldi(0);
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * GGML_RESTRICT q3 = x[i].qs; const uint8_t * GGML_RESTRICT q3 = x[i].qs;
const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4; const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
@ -2213,7 +2214,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
float sumf = 0.f; float sumf = 0.f;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * GGML_RESTRICT q3 = x[i].qs; const uint8_t * GGML_RESTRICT q3 = x[i].qs;
const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4; const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
@ -2279,7 +2280,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
__m256 accumf = (__m256)__lasx_xvldi(0); __m256 accumf = (__m256)__lasx_xvldi(0);
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * GGML_RESTRICT qs = x[i].qs; const uint8_t * GGML_RESTRICT qs = x[i].qs;
const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint8_t * GGML_RESTRICT qh = x[i].qh;
const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs; const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs;
@ -2340,7 +2341,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
float sumf = 0.f; float sumf = 0.f;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * GGML_RESTRICT qs = x[i].qs; const uint8_t * GGML_RESTRICT qs = x[i].qs;
const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint8_t * GGML_RESTRICT qh = x[i].qh;
const uint8_t * GGML_RESTRICT signs = x[i].signs; const uint8_t * GGML_RESTRICT signs = x[i].signs;
@ -2451,7 +2452,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
+ (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2; + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
} }
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
accum = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), accum); accum = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), accum);
accum1 += d * sumi1; accum1 += d * sumi1;
} }
@ -2484,7 +2485,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
qs += 4; qs += 4;
} }
sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1); sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
} }
*s = sumf; *s = sumf;
@ -2530,9 +2531,9 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2); const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
const __m256i p_1 = lasx_madd_h(p16_1, mone); const __m256i p_1 = lasx_madd_h(p16_1, mone);
const __m256i p_2 = lasx_madd_h(p16_2, mone); const __m256i p_2 = lasx_madd_h(p16_2, mone);
accum1 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[ib + 0].d)*GGML_FP16_TO_FP32(x[ib + 0].d)), accum1 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(y[ib + 0].d)*GGML_CPU_FP16_TO_FP32(x[ib + 0].d)),
__lasx_xvffint_s_w(p_1), accum1); __lasx_xvffint_s_w(p_1), accum1);
accum2 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[ib + 1].d)*GGML_FP16_TO_FP32(x[ib + 1].d)), accum2 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(y[ib + 1].d)*GGML_CPU_FP16_TO_FP32(x[ib + 1].d)),
__lasx_xvffint_s_w(p_2), accum2); __lasx_xvffint_s_w(p_2), accum2);
} }
@ -2540,7 +2541,7 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
#endif #endif
for (; ib < nb; ++ib) { for (; ib < nb; ++ib) {
const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d); const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
int sumi1 = 0, sumi2 = 0; int sumi1 = 0, sumi2 = 0;
for (int j = 0; j < QK4_NL/2; ++j) { for (int j = 0; j < QK4_NL/2; ++j) {
sumi1 += y[ib].qs[j+ 0] * kvalues_iq4nl[x[ib].qs[j] & 0xf]; sumi1 += y[ib].qs[j+ 0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
@ -2595,7 +2596,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
sumi1 = __lasx_xvadd_w(p_1, sumi1); sumi1 = __lasx_xvadd_w(p_1, sumi1);
sumi2 = __lasx_xvadd_w(p_2, sumi2); sumi2 = __lasx_xvadd_w(p_2, sumi2);
} }
accum = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ibl].d)*y[ibl].d), accum = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
__lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accum); __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accum);
} }
@ -2604,7 +2605,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
#else #else
float sumf = 0; float sumf = 0;
for (int ibl = 0; ibl < nb; ++ibl) { for (int ibl = 0; ibl < nb; ++ibl) {
const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d; const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
uint16_t h = x[ibl].scales_h; uint16_t h = x[ibl].scales_h;
const uint8_t * qs = x[ibl].qs; const uint8_t * qs = x[ibl].qs;
const int8_t * q8 = y[ibl].qs; const int8_t * q8 = y[ibl].qs;

111
ggml/src/ggml-cpu/arch/powerpc/quants.c
View File

@ -3,6 +3,7 @@
#include "ggml-quants.h" #include "ggml-quants.h"
#include "ggml-impl.h" #include "ggml-impl.h"
#include "ggml-cpu.h" #include "ggml-cpu.h"
#include "simd-mappings.h"
#include "../../quants.h" #include "../../quants.h"
#include "../../ggml-cpu-impl.h" #include "../../ggml-cpu-impl.h"
@ -67,7 +68,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
const float id = d ? 1.0f/d : 0.0f; const float id = d ? 1.0f/d : 0.0f;
const vector float vid = vec_splats(id); const vector float vid = vec_splats(id);
y[i].d = GGML_FP32_TO_FP16(d); y[i].d = GGML_CPU_FP32_TO_FP16(d);
for (int j = 0; j < 8; j++) { for (int j = 0; j < 8; j++) {
const vector float v = vec_round(vec_mul(srcv[j], vid)); const vector float v = vec_round(vec_mul(srcv[j], vid));
@ -112,7 +113,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
const float id = d ? 1.0f/d : 0.0f; const float id = d ? 1.0f/d : 0.0f;
const vector float vid = vec_splats(id); const vector float vid = vec_splats(id);
y[i].d = GGML_FP32_TO_FP16(d); y[i].d = GGML_CPU_FP32_TO_FP16(d);
vector int accv = vec_splats(0); vector int accv = vec_splats(0);
@ -127,7 +128,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
accv = vec_add(accv, vec_sld(accv, accv, 4)); accv = vec_add(accv, vec_sld(accv, accv, 4));
accv = vec_add(accv, vec_sld(accv, accv, 8)); accv = vec_add(accv, vec_sld(accv, accv, 8));
y[i].s = GGML_FP32_TO_FP16(d * vec_extract(accv, 0)); y[i].s = GGML_CPU_FP32_TO_FP16(d * vec_extract(accv, 0));
} }
#else #else
@ -170,8 +171,8 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
__builtin_prefetch(x[ib].qs, 0, 1); __builtin_prefetch(x[ib].qs, 0, 1);
__builtin_prefetch(y[ib].qs, 0, 1); __builtin_prefetch(y[ib].qs, 0, 1);
vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d)); vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d));
vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d)); vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d));
vector float vd = vec_mul(vxd, vyd); vector float vd = vec_mul(vxd, vyd);
vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs); vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
@ -214,7 +215,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
} }
int sumi = sumi0 + sumi1; int sumi = sumi0 + sumi1;
sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d); sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
} }
*s = sumf; *s = sumf;
@ -249,12 +250,12 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
__builtin_prefetch(x[ib].qs, 0, 1); __builtin_prefetch(x[ib].qs, 0, 1);
__builtin_prefetch(y[ib].qs, 0, 1); __builtin_prefetch(y[ib].qs, 0, 1);
vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d)); vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d));
vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d)); vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d));
vector float vd = vec_mul(vxd, vyd); vector float vd = vec_mul(vxd, vyd);
vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[ib].m)); vector float vxmin = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].m));
vector float vys = {GGML_FP16_TO_FP32(y[ib].s), 0.0f, 0.0f, 0.0f}; vector float vys = {GGML_CPU_FP16_TO_FP32(y[ib].s), 0.0f, 0.0f, 0.0f};
vsumf0 = vec_madd(vxmin, vys, vsumf0); vsumf0 = vec_madd(vxmin, vys, vsumf0);
vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs); vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
@ -291,7 +292,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
} }
int sumi = sumi0 + sumi1; int sumi = sumi0 + sumi1;
sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
} }
*s = sumf; *s = sumf;
@ -326,8 +327,8 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
__builtin_prefetch(x[ib].qs, 0, 1); __builtin_prefetch(x[ib].qs, 0, 1);
__builtin_prefetch(y[ib].qs, 0, 1); __builtin_prefetch(y[ib].qs, 0, 1);
vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d)); vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d));
vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d)); vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d));
vector float vd = vec_mul(vxd, vyd); vector float vd = vec_mul(vxd, vyd);
vector signed long long aux64x2_0 = {(uint64_t)(table_b2b_1[x[ib].qh[0]]), (uint64_t)(table_b2b_1[x[ib].qh[1]])}; vector signed long long aux64x2_0 = {(uint64_t)(table_b2b_1[x[ib].qh[0]]), (uint64_t)(table_b2b_1[x[ib].qh[1]])};
@ -379,7 +380,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
} }
int sumi = sumi0 + sumi1; int sumi = sumi0 + sumi1;
sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi; sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
} }
*s = sumf; *s = sumf;
@ -415,12 +416,12 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
__builtin_prefetch(x[ib].qs, 0, 1); __builtin_prefetch(x[ib].qs, 0, 1);
__builtin_prefetch(y[ib].qs, 0, 1); __builtin_prefetch(y[ib].qs, 0, 1);
vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d)); vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d));
vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d)); vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d));
vector float vd = vec_mul(vxd, vyd); vector float vd = vec_mul(vxd, vyd);
vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[ib].m)); vector float vxmin = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].m));
vector float vys = {GGML_FP16_TO_FP32(y[ib].s), 0.f, 0.f, 0.f}; vector float vys = {GGML_CPU_FP16_TO_FP32(y[ib].s), 0.f, 0.f, 0.f};
vsumf0 = vec_madd(vxmin, vys, vsumf0); vsumf0 = vec_madd(vxmin, vys, vsumf0);
vector unsigned long long aux64x2_0 = {(uint64_t)(table_b2b_0[x[ib].qh[0]]), (uint64_t)(table_b2b_0[x[ib].qh[1]])}; vector unsigned long long aux64x2_0 = {(uint64_t)(table_b2b_0[x[ib].qh[0]]), (uint64_t)(table_b2b_0[x[ib].qh[1]])};
@ -470,7 +471,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
} }
int sumi = sumi0 + sumi1; int sumi = sumi0 + sumi1;
sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
} }
*s = sumf; *s = sumf;
@ -502,8 +503,8 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
__builtin_prefetch(x[ib].qs, 0, 1); __builtin_prefetch(x[ib].qs, 0, 1);
__builtin_prefetch(y[ib].qs, 0, 1); __builtin_prefetch(y[ib].qs, 0, 1);
vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d)); vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d));
vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d)); vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d));
vector float vd = vec_mul(vxd, vyd); vector float vd = vec_mul(vxd, vyd);
vector signed char q8x0 = vec_xl( 0, x[ib].qs); vector signed char q8x0 = vec_xl( 0, x[ib].qs);
@ -542,7 +543,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumi += x[ib].qs[j]*y[ib].qs[j]; sumi += x[ib].qs[j]*y[ib].qs[j];
} }
sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)); sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
} }
*s = sumf; *s = sumf;
@ -574,11 +575,11 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
vector float vsumf3 = vec_splats(0.0f); vector float vsumf3 = vec_splats(0.0f);
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
vector float vyd = vec_splats(y[i].d); vector float vyd = vec_splats(y[i].d);
vector float vd = vec_mul(vxd, vyd); vector float vd = vec_mul(vxd, vyd);
vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin)); vector float vxmin = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].dmin));
vector float vdmin = vec_mul(vxmin, vyd); vector float vdmin = vec_mul(vxmin, vyd);
vector signed short q8ysums0 = vec_xl( 0, y[i].bsums); vector signed short q8ysums0 = vec_xl( 0, y[i].bsums);
@ -708,8 +709,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
summs += y[i].bsums[j] * (sc[j] >> 4); summs += y[i].bsums[j] * (sc[j] >> 4);
} }
const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
int isum = 0; int isum = 0;
int is = 0; int is = 0;
@ -770,7 +771,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
vector float vsumf3 = vec_splats(0.0f); vector float vsumf3 = vec_splats(0.0f);
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
vector float vyd = vec_splats(y[i].d); vector float vyd = vec_splats(y[i].d);
vector float vd = vec_mul(vxd, vyd); vector float vd = vec_mul(vxd, vyd);
@ -962,7 +963,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];
@ -1005,11 +1006,11 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
vector float vsumf3 = vec_splats(0.0f); vector float vsumf3 = vec_splats(0.0f);
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
vector float vyd = vec_splats(y[i].d); vector float vyd = vec_splats(y[i].d);
vector float vd = vec_mul(vxd, vyd); vector float vd = vec_mul(vxd, vyd);
vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin)); vector float vxmin = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].dmin));
vector float vdmin = vec_mul(vxmin, vyd); vector float vdmin = vec_mul(vxmin, vyd);
vector signed short q8ysums0 = vec_xl( 0, y[i].bsums); vector signed short q8ysums0 = vec_xl( 0, y[i].bsums);
@ -1177,9 +1178,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi; sumf -= dmin * sumi;
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];
@ -1222,11 +1223,11 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
vector float vsumf3 = vec_splats(0.0f); vector float vsumf3 = vec_splats(0.0f);
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
vector float vyd = vec_splats(y[i].d); vector float vyd = vec_splats(y[i].d);
vector float vd = vec_mul(vxd, vyd); vector float vd = vec_mul(vxd, vyd);
vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin)); vector float vxmin = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].dmin));
vector float vdmin = vec_mul(vxmin, vyd); vector float vdmin = vec_mul(vxmin, vyd);
UNUSED(kmask1); UNUSED(kmask1);
@ -1394,9 +1395,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi; sumf -= dmin * sumi;
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];
@ -1432,7 +1433,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
vector float vsumf3 = vec_splats(0.0f); vector float vsumf3 = vec_splats(0.0f);
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
vector float vyd = vec_splats(y[i].d); vector float vyd = vec_splats(y[i].d);
vector float vd = vec_mul(vxd, vyd); vector float vd = vec_mul(vxd, vyd);
@ -1591,7 +1592,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];
@ -1659,7 +1660,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs; const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
vector float vyd = vec_splats(y[i].d); vector float vyd = vec_splats(y[i].d);
vector float vd = vec_mul(vxd, vyd); vector float vd = vec_mul(vxd, vyd);
@ -1742,7 +1743,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
float sumf = 0.f; float sumf = 0.f;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint16_t * GGML_RESTRICT q2 = x[i].qs; const uint16_t * GGML_RESTRICT q2 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
int32_t bsum = 0; int32_t bsum = 0;
@ -1790,7 +1791,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs; const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
vector float vyd = vec_splats(y[i].d); vector float vyd = vec_splats(y[i].d);
vector float vd = vec_mul(vxd, vyd); vector float vd = vec_mul(vxd, vyd);
@ -1871,7 +1872,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
float sumf = 0.f; float sumf = 0.f;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint16_t * GGML_RESTRICT q2 = x[i].qs; const uint16_t * GGML_RESTRICT q2 = x[i].qs;
const uint8_t * GGML_RESTRICT sc = x[i].scales; const uint8_t * GGML_RESTRICT sc = x[i].scales;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
@ -1939,7 +1940,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
const vector signed char mask2 = (vector signed char)vec_xl( 0, k_mask2); const vector signed char mask2 = (vector signed char)vec_xl( 0, k_mask2);
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
vector float vyd = vec_splats(y[i].d); vector float vyd = vec_splats(y[i].d);
vector float vd = vec_mul(vxd, vyd); vector float vd = vec_mul(vxd, vyd);
@ -2033,7 +2034,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
float sumf = 0; float sumf = 0;
for (int i = 0; i < nb; i++) { for (int i = 0; i < nb; i++) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const int8_t * q8 = y[i].qs; const int8_t * q8 = y[i].qs;
const uint8_t * qs = x[i].qs; const uint8_t * qs = x[i].qs;
const uint8_t * qh = x[i].qh; const uint8_t * qh = x[i].qh;
@ -2096,7 +2097,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
vector float vsumf3 = vec_splats(0.0f); vector float vsumf3 = vec_splats(0.0f);
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
vector float vyd = vec_splats(y[i].d); vector float vyd = vec_splats(y[i].d);
vector float vd = vec_mul(vxd, vyd); vector float vd = vec_mul(vxd, vyd);
@ -2176,7 +2177,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
float sumf = 0.f; float sumf = 0.f;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * GGML_RESTRICT q3 = x[i].qs; const uint8_t * GGML_RESTRICT q3 = x[i].qs;
const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4; const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
@ -2236,7 +2237,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
const vector signed char mask2 = (vector signed char)vec_xl( 0, k_mask2); const vector signed char mask2 = (vector signed char)vec_xl( 0, k_mask2);
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
vector float vyd = vec_splats(y[i].d); vector float vyd = vec_splats(y[i].d);
vector float vd = vec_mul(vxd, vyd); vector float vd = vec_mul(vxd, vyd);
@ -2329,7 +2330,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
float sumf = 0.f; float sumf = 0.f;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * GGML_RESTRICT qs = x[i].qs; const uint8_t * GGML_RESTRICT qs = x[i].qs;
const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint8_t * GGML_RESTRICT qh = x[i].qh;
const uint8_t * GGML_RESTRICT signs = x[i].signs; const uint8_t * GGML_RESTRICT signs = x[i].signs;
@ -2394,7 +2395,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
vector float vsumf3 = vec_splats(0.0f); vector float vsumf3 = vec_splats(0.0f);
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
vector float vyd = vec_splats(y[i].d); vector float vyd = vec_splats(y[i].d);
vector float vd = vec_mul(vxd, vyd); vector float vd = vec_mul(vxd, vyd);
@ -2505,7 +2506,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
qs += 4; qs += 4;
} }
sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1); sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
} }
*s = sumf; *s = sumf;
@ -2546,8 +2547,8 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
__builtin_prefetch(y[ib].qs, 0, 1); __builtin_prefetch(y[ib].qs, 0, 1);
vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d)); vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d));
vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d)); vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d));
vector float vd = vec_mul(vxd, vyd); vector float vd = vec_mul(vxd, vyd);
vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs); vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
@ -2582,7 +2583,7 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
#endif #endif
for (; ib < nb; ++ib) { for (; ib < nb; ++ib) {
const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d); const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
int sumi1 = 0, sumi2 = 0; int sumi1 = 0, sumi2 = 0;
for (int j = 0; j < QK4_NL/2; ++j) { for (int j = 0; j < QK4_NL/2; ++j) {
sumi1 += y[ib].qs[j+ 0] * kvalues_iq4nl[x[ib].qs[j] & 0xf]; sumi1 += y[ib].qs[j+ 0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
@ -2620,7 +2621,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
for (int ibl = 0; ibl < nb; ++ibl) { for (int ibl = 0; ibl < nb; ++ibl) {
vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ibl].d)); vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ibl].d));
vector float vyd = vec_splats(y[ibl].d); vector float vyd = vec_splats(y[ibl].d);
vector float vd = vec_mul(vxd, vyd); vector float vd = vec_mul(vxd, vyd);
@ -2697,7 +2698,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
#else #else
float sumf = 0; float sumf = 0;
for (int ibl = 0; ibl < nb; ++ibl) { for (int ibl = 0; ibl < nb; ++ibl) {
const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d; const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
uint16_t h = x[ibl].scales_h; uint16_t h = x[ibl].scales_h;
const uint8_t * qs = x[ibl].qs; const uint8_t * qs = x[ibl].qs;
const int8_t * q8 = y[ibl].qs; const int8_t * q8 = y[ibl].qs;

83
ggml/src/ggml-cpu/arch/riscv/quants.c
View File

@ -3,6 +3,7 @@
#include "ggml-quants.h" #include "ggml-quants.h"
#include "ggml-impl.h" #include "ggml-impl.h"
#include "ggml-cpu.h" #include "ggml-cpu.h"
#include "simd-mappings.h"
#include "../../quants.h" #include "../../quants.h"
#include "../../ggml-cpu-impl.h" #include "../../ggml-cpu-impl.h"
@ -45,7 +46,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
const float d = amax / ((1 << 7) - 1); const float d = amax / ((1 << 7) - 1);
const float id = d ? 1.0f/d : 0.0f; const float id = d ? 1.0f/d : 0.0f;
y[i].d = GGML_FP32_TO_FP16(d); y[i].d = GGML_CPU_FP32_TO_FP16(d);
vfloat32m8_t x0 = __riscv_vfmul_vf_f32m8(v_x, id, vl); vfloat32m8_t x0 = __riscv_vfmul_vf_f32m8(v_x, id, vl);
@ -85,7 +86,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
const float d = amax / ((1 << 7) - 1); const float d = amax / ((1 << 7) - 1);
const float id = d ? 1.0f/d : 0.0f; const float id = d ? 1.0f/d : 0.0f;
y[i].d = GGML_FP32_TO_FP16(d); y[i].d = GGML_CPU_FP32_TO_FP16(d);
vfloat32m8_t x0 = __riscv_vfmul_vf_f32m8(v_x, id, vl); vfloat32m8_t x0 = __riscv_vfmul_vf_f32m8(v_x, id, vl);
@ -102,7 +103,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
// set y[i].s // set y[i].s
int sum = __riscv_vmv_x_s_i16m1_i16(vwrs); int sum = __riscv_vmv_x_s_i16m1_i16(vwrs);
y[i].s = GGML_FP32_TO_FP16(sum*d); y[i].s = GGML_CPU_FP32_TO_FP16(sum*d);
} }
#else #else
@ -160,7 +161,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
int sumi = __riscv_vmv_x_s_i32m1_i32(vs2); int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d); sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
} }
#endif #endif
@ -177,7 +178,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
} }
int sumi = sumi0 + sumi1; int sumi = sumi0 + sumi1;
sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d); sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
} }
*s = sumf; *s = sumf;
@ -225,7 +226,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
int sumi = __riscv_vmv_x_s_i32m1_i32(vs2); int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
} }
#endif #endif
@ -242,7 +243,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
} }
int sumi = sumi0 + sumi1; int sumi = sumi0 + sumi1;
sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
} }
*s = sumf; *s = sumf;
@ -293,7 +294,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
vint32m1_t sum = __riscv_vwredsum_vs_i16m4_i32m1(mul, zero, vl); vint32m1_t sum = __riscv_vwredsum_vs_i16m4_i32m1(mul, zero, vl);
int32_t sumi = __riscv_vmv_x_s_i32m1_i32(sum); int32_t sumi = __riscv_vmv_x_s_i32m1_i32(sum);
sumf += (GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)) * sumi; sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
} }
#endif #endif
@ -316,7 +317,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
} }
int sumi = sumi0 + sumi1; int sumi = sumi0 + sumi1;
sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi; sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
} }
*s = sumf; *s = sumf;
@ -366,7 +367,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
vint32m1_t sum = __riscv_vwredsum_vs_i16m4_i32m1(mul, zero, vl); vint32m1_t sum = __riscv_vwredsum_vs_i16m4_i32m1(mul, zero, vl);
int32_t sumi = __riscv_vmv_x_s_i32m1_i32(sum); int32_t sumi = __riscv_vmv_x_s_i32m1_i32(sum);
sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
} }
#endif #endif
@ -389,7 +390,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
} }
int sumi = sumi0 + sumi1; int sumi = sumi0 + sumi1;
sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
} }
*s = sumf; *s = sumf;
@ -427,7 +428,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
int sumi = __riscv_vmv_x_s_i32m1_i32(v_sum); int sumi = __riscv_vmv_x_s_i32m1_i32(v_sum);
sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)); sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
} }
#endif #endif
@ -438,7 +439,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumi += x[ib].qs[j]*y[ib].qs[j]; sumi += x[ib].qs[j]*y[ib].qs[j];
} }
sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)); sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
} }
*s = sumf; *s = sumf;
@ -465,8 +466,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
const uint8_t * q2 = x[i].qs; const uint8_t * q2 = x[i].qs;
const int8_t * q8 = y[i].qs; const int8_t * q8 = y[i].qs;
const uint8_t * sc = x[i].scales; const uint8_t * sc = x[i].scales;
const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
uint8_t *patmp = atmp; uint8_t *patmp = atmp;
int vsums; int vsums;
int tmp; int tmp;
@ -569,8 +570,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
const int8_t * q8 = y[i].qs; const int8_t * q8 = y[i].qs;
const uint8_t * sc = x[i].scales; const uint8_t * sc = x[i].scales;
const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
size_t vl = 16; size_t vl = 16;
@ -644,8 +645,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
const uint8_t * q2 = x[i].qs; const uint8_t * q2 = x[i].qs;
const int8_t * q8 = y[i].qs; const int8_t * q8 = y[i].qs;
const uint8_t * sc = x[i].scales; const uint8_t * sc = x[i].scales;
const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
uint8_t *patmp = atmp; uint8_t *patmp = atmp;
int vsums; int vsums;
int tmp; int tmp;
@ -750,8 +751,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
summs += y[i].bsums[j] * (sc[j] >> 4); summs += y[i].bsums[j] * (sc[j] >> 4);
} }
const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
int isum = 0; int isum = 0;
int is = 0; int is = 0;
@ -916,7 +917,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
q3 += 32; q8 += 128; scale += 8; q3 += 32; q8 += 128; scale += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
sumf += d * isum; sumf += d * isum;
} }
@ -1017,7 +1018,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
sumf += d*sum_t; sumf += d*sum_t;
@ -1134,7 +1135,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
q3 += 32; q8 += 128; scale += 8; q3 += 32; q8 += 128; scale += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
sumf += d * isum; sumf += d * isum;
} }
break; break;
@ -1202,7 +1203,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];
@ -1239,8 +1240,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
float sumf = 0; float sumf = 0;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
int tmp, tmp2, sumi; int tmp, tmp2, sumi;
__asm__ __volatile__( __asm__ __volatile__(
@ -1361,8 +1362,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
size_t vl = 8; size_t vl = 8;
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
vint16mf2_t q8sums_0 = __riscv_vlse16_v_i16mf2(y[i].bsums, 4, vl); vint16mf2_t q8sums_0 = __riscv_vlse16_v_i16mf2(y[i].bsums, 4, vl);
vint16mf2_t q8sums_1 = __riscv_vlse16_v_i16mf2(y[i].bsums+1, 4, vl); vint16mf2_t q8sums_1 = __riscv_vlse16_v_i16mf2(y[i].bsums+1, 4, vl);
@ -1422,8 +1423,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
break; break;
case 128: case 128:
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
int tmp, tmp2, sumi; int tmp, tmp2, sumi;
__asm__ __volatile__( __asm__ __volatile__(
@ -1580,9 +1581,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi; sumf -= dmin * sumi;
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];
@ -1627,8 +1628,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
const uint8_t * GGML_RESTRICT hm = x[i].qh; const uint8_t * GGML_RESTRICT hm = x[i].qh;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
vint16m1_t q8sums_0 = __riscv_vlse16_v_i16m1(y[i].bsums, 4, vl); vint16m1_t q8sums_0 = __riscv_vlse16_v_i16m1(y[i].bsums, 4, vl);
vint16m1_t q8sums_1 = __riscv_vlse16_v_i16m1(y[i].bsums+1, 4, vl); vint16m1_t q8sums_1 = __riscv_vlse16_v_i16m1(y[i].bsums+1, 4, vl);
@ -1749,9 +1750,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi; sumf -= dmin * sumi;
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];
@ -1778,7 +1779,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * restrict q6 = x[i].ql; const uint8_t * restrict q6 = x[i].ql;
const uint8_t * restrict qh = x[i].qh; const uint8_t * restrict qh = x[i].qh;
@ -1862,7 +1863,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
case 256: case 256:
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * GGML_RESTRICT q6 = x[i].ql; const uint8_t * GGML_RESTRICT q6 = x[i].ql;
const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint8_t * GGML_RESTRICT qh = x[i].qh;
@ -1943,7 +1944,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
case 128: case 128:
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * restrict q6 = x[i].ql; const uint8_t * restrict q6 = x[i].ql;
const uint8_t * restrict qh = x[i].qh; const uint8_t * restrict qh = x[i].qh;
@ -2058,7 +2059,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];

47
ggml/src/ggml-cpu/arch/riscv/repack.cpp
View File

@ -6,6 +6,7 @@
#include "ggml-impl.h" #include "ggml-impl.h"
#include "ggml-cpu.h" #include "ggml-cpu.h"
#include "ggml-cpu-impl.h" #include "ggml-cpu-impl.h"
#include "simd-mappings.h"
#include "traits.h" #include "traits.h"
#include <cmath> #include <cmath>
@ -90,16 +91,16 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4); const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
// vector version needs Zvfhmin extension // vector version needs Zvfhmin extension
const float a_scale = GGML_FP16_TO_FP32(a_ptr[l].d); const float a_scale = GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
const float b_scales[8] = { const float b_scales[8] = {
GGML_FP16_TO_FP32(b_ptr[l].d[0]), GGML_CPU_FP16_TO_FP32(b_ptr[l].d[0]),
GGML_FP16_TO_FP32(b_ptr[l].d[1]), GGML_CPU_FP16_TO_FP32(b_ptr[l].d[1]),
GGML_FP16_TO_FP32(b_ptr[l].d[2]), GGML_CPU_FP16_TO_FP32(b_ptr[l].d[2]),
GGML_FP16_TO_FP32(b_ptr[l].d[3]), GGML_CPU_FP16_TO_FP32(b_ptr[l].d[3]),
GGML_FP16_TO_FP32(b_ptr[l].d[4]), GGML_CPU_FP16_TO_FP32(b_ptr[l].d[4]),
GGML_FP16_TO_FP32(b_ptr[l].d[5]), GGML_CPU_FP16_TO_FP32(b_ptr[l].d[5]),
GGML_FP16_TO_FP32(b_ptr[l].d[6]), GGML_CPU_FP16_TO_FP32(b_ptr[l].d[6]),
GGML_FP16_TO_FP32(b_ptr[l].d[7]) GGML_CPU_FP16_TO_FP32(b_ptr[l].d[7])
}; };
const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4); const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scale, vl / 4); const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scale, vl / 4);
@ -129,7 +130,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
} }
sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
} }
} }
} }
@ -181,20 +182,20 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
// vector version needs Zvfhmin extension // vector version needs Zvfhmin extension
const float a_scales[4] = { const float a_scales[4] = {
GGML_FP16_TO_FP32(a_ptr[l].d[0]), GGML_CPU_FP16_TO_FP32(a_ptr[l].d[0]),
GGML_FP16_TO_FP32(a_ptr[l].d[1]), GGML_CPU_FP16_TO_FP32(a_ptr[l].d[1]),
GGML_FP16_TO_FP32(a_ptr[l].d[2]), GGML_CPU_FP16_TO_FP32(a_ptr[l].d[2]),
GGML_FP16_TO_FP32(a_ptr[l].d[3]) GGML_CPU_FP16_TO_FP32(a_ptr[l].d[3])
}; };
const float b_scales[8] = { const float b_scales[8] = {
GGML_FP16_TO_FP32(b_ptr[l].d[0]), GGML_CPU_FP16_TO_FP32(b_ptr[l].d[0]),
GGML_FP16_TO_FP32(b_ptr[l].d[1]), GGML_CPU_FP16_TO_FP32(b_ptr[l].d[1]),
GGML_FP16_TO_FP32(b_ptr[l].d[2]), GGML_CPU_FP16_TO_FP32(b_ptr[l].d[2]),
GGML_FP16_TO_FP32(b_ptr[l].d[3]), GGML_CPU_FP16_TO_FP32(b_ptr[l].d[3]),
GGML_FP16_TO_FP32(b_ptr[l].d[4]), GGML_CPU_FP16_TO_FP32(b_ptr[l].d[4]),
GGML_FP16_TO_FP32(b_ptr[l].d[5]), GGML_CPU_FP16_TO_FP32(b_ptr[l].d[5]),
GGML_FP16_TO_FP32(b_ptr[l].d[6]), GGML_CPU_FP16_TO_FP32(b_ptr[l].d[6]),
GGML_FP16_TO_FP32(b_ptr[l].d[7]) GGML_CPU_FP16_TO_FP32(b_ptr[l].d[7])
}; };
const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4); const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
@ -382,7 +383,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
(v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
} }
sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]); sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
} }
} }
} }

57
ggml/src/ggml-cpu/arch/s390/quants.c
View File

@ -3,6 +3,7 @@
#include "ggml-quants.h" #include "ggml-quants.h"
#include "ggml-impl.h" #include "ggml-impl.h"
#include "ggml-cpu.h" #include "ggml-cpu.h"
#include "simd-mappings.h"
#include "../../quants.h" #include "../../quants.h"
#include "../../ggml-cpu-impl.h" #include "../../ggml-cpu-impl.h"
@ -49,7 +50,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
const float d = amax / ((1 << 7) - 1); const float d = amax / ((1 << 7) - 1);
const float id = d ? 1.0f / d : 0.0f; const float id = d ? 1.0f / d : 0.0f;
y[i].d = GGML_FP32_TO_FP16(d); y[i].d = GGML_CPU_FP32_TO_FP16(d);
for (int j = 0; j < 8; j++) { for (int j = 0; j < 8; j++) {
const __vector float v = vec_mul(srcv[j], vec_splats(id)); const __vector float v = vec_mul(srcv[j], vec_splats(id));
@ -94,7 +95,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
const float d = amax / ((1 << 7) - 1); const float d = amax / ((1 << 7) - 1);
const float id = d ? 1.0f / d : 0.0f; const float id = d ? 1.0f / d : 0.0f;
y[i].d = GGML_FP32_TO_FP16(d); y[i].d = GGML_CPU_FP32_TO_FP16(d);
__vector int32_t acc = vec_splats(0); __vector int32_t acc = vec_splats(0);
@ -110,7 +111,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
acc = vec_add(acc, vi); acc = vec_add(acc, vi);
} }
y[i].s = GGML_FP32_TO_FP16(d * (acc[0] + acc[1] + acc[2] + acc[3])); y[i].s = GGML_CPU_FP32_TO_FP16(d * (acc[0] + acc[1] + acc[2] + acc[3]));
} }
#else #else
GGML_UNUSED(nb); GGML_UNUSED(nb);
@ -164,7 +165,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
__vector int16_t v_xy_ = v_xylso + v_xylse + v_xyhso + v_xyhse; v_xy_ += vec_reve(v_xy_); __vector int16_t v_xy_ = v_xylso + v_xylse + v_xyhso + v_xyhse; v_xy_ += vec_reve(v_xy_);
const __vector float v_xy = vec_float(vec_unpackh(v_xy_)); const __vector float v_xy = vec_float(vec_unpackh(v_xy_));
const __vector float v_d = vec_splats(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)); const __vector float v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
acc = vec_madd(v_xy, v_d, acc); acc = vec_madd(v_xy, v_d, acc);
} }
@ -185,7 +186,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
} }
int sumi = sumi0 + sumi1; int sumi = sumi0 + sumi1;
sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d); sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
} }
*s = sumf; *s = sumf;
@ -219,7 +220,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
__builtin_prefetch(x[ib].qs, 0, 1); __builtin_prefetch(x[ib].qs, 0, 1);
__builtin_prefetch(y[ib].qs, 0, 1); __builtin_prefetch(y[ib].qs, 0, 1);
summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s); summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);
const uint8x16_t v_x = vec_xl(0, x[ib].qs); const uint8x16_t v_x = vec_xl(0, x[ib].qs);
const int8x16_t v_xl = (const int8x16_t)(v_x & v_m); const int8x16_t v_xl = (const int8x16_t)(v_x & v_m);
@ -231,7 +232,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
const int32x4_t v_xy_ = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh); const int32x4_t v_xy_ = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh);
const float32x4_t v_xy = vec_float(v_xy_); const float32x4_t v_xy = vec_float(v_xy_);
const float32x4_t v_d = vec_splats(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)); const float32x4_t v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
acc = vec_madd(v_xy, v_d, acc); acc = vec_madd(v_xy, v_d, acc);
} }
@ -252,7 +253,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
} }
int sumi = sumi0 + sumi1; int sumi = sumi0 + sumi1;
sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
} }
*s = sumf; *s = sumf;
@ -290,7 +291,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
const int32x4_t v_xy_ = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh); const int32x4_t v_xy_ = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh);
const float32x4_t v_xy = vec_float(v_xy_); const float32x4_t v_xy = vec_float(v_xy_);
const float32x4_t v_d = vec_splats(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)); const float32x4_t v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
acc = vec_madd(v_xy, v_d, acc); acc = vec_madd(v_xy, v_d, acc);
} }
@ -305,7 +306,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumi += x[ib].qs[j]*y[ib].qs[j]; sumi += x[ib].qs[j]*y[ib].qs[j];
} }
sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)); sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
} }
*s = sumf; *s = sumf;
@ -348,7 +349,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
float sum = 0; float sum = 0;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const uint8_t * restrict x0l = x[i].qs; const uint8_t * restrict x0l = x[i].qs;
const uint8_t * restrict x0h = x[i].hmask; const uint8_t * restrict x0h = x[i].hmask;
@ -497,7 +498,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];
@ -537,8 +538,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
float sumf = 0; float sumf = 0;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
const int16x8_t v_ysumsl = vec_xl(0 , y[i].bsums); const int16x8_t v_ysumsl = vec_xl(0 , y[i].bsums);
const int16x8_t v_ysumsh = vec_xl(16, y[i].bsums); const int16x8_t v_ysumsh = vec_xl(16, y[i].bsums);
@ -647,9 +648,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi; sumf -= dmin * sumi;
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];
@ -698,8 +699,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
float sumf = 0; float sumf = 0;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
const int16x8_t v_ysumsl = vec_xl(0 , y[i].bsums); const int16x8_t v_ysumsl = vec_xl(0 , y[i].bsums);
const int16x8_t v_ysumsh = vec_xl(16, y[i].bsums); const int16x8_t v_ysumsh = vec_xl(16, y[i].bsums);
@ -819,9 +820,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi; sumf -= dmin * sumi;
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];
@ -859,7 +860,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
int8x16_t v_y[4]; int8x16_t v_y[4];
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d_all = GGML_FP16_TO_FP32(x[i].d); const float d_all = GGML_CPU_FP16_TO_FP32(x[i].d);
const uint8_t * GGML_RESTRICT x0l = x[i].ql; const uint8_t * GGML_RESTRICT x0l = x[i].ql;
const uint8_t * GGML_RESTRICT x0h = x[i].qh; const uint8_t * GGML_RESTRICT x0h = x[i].qh;
@ -1004,7 +1005,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];
@ -1071,7 +1072,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
// float sumf = 0; // float sumf = 0;
// for (int i = 0; i < nb; ++i) { // for (int i = 0; i < nb; ++i) {
// const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; // const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
// const uint16_t * GGML_RESTRICT q2 = x[i].qs; // const uint16_t * GGML_RESTRICT q2 = x[i].qs;
// const int8_t * GGML_RESTRICT q8 = y[i].qs; // const int8_t * GGML_RESTRICT q8 = y[i].qs;
@ -1121,7 +1122,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
// float sumf = 0.f; // float sumf = 0.f;
// for (int i = 0; i < nb; ++i) { // for (int i = 0; i < nb; ++i) {
// const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; // const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
// const uint16_t * GGML_RESTRICT q2 = x[i].qs; // const uint16_t * GGML_RESTRICT q2 = x[i].qs;
// const int8_t * GGML_RESTRICT q8 = y[i].qs; // const int8_t * GGML_RESTRICT q8 = y[i].qs;
// int32_t bsum = 0; // int32_t bsum = 0;
@ -1182,12 +1183,12 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
const int8x16_t v_yh = vec_xl(QK8_0/2, y0->qs); const int8x16_t v_yh = vec_xl(QK8_0/2, y0->qs);
const int32x4_t v_xy = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh); const int32x4_t v_xy = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh);
sumf += GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d) * (v_xy[0] + v_xy[1] + v_xy[2] + v_xy[3]); sumf += GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d) * (v_xy[0] + v_xy[1] + v_xy[2] + v_xy[3]);
} }
#endif #endif
for (; ib < nb; ++ib) { for (; ib < nb; ++ib) {
const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d); const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
int sumi1 = 0, sumi2 = 0; int sumi1 = 0, sumi2 = 0;
for (int j = 0; j < QK4_NL/2; ++j) { for (int j = 0; j < QK4_NL/2; ++j) {
sumi1 += y[ib].qs[j+ 0] * kvalues_iq4nl[x[ib].qs[j] & 0xf]; sumi1 += y[ib].qs[j+ 0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
@ -1257,7 +1258,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
sumi2 += (vsumi1[0] + vsumi1[1] + vsumi1[2] + vsumi1[3]) * ls2; sumi2 += (vsumi1[0] + vsumi1[1] + vsumi1[2] + vsumi1[3]) * ls2;
} }
sumf += GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2); sumf += GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2);
} }
*s = sumf; *s = sumf;
@ -1265,7 +1266,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
#else #else
float sumf = 0; float sumf = 0;
for (int ibl = 0; ibl < nb; ++ibl) { for (int ibl = 0; ibl < nb; ++ibl) {
const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d; const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
uint16_t h = x[ibl].scales_h; uint16_t h = x[ibl].scales_h;
const uint8_t * qs = x[ibl].qs; const uint8_t * qs = x[ibl].qs;
const int8_t * q8 = y[ibl].qs; const int8_t * q8 = y[ibl].qs;

59
ggml/src/ggml-cpu/arch/wasm/quants.c
View File

@ -3,6 +3,7 @@
#include "ggml-quants.h" #include "ggml-quants.h"
#include "ggml-impl.h" #include "ggml-impl.h"
#include "ggml-cpu.h" #include "ggml-cpu.h"
#include "simd-mappings.h"
#include "../../quants.h" #include "../../quants.h"
#include "../../ggml-cpu-impl.h" #include "../../ggml-cpu-impl.h"
@ -65,7 +66,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
const float d = amax / ((1 << 7) - 1); const float d = amax / ((1 << 7) - 1);
const float id = d ? 1.0f/d : 0.0f; const float id = d ? 1.0f/d : 0.0f;
y[i].d = GGML_FP32_TO_FP16(d); y[i].d = GGML_CPU_FP32_TO_FP16(d);
for (int j = 0; j < 8; j++) { for (int j = 0; j < 8; j++) {
const v128_t v = wasm_f32x4_mul(srcv[j], wasm_f32x4_splat(id)); const v128_t v = wasm_f32x4_mul(srcv[j], wasm_f32x4_splat(id));
@ -110,7 +111,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
const float d = amax / ((1 << 7) - 1); const float d = amax / ((1 << 7) - 1);
const float id = d ? 1.0f/d : 0.0f; const float id = d ? 1.0f/d : 0.0f;
y[i].d = GGML_FP32_TO_FP16(d); y[i].d = GGML_CPU_FP32_TO_FP16(d);
v128_t accv = wasm_i32x4_splat(0); v128_t accv = wasm_i32x4_splat(0);
@ -126,7 +127,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
accv = wasm_i32x4_add(accv, vi); accv = wasm_i32x4_add(accv, vi);
} }
y[i].s = GGML_FP32_TO_FP16( y[i].s = GGML_CPU_FP32_TO_FP16(
d * (wasm_i32x4_extract_lane(accv, 0) + d * (wasm_i32x4_extract_lane(accv, 0) +
wasm_i32x4_extract_lane(accv, 1) + wasm_i32x4_extract_lane(accv, 1) +
wasm_i32x4_extract_lane(accv, 2) + wasm_i32x4_extract_lane(accv, 2) +
@ -324,8 +325,8 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
); );
// Accumulate results with scaling // Accumulate results with scaling
float scale0 = GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d); float scale0 = GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d);
float scale1 = GGML_FP16_TO_FP32(x1->d) * GGML_FP16_TO_FP32(y1->d); float scale1 = GGML_CPU_FP16_TO_FP32(x1->d) * GGML_CPU_FP16_TO_FP32(y1->d);
sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(dp0), wasm_f32x4_splat(scale0))); sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(dp0), wasm_f32x4_splat(scale0)));
sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(dp1), wasm_f32x4_splat(scale1))); sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(dp1), wasm_f32x4_splat(scale1)));
@ -348,7 +349,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
} }
int sumi = sumi0 + sumi1; int sumi = sumi0 + sumi1;
sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d); sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
} }
*s = sumf; *s = sumf;
@ -428,7 +429,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
wasm_i32x4_dot_i16x8(v0lfh, v1lh)), wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl), wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
wasm_i32x4_dot_i16x8(v0hfh, v1hh)))), wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d)))); wasm_f32x4_splat(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d))));
} }
sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) + sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
@ -454,7 +455,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
} }
int sumi = sumi0 + sumi1; int sumi = sumi0 + sumi1;
sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi; sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
} }
*s = sumf; *s = sumf;
@ -491,7 +492,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
const block_q5_1 * GGML_RESTRICT x0 = &x[ib]; const block_q5_1 * GGML_RESTRICT x0 = &x[ib];
const block_q8_1 * GGML_RESTRICT y0 = &y[ib]; const block_q8_1 * GGML_RESTRICT y0 = &y[ib];
summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s); summs += GGML_CPU_FP16_TO_FP32(x0->m) * GGML_CPU_FP16_TO_FP32(y0->s);
const v128_t m4b = wasm_i8x16_splat(0x0F); const v128_t m4b = wasm_i8x16_splat(0x0F);
@ -538,7 +539,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
wasm_i32x4_dot_i16x8(v0lfh, v1lh)), wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl), wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
wasm_i32x4_dot_i16x8(v0hfh, v1hh)))), wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d)))); wasm_f32x4_splat(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d))));
} }
sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) + sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
@ -564,7 +565,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
} }
int sumi = sumi0 + sumi1; int sumi = sumi0 + sumi1;
sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
} }
*s = sumf; *s = sumf;
@ -620,7 +621,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
const v128_t sum_dots = wasm_i32x4_add(wasm_i32x4_add(dx0_0, dx0_1), wasm_i32x4_add(dx1_0, dx1_1)); const v128_t sum_dots = wasm_i32x4_add(wasm_i32x4_add(dx0_0, dx0_1), wasm_i32x4_add(dx1_0, dx1_1));
// Convert to float and accumulate // Convert to float and accumulate
const float scale = GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d); const float scale = GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d);
sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(sum_dots), wasm_f32x4_splat(scale))); sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(sum_dots), wasm_f32x4_splat(scale)));
} }
@ -635,7 +636,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumi += x[ib].qs[j]*y[ib].qs[j]; sumi += x[ib].qs[j]*y[ib].qs[j];
} }
sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)); sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
} }
*s = sumf; *s = sumf;
@ -746,8 +747,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
isum += wasm_i32x4_extract_lane(isum_vec, 0); isum += wasm_i32x4_extract_lane(isum_vec, 0);
} }
const float dall = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float dall = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
sumf += dall * isum - dmin * summs; sumf += dall * isum - dmin * summs;
} }
@ -768,8 +769,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
summs += y[i].bsums[j] * (sc[j] >> 4); summs += y[i].bsums[j] * (sc[j] >> 4);
} }
const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
int isum = 0; int isum = 0;
int is = 0; int is = 0;
@ -880,7 +881,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
} }
// Accumulate results // Accumulate results
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const v128_t v_d = wasm_f32x4_splat(d); const v128_t v_d = wasm_f32x4_splat(d);
v128_t v_sum = wasm_f32x4_add( v128_t v_sum = wasm_f32x4_add(
wasm_f32x4_mul(wasm_f32x4_convert_i32x4(v_acc0), v_d), wasm_f32x4_mul(wasm_f32x4_convert_i32x4(v_acc0), v_d),
@ -957,7 +958,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];
@ -991,8 +992,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
float sumf = 0; float sumf = 0;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); // Corrected sign const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); // Corrected sign
const uint8_t * GGML_RESTRICT q4 = x[i].qs; const uint8_t * GGML_RESTRICT q4 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
@ -1136,9 +1137,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi; sumf -= dmin * sumi;
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];
@ -1170,8 +1171,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
float sumf = 0; float sumf = 0;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); // Fixed sign const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); // Fixed sign
const uint8_t * GGML_RESTRICT q5 = x[i].qs; const uint8_t * GGML_RESTRICT q5 = x[i].qs;
const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint8_t * GGML_RESTRICT qh = x[i].qh;
@ -1331,9 +1332,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi; sumf -= dmin * sumi;
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];
@ -1420,7 +1421,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
wasm_v128_store(&aux32[0], acc0); wasm_v128_store(&aux32[0], acc0);
wasm_v128_store(&aux32[4], acc1); wasm_v128_store(&aux32[4], acc1);
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) { for (int l = 0; l < 8; ++l) {
sums[l] += d * aux32[l]; sums[l] += d * aux32[l];
} }
@ -1470,7 +1471,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];

165
ggml/src/ggml-cpu/arch/x86/quants.c
View File

@ -3,6 +3,7 @@
#include "ggml-quants.h" #include "ggml-quants.h"
#include "ggml-impl.h" #include "ggml-impl.h"
#include "ggml-cpu.h" #include "ggml-cpu.h"
#include "simd-mappings.h"
#include "../../quants.h" #include "../../quants.h"
#include "../../ggml-cpu-impl.h" #include "../../ggml-cpu-impl.h"
@ -256,9 +257,9 @@ static inline __m256 mul_sum_i8_quad_float(const __m128i x_1_0, const __m128i x_
// quad fp16 delta calculation // quad fp16 delta calculation
static inline __m256 quad_fp16_delta_float(const float x0, const float y0, const float x1, const float y1) { static inline __m256 quad_fp16_delta_float(const float x0, const float y0, const float x1, const float y1) {
// GGML_FP16_TO_FP32 is faster than Intel F16C // GGML_CPU_FP16_TO_FP32 is faster than Intel F16C
return _mm256_set_m128(_mm_set1_ps(GGML_FP16_TO_FP32(x1) * GGML_FP16_TO_FP32(y1)), return _mm256_set_m128(_mm_set1_ps(GGML_CPU_FP16_TO_FP32(x1) * GGML_CPU_FP16_TO_FP32(y1)),
_mm_set1_ps(GGML_FP16_TO_FP32(x0) * GGML_FP16_TO_FP32(y0))); _mm_set1_ps(GGML_CPU_FP16_TO_FP32(x0) * GGML_CPU_FP16_TO_FP32(y0)));
} }
#endif #endif
#elif defined(__SSSE3__) #elif defined(__SSSE3__)
@ -305,7 +306,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
// Quantize these floats // Quantize these floats
const float d = maxScalar / 127.f; const float d = maxScalar / 127.f;
y[i].d = GGML_FP32_TO_FP16(d); y[i].d = GGML_CPU_FP32_TO_FP16(d);
const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f; const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f;
const __m256 mul = _mm256_set1_ps( id ); const __m256 mul = _mm256_set1_ps( id );
@ -401,7 +402,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
// Quantize these floats // Quantize these floats
const float d = max_scalar / 127.f; const float d = max_scalar / 127.f;
y[i].d = GGML_FP32_TO_FP16(d); y[i].d = GGML_CPU_FP32_TO_FP16(d);
const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f; const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f;
const __m256 mul = _mm256_set1_ps( id ); const __m256 mul = _mm256_set1_ps( id );
@ -425,7 +426,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
#if defined(__AVX2__) #if defined(__AVX2__)
// Compute the sum of the quants and set y[i].s // Compute the sum of the quants and set y[i].s
y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3)))); y[i].s = GGML_CPU_FP32_TO_FP16(d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3))));
// Convert int32 to int16 // Convert int32 to int16
i0 = _mm256_packs_epi32( i0, i1 ); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15 i0 = _mm256_packs_epi32( i0, i1 ); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15
@ -455,7 +456,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
// Compute the sum of the quants and set y[i].s // Compute the sum of the quants and set y[i].s
const __m128i s0 = _mm_add_epi32(_mm_add_epi32(ni0, ni1), _mm_add_epi32(ni2, ni3)); const __m128i s0 = _mm_add_epi32(_mm_add_epi32(ni0, ni1), _mm_add_epi32(ni2, ni3));
const __m128i s1 = _mm_add_epi32(_mm_add_epi32(ni4, ni5), _mm_add_epi32(ni6, ni7)); const __m128i s1 = _mm_add_epi32(_mm_add_epi32(ni4, ni5), _mm_add_epi32(ni6, ni7));
y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_4(_mm_add_epi32(s0, s1))); y[i].s = GGML_CPU_FP32_TO_FP16(d * hsum_i32_4(_mm_add_epi32(s0, s1)));
// Convert int32 to int16 // Convert int32 to int16
ni0 = _mm_packs_epi32( ni0, ni1 ); ni0 = _mm_packs_epi32( ni0, ni1 );
@ -552,7 +553,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
// Main loop // Main loop
for (; ib < nb; ++ib) { for (; ib < nb; ++ib) {
/* Compute combined scale for the block */ /* Compute combined scale for the block */
const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) ); const __m256 d = _mm256_set1_ps( GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d) );
__m256i qx = bytes_from_nibbles_32(x[ib].qs); __m256i qx = bytes_from_nibbles_32(x[ib].qs);
@ -613,7 +614,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
_mm_prefetch(&y[ib] + sizeof(block_q8_0), _MM_HINT_T0); _mm_prefetch(&y[ib] + sizeof(block_q8_0), _MM_HINT_T0);
// Compute combined scale for the block 0 and 1 // Compute combined scale for the block 0 and 1
const __m128 d_0_1 = _mm_set1_ps( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) ); const __m128 d_0_1 = _mm_set1_ps( GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d) );
const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[ib].qs); const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[ib].qs);
@ -631,7 +632,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
_mm_prefetch(&y[ib] + 2 * sizeof(block_q8_0), _MM_HINT_T0); _mm_prefetch(&y[ib] + 2 * sizeof(block_q8_0), _MM_HINT_T0);
// Compute combined scale for the block 2 and 3 // Compute combined scale for the block 2 and 3
const __m128 d_2_3 = _mm_set1_ps( GGML_FP16_TO_FP32(x[ib + 1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) ); const __m128 d_2_3 = _mm_set1_ps( GGML_CPU_FP16_TO_FP32(x[ib + 1].d) * GGML_CPU_FP16_TO_FP32(y[ib + 1].d) );
const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs); const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
@ -680,7 +681,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
} }
int sumi = sumi0 + sumi1; int sumi = sumi0 + sumi1;
sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d); sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
} }
*s = sumf; *s = sumf;
@ -711,10 +712,10 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
// Main loop // Main loop
for (; ib < nb; ++ib) { for (; ib < nb; ++ib) {
const float d0 = GGML_FP16_TO_FP32(x[ib].d); const float d0 = GGML_CPU_FP16_TO_FP32(x[ib].d);
const float d1 = GGML_FP16_TO_FP32(y[ib].d); const float d1 = GGML_CPU_FP16_TO_FP32(y[ib].d);
summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s); summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);
const __m256 d0v = _mm256_set1_ps( d0 ); const __m256 d0v = _mm256_set1_ps( d0 );
const __m256 d1v = _mm256_set1_ps( d1 ); const __m256 d1v = _mm256_set1_ps( d1 );
@ -752,7 +753,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
} }
int sumi = sumi0 + sumi1; int sumi = sumi0 + sumi1;
sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
} }
*s = sumf; *s = sumf;
@ -783,7 +784,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
// Main loop // Main loop
for (; ib < nb; ++ib) { for (; ib < nb; ++ib) {
/* Compute combined scale for the block */ /* Compute combined scale for the block */
const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)); const __m256 d = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
__m256i qx = bytes_from_nibbles_32(x[ib].qs); __m256i qx = bytes_from_nibbles_32(x[ib].qs);
__m256i bxhi = bytes_from_bits_32(x[ib].qh); __m256i bxhi = bytes_from_bits_32(x[ib].qh);
@ -807,7 +808,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
// Main loop // Main loop
for (; ib < nb; ++ib) { for (; ib < nb; ++ib) {
/* Compute combined scale for the block */ /* Compute combined scale for the block */
const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)); const __m256 d = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
__m256i bx_0 = bytes_from_nibbles_32(x[ib].qs); __m256i bx_0 = bytes_from_nibbles_32(x[ib].qs);
const __m256i bxhi = bytes_from_bits_32(x[ib].qh); const __m256i bxhi = bytes_from_bits_32(x[ib].qh);
@ -851,7 +852,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
} }
int sumi = sumi0 + sumi1; int sumi = sumi0 + sumi1;
sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi; sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
} }
*s = sumf; *s = sumf;
@ -883,16 +884,16 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
// Main loop // Main loop
for (; ib < nb; ++ib) { for (; ib < nb; ++ib) {
const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d)); const __m256 dx = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d));
summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s); summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);
__m256i qx = bytes_from_nibbles_32(x[ib].qs); __m256i qx = bytes_from_nibbles_32(x[ib].qs);
__m256i bxhi = bytes_from_bits_32(x[ib].qh); __m256i bxhi = bytes_from_bits_32(x[ib].qh);
bxhi = _mm256_and_si256(bxhi, _mm256_set1_epi8(0x10)); bxhi = _mm256_and_si256(bxhi, _mm256_set1_epi8(0x10));
qx = _mm256_or_si256(qx, bxhi); qx = _mm256_or_si256(qx, bxhi);
const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[ib].d)); const __m256 dy = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib].d));
const __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs); const __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
const __m256 q = mul_sum_us8_pairs_float(qx, qy); const __m256 q = mul_sum_us8_pairs_float(qx, qy);
@ -910,9 +911,9 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
// Main loop // Main loop
for (; ib < nb; ++ib) { for (; ib < nb; ++ib) {
const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d)); const __m256 dx = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d));
summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s); summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);
__m256i bx_0 = bytes_from_nibbles_32(x[ib].qs); __m256i bx_0 = bytes_from_nibbles_32(x[ib].qs);
const __m256i bxhi = bytes_from_bits_32(x[ib].qh); const __m256i bxhi = bytes_from_bits_32(x[ib].qh);
@ -926,7 +927,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
bxh = _mm_or_si128(bxh, bxhih); bxh = _mm_or_si128(bxh, bxhih);
bx_0 = MM256_SET_M128I(bxh, bxl); bx_0 = MM256_SET_M128I(bxh, bxl);
const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[ib].d)); const __m256 dy = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib].d));
const __m256i by_0 = _mm256_loadu_si256((const __m256i *)y[ib].qs); const __m256i by_0 = _mm256_loadu_si256((const __m256i *)y[ib].qs);
const __m256 q = mul_sum_us8_pairs_float(bx_0, by_0); const __m256 q = mul_sum_us8_pairs_float(bx_0, by_0);
@ -956,7 +957,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
} }
int sumi = sumi0 + sumi1; int sumi = sumi0 + sumi1;
sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
} }
*s = sumf; *s = sumf;
@ -986,7 +987,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
// Main loop // Main loop
for (; ib < nb; ++ib) { for (; ib < nb; ++ib) {
// Compute combined scale for the block // Compute combined scale for the block
const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)); const __m256 d = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
__m256i qx = _mm256_loadu_si256((const __m256i *)x[ib].qs); __m256i qx = _mm256_loadu_si256((const __m256i *)x[ib].qs);
__m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs); __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
@ -1025,7 +1026,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
sumi += x[ib].qs[j]*y[ib].qs[j]; sumi += x[ib].qs[j]*y[ib].qs[j];
} }
sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)); sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
} }
*s = sumf; *s = sumf;
@ -1144,7 +1145,7 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
} }
const __m256i ysum = _mm256_loadu_si256((const __m256i *) y[i].bsums); const __m256i ysum = _mm256_loadu_si256((const __m256i *) y[i].bsums);
const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(x[i].d)); const __m256 d = _mm256_set1_ps(y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d));
sumi0 = _mm256_sub_epi16(sumi0, ysum); sumi0 = _mm256_sub_epi16(sumi0, ysum);
sumi0 = _mm256_add_epi16(sumi0, _mm256_add_epi16(sumi1, sumi2)); sumi0 = _mm256_add_epi16(sumi0, _mm256_add_epi16(sumi1, sumi2));
@ -1190,7 +1191,7 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
} }
} }
sumf += (float) sum * (GGML_FP16_TO_FP32(x[i].d) * y[i].d); sumf += (float) sum * (GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d);
} }
*s = sumf; *s = sumf;
@ -1244,7 +1245,7 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
} }
const __m256i ysum = _mm256_loadu_si256((const __m256i *) y[i].bsums); const __m256i ysum = _mm256_loadu_si256((const __m256i *) y[i].bsums);
const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(x[i].d)); const __m256 d = _mm256_set1_ps(y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d));
sumi0 = _mm256_add_epi16(sumi0, sumi1); sumi0 = _mm256_add_epi16(sumi0, sumi1);
sumi0 = _mm256_sub_epi16(sumi0, ysum); sumi0 = _mm256_sub_epi16(sumi0, ysum);
@ -1269,7 +1270,7 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
} }
} }
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
sumf += (float) sumi * d; sumf += (float) sumi * d;
} }
@ -1299,8 +1300,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
const uint8_t * GGML_RESTRICT q2 = x[i].qs; const uint8_t * GGML_RESTRICT q2 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
@ -1366,8 +1367,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
const uint8_t * GGML_RESTRICT q2 = x[i].qs; const uint8_t * GGML_RESTRICT q2 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
@ -1477,8 +1478,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
summs += y[i].bsums[j] * (sc[j] >> 4); summs += y[i].bsums[j] * (sc[j] >> 4);
} }
const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
int isum = 0; int isum = 0;
int is = 0; int is = 0;
@ -1533,7 +1534,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const uint8_t * GGML_RESTRICT q3 = x[i].qs; const uint8_t * GGML_RESTRICT q3 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
@ -1638,7 +1639,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const uint8_t * GGML_RESTRICT q3 = x[i].qs; const uint8_t * GGML_RESTRICT q3 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
@ -1824,7 +1825,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];
@ -1862,8 +1863,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
memcpy(utmp, x[i].scales, 12); memcpy(utmp, x[i].scales, 12);
utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4); utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
@ -1928,8 +1929,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
const uint8_t * GGML_RESTRICT q4 = x[i].qs; const uint8_t * GGML_RESTRICT q4 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
@ -2049,9 +2050,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi; sumf -= dmin * sumi;
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];
@ -2092,8 +2093,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
const uint8_t * GGML_RESTRICT q5 = x[i].qs; const uint8_t * GGML_RESTRICT q5 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
memcpy(utmp, x[i].scales, 12); memcpy(utmp, x[i].scales, 12);
utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4); utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
@ -2170,8 +2171,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
const uint8_t * GGML_RESTRICT q5 = x[i].qs; const uint8_t * GGML_RESTRICT q5 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
@ -2311,9 +2312,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi; sumf -= dmin * sumi;
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];
@ -2344,7 +2345,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const uint8_t * GGML_RESTRICT q4 = x[i].ql; const uint8_t * GGML_RESTRICT q4 = x[i].ql;
const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint8_t * GGML_RESTRICT qh = x[i].qh;
@ -2422,7 +2423,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const uint8_t * GGML_RESTRICT q4 = x[i].ql; const uint8_t * GGML_RESTRICT q4 = x[i].ql;
const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint8_t * GGML_RESTRICT qh = x[i].qh;
@ -2555,7 +2556,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];
@ -2622,7 +2623,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
__m256 accumf = _mm256_setzero_ps(); __m256 accumf = _mm256_setzero_ps();
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint16_t * GGML_RESTRICT q2 = x[i].qs; const uint16_t * GGML_RESTRICT q2 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
__m256i sumi1 = _mm256_setzero_si256(); __m256i sumi1 = _mm256_setzero_si256();
@ -2663,7 +2664,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
__m256 accumf = _mm256_setzero_ps(); __m256 accumf = _mm256_setzero_ps();
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint16_t * GGML_RESTRICT q2 = x[i].qs; const uint16_t * GGML_RESTRICT q2 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
__m128i sumi1_0 = _mm_setzero_si128(); __m128i sumi1_0 = _mm_setzero_si128();
@ -2717,7 +2718,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
float sumf = 0.f; float sumf = 0.f;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint16_t * GGML_RESTRICT q2 = x[i].qs; const uint16_t * GGML_RESTRICT q2 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
int32_t bsum = 0; int32_t bsum = 0;
@ -2792,7 +2793,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
__m256 accumf = _mm256_setzero_ps(); __m256 accumf = _mm256_setzero_ps();
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint16_t * GGML_RESTRICT q2 = x[i].qs; const uint16_t * GGML_RESTRICT q2 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
@ -2913,7 +2914,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
__m256 accumf = _mm256_setzero_ps(); __m256 accumf = _mm256_setzero_ps();
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint16_t * GGML_RESTRICT q2 = x[i].qs; const uint16_t * GGML_RESTRICT q2 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
@ -3035,7 +3036,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
float sumf = 0.f; float sumf = 0.f;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint16_t * GGML_RESTRICT q2 = x[i].qs; const uint16_t * GGML_RESTRICT q2 = x[i].qs;
const uint8_t * GGML_RESTRICT sc = x[i].scales; const uint8_t * GGML_RESTRICT sc = x[i].scales;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
@ -3104,7 +3105,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
__m256 accumf = _mm256_setzero_ps(); __m256 accumf = _mm256_setzero_ps();
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * GGML_RESTRICT qs = x[i].qs; const uint8_t * GGML_RESTRICT qs = x[i].qs;
const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint8_t * GGML_RESTRICT qh = x[i].qh;
const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].qs + QK_K/8); const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].qs + QK_K/8);
@ -3177,7 +3178,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
__m256 accumf = _mm256_setzero_ps(); __m256 accumf = _mm256_setzero_ps();
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * GGML_RESTRICT qs = x[i].qs; const uint8_t * GGML_RESTRICT qs = x[i].qs;
const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint8_t * GGML_RESTRICT qh = x[i].qh;
const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].qs + QK_K/8); const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].qs + QK_K/8);
@ -3253,7 +3254,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
float sumf = 0; float sumf = 0;
for (int i = 0; i < nb; i++) { for (int i = 0; i < nb; i++) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const int8_t * q8 = y[i].qs; const int8_t * q8 = y[i].qs;
const uint8_t * qs = x[i].qs; const uint8_t * qs = x[i].qs;
const uint8_t * qh = x[i].qh; const uint8_t * qh = x[i].qh;
@ -3313,7 +3314,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
__m256 accumf = _mm256_setzero_ps(); __m256 accumf = _mm256_setzero_ps();
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * GGML_RESTRICT q3 = x[i].qs; const uint8_t * GGML_RESTRICT q3 = x[i].qs;
const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4; const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
@ -3358,7 +3359,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
__m256 accumf = _mm256_setzero_ps(); __m256 accumf = _mm256_setzero_ps();
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * GGML_RESTRICT q3 = x[i].qs; const uint8_t * GGML_RESTRICT q3 = x[i].qs;
const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4; const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
@ -3414,7 +3415,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
float sumf = 0.f; float sumf = 0.f;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * GGML_RESTRICT q3 = x[i].qs; const uint8_t * GGML_RESTRICT q3 = x[i].qs;
const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4; const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
@ -3480,7 +3481,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
__m256 accumf = _mm256_setzero_ps(); __m256 accumf = _mm256_setzero_ps();
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * GGML_RESTRICT qs = x[i].qs; const uint8_t * GGML_RESTRICT qs = x[i].qs;
const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint8_t * GGML_RESTRICT qh = x[i].qh;
const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs; const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs;
@ -3565,7 +3566,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
__m256 accumf = _mm256_setzero_ps(); __m256 accumf = _mm256_setzero_ps();
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * GGML_RESTRICT qs = x[i].qs; const uint8_t * GGML_RESTRICT qs = x[i].qs;
const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint8_t * GGML_RESTRICT qh = x[i].qh;
const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs; const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs;
@ -3648,7 +3649,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
float sumf = 0.f; float sumf = 0.f;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * GGML_RESTRICT qs = x[i].qs; const uint8_t * GGML_RESTRICT qs = x[i].qs;
const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint8_t * GGML_RESTRICT qh = x[i].qh;
const uint8_t * GGML_RESTRICT signs = x[i].signs; const uint8_t * GGML_RESTRICT signs = x[i].signs;
@ -3753,7 +3754,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
+ (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2; + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
} }
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
accum = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(sumi), accum); accum = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(sumi), accum);
accum1 += d * sumi1; accum1 += d * sumi1;
@ -3801,7 +3802,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
+ (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2; + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
} }
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(sumi1_1, sumi1_0))), accum); accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(sumi1_1, sumi1_0))), accum);
accum1 += d * sumi1; accum1 += d * sumi1;
@ -3835,7 +3836,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
qs += 4; qs += 4;
} }
sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1); sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
} }
*s = sumf; *s = sumf;
@ -3947,7 +3948,7 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
qs += 8; qh += 4; qs += 8; qh += 4;
} }
const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(scale.f16)); const __m256 d = _mm256_set1_ps(y[i].d * GGML_CPU_FP16_TO_FP32(scale.f16));
accum1 = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi1), accum1); accum1 = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi1), accum1);
accum2 = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi2), accum2); accum2 = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi2), accum2);
@ -4033,7 +4034,7 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
qs += 8; qh += 4; qs += 8; qh += 4;
} }
const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(scale.f16)); const __m256 d = _mm256_set1_ps(y[i].d * GGML_CPU_FP16_TO_FP32(scale.f16));
accum1 = _mm256_add_ps(_mm256_mul_ps(d, _mm256_cvtepi32_ps(MM256_SET_M128I(sumi1_1, sumi1_0))), accum1); accum1 = _mm256_add_ps(_mm256_mul_ps(d, _mm256_cvtepi32_ps(MM256_SET_M128I(sumi1_1, sumi1_0))), accum1);
accum2 = _mm256_add_ps(_mm256_mul_ps(d, _mm256_cvtepi32_ps(MM256_SET_M128I(sumi2_1, sumi2_0))), accum2); accum2 = _mm256_add_ps(_mm256_mul_ps(d, _mm256_cvtepi32_ps(MM256_SET_M128I(sumi2_1, sumi2_0))), accum2);
@ -4083,7 +4084,7 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
qh += 2; qh += 2;
} }
sumf += GGML_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2); sumf += GGML_CPU_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
} }
*s = sumf; *s = sumf;
@ -4129,9 +4130,9 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2); const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
const __m256i p_1 = _mm256_madd_epi16(p16_1, mone); const __m256i p_1 = _mm256_madd_epi16(p16_1, mone);
const __m256i p_2 = _mm256_madd_epi16(p16_2, mone); const __m256i p_2 = _mm256_madd_epi16(p16_2, mone);
accum1 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 0].d)*GGML_FP16_TO_FP32(x[ib + 0].d)), accum1 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 0].d)*GGML_CPU_FP16_TO_FP32(x[ib + 0].d)),
_mm256_cvtepi32_ps(p_1), accum1); _mm256_cvtepi32_ps(p_1), accum1);
accum2 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 1].d)*GGML_FP16_TO_FP32(x[ib + 1].d)), accum2 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 1].d)*GGML_CPU_FP16_TO_FP32(x[ib + 1].d)),
_mm256_cvtepi32_ps(p_2), accum2); _mm256_cvtepi32_ps(p_2), accum2);
} }
@ -4164,7 +4165,7 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
#endif #endif
for (; ib < nb; ++ib) { for (; ib < nb; ++ib) {
const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d); const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
int sumi1 = 0, sumi2 = 0; int sumi1 = 0, sumi2 = 0;
for (int j = 0; j < QK4_NL/2; ++j) { for (int j = 0; j < QK4_NL/2; ++j) {
sumi1 += y[ib].qs[j+ 0] * kvalues_iq4nl[x[ib].qs[j] & 0xf]; sumi1 += y[ib].qs[j+ 0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
@ -4219,7 +4220,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
sumi1 = _mm256_add_epi32(p_1, sumi1); sumi1 = _mm256_add_epi32(p_1, sumi1);
sumi2 = _mm256_add_epi32(p_2, sumi2); sumi2 = _mm256_add_epi32(p_2, sumi2);
} }
accum = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(x[ibl].d)*y[ibl].d), accum = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
_mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accum); _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accum);
} }
@ -4267,7 +4268,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
} }
__m128i sumi12_0 = _mm_add_epi32(sumi1_0, sumi2_0); __m128i sumi12_0 = _mm_add_epi32(sumi1_0, sumi2_0);
__m128i sumi12_1 = _mm_add_epi32(sumi1_1, sumi2_1); __m128i sumi12_1 = _mm_add_epi32(sumi1_1, sumi2_1);
accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(x[ibl].d)*y[ibl].d), accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
_mm256_cvtepi32_ps(MM256_SET_M128I(sumi12_1, sumi12_0))), accum); _mm256_cvtepi32_ps(MM256_SET_M128I(sumi12_1, sumi12_0))), accum);
} }
@ -4276,7 +4277,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
#else #else
float sumf = 0; float sumf = 0;
for (int ibl = 0; ibl < nb; ++ibl) { for (int ibl = 0; ibl < nb; ++ibl) {
const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d; const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
uint16_t h = x[ibl].scales_h; uint16_t h = x[ibl].scales_h;
const uint8_t * qs = x[ibl].qs; const uint8_t * qs = x[ibl].qs;
const int8_t * q8 = y[ibl].qs; const int8_t * q8 = y[ibl].qs;

39
ggml/src/ggml-cpu/arch/x86/repack.cpp
View File

@ -6,6 +6,7 @@
#include "ggml-impl.h" #include "ggml-impl.h"
#include "ggml-cpu.h" #include "ggml-cpu.h"
#include "ggml-cpu-impl.h" #include "ggml-cpu-impl.h"
#include "simd-mappings.h"
#include "traits.h" #include "traits.h"
#include <cmath> #include <cmath>
@ -39,11 +40,11 @@ static inline __m512 __avx512_f32cx8x2_load(ggml_fp16_t *x, ggml_fp16_t *y) {
float tmp[16]; float tmp[16];
for (int i = 0; i < 8; i++) { for (int i = 0; i < 8; i++) {
tmp[i] = GGML_FP16_TO_FP32(x[i]); tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]);
} }
for (int i = 0; i < 8; i++) { for (int i = 0; i < 8; i++) {
tmp[i + 8] = GGML_FP16_TO_FP32(y[i]); tmp[i + 8] = GGML_CPU_FP16_TO_FP32(y[i]);
} }
return _mm512_loadu_ps(tmp); return _mm512_loadu_ps(tmp);
@ -54,10 +55,10 @@ static inline __m512 __avx512_repeat_f32cx16_load(__m128i x) {
_mm_storeu_si128((__m128i*)tmphalf, x); _mm_storeu_si128((__m128i*)tmphalf, x);
for (int i = 0; i < 4; i++) { for (int i = 0; i < 4; i++) {
tmp[i] = GGML_FP16_TO_FP32(tmphalf[i]); tmp[i] = GGML_CPU_FP16_TO_FP32(tmphalf[i]);
tmp[i + 4] = GGML_FP16_TO_FP32(tmphalf[i]); tmp[i + 4] = GGML_CPU_FP16_TO_FP32(tmphalf[i]);
tmp[i + 8] = GGML_FP16_TO_FP32(tmphalf[i]); tmp[i + 8] = GGML_CPU_FP16_TO_FP32(tmphalf[i]);
tmp[i + 12] = GGML_FP16_TO_FP32(tmphalf[i]); tmp[i + 12] = GGML_CPU_FP16_TO_FP32(tmphalf[i]);
} }
return _mm512_loadu_ps(tmp); return _mm512_loadu_ps(tmp);
@ -67,7 +68,7 @@ static inline __m256 __avx_f32cx8_load(ggml_fp16_t *x) {
float tmp[8]; float tmp[8];
for (int i = 0; i < 8; i++) { for (int i = 0; i < 8; i++) {
tmp[i] = GGML_FP16_TO_FP32(x[i]); tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]);
} }
return _mm256_loadu_ps(tmp); return _mm256_loadu_ps(tmp);
@ -76,8 +77,8 @@ static inline __m256 __avx_repeat_f32cx8_load(ggml_fp16_t *x) {
float tmp[8]; float tmp[8];
for (int i = 0; i < 4; i++) { for (int i = 0; i < 4; i++) {
tmp[i] = GGML_FP16_TO_FP32(x[i]); tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]);
tmp[i + 4] = GGML_FP16_TO_FP32(x[i]); tmp[i + 4] = GGML_CPU_FP16_TO_FP32(x[i]);
} }
return _mm256_loadu_ps(tmp); return _mm256_loadu_ps(tmp);
@ -88,7 +89,7 @@ static inline __m256 __avx_rearranged_f32cx8_load(ggml_fp16_t *x, __m128i arrang
_mm_storeu_si128((__m128i*)tmphalf, _mm_shuffle_epi8(_mm_loadu_si128((const __m128i *) x), arrangeMask)); _mm_storeu_si128((__m128i*)tmphalf, _mm_shuffle_epi8(_mm_loadu_si128((const __m128i *) x), arrangeMask));
for (int i = 0; i < 8; i++) { for (int i = 0; i < 8; i++) {
tmp[i] = GGML_FP16_TO_FP32(tmphalf[i]); tmp[i] = GGML_CPU_FP16_TO_FP32(tmphalf[i]);
} }
return _mm256_loadu_ps(tmp); return _mm256_loadu_ps(tmp);
@ -211,7 +212,7 @@ void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR
id[row_iter] = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f; //d ? 1.0f / d : 0.0f; id[row_iter] = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f; //d ? 1.0f / d : 0.0f;
// Store the scale for the individual block // Store the scale for the individual block
y[i].d[row_iter] = GGML_FP32_TO_FP16(d); y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
// Store the values in blocks of eight values - Aim is to use these later for block interleaving // Store the values in blocks of eight values - Aim is to use these later for block interleaving
srcv[row_iter][0] = v0; srcv[row_iter][0] = v0;
@ -297,7 +298,7 @@ void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR
const float d = amax / ((1 << 7) - 1); const float d = amax / ((1 << 7) - 1);
id[row_iter] = d ? 1.0f / d : 0.0f; id[row_iter] = d ? 1.0f / d : 0.0f;
y[i].d[row_iter] = GGML_FP32_TO_FP16(d); y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
} }
for (int j = 0; j < QK8_0 * 4; j++) { for (int j = 0; j < QK8_0 * 4; j++) {
@ -647,7 +648,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
const __m256 col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, changemask); const __m256 col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, changemask);
// Load and convert to FP32 scale from block_q8_0 // Load and convert to FP32 scale from block_q8_0
const __m256 row_scale_f32 = _mm256_set1_ps(GGML_FP16_TO_FP32(a_ptr[b].d)); const __m256 row_scale_f32 = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(a_ptr[b].d));
// Load the block values in block_q8_0 in batches of 16 bytes and replicate the same across 256 bit vector // Load the block values in block_q8_0 in batches of 16 bytes and replicate the same across 256 bit vector
__m256i lhs_vec_0 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)a_ptr[b].qs)); __m256i lhs_vec_0 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)a_ptr[b].qs));
@ -706,7 +707,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
} }
sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
} }
} }
} }
@ -972,13 +973,13 @@ void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
sumi2 = sumi2 * scales_1[j]; sumi2 = sumi2 * scales_1[j];
sumi += sumi1 + sumi2; sumi += sumi1 + sumi2;
} }
sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d; sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
} }
} }
for (int sb = 0; sb < 8; sb++) { for (int sb = 0; sb < 8; sb++) {
uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16; uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16;
for (int j = 0; j < ncols_interleaved; j++) { for (int j = 0; j < ncols_interleaved; j++) {
sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d; sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
} }
} }
} }
@ -1755,7 +1756,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
(v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
} }
sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]); sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
} }
} }
} }
@ -3259,7 +3260,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
sumi2 = sumi2 * scales_1[j]; sumi2 = sumi2 * scales_1[j];
sumi += sumi1 + sumi2; sumi += sumi1 + sumi2;
} }
sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m]; sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
} }
} }
} }
@ -3268,7 +3269,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
for(int m = 0; m < 4; m++) { for(int m = 0; m < 4; m++) {
const int16_t *bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6); const int16_t *bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6);
for(int j = 0; j < ncols_interleaved; j++) { for(int j = 0; j < ncols_interleaved; j++) {
sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m]; sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
} }
} }
} }

5
ggml/src/ggml-cpu/common.h
View File

@ -4,6 +4,7 @@
#include "traits.h" #include "traits.h"
#include "ggml-cpu-impl.h" #include "ggml-cpu-impl.h"
#include "ggml-impl.h" #include "ggml-impl.h"
#include "simd-mappings.h"
#ifdef __cplusplus #ifdef __cplusplus
@ -12,11 +13,11 @@
// convenience functions/macros for use in template calls // convenience functions/macros for use in template calls
// note: these won't be required after the 'traits' lookup table is used. // note: these won't be required after the 'traits' lookup table is used.
static inline ggml_fp16_t f32_to_f16(float x) { static inline ggml_fp16_t f32_to_f16(float x) {
return GGML_FP32_TO_FP16(x); return GGML_CPU_FP32_TO_FP16(x);
} }
static inline float f16_to_f32(ggml_fp16_t x) { static inline float f16_to_f32(ggml_fp16_t x) {
return GGML_FP16_TO_FP32(x); return GGML_CPU_FP16_TO_FP32(x);
} }
static inline ggml_bf16_t f32_to_bf16(float x) { static inline ggml_bf16_t f32_to_bf16(float x) {

12
ggml/src/ggml-cpu/ggml-cpu-impl.h
View File

@ -62,11 +62,17 @@ struct ggml_compute_params {
#if defined(__s390x__) && defined(__VEC__) #if defined(__s390x__) && defined(__VEC__)
#ifndef __VXE__ #ifndef __VXE__
#define __VXE__ #define __VXE__
#endif #endif // __VXE__
#ifndef __VXE2__ #ifndef __VXE2__
#define __VXE2__ #define __VXE2__
#endif #endif // __VXE2__
#endif #endif // __s390x__ && __VEC__
#if defined(__s390x__) && defined(GGML_NNPA)
#ifndef __NNPA__
#define __NNPA__
#endif // __NNPA__
#endif // __s390x__ && GGML_NNPA
#if defined(__ARM_FEATURE_SVE) #if defined(__ARM_FEATURE_SVE)
#include <sys/prctl.h> #include <sys/prctl.h>

75
ggml/src/ggml-cpu/ggml-cpu.c
View File

@ -72,6 +72,9 @@
#define UNUSED GGML_UNUSED #define UNUSED GGML_UNUSED
#define SWAP(x, y, T) do { T SWAP = x; (x) = y; (y) = SWAP; } while (0) #define SWAP(x, y, T) do { T SWAP = x; (x) = y; (y) = SWAP; } while (0)
// precomputed f32 table for f16 (256 KB) (simd-mappings.h)
float ggml_table_f32_f16[1 << 16];
#if defined(__ARM_ARCH) #if defined(__ARM_ARCH)
struct ggml_arm_arch_features_type { struct ggml_arm_arch_features_type {
int sve_cnt; int sve_cnt;
@ -736,7 +739,7 @@ struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value) {
{ {
assert(tensor->nb[0] == sizeof(ggml_fp16_t)); assert(tensor->nb[0] == sizeof(ggml_fp16_t));
for (int i = 0; i < n; i++) { for (int i = 0; i < n; i++) {
ggml_vec_set_f16(nc, (ggml_fp16_t *)(data + i*n1), GGML_FP32_TO_FP16(value)); ggml_vec_set_f16(nc, (ggml_fp16_t *)(data + i*n1), GGML_CPU_FP32_TO_FP16(value));
} }
} break; } break;
case GGML_TYPE_BF16: case GGML_TYPE_BF16:
@ -795,7 +798,7 @@ struct ggml_tensor * ggml_set_f32(struct ggml_tensor * tensor, float value) {
{ {
assert(tensor->nb[0] == sizeof(ggml_fp16_t)); assert(tensor->nb[0] == sizeof(ggml_fp16_t));
for (int i = 0; i < n; i++) { for (int i = 0; i < n; i++) {
ggml_vec_set_f16(nc, (ggml_fp16_t *)(data + i*n1), GGML_FP32_TO_FP16(value)); ggml_vec_set_f16(nc, (ggml_fp16_t *)(data + i*n1), GGML_CPU_FP32_TO_FP16(value));
} }
} break; } break;
case GGML_TYPE_BF16: case GGML_TYPE_BF16:
@ -846,7 +849,7 @@ int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i) {
case GGML_TYPE_F16: case GGML_TYPE_F16:
{ {
GGML_ASSERT(tensor->nb[0] == sizeof(ggml_fp16_t)); GGML_ASSERT(tensor->nb[0] == sizeof(ggml_fp16_t));
return GGML_FP16_TO_FP32(((ggml_fp16_t *)(tensor->data))[i]); return GGML_CPU_FP16_TO_FP32(((ggml_fp16_t *)(tensor->data))[i]);
} }
case GGML_TYPE_BF16: case GGML_TYPE_BF16:
{ {
@ -891,7 +894,7 @@ void ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value) {
case GGML_TYPE_F16: case GGML_TYPE_F16:
{ {
GGML_ASSERT(tensor->nb[0] == sizeof(ggml_fp16_t)); GGML_ASSERT(tensor->nb[0] == sizeof(ggml_fp16_t));
((ggml_fp16_t *)(tensor->data))[i] = GGML_FP32_TO_FP16(value); ((ggml_fp16_t *)(tensor->data))[i] = GGML_CPU_FP32_TO_FP16(value);
} break; } break;
case GGML_TYPE_BF16: case GGML_TYPE_BF16:
{ {
@ -920,7 +923,7 @@ int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i
case GGML_TYPE_I32: case GGML_TYPE_I32:
return ((int32_t *) data)[0]; return ((int32_t *) data)[0];
case GGML_TYPE_F16: case GGML_TYPE_F16:
return GGML_FP16_TO_FP32(((ggml_fp16_t *) data)[0]); return GGML_CPU_FP16_TO_FP32(((ggml_fp16_t *) data)[0]);
case GGML_TYPE_BF16: case GGML_TYPE_BF16:
return GGML_BF16_TO_FP32(((ggml_bf16_t *) data)[0]); return GGML_BF16_TO_FP32(((ggml_bf16_t *) data)[0]);
case GGML_TYPE_F32: case GGML_TYPE_F32:
@ -947,7 +950,7 @@ void ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2,
} break; } break;
case GGML_TYPE_F16: case GGML_TYPE_F16:
{ {
((ggml_fp16_t *)(data))[0] = GGML_FP32_TO_FP16(value); ((ggml_fp16_t *)(data))[0] = GGML_CPU_FP32_TO_FP16(value);
} break; } break;
case GGML_TYPE_BF16: case GGML_TYPE_BF16:
{ {
@ -985,7 +988,7 @@ float ggml_get_f32_1d(const struct ggml_tensor * tensor, int i) {
} }
case GGML_TYPE_F16: case GGML_TYPE_F16:
{ {
return GGML_FP16_TO_FP32(((ggml_fp16_t *)(tensor->data))[i]); return GGML_CPU_FP16_TO_FP32(((ggml_fp16_t *)(tensor->data))[i]);
} }
case GGML_TYPE_BF16: case GGML_TYPE_BF16:
{ {
@ -1024,7 +1027,7 @@ void ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value) {
} break; } break;
case GGML_TYPE_F16: case GGML_TYPE_F16:
{ {
((ggml_fp16_t *)(tensor->data))[i] = GGML_FP32_TO_FP16(value); ((ggml_fp16_t *)(tensor->data))[i] = GGML_CPU_FP32_TO_FP16(value);
} break; } break;
case GGML_TYPE_BF16: case GGML_TYPE_BF16:
{ {
@ -1051,7 +1054,7 @@ float ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2,
case GGML_TYPE_I32: case GGML_TYPE_I32:
return ((int32_t *) data)[0]; return ((int32_t *) data)[0];
case GGML_TYPE_F16: case GGML_TYPE_F16:
return GGML_FP16_TO_FP32(((ggml_fp16_t *) data)[0]); return GGML_CPU_FP16_TO_FP32(((ggml_fp16_t *) data)[0]);
case GGML_TYPE_BF16: case GGML_TYPE_BF16:
return GGML_BF16_TO_FP32(((ggml_bf16_t *) data)[0]); return GGML_BF16_TO_FP32(((ggml_bf16_t *) data)[0]);
case GGML_TYPE_F32: case GGML_TYPE_F32:
@ -1078,7 +1081,7 @@ void ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2,
} break; } break;
case GGML_TYPE_F16: case GGML_TYPE_F16:
{ {
((ggml_fp16_t *)(data))[0] = GGML_FP32_TO_FP16(value); ((ggml_fp16_t *)(data))[0] = GGML_CPU_FP32_TO_FP16(value);
} break; } break;
case GGML_TYPE_BF16: case GGML_TYPE_BF16:
{ {
@ -3141,9 +3144,24 @@ void ggml_cpu_fp32_to_fp16(const float * x, ggml_fp16_t * y, int64_t n) {
__m128i y_vec = _mm_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT); __m128i y_vec = _mm_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT);
_mm_storel_epi64((__m128i *)(y + i), y_vec); _mm_storel_epi64((__m128i *)(y + i), y_vec);
} }
#elif defined(__NNPA__)
for (; i + 7 < n; i += 8) {
float32x4_t v_xh = vec_xl(0, (const float *)(x + i + 0));
float32x4_t v_xl = vec_xl(0, (const float *)(x + i + 4));
uint16x8_t v_yd = vec_round_from_fp32(v_xh, v_xl, 0);
uint16x8_t v_y = vec_convert_to_fp16(v_yd, 0);
vec_xst(v_y, 0, (ggml_fp16_t *)(y + i));
}
for (; i + 3 < n; i += 4) {
float32x4_t v_x = vec_xl(0, (const float *)(x + i));
float32x4_t v_zero = vec_splats(0.0f);
uint16x8_t v_yd = vec_round_from_fp32(v_x, v_zero, 0);
uint16x8_t v_y = vec_convert_to_fp16(v_yd, 0);
vec_xst(v_y, 0, (ggml_fp16_t *)(y + i));
}
#endif #endif
for (; i < n; ++i) { for (; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(x[i]); y[i] = GGML_CPU_FP32_TO_FP16(x[i]);
} }
} }
@ -3167,9 +3185,25 @@ void ggml_cpu_fp16_to_fp32(const ggml_fp16_t * x, float * y, int64_t n) {
__m128 y_vec = _mm_cvtph_ps(x_vec); __m128 y_vec = _mm_cvtph_ps(x_vec);
_mm_storeu_ps(y + i, y_vec); _mm_storeu_ps(y + i, y_vec);
} }
#elif defined(__NNPA__)
for (; i + 7 < n; i += 8) {
uint16x8_t v_x = vec_xl(0, (const ggml_fp16_t *)(x + i));
uint16x8_t v_yd = vec_convert_from_fp16(v_x, 0);
float32x4_t v_yh = vec_extend_to_fp32_hi(v_yd, 0);
float32x4_t v_yl = vec_extend_to_fp32_lo(v_yd, 0);
vec_xst(v_yh, 0, (float *)(y + i + 0));
vec_xst(v_yl, 0, (float *)(y + i + 4));
}
for (; i + 3 < n; i += 4) {
uint16x8_t v_x = vec_xl(0, (const ggml_fp16_t *)(x + i));
uint16x8_t v_yd = vec_convert_from_fp16(v_x, 0);
float32x4_t v_yh = vec_extend_to_fp32_hi(v_yd, 0);
vec_xst(v_yh, 0, (float *)(y + i));
}
#endif #endif
for (; i < n; ++i) { for (; i < n; ++i) {
y[i] = GGML_FP16_TO_FP32(x[i]); y[i] = GGML_CPU_FP16_TO_FP32(x[i]);
} }
} }
@ -3369,6 +3403,14 @@ int ggml_cpu_has_vxe(void) {
#endif #endif
} }
int ggml_cpu_has_nnpa(void) {
#if defined(GGML_NNPA)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_neon(void) { int ggml_cpu_has_neon(void) {
#if defined(__ARM_ARCH) && defined(__ARM_NEON) #if defined(__ARM_ARCH) && defined(__ARM_NEON)
return 1; return 1;
@ -3418,7 +3460,7 @@ int ggml_cpu_has_sme(void) {
} }
void ggml_cpu_init(void) { void ggml_cpu_init(void) {
// needed to initialize f16 tables // needed to initialize ggml_time
{ {
struct ggml_init_params params = { 0, NULL, false }; struct ggml_init_params params = { 0, NULL, false };
struct ggml_context * ctx = ggml_init(params); struct ggml_context * ctx = ggml_init(params);
@ -3439,9 +3481,10 @@ void ggml_cpu_init(void) {
uint16_t u16; uint16_t u16;
ggml_fp16_t fp16; ggml_fp16_t fp16;
} u = {i}; } u = {i};
float f = GGML_FP16_TO_FP32(u.fp16); float f = GGML_COMPUTE_FP16_TO_FP32(u.fp16);
ggml_table_gelu_f16[i] = GGML_FP32_TO_FP16(ggml_gelu_f32(f)); ggml_table_f32_f16[i] = f;
ggml_table_gelu_quick_f16[i] = GGML_FP32_TO_FP16(ggml_gelu_quick_f32(f)); ggml_table_gelu_f16[i] = GGML_CPU_FP32_TO_FP16(ggml_gelu_f32(f));
ggml_table_gelu_quick_f16[i] = GGML_CPU_FP32_TO_FP16(ggml_gelu_quick_f32(f));
} }
const uint64_t t_end = ggml_time_us(); UNUSED(t_end); const uint64_t t_end = ggml_time_us(); UNUSED(t_end);

3
ggml/src/ggml-cpu/ggml-cpu.cpp
View File

@ -578,6 +578,9 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
if (ggml_cpu_has_vxe()) { if (ggml_cpu_has_vxe()) {
features.push_back({ "VXE", "1" }); features.push_back({ "VXE", "1" });
} }
if (ggml_cpu_has_nnpa()) {
features.push_back({ "NNPA", "1" });
}
if (ggml_cpu_has_wasm_simd()) { if (ggml_cpu_has_wasm_simd()) {
features.push_back({ "WASM_SIMD", "1" }); features.push_back({ "WASM_SIMD", "1" });
} }

5
ggml/src/ggml-cpu/llamafile/sgemm.cpp
View File

@ -52,6 +52,7 @@
#include "ggml-impl.h" #include "ggml-impl.h"
#include "ggml-cpu-impl.h" #include "ggml-cpu-impl.h"
#include "ggml-quants.h" #include "ggml-quants.h"
#include "simd-mappings.h"
#include <array> #include <array>
#include <type_traits> #include <type_traits>
@ -73,7 +74,7 @@
namespace { namespace {
inline float unhalf(ggml_fp16_t d) { inline float unhalf(ggml_fp16_t d) {
return GGML_FP16_TO_FP32(d); return GGML_CPU_FP16_TO_FP32(d);
} }
//////////////////////////////////////////////////////////////////////////////////////////////////// ////////////////////////////////////////////////////////////////////////////////////////////////////
@ -252,7 +253,7 @@ template <> inline float32x4_t load(const ggml_fp16_t * p) {
float tmp[4]; float tmp[4];
for (int i = 0; i < 4; i++) { for (int i = 0; i < 4; i++) {
tmp[i] = GGML_FP16_TO_FP32(p[i]); tmp[i] = GGML_CPU_FP16_TO_FP32(p[i]);
} }
return vec_xl(0, (const float *)(tmp)); return vec_xl(0, (const float *)(tmp));

96
ggml/src/ggml-cpu/ops.cpp
View File

@ -108,7 +108,7 @@ static void ggml_compute_forward_dup_f16(
for (int i01 = ir0; i01 < ir1; i01++) { for (int i01 = ir0; i01 < ir1; i01++) {
const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
for (int i00 = 0; i00 < ne00; i00++) { for (int i00 = 0; i00 < ne00; i00++) {
dst_ptr[id] = GGML_FP16_TO_FP32(src0_ptr[i00]); dst_ptr[id] = GGML_CPU_FP16_TO_FP32(src0_ptr[i00]);
id++; id++;
} }
} }
@ -130,7 +130,7 @@ static void ggml_compute_forward_dup_f16(
const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
for (int i00 = 0; i00 < ne00; i00++) { for (int i00 = 0; i00 < ne00; i00++) {
src0_f32[i00] = GGML_FP16_TO_FP32(src0_ptr[i00]); src0_f32[i00] = GGML_CPU_FP16_TO_FP32(src0_ptr[i00]);
} }
quantize_row_q(src0_f32, dst_ptr + id, ne00); quantize_row_q(src0_f32, dst_ptr + id, ne00);
@ -156,7 +156,7 @@ static void ggml_compute_forward_dup_f16(
for (int i00 = 0; i00 < ne00; i00++) { for (int i00 = 0; i00 < ne00; i00++) {
const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
dst_ptr[id] = GGML_FP16_TO_FP32(*src0_ptr); dst_ptr[id] = GGML_CPU_FP16_TO_FP32(*src0_ptr);
id++; id++;
} }
} }
@ -267,7 +267,7 @@ static void ggml_compute_forward_dup_f16(
const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3);
*(float *) dst_ptr = GGML_FP16_TO_FP32(*(const ggml_fp16_t *) src0_ptr); *(float *) dst_ptr = GGML_CPU_FP16_TO_FP32(*(const ggml_fp16_t *) src0_ptr);
if (++i10 == ne0) { if (++i10 == ne0) {
i10 = 0; i10 = 0;
@ -372,7 +372,7 @@ static void ggml_compute_forward_dup_bf16(
for (int i01 = ir0; i01 < ir1; i01++) { for (int i01 = ir0; i01 < ir1; i01++) {
const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
for (int i00 = 0; i00 < ne00; i00++) { for (int i00 = 0; i00 < ne00; i00++) {
dst_ptr[id] = GGML_FP32_TO_FP16(GGML_BF16_TO_FP32(src0_ptr[i00])); dst_ptr[id] = GGML_CPU_FP32_TO_FP16(GGML_BF16_TO_FP32(src0_ptr[i00]));
id++; id++;
} }
} }
@ -473,7 +473,7 @@ static void ggml_compute_forward_dup_bf16(
for (int i00 = 0; i00 < ne00; i00++) { for (int i00 = 0; i00 < ne00; i00++) {
const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
dst_ptr[id] = GGML_FP32_TO_FP16(GGML_BF16_TO_FP32(*src0_ptr)); dst_ptr[id] = GGML_CPU_FP32_TO_FP16(GGML_BF16_TO_FP32(*src0_ptr));
id++; id++;
} }
} }
@ -566,7 +566,7 @@ static void ggml_compute_forward_dup_bf16(
const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3);
*(ggml_fp16_t *) dst_ptr = GGML_FP32_TO_FP16(GGML_BF16_TO_FP32(*(const ggml_bf16_t *) src0_ptr)); *(ggml_fp16_t *) dst_ptr = GGML_CPU_FP32_TO_FP16(GGML_BF16_TO_FP32(*(const ggml_bf16_t *) src0_ptr));
if (++i10 == ne0) { if (++i10 == ne0) {
i10 = 0; i10 = 0;
@ -765,7 +765,7 @@ static void ggml_compute_forward_dup_f32(
for (int i00 = 0; i00 < ne00; i00++) { for (int i00 = 0; i00 < ne00; i00++) {
const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
dst_ptr[id] = GGML_FP32_TO_FP16(*src0_ptr); dst_ptr[id] = GGML_CPU_FP32_TO_FP16(*src0_ptr);
id++; id++;
} }
} }
@ -878,7 +878,7 @@ static void ggml_compute_forward_dup_f32(
const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3);
*(ggml_fp16_t *) dst_ptr = GGML_FP32_TO_FP16(*(const float *) src0_ptr); *(ggml_fp16_t *) dst_ptr = GGML_CPU_FP32_TO_FP16(*(const float *) src0_ptr);
if (++i10 == ne0) { if (++i10 == ne0) {
i10 = 0; i10 = 0;
@ -1419,7 +1419,7 @@ static void ggml_compute_forward_add1_f16_f32(
ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ); ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 );
ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01);
for (int i = 0; i < ne0; i++) { for (int i = 0; i < ne0; i++) {
dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + v); dst_ptr[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(src0_ptr[i]) + v);
} }
} }
} }
@ -1435,7 +1435,7 @@ static void ggml_compute_forward_add1_f16_f16(
GGML_ASSERT(ggml_is_scalar(src1)); GGML_ASSERT(ggml_is_scalar(src1));
// scalar to add // scalar to add
const float v = GGML_FP16_TO_FP32(*(ggml_fp16_t *) src1->data); const float v = GGML_CPU_FP16_TO_FP32(*(ggml_fp16_t *) src1->data);
const int ith = params->ith; const int ith = params->ith;
const int nth = params->nth; const int nth = params->nth;
@ -1467,7 +1467,7 @@ static void ggml_compute_forward_add1_f16_f16(
ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ); ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 );
ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01);
for (int i = 0; i < ne0; i++) { for (int i = 0; i < ne0; i++) {
dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + v); dst_ptr[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(src0_ptr[i]) + v);
} }
} }
} }
@ -1889,7 +1889,7 @@ static void ggml_compute_forward_sum_f16(
} }
} }
} }
((ggml_fp16_t *) dst->data)[0] = GGML_FP32_TO_FP16(sum); ((ggml_fp16_t *) dst->data)[0] = GGML_CPU_FP32_TO_FP16(sum);
} }
static void ggml_compute_forward_sum_bf16( static void ggml_compute_forward_sum_bf16(
@ -2660,7 +2660,7 @@ static void ggml_compute_forward_gelu_f16(
#ifndef NDEBUG #ifndef NDEBUG
for (int k = 0; k < nc; k++) { for (int k = 0; k < nc; k++) {
const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k]; const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k];
const float v = GGML_FP16_TO_FP32(x); const float v = GGML_CPU_FP16_TO_FP32(x);
GGML_UNUSED(v); GGML_UNUSED(v);
assert(!isnan(v)); assert(!isnan(v));
assert(!isinf(v)); assert(!isinf(v));
@ -2763,7 +2763,7 @@ static void ggml_compute_forward_gelu_erf_f16(
#ifndef NDEBUG #ifndef NDEBUG
for (int k = 0; k < nc; k++) { for (int k = 0; k < nc; k++) {
const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k]; const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k];
const float v = GGML_FP16_TO_FP32(x); const float v = GGML_CPU_FP16_TO_FP32(x);
GGML_UNUSED(v); GGML_UNUSED(v);
assert(!isnan(v)); assert(!isnan(v));
assert(!isinf(v)); assert(!isinf(v));
@ -2866,7 +2866,7 @@ static void ggml_compute_forward_gelu_quick_f16(
#ifndef NDEBUG #ifndef NDEBUG
for (int k = 0; k < nc; k++) { for (int k = 0; k < nc; k++) {
const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k]; const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k];
const float v = GGML_FP16_TO_FP32(x); const float v = GGML_CPU_FP16_TO_FP32(x);
GGML_UNUSED(v); GGML_UNUSED(v);
assert(!isnan(v)); assert(!isnan(v));
assert(!isinf(v)); assert(!isinf(v));
@ -2969,7 +2969,7 @@ static void ggml_compute_forward_silu_f16(
#ifndef NDEBUG #ifndef NDEBUG
for (int k = 0; k < nc; k++) { for (int k = 0; k < nc; k++) {
const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*(dst->nb[1])))[k]; const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*(dst->nb[1])))[k];
const float v = GGML_FP16_TO_FP32(x); const float v = GGML_CPU_FP16_TO_FP32(x);
GGML_UNUSED(v); GGML_UNUSED(v);
assert(!isnan(v)); assert(!isnan(v));
assert(!isinf(v)); assert(!isinf(v));
@ -3163,7 +3163,7 @@ static void ggml_compute_forward_silu_back_f16(
#ifndef NDEBUG #ifndef NDEBUG
for (int k = 0; k < nc; k++) { for (int k = 0; k < nc; k++) {
const float x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k]; const float x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k];
const float v = GGML_FP16_TO_FP32(x); const float v = GGML_CPU_FP16_TO_FP32(x);
GGML_UNUSED(v); GGML_UNUSED(v);
assert(!isnan(v)); assert(!isnan(v));
assert(!isinf(v)); assert(!isinf(v));
@ -4500,7 +4500,7 @@ static void ggml_compute_forward_get_rows_back_f32_f16(
for (int j = 0; j < nc; ++j) { for (int j = 0; j < nc; ++j) {
ggml_fp16_t v = ((ggml_fp16_t *) ((char *) src0->data + i*src0->nb[1]))[j]; ggml_fp16_t v = ((ggml_fp16_t *) ((char *) src0->data + i*src0->nb[1]))[j];
((float *) ((char *) dst->data + r*dst->nb[1]))[j] += GGML_FP16_TO_FP32(v); ((float *) ((char *) dst->data + r*dst->nb[1]))[j] += GGML_CPU_FP16_TO_FP32(v);
} }
} }
} }
@ -4792,7 +4792,7 @@ static void ggml_compute_forward_soft_max_f32(
if (mp_f32) { if (mp_f32) {
if (use_f16) { if (use_f16) {
for (int i = 0; i < nc; ++i) { for (int i = 0; i < nc; ++i) {
wp[i] += slope*GGML_FP16_TO_FP32(mp_f16[i]); wp[i] += slope*GGML_CPU_FP16_TO_FP32(mp_f16[i]);
} }
} else { } else {
for (int i = 0; i < nc; ++i) { for (int i = 0; i < nc; ++i) {
@ -5018,8 +5018,8 @@ static void ggml_compute_forward_clamp_f16(
ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + j*nb01); ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + j*nb01);
for (int i = 0; i < nc; i++) { for (int i = 0; i < nc; i++) {
float v = GGML_FP16_TO_FP32(src0_ptr[i]); float v = GGML_CPU_FP16_TO_FP32(src0_ptr[i]);
dst_ptr[i] = GGML_FP32_TO_FP16(MAX(MIN(v, max), min)); dst_ptr[i] = GGML_CPU_FP32_TO_FP16(MAX(MIN(v, max), min));
} }
} }
} }
@ -5476,11 +5476,11 @@ static void ggml_compute_forward_rope_f16(
const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00); const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0); ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0);
const float x0 = GGML_FP16_TO_FP32(src[0]); const float x0 = GGML_CPU_FP16_TO_FP32(src[0]);
const float x1 = GGML_FP16_TO_FP32(src[n_dims]); const float x1 = GGML_CPU_FP16_TO_FP32(src[n_dims]);
dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); dst_data[0] = GGML_CPU_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
dst_data[n_dims] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); dst_data[n_dims] = GGML_CPU_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
} }
} else { } else {
for (int64_t i0 = 0; i0 < n_dims; i0 += 2) { for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
@ -5492,11 +5492,11 @@ static void ggml_compute_forward_rope_f16(
const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00); const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0); ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0);
const float x0 = GGML_FP16_TO_FP32(src[0]); const float x0 = GGML_CPU_FP16_TO_FP32(src[0]);
const float x1 = GGML_FP16_TO_FP32(src[n_dims/2]); const float x1 = GGML_CPU_FP16_TO_FP32(src[n_dims/2]);
dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); dst_data[0] = GGML_CPU_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
dst_data[n_dims/2] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); dst_data[n_dims/2] = GGML_CPU_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
} }
} }
} else { } else {
@ -5507,11 +5507,11 @@ static void ggml_compute_forward_rope_f16(
const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
const float x0 = GGML_FP16_TO_FP32(src[0]); const float x0 = GGML_CPU_FP16_TO_FP32(src[0]);
const float x1 = GGML_FP16_TO_FP32(src[1]); const float x1 = GGML_CPU_FP16_TO_FP32(src[1]);
dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); dst_data[0] = GGML_CPU_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
dst_data[1] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); dst_data[1] = GGML_CPU_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
} }
} }
@ -5525,11 +5525,11 @@ static void ggml_compute_forward_rope_f16(
const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00); const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0); ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0);
const float x0 = GGML_FP16_TO_FP32(src[0]); const float x0 = GGML_CPU_FP16_TO_FP32(src[0]);
const float x1 = GGML_FP16_TO_FP32(src[n_dims]); const float x1 = GGML_CPU_FP16_TO_FP32(src[n_dims]);
dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); dst_data[0] = GGML_CPU_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
dst_data[n_dims] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); dst_data[n_dims] = GGML_CPU_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
} }
} else { } else {
for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) { for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
@ -5640,7 +5640,7 @@ static void ggml_compute_forward_conv_transpose_1d_f16_f32(
for (int64_t i11 = 0; i11 < ne11; i11++) { for (int64_t i11 = 0; i11 < ne11; i11++) {
const float * const src = (float *)((char *) src1->data + i11*nb11); const float * const src = (float *)((char *) src1->data + i11*nb11);
for (int64_t i10 = 0; i10 < ne10; i10++) { for (int64_t i10 = 0; i10 < ne10; i10++) {
dst_data[i10*ne11 + i11] = GGML_FP32_TO_FP16(src[i10]); dst_data[i10*ne11 + i11] = GGML_CPU_FP32_TO_FP16(src[i10]);
} }
} }
} }
@ -5933,7 +5933,7 @@ static void ggml_compute_forward_im2col_f16(
if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) { if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
dst_data[iic*(KH*KW) + ikh*KW + ikw] = 0; dst_data[iic*(KH*KW) + ikh*KW + ikw] = 0;
} else { } else {
dst_data[iic*(KH*KW) + ikh*KW + ikw] = GGML_FP32_TO_FP16(src_data[iih*IW + iiw]); dst_data[iic*(KH*KW) + ikh*KW + ikw] = GGML_CPU_FP32_TO_FP16(src_data[iih*IW + iiw]);
} }
} }
} }
@ -6109,7 +6109,7 @@ void ggml_compute_forward_conv_transpose_2d(
const float * const src = (float *)((char *) src1->data + i12*nb12 + i11*nb11); const float * const src = (float *)((char *) src1->data + i12*nb12 + i11*nb11);
ggml_fp16_t * dst_data = wdata + i11*ne10*ne12; ggml_fp16_t * dst_data = wdata + i11*ne10*ne12;
for (int i10 = 0; i10 < ne10; i10++) { for (int i10 = 0; i10 < ne10; i10++) {
dst_data[i10*ne12 + i12] = GGML_FP32_TO_FP16(src[i10]); dst_data[i10*ne12 + i12] = GGML_CPU_FP32_TO_FP16(src[i10]);
} }
} }
} }
@ -6358,7 +6358,7 @@ static void ggml_compute_forward_pool_1d_sk_p0(
case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error"); case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error");
} }
for (int ki = 0; ki < k; ++ki) { for (int ki = 0; ki < k; ++ki) {
const float srow_j = (src->type == GGML_TYPE_F32) ? ((const float*)srow)[j] : GGML_FP16_TO_FP32(((const ggml_fp16_t*)srow)[j]); const float srow_j = (src->type == GGML_TYPE_F32) ? ((const float*)srow)[j] : GGML_CPU_FP16_TO_FP32(((const ggml_fp16_t*)srow)[j]);
switch (op) { switch (op) {
case GGML_OP_POOL_AVG: drow[i] += srow_j; break; case GGML_OP_POOL_AVG: drow[i] += srow_j; break;
case GGML_OP_POOL_MAX: if (srow_j > drow[i]) drow[i] = srow_j; break; case GGML_OP_POOL_MAX: if (srow_j > drow[i]) drow[i] = srow_j; break;
@ -6450,7 +6450,7 @@ void ggml_compute_forward_pool_2d(
for (int kx = 0; kx < k0; ++kx) { for (int kx = 0; kx < k0; ++kx) {
int j = ix + kx; int j = ix + kx;
if (j < 0 || j >= src->ne[0]) continue; if (j < 0 || j >= src->ne[0]) continue;
const float srow_j = (src->type == GGML_TYPE_F32) ? ((const float*)srow)[j] : GGML_FP16_TO_FP32(((const ggml_fp16_t*)srow)[j]); const float srow_j = (src->type == GGML_TYPE_F32) ? ((const float*)srow)[j] : GGML_CPU_FP16_TO_FP32(((const ggml_fp16_t*)srow)[j]);
switch (op) { switch (op) {
case GGML_OP_POOL_AVG: *out += srow_j; break; case GGML_OP_POOL_AVG: *out += srow_j; break;
case GGML_OP_POOL_MAX: if (srow_j > *out) *out = srow_j; break; case GGML_OP_POOL_MAX: if (srow_j > *out) *out = srow_j; break;
@ -6538,7 +6538,7 @@ void ggml_compute_forward_pool_2d_back(
} }
const float val = dst->type == GGML_TYPE_F32 ? const float val = dst->type == GGML_TYPE_F32 ?
((const float *) drowf)[j] : GGML_FP16_TO_FP32(((const ggml_fp16_t *) drowf)[j]); ((const float *) drowf)[j] : GGML_CPU_FP16_TO_FP32(((const ggml_fp16_t *) drowf)[j]);
if (val <= maxval) { if (val <= maxval) {
continue; continue;
} }
@ -6558,7 +6558,7 @@ void ggml_compute_forward_pool_2d_back(
if (dst->type == GGML_TYPE_F32) { if (dst->type == GGML_TYPE_F32) {
((float *) drow)[j] += grad0; ((float *) drow)[j] += grad0;
} else { } else {
((ggml_fp16_t *) drow)[j] = GGML_FP32_TO_FP16(grad0 + GGML_FP16_TO_FP32(((const ggml_fp16_t *) drow)[j])); ((ggml_fp16_t *) drow)[j] = GGML_CPU_FP32_TO_FP16(grad0 + GGML_CPU_FP16_TO_FP32(((const ggml_fp16_t *) drow)[j]));
} }
} else if (op == GGML_OP_POOL_AVG) { } else if (op == GGML_OP_POOL_AVG) {
const float grad = grad0 / ka; const float grad = grad0 / ka;
@ -6577,7 +6577,7 @@ void ggml_compute_forward_pool_2d_back(
if (dst->type == GGML_TYPE_F32) { if (dst->type == GGML_TYPE_F32) {
((float *) drow)[j] += grad; ((float *) drow)[j] += grad;
} else { } else {
((ggml_fp16_t *) drow)[j] += GGML_FP32_TO_FP16(grad); ((ggml_fp16_t *) drow)[j] += GGML_CPU_FP32_TO_FP16(grad);
} }
} }
} }
@ -7147,7 +7147,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
// loop over n_kv and n_head_kv // loop over n_kv and n_head_kv
// ref: https://arxiv.org/pdf/2112.05682.pdf // ref: https://arxiv.org/pdf/2112.05682.pdf
for (int64_t ic = 0; ic < nek1; ++ic) { for (int64_t ic = 0; ic < nek1; ++ic) {
const float mv = mp ? slope*GGML_FP16_TO_FP32(mp[ic]) : 0.0f; const float mv = mp ? slope*GGML_CPU_FP16_TO_FP32(mp[ic]) : 0.0f;
if (mv == -INFINITY) { if (mv == -INFINITY) {
continue; continue;
} }
@ -7215,7 +7215,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
if (v->type == GGML_TYPE_F16) { if (v->type == GGML_TYPE_F16) {
for (int64_t d = 0; d < DV; ++d) { for (int64_t d = 0; d < DV; ++d) {
VKQ32[d] = GGML_FP16_TO_FP32(VKQ16[d]); VKQ32[d] = GGML_CPU_FP16_TO_FP32(VKQ16[d]);
} }
} }

49
ggml/src/ggml-cpu/quants.c
View File

@ -2,6 +2,7 @@
#include "ggml-common.h" #include "ggml-common.h"
#include "ggml-cpu-impl.h" #include "ggml-cpu-impl.h"
#include "simd-mappings.h"
#include "ggml-quants.h" #include "ggml-quants.h"
#include "quants.h" #include "quants.h"
@ -137,7 +138,7 @@ void ggml_vec_dot_q4_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, c
} }
int sumi = sumi0 + sumi1; int sumi = sumi0 + sumi1;
sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d); sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
} }
*s = sumf; *s = sumf;
@ -174,7 +175,7 @@ void ggml_vec_dot_q4_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, c
} }
int sumi = sumi0 + sumi1; int sumi = sumi0 + sumi1;
sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
} }
*s = sumf; *s = sumf;
@ -217,7 +218,7 @@ void ggml_vec_dot_q5_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, c
} }
int sumi = sumi0 + sumi1; int sumi = sumi0 + sumi1;
sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi; sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
} }
*s = sumf; *s = sumf;
@ -260,7 +261,7 @@ void ggml_vec_dot_q5_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, c
} }
int sumi = sumi0 + sumi1; int sumi = sumi0 + sumi1;
sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
} }
*s = sumf; *s = sumf;
@ -290,7 +291,7 @@ void ggml_vec_dot_q8_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, c
sumi += x[ib].qs[j]*y[ib].qs[j]; sumi += x[ib].qs[j]*y[ib].qs[j];
} }
sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)); sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
} }
*s = sumf; *s = sumf;
@ -342,7 +343,7 @@ void ggml_vec_dot_tq1_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
} }
} }
sumf += (float) sum * (GGML_FP16_TO_FP32(x[i].d) * y[i].d); sumf += (float) sum * (GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d);
} }
*s = sumf; *s = sumf;
@ -372,7 +373,7 @@ void ggml_vec_dot_tq2_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
} }
} }
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
sumf += (float) sumi * d; sumf += (float) sumi * d;
} }
@ -405,8 +406,8 @@ void ggml_vec_dot_q2_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, c
summs += y[i].bsums[j] * (sc[j] >> 4); summs += y[i].bsums[j] * (sc[j] >> 4);
} }
const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
int isum = 0; int isum = 0;
int is = 0; int is = 0;
@ -504,7 +505,7 @@ void ggml_vec_dot_q3_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, c
for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];
@ -577,9 +578,9 @@ void ggml_vec_dot_q4_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, c
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi; sumf -= dmin * sumi;
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];
@ -657,9 +658,9 @@ void ggml_vec_dot_q5_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, c
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi; sumf -= dmin * sumi;
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];
@ -714,7 +715,7 @@ void ggml_vec_dot_q6_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, c
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8; q8 += 8; a += 8;
} }
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
} }
for (int l = 0; l < 8; ++l) sumf += sums[l]; for (int l = 0; l < 8; ++l) sumf += sums[l];
@ -739,7 +740,7 @@ void ggml_vec_dot_iq2_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs
float sumf = 0.f; float sumf = 0.f;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint16_t * GGML_RESTRICT q2 = x[i].qs; const uint16_t * GGML_RESTRICT q2 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
int32_t bsum = 0; int32_t bsum = 0;
@ -778,7 +779,7 @@ void ggml_vec_dot_iq2_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
float sumf = 0.f; float sumf = 0.f;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint16_t * GGML_RESTRICT q2 = x[i].qs; const uint16_t * GGML_RESTRICT q2 = x[i].qs;
const uint8_t * GGML_RESTRICT sc = x[i].scales; const uint8_t * GGML_RESTRICT sc = x[i].scales;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
@ -829,7 +830,7 @@ void ggml_vec_dot_iq2_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
float sumf = 0; float sumf = 0;
for (int i = 0; i < nb; i++) { for (int i = 0; i < nb; i++) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const int8_t * q8 = y[i].qs; const int8_t * q8 = y[i].qs;
const uint8_t * qs = x[i].qs; const uint8_t * qs = x[i].qs;
const uint8_t * qh = x[i].qh; const uint8_t * qh = x[i].qh;
@ -882,7 +883,7 @@ void ggml_vec_dot_iq3_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs
float sumf = 0.f; float sumf = 0.f;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * GGML_RESTRICT q3 = x[i].qs; const uint8_t * GGML_RESTRICT q3 = x[i].qs;
const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4; const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
const int8_t * GGML_RESTRICT q8 = y[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs;
@ -924,7 +925,7 @@ void ggml_vec_dot_iq3_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
float sumf = 0.f; float sumf = 0.f;
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
const uint8_t * GGML_RESTRICT qs = x[i].qs; const uint8_t * GGML_RESTRICT qs = x[i].qs;
const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint8_t * GGML_RESTRICT qh = x[i].qh;
const uint8_t * GGML_RESTRICT signs = x[i].signs; const uint8_t * GGML_RESTRICT signs = x[i].signs;
@ -1002,7 +1003,7 @@ void ggml_vec_dot_iq1_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
qs += 4; qs += 4;
} }
sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1); sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
} }
*s = sumf; *s = sumf;
@ -1063,7 +1064,7 @@ void ggml_vec_dot_iq1_m_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
qh += 2; qh += 2;
} }
sumf += GGML_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2); sumf += GGML_CPU_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
} }
*s = sumf; *s = sumf;
@ -1087,7 +1088,7 @@ void ggml_vec_dot_iq4_nl_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
float sumf = 0; float sumf = 0;
for (; ib < nb; ++ib) { for (; ib < nb; ++ib) {
const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d); const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
int sumi1 = 0, sumi2 = 0; int sumi1 = 0, sumi2 = 0;
for (int j = 0; j < QK4_NL/2; ++j) { for (int j = 0; j < QK4_NL/2; ++j) {
sumi1 += y[ib].qs[j+ 0] * kvalues_iq4nl[x[ib].qs[j] & 0xf]; sumi1 += y[ib].qs[j+ 0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
@ -1113,7 +1114,7 @@ void ggml_vec_dot_iq4_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
float sumf = 0; float sumf = 0;
for (int ibl = 0; ibl < nb; ++ibl) { for (int ibl = 0; ibl < nb; ++ibl) {
const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d; const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
uint16_t h = x[ibl].scales_h; uint16_t h = x[ibl].scales_h;
const uint8_t * qs = x[ibl].qs; const uint8_t * qs = x[ibl].qs;
const int8_t * q8 = y[ibl].qs; const int8_t * q8 = y[ibl].qs;

29
ggml/src/ggml-cpu/repack.cpp
View File

@ -6,6 +6,7 @@
#include "ggml-impl.h" #include "ggml-impl.h"
#include "ggml-cpu.h" #include "ggml-cpu.h"
#include "ggml-cpu-impl.h" #include "ggml-cpu-impl.h"
#include "simd-mappings.h"
#include "traits.h" #include "traits.h"
#include "arch-fallback.h" #include "arch-fallback.h"
@ -72,7 +73,7 @@ void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GG
const float d = amax / ((1 << 7) - 1); const float d = amax / ((1 << 7) - 1);
id[row_iter] = d ? 1.0f / d : 0.0f; id[row_iter] = d ? 1.0f / d : 0.0f;
y[i].d[row_iter] = GGML_FP32_TO_FP16(d); y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
} }
for (int j = 0; j < QK8_0 * 4; j++) { for (int j = 0; j < QK8_0 * 4; j++) {
@ -110,7 +111,7 @@ void ggml_quantize_mat_q8_0_4x8_generic(const float * GGML_RESTRICT x, void * GG
const float d = amax / ((1 << 7) - 1); const float d = amax / ((1 << 7) - 1);
id[row_iter] = d ? 1.0f / d : 0.0f; id[row_iter] = d ? 1.0f / d : 0.0f;
y[i].d[row_iter] = GGML_FP32_TO_FP16(d); y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
} }
for (int j = 0; j < QK8_0 * 4; j++) { for (int j = 0; j < QK8_0 * 4; j++) {
@ -236,7 +237,7 @@ void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
} }
sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
} }
} }
} }
@ -280,7 +281,7 @@ void ggml_gemv_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
} }
sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
} }
} }
} }
@ -325,7 +326,7 @@ void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
} }
sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
} }
} }
} }
@ -396,13 +397,13 @@ void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
sumi2 = sumi2 * scales_1[j]; sumi2 = sumi2 * scales_1[j];
sumi += sumi1 + sumi2; sumi += sumi1 + sumi2;
} }
sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d; sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
} }
} }
for (int sb = 0; sb < 8; sb++) { for (int sb = 0; sb < 8; sb++) {
uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16; uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16;
for (int j = 0; j < ncols_interleaved; j++) { for (int j = 0; j < ncols_interleaved; j++) {
sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d; sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
} }
} }
} }
@ -449,7 +450,7 @@ void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])); sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
} }
sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
} }
} }
} }
@ -500,7 +501,7 @@ void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
(v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
} }
sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]); sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
} }
} }
} }
@ -555,7 +556,7 @@ void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
(v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
} }
sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]); sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
} }
} }
} }
@ -609,7 +610,7 @@ void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
(v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
} }
sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]); sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
} }
} }
} }
@ -688,7 +689,7 @@ void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
sumi2 = sumi2 * scales_1[j]; sumi2 = sumi2 * scales_1[j];
sumi += sumi1 + sumi2; sumi += sumi1 + sumi2;
} }
sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m]; sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
} }
} }
} }
@ -697,7 +698,7 @@ void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
for(int m = 0; m < 4; m++) { for(int m = 0; m < 4; m++) {
const int16_t *bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6); const int16_t *bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6);
for(int j = 0; j < ncols_interleaved; j++) { for(int j = 0; j < ncols_interleaved; j++) {
sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m]; sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
} }
} }
} }
@ -753,7 +754,7 @@ void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
(v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])); (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4]));
} }
sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]); sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
} }
} }
} }

244
ggml/src/ggml-cpu/simd-mappings.h
View File

@ -2,10 +2,167 @@
#include "ggml-cpu-impl.h" #include "ggml-cpu-impl.h"
#ifdef __ARM_FEATURE_SVE
#include <arm_sve.h>
#endif // __ARM_FEATURE_SVE
#if defined(__ARM_NEON) && !defined(__CUDACC__) && !defined(__MUSACC__)
// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
//
// $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
//
#include <arm_neon.h>
#endif
#if defined(__F16C__)
#include <immintrin.h>
#endif
#ifdef __cplusplus
extern "C" {
#endif
// //
// simd mappings // simd mappings
// //
// FP16 to FP32 conversion
// 16-bit float
// on Arm, we use __fp16
// on x86, we use uint16_t
//
// for old CUDA compilers (<= 11), we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/10616
// for MUSA compilers , we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/11843
//
#if defined(__ARM_NEON) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) && !defined(__MUSACC__)
#define GGML_CPU_COMPUTE_FP16_TO_FP32(x) neon_compute_fp16_to_fp32(x)
#define GGML_CPU_COMPUTE_FP32_TO_FP16(x) neon_compute_fp32_to_fp16(x)
#define GGML_CPU_FP16_TO_FP32(x) GGML_CPU_COMPUTE_FP16_TO_FP32(x)
static inline float neon_compute_fp16_to_fp32(ggml_fp16_t h) {
__fp16 tmp;
memcpy(&tmp, &h, sizeof(ggml_fp16_t));
return (float)tmp;
}
static inline ggml_fp16_t neon_compute_fp32_to_fp16(float f) {
ggml_fp16_t res;
__fp16 tmp = f;
memcpy(&res, &tmp, sizeof(ggml_fp16_t));
return res;
}
#elif defined(__F16C__)
#ifdef _MSC_VER
#define GGML_CPU_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
#define GGML_CPU_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
#else
#define GGML_CPU_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
#define GGML_CPU_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
#endif
#elif defined(__POWER9_VECTOR__)
#define GGML_CPU_COMPUTE_FP16_TO_FP32(x) power_compute_fp16_to_fp32(x)
#define GGML_CPU_COMPUTE_FP32_TO_FP16(x) power_compute_fp32_to_fp16(x)
/* the inline asm below is about 12% faster than the lookup method */
#define GGML_CPU_FP16_TO_FP32(x) GGML_CPU_COMPUTE_FP16_TO_FP32(x)
#define GGML_CPU_FP32_TO_FP16(x) GGML_CPU_COMPUTE_FP32_TO_FP16(x)
static inline float power_compute_fp16_to_fp32(ggml_fp16_t h) {
float f;
double d;
__asm__(
"mtfprd %0,%2\n"
"xscvhpdp %0,%0\n"
"frsp %1,%0\n" :
/* temp */ "=d"(d),
/* out */ "=f"(f):
/* in */ "r"(h));
return f;
}
static inline ggml_fp16_t power_compute_fp32_to_fp16(float f) {
double d;
ggml_fp16_t r;
__asm__( /* xscvdphp can work on double or single precision */
"xscvdphp %0,%2\n"
"mffprd %1,%0\n" :
/* temp */ "=d"(d),
/* out */ "=r"(r):
/* in */ "f"(f));
return r;
}
#elif defined(__riscv) && defined(__riscv_zfhmin)
static inline float riscv_compute_fp16_to_fp32(ggml_fp16_t h) {
float f;
__asm__(
"fmv.h.x %[f], %[h]\n\t"
"fcvt.s.h %[f], %[f]"
: [f] "=&f" (f)
: [h] "r" (h)
);
return f;
}
static inline ggml_fp16_t riscv_compute_fp32_to_fp16(float f) {
ggml_fp16_t res;
__asm__(
"fcvt.h.s %[f], %[f]\n\t"
"fmv.x.h %[h], %[f]"
: [h] "=&r" (res)
: [f] "f" (f)
);
return res;
}
#define GGML_CPU_COMPUTE_FP16_TO_FP32(x) riscv_compute_fp16_to_fp32(x)
#define GGML_CPU_COMPUTE_FP32_TO_FP16(x) riscv_compute_fp32_to_fp16(x)
#define GGML_CPU_FP16_TO_FP32(x) GGML_CPU_COMPUTE_FP16_TO_FP32(x)
#define GGML_CPU_FP32_TO_FP16(x) GGML_CPU_COMPUTE_FP32_TO_FP16(x)
#elif defined(__NNPA__)
#define GGML_CPU_COMPUTE_FP16_TO_FP32(x) nnpa_compute_fp16_to_fp32(x)
#define GGML_CPU_COMPUTE_FP32_TO_FP16(x) nnpa_compute_fp32_to_fp16(x)
#define GGML_CPU_FP16_TO_FP32(x) GGML_CPU_COMPUTE_FP16_TO_FP32(x)
#define GGML_CPU_FP32_TO_FP16(x) GGML_CPU_COMPUTE_FP32_TO_FP16(x)
static inline float nnpa_compute_fp16_to_fp32(ggml_fp16_t h) {
uint16x8_t v_h = vec_splats(h);
uint16x8_t v_hd = vec_convert_from_fp16(v_h, 0);
return vec_extend_to_fp32_hi(v_hd, 0)[0];
}
static inline ggml_fp16_t nnpa_compute_fp32_to_fp16(float f) {
float32x4_t v_f = vec_splats(f);
float32x4_t v_zero = vec_splats(0.0f);
uint16x8_t v_hd = vec_round_from_fp32(v_f, v_zero, 0);
uint16x8_t v_h = vec_convert_to_fp16(v_hd, 0);
return vec_extract(v_h, 0);
}
#endif
// precomputed f32 table for f16 (256 KB)
// defined in ggml-cpu.c, initialized in ggml_cpu_init()
extern float ggml_table_f32_f16[1 << 16];
// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
// so we define GGML_CPU_FP16_TO_FP32 and GGML_CPU_FP32_TO_FP16 elsewhere for NEON.
// This is also true for POWER9.
#if !defined(GGML_CPU_FP16_TO_FP32)
inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
uint16_t s;
memcpy(&s, &f, sizeof(uint16_t));
return ggml_table_f32_f16[s];
}
#define GGML_CPU_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
#endif
#if !defined(GGML_CPU_FP32_TO_FP16)
#define GGML_CPU_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
#endif
// we define a common set of C macros which map to specific intrinsics based on the current architecture // we define a common set of C macros which map to specific intrinsics based on the current architecture
// we then implement the fundamental computation operations below using only these macros // we then implement the fundamental computation operations below using only these macros
// adding support for new architectures requires to define the corresponding SIMD macros // adding support for new architectures requires to define the corresponding SIMD macros
@ -415,7 +572,7 @@ static inline __m256 __avx_f32cx8_load(const ggml_fp16_t * x) {
float tmp[8]; float tmp[8];
for (int i = 0; i < 8; i++) { for (int i = 0; i < 8; i++) {
tmp[i] = GGML_FP16_TO_FP32(x[i]); tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]);
} }
return _mm256_loadu_ps(tmp); return _mm256_loadu_ps(tmp);
@ -426,7 +583,7 @@ static inline void __avx_f32cx8_store(ggml_fp16_t *x, __m256 y) {
_mm256_storeu_ps(arr, y); _mm256_storeu_ps(arr, y);
for (int i = 0; i < 8; i++) for (int i = 0; i < 8; i++)
x[i] = GGML_FP32_TO_FP16(arr[i]); x[i] = GGML_CPU_FP32_TO_FP16(arr[i]);
} }
#define GGML_F32Cx8_LOAD(x) __avx_f32cx8_load(x) #define GGML_F32Cx8_LOAD(x) __avx_f32cx8_load(x)
#define GGML_F32Cx8_STORE(x, y) __avx_f32cx8_store(x, y) #define GGML_F32Cx8_STORE(x, y) __avx_f32cx8_store(x, y)
@ -574,10 +731,10 @@ static inline unsigned char ggml_endian_byte(int i) {
inline static v128_t __wasm_f16x4_load(const ggml_fp16_t * p) { inline static v128_t __wasm_f16x4_load(const ggml_fp16_t * p) {
float tmp[4]; float tmp[4];
tmp[0] = GGML_FP16_TO_FP32(p[0]); tmp[0] = GGML_CPU_FP16_TO_FP32(p[0]);
tmp[1] = GGML_FP16_TO_FP32(p[1]); tmp[1] = GGML_CPU_FP16_TO_FP32(p[1]);
tmp[2] = GGML_FP16_TO_FP32(p[2]); tmp[2] = GGML_CPU_FP16_TO_FP32(p[2]);
tmp[3] = GGML_FP16_TO_FP32(p[3]); tmp[3] = GGML_CPU_FP16_TO_FP32(p[3]);
return wasm_v128_load(tmp); return wasm_v128_load(tmp);
} }
@ -587,10 +744,10 @@ inline static void __wasm_f16x4_store(ggml_fp16_t * p, v128_t x) {
wasm_v128_store(tmp, x); wasm_v128_store(tmp, x);
p[0] = GGML_FP32_TO_FP16(tmp[0]); p[0] = GGML_CPU_FP32_TO_FP16(tmp[0]);
p[1] = GGML_FP32_TO_FP16(tmp[1]); p[1] = GGML_CPU_FP32_TO_FP16(tmp[1]);
p[2] = GGML_FP32_TO_FP16(tmp[2]); p[2] = GGML_CPU_FP32_TO_FP16(tmp[2]);
p[3] = GGML_FP32_TO_FP16(tmp[3]); p[3] = GGML_CPU_FP32_TO_FP16(tmp[3]);
} }
#define GGML_F16x4 v128_t #define GGML_F16x4 v128_t
@ -690,10 +847,10 @@ inline static void __wasm_f16x4_store(ggml_fp16_t * p, v128_t x) {
static inline __m128 __sse_f16x4_load(const ggml_fp16_t * x) { static inline __m128 __sse_f16x4_load(const ggml_fp16_t * x) {
float tmp[4]; float tmp[4];
tmp[0] = GGML_FP16_TO_FP32(x[0]); tmp[0] = GGML_CPU_FP16_TO_FP32(x[0]);
tmp[1] = GGML_FP16_TO_FP32(x[1]); tmp[1] = GGML_CPU_FP16_TO_FP32(x[1]);
tmp[2] = GGML_FP16_TO_FP32(x[2]); tmp[2] = GGML_CPU_FP16_TO_FP32(x[2]);
tmp[3] = GGML_FP16_TO_FP32(x[3]); tmp[3] = GGML_CPU_FP16_TO_FP32(x[3]);
return _mm_loadu_ps(tmp); return _mm_loadu_ps(tmp);
} }
@ -703,10 +860,10 @@ static inline void __sse_f16x4_store(ggml_fp16_t * x, __m128 y) {
_mm_storeu_ps(arr, y); _mm_storeu_ps(arr, y);
x[0] = GGML_FP32_TO_FP16(arr[0]); x[0] = GGML_CPU_FP32_TO_FP16(arr[0]);
x[1] = GGML_FP32_TO_FP16(arr[1]); x[1] = GGML_CPU_FP32_TO_FP16(arr[1]);
x[2] = GGML_FP32_TO_FP16(arr[2]); x[2] = GGML_CPU_FP32_TO_FP16(arr[2]);
x[3] = GGML_FP32_TO_FP16(arr[3]); x[3] = GGML_CPU_FP32_TO_FP16(arr[3]);
} }
#define GGML_F32Cx4 __m128 #define GGML_F32Cx4 __m128
@ -828,7 +985,7 @@ static inline void __lasx_f32cx8_store(ggml_fp16_t * x, __m256 y) {
#define GGML_F32x4_ZERO __lsx_vldi(0) #define GGML_F32x4_ZERO __lsx_vldi(0)
#define GGML_F32x4_SET1(x) __lsx_vinsgr2vr_w(__lsx_vldi(0),(x), 0) #define GGML_F32x4_SET1(x) __lsx_vinsgr2vr_w(__lsx_vldi(0),(x), 0)
#define GGML_F32x4_LOAD(x) __lsx_vld((x), 0) #define GGML_F32x4_LOAD(x) __lsx_vld((x), 0)
#define GGML_F32x4_STORE((x),(y)) __lsx_vst((y), (x), 0) #define GGML_F32x4_STORE(x, y) __lsx_vst(y, x, 0)
#define GGML_F32x4_FMA(a, b, c) __lsx_vfmadd_s(b, c, a) #define GGML_F32x4_FMA(a, b, c) __lsx_vfmadd_s(b, c, a)
#define GGML_F32x4_ADD __lsx_vfadd_s #define GGML_F32x4_ADD __lsx_vfadd_s
#define GGML_F32x4_MUL __lsx_vfmul_s #define GGML_F32x4_MUL __lsx_vfmul_s
@ -874,10 +1031,10 @@ static inline void __lasx_f32cx8_store(ggml_fp16_t * x, __m256 y) {
static inline __m128 __lsx_f16x4_load(const ggml_fp16_t * x) { static inline __m128 __lsx_f16x4_load(const ggml_fp16_t * x) {
float tmp[4]; float tmp[4];
tmp[0] = GGML_FP16_TO_FP32(x[0]); tmp[0] = GGML_CPU_FP16_TO_FP32(x[0]);
tmp[1] = GGML_FP16_TO_FP32(x[1]); tmp[1] = GGML_CPU_FP16_TO_FP32(x[1]);
tmp[2] = GGML_FP16_TO_FP32(x[2]); tmp[2] = GGML_CPU_FP16_TO_FP32(x[2]);
tmp[3] = GGML_FP16_TO_FP32(x[3]); tmp[3] = GGML_CPU_FP16_TO_FP32(x[3]);
return __lsx_vld(tmp, 0); return __lsx_vld(tmp, 0);
} }
@ -887,10 +1044,10 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
__lsx_vst(y, arr, 0); __lsx_vst(y, arr, 0);
x[0] = GGML_FP32_TO_FP16(arr[0]); x[0] = GGML_CPU_FP32_TO_FP16(arr[0]);
x[1] = GGML_FP32_TO_FP16(arr[1]); x[1] = GGML_CPU_FP32_TO_FP16(arr[1]);
x[2] = GGML_FP32_TO_FP16(arr[2]); x[2] = GGML_CPU_FP32_TO_FP16(arr[2]);
x[3] = GGML_FP32_TO_FP16(arr[3]); x[3] = GGML_CPU_FP32_TO_FP16(arr[3]);
} }
#define GGML_F32Cx4 __m128 #define GGML_F32Cx4 __m128
@ -922,7 +1079,7 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
#define GGML_F32_STEP 32 #define GGML_F32_STEP 32
#define GGML_F32_EPR 4 #define GGML_F32_EPR 4
#define GGML_F32x4 __vector float #define GGML_F32x4 float32x4_t
#define GGML_F32x4_ZERO vec_splats(0.0f) #define GGML_F32x4_ZERO vec_splats(0.0f)
#define GGML_F32x4_SET1 vec_splats #define GGML_F32x4_SET1 vec_splats
#define GGML_F32x4_LOAD(p) vec_xl(0, p) #define GGML_F32x4_LOAD(p) vec_xl(0, p)
@ -962,28 +1119,45 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
#define GGML_F16_STEP GGML_F32_STEP #define GGML_F16_STEP GGML_F32_STEP
#define GGML_F16_EPR GGML_F32_EPR #define GGML_F16_EPR GGML_F32_EPR
static inline __vector float __lzs_f16cx4_load(const ggml_fp16_t * x) { static inline float32x4_t __lzs_f16cx4_load(const ggml_fp16_t * x) {
#if defined(__NNPA__)
uint16x8_t v_x = vec_xl(0, (const ggml_fp16_t *)x);
uint16x8_t v_xd = vec_convert_from_fp16(v_x, 0);
return vec_extend_to_fp32_hi(v_xd, 0);
#else
float tmp[4]; float tmp[4];
for (int i = 0; i < 4; i++) { for (int i = 0; i < 4; i++) {
tmp[i] = GGML_FP16_TO_FP32(x[i]); tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]);
} }
// note: keep type-cast here to prevent compiler bugs // note: keep type-cast here to prevent compiler bugs
// see: https://github.com/ggml-org/llama.cpp/issues/12846 // see: https://github.com/ggml-org/llama.cpp/issues/12846
return vec_xl(0, (const float *)(tmp)); return vec_xl(0, (const float *)(tmp));
#endif
} }
static inline void __lzs_f16cx4_store(ggml_fp16_t * x, __vector float y) { static inline void __lzs_f16cx4_store(ggml_fp16_t * x, float32x4_t v_y) {
#if defined(__NNPA__)
float32x4_t v_zero = vec_splats(0.0f);
uint16x8_t v_xd = vec_round_from_fp32(v_y, v_zero, 0);
uint16x8_t v_x = vec_convert_to_fp16(v_xd, 0);
x[0] = vec_extract(v_x, 0);
x[1] = vec_extract(v_x, 1);
x[2] = vec_extract(v_x, 2);
x[3] = vec_extract(v_x, 3);
#else
float arr[4]; float arr[4];
// note: keep type-cast here to prevent compiler bugs // note: keep type-cast here to prevent compiler bugs
// see: https://github.com/ggml-org/llama.cpp/issues/12846 // see: https://github.com/ggml-org/llama.cpp/issues/12846
vec_xst(y, 0, (float *)(arr)); vec_xst(v_y, 0, (float *)(arr));
for (int i = 0; i < 4; i++) { for (int i = 0; i < 4; i++) {
x[i] = GGML_FP32_TO_FP16(arr[i]); x[i] = GGML_CPU_FP32_TO_FP16(arr[i]);
} }
#endif
} }
#define GGML_F16_VEC GGML_F32x4 #define GGML_F16_VEC GGML_F32x4
@ -1004,3 +1178,7 @@ static inline void __lzs_f16cx4_store(ggml_fp16_t * x, __vector float y) {
#define GGML_F32_ARR (GGML_F32_STEP/GGML_F32_EPR) #define GGML_F32_ARR (GGML_F32_STEP/GGML_F32_EPR)
#define GGML_F16_ARR (GGML_F16_STEP/GGML_F16_EPR) #define GGML_F16_ARR (GGML_F16_STEP/GGML_F16_EPR)
#endif #endif
#ifdef __cplusplus
}
#endif

4
ggml/src/ggml-cpu/vec.cpp
View File

@ -219,11 +219,11 @@ void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * G
// leftovers // leftovers
for (int i = np; i < n; ++i) { for (int i = np; i < n; ++i) {
sumf += (ggml_float)(GGML_FP16_TO_FP32(x[i])*GGML_FP16_TO_FP32(y[i])); sumf += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[i])*GGML_CPU_FP16_TO_FP32(y[i]));
} }
#else #else
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
sumf += (ggml_float)(GGML_FP16_TO_FP32(x[i])*GGML_FP16_TO_FP32(y[i])); sumf += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[i])*GGML_CPU_FP16_TO_FP32(y[i]));
} }
#endif #endif

90
ggml/src/ggml-cpu/vec.h
View File

@ -58,7 +58,7 @@ inline static void ggml_vec_set_bf16(const int n, ggml_bf16_t * x, const ggml_bf
inline static void ggml_vec_add_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i] + y[i]; } inline static void ggml_vec_add_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i] + y[i]; }
inline static void ggml_vec_add_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) { inline static void ggml_vec_add_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) {
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
z[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(x[i]) + GGML_FP16_TO_FP32(y[i])); z[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(x[i]) + GGML_CPU_FP16_TO_FP32(y[i]));
} }
} }
inline static void ggml_vec_add1_f32(const int n, float * z, const float * x, const float v) { for (int i = 0; i < n; ++i) z[i] = x[i] + v; } inline static void ggml_vec_add1_f32(const int n, float * z, const float * x, const float v) { for (int i = 0; i < n; ++i) z[i] = x[i] + v; }
@ -67,7 +67,7 @@ inline static void ggml_vec_acc1_f32(const int n, float * y, const float v)
inline static void ggml_vec_sub_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i] - y[i]; } inline static void ggml_vec_sub_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i] - y[i]; }
inline static void ggml_vec_sub_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) { inline static void ggml_vec_sub_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) {
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
z[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(x[i]) - GGML_FP16_TO_FP32(y[i])); z[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(x[i]) - GGML_CPU_FP16_TO_FP32(y[i]));
} }
} }
inline static void ggml_vec_set_f32 (const int n, float * x, const float v) { for (int i = 0; i < n; ++i) x[i] = v; } inline static void ggml_vec_set_f32 (const int n, float * x, const float v) { for (int i = 0; i < n; ++i) x[i] = v; }
@ -75,20 +75,20 @@ inline static void ggml_vec_cpy_f32 (const int n, float * y, const float * x)
inline static void ggml_vec_neg_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = -x[i]; } inline static void ggml_vec_neg_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = -x[i]; }
inline static void ggml_vec_neg_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { inline static void ggml_vec_neg_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(-GGML_FP16_TO_FP32(x[i])); y[i] = GGML_CPU_FP32_TO_FP16(-GGML_CPU_FP16_TO_FP32(x[i]));
} }
} }
inline static void ggml_vec_mul_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i]*y[i]; } inline static void ggml_vec_mul_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i]*y[i]; }
inline static void ggml_vec_mul_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) { inline static void ggml_vec_mul_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) {
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
z[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(x[i]) * GGML_FP16_TO_FP32(y[i])); z[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(x[i]) * GGML_CPU_FP16_TO_FP32(y[i]));
} }
} }
inline static void ggml_vec_div_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i]/y[i]; } inline static void ggml_vec_div_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i]/y[i]; }
inline static void ggml_vec_div_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) { inline static void ggml_vec_div_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) {
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
z[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(x[i]) / GGML_FP16_TO_FP32(y[i])); z[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(x[i]) / GGML_CPU_FP16_TO_FP32(y[i]));
} }
} }
@ -131,13 +131,13 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG
// leftovers // leftovers
for (int i = np; i < n; ++i) { for (int i = np; i < n; ++i) {
for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) { for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) {
sumf[j] += (ggml_float)(GGML_FP16_TO_FP32(x[j][i])*GGML_FP16_TO_FP32(y[i])); sumf[j] += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[j][i])*GGML_CPU_FP16_TO_FP32(y[i]));
} }
} }
#else #else
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) { for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) {
sumf[j] += (ggml_float)(GGML_FP16_TO_FP32(x[j][i])*GGML_FP16_TO_FP32(y[i])); sumf[j] += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[j][i])*GGML_CPU_FP16_TO_FP32(y[i]));
} }
} }
#endif #endif
@ -280,12 +280,12 @@ inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * GGML_RESTRICT y,
// leftovers // leftovers
for (int i = np; i < n; ++i) { for (int i = np; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i]) + GGML_FP16_TO_FP32(x[i])*v); y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i]) + GGML_CPU_FP16_TO_FP32(x[i])*v);
} }
#else #else
// scalar // scalar
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i]) + GGML_FP16_TO_FP32(x[i])*v); y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i]) + GGML_CPU_FP16_TO_FP32(x[i])*v);
} }
#endif #endif
} }
@ -430,12 +430,12 @@ inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float
// leftovers // leftovers
for (int i = np; i < n; ++i) { for (int i = np; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i])*v); y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
} }
#else #else
// scalar // scalar
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i])*v); y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
} }
#endif #endif
} }
@ -444,103 +444,103 @@ inline static void ggml_vec_norm_f32 (const int n, float * s, const float * x) {
inline static void ggml_vec_sqr_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i]*x[i]; } inline static void ggml_vec_sqr_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i]*x[i]; }
inline static void ggml_vec_sqr_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { inline static void ggml_vec_sqr_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
float v = GGML_FP16_TO_FP32(x[i]); float v = GGML_CPU_FP16_TO_FP32(x[i]);
y[i] = GGML_FP32_TO_FP16(v*v); y[i] = GGML_CPU_FP32_TO_FP16(v*v);
} }
} }
inline static void ggml_vec_sqrt_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sqrtf(x[i]); } inline static void ggml_vec_sqrt_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sqrtf(x[i]); }
inline static void ggml_vec_sqrt_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { inline static void ggml_vec_sqrt_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(sqrtf(GGML_FP16_TO_FP32(x[i]))); y[i] = GGML_CPU_FP32_TO_FP16(sqrtf(GGML_CPU_FP16_TO_FP32(x[i])));
} }
} }
inline static void ggml_vec_log_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = logf(x[i]); } inline static void ggml_vec_log_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = logf(x[i]); }
inline static void ggml_vec_log_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { inline static void ggml_vec_log_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(logf(GGML_FP16_TO_FP32(x[i]))); y[i] = GGML_CPU_FP32_TO_FP16(logf(GGML_CPU_FP16_TO_FP32(x[i])));
} }
} }
inline static void ggml_vec_sin_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sinf(x[i]); } inline static void ggml_vec_sin_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sinf(x[i]); }
inline static void ggml_vec_sin_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { inline static void ggml_vec_sin_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(sinf(GGML_FP16_TO_FP32(x[i]))); y[i] = GGML_CPU_FP32_TO_FP16(sinf(GGML_CPU_FP16_TO_FP32(x[i])));
} }
} }
inline static void ggml_vec_cos_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = cosf(x[i]); } inline static void ggml_vec_cos_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = cosf(x[i]); }
inline static void ggml_vec_cos_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { inline static void ggml_vec_cos_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(cosf(GGML_FP16_TO_FP32(x[i]))); y[i] = GGML_CPU_FP32_TO_FP16(cosf(GGML_CPU_FP16_TO_FP32(x[i])));
} }
} }
inline static void ggml_vec_abs_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fabsf(x[i]); } inline static void ggml_vec_abs_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fabsf(x[i]); }
inline static void ggml_vec_abs_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { inline static void ggml_vec_abs_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(fabsf(GGML_FP16_TO_FP32(x[i]))); y[i] = GGML_CPU_FP32_TO_FP16(fabsf(GGML_CPU_FP16_TO_FP32(x[i])));
} }
} }
inline static void ggml_vec_sgn_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : ((x[i] < 0.f) ? -1.f : 0.f); } inline static void ggml_vec_sgn_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : ((x[i] < 0.f) ? -1.f : 0.f); }
inline static void ggml_vec_sgn_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { inline static void ggml_vec_sgn_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
float v = GGML_FP16_TO_FP32(x[i]); float v = GGML_CPU_FP16_TO_FP32(x[i]);
y[i] = GGML_FP32_TO_FP16((v > 0.f) ? 1.f : ((v < 0.f) ? -1.f : 0.f)); y[i] = GGML_CPU_FP32_TO_FP16((v > 0.f) ? 1.f : ((v < 0.f) ? -1.f : 0.f));
} }
} }
inline static void ggml_vec_step_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : 0.f; } inline static void ggml_vec_step_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : 0.f; }
inline static void ggml_vec_step_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { inline static void ggml_vec_step_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16((GGML_FP16_TO_FP32(x[i]) > 0.f) ? 1.f : 0.f); y[i] = GGML_CPU_FP32_TO_FP16((GGML_CPU_FP16_TO_FP32(x[i]) > 0.f) ? 1.f : 0.f);
} }
} }
inline static void ggml_vec_tanh_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = tanhf(x[i]); } inline static void ggml_vec_tanh_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = tanhf(x[i]); }
inline static void ggml_vec_tanh_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { inline static void ggml_vec_tanh_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(tanhf(GGML_FP16_TO_FP32(x[i]))); y[i] = GGML_CPU_FP32_TO_FP16(tanhf(GGML_CPU_FP16_TO_FP32(x[i])));
} }
} }
inline static void ggml_vec_elu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expm1f(x[i]); } inline static void ggml_vec_elu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expm1f(x[i]); }
inline static void ggml_vec_elu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { inline static void ggml_vec_elu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(expm1f(GGML_FP16_TO_FP32(x[i]))); y[i] = GGML_CPU_FP32_TO_FP16(expm1f(GGML_CPU_FP16_TO_FP32(x[i])));
} }
} }
inline static void ggml_vec_relu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.f; } inline static void ggml_vec_relu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.f; }
inline static void ggml_vec_relu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { inline static void ggml_vec_relu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
float v = GGML_FP16_TO_FP32(x[i]); float v = GGML_CPU_FP16_TO_FP32(x[i]);
y[i] = GGML_FP32_TO_FP16((v > 0.f) ? v : 0.f); y[i] = GGML_CPU_FP32_TO_FP16((v > 0.f) ? v : 0.f);
} }
} }
inline static void ggml_vec_leaky_relu_f32 (const int n, float * y, const float * x, const float ns) { for (int i = 0; i < n; ++i) y[i] = ((x[i] > 0.f) ? x[i] : 0.f) + ns * ((x[i] < 0.0f) ? x[i] : 0.f); } inline static void ggml_vec_leaky_relu_f32 (const int n, float * y, const float * x, const float ns) { for (int i = 0; i < n; ++i) y[i] = ((x[i] > 0.f) ? x[i] : 0.f) + ns * ((x[i] < 0.0f) ? x[i] : 0.f); }
inline static void ggml_vec_leaky_relu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const float ns) { inline static void ggml_vec_leaky_relu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const float ns) {
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
float v = GGML_FP16_TO_FP32(x[i]); float v = GGML_CPU_FP16_TO_FP32(x[i]);
y[i] = GGML_FP32_TO_FP16(((v > 0.f) ? v : 0.f) + ns * ((v < 0.0f) ? v : 0.f)); y[i] = GGML_CPU_FP32_TO_FP16(((v > 0.f) ? v : 0.f) + ns * ((v < 0.0f) ? v : 0.f));
} }
} }
inline static void ggml_vec_sigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = 1.f / (1.f + expf(-x[i])); } inline static void ggml_vec_sigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = 1.f / (1.f + expf(-x[i])); }
inline static void ggml_vec_sigmoid_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { inline static void ggml_vec_sigmoid_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(1.f / (1.f + expf(-GGML_FP16_TO_FP32(x[i])))); y[i] = GGML_CPU_FP32_TO_FP16(1.f / (1.f + expf(-GGML_CPU_FP16_TO_FP32(x[i]))));
} }
} }
// TODO: optimize performance // TODO: optimize performance
inline static void ggml_vec_hardswish_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i] * fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); } inline static void ggml_vec_hardswish_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i] * fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); }
inline static void ggml_vec_hardswish_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { inline static void ggml_vec_hardswish_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
float v = GGML_FP16_TO_FP32(x[i]); float v = GGML_CPU_FP16_TO_FP32(x[i]);
y[i] = GGML_FP32_TO_FP16(v * fminf(1.0f, fmaxf(0.0f, (v + 3.0f) / 6.0f))); y[i] = GGML_CPU_FP32_TO_FP16(v * fminf(1.0f, fmaxf(0.0f, (v + 3.0f) / 6.0f)));
} }
} }
inline static void ggml_vec_hardsigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); } inline static void ggml_vec_hardsigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); }
inline static void ggml_vec_hardsigmoid_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { inline static void ggml_vec_hardsigmoid_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(fminf(1.0f, fmaxf(0.0f, (GGML_FP16_TO_FP32(x[i]) + 3.0f) / 6.0f))); y[i] = GGML_CPU_FP32_TO_FP16(fminf(1.0f, fmaxf(0.0f, (GGML_CPU_FP16_TO_FP32(x[i]) + 3.0f) / 6.0f)));
} }
} }
inline static void ggml_vec_exp_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = expf(x[i]); } inline static void ggml_vec_exp_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = expf(x[i]); }
inline static void ggml_vec_exp_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { inline static void ggml_vec_exp_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(expf(GGML_FP16_TO_FP32(x[i]))); y[i] = GGML_CPU_FP32_TO_FP16(expf(GGML_CPU_FP16_TO_FP32(x[i])));
} }
} }
@ -562,9 +562,9 @@ inline static void ggml_vec_gelu_f16(const int n, ggml_fp16_t * y, const ggml_fp
inline static void ggml_vec_gelu_erf_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { inline static void ggml_vec_gelu_erf_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
float xi = GGML_FP16_TO_FP32(x[i]); float xi = GGML_CPU_FP16_TO_FP32(x[i]);
float res = 0.5f*xi*(1.0f + erff(xi*SQRT_2_INV)); float res = 0.5f*xi*(1.0f + erff(xi*SQRT_2_INV));
y[i] = GGML_FP32_TO_FP16(res); y[i] = GGML_CPU_FP32_TO_FP16(res);
} }
} }
@ -577,9 +577,9 @@ inline static void ggml_vec_gelu_f32(const int n, float * y, const float * x) {
} else if (x[i] >= 10.0f) { } else if (x[i] >= 10.0f) {
y[i] = x[i]; y[i] = x[i];
} else { } else {
ggml_fp16_t fp16 = GGML_FP32_TO_FP16(x[i]); ggml_fp16_t fp16 = GGML_CPU_FP32_TO_FP16(x[i]);
memcpy(&t, &fp16, sizeof(uint16_t)); memcpy(&t, &fp16, sizeof(uint16_t));
y[i] = GGML_FP16_TO_FP32(ggml_table_gelu_f16[t]); y[i] = GGML_CPU_FP16_TO_FP32(ggml_table_gelu_f16[t]);
} }
} }
} }
@ -613,9 +613,9 @@ inline static float ggml_gelu_quick_f32(float x) {
inline static void ggml_vec_gelu_quick_f32(const int n, float * y, const float * x) { inline static void ggml_vec_gelu_quick_f32(const int n, float * y, const float * x) {
uint16_t t; uint16_t t;
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
ggml_fp16_t fp16 = GGML_FP32_TO_FP16(x[i]); ggml_fp16_t fp16 = GGML_CPU_FP32_TO_FP16(x[i]);
memcpy(&t, &fp16, sizeof(uint16_t)); memcpy(&t, &fp16, sizeof(uint16_t));
y[i] = GGML_FP16_TO_FP32(ggml_table_gelu_quick_f16[t]); y[i] = GGML_CPU_FP16_TO_FP32(ggml_table_gelu_quick_f16[t]);
} }
} }
#else #else
@ -628,8 +628,8 @@ inline static void ggml_vec_gelu_quick_f32(const int n, float * y, const float *
inline static void ggml_vec_gelu_quick_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { inline static void ggml_vec_gelu_quick_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
float v = GGML_FP16_TO_FP32(x[i]); float v = GGML_CPU_FP16_TO_FP32(x[i]);
y[i] = GGML_FP32_TO_FP16(v*(1.0f/(1.0f+expf(GELU_QUICK_COEF*v)))); y[i] = GGML_CPU_FP32_TO_FP16(v*(1.0f/(1.0f+expf(GELU_QUICK_COEF*v))));
} }
} }
@ -638,8 +638,8 @@ inline static float ggml_silu_f32(float x) {
return x/(1.0f + expf(-x)); return x/(1.0f + expf(-x));
} }
inline static ggml_fp16_t ggml_silu_f16(ggml_fp16_t x) { inline static ggml_fp16_t ggml_silu_f16(ggml_fp16_t x) {
float v = GGML_FP16_TO_FP32(x); float v = GGML_CPU_FP16_TO_FP32(x);
return GGML_FP32_TO_FP16(v/(1.0f + expf(-v))); return GGML_CPU_FP32_TO_FP16(v/(1.0f + expf(-v)));
} }
#if __FINITE_MATH_ONLY__ #if __FINITE_MATH_ONLY__
@ -888,9 +888,9 @@ inline static float ggml_silu_backward_f32(float x, float dy) {
} }
inline static ggml_fp16_t ggml_silu_backward_f16(ggml_fp16_t x, ggml_fp16_t dy) { inline static ggml_fp16_t ggml_silu_backward_f16(ggml_fp16_t x, ggml_fp16_t dy) {
const float v = GGML_FP16_TO_FP32(x); const float v = GGML_CPU_FP16_TO_FP32(x);
const float s = 1.0f/(1.0f + expf(-v)); const float s = 1.0f/(1.0f + expf(-v));
return GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(dy)*s*(1.0f + v*(1.0f - s))); return GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(dy)*s*(1.0f + v*(1.0f - s)));
} }
inline static void ggml_vec_silu_backward_f32(const int n, float * dx, const float * x, const float * dy) { inline static void ggml_vec_silu_backward_f32(const int n, float * dx, const float * x, const float * dy) {
@ -928,7 +928,7 @@ inline static void ggml_vec_sum_f32_ggf(const int n, ggml_float * s, const float
inline static void ggml_vec_sum_f16_ggf(const int n, float * s, const ggml_fp16_t * x) { inline static void ggml_vec_sum_f16_ggf(const int n, float * s, const ggml_fp16_t * x) {
float sum = 0.0f; float sum = 0.0f;
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
sum += GGML_FP16_TO_FP32(x[i]); sum += GGML_CPU_FP16_TO_FP32(x[i]);
} }
*s = sum; *s = sum;
} }

262
ggml/src/ggml-impl.h
View File

@ -317,203 +317,81 @@ struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph, int i0, int i1);
GGML_API void * ggml_aligned_malloc(size_t size); GGML_API void * ggml_aligned_malloc(size_t size);
GGML_API void ggml_aligned_free(void * ptr, size_t size); GGML_API void ggml_aligned_free(void * ptr, size_t size);
// FP16 to FP32 conversion // FP16 <-> FP32
// ref: https://github.com/Maratyszcza/FP16
// 16-bit float static inline float fp32_from_bits(uint32_t w) {
// on Arm, we use __fp16 union {
// on x86, we use uint16_t uint32_t as_bits;
// float as_value;
// for old CUDA compilers (<= 11), we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/10616 } fp32;
// for MUSA compilers , we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/11843 fp32.as_bits = w;
// return fp32.as_value;
#if defined(__ARM_NEON) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) && !defined(__MUSACC__)
#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
__fp16 tmp;
memcpy(&tmp, &h, sizeof(ggml_fp16_t));
return (float)tmp;
}
static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
ggml_fp16_t res;
__fp16 tmp = f;
memcpy(&res, &tmp, sizeof(ggml_fp16_t));
return res;
}
#elif defined(__F16C__)
#ifdef _MSC_VER
#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
#else
#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
#endif
#elif defined(__POWER9_VECTOR__)
#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
/* the inline asm below is about 12% faster than the lookup method */
#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
float f;
double d;
__asm__(
"mtfprd %0,%2\n"
"xscvhpdp %0,%0\n"
"frsp %1,%0\n" :
/* temp */ "=d"(d),
/* out */ "=f"(f):
/* in */ "r"(h));
return f;
}
static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
double d;
ggml_fp16_t r;
__asm__( /* xscvdphp can work on double or single precision */
"xscvdphp %0,%2\n"
"mffprd %1,%0\n" :
/* temp */ "=d"(d),
/* out */ "=r"(r):
/* in */ "f"(f));
return r;
}
#elif defined(__riscv) && defined(__riscv_zfhmin)
static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
float f;
__asm__(
"fmv.h.x %[f], %[h]\n\t"
"fcvt.s.h %[f], %[f]"
: [f] "=&f" (f)
: [h] "r" (h)
);
return f;
}
static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
ggml_fp16_t res;
__asm__(
"fcvt.h.s %[f], %[f]\n\t"
"fmv.x.h %[h], %[f]"
: [h] "=&r" (res)
: [f] "f" (f)
);
return res;
}
#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
#else
// FP16 <-> FP32
// ref: https://github.com/Maratyszcza/FP16
static inline float fp32_from_bits(uint32_t w) {
union {
uint32_t as_bits;
float as_value;
} fp32;
fp32.as_bits = w;
return fp32.as_value;
}
static inline uint32_t fp32_to_bits(float f) {
union {
float as_value;
uint32_t as_bits;
} fp32;
fp32.as_value = f;
return fp32.as_bits;
}
static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
const uint32_t w = (uint32_t) h << 16;
const uint32_t sign = w & UINT32_C(0x80000000);
const uint32_t two_w = w + w;
const uint32_t exp_offset = UINT32_C(0xE0) << 23;
#if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)) && (!defined(__cplusplus) || __cplusplus >= 201703L)
const float exp_scale = 0x1.0p-112f;
#else
const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
#endif
const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
const uint32_t magic_mask = UINT32_C(126) << 23;
const float magic_bias = 0.5f;
const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
const uint32_t result = sign |
(two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
return fp32_from_bits(result);
}
static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
#if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)) && (!defined(__cplusplus) || __cplusplus >= 201703L)
const float scale_to_inf = 0x1.0p+112f;
const float scale_to_zero = 0x1.0p-110f;
#else
const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
#endif
float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
const uint32_t w = fp32_to_bits(f);
const uint32_t shl1_w = w + w;
const uint32_t sign = w & UINT32_C(0x80000000);
uint32_t bias = shl1_w & UINT32_C(0xFF000000);
if (bias < UINT32_C(0x71000000)) {
bias = UINT32_C(0x71000000);
}
base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
const uint32_t bits = fp32_to_bits(base);
const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
const uint32_t nonsign = exp_bits + mantissa_bits;
return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
}
#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
#endif // defined(__ARM_NEON) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) && !defined(__MUSACC__)
// precomputed f32 table for f16 (256 KB)
// defined in ggml.c, initialized in ggml_init()
GGML_API float ggml_table_f32_f16[1 << 16];
// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
// This is also true for POWER9.
#if !defined(GGML_FP16_TO_FP32)
inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
uint16_t s;
memcpy(&s, &f, sizeof(uint16_t));
return ggml_table_f32_f16[s];
} }
#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x) static inline uint32_t fp32_to_bits(float f) {
#endif union {
float as_value;
uint32_t as_bits;
} fp32;
fp32.as_value = f;
return fp32.as_bits;
}
#if !defined(GGML_FP32_TO_FP16) static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x) const uint32_t w = (uint32_t) h << 16;
const uint32_t sign = w & UINT32_C(0x80000000);
const uint32_t two_w = w + w;
const uint32_t exp_offset = UINT32_C(0xE0) << 23;
#if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)) && (!defined(__cplusplus) || __cplusplus >= 201703L)
const float exp_scale = 0x1.0p-112f;
#else
const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
#endif #endif
const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
const uint32_t magic_mask = UINT32_C(126) << 23;
const float magic_bias = 0.5f;
const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
const uint32_t result = sign |
(two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
return fp32_from_bits(result);
}
static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
#if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)) && (!defined(__cplusplus) || __cplusplus >= 201703L)
const float scale_to_inf = 0x1.0p+112f;
const float scale_to_zero = 0x1.0p-110f;
#else
const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
#endif
float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
const uint32_t w = fp32_to_bits(f);
const uint32_t shl1_w = w + w;
const uint32_t sign = w & UINT32_C(0x80000000);
uint32_t bias = shl1_w & UINT32_C(0xFF000000);
if (bias < UINT32_C(0x71000000)) {
bias = UINT32_C(0x71000000);
}
base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
const uint32_t bits = fp32_to_bits(base);
const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
const uint32_t nonsign = exp_bits + mantissa_bits;
return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
}
#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
/** /**
* Converts brain16 to float32. * Converts brain16 to float32.

11
ggml/src/ggml.c
View File

@ -61,9 +61,6 @@
#define m512i(p) (__m512i)(p) #define m512i(p) (__m512i)(p)
#endif #endif
// precomputed f32 table for f16 (256 KB) (ggml-impl.h)
float ggml_table_f32_f16[1 << 16];
#if defined(__linux__) || \ #if defined(__linux__) || \
defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__) || \ defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__) || \
(defined(__APPLE__) && !TARGET_OS_TV && !TARGET_OS_WATCH) (defined(__APPLE__) && !TARGET_OS_TV && !TARGET_OS_WATCH)
@ -1422,14 +1419,6 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
// initialize time system (required on Windows) // initialize time system (required on Windows)
ggml_time_init(); ggml_time_init();
for (int i = 0; i < (1 << 16); ++i) {
union {
uint16_t u16;
ggml_fp16_t fp16;
} u = {i};
ggml_table_f32_f16[i] = GGML_COMPUTE_FP16_TO_FP32(u.fp16);
}
is_first_call = false; is_first_call = false;
} }