mirror of
https://github.com/ggerganov/whisper.cpp.git
synced 2025-08-14 17:18:43 +02:00
whisper : add integer quantization support (#540)
* whisper : add integer quantization support * examples : add common-ggml + prepare to add "quantize" tool * whisper : quantization tool ready * whisper : fix F32 support * whisper : try to fix shared lib linkage * wasm : update quantized models to Q5 * bench.wasm : remove "medium" button * bench.wasm : fix custom model button * ggml : add Q5_0 and Q5_1 WASM SIMD * wasm : add quantized models to all WASM examples * wasm : bump DB version number to 2 * talk-llama : update example to latest llama.cpp * node : increase test timeout to 10s * readme : add information for model quantization * wasm : add links to other examples
This commit is contained in:
Georgi Gerganov
GitHub
162
ggml.c
162
ggml.c
@ -330,7 +330,7 @@ static ggml_fp16_t table_exp_f16[1 << 16];
|
||||
// precomputed f32 table for f16 (256 KB)
|
||||
static float table_f32_f16[1 << 16];
|
||||
|
||||
#if defined(__ARM_NEON)
|
||||
#if defined(__ARM_NEON) || defined(__wasm_simd128__)
|
||||
#define B1(c,s,n) 0x ## n ## c , 0x ## n ## s
|
||||
#define B2(c,s,n) B1(c,s,n ## c), B1(c,s,n ## s)
|
||||
#define B3(c,s,n) B2(c,s,n ## c), B2(c,s,n ## s)
|
||||
@ -3180,6 +3180,72 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void *
|
||||
}
|
||||
|
||||
*s = vaddvq_f32(sumv);
|
||||
#elif defined(__wasm_simd128__)
|
||||
v128_t sumv = wasm_f32x4_splat(0.0f);
|
||||
|
||||
uint64_t tmp[4];
|
||||
|
||||
for (int i = 0; i < nb; ++i) {
|
||||
const block_q5_0 * restrict x0 = &x[i];
|
||||
const block_q8_0 * restrict y0 = &y[i];
|
||||
|
||||
const v128_t m4b = wasm_i8x16_splat(0x0F);
|
||||
const v128_t s16b = wasm_i8x16_splat(0x10);
|
||||
|
||||
// extract the 5th bit
|
||||
uint32_t qh;
|
||||
memcpy(&qh, x0->qh, sizeof(qh));
|
||||
|
||||
tmp[0] = table_b2b_u[(qh >> 0) & 0xFF];
|
||||
tmp[1] = table_b2b_u[(qh >> 8) & 0xFF];
|
||||
tmp[2] = table_b2b_u[(qh >> 16) & 0xFF];
|
||||
tmp[3] = table_b2b_u[(qh >> 24) ];
|
||||
|
||||
const v128_t qhl = wasm_v128_load(tmp + 0);
|
||||
const v128_t qhh = wasm_v128_load(tmp + 2);
|
||||
|
||||
const v128_t v0 = wasm_v128_load(x0->qs);
|
||||
|
||||
// 4-bit -> 8-bit
|
||||
const v128_t v0l = wasm_v128_and (v0, m4b);
|
||||
const v128_t v0h = wasm_u8x16_shr(v0, 4);
|
||||
|
||||
// interleave
|
||||
const v128_t v0lz = wasm_v8x16_shuffle(v0l, v0h, 0, 16, 1, 17, 2, 18, 3, 19, 4, 20, 5, 21, 6, 22, 7, 23);
|
||||
const v128_t v0hz = wasm_v8x16_shuffle(v0l, v0h, 8, 24, 9, 25, 10, 26, 11, 27, 12, 28, 13, 29, 14, 30, 15, 31);
|
||||
|
||||
// add high bit and sub 16
|
||||
const v128_t v0lf = wasm_i8x16_sub(wasm_v128_or(v0lz, qhl), s16b);
|
||||
const v128_t v0hf = wasm_i8x16_sub(wasm_v128_or(v0hz, qhh), s16b);
|
||||
|
||||
// load y
|
||||
const v128_t v1l = wasm_v128_load(y0->qs);
|
||||
const v128_t v1h = wasm_v128_load(y0->qs + 16);
|
||||
|
||||
// int8x16 -> int16x8
|
||||
const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf);
|
||||
const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf);
|
||||
const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf);
|
||||
const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf);
|
||||
|
||||
const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l);
|
||||
const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l);
|
||||
const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h);
|
||||
const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h);
|
||||
|
||||
const float x0d = GGML_FP16_TO_FP32(x0->d);
|
||||
|
||||
// dot product
|
||||
sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(
|
||||
wasm_i32x4_add(
|
||||
wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll),
|
||||
wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
|
||||
wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
|
||||
wasm_i32x4_dot_i16x8(v0hfh, v1hh)))), wasm_f32x4_splat(x0d*y0->d)));
|
||||
}
|
||||
|
||||
*s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
|
||||
wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3);
|
||||
#elif defined(__AVX2__)
|
||||
// Initialize accumulator with zeros
|
||||
__m256 acc = _mm256_setzero_ps();
|
||||
@ -3311,6 +3377,77 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void *
|
||||
}
|
||||
|
||||
*s = vaddvq_f32(sumv) + summs;
|
||||
#elif defined(__wasm_simd128__)
|
||||
v128_t sumv = wasm_f32x4_splat(0.0f);
|
||||
|
||||
float summs = 0.0f;
|
||||
|
||||
uint64_t tmp[4];
|
||||
|
||||
for (int i = 0; i < nb; ++i) {
|
||||
const block_q5_1 * restrict x0 = &x[i];
|
||||
const block_q8_1 * restrict y0 = &y[i];
|
||||
|
||||
summs += GGML_FP16_TO_FP32(x0->m) * (y0->s0 + y0->s1);
|
||||
|
||||
const v128_t m4b = wasm_i8x16_splat(0x0F);
|
||||
|
||||
// extract the 5th bit
|
||||
uint32_t qh;
|
||||
memcpy(&qh, x0->qh, sizeof(qh));
|
||||
|
||||
tmp[0] = table_b2b_u[(qh >> 0) & 0xFF];
|
||||
tmp[1] = table_b2b_u[(qh >> 8) & 0xFF];
|
||||
tmp[2] = table_b2b_u[(qh >> 16) & 0xFF];
|
||||
tmp[3] = table_b2b_u[(qh >> 24) ];
|
||||
|
||||
const v128_t qhl = wasm_v128_load(tmp + 0);
|
||||
const v128_t qhh = wasm_v128_load(tmp + 2);
|
||||
|
||||
const v128_t v0 = wasm_v128_load(x0->qs);
|
||||
|
||||
// 4-bit -> 8-bit
|
||||
const v128_t v0l = wasm_v128_and (v0, m4b);
|
||||
const v128_t v0h = wasm_u8x16_shr(v0, 4);
|
||||
|
||||
static bool x = true;
|
||||
|
||||
// interleave
|
||||
const v128_t v0lz = wasm_v8x16_shuffle(v0l, v0h, 0, 16, 1, 17, 2, 18, 3, 19, 4, 20, 5, 21, 6, 22, 7, 23);
|
||||
const v128_t v0hz = wasm_v8x16_shuffle(v0l, v0h, 8, 24, 9, 25, 10, 26, 11, 27, 12, 28, 13, 29, 14, 30, 15, 31);
|
||||
|
||||
// add high bit
|
||||
const v128_t v0lf = wasm_v128_or(v0lz, qhl);
|
||||
const v128_t v0hf = wasm_v128_or(v0hz, qhh);
|
||||
|
||||
// load y
|
||||
const v128_t v1l = wasm_v128_load(y0->qs);
|
||||
const v128_t v1h = wasm_v128_load(y0->qs + 16);
|
||||
|
||||
// int8x16 -> int16x8
|
||||
const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf);
|
||||
const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf);
|
||||
const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf);
|
||||
const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf);
|
||||
|
||||
const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l);
|
||||
const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l);
|
||||
const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h);
|
||||
const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h);
|
||||
|
||||
const float x0d = GGML_FP16_TO_FP32(x0->d);
|
||||
|
||||
// dot product
|
||||
sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(
|
||||
wasm_i32x4_add(
|
||||
wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll),
|
||||
wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
|
||||
wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
|
||||
wasm_i32x4_dot_i16x8(v0hfh, v1hh)))), wasm_f32x4_splat(x0d*y0->d)));
|
||||
}
|
||||
|
||||
*s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
|
||||
wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3) + summs;
|
||||
#elif defined(__AVX2__)
|
||||
// Initialize accumulator with zeros
|
||||
__m256 acc = _mm256_setzero_ps();
|
||||
@ -3827,6 +3964,7 @@ static const char * GGML_OP_LABEL[GGML_OP_COUNT] = {
|
||||
"DIAG_MASK_INF",
|
||||
"SOFT_MAX",
|
||||
"ROPE",
|
||||
"ALIBI",
|
||||
"CONV_1D_1S",
|
||||
"CONV_1D_2S",
|
||||
|
||||
@ -3875,6 +4013,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
|
||||
"diag_mask_inf(x)",
|
||||
"soft_max(x)",
|
||||
"rope(x)",
|
||||
"alibi(x)",
|
||||
"conv_1d_1s(x)",
|
||||
"conv_1d_2s(x)",
|
||||
|
||||
@ -4055,6 +4194,27 @@ bool ggml_is_quantized(enum ggml_type type) {
|
||||
return GGML_IS_QUANTIZED[type];
|
||||
}
|
||||
|
||||
enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
|
||||
enum ggml_type wtype = GGML_TYPE_COUNT;
|
||||
|
||||
switch (ftype) {
|
||||
case GGML_FTYPE_ALL_F32: wtype = GGML_TYPE_F32; break;
|
||||
case GGML_FTYPE_MOSTLY_F16: wtype = GGML_TYPE_F16; break;
|
||||
case GGML_FTYPE_MOSTLY_Q4_0: wtype = GGML_TYPE_Q4_0; break;
|
||||
case GGML_FTYPE_MOSTLY_Q4_1: wtype = GGML_TYPE_Q4_1; break;
|
||||
case GGML_FTYPE_MOSTLY_Q4_2: wtype = GGML_TYPE_Q4_2; break;
|
||||
case GGML_FTYPE_MOSTLY_Q5_0: wtype = GGML_TYPE_Q5_0; break;
|
||||
case GGML_FTYPE_MOSTLY_Q5_1: wtype = GGML_TYPE_Q5_1; break;
|
||||
case GGML_FTYPE_MOSTLY_Q8_0: wtype = GGML_TYPE_Q8_0; break;
|
||||
case GGML_FTYPE_UNKNOWN: wtype = GGML_TYPE_COUNT; break;
|
||||
case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16: wtype = GGML_TYPE_COUNT; break;
|
||||
}
|
||||
|
||||
GGML_ASSERT(wtype != GGML_TYPE_COUNT);
|
||||
|
||||
return wtype;
|
||||
}
|
||||
|
||||
static inline bool ggml_is_transposed(const struct ggml_tensor * tensor) {
|
||||
return tensor->nb[0] > tensor->nb[1];
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user