extern/whisper.cpp
1
0
Fork 1
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-08-19 06:17:03 +02:00
Code Issues Packages Projects Releases Wiki Activity

whisper : add full CUDA and Metal offloading (#1472)

Browse Source 
* whisper : migrate to ggml-backend

* whisper : fix logit reading

* whisper : fix tensor allocation during load

* whisper : fix beam-search with CUDA

* whisper : free backends + fix compile warning

* whisper : print when CUDA is enabled

* whisper : fix CoreML

* make : clean-up

* talk : fix compile warning

* whisper : support ggml_conv with CUDA and Metal (#1473)

* ggml : add CUDA support for ggml_conv

* whisper : remove ggml_repeat for conv bias + single backend

* cuda : fix im2col kernel

* metal : add im2col support + mul mat-vec f16 x f16

* bench-all : add q4 models

* whisper : clean-up

* quantize-all : fix

* ggml : im2col opts

* whisper : avoid whisper_model_data wrapper

* whisper : add note that ggml_mul_mat_pad does not work with CUDA

* whisper : factor out graph compute in common function

* whisper : fixes

* whisper : fix UB with measure buffers

* whisper : try to fix the parallel whisper_state functionality (#1479)

* whisper : try to fix the parallel whisper_state functionality

* whisper : fix multi-state Metal

* whisper : free backend instances in whisper_state
This commit is contained in:
Georgi Gerganov
2023-11-12 15:31:08 +02:00
committed by GitHub
parent ec7a6f04f9
commit b0502836b8
14 changed files with 916 additions and 1594 deletions
Download Patch File Download Diff File Expand all files Collapse all files

96
ggml-cuda.cu
View File

@@ -4476,6 +4476,13 @@ static __device__ void cpy_1_f32_f16(const char * cxi, char * cdsti) {
*dsti = __float2half(*xi);
}
static __device__ void cpy_1_f16_f16(const char * cxi, char * cdsti) {
const half * xi = (const half *) cxi;
half * dsti = (half *) cdsti;
*dsti = *xi;
}
template <cpy_kernel_t cpy_1>
static __global__ void cpy_f32_f16(const char * cx, char * cdst, const int ne,
const int ne00, const int ne01, const int nb00, const int nb01, const int nb02,
@@ -4729,6 +4736,25 @@ static __global__ void clamp_f32(const float * x, float * dst, const float min,
dst[i] = x[i] < min ? min : (x[i] > max ? max : x[i]);
}
static __global__ void im2col_f32_f16(
const float * x, half * dst,
int ofs0, int ofs1, int IW, int IH, int CHW,
int s0, int s1, int p0, int p1, int d0, int d1) {
const int iiw = blockIdx.z * s0 + threadIdx.z * d0 - p0;
const int iih = blockIdx.y * s1 + threadIdx.y * d1 - p1;
const int offset_dst =
(threadIdx.x * gridDim.y * gridDim.z + blockIdx.y * gridDim.z + blockIdx.z) * CHW +
(blockIdx.x * (blockDim.y * blockDim.z) + threadIdx.y * blockDim.z + threadIdx.z);
if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
dst[offset_dst] = __float2half(0.0f);
} else {
const int offset_src = threadIdx.x * ofs0 + blockIdx.x * ofs1;
dst[offset_dst] = __float2half(x[offset_src + iih * IW + iiw]);
}
}
template<int qk, int qr, dequantize_kernel_t dq>
static void get_rows_cuda(const void * x, const int32_t * y, float * dst, const int nrows, const int ncols, cudaStream_t stream) {
const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
@@ -5618,6 +5644,16 @@ static void ggml_cpy_f32_f16_cuda(
(cx, cdst, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12);
}
static void ggml_cpy_f16_f16_cuda(
const char * cx, char * cdst, const int ne,
const int ne00, const int ne01, const int nb00, const int nb01, const int nb02,
const int ne10, const int ne11, const int nb10, const int nb11, const int nb12, cudaStream_t stream) {
const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
cpy_f32_f16<cpy_1_f16_f16><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
(cx, cdst, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12);
}
static void scale_f32_cuda(const float * x, float * dst, const float scale, const int k, cudaStream_t stream) {
const int num_blocks = (k + CUDA_SCALE_BLOCK_SIZE - 1) / CUDA_SCALE_BLOCK_SIZE;
scale_f32<<<num_blocks, CUDA_SCALE_BLOCK_SIZE, 0, stream>>>(x, dst, scale, k);
@@ -5701,6 +5737,15 @@ static void soft_max_f32_cuda(const float * x, float * dst, const int ncols_x, c
soft_max_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols_x);
}
static void im2col_f32_f16_cuda(const float * x, half * dst,
int OH, int IW, int IH, int OW, int IC,
int KH, int KW, int N, int ofs0, int ofs1,
int s0, int s1, int p0, int p1, int d0, int d1, cudaStream_t stream) {
dim3 block_nums(IC, OH, OW);
dim3 block_dims(N, KH, KW);
im2col_f32_f16<<<block_nums, block_dims, 0, stream>>>(x, dst, ofs0, ofs1, IW, IH, (IC * KH * KW), s0, s1, p0, p1, d0, d1);
}
// buffer pool for cuda
#define MAX_CUDA_BUFFERS 256
@@ -6483,7 +6528,7 @@ inline void ggml_cuda_op_mul_mat_cublas(
src1_as_f16 = (half *) ggml_cuda_pool_malloc_async(ne * sizeof(half), &src1_as, id, stream);
to_fp16_cuda(src1_ddf_i, src1_as_f16, ne, stream);
}
const half * src1_ptr = src1->type == GGML_TYPE_F16 ? (const half *) src1_ddq_i : src1_as_f16;
const half * src1_ptr = src1->type == GGML_TYPE_F16 ? (const half *) src1_ddf_i : src1_as_f16;
size_t dst_f16_as = 0;
half * dst_f16 = (half *) ggml_cuda_pool_malloc_async(row_diff*src1_ncols * sizeof(half), &dst_f16_as, id, stream);
@@ -6659,6 +6704,45 @@ inline void ggml_cuda_op_alibi(
(void) src1_dd;
}
inline void ggml_cuda_op_im2col(
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
GGML_ASSERT(src0->type == GGML_TYPE_F16);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F16);
const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
const int32_t s1 = ((const int32_t*)(dst->op_params))[1];
const int32_t p0 = ((const int32_t*)(dst->op_params))[2];
const int32_t p1 = ((const int32_t*)(dst->op_params))[3];
const int32_t d0 = ((const int32_t*)(dst->op_params))[4];
const int32_t d1 = ((const int32_t*)(dst->op_params))[5];
const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1;
const int64_t N = src1->ne[is_2D ? 3 : 2];
const int64_t IC = src1->ne[is_2D ? 2 : 1];
const int64_t IH = is_2D ? src1->ne[1] : 1;
const int64_t IW = src1->ne[0];
const int64_t KH = is_2D ? src0->ne[1] : 1;
const int64_t KW = src0->ne[0];
const int64_t OH = is_2D ? dst->ne[2] : 1;
const int64_t OW = dst->ne[1];
const size_t ofs0 = src1->nb[is_2D ? 3 : 2] / 4; // nb is byte offset, src is type float32
const size_t ofs1 = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32
im2col_f32_f16_cuda(src1_dd, (half*) dst_dd,
OH, IW, IH, OW, IC, KH, KW, N,
ofs0, ofs1, s0, s1, p0, p1, d0, d1, main_stream);
(void) src0;
(void) src0_dd;
}
inline void ggml_cuda_op_diag_mask_inf(
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
@@ -7549,6 +7633,9 @@ static void ggml_cuda_cpy(const ggml_tensor * src0, const ggml_tensor * src1, gg
} else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
ggml_cpy_f32_f16_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02,
ne10, ne11, nb10, nb11, nb12, main_stream);
} else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
ggml_cpy_f16_f16_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02,
ne10, ne11, nb10, nb11, nb12, main_stream);
} else {
fprintf(stderr, "%s: unsupported type combination (%s to %s)\n", __func__,
ggml_type_name(src0->type), ggml_type_name(src1->type));
@@ -7580,6 +7667,10 @@ static void ggml_cuda_alibi(const ggml_tensor * src0, const ggml_tensor * src1,
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_alibi);
}
void ggml_cuda_im2col(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_im2col);
}
static void ggml_cuda_nop(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
(void) src0;
(void) src1;
@@ -7943,6 +8034,9 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_
case GGML_OP_ALIBI:
func = ggml_cuda_alibi;
break;
case GGML_OP_IM2COL:
func = ggml_cuda_im2col;
break;
default:
return false;
}