CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-21 12:16:57 +00:00
Code Issues Packages Projects Releases Wiki Activity

opencl: add mul_mat_f32_f32_l4_lm and mul_mat_f16_f32_l4_lm (#14809)

Browse Source
This commit is contained in:
lhez
2025-07-30 14:56:55 -07:00
committed by GitHub
parent e9192bec56
commit 6e6725459a
4 changed files with 399 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

144
ggml/src/ggml-opencl/ggml-opencl.cpp
View File

@@ -33,6 +33,7 @@
#undef MAX
#define MIN(a, b) ((a) < (b) ? (a) : (b))
#define MAX(a, b) ((a) > (b) ? (a) : (b))
#define CEIL_DIV(M, N) (((M) + (N)-1) / (N))
#define UNUSED(x) (void)(x)
@@ -396,6 +397,8 @@ struct ggml_backend_opencl_context {
cl_program program_conv_2d_f16_f32;
cl_program program_tsembd;
cl_program program_mul_mv_id_q4_0_f32_8x_flat;
cl_program program_mul_mm_f32_f32_l4_lm;
cl_program program_mul_mm_f16_f32_l4_lm;
cl_kernel kernel_add, kernel_add_row;
cl_kernel kernel_mul, kernel_mul_row;
@@ -450,6 +453,8 @@ struct ggml_backend_opencl_context {
cl_kernel kernel_conv_2d_f16_f32;
cl_kernel kernel_timestep_embedding;
cl_kernel kernel_mul_mv_id_q4_0_f32_8x_flat;
cl_kernel kernel_mul_mm_f32_f32_l4_lm;
cl_kernel kernel_mul_mm_f16_f32_l4_lm;
std::vector<ProfilingInfo> profiling_info;
@@ -1040,6 +1045,38 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
GGML_LOG_CONT(".");
}
// mul_mm_f32_f32_l4_lm
{
#ifdef GGML_OPENCL_EMBED_KERNELS
const std::string kernel_src {
#include "mul_mm_f32_f32_l4_lm.cl.h"
};
#else
const std::string kernel_src = read_file("mul_mm_f32_f32_l4_lm.cl");
#endif
backend_ctx->program_mul_mm_f32_f32_l4_lm =
build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
CL_CHECK((backend_ctx->kernel_mul_mm_f32_f32_l4_lm = clCreateKernel(backend_ctx->program_mul_mm_f32_f32_l4_lm, "kernel_mul_mm_f32_f32_l4_lm", &err), err));
GGML_LOG_CONT(".");
}
// mul_mm_f16_f32_l4_lm
{
#ifdef GGML_OPENCL_EMBED_KERNELS
const std::string kernel_src {
#include "mul_mm_f16_f32_l4_lm.cl.h"
};
#else
const std::string kernel_src = read_file("mul_mm_f16_f32_l4_lm.cl");
#endif
backend_ctx->program_mul_mm_f16_f32_l4_lm =
build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
CL_CHECK((backend_ctx->kernel_mul_mm_f16_f32_l4_lm = clCreateKernel(backend_ctx->program_mul_mm_f16_f32_l4_lm, "kernel_mul_mm_f16_f32_l4_lm", &err), err));
GGML_LOG_CONT(".");
}
// mul
{
#ifdef GGML_OPENCL_EMBED_KERNELS
@@ -5297,18 +5334,6 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
if (src0t == GGML_TYPE_F16 && src1t == GGML_TYPE_F32 &&
src0->ne[1] > 32 && // M > 32
src1->ne[1] > 32 && // N > 32
src0->ne[0] > 32 && // K > 32
src0->ne[2] == 1 && src0->ne[3] == 1 &&
src1->ne[2] == 1 && src1->ne[3] == 1 &&
ggml_is_contiguous(src0) && ggml_is_contiguous(src1) &&
backend_ctx->kernel_mul_mat_f16_f32_tiled != NULL) {
ggml_cl_mul_mat_f16_f32_tiled(backend, src0, src1, dst);
return;
}
ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
@@ -5655,6 +5680,101 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
} // if (ne01 && ne1)
#endif // GGML_OPENCL_USE_ADRENO_KERNELS
// GEMM using local memory
// Current BK = 16, so ne00 % 16 == 0
if (ggml_is_contiguous(src0) &&
ggml_is_contiguous(src1) &&
src1t == GGML_TYPE_F32 &&
ne00 % 16 == 0 &&
ne11 > 1) {
switch(src0t) {
case GGML_TYPE_F32: {
kernel = backend_ctx->kernel_mul_mm_f32_f32_l4_lm;
nth0 = 128; // calculated as (BM*BN)/(TM*TN)
int batch_stride_a = ne00*ne01;
int batch_stride_b = ne10*ne11;
int batch_stride_d = ne0*ne1;
CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra0->data_device));
CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra1->data_device));
CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset1));
CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &extrad->data_device));
CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int), &ne00));
CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int), &ne01));
CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int), &ne02));
CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int), &ne11));
CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int), &ne12));
CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int), &ne10)); // stride_a
CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int), &ne10)); // stride_b
CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int), &ne01)); // stride_d
CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int), &batch_stride_a));
CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int), &batch_stride_b));
CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int), &batch_stride_d));
CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int), &r2));
CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int), &r3));
// 64 is block tile size BM and BN - change here when BM and BN in the kernel are changed.
size_t global_work_size[] = {(size_t)(CEIL_DIV(ne01, 64)*nth0), (size_t)(CEIL_DIV(ne11, 64)), (size_t)ne12*ne13};
size_t local_work_size[] = {(size_t)nth0, 1, 1};
backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
return;
}
case GGML_TYPE_F16: {
kernel = backend_ctx->kernel_mul_mm_f16_f32_l4_lm;
nth0 = 128; // calculated as (BM*BN)/(TM*TN)
int batch_stride_a = ne00*ne01;
int batch_stride_b = ne10*ne11;
int batch_stride_d = ne0*ne1;
CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra0->data_device));
CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra1->data_device));
CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset1));
CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &extrad->data_device));
CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int), &ne00));
CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int), &ne01));
CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int), &ne02));
CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int), &ne11));
CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int), &ne12));
CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int), &ne10)); // stride_a
CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int), &ne10)); // stride_b
CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int), &ne01)); // stride_d
CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int), &batch_stride_a));
CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int), &batch_stride_b));
CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int), &batch_stride_d));
CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int), &r2));
CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int), &r3));
// 64 is block tile size BM and BN - change here when BM and BN in the kernel are changed.
size_t global_work_size[] = {(size_t)(CEIL_DIV(ne01, 64)*nth0), (size_t)(CEIL_DIV(ne11, 64)), (size_t)ne12*ne13};
size_t local_work_size[] = {(size_t)nth0, 1, 1};
backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
return;
}
default:
break;
}
}
if (src0t == GGML_TYPE_F16 && src1t == GGML_TYPE_F32 &&
src0->ne[1] > 32 && // M > 32
src1->ne[1] > 32 && // N > 32
src0->ne[0] > 32 && // K > 32
src0->ne[2] == 1 && src0->ne[3] == 1 &&
src1->ne[2] == 1 && src1->ne[3] == 1 &&
ggml_is_contiguous(src0) && ggml_is_contiguous(src1) &&
backend_ctx->kernel_mul_mat_f16_f32_tiled != NULL) {
ggml_cl_mul_mat_f16_f32_tiled(backend, src0, src1, dst);
return;
}
if (!ggml_is_transposed(src0) &&
!ggml_is_transposed(src1) &&
src1t == GGML_TYPE_F32 &&