Skip to content

Commit 81542bb

Browse files
shawngu-quicshawngu-quic
committed
transposed gemm/gemv moe kernel with mxfp4,f32
1 parent 28d3073 commit 81542bb

File tree

5 files changed

+504
-4
lines changed

5 files changed

+504
-4
lines changed

‎ggml/src/ggml-opencl/CMakeLists.txt‎

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -91,6 +91,8 @@ set(GGML_OPENCL_KERNELS
9191
mul_mv_id_q8_0_f32_flat
9292
mul_mv_id_mxfp4_f32
9393
mul_mv_id_mxfp4_f32_flat
94+
gemm_moe_mxfp4_f32
95+
gemv_moe_mxfp4_f32
9496
mul_mm_f32_f32_l4_lm
9597
mul_mm_f16_f32_l4_lm
9698
mul_mm_q8_0_f32_l4_lm

‎ggml/src/ggml-opencl/ggml-opencl.cpp‎

Lines changed: 167 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -402,6 +402,7 @@ struct ggml_backend_opencl_context {
402402
cl_program program_conv_2d_f32;
403403
cl_program program_conv_2d_f16_f32;
404404
cl_program program_tsembd;
405+
cl_program program_gemv_moe_mxfp4_f32, program_gemm_moe_mxfp4_f32;
405406
cl_program program_mul_mv_id_q4_0_f32_8x_flat;
406407
cl_program program_mul_mv_id_q8_0_f32, program_mul_mv_id_q8_0_f32_flat;
407408
cl_program program_mul_mv_id_mxfp4_f32;
@@ -452,7 +453,7 @@ struct ggml_backend_opencl_context {
452453
cl_kernel kernel_mul_mat_f16_f32_tiled;
453454
cl_kernel kernel_mul_mat_q4_0_f32, kernel_mul_mat_q4_0_f32_v;
454455
cl_kernel kernel_convert_block_q4_0, kernel_restore_block_q4_0;
455-
cl_kernel kernel_convert_block_mxfp4, kernel_restore_block_mxfp4;
456+
cl_kernel kernel_convert_block_mxfp4, kernel_convert_block_mxfp4_trans, kernel_restore_block_mxfp4;
456457
cl_kernel kernel_convert_block_q8_0, kernel_restore_block_q8_0;
457458
cl_kernel kernel_mul_mat_q4_0_f32_8x_flat;
458459
cl_kernel kernel_convert_block_q4_0_noshuffle;
@@ -475,6 +476,7 @@ struct ggml_backend_opencl_context {
475476
cl_kernel kernel_conv_2d_f32;
476477
cl_kernel kernel_conv_2d_f16_f32;
477478
cl_kernel kernel_timestep_embedding;
479+
cl_kernel kernel_gemv_moe_mxfp4_f32, kernel_gemm_moe_mxfp4_f32;
478480
cl_kernel kernel_mul_mv_id_q4_0_f32_8x_flat;
479481
cl_kernel kernel_mul_mv_id_q8_0_f32, kernel_mul_mv_id_q8_0_f32_flat;
480482
cl_kernel kernel_mul_mv_id_mxfp4_f32;
@@ -777,6 +779,7 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
777779
CL_CHECK((backend_ctx->kernel_convert_block_q4_0 = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q4_0", &err), err));
778780
CL_CHECK((backend_ctx->kernel_restore_block_q4_0 = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q4_0", &err), err));
779781
CL_CHECK((backend_ctx->kernel_convert_block_mxfp4 = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_mxfp4", &err), err));
782+
CL_CHECK((backend_ctx->kernel_convert_block_mxfp4_trans = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_mxfp4_trans", &err), err));
780783
CL_CHECK((backend_ctx->kernel_restore_block_mxfp4 = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_mxfp4", &err), err));
781784
CL_CHECK((backend_ctx->kernel_convert_block_q8_0 = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q8_0", &err), err));
782785
CL_CHECK((backend_ctx->kernel_restore_block_q8_0 = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q8_0", &err), err));
@@ -1991,6 +1994,43 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
19911994
CL_CHECK((backend_ctx->CL_mul_mat_Ab_Bi_8x4 = clCreateKernel(backend_ctx->program_CL_gemm, "kernel_mul_mat_Ab_Bi_8x4", &err), err));
19921995
GGML_LOG_CONT(".");
19931996
}
1997+
1998+
std::string CL_moe_compile_opts = std::string("-cl-std=") + opencl_c_std +
1999+
" -cl-mad-enable "
2000+
" -qcom-disable-promote-pointer-to-texture"
2001+
" -cl-fast-relaxed-math";
2002+
2003+
// gemv_moe_mxfp4_f32
2004+
{
2005+
#ifdef GGML_OPENCL_EMBED_KERNELS
2006+
const std::string kernel_src {
2007+
#include "gemv_moe_mxfp4_f32.cl.h"
2008+
};
2009+
#else
2010+
const std::string kernel_src = read_file("gemv_moe_mxfp4_f32.cl");
2011+
#endif
2012+
backend_ctx->program_gemv_moe_mxfp4_f32 =
2013+
build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), CL_moe_compile_opts);
2014+
2015+
CL_CHECK((backend_ctx->kernel_gemv_moe_mxfp4_f32 = clCreateKernel(backend_ctx->program_gemv_moe_mxfp4_f32, "kernel_gemv_moe_mxfp4_f32", &err), err));
2016+
GGML_LOG_CONT(".");
2017+
}
2018+
2019+
// gemm_moe_mxfp4_f32
2020+
{
2021+
#ifdef GGML_OPENCL_EMBED_KERNELS
2022+
const std::string kernel_src {
2023+
#include "gemm_moe_mxfp4_f32.cl.h"
2024+
};
2025+
#else
2026+
const std::string kernel_src = read_file("gemm_moe_mxfp4_f32.cl");
2027+
#endif
2028+
backend_ctx->program_gemm_moe_mxfp4_f32 =
2029+
build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), CL_moe_compile_opts);
2030+
2031+
CL_CHECK((backend_ctx->kernel_gemm_moe_mxfp4_f32 = clCreateKernel(backend_ctx->program_gemm_moe_mxfp4_f32, "kernel_gemm_moe_mxfp4_f32", &err), err));
2032+
GGML_LOG_CONT(".");
2033+
}
19942034
#endif // GGML_OPENCL_USE_ADRENO_KERNELS
19952035
GGML_LOG_CONT("\n");
19962036
}
@@ -3596,14 +3636,40 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
35963636
CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
35973637
CL_CHECK(err);
35983638

3639+
#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
3640+
if (strstr(tensor->name, "ffn") != NULL) {
3641+
int ne00 = tensor->ne[0];
3642+
int ne01 = tensor->ne[1];
3643+
int ne02 = tensor->ne[2];
3644+
3645+
cl_kernel kernel = backend_ctx->kernel_convert_block_mxfp4_trans;
3646+
3647+
CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &data_device));
3648+
CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->q));
3649+
CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->e));
3650+
CL_CHECK(clSetKernelArg(kernel, 3, sizeof(int), &ne00));
3651+
CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int), &ne01));
3652+
3653+
size_t global_work_size[3] = {static_cast<size_t>(((ne01 + 63) / 64) * 64), static_cast<size_t>(ne00 / 32), static_cast<size_t>(ne02)};
3654+
size_t local_work_size[3] = {64, 2, 1};
3655+
3656+
cl_event evt;
3657+
CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
3658+
CL_CHECK(clWaitForEvents(1, &evt));
3659+
CL_CHECK(clReleaseMemObject(data_device));
3660+
tensor->extra = extra;
3661+
3662+
return;
3663+
}
3664+
#endif
35993665
cl_kernel kernel = backend_ctx->kernel_convert_block_mxfp4;
36003666

36013667
CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &data_device));
36023668
CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->q));
36033669
CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->e));
36043670

3605-
size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
3606-
size_t local_work_size[] = {64, 1, 1};
3671+
size_t global_work_size[3] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
3672+
size_t local_work_size[3] = {64, 1, 1};
36073673

36083674
cl_event evt;
36093675
CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
@@ -3619,7 +3685,6 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
36193685
{ extra->q }
36203686
};
36213687
extra->q_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_format_q, &img_desc_q, NULL, &err);
3622-
36233688
tensor->extra = extra;
36243689

36253690
return;
@@ -7545,6 +7610,7 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
75457610
const int ne21 = src2->ne[1];
75467611

75477612
const cl_ulong nb21 = src2->nb[1];
7613+
const cl_ulong nb20 = src2->nb[0];
75487614

75497615
const int ne0 = dst->ne[0];
75507616
const int ne1 = dst->ne[1];
@@ -7684,6 +7750,103 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
76847750
break;
76857751
}
76867752
case GGML_TYPE_MXFP4: {
7753+
#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
7754+
if (true) { // condition todo
7755+
cl_int status;
7756+
7757+
size_t local_size[3] = {64, 4, 1};
7758+
size_t global_size[3] = {64, 4, 1};
7759+
7760+
cl_mem src1_sub_buffer, buf_src1_image, buf_src2;
7761+
7762+
if (ne12 == 1) { // for gemv
7763+
kernel = backend_ctx->kernel_gemv_moe_mxfp4_f32;
7764+
7765+
// create a sub_buffer for src2
7766+
cl_buffer_region region;
7767+
region.origin = offset2;
7768+
region.size = ne20 * ne21 * sizeof(int);
7769+
buf_src2 = clCreateSubBuffer(extra2->data_device, 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &status);
7770+
CL_CHECK(status);
7771+
7772+
// set thread grid
7773+
global_size[0] = static_cast<size_t>(ne01);
7774+
global_size[2] = static_cast<size_t>(ne20);
7775+
} else { // for gemm
7776+
kernel = backend_ctx->kernel_gemm_moe_mxfp4_f32;
7777+
7778+
// preprocess router table
7779+
int tile_size = 320;
7780+
int num_tiles_per_expert = (ne01 + tile_size - 1) / tile_size;
7781+
void * host_src2_reorder = malloc(ne20 * ne21 * 4 * num_tiles_per_expert * sizeof(short));
7782+
void * host_src2 = malloc(ne21 * nb21);
7783+
CL_CHECK(clEnqueueReadBuffer(backend_ctx->queue, extra2->data_device, CL_TRUE, offset2, ne21 * nb21, host_src2, 0, NULL, NULL));
7784+
int total_experts = nb21 / nb20;
7785+
int out_idx = 0;
7786+
for (int i_expert = 0; i_expert < ne02; i_expert++) {
7787+
for (int i_tile = 0; i_tile < num_tiles_per_expert; i_tile++) {
7788+
for (int j = 0; j < ne21; j++) {
7789+
for (int i = 0; i < ne20; i++) {
7790+
int expert = ((int *)host_src2)[j * total_experts + i];
7791+
if (i_expert == expert) {
7792+
((short *)host_src2_reorder)[out_idx] = static_cast<short>(expert);
7793+
((short *)host_src2_reorder)[out_idx + 1] = static_cast<short>(j * ne11 + (i % ne11));
7794+
((short *)host_src2_reorder)[out_idx + 2] = static_cast<short>(j * ne20 + i);
7795+
((short *)host_src2_reorder)[out_idx + 3] = static_cast<short>(i_tile);
7796+
out_idx += 4;
7797+
}
7798+
}
7799+
}
7800+
}
7801+
}
7802+
buf_src2 = clCreateBuffer(backend_ctx->context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, ne20 * ne21 * 4 * num_tiles_per_expert * sizeof(short), host_src2_reorder, &status);
7803+
CL_CHECK(status);
7804+
7805+
// set thread grid
7806+
global_size[0] = static_cast<size_t>(tile_size);
7807+
global_size[2] = static_cast<size_t>(ne20 * ne21 * num_tiles_per_expert);
7808+
}
7809+
7810+
// create a sub_buffer for src1
7811+
cl_buffer_region region;
7812+
region.origin = offset1;
7813+
region.size = ne10 * ne11 * ne12 * sizeof(float);
7814+
src1_sub_buffer = clCreateSubBuffer(extra1->data_device, 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &status);
7815+
CL_CHECK(status);
7816+
7817+
// create image for src1
7818+
cl_image_format image_format_buf_src1 = {CL_RGBA, CL_FLOAT};
7819+
cl_image_desc image_desc_buf_src1 = {CL_MEM_OBJECT_IMAGE1D_BUFFER, static_cast<size_t>(ne10 * ne11 * ne12 / 4), 0,0,0,0,0,0,0, src1_sub_buffer};
7820+
buf_src1_image = clCreateImage(backend_ctx->context, CL_MEM_READ_ONLY, &image_format_buf_src1, &image_desc_buf_src1, NULL, &status);
7821+
CL_CHECK(status);
7822+
7823+
// Set kernel args
7824+
int arg_idx = 0;
7825+
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_mem), &extra0_mxfp4->q));
7826+
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_mem), &extra0_mxfp4->e));
7827+
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_mem), &buf_src1_image));
7828+
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_mem), &buf_src2));
7829+
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_mem), &extrad->data_device));
7830+
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_ulong), &offsetd));
7831+
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int), &ne00));
7832+
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int), &ne01));
7833+
if (ne12 == 1) {
7834+
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int), &ne11));
7835+
} else {
7836+
CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int), &ne02));
7837+
}
7838+
7839+
// launch kernel
7840+
backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_size, local_size, dst);
7841+
7842+
// deallocate sub buffers and images
7843+
CL_CHECK(clReleaseMemObject(src1_sub_buffer));
7844+
CL_CHECK(clReleaseMemObject(buf_src1_image));
7845+
CL_CHECK(clReleaseMemObject(buf_src2));
7846+
return;
7847+
} // else fallback to generic kernel
7848+
#endif // GGML_OPENCL_USE_ADRENO_KERNELS
7849+
76877850
#ifdef GGML_OPENCL_SOA_Q
76887851
kernel = backend_ctx->kernel_mul_mv_id_mxfp4_f32_flat;
76897852

‎ggml/src/ggml-opencl/kernels/cvt.cl‎

Lines changed: 21 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -147,6 +147,27 @@ kernel void kernel_convert_block_mxfp4(
147147
}
148148
}
149149

150+
kernel void kernel_convert_block_mxfp4_trans(
151+
global struct block_mxfp4 * src0,
152+
__global uint4 * dst_q,
153+
__global uchar * dst_e,
154+
uint ne00,
155+
uint ne01
156+
) {
157+
int i00 = get_global_id(1);
158+
uint i01 = get_global_id(0);
159+
uint i02 = get_global_id(2);
160+
161+
uint ne00_blk = ne00 / QK_MXFP4;
162+
uint src_blk_offset = i00 + i01 * ne00_blk + i02 * ne00_blk * ne01;
163+
uint dst_blk_offset = i01 + i00 * ne01 + i02 * ne00_blk * ne01;
164+
165+
global struct block_mxfp4 * b = src0 + src_blk_offset;
166+
167+
dst_q[dst_blk_offset] = ((global uint4 *)(&(b->qs[0])))[0];
168+
dst_e[dst_blk_offset] = b->e;
169+
}
170+
150171
kernel void kernel_restore_block_mxfp4(
151172
global uchar * src_q,
152173
global half * src_e,

0 commit comments

Comments
 (0)