Skip to content

Commit c59ea5e

Browse files
lhezqnixsynapse
lhez
authored and
qnixsynapse
committed
opencl : add GEGLU, REGLU, SWIGLU (ggml-org#14456)
1 parent 27b1879 commit c59ea5e

File tree

1 file changed

+124
-0
lines changed

1 file changed

+124
-0
lines changed

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

Lines changed: 124 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -351,6 +351,7 @@ struct ggml_backend_opencl_context {
351351
cl_program program_gemv_noshuffle_general;
352352
cl_program program_gemv_noshuffle;
353353
cl_program program_get_rows;
354+
cl_program program_glu;
354355
cl_program program_im2col_f16;
355356
cl_program program_im2col_f32;
356357
cl_program program_mul_mat_Ab_Bi_8x4;
@@ -401,6 +402,8 @@ struct ggml_backend_opencl_context {
401402
cl_kernel kernel_relu;
402403
cl_kernel kernel_sigmoid_f32, kernel_sigmoid_f16;
403404
cl_kernel kernel_clamp;
405+
cl_kernel kernel_geglu, kernel_reglu, kernel_swiglu,
406+
kernel_geglu_f16, kernel_reglu_f16, kernel_swiglu_f16;
404407
cl_kernel kernel_norm;
405408
cl_kernel kernel_rms_norm;
406409
cl_kernel kernel_group_norm;
@@ -738,6 +741,27 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
738741
GGML_LOG_CONT(".");
739742
}
740743

744+
// glu
745+
{
746+
#ifdef GGML_OPENCL_EMBED_KERNELS
747+
const std::string kernel_src {
748+
#include "glu.cl.h"
749+
};
750+
#else
751+
const std::string kernel_src = read_file("glu.cl");
752+
#endif
753+
backend_ctx->program_glu =
754+
build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
755+
756+
CL_CHECK((backend_ctx->kernel_geglu = clCreateKernel(backend_ctx->program_glu, "kernel_geglu", &err), err));
757+
CL_CHECK((backend_ctx->kernel_reglu = clCreateKernel(backend_ctx->program_glu, "kernel_reglu", &err), err));
758+
CL_CHECK((backend_ctx->kernel_swiglu = clCreateKernel(backend_ctx->program_glu, "kernel_swiglu", &err), err));
759+
CL_CHECK((backend_ctx->kernel_geglu_f16 = clCreateKernel(backend_ctx->program_glu, "kernel_geglu_f16", &err), err));
760+
CL_CHECK((backend_ctx->kernel_reglu_f16 = clCreateKernel(backend_ctx->program_glu, "kernel_reglu_f16", &err), err));
761+
CL_CHECK((backend_ctx->kernel_swiglu_f16 = clCreateKernel(backend_ctx->program_glu, "kernel_swiglu_f16", &err), err));
762+
GGML_LOG_CONT(".");
763+
}
764+
741765
// get_rows
742766
{
743767
#ifdef GGML_OPENCL_EMBED_KERNELS
@@ -2242,6 +2266,15 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
22422266
default:
22432267
return false;
22442268
}
2269+
case GGML_OP_GLU:
2270+
switch (ggml_get_glu_op(op)) {
2271+
case GGML_GLU_OP_GEGLU:
2272+
case GGML_GLU_OP_REGLU:
2273+
case GGML_GLU_OP_SWIGLU:
2274+
return ggml_is_contiguous_1(op->src[0]) && (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16);
2275+
default:
2276+
return false;
2277+
}
22452278
case GGML_OP_CLAMP:
22462279
return op->src[0]->type == GGML_TYPE_F32;
22472280
case GGML_OP_SOFT_MAX:
@@ -6143,6 +6176,91 @@ static void ggml_cl_sum_rows(ggml_backend_t backend, const ggml_tensor * src0, c
61436176
backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
61446177
}
61456178

6179+
static void ggml_cl_glu(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
6180+
GGML_ASSERT(src0);
6181+
GGML_ASSERT(src0->extra);
6182+
GGML_ASSERT(dst);
6183+
GGML_ASSERT(dst->extra);
6184+
6185+
GGML_ASSERT(ggml_is_contiguous_1(src0));
6186+
6187+
if (src1) {
6188+
GGML_ASSERT(src1);
6189+
GGML_ASSERT(src1->extra);
6190+
GGML_ASSERT(ggml_are_same_shape(src0, src1));
6191+
}
6192+
6193+
ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
6194+
6195+
cl_kernel kernel;
6196+
switch (ggml_get_glu_op(dst)) {
6197+
case GGML_GLU_OP_GEGLU:
6198+
if (dst->type == GGML_TYPE_F32) {
6199+
kernel = backend_ctx->kernel_geglu;
6200+
} else {
6201+
kernel = backend_ctx->kernel_geglu_f16;
6202+
}
6203+
break;
6204+
case GGML_GLU_OP_REGLU:
6205+
if (dst->type == GGML_TYPE_F32) {
6206+
kernel = backend_ctx->kernel_reglu;
6207+
} else {
6208+
kernel = backend_ctx->kernel_reglu_f16;
6209+
}
6210+
break;
6211+
case GGML_GLU_OP_SWIGLU:
6212+
if (dst->type == GGML_TYPE_F32) {
6213+
kernel = backend_ctx->kernel_swiglu;
6214+
} else {
6215+
kernel = backend_ctx->kernel_swiglu_f16;
6216+
}
6217+
break;
6218+
default:
6219+
GGML_ABORT("Unsupported glu op");
6220+
}
6221+
6222+
ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
6223+
ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
6224+
6225+
ggml_tensor_extra_cl * extra1 = src1 ? (ggml_tensor_extra_cl *)src1->extra : nullptr;
6226+
6227+
cl_ulong offset0 = extra0->offset + src0->view_offs;
6228+
cl_ulong offsetd = extrad->offset + dst->view_offs;
6229+
6230+
cl_ulong offset1 = extra1 ? extra1->offset + src1->view_offs : offset0;
6231+
6232+
const int ne0 = dst->ne[0];
6233+
6234+
const cl_ulong nb01 = src0->nb[1];
6235+
const cl_ulong nb11 = src1 ? src1->nb[1] : nb01;
6236+
6237+
const cl_ulong nb1 = dst->nb[1];
6238+
6239+
const int swp = ((const int32_t *) dst->op_params)[1];
6240+
const int ne00_off = src1 ? 0 : (swp ? ne0 : 0);
6241+
const int ne10_off = src1 ? 0 : (swp ? 0 : ne0);
6242+
6243+
CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra0->data_device));
6244+
CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
6245+
CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), src1 ? &extra1->data_device : &extra0->data_device));
6246+
CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset1));
6247+
CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &extrad->data_device));
6248+
CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
6249+
CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_ulong), &nb01));
6250+
CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_ulong), &nb11));
6251+
CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int), &ne0));
6252+
CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_ulong), &nb1));
6253+
CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int), &ne00_off));
6254+
CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int), &ne10_off));
6255+
6256+
const size_t nrows = ggml_nrows(src0);
6257+
size_t nth = 512;
6258+
size_t global_work_size[] = {nrows*nth, 1, 1};
6259+
size_t local_work_size[] = {nth, 1, 1};
6260+
6261+
backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
6262+
}
6263+
61466264
//------------------------------------------------------------------------------
61476265
// Op offloading
61486266
//------------------------------------------------------------------------------
@@ -6244,6 +6362,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
62446362
default:
62456363
return false;
62466364
} break;
6365+
case GGML_OP_GLU:
6366+
if (!any_on_device) {
6367+
return false;
6368+
}
6369+
func = ggml_cl_glu;
6370+
break;
62476371
case GGML_OP_CLAMP:
62486372
if (!any_on_device) {
62496373
return false;

0 commit comments

Comments
 (0)