Skip to content

Commit 0b88557

Browse files
lhezlhez
authored
opencl: add set_rows for f16 and f32 (#14547)
* opencl: add `set_rows` for `f16` and `f32` * opencl: better choose workgroup size for `set_rows`
1 parent 4bb625b commit 0b88557

File tree

3 files changed

+247
-2
lines changed

3 files changed

+247
-2
lines changed

‎ggml/src/ggml-opencl/CMakeLists.txt

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -88,6 +88,7 @@ set(GGML_OPENCL_KERNELS
8888
rms_norm
8989
rope
9090
scale
91+
set_rows
9192
sigmoid
9293
silu
9394
softmax_4_f32

‎ggml/src/ggml-opencl/ggml-opencl.cpp

Lines changed: 151 additions & 2 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_set_rows;
354355
cl_program program_glu;
355356
cl_program program_im2col_f16;
356357
cl_program program_im2col_f32;
@@ -412,6 +413,7 @@ struct ggml_backend_opencl_context {
412413
cl_kernel kernel_soft_max, kernel_soft_max_4;
413414
cl_kernel kernel_soft_max_f16, kernel_soft_max_4_f16;
414415
cl_kernel kernel_get_rows_f32, kernel_get_rows_f16, kernel_get_rows_q4_0;
416+
cl_kernel kernel_set_rows_f32, kernel_set_rows_f16;
415417
cl_kernel kernel_rope_norm_f32, kernel_rope_norm_f16, kernel_rope_neox_f32, kernel_rope_neox_f16;
416418
cl_kernel kernel_rope_multi_f32, kernel_rope_multi_f16, kernel_rope_vision_f32, kernel_rope_vision_f16;
417419
cl_kernel kernel_cpy_f16_f16, kernel_cpy_f16_f32, kernel_cpy_f32_f16, kernel_cpy_f32_f32;
@@ -529,6 +531,16 @@ struct ggml_backend_opencl_context {
529531
fclose(ftrace);
530532
}
531533

534+
size_t get_kernel_workgroup_size(cl_kernel kernel) const {
535+
size_t workgroup_size = 0;
536+
size_t ret_size = 0;
537+
CL_CHECK(
538+
clGetKernelWorkGroupInfo(kernel, device, CL_KERNEL_WORK_GROUP_SIZE,
539+
sizeof(size_t), &workgroup_size, &ret_size));
540+
GGML_ASSERT(sizeof(size_t) == ret_size);
541+
return workgroup_size;
542+
}
543+
532544
void enqueue_ndrange_kernel(cl_kernel kernel, cl_uint work_dim, size_t *global_work_size, size_t *local_work_size, const ggml_tensor * tensor) {
533545
#ifdef GGML_OPENCL_PROFILING
534546
cl_event evt;
@@ -1431,6 +1443,23 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
14311443
}
14321444
}
14331445

1446+
// set_rows
1447+
{
1448+
#ifdef GGML_OPENCL_EMBED_KERNELS
1449+
const std::string kernel_src {
1450+
#include "set_rows.cl.h"
1451+
};
1452+
#else
1453+
const std::string kernel_src = read_file("set_rows.cl");
1454+
#endif
1455+
backend_ctx->program_set_rows =
1456+
build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
1457+
1458+
CL_CHECK((backend_ctx->kernel_set_rows_f32 = clCreateKernel(backend_ctx->program_set_rows, "kernel_set_rows_f32", &err), err));
1459+
CL_CHECK((backend_ctx->kernel_set_rows_f16 = clCreateKernel(backend_ctx->program_set_rows, "kernel_set_rows_f16", &err), err));
1460+
GGML_LOG_CONT(".");
1461+
}
1462+
14341463
// mul_mv_id_q4_0_f32_8x_flat
14351464
{
14361465
#ifdef GGML_OPENCL_EMBED_KERNELS
@@ -2233,8 +2262,17 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
22332262
{
22342263
// TODO: add support
22352264
// ref: https://github.com/ggml-org/llama.cpp/pull/14274
2236-
return false;
2237-
} break;
2265+
if (op->src[0]->type != GGML_TYPE_F32) {
2266+
return false;
2267+
}
2268+
switch (op->type) {
2269+
case GGML_TYPE_F16:
2270+
case GGML_TYPE_F32:
2271+
return true;
2272+
default:
2273+
return false;
2274+
}
2275+
}
22382276
case GGML_OP_CPY:
22392277
case GGML_OP_DUP:
22402278
case GGML_OP_CONT:
@@ -3374,6 +3412,111 @@ static void ggml_cl_get_rows(ggml_backend_t backend, const ggml_tensor * src0, c
33743412
backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
33753413
}
33763414

3415+
static void ggml_cl_set_rows(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
3416+
GGML_ASSERT(src0);
3417+
GGML_ASSERT(src0->extra);
3418+
GGML_ASSERT(src1);
3419+
GGML_ASSERT(src1->extra);
3420+
GGML_ASSERT(dst);
3421+
GGML_ASSERT(dst->extra);
3422+
3423+
// ne0 = ne00
3424+
// ne2 = ne02
3425+
// ne3 = ne03
3426+
3427+
const int ne01 = src0->ne[1];
3428+
const int ne02 = src0->ne[2];
3429+
const int ne03 = src0->ne[3];
3430+
3431+
const cl_ulong nb01 = src0->nb[1];
3432+
const cl_ulong nb02 = src0->nb[2];
3433+
const cl_ulong nb03 = src0->nb[3];
3434+
3435+
const int ne11 = src1->ne[1];
3436+
const int ne12 = src1->ne[2];
3437+
3438+
const cl_ulong nb10 = src1->nb[0];
3439+
const cl_ulong nb11 = src1->nb[1];
3440+
const cl_ulong nb12 = src1->nb[2];
3441+
3442+
const int ne0 = dst->ne[0];
3443+
3444+
const cl_ulong nb1 = dst->nb[1];
3445+
const cl_ulong nb2 = dst->nb[2];
3446+
const cl_ulong nb3 = dst->nb[3];
3447+
3448+
const int nblk0 = ne0/ggml_blck_size(dst->type);
3449+
3450+
ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
3451+
3452+
ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
3453+
ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
3454+
ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
3455+
3456+
cl_ulong offset0 = extra0->offset + src0->view_offs;
3457+
cl_ulong offset1 = extra1->offset + src1->view_offs;
3458+
cl_ulong offsetd = extrad->offset + dst->view_offs;
3459+
3460+
cl_kernel kernel;
3461+
3462+
switch (dst->type) {
3463+
case GGML_TYPE_F32:
3464+
kernel = backend_ctx->kernel_set_rows_f32;
3465+
break;
3466+
case GGML_TYPE_F16:
3467+
kernel = backend_ctx->kernel_set_rows_f16;
3468+
break;
3469+
default:
3470+
GGML_ABORT("not implemented");
3471+
}
3472+
3473+
CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra0->data_device));
3474+
CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
3475+
CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra1->data_device));
3476+
CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset1));
3477+
CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &extrad->data_device));
3478+
CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
3479+
CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int), &ne01));
3480+
CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_ulong), &nb01));
3481+
CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_ulong), &nb02));
3482+
CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_ulong), &nb03));
3483+
CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int), &ne11));
3484+
CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int), &ne12));
3485+
CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb10));
3486+
CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb11));
3487+
CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb12));
3488+
CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int), &nblk0));
3489+
CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb1));
3490+
CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &nb2));
3491+
CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb3));
3492+
3493+
int nth0 = 64;
3494+
if (backend_ctx->gpu_family == INTEL) {
3495+
nth0 = 32;
3496+
} else if (backend_ctx->gpu_family == ADRENO) {
3497+
nth0 = 64;
3498+
}
3499+
3500+
int max_workgroup_size = backend_ctx->get_kernel_workgroup_size(kernel);
3501+
while (nth0 < nblk0 && nth0 < max_workgroup_size) {
3502+
nth0 *= 2;
3503+
}
3504+
3505+
int rows_per_workgroup = 1;
3506+
if (nth0 > nblk0) {
3507+
rows_per_workgroup = nth0 / nblk0;
3508+
nth0 = nblk0;
3509+
}
3510+
3511+
size_t global_work_size[] = {
3512+
(size_t)(ne01 + rows_per_workgroup - 1)/rows_per_workgroup*nth0,
3513+
(size_t)ne02*rows_per_workgroup,
3514+
(size_t)ne03};
3515+
size_t local_work_size[] = {(size_t)nth0, (size_t)rows_per_workgroup, 1};
3516+
3517+
backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
3518+
}
3519+
33773520
static void ggml_cl_add(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
33783521
GGML_ASSERT(src0);
33793522
GGML_ASSERT(src0->extra);
@@ -6388,6 +6531,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
63886531
}
63896532
func = ggml_cl_get_rows;
63906533
break;
6534+
case GGML_OP_SET_ROWS:
6535+
if (!any_on_device) {
6536+
return false;
6537+
}
6538+
func = ggml_cl_set_rows;
6539+
break;
63916540
case GGML_OP_CPY:
63926541
if (!any_on_device) {
63936542
return false;

‎ggml/src/ggml-opencl/kernels/set_rows.cl

Lines changed: 95 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,95 @@
1+
#pragma OPENCL EXTENSION cl_khr_fp16 : enable
2+
3+
kernel void kernel_set_rows_f32(
4+
global char * src0,
5+
ulong offset0,
6+
global char * src1,
7+
ulong offset1,
8+
global char * dst,
9+
ulong offsetd,
10+
int ne01,
11+
ulong nb01,
12+
ulong nb02,
13+
ulong nb03,
14+
int ne11,
15+
int ne12,
16+
ulong nb10,
17+
ulong nb11,
18+
ulong nb12,
19+
int nblk0,
20+
ulong nb1,
21+
ulong nb2,
22+
ulong nb3
23+
) {
24+
src0 = src0 + offset0;
25+
src1 = src1 + offset1;
26+
dst = dst + offsetd;
27+
28+
int i03 = get_group_id(2);
29+
int i02 = get_group_id(1);
30+
int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1);
31+
32+
if (i01 >= ne01) {
33+
return;
34+
}
35+
36+
int i12 = i03%ne12;
37+
int i11 = i02%ne11;
38+
39+
int i10 = i01;
40+
long i1 = ((global long *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
41+
42+
global float * dst_row = (global float *) (dst + i1*nb1 + i02*nb2 + i03*nb3);
43+
global float * src_row = (global float *) (src0 + i01*nb01 + i02*nb02 + i03*nb03);
44+
45+
for (int ind = get_local_id(0); ind < nblk0; ind += get_local_size(0)) {
46+
dst_row[ind] = (float)src_row[ind];
47+
}
48+
}
49+
50+
kernel void kernel_set_rows_f16(
51+
global char * src0,
52+
ulong offset0,
53+
global char * src1,
54+
ulong offset1,
55+
global char * dst,
56+
ulong offsetd,
57+
int ne01,
58+
ulong nb01,
59+
ulong nb02,
60+
ulong nb03,
61+
int ne11,
62+
int ne12,
63+
ulong nb10,
64+
ulong nb11,
65+
ulong nb12,
66+
int nblk0,
67+
ulong nb1,
68+
ulong nb2,
69+
ulong nb3
70+
) {
71+
src0 = src0 + offset0;
72+
src1 = src1 + offset1;
73+
dst = dst + offsetd;
74+
75+
int i03 = get_group_id(2);
76+
int i02 = get_group_id(1);
77+
int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1);
78+
79+
if (i01 >= ne01) {
80+
return;
81+
}
82+
83+
int i12 = i03%ne12;
84+
int i11 = i02%ne11;
85+
86+
int i10 = i01;
87+
long i1 = ((global long *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
88+
89+
global half * dst_row = (global half *) (dst + i1*nb1 + i02*nb2 + i03*nb3);
90+
global float * src_row = (global float *) (src0 + i01*nb01 + i02*nb02 + i03*nb03);
91+
92+
for (int ind = get_local_id(0); ind < nblk0; ind += get_local_size(0)) {
93+
dst_row[ind] = src_row[ind];
94+
}
95+
}

0 commit comments

Comments
 (0)