Skip to content

Commit 5d4206d

Browse files
lhezlhez
committed
opencl: add set_rows for f16 and f32 
1 parent a0374a6 commit 5d4206d

File tree

3 files changed

+209
-2
lines changed

3 files changed

+209
-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: 121 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;
@@ -1431,6 +1433,23 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
14311433
}
14321434
}
14331435

1436+
// set_rows
1437+
{
1438+
#ifdef GGML_OPENCL_EMBED_KERNELS
1439+
const std::string kernel_src {
1440+
#include "set_rows.cl.h"
1441+
};
1442+
#else
1443+
const std::string kernel_src = read_file("set_rows.cl");
1444+
#endif
1445+
backend_ctx->program_set_rows =
1446+
build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
1447+
1448+
CL_CHECK((backend_ctx->kernel_set_rows_f32 = clCreateKernel(backend_ctx->program_set_rows, "kernel_set_rows_f32", &err), err));
1449+
CL_CHECK((backend_ctx->kernel_set_rows_f16 = clCreateKernel(backend_ctx->program_set_rows, "kernel_set_rows_f16", &err), err));
1450+
GGML_LOG_CONT(".");
1451+
}
1452+
14341453
// mul_mv_id_q4_0_f32_8x_flat
14351454
{
14361455
#ifdef GGML_OPENCL_EMBED_KERNELS
@@ -2233,8 +2252,17 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
22332252
{
22342253
// TODO: add support
22352254
// ref: https://github.com/ggml-org/llama.cpp/pull/14274
2236-
return false;
2237-
} break;
2255+
if (op->src[0]->type != GGML_TYPE_F32) {
2256+
return false;
2257+
}
2258+
switch (op->type) {
2259+
case GGML_TYPE_F16:
2260+
case GGML_TYPE_F32:
2261+
return true;
2262+
default:
2263+
return false;
2264+
}
2265+
}
22382266
case GGML_OP_CPY:
22392267
case GGML_OP_DUP:
22402268
case GGML_OP_CONT:
@@ -3374,6 +3402,91 @@ static void ggml_cl_get_rows(ggml_backend_t backend, const ggml_tensor * src0, c
33743402
backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
33753403
}
33763404

3405+
static void ggml_cl_set_rows(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
3406+
GGML_ASSERT(src0);
3407+
GGML_ASSERT(src0->extra);
3408+
GGML_ASSERT(src1);
3409+
GGML_ASSERT(src1->extra);
3410+
GGML_ASSERT(dst);
3411+
GGML_ASSERT(dst->extra);
3412+
3413+
// ne0 = ne00
3414+
// ne2 = ne02
3415+
// ne3 = ne03
3416+
3417+
const int ne01 = src0->ne[1];
3418+
const int ne02 = src0->ne[2];
3419+
const int ne03 = src0->ne[3];
3420+
3421+
const cl_ulong nb01 = src0->nb[1];
3422+
const cl_ulong nb02 = src0->nb[2];
3423+
const cl_ulong nb03 = src0->nb[3];
3424+
3425+
const int ne11 = src1->ne[1];
3426+
const int ne12 = src1->ne[2];
3427+
3428+
const cl_ulong nb10 = src1->nb[0];
3429+
const cl_ulong nb11 = src1->nb[1];
3430+
const cl_ulong nb12 = src1->nb[2];
3431+
3432+
const int ne0 = dst->ne[0];
3433+
3434+
const cl_ulong nb1 = dst->nb[1];
3435+
const cl_ulong nb2 = dst->nb[2];
3436+
const cl_ulong nb3 = dst->nb[3];
3437+
3438+
const int nblk0 = ne0/ggml_blck_size(dst->type);
3439+
3440+
ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
3441+
3442+
ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
3443+
ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
3444+
ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
3445+
3446+
cl_ulong offset0 = extra0->offset + src0->view_offs;
3447+
cl_ulong offset1 = extra1->offset + src1->view_offs;
3448+
cl_ulong offsetd = extrad->offset + dst->view_offs;
3449+
3450+
cl_kernel kernel;
3451+
3452+
switch (dst->type) {
3453+
case GGML_TYPE_F32:
3454+
kernel = backend_ctx->kernel_set_rows_f32;
3455+
break;
3456+
case GGML_TYPE_F16:
3457+
kernel = backend_ctx->kernel_set_rows_f16;
3458+
break;
3459+
default:
3460+
GGML_ABORT("not implemented");
3461+
}
3462+
3463+
CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra0->data_device));
3464+
CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
3465+
CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra1->data_device));
3466+
CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset1));
3467+
CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &extrad->data_device));
3468+
CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
3469+
CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int), &ne01));
3470+
CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_ulong), &nb01));
3471+
CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_ulong), &nb02));
3472+
CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_ulong), &nb03));
3473+
CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int), &ne11));
3474+
CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int), &ne12));
3475+
CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb10));
3476+
CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb11));
3477+
CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb12));
3478+
CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int), &nblk0));
3479+
CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb1));
3480+
CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &nb2));
3481+
CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb3));
3482+
3483+
int nth0 = 256;
3484+
size_t global_work_size[] = {(size_t)ne01*nth0, (size_t)ne02, (size_t)ne03};
3485+
size_t local_work_size[] = {(size_t)nth0, 1, 1};
3486+
3487+
backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
3488+
}
3489+
33773490
static void ggml_cl_add(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
33783491
GGML_ASSERT(src0);
33793492
GGML_ASSERT(src0->extra);
@@ -6385,6 +6498,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
63856498
}
63866499
func = ggml_cl_get_rows;
63876500
break;
6501+
case GGML_OP_SET_ROWS:
6502+
if (!any_on_device) {
6503+
return false;
6504+
}
6505+
func = ggml_cl_set_rows;
6506+
break;
63886507
case GGML_OP_CPY:
63896508
if (!any_on_device) {
63906509
return false;

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

Lines changed: 87 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,87 @@
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);
31+
32+
int i12 = i03%ne12;
33+
int i11 = i02%ne11;
34+
35+
int i10 = i01;
36+
long i1 = ((global long *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
37+
38+
global float * dst_row = (global float *) (dst + i1*nb1 + i02*nb2 + i03*nb3);
39+
global float * src_row = (global float *) (src0 + i01*nb01 + i02*nb02 + i03*nb03);
40+
41+
for (int ind = get_local_id(0); ind < nblk0; ind += get_local_size(0)) {
42+
dst_row[ind] = (float)src_row[ind];
43+
}
44+
}
45+
46+
kernel void kernel_set_rows_f16(
47+
global char * src0,
48+
ulong offset0,
49+
global char * src1,
50+
ulong offset1,
51+
global char * dst,
52+
ulong offsetd,
53+
int ne01,
54+
ulong nb01,
55+
ulong nb02,
56+
ulong nb03,
57+
int ne11,
58+
int ne12,
59+
ulong nb10,
60+
ulong nb11,
61+
ulong nb12,
62+
int nblk0,
63+
ulong nb1,
64+
ulong nb2,
65+
ulong nb3
66+
) {
67+
src0 = src0 + offset0;
68+
src1 = src1 + offset1;
69+
dst = dst + offsetd;
70+
71+
int i03 = get_group_id(2);
72+
int i02 = get_group_id(1);
73+
int i01 = get_group_id(0);
74+
75+
int i12 = i03%ne12;
76+
int i11 = i02%ne11;
77+
78+
int i10 = i01;
79+
long i1 = ((global long *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
80+
81+
global half * dst_row = (global half *) (dst + i1*nb1 + i02*nb2 + i03*nb3);
82+
global float * src_row = (global float *) (src0 + i01*nb01 + i02*nb02 + i03*nb03);
83+
84+
for (int ind = get_local_id(0); ind < nblk0; ind += get_local_size(0)) {
85+
dst_row[ind] = src_row[ind];
86+
}
87+
}

0 commit comments

Comments
 (0)