Skip to content

Commit 87ad642

Browse files
qnixsynapseqnixsynapse
committed
binbcast: move to a separate file
1 parent 1ea57a8 commit 87ad642

File tree

7 files changed

+329
-352
lines changed

7 files changed

+329
-352
lines changed

‎ggml/src/ggml-sycl/backend.hpp‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -29,6 +29,7 @@
2929
#include "wkv6.hpp"
3030
#include "outprod.hpp"
3131
#include "element_wise.hpp"
32+
#include "binbcast.hpp"
3233
#include "gla.hpp"
3334

3435
#endif // GGML_SYCL_BACKEND_HPP

‎ggml/src/ggml-sycl/binbcast.cpp‎

Lines changed: 311 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,311 @@
1+
#include "binbcast.hpp"
2+
#include "common.hpp"
3+
4+
static __dpct_inline__ float op_repeat(const float a, const float b) {
5+
return b;
6+
GGML_UNUSED(a);
7+
}
8+
9+
static __dpct_inline__ float op_add(const float a, const float b) {
10+
return a + b;
11+
}
12+
13+
static __dpct_inline__ float op_sub(const float a, const float b) {
14+
return a - b;
15+
}
16+
17+
static __dpct_inline__ float op_mul(const float a, const float b) {
18+
return a * b;
19+
}
20+
21+
static __dpct_inline__ float op_div(const float a, const float b) {
22+
return a / b;
23+
}
24+
25+
template <float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
26+
static void k_bin_bcast(const src0_t * src0, const src1_t * src1, dst_t * dst, int ne0, int ne1, int ne2, int ne3,
27+
int ne10, int ne11, int ne12, int ne13,
28+
/*int s0, */ int s1, int s2, int s3,
29+
/*int s10,*/ int s11, int s12, int s13, const sycl::nd_item<3> & item_ct1) {
30+
const int i0s = item_ct1.get_local_range(2) * item_ct1.get_group(2) + item_ct1.get_local_id(2);
31+
const int i1 = (item_ct1.get_local_range(1) * item_ct1.get_group(1) + item_ct1.get_local_id(1));
32+
const int i2 = (item_ct1.get_local_range(0) * item_ct1.get_group(0) + item_ct1.get_local_id(0)) / ne3;
33+
const int i3 = (item_ct1.get_local_range(0) * item_ct1.get_group(0) + item_ct1.get_local_id(0)) % ne3;
34+
35+
if (i0s >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
36+
return;
37+
}
38+
39+
const int i11 = i1 % ne11;
40+
const int i12 = i2 % ne12;
41+
const int i13 = i3 % ne13;
42+
43+
const size_t i_src0 = i3 * s3 + i2 * s2 + i1 * s1;
44+
const size_t i_src1 = i13 * s13 + i12 * s12 + i11 * s11;
45+
const size_t i_dst = i_src0;
46+
47+
const src0_t * src0_row = src0 + i_src0;
48+
const src1_t * src1_row = src1 + i_src1;
49+
dst_t * dst_row = dst + i_dst;
50+
51+
for (int i0 = i0s; i0 < ne0; i0 += item_ct1.get_local_range(2) * item_ct1.get_group_range(2)) {
52+
const int i10 = i0 % ne10;
53+
dst_row[i0] = (dst_t) bin_op(src0 ? (float) src0_row[i0] : 0.0f, (float) src1_row[i10]);
54+
}
55+
}
56+
57+
template <float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
58+
static void k_bin_bcast_unravel(const src0_t * src0, const src1_t * src1, dst_t * dst, int ne0, int ne1, int ne2,
59+
int ne3, int ne10, int ne11, int ne12, int ne13,
60+
/*int s0, */ int s1, int s2, int s3,
61+
/*int s10,*/ int s11, int s12, int s13, const sycl::nd_item<3> & item_ct1) {
62+
const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + item_ct1.get_local_id(2);
63+
64+
const int i3 = i / (ne2 * ne1 * ne0);
65+
const int i2 = (i / (ne1 * ne0)) % ne2;
66+
const int i1 = (i / ne0) % ne1;
67+
const int i0 = i % ne0;
68+
69+
if (i0 >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
70+
return;
71+
}
72+
73+
const int i11 = i1 % ne11;
74+
const int i12 = i2 % ne12;
75+
const int i13 = i3 % ne13;
76+
77+
const size_t i_src0 = i3 * s3 + i2 * s2 + i1 * s1;
78+
const size_t i_src1 = i13 * s13 + i12 * s12 + i11 * s11;
79+
const size_t i_dst = i_src0;
80+
81+
const src0_t * src0_row = src0 + i_src0;
82+
const src1_t * src1_row = src1 + i_src1;
83+
dst_t * dst_row = dst + i_dst;
84+
85+
const int i10 = i0 % ne10;
86+
dst_row[i0] = (dst_t) bin_op(src0 ? (float) src0_row[i0] : 0.0f, (float) src1_row[i10]);
87+
}
88+
89+
template <float (*bin_op)(const float, const float)> struct bin_bcast_sycl {
90+
template <typename src0_t, typename src1_t, typename dst_t>
91+
void operator()(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst,
92+
const src0_t * src0_dd, const src1_t * src1_dd, dst_t * dst_dd, queue_ptr stream) {
93+
GGML_TENSOR_BINARY_OP_LOCALS
94+
95+
int nr0 = ne10 / ne0;
96+
int nr1 = ne11 / ne1;
97+
int nr2 = ne12 / ne2;
98+
int nr3 = ne13 / ne3;
99+
100+
int nr[4] = { nr0, nr1, nr2, nr3 };
101+
102+
// collapse dimensions until first broadcast dimension
103+
int64_t cne0[] = { ne0, ne1, ne2, ne3 };
104+
int64_t cne1[] = { ne10, ne11, ne12, ne13 };
105+
size_t cnb0[] = { nb0, nb1, nb2, nb3 };
106+
size_t cnb1[] = { nb10, nb11, nb12, nb13 };
107+
auto collapse = [](int64_t cne[]) {
108+
cne[0] *= cne[1];
109+
cne[1] = cne[2];
110+
cne[2] = cne[3];
111+
cne[3] = 1;
112+
};
113+
114+
auto collapse_nb = [](size_t cnb[], int64_t cne[]) {
115+
cnb[1] *= cne[1];
116+
cnb[2] *= cne[2];
117+
cnb[3] *= cne[3];
118+
};
119+
120+
for (int i = 0; i < 4; i++) {
121+
if (nr[i] != 1) {
122+
break;
123+
}
124+
if (i > 0) {
125+
collapse_nb(cnb0, cne0);
126+
collapse_nb(cnb1, cne1);
127+
collapse(cne0);
128+
collapse(cne1);
129+
}
130+
}
131+
{
132+
int64_t ne0 = cne0[0];
133+
int64_t ne1 = cne0[1];
134+
int64_t ne2 = cne0[2];
135+
int64_t ne3 = cne0[3];
136+
137+
int64_t ne10 = cne1[0];
138+
int64_t ne11 = cne1[1];
139+
int64_t ne12 = cne1[2];
140+
int64_t ne13 = cne1[3];
141+
142+
size_t nb0 = cnb0[0];
143+
size_t nb1 = cnb0[1];
144+
size_t nb2 = cnb0[2];
145+
size_t nb3 = cnb0[3];
146+
147+
size_t nb10 = cnb1[0];
148+
size_t nb11 = cnb1[1];
149+
size_t nb12 = cnb1[2];
150+
size_t nb13 = cnb1[3];
151+
152+
size_t s0 = nb0 / sizeof(dst_t);
153+
size_t s1 = nb1 / sizeof(dst_t);
154+
size_t s2 = nb2 / sizeof(dst_t);
155+
size_t s3 = nb3 / sizeof(dst_t);
156+
157+
size_t s10 = nb10 / sizeof(src1_t);
158+
size_t s11 = nb11 / sizeof(src1_t);
159+
size_t s12 = nb12 / sizeof(src1_t);
160+
size_t s13 = nb13 / sizeof(src1_t);
161+
162+
GGML_ASSERT(s0 == 1);
163+
GGML_ASSERT(s10 == 1);
164+
165+
const int block_size = 128;
166+
167+
int64_t hne0 = std::max(ne0 / 2LL, 1LL);
168+
169+
sycl::range<3> block_dims(1, 1, 1);
170+
block_dims[2] = std::min<unsigned int>(hne0, block_size);
171+
block_dims[1] = std::min<unsigned int>(ne1, block_size / (unsigned int) block_dims[2]);
172+
block_dims[0] = std::min(std::min<unsigned int>(ne2 * ne3, block_size / (unsigned int) block_dims[2] /
173+
(unsigned int) block_dims[1]),
174+
64U);
175+
176+
sycl::range<3> block_nums((ne2 * ne3 + block_dims[0] - 1) / block_dims[0],
177+
(ne1 + block_dims[1] - 1) / block_dims[1],
178+
(hne0 + block_dims[2] - 1) / block_dims[2]);
179+
180+
if (block_nums[0] > 65535) {
181+
// this is the maximum number of blocks in z direction, fallback to 1D grid kernel
182+
int block_num = (ne0 * ne1 * ne2 * ne3 + block_size - 1) / block_size;
183+
{
184+
dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
185+
186+
stream->parallel_for(
187+
sycl::nd_range<3>(sycl::range<3>(1, 1, block_num) * sycl::range<3>(1, 1, block_size),
188+
sycl::range<3>(1, 1, block_size)),
189+
[=](sycl::nd_item<3> item_ct1) {
190+
k_bin_bcast_unravel<bin_op>(src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3, ne10, ne11, ne12,
191+
ne13, s1, s2, s3, s11, s12, s13, item_ct1);
192+
});
193+
}
194+
} else {
195+
/*
196+
DPCT1049:16: The work-group size passed to the SYCL kernel may
197+
exceed the limit. To get the device limit, query
198+
info::device::max_work_group_size. Adjust the work-group size if
199+
needed.
200+
*/
201+
dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
202+
203+
stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
204+
[=](sycl::nd_item<3> item_ct1) {
205+
k_bin_bcast<bin_op>(src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3, ne10, ne11,
206+
ne12, ne13, s1, s2, s3, s11, s12, s13, item_ct1);
207+
});
208+
}
209+
}
210+
}
211+
};
212+
213+
template <class op>
214+
inline void ggml_sycl_op_bin_bcast(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
215+
const void * src0_dd, const void * src1_dd, void * dst_dd,
216+
const queue_ptr & main_stream) {
217+
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
218+
op()(src0, src1, dst, (const float *) src0_dd, (const float *) src1_dd, (float *) dst_dd, main_stream);
219+
} else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
220+
op()(src0, src1, dst, (const sycl::half *) src0_dd, (const float *) src1_dd, (sycl::half *) dst_dd,
221+
main_stream);
222+
} else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
223+
op()(src0, src1, dst, (const sycl::half *) src0_dd, (const float *) src1_dd, (float *) dst_dd, main_stream);
224+
} else if (src0->type == GGML_TYPE_I32 && dst->type == GGML_TYPE_I32) {
225+
op()(src0, src1, dst, (const int32_t *) src0_dd, (const int32_t *) src1_dd, (int32_t *) dst_dd, main_stream);
226+
} else if (src0->type == GGML_TYPE_I16 && dst->type == GGML_TYPE_I16) {
227+
op()(src0, src1, dst, (const int16_t *) src0_dd, (const int16_t *) src1_dd, (int16_t *) dst_dd, main_stream);
228+
} else {
229+
fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__, ggml_type_name(dst->type),
230+
ggml_type_name(src0->type), ggml_type_name(src1->type));
231+
GGML_ABORT("fatal error");
232+
}
233+
}
234+
235+
inline void ggml_sycl_op_add(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
236+
const void * src0_dd = static_cast<void *>(dst->src[0]->data);
237+
const void * src1_dd = static_cast<void *>(dst->src[1]->data);
238+
void * dst_dd = static_cast<void *>(dst->data);
239+
const dpct::queue_ptr main_stream = ctx.stream();
240+
241+
ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_add>>(dst->src[0], dst->src[1], dst, src0_dd, src1_dd, dst_dd,
242+
main_stream);
243+
}
244+
245+
inline void ggml_sycl_op_sub(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
246+
const void * src0_dd = static_cast<void *>(dst->src[0]->data);
247+
const void * src1_dd = static_cast<void *>(dst->src[1]->data);
248+
void * dst_dd = static_cast<void *>(dst->data);
249+
const dpct::queue_ptr main_stream = ctx.stream();
250+
251+
ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_sub>>(dst->src[0], dst->src[1], dst, src0_dd, src1_dd, dst_dd,
252+
main_stream);
253+
}
254+
255+
inline void ggml_sycl_op_mul(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
256+
const void * src0_dd = static_cast<void *>(dst->src[0]->data);
257+
const void * src1_dd = static_cast<void *>(dst->src[1]->data);
258+
void * dst_dd = static_cast<void *>(dst->data);
259+
const dpct::queue_ptr main_stream = ctx.stream();
260+
261+
ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_mul>>(dst->src[0], dst->src[1], dst, src0_dd, src1_dd, dst_dd,
262+
main_stream);
263+
}
264+
265+
inline void ggml_sycl_op_div(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
266+
const void * src0_dd = static_cast<void *>(dst->src[0]->data);
267+
const void * src1_dd = static_cast<void *>(dst->src[1]->data);
268+
void * dst_dd = static_cast<void *>(dst->data);
269+
const dpct::queue_ptr main_stream = ctx.stream();
270+
271+
ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_div>>(dst->src[0], dst->src[1], dst, src0_dd, src1_dd, dst_dd,
272+
main_stream);
273+
}
274+
275+
inline void ggml_sycl_op_repeat(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
276+
const void * src0_d = static_cast<void *>(dst->src[0]->data);
277+
void * dst_d = static_cast<void *>(dst->data);
278+
dpct::queue_ptr main_stream = ctx.stream();
279+
280+
ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_repeat>>(dst, dst->src[0], dst, nullptr, src0_d, dst_d, main_stream);
281+
}
282+
283+
void ggml_sycl_add(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
284+
GGML_SYCL_DEBUG("call %s\n", __func__);
285+
ggml_sycl_op_add(ctx, dst);
286+
GGML_SYCL_DEBUG("call %s done\n", __func__);
287+
}
288+
289+
void ggml_sycl_sub(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
290+
GGML_SYCL_DEBUG("call %s\n", __func__);
291+
ggml_sycl_op_sub(ctx, dst);
292+
GGML_SYCL_DEBUG("call %s done\n", __func__);
293+
}
294+
295+
void ggml_sycl_mul(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
296+
GGML_SYCL_DEBUG("call %s\n", __func__);
297+
ggml_sycl_op_mul(ctx, dst);
298+
GGML_SYCL_DEBUG("call %s done\n", __func__);
299+
}
300+
301+
void ggml_sycl_div(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
302+
GGML_SYCL_DEBUG("call %s\n", __func__);
303+
ggml_sycl_op_div(ctx, dst);
304+
GGML_SYCL_DEBUG("call %s done\n", __func__);
305+
}
306+
307+
void ggml_sycl_repeat(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
308+
GGML_SYCL_DEBUG("call %s\n", __func__);
309+
ggml_sycl_op_repeat(ctx, dst);
310+
GGML_SYCL_DEBUG("call %s done\n", __func__);
311+
}

‎ggml/src/ggml-sycl/binbcast.hpp‎

Lines changed: 16 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,16 @@
1+
#ifndef GGML_SYCL_BINBCAST_HPP
2+
#define GGML_SYCL_BINBCAST_HPP
3+
4+
#include "common.hpp"
5+
6+
void ggml_sycl_add(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
7+
8+
void ggml_sycl_sub(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
9+
10+
void ggml_sycl_mul(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
11+
12+
void ggml_sycl_div(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
13+
14+
void ggml_sycl_repeat(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
15+
16+
#endif // GGML_SYCL_BINBCAST_HPP

0 commit comments

Comments
 (0)