Skip to content

Commit 37d1d51

Browse files
dkhaldidkhaldi
authored
[SYCL][E2E][Joint Matrix] update performance test to add SLM (#15229)
1 parent 594ae74 commit 37d1d51

File tree

6 files changed

+217
-51
lines changed

6 files changed

+217
-51
lines changed

‎sycl/test-e2e/Matrix/SG32/joint_matrix_bf16_fill_k_cache_SLM.cpp

Lines changed: 23 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,23 @@
1+
//==--- joint_matrix_bf16_fill_k_cache_SLM.cpp - DPC++ joint_matrix--------==//
2+
//
3+
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4+
// See https://llvm.org/LICENSE.txt for license information.
5+
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6+
//
7+
//===----------------------------------------------------------------------===//
8+
// SG size = 32 is not currently supported for SYCL Joint Matrix by IGC on DG2
9+
// UNSUPPORTED: gpu-intel-dg2
10+
// REQUIRES: aspect-ext_intel_matrix, gpu
11+
12+
// RUN: %{build} -o %t_gpu_vnni.out -ffp-model=precise -DSLM -DVNNI
13+
// RUN: %{run} %t_gpu_vnni.out
14+
15+
// TODO: add row major compilation and run once Sub-group size 32
16+
// support becomes available in IGC for row major
17+
18+
// -ffp-model=precise is added to not depend on compiler defaults.
19+
20+
#include "../common.hpp"
21+
#define SG_SZ 32
22+
23+
#include "../joint_matrix_bf16_fill_k_cache_impl.hpp"

‎sycl/test-e2e/Matrix/SPVCooperativeMatrix/SG32/joint_matrix_bf16_fill_k_cache_SLM.cpp

Lines changed: 23 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,23 @@
1+
//==--- joint_matrix_bf16_fill_k_cache_SLM.cpp - DPC++ joint_matrix--------==//
2+
//
3+
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4+
// See https://llvm.org/LICENSE.txt for license information.
5+
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6+
//
7+
//===----------------------------------------------------------------------===//
8+
// SG size = 32 is not currently supported for SYCL Joint Matrix by IGC on DG2
9+
// UNSUPPORTED: gpu-intel-dg2
10+
// REQUIRES: aspect-ext_intel_matrix, gpu
11+
12+
// RUN: %{build} -D__SPIRV_USE_COOPERATIVE_MATRIX -o %t_gpu_vnni.out -ffp-model=precise -DSLM -DVNNI
13+
// RUN: %{run} %t_gpu_vnni.out
14+
15+
// TODO: add row major compilation and run once Sub-group size 32
16+
// support becomes available in IGC for row major
17+
18+
// -ffp-model=precise is added to not depend on compiler defaults.
19+
20+
#include "../../common.hpp"
21+
#define SG_SZ 32
22+
23+
#include "../../joint_matrix_bf16_fill_k_cache_impl.hpp"

‎sycl/test-e2e/Matrix/SPVCooperativeMatrix/joint_matrix_bf16_fill_k_cache_SLM.cpp

Lines changed: 19 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,19 @@
1+
//==--- joint_matrix_bf16_fill_k_cache_SLM.cpp - DPC++ joint_matrix--------==//
2+
//
3+
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4+
// See https://llvm.org/LICENSE.txt for license information.
5+
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6+
//
7+
//===----------------------------------------------------------------------===//
8+
// REQUIRES: aspect-ext_intel_matrix, gpu
9+
10+
// RUN: %{build} -D__SPIRV_USE_COOPERATIVE_MATRIX -o %t_gpu_vnni.out -ffp-model=precise -DSLM -DVNNI
11+
// RUN: %{run} %t_gpu_vnni.out
12+
13+
// RUN: %{build} -D__SPIRV_USE_COOPERATIVE_MATRIX -o %t_gpu.out -ffp-model=precise -DSLM
14+
// RUN: %{run} %t_gpu.out
15+
16+
// -ffp-model=precise is added to not depend on compiler defaults.
17+
18+
#include "../common.hpp"
19+
#include "../joint_matrix_bf16_fill_k_cache_impl.hpp"

‎sycl/test-e2e/Matrix/joint_matrix_bf16_fill_k_cache_SLM.cpp

Lines changed: 19 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,19 @@
1+
//==--- joint_matrix_bf16_fill_k_cache_SLM.cpp - DPC++ joint_matrix--------==//
2+
//
3+
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4+
// See https://llvm.org/LICENSE.txt for license information.
5+
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6+
//
7+
//===----------------------------------------------------------------------===//
8+
// REQUIRES: aspect-ext_intel_matrix, gpu
9+
10+
// RUN: %{build} -o %t_gpu_vnni.out -ffp-model=precise -DSLM -DVNNI
11+
// RUN: %{run} %t_gpu_vnni.out
12+
13+
// RUN: %{build} -o %t_gpu.out -ffp-model=precise -DSLM
14+
// RUN: %{run} %t_gpu.out
15+
16+
// -ffp-model=precise is added to not depend on compiler defaults.
17+
18+
#include "common.hpp"
19+
#include "joint_matrix_bf16_fill_k_cache_impl.hpp"

‎sycl/test-e2e/Matrix/joint_matrix_bf16_fill_k_cache_impl.hpp

Lines changed: 63 additions & 51 deletions
Original file line numberDiff line numberDiff line change
@@ -9,6 +9,10 @@
99
#include <random>
1010
#include <sycl/usm.hpp>
1111

12+
#ifdef SLM
13+
#include "slm_utils.hpp"
14+
#endif
15+
1216
// number of test iterations
1317
constexpr unsigned int testIterations = 100;
1418
// start recording time after X iterations
@@ -51,6 +55,12 @@ double joint_matmul(TOperand *A, TOperand *B, TResult *C, queue &q, int i) {
5155
std::chrono::high_resolution_clock::now();
5256

5357
q.submit([&](handler &h) {
58+
#ifdef SLM
59+
local_accessor<TOperand, 2> tileA{{MCache2, KCache2}, h};
60+
local_accessor<TOperand, 2> tileB{
61+
{KCache2 / vnniFactor, NCache2 * vnniFactor}, h};
62+
#endif
63+
5464
h.parallel_for<MatMul<TM, TN, TK>>( // cache layer#1
5565
nd_range<2>{global, cachelocal},
5666
// loop global
@@ -60,15 +70,16 @@ double joint_matmul(TOperand *A, TOperand *B, TResult *C, queue &q, int i) {
6070
[[intel::reqd_sub_group_size(SG_SZ)]]
6171
#endif // SG_SZ
6272
{
73+
// sg::load and sg::store expect decorations to be ON
6374
auto pA =
6475
address_space_cast<sycl::access::address_space::global_space,
65-
sycl::access::decorated::no>(A);
76+
sycl::access::decorated::yes>(A);
6677
auto pB =
6778
address_space_cast<sycl::access::address_space::global_space,
68-
sycl::access::decorated::no>(B);
79+
sycl::access::decorated::yes>(B);
6980
auto pC =
7081
address_space_cast<sycl::access::address_space::global_space,
71-
sycl::access::decorated::no>(C);
82+
sycl::access::decorated::yes>(C);
7283
auto m2 = it.get_group(0);
7384
auto n2 = it.get_group(1);
7485
auto m1 = it.get_local_id(0);
@@ -112,7 +123,6 @@ double joint_matmul(TOperand *A, TOperand *B, TResult *C, queue &q, int i) {
112123
colsA, layout::row_major,
113124
syclex::properties{syclex::prefetch_hint_L1});
114125

115-
#ifdef VNNI
116126
for (int p = 0; p < prefDistance; p++)
117127
joint_matrix_prefetch<prefRow, prefCol>(
118128
sg,
@@ -122,15 +132,6 @@ double joint_matmul(TOperand *A, TOperand *B, TResult *C, queue &q, int i) {
122132
(n2 * NCache2 * vnniFactor + pn1B * prefCol),
123133
colsB * vnniFactor, layout::row_major,
124134
syclex::properties{syclex::prefetch_hint_L1});
125-
#else // VNNI
126-
for (int p = 0; p < prefDistance; p++)
127-
joint_matrix_prefetch<prefRow, prefCol>(
128-
sg,
129-
B + (p * KCache2 + pm1B * prefRow) * colsB + n2 * NCache2 +
130-
pn1B * prefCol,
131-
colsB, layout::row_major,
132-
syclex::properties{syclex::prefetch_hint_L1});
133-
#endif // VNNI
134135
#endif // PREFETCH
135136

136137
joint_matrix<sub_group, TResult, use::accumulator, TM, TN>
@@ -157,7 +158,16 @@ double joint_matmul(TOperand *A, TOperand *B, TResult *C, queue &q, int i) {
157158
}
158159
#endif // MANUAL_UNROLL
159160

161+
#ifdef SLM
162+
constexpr unsigned int SGs =
163+
(MCache2 / MCache1) * (NCache2 / NCache1);
164+
#endif // SLM
160165
for (unsigned int k2 = 0; k2 < colsA / KCache2; k2++) {
166+
#ifdef SLM
167+
slm_read_write<colsA, colsB, MCache2, NCache2, KCache2, vnniFactor,
168+
SGs>(pA, pB, tileA, tileB, sg, k2, m2, n2, sgSize);
169+
it.barrier(access::fence_space::local_space);
170+
#endif // SLM
161171
joint_matrix<sub_group, TOperand, use::a, TM, TK, layout::row_major>
162172
tA[MCache1 / TM][KCache2 / KCache1]
163173
#ifdef INIT_LIST
@@ -192,6 +202,14 @@ double joint_matmul(TOperand *A, TOperand *B, TResult *C, queue &q, int i) {
192202
#else // MANUAL_UNROLL
193203
for (unsigned int m = 0; m < MCache1 / TM; m++) {
194204
#endif // MANUAL_UNROLL
205+
#ifdef SLM
206+
joint_matrix_load(sg, tA[m][k1],
207+
tileA.template get_multi_ptr<
208+
sycl::access::decorated::no>() +
209+
(m1 * MCache1 + m * TM) * KCache2 +
210+
k1 * TK,
211+
KCache2);
212+
#else // SLM
195213
#ifdef OOB
196214
ext::intel::experimental::matrix::joint_matrix_load_checked(
197215
sg, tA[m][k1], pA, colsA, rowsA, colsA,
@@ -203,6 +221,7 @@ double joint_matmul(TOperand *A, TOperand *B, TResult *C, queue &q, int i) {
203221
k * TK,
204222
colsA);
205223
#endif // OOB
224+
#endif // SLM
206225
#ifdef MANUAL_UNROLL
207226
}); // m
208227
#else // MANUAL_UNROLL
@@ -213,32 +232,28 @@ double joint_matmul(TOperand *A, TOperand *B, TResult *C, queue &q, int i) {
213232
#else // MANUAL_UNROLL
214233
for (unsigned int n = 0; n < NCache1 / TN; n++) {
215234
#endif // MANUAL_UNROLL
235+
#ifdef SLM
236+
joint_matrix_load(sg, tB[n][k1],
237+
tileB.template get_multi_ptr<
238+
sycl::access::decorated::no>() +
239+
(k1 * TK / vnniFactor) *
240+
(NCache2 * vnniFactor) +
241+
(n1 * NCache1 + n * TN) * vnniFactor,
242+
NCache2 * vnniFactor);
243+
#else // SLM
216244
#ifdef OOB
217-
#ifdef VNNI
218245
ext::intel::experimental::matrix::joint_matrix_load_checked(
219246
sg, tB[n][k1], pB, colsB * vnniFactor, rowsB / vnniFactor,
220247
colsB * vnniFactor, k * TK / vnniFactor,
221248
(n2 * NCache2 + n1 * NCache1 + n * TN) * vnniFactor);
222-
#else // VNNI
223-
ext::intel::experimental::matrix::joint_matrix_load_checked(
224-
sg, tB[n][k1], pB, colsB, rowsB, colsB, k * TK,
225-
n2 * NCache2 + n1 * NCache1 + n * TN);
226-
227-
#endif // VNNI
228249
#else // OOB
229-
#ifdef VNNI
230250
joint_matrix_load(
231251
sg, tB[n][k1],
232252
pB + (k * TK / vnniFactor) * (colsB * vnniFactor) +
233253
(n2 * NCache2 + n1 * NCache1 + n * TN) * vnniFactor,
234254
colsB * vnniFactor);
235-
#else // VNNI
236-
joint_matrix_load(sg, tB[n][k1],
237-
pB + (k * TK) * (colsB) +
238-
(n2 * NCache2 + n1 * NCache1 + n * TN),
239-
colsB);
240-
#endif // VNNI
241255
#endif // OOB
256+
#endif // SLM
242257
#ifdef MANUAL_UNROLL
243258
}); // n
244259
#else // MANUAL_UNROLL
@@ -266,6 +281,9 @@ double joint_matmul(TOperand *A, TOperand *B, TResult *C, queue &q, int i) {
266281
} // m
267282
} // k1
268283
#endif // MANUAL_UNROLL
284+
#ifdef SLM
285+
it.barrier(access::fence_space::local_space);
286+
#endif // SLM
269287
#ifdef PREFETCH
270288
auto prefetch_offsetA = (m2 * MCache2 + sgId * prefRow) * colsA +
271289
(k2 + prefDistance) * prefCol;
@@ -275,7 +293,6 @@ double joint_matmul(TOperand *A, TOperand *B, TResult *C, queue &q, int i) {
275293
sg, A + prefetch_offsetA, colsA, layout::row_major,
276294
syclex::properties{syclex::prefetch_hint_L1});
277295

278-
#ifdef VNNI
279296
auto prefetch_offsetB =
280297
((k2 + prefDistance) * (KCache2 / vnniFactor) +
281298
pm1B * prefRow) *
@@ -287,16 +304,6 @@ double joint_matmul(TOperand *A, TOperand *B, TResult *C, queue &q, int i) {
287304
sg, B + prefetch_offsetB, colsB * vnniFactor,
288305
layout::row_major,
289306
syclex::properties{syclex::prefetch_hint_L1});
290-
#else // VNNI
291-
auto prefetch_offsetB =
292-
((k2 + prefDistance) * KCache2 + pm1B * prefRow) * (colsB) +
293-
(n2 * NCache2 + pn1B * prefCol);
294-
if ((prefetch_offsetB + (prefRow * MATRIX_SIZE) + prefCol) <
295-
(MATRIX_SIZE * MATRIX_SIZE))
296-
joint_matrix_prefetch<prefRow, prefCol>(
297-
sg, B + prefetch_offsetB, colsB, layout::row_major,
298-
syclex::properties{syclex::prefetch_hint_L1});
299-
#endif // VNNI
300307
#endif // PREFETCH
301308
} // for k2
302309
#ifdef MANUAL_UNROLL
@@ -411,29 +418,33 @@ int main() {
411418
constexpr size_t NCache2 = 256;
412419
constexpr size_t KCache2 = 32;
413420

421+
#ifdef VNNI
422+
constexpr unsigned int VnniFactor = 2;
423+
#else // VNNI
424+
constexpr unsigned int VnniFactor = 1;
425+
#endif // VNNI
426+
414427
for (unsigned int i = 0; i < combinations.size(); i++) {
415428
if (combinations[i].nsize == 0) { // Intel AMX
416429
constexpr size_t NCache1 = 32;
417430
constexpr size_t KCache1 = 32;
418-
419-
test<bfloat16, float, 2, /*TM*/ 16, /*TN*/ 16, /*TK*/ 32, MCache1,
420-
NCache1, KCache1, MCache2, NCache2, KCache2>();
431+
test<bfloat16, float, VnniFactor, /*TM*/ 16, /*TN*/ 16, /*TK*/ 32,
432+
MCache1, NCache1, KCache1, MCache2, NCache2, KCache2>();
421433
break;
422434
}
423435

424436
if (combinations[i].nsize == 16) { // architecture::intel_gpu_pvc
425437
constexpr size_t NCache1 = 4 * /*TN*/ 16;
426438
constexpr size_t KCache1 = 16;
427-
428-
test<bfloat16, float, 2, /*TM*/ 8, /*TN*/ 16, /*TK*/ 16, MCache1, NCache1,
429-
KCache1, MCache2, NCache2, KCache2>();
439+
test<bfloat16, float, VnniFactor, /*TM*/ 8, /*TN*/ 16, /*TK*/ 16, MCache1,
440+
NCache1, KCache1, MCache2, NCache2, KCache2>();
430441
#if (!defined(SG_SZ) || SG_SZ != 32)
431442
// These combination are not currently supported for subgroup size = 32 in
432443
// IGC
433-
test<bfloat16, float, 2, /*TM*/ 16, /*TN*/ 16, /*TK*/ 16, MCache1,
434-
NCache1, KCache1, MCache2, NCache2, KCache2>();
435-
test<bfloat16, float, 2, /*TM*/ 32, /*TN*/ 64, /*TK*/ 16, MCache1,
436-
NCache1, KCache1, MCache2, NCache2, KCache2>();
444+
test<bfloat16, float, VnniFactor, /*TM*/ 16, /*TN*/ 16, /*TK*/ 16,
445+
MCache1, NCache1, KCache1, MCache2, NCache2, KCache2>();
446+
test<bfloat16, float, VnniFactor, /*TM*/ 32, /*TN*/ 64, /*TK*/ 16,
447+
MCache1, NCache1, KCache1, MCache2, NCache2, KCache2>();
437448
#endif
438449
break;
439450
}
@@ -442,9 +453,10 @@ int main() {
442453
constexpr size_t NCache1 = 4 * /*TN*/ 8;
443454
constexpr size_t KCache1 = 16;
444455

445-
test<bfloat16, float, 2, /*TM*/ 8, /*TN*/ 8, /*TK*/ 16, MCache1, NCache1,
446-
KCache1, MCache2, NCache2, KCache2>();
447-
// test<bfloat16, float, 2, /*TM*/ 32, /*TN*/ 32, /*TK*/ 16, MCache1,
456+
test<bfloat16, float, VnniFactor, /*TM*/ 8, /*TN*/ 8, /*TK*/ 16, MCache1,
457+
NCache1, KCache1, MCache2, NCache2, KCache2>();
458+
// test<bfloat16, float, VnniFactor, /*TM*/ 32, /*TN*/ 32, /*TK*/ 16,
459+
// MCache1,
448460
// NCache1, KCache1, MCache2, NCache2, KCache2>();
449461
break;
450462
}

0 commit comments

Comments
 (0)