Repository navigation
Expand file tree
/
Copy pathcublasdx_pipeline.cpp
More file actions
526 lines (449 loc) · 23.3 KB
/
Copy pathcublasdx_pipeline.cpp
File metadata and controls
526 lines (449 loc) · 23.3 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
/*
* SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
* SPDX-License-Identifier: Apache-2.0
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#include <cuda.h>
#include <cuda_runtime.h>
#include <libcublasdx.h>
#include <nvJitLink.h>
#include <nvrtc.h>
#include <array>
#include <random>
#include <string>
#include <unordered_map>
#include <vector>
#include "arch.hpp"
#include "common_examples.hpp"
#include "macros.hpp"
using namespace examples;
int main() {
long long int m = 512;
long long int n = 512;
long long int k = 512;
long long int tile_m = 128;
long long int tile_n = 128;
long long int tile_k = 32;
long long int num_threads = 128;
arch_t dx_sm = maybe_accelerated_dx(get_dx_cc());
arch_t target_sm = maybe_accelerated_target(get_target_cc());
auto dx_sm_array = dx_sm.to_array();
auto target_sm_array = target_sm.to_array();
auto [nvrtc_major, nvrtc_minor] = get_nvrtc_version();
if (nvrtc_major == 13 && nvrtc_minor == 0 && target_sm.cc >= cc_t { 10, 0 }) {
printf("Pipeline examples on SM100+ requires NVRTC 13.1 or above.\n");
return 0;
}
if (dx_sm.cc < cc_t { 7, 5 }) {
printf("Pipeline examples require SM75 or above.\n");
return 0;
}
/**
* Create the cuBLASDx descriptor
*/
cublasdxDescriptor h { 0 };
LIBMATHDX_CHECK(cublasdxCreateDescriptor(&h));
LIBMATHDX_CHECK(cublasdxSetOperatorInt64(
h, cublasdxOperatorType::CUBLASDX_OPERATOR_FUNCTION, cublasdxFunction::CUBLASDX_FUNCTION_MM));
LIBMATHDX_CHECK(cublasdxSetOperatorInt64(
h, cublasdxOperatorType::CUBLASDX_OPERATOR_EXECUTION, commondxExecution::COMMONDX_EXECUTION_BLOCK));
LIBMATHDX_CHECK(
cublasdxSetOperatorInt64(h, cublasdxOperatorType::CUBLASDX_OPERATOR_API, cublasdxApi::CUBLASDX_API_TENSORS));
std::array<long long int, 3> prec = { COMMONDX_PRECISION_I8, COMMONDX_PRECISION_I8, COMMONDX_PRECISION_I32 };
LIBMATHDX_CHECK(
cublasdxSetOperatorInt64s(h, cublasdxOperatorType::CUBLASDX_OPERATOR_PRECISION, prec.size(), prec.data()));
LIBMATHDX_CHECK(cublasdxSetOperatorInt64s(
h, cublasdxOperatorType::CUBLASDX_OPERATOR_SM, dx_sm_array.size(), dx_sm_array.data()));
LIBMATHDX_CHECK(
cublasdxSetOperatorInt64(h, cublasdxOperatorType::CUBLASDX_OPERATOR_TYPE, cublasdxType::CUBLASDX_TYPE_REAL));
std::array<long long int, 3> block_dim = { num_threads, 1, 1 };
LIBMATHDX_CHECK(cublasdxSetOperatorInt64s(
h, cublasdxOperatorType::CUBLASDX_OPERATOR_BLOCK_DIM, block_dim.size(), block_dim.data()));
std::array<long long int, 3> size = { tile_m, tile_n, tile_k };
LIBMATHDX_CHECK(
cublasdxSetOperatorInt64s(h, cublasdxOperatorType::CUBLASDX_OPERATOR_SIZE, size.size(), size.data()));
std::array<long long int, 3> arrangement = { CUBLASDX_ARRANGEMENT_ROW_MAJOR,
CUBLASDX_ARRANGEMENT_COL_MAJOR,
CUBLASDX_ARRANGEMENT_ROW_MAJOR };
LIBMATHDX_CHECK(cublasdxSetOperatorInt64s(
h, cublasdxOperatorType::CUBLASDX_OPERATOR_ARRANGEMENT, arrangement.size(), arrangement.data()));
std::array<long long int, 3> alignment = { 16, 16, 16 };
LIBMATHDX_CHECK(cublasdxSetOperatorInt64s(
h, cublasdxOperatorType::CUBLASDX_OPERATOR_ALIGNMENT, alignment.size(), alignment.data()));
LIBMATHDX_CHECK(cublasdxSetOperatorInt64(
h, cublasdxOperatorType::CUBLASDX_OPERATOR_ENABLE_INPUT_STREAMING, CUBLASDX_INPUT_STREAMING_ENABLED));
LIBMATHDX_CHECK(cublasdxSetOperatorInt64(
h, cublasdxOperatorType::CUBLASDX_OPERATOR_WITH_PIPELINE, CUBLASDX_WITH_PIPELINE_ENABLED));
LIBMATHDX_CHECK(cublasdxSetOptionStr(h, commondxOption::COMMONDX_OPTION_SYMBOL_NAME, "matmul"));
/**
* Define the input and output tensors
*/
int8_t* d_a {};
int8_t* d_b {};
int32_t* d_c {};
CUDA_CHECK(cudaMallocManaged(&d_a, m * k * sizeof(int8_t)));
CUDA_CHECK(cudaMallocManaged(&d_b, k * n * sizeof(int8_t)));
CUDA_CHECK(cudaMallocManaged(&d_c, m * n * sizeof(int32_t)));
CUDA_CHECK(cudaMemset(d_c, 0, m * n * sizeof(int32_t)));
std::random_device rd;
std::mt19937 gen(rd());
std::uniform_int_distribution<> dis(0, 10);
for (int i = 0; i < m * k; i++) {
d_a[i] = static_cast<int8_t>(dis(gen));
}
for (int i = 0; i < k * n; i++) {
d_b[i] = static_cast<int8_t>(dis(gen));
}
cublasdxTensor matrix_a { 0 };
cublasdxTensor matrix_b { 0 };
cublasdxTensor tile_gemm_c { 0 };
std::vector<long long int> shape_a = { m, k };
std::vector<long long int> shape_b = { k, n };
// Importantly, the shape of the output tensor is the tile size, not the global size
// the raw pointer must be offset for each CTA tile as shown in the kernel code
std::vector<long long int> shape_c = { tile_m, tile_n };
std::vector<long long int> strides_a = { k, 1 };
std::vector<long long int> strides_b = { 1, k };
// while the shape of the tile gemm c is the tile size, the strides are the global size
std::vector<long long int> strides_c = { n, 1 };
LIBMATHDX_CHECK(cublasdxCreateTensorStrided(CUBLASDX_MEMORY_SPACE_GMEM,
COMMONDX_R_8I,
nullptr,
shape_a.size(),
shape_a.data(),
strides_a.data(),
&matrix_a));
LIBMATHDX_CHECK(cublasdxCreateTensorStrided(CUBLASDX_MEMORY_SPACE_GMEM,
COMMONDX_R_8I,
nullptr,
shape_b.size(),
shape_b.data(),
strides_b.data(),
&matrix_b));
LIBMATHDX_CHECK(cublasdxCreateTensorStrided(CUBLASDX_MEMORY_SPACE_GMEM,
COMMONDX_R_32I,
nullptr,
shape_c.size(),
shape_c.data(),
strides_c.data(),
&tile_gemm_c));
cublasdxTensor acc_c { 0 };
LIBMATHDX_CHECK(cublasdxCreateTensor(h, CUBLASDX_TENSOR_SUGGESTED_ACCUMULATOR_C, &acc_c));
cublasdxPipeline device_pipeline { 0 };
cublasdxPipeline tile_pipeline { 0 };
LIBMATHDX_CHECK(cublasdxCreateDevicePipeline(h,
CUBLASDX_DEVICE_PIPELINE_SUGGESTED,
4,
CUBLASDX_BLOCK_SIZE_STRATEGY_FIXED,
matrix_a,
matrix_b,
&device_pipeline));
LIBMATHDX_CHECK(cublasdxCreateTilePipeline(h, CUBLASDX_TILE_PIPELINE_DEFAULT, device_pipeline, &tile_pipeline));
std::array tensors = { matrix_a, matrix_b, tile_gemm_c, acc_c };
std::array pipelines = { device_pipeline, tile_pipeline };
LIBMATHDX_CHECK(cublasdxFinalize(tensors.size(), tensors.data(), pipelines.size(), pipelines.data()));
// size and alignment of accumulator tensor
long long int acc_size = 0;
LIBMATHDX_CHECK(cublasdxGetTensorTraitInt64(acc_c, CUBLASDX_TENSOR_TRAIT_STORAGE_BYTES, &acc_size));
long long int acc_alignment = 0;
LIBMATHDX_CHECK(cublasdxGetTensorTraitInt64(acc_c, CUBLASDX_TENSOR_TRAIT_ALIGNMENT_BYTES, &acc_alignment));
// size and alignment of the device pipeline
long long int device_pipeline_storage_size = 0;
LIBMATHDX_CHECK(cublasdxGetPipelineTraitInt64(
device_pipeline, CUBLASDX_PIPELINE_TRAIT_STORAGE_BYTES, &device_pipeline_storage_size));
long long int device_pipeline_storage_alignment = 0;
LIBMATHDX_CHECK(cublasdxGetPipelineTraitInt64(
device_pipeline, CUBLASDX_PIPELINE_TRAIT_STORAGE_ALIGNMENT_BYTES, &device_pipeline_storage_alignment));
// size and alignment of the shared memory buffer (on sm_90a and above (accelerated arches with suffix 'a') should
// be 128B aligned for TMA usage)
long long int shared_memory_buffer_size = 0;
LIBMATHDX_CHECK(cublasdxGetPipelineTraitInt64(
device_pipeline, CUBLASDX_PIPELINE_TRAIT_BUFFER_SIZE, &shared_memory_buffer_size));
long long int shared_memory_buffer_alignment = 0;
LIBMATHDX_CHECK(cublasdxGetPipelineTraitInt64(
device_pipeline, CUBLASDX_PIPELINE_TRAIT_BUFFER_ALIGNMENT_BYTES, &shared_memory_buffer_alignment));
// block dimension for GEMM kernel launch (might be different than operator block dim)
std::array<long long int, 3> device_pipeline_block_dim = { 0, 0, 0 };
LIBMATHDX_CHECK(cublasdxGetPipelineTraitInt64s(
device_pipeline, CUBLASDX_PIPELINE_TRAIT_BLOCK_DIM, 3, device_pipeline_block_dim.data()));
// size and alignment of the tile pipeline
long long int tile_pipeline_storage_size = 0;
LIBMATHDX_CHECK(cublasdxGetPipelineTraitInt64(
tile_pipeline, CUBLASDX_PIPELINE_TRAIT_STORAGE_BYTES, &tile_pipeline_storage_size));
long long int tile_pipeline_storage_alignment = 0;
LIBMATHDX_CHECK(cublasdxGetPipelineTraitInt64(
tile_pipeline, CUBLASDX_PIPELINE_TRAIT_STORAGE_ALIGNMENT_BYTES, &tile_pipeline_storage_alignment));
/**
* Define a function operating on those input and output tensors.
*
* The device function output is an opaque and stateful accumulator.
*/
cublasdxDeviceFunction init_acc { 0 };
LIBMATHDX_CHECK(cublasdxCreateDeviceFunctionWithPipelines(
h, CUBLASDX_DEVICE_FUNCTION_CREATE, 1, &acc_c, 1, &tile_pipeline, &init_acc));
cublasdxDeviceFunction init_device_pipeline { 0 };
LIBMATHDX_CHECK(cublasdxCreateDeviceFunctionWithPipelines(
h, CUBLASDX_DEVICE_FUNCTION_CREATE, 0, nullptr, 1, &device_pipeline, &init_device_pipeline));
cublasdxDeviceFunction init_tile_pipeline { 0 };
LIBMATHDX_CHECK(cublasdxCreateDeviceFunctionWithPipelines(
h, CUBLASDX_DEVICE_FUNCTION_CREATE, 0, nullptr, 1, &tile_pipeline, &init_tile_pipeline));
cublasdxDeviceFunction destroy_acc { 0 };
LIBMATHDX_CHECK(cublasdxCreateDeviceFunction(h, CUBLASDX_DEVICE_FUNCTION_DESTROY, 1, &acc_c, &destroy_acc));
cublasdxDeviceFunction destroy_device_pipeline { 0 };
LIBMATHDX_CHECK(cublasdxCreateDeviceFunctionWithPipelines(
h, CUBLASDX_DEVICE_FUNCTION_DESTROY, 0, nullptr, 1, &device_pipeline, &destroy_device_pipeline));
cublasdxDeviceFunction destroy_tile_pipeline { 0 };
LIBMATHDX_CHECK(cublasdxCreateDeviceFunctionWithPipelines(
h, CUBLASDX_DEVICE_FUNCTION_DESTROY, 0, nullptr, 1, &tile_pipeline, &destroy_tile_pipeline));
cublasdxDeviceFunction execute_pipeline { 0 };
LIBMATHDX_CHECK(cublasdxCreateDeviceFunctionWithPipelines(
h, CUBLASDX_DEVICE_FUNCTION_EXECUTE, 1, &acc_c, 1, &tile_pipeline, &execute_pipeline));
// Finalize the reusable accumulator after execute and before reading it back with copy
// (the non-epilogue read path does not finalize on its own).
cublasdxDeviceFunction finish_accumulation { 0 };
LIBMATHDX_CHECK(
cublasdxCreateDeviceFunction(h, CUBLASDX_DEVICE_FUNCTION_FINISH_ACCUMULATION, 1, &acc_c, &finish_accumulation));
std::array copy_tensors = { acc_c, tile_gemm_c };
cublasdxDeviceFunction copy_acc_big_gmem_c { 0 };
LIBMATHDX_CHECK(cublasdxCreateDeviceFunction(
h, CUBLASDX_DEVICE_FUNCTION_COPY, copy_tensors.size(), copy_tensors.data(), ©_acc_big_gmem_c));
std::vector<cublasdxDeviceFunction> functions = {
execute_pipeline, finish_accumulation, init_acc,
init_device_pipeline, init_tile_pipeline, destroy_acc,
destroy_device_pipeline, destroy_tile_pipeline, copy_acc_big_gmem_c
};
std::unordered_map<cublasdxDeviceFunction, std::string> function_symbols;
for (auto f : functions) {
size_t symbol_size { 0 };
LIBMATHDX_CHECK(cublasdxGetDeviceFunctionTraitStrSize(f, CUBLASDX_DEVICE_FUNCTION_TRAIT_SYMBOL, &symbol_size));
std::vector<char> symbol(symbol_size, '\0');
LIBMATHDX_CHECK(
cublasdxGetDeviceFunctionTraitStr(f, CUBLASDX_DEVICE_FUNCTION_TRAIT_SYMBOL, symbol.size(), symbol.data()));
printf("Device function %lld: symbol: %s\n", static_cast<long long int>(f), symbol.data());
function_symbols[f] = std::string(symbol.data());
}
/**
* Compile the device functions to LTOIR
*/
commondxCode code { 0 };
LIBMATHDX_CHECK(commondxCreateCode(&code));
LIBMATHDX_CHECK(
commondxSetCodeOptionInt64s(code, COMMONDX_OPTION_TARGET_SM, target_sm_array.size(), target_sm_array.data()));
LIBMATHDX_CHECK(cublasdxFinalizeDeviceFunctions(code, functions.size(), functions.data()));
/**
* Extract the LTOIR
*/
size_t lto_size = 0;
LIBMATHDX_CHECK(commondxGetCodeLTOIRSize(code, <o_size));
std::vector<char> lto(lto_size);
LIBMATHDX_CHECK(commondxGetCodeLTOIR(code, lto_size, lto.data()));
printf("Generated LTOIR for GEMM pipeline, %zu bytes\n", lto.size());
const char kernel_template[] = R"(
struct libmathdx_tensor_0s_0s { void* ptr; };
struct libmathdx_pipeline { void* ptr; };
#define M %d
#define N %d
#define K %d
#define tile_m %d
#define tile_n %d
#define tile_k %d
#define acc_name libmathdx_tensor_0s_0s
#define ga_name libmathdx_tensor_0s_0s
#define gb_name libmathdx_tensor_0s_0s
#define gc_name libmathdx_tensor_0s_0s
#define device_pipeline_name libmathdx_pipeline
#define tile_pipeline_name libmathdx_pipeline
constexpr unsigned acc_size = %lld;
constexpr unsigned acc_alignment = %lld;
constexpr unsigned block_size = %lld;
constexpr unsigned smem_alignment = %lld;
constexpr unsigned tile_pipeline_size = %lld;
constexpr unsigned tile_pipeline_alignment = %lld;
#define execute_pipeline_acc %s
#define finish_acc %s
#define copy_acc_gc %s
#define create_dev_pipe %s
#define create_tile_pipe %s
#define create_acc %s
#define destroy_dev_pipe %s
#define destroy_tile_pipe %s
#define destroy_acc %s
using C_VALUE_TYPE = signed int;
extern "C" __device__ void execute_pipeline_acc(tile_pipeline_name, acc_name);
extern "C" __device__ void finish_acc(acc_name);
extern "C" __device__ void copy_acc_gc(acc_name, gc_name);
extern "C" __device__ void create_dev_pipe(device_pipeline_name, ga_name, gb_name);
extern "C" __device__ void create_tile_pipe(device_pipeline_name, tile_pipeline_name, char*, int*, int*);
extern "C" __device__ void create_acc(tile_pipeline_name, acc_name);
extern "C" __device__ void destroy_dev_pipe(device_pipeline_name);
extern "C" __device__ void destroy_tile_pipe(tile_pipeline_name);
extern "C" __device__ void destroy_acc(acc_name);
// Create the device pipeline
extern "C" __global__ void create_device_pipeline(void* device_pipeline_ptr, void* ga_storage, void* gb_storage) {
if(threadIdx.x == 0) {
auto ga = ga_name { ga_storage };
auto gb = gb_name { gb_storage };
auto device_pipeline = device_pipeline_name { device_pipeline_ptr };
create_dev_pipe(device_pipeline, ga, gb);
}
}
// Perform the GEMM
extern "C" __launch_bounds__(block_size, 1) __global__ void gemm(void* device_pipeline_ptr, void* gc_storage)
{
int row_id = blockIdx.x;
int col_id = blockIdx.y;
C_VALUE_TYPE* gmem_c = reinterpret_cast<C_VALUE_TYPE*>(gc_storage) + ((row_id * tile_m) * N + (col_id * tile_n));
// Allocate dynamic shared memory for tiles
extern __shared__ __align__(smem_alignment) char smem[];
// Allocate local memory
alignas(acc_alignment) char acc_storage[acc_size];
alignas(tile_pipeline_alignment) char tile_pipeline_storage[tile_pipeline_size];
// Create opaque types
auto acc = acc_name { acc_storage };
auto gc = gc_name { gmem_c };
auto device_pipeline = device_pipeline_name { device_pipeline_ptr };
auto tile_pipeline = tile_pipeline_name { tile_pipeline_storage };
// Perform the GEMM
create_tile_pipe(device_pipeline, tile_pipeline, smem, &row_id, &col_id);
create_acc(tile_pipeline, acc);
execute_pipeline_acc(tile_pipeline, acc);
finish_acc(acc);
copy_acc_gc(acc, gc);
// Destroy the accumulator and the tile pipeline
destroy_acc(acc);
destroy_tile_pipe(tile_pipeline);
}
// Destroy the device pipeline
extern "C" __global__ void destroy_device_pipeline(void* device_pipeline_ptr) {
if(threadIdx.x == 0) {
auto device_pipeline = device_pipeline_name { device_pipeline_ptr };
destroy_dev_pipe(device_pipeline);
}
}
)";
std::string cpp = strprintf(kernel_template,
m,
n,
k,
tile_m,
tile_n,
tile_k,
acc_size,
acc_alignment,
device_pipeline_block_dim[0],
shared_memory_buffer_alignment,
tile_pipeline_storage_size,
tile_pipeline_storage_alignment,
function_symbols[execute_pipeline].data(),
function_symbols[finish_accumulation].data(),
function_symbols[copy_acc_big_gmem_c].data(),
function_symbols[init_device_pipeline].data(),
function_symbols[init_tile_pipeline].data(),
function_symbols[init_acc].data(),
function_symbols[destroy_device_pipeline].data(),
function_symbols[destroy_tile_pipeline].data(),
function_symbols[destroy_acc].data());
std::vector<char> cubin = compile_and_link(cpp, lto, target_sm);
CUmodule module {};
CUfunction kernel, kernel2, kernel3 {};
CUDA_CHECK(cudaSetDevice(0));
CU_CHECK(cuModuleLoadDataEx(&module, cubin.data(), 0, 0, 0));
CU_CHECK(cuModuleGetFunction(&kernel, module, "create_device_pipeline"));
CU_CHECK(cuModuleGetFunction(&kernel2, module, "gemm"));
CU_CHECK(cuModuleGetFunction(&kernel3, module, "destroy_device_pipeline"));
// create the device pipeline using trait
uint8_t* device_pipeline_ptr {};
CUDA_CHECK(cudaMallocManaged(&device_pipeline_ptr, device_pipeline_storage_size));
CUDA_CHECK(cudaMemset(device_pipeline_ptr, 0, device_pipeline_storage_size));
{
std::vector<void*> kernel_args = { reinterpret_cast<void*>(&device_pipeline_ptr),
reinterpret_cast<void*>(&d_a),
reinterpret_cast<void*>(&d_b) };
CU_CHECK(cuLaunchKernel(kernel, 1, 1, 1, 1, 1, 1, 0, nullptr, kernel_args.data(), nullptr));
CUDA_CHECK(cudaDeviceSynchronize());
}
{
printf("Checking device pipeline bytes\n");
printf("Pointers to gmem A & B: %p %p\n", d_a, d_b);
printf("Device pipeline bytes: \n");
for (long long int i = 0; i < device_pipeline_storage_size; i++) {
if (i > 0 && i % 32 == 0) printf("\n");
printf("%02x ", device_pipeline_ptr[i]);
}
printf("\n");
}
{
CU_CHECK(cuFuncSetAttribute(
kernel2, CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES, static_cast<int>(shared_memory_buffer_size)));
std::vector<void*> kernel_args = { reinterpret_cast<void*>(&device_pipeline_ptr),
reinterpret_cast<void*>(&d_c) };
CU_CHECK(cuLaunchKernel(kernel2,
static_cast<unsigned int>(m / tile_m),
static_cast<unsigned int>(n / tile_n),
1,
static_cast<unsigned int>(device_pipeline_block_dim[0]),
static_cast<unsigned int>(device_pipeline_block_dim[1]),
static_cast<unsigned int>(device_pipeline_block_dim[2]),
static_cast<unsigned int>(shared_memory_buffer_size),
nullptr,
kernel_args.data(),
nullptr));
CUDA_CHECK(cudaDeviceSynchronize());
}
{
std::vector<void*> kernel_args = { reinterpret_cast<void*>(&device_pipeline_ptr) };
CU_CHECK(cuLaunchKernel(kernel3, 1, 1, 1, 1, 1, 1, 0, nullptr, kernel_args.data(), nullptr));
CUDA_CHECK(cudaDeviceSynchronize());
}
std::vector<int32_t> h_c_ref(m * n, 0);
for (int l = 0; l < k; l++) {
for (int i = 0; i < m; i++) {
for (int j = 0; j < n; j++) {
h_c_ref[i * n + j] += d_a[i * k + l] * d_b[j * k + l];
}
}
}
for (int i = 0; i < m * n; i++) {
if (h_c_ref[i] != d_c[i]) {
printf("Error at %d: h_c_ref[%d] = %d, d_c[%d] = %d\n", i, i, h_c_ref[i], i, d_c[i]);
abort();
}
}
printf("Successfully ran the kernel\n");
/**
* Destroy handles
*/
CU_CHECK(cuModuleUnload(module));
CUDA_CHECK(cudaFree(device_pipeline_ptr));
CUDA_CHECK(cudaFree(d_a));
CUDA_CHECK(cudaFree(d_b));
CUDA_CHECK(cudaFree(d_c));
LIBMATHDX_CHECK(cublasdxDestroyTensor(matrix_a));
LIBMATHDX_CHECK(cublasdxDestroyTensor(matrix_b));
LIBMATHDX_CHECK(cublasdxDestroyTensor(acc_c));
LIBMATHDX_CHECK(cublasdxDestroyTensor(tile_gemm_c));
LIBMATHDX_CHECK(cublasdxDestroyPipeline(device_pipeline));
LIBMATHDX_CHECK(cublasdxDestroyPipeline(tile_pipeline));
LIBMATHDX_CHECK(cublasdxDestroyDeviceFunction(init_acc));
LIBMATHDX_CHECK(cublasdxDestroyDeviceFunction(init_device_pipeline));
LIBMATHDX_CHECK(cublasdxDestroyDeviceFunction(init_tile_pipeline));
LIBMATHDX_CHECK(cublasdxDestroyDeviceFunction(destroy_acc));
LIBMATHDX_CHECK(cublasdxDestroyDeviceFunction(destroy_device_pipeline));
LIBMATHDX_CHECK(cublasdxDestroyDeviceFunction(destroy_tile_pipeline));
LIBMATHDX_CHECK(cublasdxDestroyDeviceFunction(execute_pipeline));
LIBMATHDX_CHECK(cublasdxDestroyDeviceFunction(finish_accumulation));
LIBMATHDX_CHECK(cublasdxDestroyDeviceFunction(copy_acc_big_gmem_c));
LIBMATHDX_CHECK(commondxDestroyCode(code));
LIBMATHDX_CHECK(cublasdxDestroyDescriptor(h));
}