forked from pytorch/pytorch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
executor_utils.cpp
418 lines (359 loc) · 12.9 KB
/
executor_utils.cpp
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
#include <ATen/CUDAGeneratorImpl.h>
#include <ATen/cuda/CUDAContext.h>
#include <ATen/cuda/nvrtc_stub/ATenNVRTC.h>
#include <c10/cuda/CUDACachingAllocator.h>
#include <torch/csrc/jit/codegen/cuda/executor_utils.h>
#include <torch/csrc/jit/codegen/cuda/instrumentation.h>
#include <torch/csrc/jit/codegen/cuda/ir_all_nodes.h>
#include <torch/csrc/jit/codegen/cuda/kernel_resource_strings.h>
#include <torch/csrc/jit/resource_guard.h>
#include <fstream>
namespace torch {
namespace jit {
namespace fuser {
namespace cuda {
namespace executor_utils {
std::string kernelPreamble() {
std::stringstream ss;
ss << code_template_tensor_struct << "\n"
<< code_fp16_support << "\n"
<< code_random_number_gen << "\n"
<< code_helper_funcs << "\n"
<< code_template_block_reduction << "\n"
<< code_template_grid_reduction << "\n"
<< code_template_block_broadcast << "\n";
return ss.str();
}
namespace {
// return false if arg's type, number of dimensions, and device, doesn't match
// param and provided c10:device
bool validateKernelArgTensor(
const at::Tensor& arg,
const Val* param,
const c10::Device& device,
std::stringstream& msg) {
// Arg is a tensor. Param must be a tensor too.
if (*param->getValType() != ValType::TensorView) {
msg << "Argument is a tensor, but the parameter is not.\n";
return false;
}
// Check the rank of the tensors.
size_t arg_dim = arg.dim();
// Note: This requires current Fusion to be active.
size_t param_dim =
TensorDomain::noReductions(param->as<TensorView>()->getRootDomain())
.size();
// see [Note - broadcast support in integration]
// Because of broadcasting support handled in integration, we relax the rank
// check as necessary.
if (arg_dim > param_dim) {
msg << "Argument tensor's rank is " << arg_dim << ", but the parameter is "
<< param_dim << "\n";
return false;
}
if (arg.device() != device) {
msg << "Argument is on device that is not compiled for."
<< "\n";
return false;
}
// Check element type
at::ScalarType arg_data_type = arg.scalar_type();
DataType param_data_type = *param->getDataType();
bool match = false;
switch (arg_data_type) {
case at::ScalarType::Half:
match = param_data_type == DataType::Half;
break;
case at::ScalarType::Float:
match = param_data_type == DataType::Float;
break;
case at::ScalarType::Bool:
match = param_data_type == DataType::Bool;
break;
default:
msg << "Argument element type, " << arg_data_type << ", is not supported."
<< "\n";
return false;
}
if (!match)
msg << "Argument element type is " << arg_data_type
<< ", but the parameter is " << param_data_type << "\n";
return match;
}
// Return false if arg_type doesn't match the type in param
bool validateKernelArgScalar(
const c10::TypePtr& arg_type,
const Val* param,
std::stringstream& msg) {
if (!param->isScalar()) {
msg << "Argument is a scalar, but the parameter is not."
<< "\n";
return false;
}
DataType param_type = *param->getDataType();
bool match = false;
switch (arg_type->kind()) {
case c10::TypeKind::IntType:
match = param_type == DataType::Int;
break;
case c10::TypeKind::FloatType:
match = param_type == DataType::Float;
break;
case c10::TypeKind::BoolType:
match = param_type == DataType::Bool;
break;
default:
match = false;
}
if (!match) {
msg << "Argument type is " << *arg_type << ", but the parameter is "
<< param_type << "\n";
}
return match;
}
// Return false if arg and param don't match up and if arg's device (if a
// tensor) doesn't match provided device
bool validateKernelArg(
const c10::IValue& arg,
const Val* param,
const c10::Device& device,
std::stringstream& msg) {
if (arg.isTensor()) {
return validateKernelArgTensor(arg.toTensor(), param, device, msg);
} else {
return validateKernelArgScalar(arg.type(), param, msg);
}
}
} // namespace
void validateKernelInputs(
Fusion* fusion,
const at::ArrayRef<IValue>& inputs,
const c10::Device& device) {
FUSER_PERF_SCOPE("validateKernelInputs");
// This is necessary as we were traversing the fusion graph later in the check
FusionGuard fg(fusion);
// Check inputs
TORCH_INTERNAL_ASSERT(
inputs.size() == fusion->inputs().size(),
"Wrong number of kernel inputs.");
std::stringstream msg;
bool mismatch = false;
for (size_t i = 0; i < inputs.size(); ++i) {
const IValue& arg = inputs[i];
const Val* param = fusion->inputs()[i];
mismatch = !validateKernelArg(arg, param, device, msg) || mismatch;
}
TORCH_INTERNAL_ASSERT(
!mismatch, "Found one or more invalid arguments: ", msg.str());
}
void validateKernelOutputs(
Fusion* fusion,
const std::vector<at::Tensor>& outputs,
const c10::Device& device) {
FUSER_PERF_SCOPE("validateKernelOutputs");
TORCH_INTERNAL_ASSERT(
fusion->outputs().size() != 0,
"Kernel should have at least one output tensor.");
TORCH_INTERNAL_ASSERT(
outputs.size() == fusion->outputs().size(),
"Wrong number of kernel outputs.");
std::stringstream msg;
bool mismatch = false;
for (size_t i = 0; i < outputs.size(); ++i) {
const at::Tensor& arg = outputs[i];
const Val* param = fusion->outputs()[i];
mismatch = !validateKernelArg(arg, param, device, msg) || mismatch;
}
TORCH_INTERNAL_ASSERT(
!mismatch, "Found one or more invalid arguments: ", msg.str());
}
StatefulExpressionEvaluator statefulBindInputs(
const at::ArrayRef<IValue>& aten_inputs,
Fusion* fusion,
GpuLower* lower) {
FUSER_PERF_SCOPE("statefulBindInputs");
TORCH_INTERNAL_ASSERT(
fusion->inputs().size() == aten_inputs.size(),
"Something went wrong configuring launch. Inputs no longer match.");
auto fusion_inputs = fusion->inputs();
StatefulExpressionEvaluator evaluator(fusion);
// This should probably move to EvaluationContext as we may want to bind
// input values frequently. Bind fusion input values to runtime values.
for (size_t i = 0; i < fusion->inputs().size(); i++) {
if (fusion->inputs()[i]->getValType() == ValType::TensorView) {
TensorView* cg_tensor = fusion->inputs()[i]->as<TensorView>();
TORCH_INTERNAL_ASSERT(
aten_inputs[i].isTensor(),
"Something went wrong configuring launch. Inputs no longer match.");
auto aten_tensor = aten_inputs[i].toTensor();
auto root_dom = TensorDomain::noReductions(cg_tensor->getRootDomain());
TORCH_INTERNAL_ASSERT(
aten_tensor.ndimension() == (int64_t)root_dom.size(),
"Something went wrong configuring launch. Inputs no longer match.");
for (size_t dim = 0; dim < root_dom.size(); dim++) {
evaluator.safeBind(
root_dom[dim]->extent(), aten_tensor.sizes()[dim], lower);
}
} else if (
fusion->inputs()[i]->getValType().value() == ValType::Scalar &&
fusion->inputs()[i]->getDataType().value() == DataType::Int) {
TORCH_INTERNAL_ASSERT(
aten_inputs[i].type()->kind() == c10::TypeKind::IntType);
evaluator.safeBind(fusion->inputs()[i], aten_inputs[i].toInt(), lower);
}
}
return evaluator;
}
NvrtcFunction nvrtcCompile(
const std::string& code,
const std::string& func_name,
int id) {
FUSER_PERF_SCOPE("NVRTC");
// lazily construct context if non-existing yet;
CUcontext pctx = nullptr;
AT_CUDA_DRIVER_CHECK(at::globalContext().getNVRTC().cuCtxGetCurrent(&pctx));
if (!pctx) {
std::unique_lock<std::mutex> cudaFreeMutexLock(
*(c10::cuda::CUDACachingAllocator::getFreeMutex()));
cudaFree(nullptr);
}
const auto prop = at::cuda::getCurrentDeviceProperties();
int nvrtc_major, nvrtc_minor;
AT_CUDA_NVRTC_CHECK(
at::globalContext().getNVRTC().nvrtcVersion(&nvrtc_major, &nvrtc_minor));
// Short-circuits if NVRTC version too low
TORCH_INTERNAL_ASSERT(nvrtc_major >= 6);
// Major and minor is determined by device properties and
// possibly "downcompiled" to a lower (compatible) compute architecture
// based on the NVRTC version
const int major = prop->major;
const int minor = prop->minor;
nvrtcProgram program;
{
FUSER_PERF_SCOPE("nvrtcCreateProgram");
AT_CUDA_NVRTC_CHECK(at::globalContext().getNVRTC().nvrtcCreateProgram(
&program, code.c_str(), nullptr, 0, nullptr, nullptr));
}
ResourceGuard holdProgram([&] {
FUSER_PERF_SCOPE("nvrtcDestroyProgram");
AT_CUDA_NVRTC_CHECK(
at::globalContext().getNVRTC().nvrtcDestroyProgram(&program));
});
const std::string compute = "--gpu-architecture=compute_" +
std::to_string(major) + std::to_string(minor);
std::vector<const char*> args = {
"--std=c++14", compute.c_str(), "-default-device"};
const char* disable_fma = getenv("PYTORCH_CUDA_FUSER_DISABLE_FMA");
// int disable_fma_flag = disable_fma ? atoi(disable_fma) : 0;
if (disable_fma && atoi(disable_fma)) {
args.push_back("--fmad=false");
}
const char* ptxas_opt_level = getenv("PYTORCH_CUDA_FUSER_JIT_OPT_LEVEL");
uint32_t jit_opt_level;
std::vector<CUjit_option> options;
std::vector<void*> option_vals;
if (ptxas_opt_level) {
int val = atoi(ptxas_opt_level);
if (val <= 4 && val >= 0) {
jit_opt_level = static_cast<uint32_t>(val);
options.push_back(CU_JIT_OPTIMIZATION_LEVEL);
option_vals.emplace_back(&jit_opt_level);
} else {
TORCH_WARN_ONCE(
"acceptable range for PYTORCH_CUDA_FUSER_JIT_OPT_LEVEL is between 0 and 4, but received ",
jit_opt_level,
", ignoring the option");
}
}
at::globalContext().getNVRTC().nvrtcAddNameExpression(
program, func_name.c_str());
{
FUSER_PERF_SCOPE("nvrtcCompileProgram");
const auto result = at::globalContext().getNVRTC().nvrtcCompileProgram(
program, args.size(), args.data());
if (result != NVRTC_SUCCESS) {
size_t logsize;
at::globalContext().getNVRTC().nvrtcGetProgramLogSize(program, &logsize);
std::vector<char> log(logsize);
at::globalContext().getNVRTC().nvrtcGetProgramLog(program, log.data());
TORCH_INTERNAL_ASSERT(
false, code.c_str(), "\nCUDA NVRTC compile error: ", log.data());
}
AT_CUDA_NVRTC_CHECK(result);
}
const char* lowered_kernel_name = nullptr;
at::globalContext().getNVRTC().nvrtcGetLoweredName(
program, func_name.c_str(), &lowered_kernel_name);
size_t ptx_size = 0;
std::vector<char> ptx;
{
FUSER_PERF_SCOPE("get PTX");
AT_CUDA_NVRTC_CHECK(
at::globalContext().getNVRTC().nvrtcGetPTXSize(program, &ptx_size));
ptx.resize(ptx_size);
AT_CUDA_NVRTC_CHECK(
at::globalContext().getNVRTC().nvrtcGetPTX(program, ptx.data()));
}
NvrtcFunction compiled_kernel_;
// TODO: We do go through different code path, should investigate whether this
// has an impact on generated binary.
const char* prefix_env = getenv("PYTORCH_CUDA_FUSER_CUBIN");
if (prefix_env) {
FUSER_PERF_SCOPE("load CUBIN");
// Output ptx file
std::stringstream ptx_file_name;
ptx_file_name << prefix_env << "_" << id << ".ptx";
std::ofstream myPtxFile(ptx_file_name.str().c_str(), std::ios::out);
if (myPtxFile.is_open()) {
myPtxFile.write(ptx.data(), ptx.size());
myPtxFile.close();
}
CUlinkState linkState;
AT_CUDA_DRIVER_CHECK(at::globalContext().getNVRTC().cuLinkCreate(
0, nullptr, nullptr, &linkState));
AT_CUDA_DRIVER_CHECK(at::globalContext().getNVRTC().cuLinkAddData(
linkState,
CU_JIT_INPUT_PTX,
ptx.data(),
ptx_size,
"compiling PTX",
options.size(),
options.data(),
option_vals.data()));
size_t cubinSize;
void* cubin;
AT_CUDA_DRIVER_CHECK(at::globalContext().getNVRTC().cuLinkComplete(
linkState, &cubin, &cubinSize));
// Output binary file
std::stringstream cubin_file_name;
cubin_file_name << prefix_env << "_" << id << ".cubin";
std::ofstream myCubinFile(
cubin_file_name.str().c_str(), std::ios::out | std::ios::binary);
if (myCubinFile.is_open()) {
myCubinFile.write(static_cast<const char*>(cubin), cubinSize);
myCubinFile.close();
}
// load compiled cubin
AT_CUDA_DRIVER_CHECK(at::globalContext().getNVRTC().cuModuleLoadData(
&(compiled_kernel_.module), cubin));
} else {
FUSER_PERF_SCOPE("load PTX");
// load ptx directly
AT_CUDA_DRIVER_CHECK(at::globalContext().getNVRTC().cuModuleLoadDataEx(
&(compiled_kernel_.module),
ptx.data(),
options.size(),
options.data(),
option_vals.data()));
}
AT_CUDA_DRIVER_CHECK(at::globalContext().getNVRTC().cuModuleGetFunction(
&(compiled_kernel_.function),
compiled_kernel_.module,
lowered_kernel_name));
return compiled_kernel_;
}
} // namespace executor_utils
} // namespace cuda
} // namespace fuser
} // namespace jit
} // namespace torch