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
|
#include <torch/csrc/jit/codegen/cuda/executor.h>
#include <torch/csrc/jit/codegen/cuda/fusion.h>
#include <torch/csrc/jit/codegen/cuda/ir_all_nodes.h>
#include <torch/csrc/jit/codegen/cuda/ir_builder.h>
#include <torch/csrc/jit/codegen/cuda/ir_utils.h>
#include <torch/csrc/jit/codegen/cuda/lower2device.h>
#include <torch/csrc/jit/codegen/cuda/ops/all_ops.h>
#include <torch/csrc/jit/codegen/cuda/scheduler/all_schedulers.h>
#include <benchmark/benchmark.h>
#include <cuda_runtime.h>
#include <benchmarks/cpp/nvfuser/utils.h>
using namespace torch::jit::fuser::cuda;
static auto getLayerBackwardNormRuntime(
std::unique_ptr<Fusion> fusion_ptr,
std::unique_ptr<FusionExecutorCache>& fec,
std::vector<at::IValue>& aten_inputs,
std::vector<int64_t>& shape,
std::vector<int64_t>& norm_shape) {
Fusion& fusion = *fusion_ptr.get();
const size_t kM = shape.size();
const size_t kN = norm_shape.size();
const size_t kOuterNumDims = kM - kN;
std::vector<int64_t> outer_shape;
for (size_t idx = 0; idx < kOuterNumDims; ++idx) {
outer_shape.push_back(shape[idx]);
}
for (size_t idx = kOuterNumDims; idx < kM; ++idx) {
outer_shape.push_back(1);
}
auto grad_out = makeSymbolicTensor(shape.size());
auto input = makeSymbolicTensor(shape.size());
auto mean = makeConcreteTensor(outer_shape);
auto rstd = makeConcreteTensor(outer_shape);
auto weight = makeSymbolicTensor(norm_shape.size());
auto bias = makeSymbolicTensor(norm_shape.size());
fusion.addInput(grad_out);
fusion.addInput(input);
fusion.addInput(mean);
fusion.addInput(rstd);
fusion.addInput(weight);
fusion.addInput(bias);
auto grads = layer_norm_backward(
grad_out,
input,
norm_shape,
mean,
rstd,
weight,
bias,
{true, true, true});
fusion.addOutput(grads.grad_input);
fusion.addOutput(grads.grad_weight);
fusion.addOutput(grads.grad_bias);
auto options = at::TensorOptions().dtype(at::kFloat).device(at::kCUDA, 0);
at::Tensor aten_grad_out = at::randn(shape, options);
at::Tensor aten_input = at::randn(shape, options);
at::Tensor aten_weight = at::randn(norm_shape, options);
at::Tensor aten_bias = at::randn(norm_shape, options);
auto at_weight = c10::optional<at::Tensor>(aten_weight);
auto at_bias = c10::optional<at::Tensor>(aten_bias);
const float kEps = 1e-5;
auto aten_results =
at::native_layer_norm(aten_input, norm_shape, at_weight, at_bias, kEps);
auto aten_output = std::get<0>(aten_results);
auto aten_mean = std::get<1>(aten_results);
auto aten_rstd = std::get<2>(aten_results);
fec = std::make_unique<FusionExecutorCache>(std::move(fusion_ptr));
aten_inputs = {
aten_grad_out, aten_input, aten_mean, aten_rstd, aten_weight, aten_bias};
auto cg_outputs = fec->runFusionWithInputs(aten_inputs);
return fec->getMostRecentKernelRuntime();
}
void LayerNormBackward_ShapeInference_Base(
benchmark::State& benchmark_state,
bool disable_launch_parameter_cache) {
std::unique_ptr<Fusion> fusion_ptr = std::make_unique<Fusion>();
FusionGuard fg(fusion_ptr.get());
// PreAllocate
std::unique_ptr<FusionExecutorCache> fec;
std::vector<at::IValue> aten_inputs;
std::vector<int64_t> shape{20, 100, 35, 67};
std::vector<int64_t> norm_shape{67};
auto runtime = getLayerBackwardNormRuntime(
std::move(fusion_ptr), fec, aten_inputs, shape, norm_shape);
KernelArgumentHolder args = KernelArgumentHolder::createKernelArgumentHolder(aten_inputs);
TORCH_INTERNAL_ASSERT(
runtime->getMaybeHeuristicsFor(args).has_value());
fec->profile(true);
fec->disableKernelLaunch();
fec->runFusionWithInputs(aten_inputs);
if (disable_launch_parameter_cache) {
fec->disableLaunchParamCache();
}
for (auto _ : benchmark_state) {
// Setup (not included in the measurement)
fec->runFusionWithInputs(aten_inputs);
}
}
static void LayerNormBackward_ShapeInference(
benchmark::State& benchmark_state) {
LayerNormBackward_ShapeInference_Base(benchmark_state, true);
}
static void LayerNormBackward_NoShapeInferenceCachedBaseline(
benchmark::State& benchmark_state) {
LayerNormBackward_ShapeInference_Base(benchmark_state, false);
}
static auto getLayerForwardNormRuntime(
std::unique_ptr<Fusion> fusion_ptr,
std::unique_ptr<FusionExecutorCache>& fec,
std::vector<at::IValue>& aten_inputs,
std::vector<int64_t>& shape,
std::vector<int64_t>& norm_shape) {
Fusion& fusion = *fusion_ptr.get();
const float kEps = 1e-5;
Double* eps_ptr = IrBuilder::create<Double>(kEps);
auto input = makeSymbolicTensor(shape.size());
fusion.addInput(input);
auto result = layer_norm(input, norm_shape, nullptr, nullptr, eps_ptr);
fusion.addOutput(result.output);
fusion.addOutput(result.mean);
fusion.addOutput(result.invstd);
auto options = at::TensorOptions().dtype(at::kFloat).device(at::kCUDA, 0);
at::Tensor aten_input = at::randn(shape, options);
fec = std::make_unique<FusionExecutorCache>(std::move(fusion_ptr));
aten_inputs = {aten_input};
auto cg_outputs = fec->runFusionWithInputs(aten_inputs);
return fec->getMostRecentKernelRuntime();
}
void LayerNormForward_ShapeInferenceBase(
benchmark::State& benchmark_state,
bool disable_launch_param_cache) {
std::unique_ptr<Fusion> fusion_ptr = std::make_unique<Fusion>();
FusionGuard fg(fusion_ptr.get());
// PreAllocate
std::unique_ptr<FusionExecutorCache> fec;
std::vector<at::IValue> aten_inputs;
std::vector<int64_t> shape{20, 100, 35, 67};
std::vector<int64_t> norm_shape{67};
auto runtime = getLayerForwardNormRuntime(
std::move(fusion_ptr), fec, aten_inputs, shape, norm_shape);
KernelArgumentHolder args = KernelArgumentHolder::createKernelArgumentHolder(aten_inputs);
TORCH_INTERNAL_ASSERT(
runtime->getMaybeHeuristicsFor(args).has_value());
fec->profile(true);
fec->disableKernelLaunch();
fec->runFusionWithInputs(aten_inputs);
if (disable_launch_param_cache) {
fec->disableLaunchParamCache();
}
for (auto _ : benchmark_state) {
// Setup (not included in the measurement)
fec->runFusionWithInputs(aten_inputs);
}
}
static void LayerNormForward_NoShapeInferenceCachedBaseline(
benchmark::State& benchmark_state) {
LayerNormForward_ShapeInferenceBase(benchmark_state, false);
}
static void LayerNormForward_ShapeInference(benchmark::State& benchmark_state) {
LayerNormForward_ShapeInferenceBase(benchmark_state, true);
}
BENCHMARK(LayerNormBackward_ShapeInference)->Unit(benchmark::kMicrosecond);
BENCHMARK(LayerNormForward_ShapeInference)->Unit(benchmark::kMicrosecond);
BENCHMARK(LayerNormBackward_NoShapeInferenceCachedBaseline)
->Unit(benchmark::kMicrosecond);
BENCHMARK(LayerNormForward_NoShapeInferenceCachedBaseline)
->Unit(benchmark::kMicrosecond);
|
