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
|
#include <ATen/core/jit_type.h>
#ifdef USE_VULKAN
#include <ATen/native/vulkan/VulkanOpContext.h>
#endif
#include <torch/csrc/jit/ir/ir.h>
#include <torch/csrc/jit/ir/subgraph_matcher.h>
#include <torch/csrc/jit/passes/constant_pooling.h>
#include <torch/csrc/jit/passes/fold_conv_bn.h>
#include <torch/csrc/jit/passes/freeze_module.h>
#include <torch/csrc/jit/passes/fuse_linear.h>
#include <torch/csrc/jit/passes/graph_rewrite_helper.h>
#include <torch/csrc/jit/passes/prepack_folding.h>
#include <torch/csrc/jit/passes/remove_dropout.h>
#include <torch/csrc/jit/passes/subgraph_rewrite.h>
#include <torch/csrc/jit/passes/vulkan_rewrite.h>
namespace torch {
namespace jit {
#ifdef USE_VULKAN
namespace {
void insertPrePackedConv2dOp(std::shared_ptr<Graph>& graph) {
graph_rewrite_helper::replaceConvolutionWithAtenConv(graph);
std::string conv_2d_pattern = R"(
graph(%input, %weight, %bias, %stride:int[], %padding:int[], %dilation:int[], %groups:int):
%r = aten::conv2d(%input, %weight, %bias, %stride, %padding, %dilation, %groups)
return (%r) )";
std::string prepacked_ops_conv2d_pattern = R"(
graph(%input, %weight, %bias, %stride:int[], %padding:int[], %dilation:int[], %groups:int):
%output_min_max : None = prim::Constant()
%packed_weight_bias = vulkan_prepack::conv2d_clamp_prepack(
%weight, %bias, %stride, %padding, %dilation, %groups,
%output_min_max, %output_min_max)
%r = vulkan_prepack::conv2d_clamp_run(%input, %packed_weight_bias)
return (%r) )";
SubgraphRewriter rewriter;
rewriter.RegisterRewritePattern(
conv_2d_pattern, prepacked_ops_conv2d_pattern);
rewriter.runOnGraph(graph);
}
void fuseHardtanhWithPackedOps(std::shared_ptr<Graph>& graph) {
SubgraphRewriter rewriter;
std::string conv2d_prepack_run_hardtanh_fused = R"(
graph(%input, %weight, %bias, %stride:int[], %padding:int[],
%dilation:int[], %groups:int, %output_min, %output_max, %dummy_min_max):
%packed_weight_bias : __torch__.torch.classes.vulkan.Conv2dOpContext = vulkan_prepack::conv2d_clamp_prepack(
%weight, %bias, %stride, %padding, %dilation, %groups,
%output_min, %output_max)
%r = vulkan_prepack::conv2d_clamp_run(%input, %packed_weight_bias)
return (%r) )";
std::string conv2d_prepack_run_hardtanh = R"(
graph(%input, %weight, %bias, %stride:int[], %padding:int[],
%dilation:int[], %groups:int, %output_min, %output_max, %dummy_min_max):
%packed_weight_bias = vulkan_prepack::conv2d_clamp_prepack(
%weight, %bias, %stride, %padding, %dilation, %groups,
%dummy_min_max, %dummy_min_max)
%conv2d_res = vulkan_prepack::conv2d_clamp_run(%input, %packed_weight_bias)
%r = aten::hardtanh(%conv2d_res, %output_min, %output_max)
return (%r) )";
rewriter.RegisterRewritePattern(
conv2d_prepack_run_hardtanh, conv2d_prepack_run_hardtanh_fused);
std::string conv2d_prepack_run_hardtanh_inplace = R"(
graph(%input, %weight, %bias, %stride:int[], %padding:int[],
%dilation:int[], %groups:int, %output_min, %output_max, %dummy_min_max):
%packed_weight_bias = vulkan_prepack::conv2d_clamp_prepack(
%weight, %bias, %stride, %padding, %dilation, %groups,
%dummy_min_max, %dummy_min_max)
%conv2d_res = vulkan_prepack::conv2d_clamp_run(%input, %packed_weight_bias)
%r = aten::hardtanh_(%conv2d_res, %output_min, %output_max)
return (%r) )";
rewriter.RegisterRewritePattern(
conv2d_prepack_run_hardtanh_inplace, conv2d_prepack_run_hardtanh_fused);
rewriter.runOnGraph(graph, torch::jit::graph_rewrite_helper::isClampFusable);
}
void fuseReluWithPackedOps(std::shared_ptr<Graph>& graph) {
SubgraphRewriter rewriter;
std::string conv2d_prepack_run_relu_fused = R"(
graph(%input, %weight, %bias, %stride:int[], %padding:int[],
%dilation:int[], %groups:int, %dummy_min_max):
%output_min: float = prim::Constant[value=0.0]()
%output_max: None = prim::Constant()
%packed_weight_bias : __torch__.torch.classes.vulkan.Conv2dOpContext = vulkan_prepack::conv2d_clamp_prepack(
%weight, %bias, %stride, %padding, %dilation, %groups,
%output_min, %output_max)
%r = vulkan_prepack::conv2d_clamp_run(%input, %packed_weight_bias)
return (%r) )";
std::string conv2d_prepack_run_relu = R"(
graph(%input, %weight, %bias, %stride:int[], %padding:int[],
%dilation:int[], %groups:int, %dummy_min_max):
%packed_weight_bias = vulkan_prepack::conv2d_clamp_prepack(
%weight, %bias, %stride, %padding, %dilation, %groups,
%dummy_min_max, %dummy_min_max)
%conv2d_res = vulkan_prepack::conv2d_clamp_run(%input, %packed_weight_bias)
%r = aten::relu(%conv2d_res)
return (%r) )";
rewriter.RegisterRewritePattern(
conv2d_prepack_run_relu, conv2d_prepack_run_relu_fused);
std::string conv2d_prepack_run_relu_inplace = R"(
graph(%input, %weight, %bias, %stride:int[], %padding:int[],
%dilation:int[], %groups:int, %dummy_min_max):
%packed_weight_bias = vulkan_prepack::conv2d_clamp_prepack(
%weight, %bias, %stride, %padding, %dilation, %groups,
%dummy_min_max, %dummy_min_max)
%conv2d_res = vulkan_prepack::conv2d_clamp_run(%input, %packed_weight_bias)
%r = aten::relu_(%conv2d_res)
return (%r) )";
rewriter.RegisterRewritePattern(
conv2d_prepack_run_relu_inplace, conv2d_prepack_run_relu_fused);
rewriter.runOnGraph(graph, torch::jit::graph_rewrite_helper::isClampFusable);
}
} // namespace
void vulkanInsertPrePackedOps(std::shared_ptr<Graph>& graph) {
insertPrePackedConv2dOp(graph);
}
void vulkanInsertPrePackedOps(script::Module& module) {
for (auto& method : module.get_methods()) {
auto graph = method.graph();
vulkanInsertPrePackedOps(graph);
}
for (script::Module m : module.children()) {
vulkanInsertPrePackedOps(m);
}
}
void vulkanFusePrePackedConvWithClamp(script::Module& module) {
auto graph = module.get_method("forward").graph();
fuseReluWithPackedOps(graph);
fuseHardtanhWithPackedOps(graph);
}
void vulkanFoldPrePackingOps(script::Module& m) {
PrePackingOpsFilterFn filter_fn = [](const Node* n) -> bool {
return (
n->kind() ==
Symbol::fromQualString("vulkan_prepack::conv2d_clamp_prepack"));
};
PrePackingOpsFolder(m, filter_fn, "prepack_folding");
}
script::Module vulkanOptimizeForMobile(
const script::Module& m,
const std::vector<std::string>& preserved_methods) {
auto cloned_module = m.clone();
cloned_module.eval();
cloned_module = FoldConvBatchNorm(cloned_module);
vulkanInsertPrePackedOps(cloned_module);
cloned_module = freeze_module(cloned_module, preserved_methods);
vulkanFusePrePackedConvWithClamp(cloned_module);
vulkanFoldPrePackingOps(cloned_module);
removeDropout(cloned_module);
return cloned_module;
}
#else
void vulkanInsertPrePackedOps(std::shared_ptr<Graph>& graph) {
TORCH_INTERNAL_ASSERT(
"Vulkan is not enabled. Please build with USE_VULKAN=1");
}
void vulkanInsertPrePackedOps(script::Module& module) {
TORCH_INTERNAL_ASSERT(
"Vulkan is not enabled. Please build with USE_VULKAN=1");
}
void vulkanFusePrePackedConvWithClamp(script::Module& module) {
TORCH_INTERNAL_ASSERT(
"Vulkan is not enabled. Please build with USE_VULKAN=1");
}
void vulkanFoldPrePackingOps(script::Module& m) {
TORCH_INTERNAL_ASSERT(
"Vulkan is not enabled. Please build with USE_VULKAN=1");
}
script::Module vulkanOptimizeForMobile(
const script::Module& module,
const std::vector<std::string>& preserved_methods) {
TORCH_INTERNAL_ASSERT(
"Mobile optimizaiton only available with Vulkan at the moment. "
"Vulkan is not enabled. Please build with USE_VULKAN=1");
return module;
}
#endif
} // namespace jit
} // namespace torch
|
