Skip to content

Commit 956b0c5

Browse files
narendasannarendasan
committed
feat(//core/conversion/converters/impl/reduce): adds the rest of TRT's
basic reduction operators Signed-off-by: Naren Dasan <naren@narendasan.com> Signed-off-by: Naren Dasan <narens@nvidia.com>
1 parent 391af52 commit 956b0c5

File tree

7 files changed

+248
-156
lines changed

7 files changed

+248
-156
lines changed

‎.gitignore‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -13,4 +13,4 @@ experiments/
1313
py/build/
1414
py/tmp/
1515
py/.eggs
16-
.vscode/
16+
.vscode/

‎core/conversion/converters/impl/reduce.cpp‎

Lines changed: 130 additions & 62 deletions
Original file line numberDiff line numberDiff line change
@@ -8,7 +8,10 @@ namespace conversion {
88
namespace converters {
99
namespace impl {
1010
namespace {
11-
auto reduced_registrations = RegisterNodeConversionPatterns()
11+
12+
13+
14+
auto reduce_registrations = RegisterNodeConversionPatterns()
1215
.pattern({
1316
"aten::mean(Tensor self, *, ScalarType? dtype=None) -> (Tensor)",
1417
[](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
@@ -36,7 +39,7 @@ auto reduced_registrations = RegisterNodeConversionPatterns()
3639
LOG_DEBUG("Dim to reduce:" << util::toDims(dims)); // Some abuse of toDim but just for debug info
3740

3841
uint32_t axis_mask = 0;
39-
for (int d = 0; d < dims.size(); d++) {
42+
for (size_t d = 0; d < dims.size(); d++) {
4043
axis_mask |= 1 << dims[d];
4144
}
4245
LOG_DEBUG("Axis Mask" << std::bitset<32>(axis_mask));
@@ -52,6 +55,131 @@ auto reduced_registrations = RegisterNodeConversionPatterns()
5255
mean_layer->setName(util::node_info(n).c_str());
5356
auto out_tensor = ctx->AssociateValueAndTensor(n->outputs()[0], mean_layer->getOutput(0));
5457

58+
LOG_DEBUG("Output shape: " << out_tensor->getDimensions());
59+
return true;
60+
}
61+
}).pattern({
62+
"aten::sum(Tensor self, *, ScalarType? dtype=None) -> Tensor",
63+
[](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
64+
auto in_tensor = args[0].ITensor();
65+
auto in_dims = util::toVec(in_tensor->getDimensions());
66+
LOG_WARNING("Sum Converter disregards dtype");
67+
68+
uint32_t axis_mask = (uint32_t)(((uint64_t)1 << in_dims.size()) - 1);
69+
70+
auto sum_layer = ctx->net->addReduce(*in_tensor, nvinfer1::ReduceOperation::kSUM, axis_mask, false);
71+
72+
TRTORCH_CHECK(sum_layer, "Unable to create sum layer from node: " << *n);
73+
74+
sum_layer->setName(util::node_info(n).c_str());
75+
auto out_tensor = ctx->AssociateValueAndTensor(n->outputs()[0], sum_layer->getOutput(0));
76+
77+
LOG_DEBUG("Output shape: " << out_tensor->getDimensions());
78+
return true;
79+
}
80+
}).pattern({
81+
"aten::sum.dim_IntList(Tensor self, int[1] dim, bool keepdim=False, *, ScalarType? dtype=None) -> Tensor",
82+
[](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
83+
auto in_tensor = args[0].ITensor();
84+
auto dims = args[1].unwrapToIntList();
85+
LOG_DEBUG("Dim to reduce:" << util::toDims(dims)); // Some abuse of toDim but just for debug info
86+
87+
uint32_t axis_mask = 0;
88+
for (size_t d = 0; d < dims.size(); d++) {
89+
axis_mask |= 1 << dims[d];
90+
}
91+
LOG_DEBUG("Axis Mask" << std::bitset<32>(axis_mask));
92+
93+
auto keepdim = args[2].unwrapToBool();
94+
LOG_DEBUG("Keep dims :" << keepdim);
95+
96+
LOG_WARNING("Sum converter disregards dtype");
97+
auto sum_layer = ctx->net->addReduce(*in_tensor, nvinfer1::ReduceOperation::kSUM, axis_mask, keepdim);
98+
99+
TRTORCH_CHECK(sum_layer, "Unable to create sum layer from node: " << *n);
100+
101+
sum_layer->setName(util::node_info(n).c_str());
102+
auto out_tensor = ctx->AssociateValueAndTensor(n->outputs()[0], sum_layer->getOutput(0));
103+
104+
LOG_DEBUG("Output shape: " << out_tensor->getDimensions());
105+
return true;
106+
}
107+
}).pattern({
108+
"aten::prod(Tensor self, *, ScalarType? dtype=None) -> Tensor",
109+
[](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
110+
auto in_tensor = args[0].ITensor();
111+
auto in_dims = util::toVec(in_tensor->getDimensions());
112+
LOG_WARNING("Prod Converter disregards dtype");
113+
114+
uint32_t axis_mask = (uint32_t)(((uint64_t)1 << in_dims.size()) - 1);
115+
116+
auto prod_layer = ctx->net->addReduce(*in_tensor, nvinfer1::ReduceOperation::kPROD, axis_mask, false);
117+
118+
TRTORCH_CHECK(prod_layer, "Unable to create sum layer from node: " << *n);
119+
120+
prod_layer->setName(util::node_info(n).c_str());
121+
auto out_tensor = ctx->AssociateValueAndTensor(n->outputs()[0], prod_layer->getOutput(0));
122+
123+
LOG_DEBUG("Output shape: " << out_tensor->getDimensions());
124+
return true;
125+
}
126+
}).pattern({
127+
"aten::prod.dim_int(Tensor self, int dim, bool keepdim=False, *, ScalarType? dtype=None) -> Tensor",
128+
[](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
129+
auto in_tensor = args[0].ITensor();
130+
auto dim = args[1].unwrapToInt();
131+
LOG_DEBUG("Dim to reduce:" << dim); // Some abuse of toDim but just for debug info
132+
133+
uint32_t axis_mask = 1 << dim;
134+
LOG_DEBUG("Axis Mask" << std::bitset<32>(axis_mask));
135+
136+
auto keepdim = args[2].unwrapToBool();
137+
LOG_DEBUG("Keep dims :" << keepdim);
138+
139+
LOG_WARNING("Prod converter disregards dtype");
140+
auto prod_layer = ctx->net->addReduce(*in_tensor, nvinfer1::ReduceOperation::kPROD, axis_mask, keepdim);
141+
142+
TRTORCH_CHECK(prod_layer, "Unable to create mean layer from node: " << *n);
143+
144+
prod_layer->setName(util::node_info(n).c_str());
145+
auto out_tensor = ctx->AssociateValueAndTensor(n->outputs()[0], prod_layer->getOutput(0));
146+
147+
LOG_DEBUG("Output shape: " << out_tensor->getDimensions());
148+
return true;
149+
}
150+
}).pattern({
151+
"aten::max(Tensor self) -> Tensor",
152+
[](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
153+
auto in_tensor = args[0].ITensor();
154+
auto in_dims = util::toVec(in_tensor->getDimensions());
155+
156+
uint32_t axis_mask = (uint32_t)(((uint64_t)1 << in_dims.size()) - 1);
157+
158+
auto max_layer = ctx->net->addReduce(*in_tensor, nvinfer1::ReduceOperation::kMAX, axis_mask, false);
159+
160+
TRTORCH_CHECK(max_layer, "Unable to create max layer from node: " << *n);
161+
162+
max_layer->setName(util::node_info(n).c_str());
163+
auto out_tensor = ctx->AssociateValueAndTensor(n->outputs()[0], max_layer->getOutput(0));
164+
165+
LOG_DEBUG("Output shape: " << out_tensor->getDimensions());
166+
return true;
167+
}
168+
}).pattern({
169+
"aten::min(Tensor self) -> Tensor",
170+
[](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
171+
auto in_tensor = args[0].ITensor();
172+
auto in_dims = util::toVec(in_tensor->getDimensions());
173+
174+
uint32_t axis_mask = (uint32_t)(((uint64_t)1 << in_dims.size()) - 1);
175+
176+
auto min_layer = ctx->net->addReduce(*in_tensor, nvinfer1::ReduceOperation::kMIN, axis_mask, false);
177+
178+
TRTORCH_CHECK(min_layer, "Unable to create min layer from node: " << *n);
179+
180+
min_layer->setName(util::node_info(n).c_str());
181+
auto out_tensor = ctx->AssociateValueAndTensor(n->outputs()[0], min_layer->getOutput(0));
182+
55183
LOG_DEBUG("Output shape: " << out_tensor->getDimensions());
56184
return true;
57185
}
@@ -62,63 +190,3 @@ auto reduced_registrations = RegisterNodeConversionPatterns()
62190
} // namespace conversion
63191
} // namespace core
64192
} // namespace trtorch
65-
66-
// #include "core/util/prelude.h"
67-
// #include "core/conversion/converters/converters.h"
68-
69-
// namespace trtorch {
70-
// namespace core {
71-
// namespace conversion {
72-
// namespace converters {
73-
// namespace impl {
74-
// namespace {
75-
76-
// #define convert(unary, trt_type) \
77-
// auto unary##_registrations TRTORCH_UNUSED = \
78-
// RegisterNodeConversionPatterns().pattern( \
79-
// {"aten::" #unary "(Tensor self) -> Tensor", \
80-
// [](ConversionCtx *ctx, const torch::jit::Node *n, \
81-
// args &args) -> bool { \
82-
// auto in = args[0].ITensor(); \
83-
// auto unary = \
84-
// ctx->net->addUnary(*in, nvinfer1::UnaryOperation::trt_type); \
85-
// \
86-
// TRTORCH_CHECK( \
87-
// unary, \
88-
// "Unable to create " #unary " layer from node: " << *n); \
89-
// \
90-
// unary->setName(util::node_info(n).c_str()); \
91-
// auto out_tensor = ctx->AssociateValueAndTensor( \
92-
// n->outputs()[0], \
93-
// unary->getOutput(0)); \
94-
// LOG_DEBUG( \
95-
// "Output tensor shape: " << out_tensor->getDimensions()); \
96-
// \
97-
// return true; \
98-
// }});
99-
100-
// convert(cos, kCOS);
101-
// convert(acos, kACOS);
102-
// convert(cosh, kCOSH);
103-
// convert(sin, kSIN);
104-
// convert(asin, kASIN);
105-
// convert(sinh, kSINH);
106-
// convert(tan, kTAN);
107-
// convert(atan, kATAN);
108-
// convert(abs, kABS);
109-
// convert(floor, kFLOOR);
110-
// convert(reciprocal, kRECIP);
111-
// convert(log, kLOG);
112-
// convert(ceil, kCEIL);
113-
// convert(sqrt, kSQRT);
114-
// convert(exp, kEXP);
115-
// convert(neg, kNEG);
116-
117-
// #undef convert
118-
119-
// } // namespace
120-
// } // namespace impl
121-
// } // namespace converters
122-
// } // namespace conversion
123-
// } // namespace core
124-
// } // namespace trtorch

‎tests/core/converters/test_linear.cpp‎

Lines changed: 0 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -4,8 +4,6 @@
44
#include "tests/util/util.h"
55
#include "core/compiler.h"
66

7-
#define LAYER_TEST //used to set threshold for diff
8-
97
TEST(Converters, ATenLinearNoBiasConvertsCorrectly) {
108
const auto graph = R"IR(
119
graph(%0 : Tensor,

‎tests/core/converters/test_pooling.cpp‎

Lines changed: 0 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -4,8 +4,6 @@
44
#include "tests/util/util.h"
55
#include "core/compiler.h"
66

7-
#define LAYER_TEST //used to set threshold for diff
8-
97
TEST(Converters, ATenMaxPool2DConvertsCorrectly) {
108
const auto graph = R"IR(
119
graph(%0 : Tensor):

0 commit comments

Comments
 (0)