Movatterモバイル変換

[0]ホーム
URL:

Skip to content

Navigation Menu

Sign in
Appearance settings

Search code, repositories, users, issues, pull requests...

Provide feedback

We read every piece of feedback, and take your input very seriously.

Saved searches

Use saved searches to filter your results more quickly

 Sign up
Appearance settings

Commitd0d6b1f

Browse files
larryliu0820pytorchmergebot
authored andcommitted
[torchgen] Generate out variant for functional operator (#81437)
Summary:Previously we don't generate out variant (both schema and kernel) for an operator with functional variant only. This adds support for that and adds test.## Changes on `native_function_generation.py`We are generating out variant for all functional variants if possible. This PR introduces a lot of newly generated out variants and `native_functions.yaml` needs to incorporate the changes by adding `autogen` keywords.The logic for determining what operators we should generate an out variant for is the following:1. No existing out variant for this `NativeFunction`2. Contains an existing in place, mutable or functional variant3. Contains at least 1 tensor like return(s)For operators matching the first two conditions but failing the third, I listed them in `FUNCTIONAL_OPS_THAT_CANNOT_GET_AN_OUT_VARIANT`.## Special handlingThe following operators satisfy all 3 criteria above but we chose to not autogen them, with some reasons.* `mkldnn_adaptive_avg_pool2d`, the generated out variant `mkldnn_adaptive_avg_pool2d.out` is colliding with the `mkldnn_adaptive_avg_pool2d_out` kernel in `adaptive_avg_pool2d.out` operator. I manually created `mkldnn_adaptive_avg_pool2d.out` and renamed `mkldnn_adaptive_avg_pool2d_out` to `mkldnn_adaptive_avg_pool2d_out_stub`.* `min`, `max` and `mean`. There already exist `min.out`, `max.out` and `mean.out` but they are having different semantics with the functional ones. I manually created `min.unary_out`, `max.unary_out` and `mean.dtype_out` to disambiguate.## Autograd ChangesWe introduced a logic to not match derivatives info in `derivatives.yaml` to out variant, since we are generating `NOT_IMPLEMENTED` kernels for those out variants anyway. The issue we are seeing with the original logic is that it doesn't handle `TensorOption` arguments really well. For example we have these two operators:* `_to_copy(Tensor self, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None, bool non_blocking=False, MemoryFormat? memory_format=None) -> Tensor`* `_to_copy.out(Tensor self, *, bool non_blocking=False, MemoryFormat? memory_format=None, Tensor(a!) out) -> Tensor(a!)`If we uses `_to_copy` derivative info, there will be compilation error since `dtype` is missing from `_to_copy.out` signature.Test Plan: Rely on unit testDifferential Revision: D37832342Pull Requestresolved:#81437Approved by:https://github.com/iseeyuan,https://github.com/bdhirsh
1 parentbb1e3d8 commitd0d6b1f

File tree

11 files changed

+647
-56
lines changed

11 files changed

+647
-56
lines changed

‎aten/src/ATen/native/RangeFactories.cpp‎

Lines changed: 4 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -142,6 +142,10 @@ Tensor& range_out(const Scalar& start, const Scalar& end, const Scalar& step, Te
142142
return result;
143143
}
144144

145+
Tensor&range_out_no_step(const Scalar& start,const Scalar& end, Tensor& result) {
146+
returnrange_out(start, end,/*step =*/1, result);
147+
}
148+
145149
Tensor&arange_out(const Scalar& start,const Scalar& end,const Scalar& step, Tensor& result) {
146150
AT_DISPATCH_ALL_TYPES_AND2(kHalf,kBFloat16, result.scalar_type(),"arange_cpu", [&]() {
147151
usingaccscalar_t = at::acc_type<scalar_t,false>;

‎aten/src/ATen/native/ReduceAllOps.cpp‎

Lines changed: 15 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1,4 +1,5 @@
11
#include<ATen/native/ReduceAllOps.h>
2+
#include<ATen/native/Resize.h>
23

34
#include<ATen/ATen.h>
45
#include<ATen/NativeFunctions.h>
@@ -17,6 +18,13 @@ Tensor min(const Tensor &self) {
1718
return result;
1819
}
1920

21+
Tensor&min_unary_out(const Tensor &self, Tensor& out) {
22+
Tensor tmp_output =at::min(self);
23+
at::native::resize_output(out, tmp_output.sizes());
24+
out.copy_(tmp_output);
25+
return out;
26+
}
27+
2028
Tensormax(const Tensor &self) {
2129
TORCH_CHECK(self.numel() >0,
2230
"max(): Expected reduction dim to be specified for input.numel() == 0. Specify the reduction dim with the 'dim' argument.");
@@ -25,6 +33,13 @@ Tensor max(const Tensor &self) {
2533
return result;
2634
}
2735

36+
Tensor&max_unary_out(const Tensor &self, Tensor& out) {
37+
Tensor tmp_output =at::max(self);
38+
at::native::resize_output(out, tmp_output.sizes());
39+
out.copy_(tmp_output);
40+
return out;
41+
}
42+
2843
// DEPRECATED: Use at::aminmax instead
2944
std::tuple<Tensor, Tensor>_aminmax_all(const Tensor &self) {
3045
TORCH_WARN_ONCE("_aminmax is deprecated as of PyTorch 1.11 and will be removed in a future release. Use aminmax instead."

‎aten/src/ATen/native/TensorFactories.cpp‎

Lines changed: 5 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -101,8 +101,12 @@ Tensor arange(
101101
returnat::arange_out(result, start, end, step);
102102
}
103103

104+
Tensor&arange_start_out(const Scalar& start,const Scalar& end, Tensor& result) {
105+
returnat::arange_out(result, start, end,/*step=*/1);
106+
}
107+
104108
Tensor&arange_out(const Scalar& end, Tensor& result) {
105-
returnat::arange_out(result,/*start=*/0, end);
109+
returnat::arange_out(result,/*start=*/0, end,/*step=*/1);
106110
}
107111

108112
Tensor&arange_out(Tensor& result,const Scalar& start,const Scalar& end) {

‎aten/src/ATen/native/mkldnn/Pooling.cpp‎

Lines changed: 17 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -2,6 +2,7 @@
22
#include<ATen/Config.h>
33
#include<ATen/NativeFunctions.h>
44
#include<ATen/core/grad_mode.h>
5+
#include<ATen/native/Resize.h>
56
#include<ATen/native/utils/ParamUtils.h>
67
#include<c10/util/irange.h>
78
#include<tuple>
@@ -80,6 +81,12 @@ Tensor mkldnn_adaptive_avg_pool2d(Tensor const& input, IntArrayRef output_size)
8081
TORCH_CHECK(false,"mkldnn_adaptive_avg_pool2d: ATen not compiled with MKLDNN support");
8182
}
8283

84+
Tensor&mkldnn_adaptive_avg_pool2d_out_stub(const Tensor& input,
85+
IntArrayRef output_size,
86+
Tensor& output) {
87+
TORCH_CHECK(false,"mkldnn_adaptive_avg_pool2d_out_stub: ATen not compiled with MKLDNN support");
88+
}
89+
8390
Tensor&mkldnn_adaptive_avg_pool2d_out(const Tensor& input,
8491
IntArrayRef output_size,
8592
Tensor& output) {
@@ -498,10 +505,19 @@ Tensor mkldnn_adaptive_avg_pool2d(
498505
/*algo*/ ideep::algorithm::pooling_avg);
499506
}
500507

508+
Tensor&mkldnn_adaptive_avg_pool2d_out_stub(const Tensor& input,
509+
IntArrayRef output_size,
510+
Tensor& output) {
511+
TORCH_CHECK(false,"mkldnn_adaptive_avg_pool2d_out_stub: in-place mkldnn operations are not supported yet");
512+
}
513+
501514
Tensor&mkldnn_adaptive_avg_pool2d_out(const Tensor& input,
502515
IntArrayRef output_size,
503516
Tensor& output) {
504-
TORCH_CHECK(false,"mkldnn_adaptive_avg_pool2d_out: in-place mkldnn operations are not supported yet");
517+
auto tmp_output =at::native::mkldnn_adaptive_avg_pool2d(input, output_size);
518+
at::native::resize_output(output, tmp_output.sizes());
519+
output.copy_(tmp_output);
520+
return output;
505521
}
506522

507523
Tensormkldnn_max_pool2d_backward(

0 commit comments

Comments
 (0)
[8]ページ先頭
©2009-2025 Movatter.jp