This repository has been archived by the owner on Nov 17, 2023. It is now read-only.
[Bugfix] [Numpy] Add kAddTo and kNullOp to Transpose
#16979
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -269,8 +269,10 @@ struct TransposeParam : public dmlc::Parameter<TransposeParam> { | |
| * \param out output tensor | ||
| * \param row shape of dim 0 of input | ||
| * \param col shape of dim 1 of input | ||
| * \tparam DType Data type | ||
| * \tparam is_addto | ||
| */ | ||
| template<typename DType, bool is_addto> | ||
| MSHADOW_XINLINE void Transpose2D(const DType *in, DType *out, index_t row, index_t col) { | ||
| // ensure cache line hits and prevent cache miss for any configuration | ||
| // L1 cache size to be utilized = 32kb = 2^15 | ||
|
|
@@ -282,7 +284,7 @@ MSHADOW_XINLINE void Transpose2D(const DType *in, DType *out, index_t row, index | |
| // Block-size - 2^5 v 2^5 (32 v 32) with potential 4 pragma for loop unrolled | ||
| // blocksize * blocksize * num_threads = cache_size / dtype_size | ||
| // Instead of explicit unroll, let compiler figure out optimal unroll factor | ||
| const index_t blocksize = 32; | ||
|
|
||
| // collapse 2 parallelizes 2 for loops | ||
| // inner 2 for loops aren't parallelized to prevent cache miss | ||
|
|
@@ -299,14 +301,25 @@ MSHADOW_XINLINE void Transpose2D(const DType *in, DType *out, index_t row, index | |
| // transpose the block | ||
| for (index_t a = j; (a < blocksize + j) && (a < col); ++a) { | ||
| for (index_t b = i; (b < blocksize + i) && (b < row); ++b) { | ||
| if (!is_addto) { | ||
| out[a * row + b] = in[b * col + a]; | ||
| } else { | ||
| out[a * row + b] += in[b * col + a]; | ||
| } | ||
| } | ||
| } | ||
| } | ||
| } | ||
| } | ||
|
|
||
| inline bool IsIdentityTranspose(const TShape& axes) { | ||
| for (dim_t i = 0; i < axes.ndim(); i++) { | ||
| if (axes[i] != i) return false; | ||
| } | ||
| return true; | ||
| } | ||
|
|
||
| template<typename xpu, bool is_addto = false> | ||
| void TransposeImpl(RunContext ctx, | ||
| const TBlob& src, | ||
| const TBlob& ret, | ||
|
|
@@ -323,62 +336,79 @@ void TransposeImpl(RunContext ctx, | |
| // Example: (0, 2, 3, 1) or (0, 3, 1, 2), but not (0, 2, 1, 3). | ||
| if (isPseudo2DTranspose(axes)) { | ||
| MSHADOW_TYPE_SWITCH(ret.type_flag_, DType, { | ||
| transpose_pseudo2D<DType, is_addto>(ret, src, axes, s); | ||
| }); | ||
| return; | ||
| } | ||
| #endif | ||
| // Special handle the identity case | ||
| if (IsIdentityTranspose(axes)) { | ||
| MSHADOW_TYPE_SWITCH(ret.type_flag_, DType, { | ||
| Tensor<xpu, 1, DType> in = src.get_with_shape<xpu, 1, DType>(mshadow::Shape1(src.Size()), s); | ||
| Tensor<xpu, 1, DType> out = ret.get_with_shape<xpu, 1, DType>(mshadow::Shape1(ret.Size()), s); | ||
| if (!is_addto) { | ||
| // Use memcpy to accelerate the speed | ||
| Copy(out, in, s); | ||
| } else { | ||
| mxnet_op::Kernel<mxnet_op::op_with_req<mshadow_op::identity, kAddTo>, xpu>::Launch( | ||
| s, ret.Size(), out.dptr_, in.dptr_); | ||
| } | ||
| }); | ||
| return; | ||
| } | ||
| // Handle the general transpose case | ||
| MSHADOW_TYPE_SWITCH(ret.type_flag_, DType, { | ||
| switch (axes.ndim()) { | ||
| case 2: { | ||
|
There was a problem hiding this comment. PR for 2D transpose with GPU is still WIP #16706 There was a problem hiding this comment. @ChaiBapchya We have already had |
||
| Tensor<xpu, 2, DType> in = src.get<xpu, 2, DType>(s); | ||
| Tensor<xpu, 2, DType> out = ret.get<xpu, 2, DType>(s); | ||
| if (ctx.get_ctx().dev_mask() == cpu::kDevMask) { | ||
| Transpose2D<DType, is_addto>(in.dptr_, out.dptr_, in.shape_[0], in.shape_[1]); | ||
| } else { | ||
| LOG(FATAL) << "Not Implemented. We should never reach here because the 2D case " | ||
| "in GPU has been covered by transpose_pseudo2D." | ||
| " Report an issue in Github."; | ||
| } | ||
| break; | ||
| } | ||
| case 3: { | ||
| Tensor<xpu, 3, DType> in = src.get<xpu, 3, DType>(s); | ||
| Tensor<xpu, 3, DType> out = ret.get<xpu, 3, DType>(s); | ||
| if (!is_addto) { | ||
| out = transpose(in, axes.get<3>()); | ||
| } else { | ||
| out += transpose(in, axes.get<3>()); | ||
| } | ||
| break; | ||
| } | ||
| case 4: { | ||
| Tensor<xpu, 4, DType> in = src.get<xpu, 4, DType>(s); | ||
| Tensor<xpu, 4, DType> out = ret.get<xpu, 4, DType>(s); | ||
| if (!is_addto) { | ||
| out = transpose(in, axes.get<4>()); | ||
| } else { | ||
| out += transpose(in, axes.get<4>()); | ||
| } | ||
| break; | ||
| } | ||
| case 5: { | ||
| Tensor<xpu, 5, DType> in = src.get<xpu, 5, DType>(s); | ||
| Tensor<xpu, 5, DType> out = ret.get<xpu, 5, DType>(s); | ||
| if (!is_addto) { | ||
| out = transpose(in, axes.get<5>()); | ||
| } else { | ||
| out += transpose(in, axes.get<5>()); | ||
| } | ||
| break; | ||
| } | ||
| case 6: { | ||
| Tensor<xpu, 6, DType> in = src.get<xpu, 6, DType>(s); | ||
| Tensor<xpu, 6, DType> out = ret.get<xpu, 6, DType>(s); | ||
| if (!is_addto) { | ||
| out = transpose(in, axes.get<6>()); | ||
| } else { | ||
| out += transpose(in, axes.get<6>()); | ||
| } | ||
| break; | ||
| } | ||
| default: | ||
|
|
@@ -399,15 +429,21 @@ void Transpose(const nnvm::NodeAttrs& attrs, | |
| return; | ||
| } | ||
| const TransposeParam& param = nnvm::get<TransposeParam>(attrs.parsed); | ||
| CHECK(req[0] == kWriteTo || req[0] == kAddTo) | ||
| << "Transpose only supports kNullOp, kWriteTo and kAddTo"; | ||
| mxnet::TShape axes; | ||
| if (param.axes.ndim() == 0) { | ||
| axes = mxnet::TShape(inputs[0].ndim(), -1); | ||
| for (int i = 0; i < axes.ndim(); ++i) { | ||
| axes[i] = axes.ndim() - 1 - i; | ||
| } | ||
| } else { | ||
| axes = common::CanonicalizeAxes(param.axes); | ||
| } | ||
| if (req[0] == kAddTo) { | ||
| TransposeImpl<xpu, true>(ctx.run_ctx, inputs[0], outputs[0], axes); | ||
| } else { | ||
| TransposeImpl<xpu, false>(ctx.run_ctx, inputs[0], outputs[0], axes); | ||
| } | ||
| } | ||
|
|
||
|
|
||
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
why was this if case removed?mx.nd.transpose(a,axes=(0,1)) is basically No Transpose (and it is handled in this scenario)
Realized you moved the Copy function above and called it
IdentityTransposeMakes sense now