Add aten::_thnn_fused_gru_cell and _thnn_fused_lstm_cell #926
Conversation
yucai-intel
commented
Sep 20, 2024
•
yucai-intel
commented
- thnn_fused_gru_cell_forward
- thnn_fused_gru_cell_backward
- thnn_fused_lstm_cell_forward
- thnn_fused_lstm_cell_backward
src/ATen/native/xpu/GRUFusedCell.cpp
Outdated
| #include <ATen/core/op_registration/adaption.h> | ||
| #include <ATen/native/cpu/mixed_data_type.h> | ||
| #include <ATen/native/xpu/sycl/GRUFusedCellKernels.h> | ||
| #include <ATen/xpu/XPUNativeFunctions.h> |
There was a problem hiding this comment.
This header file seems to be non-existent from the build logs.
Would changing to XPUFunctions.h instead of XPUNativeFunctions.h make it work?
| #include <ATen/core/op_registration/adaption.h> | ||
| #include <ATen/native/cpu/mixed_data_type.h> | ||
| #include <ATen/native/xpu/sycl/RNNKernels.h> | ||
|
|
There was a problem hiding this comment.
#include <ATen/core/op_registration/adaption.h>
#include <ATen/native/cpu/mixed_data_type.h>
please check if we need head files
|
|
||
| auto hy = at::empty_like(hx, LEGACY_CONTIGUOUS_MEMORY_FORMAT); | ||
|
|
||
| AT_DISPATCH_FLOATING_TYPES_AND2( |
There was a problem hiding this comment.
How we handle non-contiguous input ? CUDA uses tensorInfo, but we just use raw data pointer, without stride information
| const Tensor& hidden_bias = | ||
| c10::value_or_else(hidden_bias_opt, [] { return Tensor(); }); | ||
|
|
||
| auto batched_input = true; |
There was a problem hiding this comment.
If hard code true, can we remove it
yucai-intel
changed the title
|
@yucai-intel Pls show the testcase |
|
2024-11-12T10:33:51.5087379Z test_modules_xpu.py::TestModuleXPU::test_device_ctx_init_nn_LSTMCell_xpu_float32 SKIPPED [ 11%] |
|
2024-11-12T10:33:51.6360467Z test_modules_xpu.py::TestModuleXPU::test_errors_nn_GRUCell_xpu_float32 PASSED [ 15%] |
