Add quantized_maxpool_2d for xpu #1049
Conversation
gaopengff
changed the title
gaopengff
changed the title
|
The specific unit test cases have passed in latest CI. 2024-11-07T09:10:50.0722836Z quantization/core/test_quantized_op_xpu.py::TestQuantizedOpsXPU::test_max_pool2d_nhwc_xpu PASSED [ 48%]
2024-11-07T09:10:50.1551494Z quantization/core/test_quantized_op_xpu.py::TestQuantizedOpsXPU::test_max_pool2d_pt2e_xpu PASSED [ 50%]
2024-11-07T09:10:50.5847030Z quantization/core/test_quantized_op_xpu.py::TestQuantizedOpsXPU::test_max_pool2d_xpu PASSED [ 51%] |
| w_start += dW_; | ||
|
|
||
| // Stock pytorch's cpu implementation use vectorized instructions | ||
| // through channels such as AVX-512. We use for-loop directly. |
There was a problem hiding this comment.
just for confirmation, the work-item indexing is optimized for cl right? I mean, tensor is cl, and the inner-most dim of work-item is on channels
There was a problem hiding this comment.
It's not optimized for cl. The innter-most dim is nbatch like DilatedMaxPool2d implementation. On our gpu we do not have vectorized method, so I use unrolled loop to simulate it. If compiler compiled and optimized the loop to vectorized codes, it will help. I just followed the stock's implementation.
There was a problem hiding this comment.
Great thanks for your explanations. I am OK that we can keep this implementation. Functionality is preferred now.
hi, @EikanWang FYI, if perf is also important for us currently, could we add further optimization after this PR?
ZhiweiYan-96
requested review from
EikanWang,
fengyuan14,
xytintel and
CuiYifeng
Now we only support datatype of uint8(Byte). Referring the stock pytorch cpu implementation at code.
Waiting #921 to be merged.