Add packing for int8 1x1 convolution and support the int8 group convolution on X86 - #2991
Conversation
|
cc @yzhliu @yidawang @anijain2305 @ajtulloch |
|
@llyfacebook Please check the testcase, because of not all test machines support avx512, we need to optionally skip them. |
|
@yzhliu @kevinthesun can you help follow up and review this PR? |
|
Thanks, I will skip the tests. |
yzhliu
left a comment
yzhliu
left a comment
There was a problem hiding this comment.
@llyfacebook would you please help to elaborate the data/kernel layout for each scenario? It's pretty confusing to me so far.
| raise ValueError("not support this layout {} yet".format(data_layout)) | ||
|
|
||
|
|
||
| if data_layout == 'NHWC': |
There was a problem hiding this comment.
shall we use kernel_layout instead? as data_layout might not be necessarily binded to kernal_layout.
I'm actually a bit confused with the int8 conv layout, for NHWC data, what kernal layout is it?
There was a problem hiding this comment.
I think I was mainly following the data layout and kernel layout corresponding relationship here: https://github.com/dmlc/tvm/blob/147ea3b0ca147b527086228d524a2f68f872112d/topi/python/topi/nn/conv2d.py#L284
| oc, _, kh, kw = kshape | ||
| elif layout == 'NHWC': | ||
| n, h, w, ic = dshape | ||
| oc, _, kh, kw = kshape |
| strides = strides if isinstance(strides, (tuple, list)) else (strides, strides) | ||
| dilation = dilation if isinstance(dilation, (tuple, list)) else (dilation, dilation) | ||
|
|
||
| _, _, kh, kw = get_const_tuple(kernel.shape) |
There was a problem hiding this comment.
so, for data=NHWC & fp32, kernel=HWIO, while for data=NHWC & int8, kernel=OIHW?
There was a problem hiding this comment.
Thanks. I will spend some time this week unifying them.
| conv2d_avx_1x1._schedule_conv_nhwc_pack_int8(*args) | ||
| else: | ||
| raise ValueError("Only support 1x1 kernel with " | ||
| "schedule template.") |
There was a problem hiding this comment.
please make the fatal msg more detailed other than just "schedule template"
|
@llyfacebook please update as per review comments |
|
@yzhliu please followup on this PR |
| _get_default_config(cfg, data, kernel, strides, padding, out_dtype) | ||
| return _declaration_conv_impl(cfg, data, kernel, strides, | ||
| padding, dilation, layout, out_dtype) | ||
| # KHOI kernel layout is for NHWC and HWCN |
There was a problem hiding this comment.
| # KHOI kernel layout is for NHWC and HWCN | |
| # HWOI kernel layout is for NHWC and HWCN |
yzhliu
added
status: accepted
and removed
status: need review
status: need update
|
Thanks @llyfacebook This is merged |
…lution on X86 (apache#2991) * Support the 1x1 int8 conv with NHWC layout and weight packing fix linter * fix the memoize issue * fix the failed nhwc test * add the schedule for pack to unbreak other tests * skip avx512 compile * Support the 1x1 int8 conv with NHWC layout and weight packing fix linter * fix the memoize issue * fix the failed nhwc test * add the schedule for pack to unbreak other tests * skip avx512 compile * Unify the data_layout and kernel_layout relation * add asf header * fix the comment * retrigger the build/test
…lution on X86 (apache#2991) * Support the 1x1 int8 conv with NHWC layout and weight packing fix linter * fix the memoize issue * fix the failed nhwc test * add the schedule for pack to unbreak other tests * skip avx512 compile * Support the 1x1 int8 conv with NHWC layout and weight packing fix linter * fix the memoize issue * fix the failed nhwc test * add the schedule for pack to unbreak other tests * skip avx512 compile * Unify the data_layout and kernel_layout relation * add asf header * fix the comment * retrigger the build/test
3 participants
Thanks for contributing to TVM! Please refer to guideline https://docs.tvm.ai/contribute/ for useful information and tips. After the pull request is submitted, please request code reviews from Reviewers.