youngkingdom/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge remote-tracking branch 'origin/master' into small_alignment

Browse Source
This commit is contained in:
Haicheng Wu
2021-09-07 20:39:38 +00:00
parent 598e35401c 6c2f8f2fb8
commit 4e8af93da1
55 changed files with 317 additions and 315 deletions
Download Patch File Download Diff File Expand all files Collapse all files

43
test/unit/conv/device/conv2d_problems.h
View File

@ -281,6 +281,22 @@ struct TestbedConv2dProblemSizes {
{1, 1} // dilation (dilation_h, dilation_w)
));
conv2d_default_sizes.push_back(cutlass::conv::Conv2dProblemSize(
{1, 8, 8, 8}, // input size (NHWC)
{8, 3, 3, 8}, // filter size (KRSC)
{1, 1, 1, 1}, // padding (pad_h, _, pad_w, _)
{3, 3}, // stride (stride_h, stride_w)
{1, 1} // dilation (dilation_h, dilation_w)
));
conv2d_default_sizes.push_back(cutlass::conv::Conv2dProblemSize(
{1, 16, 16, 8}, // input size (NHWC)
{8, 3, 3, 8}, // filter size (KRSC)
{3, 3, 3, 3}, // padding (pad_h, _, pad_w, _)
{3, 3}, // stride (stride_h, stride_w)
{1, 1} // dilation (dilation_h, dilation_w)
));
////////////////////////////////////////////////////////////////////////////////////
// Medium input size (1x16x16x128), filter size (1x1, 2x2, 3x3, 5x5), stride (1, 1)
////////////////////////////////////////////////////////////////////////////////////
@ -389,6 +405,25 @@ struct TestbedConv2dProblemSizes {
{1, 1} // dilation (dilation_h, dilation_w)
));
////////////////////////////////////////////////////////////////////////////////////
// Medium input size padding > stride, asymmetric filter, padding and striding
////////////////////////////////////////////////////////////////////////////////////
conv2d_default_sizes.push_back(cutlass::conv::Conv2dProblemSize(
{1, 27, 31, 256}, // input size (NHWC)
{512, 3, 3, 256}, // filter size (KRSC)
{5, 5, 7, 7}, // padding (pad_h, _, pad_w, _)
{3, 4}, // stride (stride_h, stride_w)
{1, 1} // dilation (dilation_h, dilation_w)
));
conv2d_default_sizes.push_back(cutlass::conv::Conv2dProblemSize(
{1, 27, 35, 256}, // input size (NHWC)
{512, 7, 5, 256}, // filter size (KRSC)
{11, 11, 7, 7}, // padding (pad_h, _, pad_w, _)
{3, 5}, // stride (stride_h, stride_w)
{1, 1} // dilation (dilation_h, dilation_w)
));
////////////////////////////////////////////////////////////////////////////////////
// Medium input size *mixed* stride (1, 2) and (2, 1),
// filter (3, 3), default padding
@ -419,6 +454,14 @@ struct TestbedConv2dProblemSizes {
{2, 2}, // stride (stride_h, stride_w)
{1, 1} // dilation (dilation_h, dilation_w)
));
conv2d_default_sizes.push_back(cutlass::conv::Conv2dProblemSize(
{1, 32, 32, 16}, // input size (NHWC)
{32, 3, 3, 16}, // filter size (KRSC)
{1, 1, 1, 1}, // padding (pad_h, _, pad_w, _)
{6, 2}, // stride (stride_h, stride_w)
{1, 1} // dilation (dilation_h, dilation_w)
));
conv2d_default_sizes.push_back(cutlass::conv::Conv2dProblemSize(
{32, 32, 32, 32}, // input size (NHWC)

20
test/unit/conv/device/conv2d_strided_dgrad_implicit_gemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32_sm80.cu
View File

@ -78,23 +78,15 @@ TEST(SM80_Device_Conv2d_Strided_Dgrad_Analytic_ImplicitGemm_f16nhwc_f16nhwc_f32n
test::conv::device::Conv2dProblemVector problem_size_list;
#if 0 // run specific problem size in the unit test first
problem_size_list.push_back(cutlass::conv::Conv2dProblemSize(
{1, 56, 56, 8}, // input size (NHWC)
{8, 1, 1, 8}, // filter size (KRSC)
{0, 0, 0, 0}, // padding (pad_h, _, pad_w, _)
{2, 2}, // stride (stride_h, stride_w)
{1, 1} // dilation (dilation_h, dilation_w)
{1, 4, 4, 8}, // input size (NHWC)
{8, 3, 3, 8}, // filter size (KRSC)
{0, 0, 0, 0}, // padding (pad_h, _, pad_w, _)
{3, 3}, // stride (stride_h, stride_w)
{1, 1} // dilation (dilation_h, dilation_w)
));
problem_size_list.push_back(cutlass::conv::Conv2dProblemSize(
{1, 55, 55, 8}, // input size (NHWC)
{8, 1, 1, 8}, // filter size (KRSC)
{0, 0, 0, 0}, // padding (pad_h, _, pad_w, _)
{2, 2}, // stride (stride_h, stride_w)
{1, 1} // dilation (dilation_h, dilation_w)
));
#endif
/// Run all unit test sizes with device-level Conv2d instance