Compiling failure XdlopsBackwardData for fp16

[zjing14]

./bin/test_conv2d --half --cmode conv --pmode default --group-count 1 --disable-backward-weights --input 16, 16, 7, 7 --weights 4, 16, 1, 1 --pads_strides_dilations 0 0 1 1 1 1 --trans_output_pads 0 0
./gridwise_convolution_backward_data_implicit_gemm_v1r1_xdlops_gnchw_gkcyx_gnkhw.hpp:68:9: error: static_assert failed due to requirement 'K % 8 == 0' "K needs to be in multiple of KPACK"
./gridwise_gemm_xdlops_fp16_bfp16.hpp:797:9: error: static_assert failed due to requirement 'M % 256 == 0 && N % 256 == 0 && K % 8 == 0' "wrong! cannot divide work evenly among block"

./bin/test_conv2d --half --cmode conv --pmode default --group-count 1 --disable-backward-weights --input 16, 16, 14, 14 --weights 4, 16, 3, 3 --pads_strides_dilations 1 1 2 2 1 1 --trans_output_pads 0 0
./gridwise_convolution_backward_data_implicit_gemm_v1r1_xdlops_gnchw_gkcyx_gnkhw.hpp:68:9: error: static_assert failed due to requirement 'K % 8 == 0' "K needs to be in multiple of KPACK"
./gridwise_gemm_xdlops_fp16_bfp16.hpp:797:9: error: static_assert failed due to requirement 'M % 256 == 0 && N % 256 == 0 && K % 8 == 0' "wrong! cannot divide work evenly among block"

./bin/test_conv2d --half --cmode conv --pmode default --group-count 1 --disable-backward-weights --input 16, 16, 7, 7 --weights 8, 16, 1, 1 --pads_strides_dilations 0 0 1 1 1 1 --trans_output_pads 0 0
./gridwise_gemm_xdlops_fp16_bfp16.hpp:797:9: error: static_assert failed due to requirement 'M % 256 == 0 && N % 256 == 0 && K % 8 == 0' "wrong! cannot divide work evenly among block"

./bin/test_conv2d --half --cmode conv --pmode default --group-count 1 --disable-backward-weights --input 16, 16, 14, 14 --weights 8, 16, 3, 3 --pads_strides_dilations 1 1 2 2 1 1 --trans_output_pads 0 0
./gridwise_gemm_xdlops_fp16_bfp16.hpp:797:9: error: static_assert failed due to requirement 'M % 256 == 0 && N % 256 == 0 && K % 8 == 0' "wrong! cannot divide work evenly among block"

./bin/test_conv2d --half --cmode conv --pmode default --group-count 1 --disable-backward-weights --input 16, 32, 7, 7 --weights 4, 32, 1, 1 --pads_strides_dilations 0 0 1 1 1 1 --trans_output_pads 0 0
./gridwise_convolution_backward_data_implicit_gemm_v1r1_xdlops_gnchw_gkcyx_gnkhw.hpp:68:9: error: static_assert failed due to requirement 'K % 8 == 0' "K needs to be in multiple of KPACK"
./gridwise_gemm_xdlops_fp16_bfp16.hpp:797:9: error: static_assert failed due to requirement 'M % 256 == 0 && N % 256 == 0 && K % 8 == 0' "wrong! cannot divide work evenly among block"

./bin/test_conv2d --half --cmode conv --pmode default --group-count 1 --disable-backward-weights --input 16, 32, 14, 14 --weights 4, 32, 3, 3 --pads_strides_dilations 1 1 2 2 1 1 --trans_output_pads 0 0
./gridwise_convolution_backward_data_implicit_gemm_v1r1_xdlops_gnchw_gkcyx_gnkhw.hpp:68:9: error: static_assert failed due to requirement 'K % 8 == 0' "K needs to be in multiple of KPACK"
./gridwise_gemm_xdlops_fp16_bfp16.hpp:797:9: error: static_assert failed due to requirement 'M % 256 == 0 && N % 256 == 0 && K % 8 == 0' "wrong! cannot divide work evenly among block"

./bin/test_conv2d --half --cmode conv --pmode default --group-count 1 --disable-backward-weights --input 16, 32, 7, 7 --weights 8, 32, 1, 1 --pads_strides_dilations 0 0 1 1 1 1 --trans_output_pads 0 0
./gridwise_gemm_xdlops_fp16_bfp16.hpp:797:9: error: static_assert failed due to requirement 'M % 256 == 0 && N % 256 == 0 && K % 8 == 0' "wrong! cannot divide work evenly among block"

./bin/test_conv2d --half --cmode conv --pmode default --group-count 1 --disable-backward-weights --input 16, 32, 7, 7 --weights 8, 32, 1, 1 --pads_strides_dilations 0 0 1 1 1 1 --trans_output_pads 0 0
./gridwise_gemm_xdlops_fp16_bfp16.hpp:797:9: error: static_assert failed due to requirement 'M % 256 == 0 && N % 256 == 0 && K % 8 == 0' "wrong! cannot divide work evenly among block"

[asroy]

we should deprecate this solver/kernel, because we would use following instead: ConvHipImplicitGemmBwdDataV1R1Xdlops gridwise_convolution_backward_data_implicit_gemm_v1r1_xdlops_nchw_kcyx_nkhw.cpp

[atamazov]

Good news. Just in case -- if we need to quickly disable it, then changing the default value of respective env variable will do the job ( IsDisabled() -> !IsEnabled()).

[aserio]

@zjing14 and @asroy, what is the status of this issue?

[aserio]

@zjing14 please close if new PR resolves this issue

[atamazov]

Workaround by #457. Full blown fix is not yet implemented, so let's keep it open (until W/A is removed) but lower priority. And this is not a bug anymore.

[ppanchad-amd]

@zjing14 Is this still an issue? If not, can we close this ticket? Thanks!

[huanrwan-amd]

Hi @atamazov, is it OK to close the issue? As we are not observing the failure in the original test: "failed due to requirement 'K % 8 == 0' "K needs to be in multiple of KPACK""