Deprecate static kernel (#42)

* deprecate static kernels

Deprecate static kernel (#42)
* deprecate static kernels
81c942cd · Chao Liu · GitHub · b8b2d0a6 · b8b2d0a6 · b8b2d0a6
Unverified Commit 81c942cd authored Jul 08, 2021 by Chao Liu Committed by GitHub Jul 08, 2021
15 changed files
--- a/driver/include/device_convolution_backward_data_implicit_gemm_v1r1_nchw_kcyx_nkhw.hpp
+++ b/driver/include/device_convolution_backward_data_implicit_gemm_v1r1_nchw_kcyx_nkhw.hpp
--- a/driver/include/device_convolution_backward_data_implicit_gemm_v1r2_nchw_kcyx_nkhw.hpp
+++ b/driver/include/device_convolution_backward_data_implicit_gemm_v1r2_nchw_kcyx_nkhw.hpp
--- a/driver/include/device_convolution_backward_data_implicit_gemm_v4r1_nchw_kcyx_nkhw.hpp
+++ b/driver/include/device_convolution_backward_data_implicit_gemm_v4r1_nchw_kcyx_nkhw.hpp
--- a/driver/include/device_convolution_backward_data_implicit_gemm_v5r1_nhwc_kyxc_nhwk.hpp
+++ b/driver/include/device_convolution_backward_data_implicit_gemm_v5r1_nhwc_kyxc_nhwk.hpp
--- a/driver/include/device_convolution_forward_implicit_gemm_v4r1_nchw_kcyx_nkhw.hpp
+++ b/driver/include/device_convolution_forward_implicit_gemm_v4r1_nchw_kcyx_nkhw.hpp
--- a/driver/include/device_convolution_forward_implicit_gemm_v4r4_nchw_kcyx_nkhw.hpp
+++ b/driver/include/device_convolution_forward_implicit_gemm_v4r4_nchw_kcyx_nkhw.hpp
--- a/driver/include/device_convolution_forward_implicit_gemm_v4r4_nhwc_kyxc_nhwk.hpp
+++ b/driver/include/device_convolution_forward_implicit_gemm_v4r4_nhwc_kyxc_nhwk.hpp
--- a/driver/include/device_dynamic_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw.hpp
+++ b/driver/include/device_dynamic_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw.hpp
@@ -257,7 +257,6 @@ void device_dynamic_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nk
            GemmKPerBlock,
            GemmMPerWave,
            GemmNPerWave,
-            GemmK1,
            MRepeat,
            NRepeat,
            GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1,

--- a/driver/include/device_dynamic_convolution_forward_implicit_gemm_v4r4r2_xdlops_nhwc_kyxc_nhwk.hpp
+++ b/driver/include/device_dynamic_convolution_forward_implicit_gemm_v4r4r2_xdlops_nhwc_kyxc_nhwk.hpp
--- a/driver/include/device_dynamic_convolution_forward_implicit_gemm_v4r4r3_xdlops_nhwc_kyxc_nhwk.hpp
+++ b/driver/include/device_dynamic_convolution_forward_implicit_gemm_v4r4r3_xdlops_nhwc_kyxc_nhwk.hpp
--- a/driver/include/device_dynamic_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk.hpp
+++ b/driver/include/device_dynamic_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk.hpp
--- a/driver/include/device_dynamic_convolution_forward_implicit_gemm_v4r5r2_nchw_kcyx_nkhw.hpp
+++ b/driver/include/device_dynamic_convolution_forward_implicit_gemm_v4r5r2_nchw_kcyx_nkhw.hpp
--- a/driver/include/device_tensor.hpp
+++ b/driver/include/device_tensor.hpp
--- a/script/cmake-cuda.sh
+++ b/script/cmake-cuda.sh
--- a/script/extract_asm-cuda.sh
+++ b/script/extract_asm-cuda.sh
-DRIVER=$1
-ARCH=$2
-cuobjdump -xelf $ARCH ./driver/$DRIVER && nvdisasm --print-code -g $DRIVER.$ARCH.cubin > $DRIVER.$ARCH.asm