tuning para,

driver/include/device_convolution_backward_data_implicit_gemm_v1r1_nchw_kcyx_nkhw.hpp

@@ -49,7 +49,7 @@ void device_convolution_backward_data_implicit_gemm_v1r1_nchw_kcyx_nkhw(InDesc i
     wei_kcyx_device_buf.ToDevice(wei_kcyx.mData.data());
     out_nkhw_device_buf.ToDevice(out_nkhw.mData.data());
 
-#if 0
+#if 1
     // BlockSize = 256, each thread hold 64 data
     constexpr index_t BlockSize = 256;
 
@@ -83,6 +83,36 @@ void device_convolution_backward_data_implicit_gemm_v1r1_nchw_kcyx_nkhw(InDesc i
     // BlockSize = 256, each thread hold 64 data
     constexpr index_t BlockSize = 256;
 
+    constexpr index_t GemmMPerBlock              = 128;
+    constexpr index_t GemmNPerBlock              = 128;
+    constexpr index_t GemmKPerBlock              = 8;
+    constexpr index_t GemmMPerThread         = 4;
+    constexpr index_t GemmNPerThread         = 4;
+    constexpr index_t GemmKPerThread         = 1;
+    constexpr index_t GemmMLevel0Cluster         = 4;
+    constexpr index_t GemmNLevel0Cluster         = 4;
+    constexpr index_t GemmMLevel1Cluster         = 4;
+    constexpr index_t GemmNLevel1Cluster         = 4;
+    constexpr index_t GemmThreadGemmDataPerReadM = 4;
+    constexpr index_t GemmThreadGemmDataPerReadN = 4;
+
+    using GemmABlockCopyThreadSliceLengths_GemmK_GemmM   = Sequence<1, 4>;
+    using GemmABlockCopyThreadClusterLengths_GemmK_GemmM = Sequence<8, 32>;
+
+    constexpr index_t GemmABlockCopySrcDataPerRead_GemmM  = 4;
+    constexpr index_t GemmABlockCopyDstDataPerWrite_GemmM = 4;
+
+    using GemmBBlockCopyThreadSliceLengths_GemmK_GemmN   = Sequence<1, 4>;
+    using GemmBBlockCopyThreadClusterLengths_GemmK_GemmN = Sequence<8, 32>;
+
+    constexpr index_t GemmBBlockCopySrcDataPerRead_GemmN  = 4;
+    constexpr index_t GemmBBlockCopyDstDataPerWrite_GemmN = 4;
+
+    constexpr index_t GemmCThreadCopyDstDataPerWrite_GemmN1 = 4;
+#elif 1
+    // BlockSize = 256, each thread hold 64 data
+    constexpr index_t BlockSize = 256;
+
     constexpr index_t GemmMPerBlock              = 128;
     constexpr index_t GemmNPerBlock              = 128;
     constexpr index_t GemmKPerBlock              = 16;

driver/include/device_convolution_backward_data_implicit_gemm_v4r1_nchw_kcyx_nkhw.hpp

@@ -62,7 +62,7 @@ void device_convolution_backward_data_implicit_gemm_v4r1_nchw_kcyx_nkhw(InDesc i
     wei_kcyx_device_buf.ToDevice(wei_kcyx.mData.data());
     out_nkhw_device_buf.ToDevice(out_nkhw.mData.data());
 
-#if 1
+#if 0
     // BlockSize = 256, each thread hold 64 data
     constexpr index_t BlockSize = 256;
 
@@ -109,19 +109,19 @@ void device_convolution_backward_data_implicit_gemm_v4r1_nchw_kcyx_nkhw(InDesc i
     constexpr index_t GemmThreadGemmDataPerReadM = 4;
     constexpr index_t GemmThreadGemmDataPerReadN = 4;
 
-    using GemmABlockCopyThreadSliceLengths_GemmK_GemmM   = Sequence<8, 1>;
-    using GemmABlockCopyThreadClusterLengths_GemmK_GemmM = Sequence<2, 128>;
+    using GemmABlockCopyThreadSliceLengths_GemmK_GemmM   = Sequence<2, 4>;
+    using GemmABlockCopyThreadClusterLengths_GemmK_GemmM = Sequence<8, 32>;
 
-    constexpr index_t GemmABlockCopySrcDataPerRead_GemmM  = 1;
-    constexpr index_t GemmABlockCopyDstDataPerWrite_GemmM = 1;
+    constexpr index_t GemmABlockCopySrcDataPerRead_GemmM  = 4;
+    constexpr index_t GemmABlockCopyDstDataPerWrite_GemmM = 4;
 
-    using GemmBBlockCopyThreadSliceLengths_GemmK_GemmN   = Sequence<8, 1>;
-    using GemmBBlockCopyThreadClusterLengths_GemmK_GemmN = Sequence<2, 128>;
+    using GemmBBlockCopyThreadSliceLengths_GemmK_GemmN   = Sequence<2, 4>;
+    using GemmBBlockCopyThreadClusterLengths_GemmK_GemmN = Sequence<8, 32>;
 
-    constexpr index_t GemmBBlockCopySrcDataPerRead_GemmN  = 1;
-    constexpr index_t GemmBBlockCopyDstDataPerWrite_GemmN = 1;
+    constexpr index_t GemmBBlockCopySrcDataPerRead_GemmN  = 4;
+    constexpr index_t GemmBBlockCopyDstDataPerWrite_GemmN = 4;
 
-    constexpr index_t GemmCThreadCopyDstDataPerWrite_GemmN1 = 1;
+    constexpr index_t GemmCThreadCopyDstDataPerWrite_GemmN1 = 4;
 #endif
 
     constexpr index_t GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);

driver/include/device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw.hpp

@@ -54,7 +54,7 @@ void device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw(InDesc,
     wei_kcyx_device_buf.ToDevice(wei_kcyx.mData.data());
     out_nkhw_device_buf.ToDevice(out_nkhw.mData.data());
 
-#if 0
+#if 1
     // BlockSize = 256, EperBlock = 8, each thread hold 64 data
     constexpr index_t BlockSize = 256;
 

driver/src/conv_bwd_data_driver.cpp

@@ -23,7 +23,7 @@ int main(int argc, char* argv[])
 {
     using namespace launcher;
 
-#if 1
+#if 0
     constexpr index_t N  = 64;
     constexpr index_t C  = 256;
     constexpr index_t HI = 56;
@@ -190,10 +190,10 @@ int main(int argc, char* argv[])
 #elif 1
     // 3x3 filter, 2x2 stride, 35x35 input, 17x17 output
     constexpr index_t N  = 128;
-    constexpr index_t C  = 1024;
+    constexpr index_t C  = 128;
     constexpr index_t HI = 35;
     constexpr index_t WI = 35;
-    constexpr index_t K  = 128;
+    constexpr index_t K  = 1024;
     constexpr index_t Y  = 3;
     constexpr index_t X  = 3;
 
@@ -247,7 +247,7 @@ int main(int argc, char* argv[])
 
 #if 1
     device_convolution_backward_data_implicit_gemm_v1r1_nchw_kcyx_nkhw
-#elif 0
+#elif 1
     device_convolution_backward_data_implicit_gemm_v1r2_nchw_kcyx_nkhw
 #elif 0
     device_convolution_backward_data_implicit_gemm_v2r1_nchw_kcyx_nkhw

driver/src/conv_driver.cpp

@@ -29,13 +29,13 @@ int main(int argc, char* argv[])
 {
     using namespace ck;
 
-#if 1
+#if 0
     // 1x1
-    constexpr index_t N  = 64;
-    constexpr index_t C  = 64;
-    constexpr index_t HI = 56;
-    constexpr index_t WI = 56;
-    constexpr index_t K  = 256;
+    constexpr index_t N  = 128;
+    constexpr index_t C  = 256;
+    constexpr index_t HI = 14;
+    constexpr index_t WI = 14;
+    constexpr index_t K  = 1024;
     constexpr index_t Y  = 1;
     constexpr index_t X  = 1;
 
@@ -47,10 +47,10 @@ int main(int argc, char* argv[])
 #elif 0
     // 1x7
     constexpr index_t N  = 128;
-    constexpr index_t C  = 1024;
+    constexpr index_t C  = 128;
     constexpr index_t HI = 17;
     constexpr index_t WI = 17;
-    constexpr index_t K  = 1024;
+    constexpr index_t K  = 128;
     constexpr index_t Y  = 1;
     constexpr index_t X  = 7;
 
@@ -59,7 +59,7 @@ int main(int argc, char* argv[])
 
     using LeftPads  = Sequence<0, 3>;
     using RightPads = Sequence<0, 3>;
-#elif 0
+#elif 1
     // 3x3, 34x34
     constexpr index_t N  = 64;
     constexpr index_t C  = 256;
@@ -74,7 +74,7 @@ int main(int argc, char* argv[])
 
     using LeftPads  = Sequence<0, 0>;
     using RightPads = Sequence<0, 0>;
-#elif 0
+#elif 1
     // 3x3 filter, 2x2 stride, 35x35 input, 17x17 output
     constexpr index_t N  = 128;
     constexpr index_t C  = 128;
@@ -434,7 +434,7 @@ int main(int argc, char* argv[])
                                                                     ConvStrides{},
                                                                     ConvDilations{},
                                                                     nrepeat);
-#elif 0
+#elif 1
     device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw(in_nchw_desc,
                                                          in_nchw,
                                                          wei_kcyx_desc,