add some configs

driver/include/device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw.hpp

@@ -522,7 +522,7 @@ void device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw(InDesc,
 
     constexpr index_t WeiBlockCopySrcDataPerRead_E  = 4;
     constexpr index_t WeiBlockCopyDstDataPerWrite_K = 1;
-#elif 1
+#elif 0
     // cdata = 64, BlockSize = 32, 32x64x3
     constexpr index_t BlockSize = 32;
 
@@ -559,6 +559,45 @@ void device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw(InDesc,
     using WeiBlockCopySrcAccessOrder            = Sequence<1, 0>; // [K, E]
     using WeiBlockCopyDstAccessOrder            = Sequence<0, 1>; // [E, K]
 
+    constexpr index_t WeiBlockCopySrcDataPerRead_E  = 1;
+    constexpr index_t WeiBlockCopyDstDataPerWrite_K = 1;
+#elif 1
+    // cdata = 64, BlockSize = 64, 32x128x3
+    constexpr index_t BlockSize = 64;
+
+    constexpr index_t KPerBlock = 32;
+    constexpr index_t BPerBlock = 16;
+    constexpr index_t EPerBlock = 3;
+
+    constexpr index_t GemmNRepeat = 2;
+
+    constexpr index_t GemmMPerThreadSubC = 4;
+    constexpr index_t GemmNPerThreadSubC = 4;
+    constexpr index_t GemmKPerThreadLoop = 1;
+
+    constexpr index_t GemmMLevel0Cluster = 4;
+    constexpr index_t GemmNLevel0Cluster = 4;
+    constexpr index_t GemmMLevel1Cluster = 1;
+    constexpr index_t GemmNLevel1Cluster = 4;
+
+    constexpr index_t GemmDataPerReadA   = 4;
+    constexpr index_t GemmDataPerReadB   = 4;
+
+    using InBlockCopySubLengths_E_N1_B_N2      = Sequence<3, 1, 1, 2>;
+    using InBlockCopyClusterLengths_E_N1_B_N2  = Sequence<1, 2, 16, 2>;
+    using InBlockCopyThreadClusterArrangeOrder = Sequence<0, 1, 3, 2>; // [E, N1, N2, B]
+    using InBlockCopySrcAccessOrder            = Sequence<0, 2, 1, 3>; // [E, B, N1, N2]
+    using InBlockCopyDstAccessOrder            = Sequence<0, 1, 2, 3>; // [E, N1, B, N2]
+
+    constexpr index_t InBlockCopySrcDataPerRead_B   = 1;
+    constexpr index_t InBlockCopyDstDataPerWrite_N2 = 2;
+
+    using WeiBlockCopySubLengths_E_K            = Sequence<3, 1>;
+    using WeiBlockCopyClusterLengths_E_K        = Sequence<1, 32>;
+    using WeiBlockCopyThreadClusterArrangeOrder = Sequence<1, 0>; // [K, E]
+    using WeiBlockCopySrcAccessOrder            = Sequence<1, 0>; // [K, E]
+    using WeiBlockCopyDstAccessOrder            = Sequence<0, 1>; // [E, K]
+
     constexpr index_t WeiBlockCopySrcDataPerRead_E  = 1;
     constexpr index_t WeiBlockCopyDstDataPerWrite_K = 1;
 #elif 0

driver/include/device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw.hpp

@@ -758,7 +758,7 @@ void device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw(InDesc,
     constexpr index_t GemmBBlockCopyDstDataPerWrite_GemmN = 1;
 
     constexpr index_t GemmCThreadCopyDstDataPerWrite_GemmN1 = 1;
-#elif 1
+#elif 0
     // cdata = 64, BlockSize = 32, 32x64x3
     constexpr index_t BlockSize = 32;
 
@@ -790,6 +790,39 @@ void device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw(InDesc,
     constexpr index_t GemmBBlockCopySrcDataPerRead_GemmN  = 1;
     constexpr index_t GemmBBlockCopyDstDataPerWrite_GemmN = 2;
 
+    constexpr index_t GemmCThreadCopyDstDataPerWrite_GemmN1 = 1;
+#elif 1
+    // cdata = 64, BlockSize = 64, 32x128x3
+    constexpr index_t BlockSize = 64;
+
+    constexpr index_t GemmMPerBlock = 32;
+    constexpr index_t GemmNPerBlock = 128;
+    constexpr index_t GemmKPerBlock = 3;
+
+    constexpr index_t GemmMPerThreadSubC     = 4;
+    constexpr index_t GemmNPerThreadSubC     = 4;
+    constexpr index_t GemmKPerThreadLoop     = 1;
+
+    constexpr index_t GemmMLevel0Cluster     = 2;
+    constexpr index_t GemmNLevel0Cluster     = 2;
+    constexpr index_t GemmMLevel1Cluster     = 2;
+    constexpr index_t GemmNLevel1Cluster     = 8;
+
+    constexpr index_t ThreadGemmDataPerReadM = 4;
+    constexpr index_t ThreadGemmDataPerReadN = 4;
+
+    using GemmABlockCopyThreadSliceLengths_GemmK_GemmM   = Sequence<3, 1>;
+    using GemmABlockCopyThreadClusterLengths_GemmK_GemmM = Sequence<1, 32>;
+
+    constexpr index_t GemmABlockCopySrcDataPerRead_GemmK  = 1;
+    constexpr index_t GemmABlockCopyDstDataPerWrite_GemmM = 1;
+
+    using GemmBBlockCopyThreadSliceLengths_GemmK_GemmN   = Sequence<3, 2>;
+    using GemmBBlockCopyThreadClusterLengths_GemmK_GemmN = Sequence<1, 64>;
+
+    constexpr index_t GemmBBlockCopySrcDataPerRead_GemmN  = 1;
+    constexpr index_t GemmBBlockCopyDstDataPerWrite_GemmN = 2;
+
     constexpr index_t GemmCThreadCopyDstDataPerWrite_GemmN1 = 1;
 #elif 1
     // cdata = 64, BlockSize = 64, 64x64x3