experimenting

driver/device_implicit_gemm_convolution_2_chwn_cyxk_khwn.hpp

@@ -160,7 +160,7 @@ void device_implicit_gemm_convolution_2_chwn_cyxk_khwn(InDesc,
     constexpr unsigned WeiBlockCopyDataPerRead = 4;
 
     constexpr unsigned BlockSize = 128;
-#elif 1
+#elif 0
     // 1x1, 28x28, 256 thread
     constexpr unsigned BPerBlock = 128;
     constexpr unsigned KPerBlock = 128;
@@ -211,7 +211,7 @@ void device_implicit_gemm_convolution_2_chwn_cyxk_khwn(InDesc,
     for(unsigned i = 0; i < nrepeat; ++i)
     {
         float time = launch_kernel(
-#if 0
+#if 1
             gridwise_implicit_gemm_convolution_2_chwn_cyxk_khwn
 #else
             gridwise_implicit_gemm_convolution_2_chwn_cyxk_khwn_lds_double_buffer

driver/driver.hip.cpp

@@ -661,9 +661,9 @@ int main(int argc, char* argv[])
     device_direct_convolution_2_nchw_kcyx_nkhw
 #elif 0
     device_direct_convolution_2_vectorized_nchw_kcyx_nkhw
-#elif 1
-    device_implicit_gemm_convolution_1_chwn_cyxk_khwn
 #elif 0
+    device_implicit_gemm_convolution_1_chwn_cyxk_khwn
+#elif 1
     device_implicit_gemm_convolution_2_chwn_cyxk_khwn
 #endif
     (in_nchw_desc, in_nchw, wei_kcyx_desc, wei_kcyx, out_nkhw_desc, out_nkhw_device, nrepeat);

src/include/gridwise_implicit_gemm_convolution_1_chwn_cyxk_khwn.hip.hpp

@@ -6,7 +6,7 @@
 #include "blockwise_2d_tensor_op.hip.hpp"
 #include "threadwise_nd_tensor_op.hip.hpp"
 #include "threadwise_4d_tensor_op.hip.hpp"
-#include "blockwise_gemm.hip.hpp"
+#include "blockwise_batched_gemm.hip.hpp"
 
 template <unsigned GridSize,
           unsigned BlockSize,
@@ -211,8 +211,6 @@ gridwise_implicit_gemm_convolution_1_chwn_cyxk_khwn(const Float* const __restric
             {
 #if 0
                 blockwise_batch_gemm.Run
-#elif 0
-                blockwise_batch_gemm.Run_v2
 #elif 1
                 blockwise_batch_gemm.Run_v3
 #endif

src/include/gridwise_implicit_gemm_convolution_2_chwn_cyxk_khwn.hip.hpp

@@ -236,9 +236,13 @@ gridwise_implicit_gemm_convolution_2_chwn_cyxk_khwn(const Float* const __restric
             for(unsigned x = 0; x < X; ++x)
             {
                 auto f_accum = [](auto& acc, const auto&& v) { acc += v; };
-#if 1
+#if 0
                 blockwise_gemm.Run
-#else
+#elif 1
+                blockwise_gemm.Run_asm
+#elif 0
+                blockwise_gemm.Run_v2
+#elif 0
                 blockwise_gemm.Run_RegisterDoubleBuffer
 #endif
                 (p_wei_block + wei_cyxk_block_desc.Get1dIndex(0, y, x, 0),

src/include/gridwise_implicit_gemm_convolution_2_chwn_cyxk_khwn_lds_double_buffer.hip.hpp

@@ -34,7 +34,11 @@ template <unsigned GridSize,
           unsigned WeiBlockCopyThreadPerDim1,
           unsigned InBlockCopyDataPerRead,
           unsigned WeiBlockCopyDataPerRead>
-__global__ void gridwise_implicit_gemm_convolution_2_chwn_cyxk_khwn_lds_double_buffer(
+__global__ void
+#if 0
+__launch_bounds__(256,2)
+#endif
+gridwise_implicit_gemm_convolution_2_chwn_cyxk_khwn_lds_double_buffer(
     const Float* const __restrict__ p_in_global,
     const Float* const __restrict__ p_wei_global,
     Float* const __restrict__ p_out_global)
@@ -280,7 +284,7 @@ __global__ void gridwise_implicit_gemm_convolution_2_chwn_cyxk_khwn_lds_double_b
             for(unsigned x = 0; x < X; ++x)
             {
                 auto f_accum = [](auto& acc, const auto&& v) { acc += v; };
-#if 1
+#if 0
                 blockwise_gemm.Run
 #else
                 blockwise_gemm.Run_RegisterDoubleBuffer