merge master

composable_kernel/include/utility/config.amd.hpp.in

@@ -14,11 +14,11 @@
 #define CK_DEVICE_BACKEND_AMD 1
 
 // GPU ID
-#if 1
+#if 0
 #define CK_AMD_GPU_GFX906 1
 #elif 0
 #define CK_AMD_GPU_GFX908 1
-#elif 0
+#elif 1
 #define CK_AMD_GPU_GFX1030 1
 #endif
 
@@ -28,10 +28,10 @@
 #endif
 
 // launch bounds
-#define CK_USE_LAUNCH_BOUNDS 0
+#define CK_USE_LAUNCH_BOUNDS 1
 
 #ifdef CK_USE_LAUNCH_BOUNDS
-#define CK_MAX_THREAD_PER_BLOCK 256
+#define CK_MAX_THREAD_PER_BLOCK 64
 #define CK_MIN_BLOCK_PER_CU 1
 #endif
 

driver/include/device_dynamic_convolution_forward_implicit_gemm_v5r1_nchw_kcyx_nkhw.hpp

@@ -118,11 +118,11 @@ void device_dynamic_convolution_forward_implicit_gemm_v5r1_nchw_kcyx_nkhw(
 
     constexpr index_t KPerBlock  = K;
     constexpr index_t HoPerBlock = 8;
-    constexpr index_t WoPerBlock = 32;
+    constexpr index_t WoPerBlock = 64;
     constexpr index_t EPerBlock  = C0;
 
     constexpr index_t KPerThread  = KPerBlock;
-    constexpr index_t HoPerThread = 2;
+    constexpr index_t HoPerThread = 4;
     constexpr index_t WoPerThread = 2;
     constexpr index_t EPerThread  = EPerBlock;
 

driver/src/conv_driver.cpp

@@ -92,7 +92,7 @@ int main(int argc, char* argv[])
 
     using LeftPads  = Sequence<1, 1>;
     using RightPads = Sequence<1, 1>;
-#elif 0
+#elif 1
     constexpr index_t N  = 1;
     constexpr index_t C  = 16;
     constexpr index_t HI = 540;
@@ -630,7 +630,7 @@ int main(int argc, char* argv[])
     print_array("ConvStrides", to_multi_index(ConvStrides{}));
     print_array("ConvDilations", to_multi_index(ConvDilations{}));
 
-#if 1
+#if 0
     using in_data_t                  = float;
     constexpr index_t in_vector_size = 1;
     using acc_data_t                 = float;
@@ -724,7 +724,7 @@ int main(int argc, char* argv[])
                                                                  LeftPads{},
                                                                  RightPads{},
                                                                  nrepeat);
-#elif 1
+#elif 0
     device_dynamic_convolution_forward_implicit_gemm_v4r4_nchw_kcyx_nkhw<in_data_t,
                                                                          in_vector_size,
                                                                          acc_data_t,