test

driver/device_convolution_implicit_gemm_v4_nchw_kcyx_nkhw.hpp

@@ -29,8 +29,7 @@ void device_convolution_implicit_gemm_v4_nchw_kc1x1_nkhw(InDesc,
     constexpr auto I3 = Number<3>{};
 
     constexpr auto in_nchw_desc = InDesc{};
-    static_assert(WeiDesc{}.GetLength(I2) == 1, "1x1 filter only");
-    static_assert(WeiDesc{}.GetLength(I3) == 1, "1x1 filter only");
+    static_assert(WeiDesc{}.GetLength(I2) == 1 && WeiDesc{}.GetLength(I3) == 1, "1x1 filter only");
     constexpr auto wei_kc_desc   = WeiDesc{}.Extract(Sequence<0, 1>{});
     constexpr auto out_nkhw_desc = OutDesc{};
 

driver/driver.hip.cpp

@@ -493,13 +493,13 @@ void check_error(const Tensor<T>& ref, const Tensor<T>& result)
 
 int main(int argc, char* argv[])
 {
-    constexpr index_t HStride = 2;
-    constexpr index_t WStride = 2;
+    constexpr index_t HStride = 1;
+    constexpr index_t WStride = 1;
 
     constexpr index_t HDilation = 1;
     constexpr index_t WDilation = 1;
 
-    constexpr index_t Direction = 0; // 1: Forward; 0:Backward
+    constexpr index_t Direction = 2; // 1: Forward; 0:Backward
 #if 0
     constexpr index_t N  = 32;
     constexpr index_t C  = 128;
@@ -550,10 +550,10 @@ int main(int argc, char* argv[])
 #elif 1
     // 1x1 filter, 28x28 image
     constexpr index_t N  = 128;
-    constexpr index_t C  = 128;
-    constexpr index_t HI = 7;
-    constexpr index_t WI = 7;
-    constexpr index_t K  = 128;
+    constexpr index_t C  = 256;
+    constexpr index_t HI = 56;
+    constexpr index_t WI = 56;
+    constexpr index_t K  = 256;
     constexpr index_t Y  = 1;
     constexpr index_t X  = 1;