expect n_per_wmma and m_per_wmma to be 32 for gfx950

Author: kabraham-streamhpc

experimental/builder/include/ck_tile/builder/conv_algorithm_limits.hpp

@@ -49,7 +49,9 @@ concept IsValidWmmaConvolutionTileing =
     (thread_block.tile_size.m % (gridwise_gemm.m_per_wmma * gridwise_gemm.m_wmma_per_wave) == 0) &&
     (thread_block.tile_size.n >= gridwise_gemm.m_per_wmma * gridwise_gemm.n_wmma_per_wave) &&
     (thread_block.tile_size.n % (gridwise_gemm.m_per_wmma * gridwise_gemm.n_wmma_per_wave) == 0) &&
-    (gridwise_gemm.m_per_wmma == 16) && // Check if this holds on MI400 and add switch if it dosen't
-    (gridwise_gemm.n_per_wmma == 16);   // Check if this holds on MI400 and add switch if it dosen't
-
+#if defined(__gfx950__)
+    (gridwise_gemm.m_per_wmma == 32) && (gridwise_gemm.n_per_wmma == 32);
+#else
+    (gridwise_gemm.m_per_wmma == 16) && (gridwise_gemm.n_per_wmma == 16);
+#endif
 } // namespace ck_tile::builder