From 4543d17a7191c16431e0b26876e59e6c35a2a692 Mon Sep 17 00:00:00 2001
From: Chao Liu <lc.roy86@gmail.com>
Date: Tue, 19 Feb 2019 22:07:15 -0600
Subject: [PATCH] refactor

---
 .../device_implicit_gemm_convolution_1_chwn_csrk_khwn.hpp | 3 ---
 driver/driver.hip.cpp                                     | 2 +-
 ...ise_implicit_gemm_convolution_1_chwn_csrk_khwn.hip.hpp | 8 +++++---
 3 files changed, 6 insertions(+), 7 deletions(-)

diff --git a/driver/device_implicit_gemm_convolution_1_chwn_csrk_khwn.hpp b/driver/device_implicit_gemm_convolution_1_chwn_csrk_khwn.hpp
index 81fef7dcbd..14e4d29f74 100644
--- a/driver/device_implicit_gemm_convolution_1_chwn_csrk_khwn.hpp
+++ b/driver/device_implicit_gemm_convolution_1_chwn_csrk_khwn.hpp
@@ -227,9 +227,6 @@ void device_implicit_gemm_convolution_1_chwn_csrk_khwn(InDesc,
     constexpr unsigned HoPerThread = 1;
     constexpr unsigned WoPerThread = 1;
 
-    constexpr unsigned WeiBlockCopyThreadPerDim0 = 4;
-    constexpr unsigned WeiBlockCopyThreadPerDim1 = 32;
-
     constexpr unsigned InBlockCopy_ThreadPerDimC = 8;
     constexpr unsigned InBlockCopy_ThreadPerDimH = 2;
     constexpr unsigned InBlockCopy_ThreadPerDimW = 2;
diff --git a/driver/driver.hip.cpp b/driver/driver.hip.cpp
index 77be2059a3..cc234f5091 100644
--- a/driver/driver.hip.cpp
+++ b/driver/driver.hip.cpp
@@ -491,7 +491,7 @@ int main(int argc, char* argv[])
 
     constexpr unsigned HPad = 1;
     constexpr unsigned WPad = 1;
-#elif 1
+#elif 0
     // 1x1 filter, 28x28 image
     constexpr unsigned N  = 16;
     constexpr unsigned C  = 256;
diff --git a/src/include/gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn.hip.hpp b/src/include/gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn.hip.hpp
index cc8c08e8d5..f11be63954 100644
--- a/src/include/gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn.hip.hpp
+++ b/src/include/gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn.hip.hpp
@@ -94,8 +94,8 @@ gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn(const Float* const __restric
 
     // tensor view of blockwise input and weight in LDS
     //   be careful of alignment
-    constexpr auto in_chwn_block_desc =
-        make_ConstantTensorDescriptor(Sequence<CPerBlock, HiPerBlock, WiPerBlock, NPerBlock>{});
+    constexpr auto in_chwn_block_desc = make_ConstantTensorDescriptor_aligned(
+        Sequence<CPerBlock, HiPerBlock, WiPerBlock, NPerBlock>{}, Number<InBlockCopyDataPerRead>{});
 
     constexpr auto wei_ek_block_desc = make_ConstantTensorDescriptor_aligned(
         Sequence<CPerBlock * S * R, KPerBlock>{}, Number<WeiBlockCopyDataPerRead>{});
@@ -164,7 +164,9 @@ gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn(const Float* const __restric
         HoPerThread>{};
 
     // LDS: be careful of alignment
-    constexpr unsigned in_block_size = in_chwn_block_desc.GetElementSpace();
+    constexpr unsigned in_block_size =
+        in_chwn_block_desc.GetElementSpace(Number<InBlockCopyDataPerRead>{});
+
     constexpr unsigned wei_block_size =
         wei_csrk_block_desc.GetElementSpace(Number<WeiBlockCopyDataPerRead>{});