From e55cfe1536e44cccfe814f29a6b1f37255d02b3b Mon Sep 17 00:00:00 2001
From: Chao Liu <lc.roy86@gmail.com>
Date: Mon, 24 Jun 2019 11:49:13 -0500
Subject: [PATCH] debugging vector load for generic tensor copy

---
 ...lution_implicit_gemm_v4_nchw_kcyx_nkhw.hpp | 40 +++++++++++++++++--
 driver/src/driver.cpp                         |  3 ++
 2 files changed, 40 insertions(+), 3 deletions(-)

diff --git a/driver/include/device_convolution_implicit_gemm_v4_nchw_kcyx_nkhw.hpp b/driver/include/device_convolution_implicit_gemm_v4_nchw_kcyx_nkhw.hpp
index 94d8f203f1..c8f9d17648 100644
--- a/driver/include/device_convolution_implicit_gemm_v4_nchw_kcyx_nkhw.hpp
+++ b/driver/include/device_convolution_implicit_gemm_v4_nchw_kcyx_nkhw.hpp
@@ -64,7 +64,7 @@ void device_convolution_implicit_gemm_v4_nchw_kcyx_nkhw(InDesc,
 
     constexpr index_t BPerBlock = 16;
     constexpr index_t KPerBlock = 128;
-    constexpr index_t CPerBlock = 8;
+    constexpr index_t EPerBlock = 8;
 
     constexpr index_t GemmMPerThreadSubC = 4;
     constexpr index_t GemmNPerThreadSubC = 4;
@@ -98,7 +98,41 @@ void device_convolution_implicit_gemm_v4_nchw_kcyx_nkhw(InDesc,
 
     constexpr index_t BPerBlock = 16;
     constexpr index_t KPerBlock = 128;
-    constexpr index_t CPerBlock = 8;
+    constexpr index_t EPerBlock = 8;
+
+    constexpr index_t GemmMPerThreadSubC = 4;
+    constexpr index_t GemmNPerThreadSubC = 4;
+    constexpr index_t GemmMLevel0Cluster = 4;
+    constexpr index_t GemmNLevel0Cluster = 4;
+    constexpr index_t GemmMLevel1Cluster = 4;
+    constexpr index_t GemmNLevel1Cluster = 4;
+    constexpr index_t GemmKPerThreadLoop = 1;
+    constexpr index_t GemmDataPerReadA   = 4;
+    constexpr index_t GemmDataPerReadB   = 4;
+
+    using InBlockCopySubLengths_E_N1_B_N2      = Sequence<1, 1, 4, 1>;
+    using InBlockCopyClusterLengths_E_N1_B_N2  = Sequence<8, 2, 4, 4>;
+    using InBlockCopyThreadClusterArrangeOrder = Sequence<0, 1, 3, 2>; // [E, N1, N2, B]
+    using InBlockCopySrcAccessOrder            = Sequence<0, 1, 3, 2>; // [E, N1, N2, B]
+    using InBlockCopyDstAccessOrder            = Sequence<0, 1, 2, 3>; // [E, N1, B, N2]
+
+    constexpr index_t InBlockCopySrcDataPerRead_B   = 4;
+    constexpr index_t InBlockCopyDstDataPerWrite_N2 = 1;
+
+    using WeiBlockCopySubLengths_E_K            = Sequence<4, 1>;
+    using WeiBlockCopyClusterLengths_E_K        = Sequence<2, 128>;
+    using WeiBlockCopyThreadClusterArrangeOrder = Sequence<1, 0>; // [K, E]
+    using WeiBlockCopySrcAccessOrder            = Sequence<1, 0>; // [K, E]
+    using WeiBlockCopyDstAccessOrder            = Sequence<0, 1>; // [E, K]
+
+    constexpr index_t WeiBlockCopySrcDataPerRead_E  = 4;
+    constexpr index_t WeiBlockCopyDstDataPerWrite_K = 1;
+#elif 1
+    constexpr index_t BlockSize = 256;
+
+    constexpr index_t BPerBlock = 16;
+    constexpr index_t KPerBlock = 128;
+    constexpr index_t EPerBlock = 8;
 
     constexpr index_t GemmMPerThreadSubC = 4;
     constexpr index_t GemmNPerThreadSubC = 4;
@@ -152,7 +186,7 @@ void device_convolution_implicit_gemm_v4_nchw_kcyx_nkhw(InDesc,
              ConvDilations,
              BPerBlock,
              KPerBlock,
-             CPerBlock,
+             EPerBlock,
              N1,
              N2,
              GemmMPerThreadSubC,
diff --git a/driver/src/driver.cpp b/driver/src/driver.cpp
index 8826b62ab3..d2c5b607b8 100644
--- a/driver/src/driver.cpp
+++ b/driver/src/driver.cpp
@@ -566,6 +566,9 @@ int main(int argc, char* argv[])
     constexpr index_t Y  = 1;
     constexpr index_t X  = 1;
 
+    using ConvStrides   = Sequence<1, 1>;
+    using ConvDilations = Sequence<1, 1>;
+
     constexpr index_t HPad = 0;
     constexpr index_t WPad = 0;
 #elif 0