fix host bug for bwd data

driver/include/host_conv_bwd_data.hpp

@@ -8,13 +8,13 @@ template <typename TIn,
           typename ConvDilations,
           typename LeftPads,
           typename RightPads>
-void host_direct_convolution_bwd_data(Tensor<TIn>& in_nchw,
-                                      const Tensor<TWei>& wei_kcyx,
-                                      const Tensor<TOut>& out_nkhw,
-                                      ConvStrides,
-                                      ConvDilations,
-                                      LeftPads,
-                                      RightPads)
+void host_direct_convolution_backward_data(Tensor<TIn>& in_nchw,
+                                           const Tensor<TWei>& wei_kcyx,
+                                           const Tensor<TOut>& out_nkhw,
+                                           ConvStrides,
+                                           ConvDilations,
+                                           LeftPads,
+                                           RightPads)
 {
     using namespace ck;
 
@@ -37,21 +37,27 @@ void host_direct_convolution_bwd_data(Tensor<TIn>& in_nchw,
         {
             int h_tmp = hi + LeftPads{}[0] - y * ConvDilations{}[0];
 
-            if(h_tmp >= 0 && h_tmp < HI && h_tmp % ConvStrides{}[0] == 0)
+            if(h_tmp % ConvStrides{}[0] == 0)
             {
                 int ho = h_tmp / ConvStrides{}[0];
 
-                for(int x = 0; x < X; ++x)
+                if(ho >= 0 && ho < HO)
                 {
-                    int w_tmp = wi + LeftPads{}[1] - x * ConvDilations{}[1];
-
-                    if(w_tmp >= 0 && w_tmp < WI && w_tmp % ConvStrides{}[1] == 0)
+                    for(int x = 0; x < X; ++x)
                     {
-                        int wo = w_tmp / ConvStrides{}[1];
+                        int w_tmp = wi + LeftPads{}[1] - x * ConvDilations{}[1];
 
-                        for(int k = 0; k < K; ++k)
+                        if(w_tmp % ConvStrides{}[1] == 0)
                         {
-                            v += out_nkhw(n, k, ho, wo) * wei_kcyx(k, c, y, x);
+                            int wo = w_tmp / ConvStrides{}[1];
+
+                            if(wo >= 0 && wo < WO)
+                            {
+                                for(int k = 0; k < K; ++k)
+                                {
+                                    v += out_nkhw(n, k, ho, wo) * wei_kcyx(k, c, y, x);
+                                }
+                            }
                         }
                     }
                 }

driver/src/conv_bwd_data_driver.cpp

@@ -21,15 +21,15 @@ int main(int argc, char* argv[])
     using namespace ck;
 
 #if 0
-    constexpr index_t N  = 128;
-    constexpr index_t C  = 256;
-    constexpr index_t HI = 35;
-    constexpr index_t WI = 35;
-    constexpr index_t K  = 384;
-    constexpr index_t Y  = 3;
-    constexpr index_t X  = 3;
+    constexpr index_t N  = 4;
+    constexpr index_t C  = 8;
+    constexpr index_t HI = 11;
+    constexpr index_t WI = 11;
+    constexpr index_t K  = 8;
+    constexpr index_t Y  = 4;
+    constexpr index_t X  = 4;
 
-    using ConvStrides   = Sequence<2, 2>;
+    using ConvStrides   = Sequence<1, 1>;
     using ConvDilations = Sequence<1, 1>;
 
     using LeftPads  = Sequence<0, 0>;
@@ -49,7 +49,7 @@ int main(int argc, char* argv[])
 
     using LeftPads  = Sequence<0, 0>;
     using RightPads = Sequence<0, 0>;
-#elif 1
+#elif 0
     // 1x1 filter, 8x8 image
     // cudnn@V100 68%, ck@V100 72%, ck@P100 52%, ck@VII 42%
     constexpr index_t N  = 64;
@@ -241,7 +241,7 @@ int main(int argc, char* argv[])
 
     using LeftPads  = Sequence<0, 0>;
     using RightPads = Sequence<0, 0>;
-#elif 1
+#elif 0
     // 3x3 filter, 2x2 stride, 35x35 input, 17x17 output
     // cudnn@V100 90%, ck@V100 93%, ck@P100 83%, ck@VII 81%
     constexpr index_t N  = 128;
@@ -287,7 +287,7 @@ int main(int argc, char* argv[])
 
     using LeftPads  = Sequence<3, 0>;
     using RightPads = Sequence<3, 0>;
-#elif 0
+#elif 1
     // 1x7 filter, 0x3 pad, 17x17 input
     constexpr index_t N  = 128;
     constexpr index_t C  = 128;
@@ -364,13 +364,13 @@ int main(int argc, char* argv[])
 
     if(do_verification)
     {
-        host_direct_convolution_bwd_data(in_nchw_host,
-                                         wei_kcyx,
-                                         out_nkhw,
-                                         ConvStrides{},
-                                         ConvDilations{},
-                                         LeftPads{},
-                                         RightPads{});
+        host_direct_convolution_backward_data(in_nchw_host,
+                                              wei_kcyx,
+                                              out_nkhw,
+                                              ConvStrides{},
+                                              ConvDilations{},
+                                              LeftPads{},
+                                              RightPads{});
 
         check_error(in_nchw_host, in_nchw_device);
 

script/compile-hip.sh

@@ -4,5 +4,5 @@
  export KMDUMPLLVM=1
  export KMDUMPDIR=$PWD
 
- make -j driver
+ make -j $1
 #/opt/rocm/hcc/bin/llvm-objdump -mcpu=gfx906 -source -line-numbers driver/dump-gfx906.isabin > driver/dump-gfx906.isabin.asm

script/docker-cuda.sh

+WORKSPACE=$1
+echo "workspace: " $WORKSPACE
+sudo docker run  -it  -v $WORKSPACE:/root/workspace --group-add sudo --runtime=nvidia     asroy/cuda:10.1-cudnn7-devel-ubuntu18.04-latest /bin/bash

script/ds_read_offset.sh

-for((i=0;i<=4096;i=i+64))
-do
-    OFFSET=$i
-    echo "if(offset == $OFFSET)"
-    echo "{"
-    echo "    asm volatile(\"\\n \\"
-    echo "        ds_read_b128 %0, %1 offset:$OFFSET\n \\"
-    echo "        \""
-    echo "    : \"=v\"(r)"
-    echo "    : \"v\"(__to_local(lds)));"
-    echo "}"
-done