revert driver changes that are irrelavent to magic number division PR

driver/include/device_dynamic_convolution_forward_implicit_gemm_v4r4_nchw_kcyx_nkhw.hpp

@@ -40,7 +40,7 @@ void device_dynamic_convolution_forward_implicit_gemm_v4r4_nchw_kcyx_nkhw(
     wei_k_c_y_x_device_buf.ToDevice(wei_k_c_y_x.mData.data());
     out_n_k_ho_wo_device_buf.ToDevice(out_n_k_ho_wo.mData.data());
 
-#if 1
+#if 0
     // run-time variables
     const auto in_n_c_hi_wi_desc =
         make_dynamic_naive_tensor_descriptor_packed_v2(to_multi_index(InDesc::GetLengths()));
@@ -167,7 +167,7 @@ void device_dynamic_convolution_forward_implicit_gemm_v4r4_nchw_kcyx_nkhw(
     constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmN = 1;
 
     constexpr index_t GemmCThreadTransferDstScalarPerVector_GemmN1 = 4;
-#elif 0
+#elif 1
     // cdata = 64, BlockSize 64, 16x256x4
     constexpr index_t BlockSize = 64;
 

driver/include/device_dynamic_convolution_forward_implicit_gemm_v5r1_nchw_kcyx_nkhw.hpp

@@ -83,10 +83,10 @@ void device_dynamic_convolution_forward_implicit_gemm_v5r1_nchw_kcyx_nkhw(
     const auto out_n_k0_ho_wo_k1_desc =
         make_dynamic_naive_tensor_descriptor_packed_v2(make_tuple(N, K0, Ho, Wo, K1));
 
-    const auto conv_strides   = sequence_to_tuple_of_number(ConvStrides{});
-    const auto conv_dilations = sequence_to_tuple_of_number(ConvDilations{});
-    const auto in_left_pads   = sequence_to_tuple_of_number(InLeftPads{});
-    const auto in_right_pads  = sequence_to_tuple_of_number(InRightPads{});
+    const auto conv_strides     = sequence_to_tuple_of_number(ConvStrides{});
+    const auto conv_dilations   = sequence_to_tuple_of_number(ConvDilations{});
+    const auto in_left_pads     = sequence_to_tuple_of_number(InLeftPads{});
+    const auto in_right_pads    = sequence_to_tuple_of_number(InRightPads{});
 #endif
 
     Tensor<TInWei> in_n_c0_hi_wi_c1(make_HostTensorDescriptor(

driver/src/conv_driver.cpp

@@ -48,8 +48,8 @@ int main(int argc, char* argv[])
     using ConvStrides   = Sequence<1, 1>;
     using ConvDilations = Sequence<1, 1>;
 
-    using LeftPads  = Sequence<0, 0>;
-    using RightPads = Sequence<0, 0>;
+    using LeftPads                   = Sequence<0, 0>;
+    using RightPads                  = Sequence<0, 0>;
 #elif 0
     constexpr index_t N  = 1;
     constexpr index_t C  = 16;
@@ -62,9 +62,9 @@ int main(int argc, char* argv[])
     using ConvStrides   = Sequence<1, 1>;
     using ConvDilations = Sequence<1, 1>;
 
-    using LeftPads  = Sequence<0, 0>;
-    using RightPads = Sequence<0, 0>;
-#elif 0
+    using LeftPads                   = Sequence<0, 0>;
+    using RightPads                  = Sequence<0, 0>;
+#elif 1
     constexpr index_t N  = 1;
     constexpr index_t C  = 16;
     constexpr index_t HI = 1080;
@@ -135,7 +135,7 @@ int main(int argc, char* argv[])
 
     using LeftPads  = Sequence<0, 0>;
     using RightPads = Sequence<0, 0>;
-#elif 1
+#elif 0
     // 3x3, 35x35, stride 2
     constexpr index_t N  = 128;
     constexpr index_t C  = 192;
@@ -150,7 +150,7 @@ int main(int argc, char* argv[])
 
     using LeftPads  = Sequence<0, 0>;
     using RightPads = Sequence<0, 0>;
-#elif 1
+#elif 0
     // 3x3, 71x71
     constexpr index_t N  = 128;
     constexpr index_t C  = 192;
@@ -630,17 +630,17 @@ int main(int argc, char* argv[])
     print_array("ConvStrides", to_multi_index(ConvStrides{}));
     print_array("ConvDilations", to_multi_index(ConvDilations{}));
 
-#if 1
+#if 0
     using in_data_t                  = float;
     constexpr index_t in_vector_size = 1;
     using acc_data_t                 = float;
     using out_data_t                 = float;
 #elif 0
-    using in_data_t                  = int8_t;
-    constexpr index_t in_vector_size = 4;
-    using acc_data_t                 = int32_t;
+    using in_data_t                  = float;
+    constexpr index_t in_vector_size = 1;
+    using acc_data_t                 = float;
     using out_data_t                 = int8_t;
-#elif 0
+#elif 1
     using in_data_t                  = int8_t;
     constexpr index_t in_vector_size = 16;
     using acc_data_t                 = int32_t;
@@ -724,7 +724,7 @@ int main(int argc, char* argv[])
                                                                  LeftPads{},
                                                                  RightPads{},
                                                                  nrepeat);
-#elif 1
+#elif 0
     device_dynamic_convolution_forward_implicit_gemm_v4r4_nchw_kcyx_nkhw<in_data_t,
                                                                          in_vector_size,
                                                                          acc_data_t,
@@ -741,7 +741,7 @@ int main(int argc, char* argv[])
          LeftPads{},
          RightPads{},
          nrepeat);
-#elif 1
+#elif 0
     device_dynamic_convolution_forward_implicit_gemm_v4r4_nhwc_kyxc_nhwk<in_data_t,
                                                                          in_vector_size,
                                                                          acc_data_t,