Add kernel_mode option to im2col, col2im, and convgemm functions

5 files changed, 58 insertions, 29 deletions

diff --git a/test/correctness/misc/override_parameters.cpp b/test/correctness/misc/override_parameters.cpp
index 54229c5e..7ed4faff 100644
--- a/test/correctness/misc/override_parameters.cpp
+++ b/test/correctness/misc/override_parameters.cpp
@@ -60,6 +60,7 @@ size_t RunOverrideTests(int argc, char *argv[], const bool silent, const std::st
   args.layout = GetArgument(arguments, help, kArgLayout, Layout::kRowMajor);
   args.a_transpose = GetArgument(arguments, help, kArgATransp, Transpose::kNo);
   args.b_transpose = GetArgument(arguments, help, kArgBTransp, Transpose::kNo);
+  args.kernel_mode = GetArgument(arguments, help, kArgKernelMode, KernelMode::kCrossCorrelation);
   args.alpha = GetArgument(arguments, help, kArgAlpha, GetScalar<T>());
   args.beta  = GetArgument(arguments, help, kArgBeta, GetScalar<T>());
 
diff --git a/test/correctness/testblas.hpp b/test/correctness/testblas.hpp
index 137df30f..b2dc6e7a 100644
--- a/test/correctness/testblas.hpp
+++ b/test/correctness/testblas.hpp
@@ -63,6 +63,7 @@ class TestBlas: public Tester<T,U> {
   static const std::vector<size_t> kNumKernels;
   static const std::vector<size_t> kStrideValues;
   static const std::vector<size_t> kChannelValues;
+  static const std::vector<KernelMode> kKernelModes;
   const std::vector<size_t> kOffsets;
   const std::vector<U> kAlphaValues;
   const std::vector<U> kBetaValues;
@@ -142,6 +143,7 @@ template <typename T, typename U> const std::vector<size_t> TestBlas<T,U>::kKern
 template <typename T, typename U> const std::vector<size_t> TestBlas<T,U>::kNumKernels = { 1, 6 };
 template <typename T, typename U> const std::vector<size_t> TestBlas<T,U>::kStrideValues = { 1, 3 };
 template <typename T, typename U> const std::vector<size_t> TestBlas<T,U>::kChannelValues = { 1, 2 };
+template <typename T, typename U> const std::vector<KernelMode> TestBlas<T,U>::kKernelModes = { KernelMode::kCrossCorrelation, KernelMode::kConvolution };
 
 // Test settings for the invalid tests
 template <typename T, typename U> const std::vector<size_t> TestBlas<T,U>::kInvalidIncrements = { 0, 1 };
@@ -168,6 +170,7 @@ static StatusCode ReferenceNotAvailable(const Arguments<U> &, BufferType &, Queu
 template <typename C, typename T, typename U>
 void handle_remaining_of_options(std::vector<Arguments<U>> &regular_test_vector, Arguments<U> &r_args,
                                  TestBlas<T,U> &tester,
+                                 const std::vector<KernelMode> &kernel_modes,
                                  const std::vector<size_t> &channelss,
                                  const std::vector<size_t> &heights,
                                  const std::vector<size_t> &widths,
@@ -181,21 +184,23 @@ void handle_remaining_of_options(std::vector<Arguments<U>> &regular_test_vector,
                                  const std::vector<size_t> &dilation_ws,
                                  const std::vector<size_t> &batch_counts,
                                  const std::vector<size_t> &num_kernelss) {
-  for (auto &channels: channelss) { r_args.channels = channels;
-    for (auto &height: heights) { r_args.height = height;
-      for (auto &width: widths) { r_args.width = width;
-        for (auto &kernel_h: kernel_hs) { r_args.kernel_h = kernel_h;
-          for (auto &kernel_w: kernel_ws) { r_args.kernel_w = kernel_w;
-            for (auto &pad_h: pad_hs) { r_args.pad_h = pad_h;
-              for (auto &pad_w: pad_ws) { r_args.pad_w = pad_w;
-                for (auto &stride_h: stride_hs) { r_args.stride_h = stride_h;
-                  for (auto &stride_w: stride_ws) { r_args.stride_w = stride_w;
-                    for (auto &dilation_h: dilation_hs) { r_args.dilation_h = dilation_h;
-                      for (auto &dilation_w: dilation_ws) { r_args.dilation_w = dilation_w;
-                        for (auto &batch_count: batch_counts) { r_args.batch_count = batch_count;
-                          for (auto &num_kernels: num_kernelss) { r_args.num_kernels = num_kernels;
-                            C::SetSizes(r_args, tester.queue_);
-                            regular_test_vector.push_back(r_args);
+  for (auto &kernel_mode: kernel_modes) { r_args.kernel_mode = kernel_mode;
+    for (auto &channels: channelss) { r_args.channels = channels;
+      for (auto &height: heights) { r_args.height = height;
+        for (auto &width: widths) { r_args.width = width;
+          for (auto &kernel_h: kernel_hs) { r_args.kernel_h = kernel_h;
+            for (auto &kernel_w: kernel_ws) { r_args.kernel_w = kernel_w;
+              for (auto &pad_h: pad_hs) { r_args.pad_h = pad_h;
+                for (auto &pad_w: pad_ws) { r_args.pad_w = pad_w;
+                  for (auto &stride_h: stride_hs) { r_args.stride_h = stride_h;
+                    for (auto &stride_w: stride_ws) { r_args.stride_w = stride_w;
+                      for (auto &dilation_h: dilation_hs) { r_args.dilation_h = dilation_h;
+                        for (auto &dilation_w: dilation_ws) { r_args.dilation_w = dilation_w;
+                          for (auto &batch_count: batch_counts) { r_args.batch_count = batch_count;
+                            for (auto &num_kernels: num_kernelss) { r_args.num_kernels = num_kernels;
+                              C::SetSizes(r_args, tester.queue_);
+                              regular_test_vector.push_back(r_args);
+                            }
                           }
                         }
                       }
@@ -284,6 +289,7 @@ size_t RunTests(int argc, char *argv[], const bool silent, const std::string &na
   auto imax_offsets = std::vector<size_t>{args.imax_offset};
   auto alphas = std::vector<U>{args.alpha};
   auto betas = std::vector<U>{args.beta};
+  auto kernel_modes = std::vector<KernelMode>{args.kernel_mode};
   auto channelss = std::vector<size_t>{args.channels};
   auto heights = std::vector<size_t>{args.height};
   auto widths = std::vector<size_t>{args.width};
@@ -340,6 +346,7 @@ size_t RunTests(int argc, char *argv[], const bool silent, const std::string &na
     if (option == kArgImaxOffset) { imax_offsets = tester.kOffsets; }
     if (option == kArgAlpha) { alphas = tester.kAlphaValues; }
     if (option == kArgBeta) { betas = tester.kBetaValues; }
+    if (option == kArgKernelMode) { kernel_modes = tester.kKernelModes; }
     if (option == kArgChannels) { channelss = tester.kChannelValues; }
     if (option == kArgHeight) { heights = tester.kMatrixDims; }
     if (option == kArgWidth) { widths = tester.kMatrixDims; }
@@ -397,6 +404,7 @@ size_t RunTests(int argc, char *argv[], const bool silent, const std::string &na
                                                         for (auto &beta: betas) { r_args.beta = beta;
                                                           // Cannot have more for-loops because of MSVC's C1061 error
                                                           handle_remaining_of_options<C>(regular_test_vector, r_args, tester,
+                                                                                      kernel_modes,
                                                                                       channelss, heights, widths, kernel_hs, kernel_ws,
                                                                                       pad_hs, pad_ws, stride_hs, stride_ws,
                                                                                       dilation_hs, dilation_ws,
diff --git a/test/routines/levelx/xcol2im.hpp b/test/routines/levelx/xcol2im.hpp
index 176fceae..c28727e7 100644
--- a/test/routines/levelx/xcol2im.hpp
+++ b/test/routines/levelx/xcol2im.hpp
@@ -31,7 +31,8 @@ public:
 
   // The list of arguments relevant for this routine
   static std::vector<std::string> GetOptions() {
-    return {kArgChannels, kArgHeight, kArgWidth, kArgKernelH, kArgKernelW, kArgPadH, kArgPadW,
+    return {kArgKernelMode,
+            kArgChannels, kArgHeight, kArgWidth, kArgKernelH, kArgKernelW, kArgPadH, kArgPadW,
             kArgStrideH, kArgStrideW, kArgDilationH, kArgDilationW,
             kArgAOffset, kArgBOffset};
   }
@@ -87,7 +88,8 @@ public:
     #ifdef OPENCL_API
       auto queue_plain = queue();
       auto event = cl_event{};
-      auto status = Col2im<T>(args.channels, args.height, args.width,
+      auto status = Col2im<T>(args.kernel_mode,
+                              args.channels, args.height, args.width,
                               args.kernel_h, args.kernel_w,
                               args.pad_h, args.pad_w,
                               args.stride_h, args.stride_w,
@@ -97,7 +99,8 @@ public:
                               &queue_plain, &event);
       if (status == StatusCode::kSuccess) { clWaitForEvents(1, &event); clReleaseEvent(event); }
     #elif CUDA_API
-      auto status = Col2im<T>(args.channels, args.height, args.width,
+      auto status = Col2im<T>(args.kernel_mode,
+                              args.channels, args.height, args.width,
                               args.kernel_h, args.kernel_w,
                               args.pad_h, args.pad_w,
                               args.stride_h, args.stride_w,
@@ -167,7 +170,10 @@ StatusCode RunReference(const Arguments<T> &args, BuffersHost<T> &buffers_host)
           for (auto w_id = size_t{0}; w_id < col_w; ++w_id) { // image width
 
             // Reads the input value
-            const auto kernel_index = kw_id + args.kernel_w * kh_id;
+            const auto kernel_index
+                    = (args.kernel_mode == KernelMode::kConvolution)
+                    ? args.kernel_h * args.kernel_w - kw_id - args.kernel_w * kh_id - 1
+                    : kw_id + args.kernel_w * kh_id;
             const auto patch_index = w_id + col_w * h_id;
             const auto col_index = patch_index + kernel_index * col_w * col_h +
                                    c_id * col_w * col_h * args.kernel_h * args.kernel_w;
diff --git a/test/routines/levelx/xconvgemm.hpp b/test/routines/levelx/xconvgemm.hpp
index 7fa4e701..e67b8174 100644
--- a/test/routines/levelx/xconvgemm.hpp
+++ b/test/routines/levelx/xconvgemm.hpp
@@ -91,7 +91,8 @@ public:
 #ifdef OPENCL_API
     auto queue_plain = queue();
     auto event = cl_event{};
-    auto status = Convgemm<T>(args.channels, args.height, args.width,
+    auto status = Convgemm<T>(args.kernel_mode,
+                              args.channels, args.height, args.width,
                               args.kernel_h, args.kernel_w,
                               args.pad_h, args.pad_w,
                               args.stride_h, args.stride_w,
@@ -103,7 +104,8 @@ public:
                               &queue_plain, &event);
     if (status == StatusCode::kSuccess) { clWaitForEvents(1, &event); clReleaseEvent(event); }
 #elif CUDA_API
-    auto status = Convgemm<T>(args.channels, args.height, args.width,
+    auto status = Convgemm<T>(args.kernel_mode,
+                              args.channels, args.height, args.width,
                               args.kernel_h, args.kernel_w,
                               args.pad_h, args.pad_w,
                               args.stride_h, args.stride_w,
@@ -189,10 +191,16 @@ StatusCode RunReference(const Arguments<T> &args, BuffersHost<T> &buffers_host)
                   const auto input_value = buffers_host.a_mat[input_index + args.a_offset];
 
                   // Multiplies with the kernel tensor
-                  const auto kernel_index = kw_id + args.kernel_w * (
-                                            kh_id + args.kernel_h * (
-                                            ci_id + args.channels * (
-                                            co_id)));
+                  const auto kernel_index
+                          = (args.kernel_mode == KernelMode::kConvolution)
+                          ? (args.kernel_w - kw_id - 1) + args.kernel_w * (
+                            (args.kernel_h - kh_id - 1) + args.kernel_h * (
+                            ci_id + args.channels * (
+                            co_id)))
+                          : kw_id + args.kernel_w * (
+                            kh_id + args.kernel_h * (
+                            ci_id + args.channels * (
+                            co_id)));
                   const auto kernel_value = buffers_host.b_mat[kernel_index + args.b_offset];
                   result += input_value * kernel_value;
 
diff --git a/test/routines/levelx/xim2col.hpp b/test/routines/levelx/xim2col.hpp
index acf7998b..2a3577c3 100644
--- a/test/routines/levelx/xim2col.hpp
+++ b/test/routines/levelx/xim2col.hpp
@@ -31,7 +31,8 @@ public:
 
   // The list of arguments relevant for this routine
   static std::vector<std::string> GetOptions() {
-    return {kArgChannels, kArgHeight, kArgWidth, kArgKernelH, kArgKernelW, kArgPadH, kArgPadW,
+    return {kArgKernelMode,
+            kArgChannels, kArgHeight, kArgWidth, kArgKernelH, kArgKernelW, kArgPadH, kArgPadW,
             kArgStrideH, kArgStrideW, kArgDilationH, kArgDilationW,
             kArgAOffset, kArgBOffset};
   }
@@ -87,7 +88,8 @@ public:
     #ifdef OPENCL_API
       auto queue_plain = queue();
       auto event = cl_event{};
-      auto status = Im2col<T>(args.channels, args.height, args.width,
+      auto status = Im2col<T>(args.kernel_mode,
+                              args.channels, args.height, args.width,
                               args.kernel_h, args.kernel_w,
                               args.pad_h, args.pad_w,
                               args.stride_h, args.stride_w,
@@ -97,7 +99,8 @@ public:
                               &queue_plain, &event);
       if (status == StatusCode::kSuccess) { clWaitForEvents(1, &event); clReleaseEvent(event); }
     #elif CUDA_API
-      auto status = Im2col<T>(args.channels, args.height, args.width,
+      auto status = Im2col<T>(args.kernel_mode,
+                              args.channels, args.height, args.width,
                               args.kernel_h, args.kernel_w,
                               args.pad_h, args.pad_w,
                               args.stride_h, args.stride_w,
@@ -175,7 +178,10 @@ StatusCode RunReference(const Arguments<T> &args, BuffersHost<T> &buffers_host)
             }
 
             // Sets the output value
-            const auto kernel_index = kw_id + args.kernel_w * kh_id;
+            const auto kernel_index
+                    = (args.kernel_mode == KernelMode::kConvolution)
+                    ? args.kernel_h * args.kernel_w - kw_id - args.kernel_w * kh_id - 1
+                    : kw_id + args.kernel_w * kh_id;
             const auto patch_index = w_id + col_w * h_id;
             const auto col_index = patch_index + kernel_index * col_w * col_h +
                                    c_id * col_w * col_h * args.kernel_h * args.kernel_w;