Add kernel_mode option to im2col, col2im, and convgemm functions

6 files changed, 25 insertions, 12 deletions

diff --git a/src/routines/levelx/xcol2im.cpp b/src/routines/levelx/xcol2im.cpp
index 7a0c36b7..d285e5c0 100644
--- a/src/routines/levelx/xcol2im.cpp
+++ b/src/routines/levelx/xcol2im.cpp
@@ -31,13 +31,17 @@ Xcol2im<T>::Xcol2im(Queue &queue, EventPointer event, const std::string &name):
 
 // The main routine
 template <typename T>
-void Xcol2im<T>::DoCol2im(const size_t channels, const size_t height, const size_t width,
+void Xcol2im<T>::DoCol2im(const KernelMode kernel_mode,
+                          const size_t channels, const size_t height, const size_t width,
                           const size_t kernel_h, const size_t kernel_w, const size_t pad_h,
                           const size_t pad_w, const size_t stride_h, const size_t stride_w,
                           const size_t dilation_h, const size_t dilation_w,
                           const Buffer<T> &col_buffer, const size_t col_offset,
                           const Buffer<T> &im_buffer, const size_t im_offset) {
 
+  // Flip the output along kernel_h and kernel_w, or not.
+  const auto kernel_name = (kernel_mode == KernelMode::kConvolution) ? "Xcol2imKernelFlip" : "Xcol2imKernelNormal";
+
   // Makes sure all dimensions are larger than zero
   if ((channels == 0) || (height == 0) || (width == 0)) { throw BLASError(StatusCode::kInvalidDimension); }
 
@@ -59,7 +63,7 @@ void Xcol2im<T>::DoCol2im(const size_t channels, const size_t height, const size
   EuclidGCD(static_cast<int>(stride_w), static_cast<int>(dilation_w), stride_bez_w, dilation_bez_w, gcd_w);
 
   // Retrieves the kernel from the compiled binary
-  auto kernel = Kernel(program_, "col2im");
+  auto kernel = Kernel(program_, kernel_name);
 
   // Sets the kernel arguments
   kernel.SetArgument(0, static_cast<int>(height));
diff --git a/src/routines/levelx/xcol2im.hpp b/src/routines/levelx/xcol2im.hpp
index 86d68c45..522c717e 100644
--- a/src/routines/levelx/xcol2im.hpp
+++ b/src/routines/levelx/xcol2im.hpp
@@ -29,7 +29,8 @@ class Xcol2im: public Routine {
   Xcol2im(Queue &queue, EventPointer event, const std::string &name = "COL2IM");
 
   // Templated-precision implementation of the routine
-  void DoCol2im(const size_t channels, const size_t height, const size_t width,
+  void DoCol2im(const KernelMode kernel_mode,
+                const size_t channels, const size_t height, const size_t width,
                 const size_t kernel_h, const size_t kernel_w,
                 const size_t pad_h, const size_t pad_w,
                 const size_t stride_h, const size_t stride_w,
diff --git a/src/routines/levelx/xconvgemm.cpp b/src/routines/levelx/xconvgemm.cpp
index f26f23a7..88127b0f 100644
--- a/src/routines/levelx/xconvgemm.cpp
+++ b/src/routines/levelx/xconvgemm.cpp
@@ -43,7 +43,8 @@ Xconvgemm<T>::Xconvgemm(Queue &queue, EventPointer event, const std::string &nam
 // =================================================================================================
 
 template <typename T>
-void Xconvgemm<T>::DoConvgemm(const size_t channels, const size_t height, const size_t width,
+void Xconvgemm<T>::DoConvgemm(const KernelMode kernel_mode,
+                              const size_t channels, const size_t height, const size_t width,
                               const size_t kernel_h, const size_t kernel_w, const size_t pad_h,
                               const size_t pad_w, const size_t stride_h, const size_t stride_w,
                               const size_t dilation_h, const size_t dilation_w,
@@ -94,7 +95,8 @@ void Xconvgemm<T>::DoConvgemm(const size_t channels, const size_t height, const
       const auto col_batch_offset = batch_id * patch_size * num_patches;
       auto im2col_event = Event();
       auto im2col = Xim2col<T>(queue_, im2col_event.pointer());
-      im2col.DoIm2col(channels, height, width, kernel_h, kernel_w,
+      im2col.DoIm2col(kernel_mode,
+                      channels, height, width, kernel_h, kernel_w,
                       pad_h, pad_w, stride_h, stride_w, dilation_h, dilation_w,
                       im_buffer, im_batch_offset,
                       col_buffer, col_batch_offset);
diff --git a/src/routines/levelx/xconvgemm.hpp b/src/routines/levelx/xconvgemm.hpp
index 9d11ccee..20cfff60 100644
--- a/src/routines/levelx/xconvgemm.hpp
+++ b/src/routines/levelx/xconvgemm.hpp
@@ -32,7 +32,8 @@ class Xconvgemm: public Routine {
             const ConvGemmMethod method = ConvGemmMethod::kWithIm2Col);
 
   // Templated-precision implementation of the routine
-  void DoConvgemm(const size_t channels, const size_t height, const size_t width,
+  void DoConvgemm(const KernelMode kernel_mode,
+                  const size_t channels, const size_t height, const size_t width,
                   const size_t kernel_h, const size_t kernel_w,
                   const size_t pad_h, const size_t pad_w,
                   const size_t stride_h, const size_t stride_w,
diff --git a/src/routines/levelx/xim2col.cpp b/src/routines/levelx/xim2col.cpp
index 09dcc42c..0f786974 100644
--- a/src/routines/levelx/xim2col.cpp
+++ b/src/routines/levelx/xim2col.cpp
@@ -22,22 +22,26 @@ namespace clblast {
 // Constructor: forwards to base class constructor
 template <typename T>
 Xim2col<T>::Xim2col(Queue &queue, EventPointer event, const std::string &name):
-        Routine(queue, event, name, {"Copy"}, PrecisionValue<T>(), {}, {
-#include "../../kernels/levelx/im2col.opencl"
-        }) {
+    Routine(queue, event, name, {"Copy"}, PrecisionValue<T>(), {}, {
+        #include "../../kernels/levelx/im2col.opencl"
+    }) {
 }
 
 // =================================================================================================
 
 // The main routine
 template <typename T>
-void Xim2col<T>::DoIm2col(const size_t channels, const size_t height, const size_t width,
+void Xim2col<T>::DoIm2col(const KernelMode kernel_mode,
+                          const size_t channels, const size_t height, const size_t width,
                           const size_t kernel_h, const size_t kernel_w, const size_t pad_h,
                           const size_t pad_w, const size_t stride_h, const size_t stride_w,
                           const size_t dilation_h, const size_t dilation_w,
                           const Buffer<T> &im_buffer, const size_t im_offset,
                           const Buffer<T> &col_buffer, const size_t col_offset) {
 
+  // Flip the output along kernel_h and kernel_w, or not.
+  const auto kernel_name = (kernel_mode == KernelMode::kConvolution) ? "Xim2colKernelFlip" : "Xim2colKernelNormal";
+
   // Makes sure all dimensions are larger than zero
   if ((channels == 0) || (height == 0) || (width == 0)) { throw BLASError(StatusCode::kInvalidDimension); }
 
@@ -50,7 +54,7 @@ void Xim2col<T>::DoIm2col(const size_t channels, const size_t height, const size
   const auto col_w = (size_w >= padding_w) ? (size_w - padding_w) / stride_w + 1 : 1;
 
   // Retrieves the kernel from the compiled binary
-  auto kernel = Kernel(program_, "im2col");
+  auto kernel = Kernel(program_, kernel_name);
 
   // Sets the kernel arguments
   kernel.SetArgument(0, static_cast<int>(height));
diff --git a/src/routines/levelx/xim2col.hpp b/src/routines/levelx/xim2col.hpp
index 2c03b169..77cc32eb 100644
--- a/src/routines/levelx/xim2col.hpp
+++ b/src/routines/levelx/xim2col.hpp
@@ -29,7 +29,8 @@ class Xim2col: public Routine {
   Xim2col(Queue &queue, EventPointer event, const std::string &name = "IM2COL");
 
   // Templated-precision implementation of the routine
-  void DoIm2col(const size_t channels, const size_t height, const size_t width,
+  void DoIm2col(const KernelMode kernel_mode,
+                const size_t channels, const size_t height, const size_t width,
                 const size_t kernel_h, const size_t kernel_w,
                 const size_t pad_h, const size_t pad_w,
                 const size_t stride_h, const size_t stride_w,