Merge pull request #343 from vbkaisetsu/feature/convgemm-single

Fix single kernel version of convgemm

11 files changed, 352 insertions, 37 deletions

diff --git a/CHANGELOG b/CHANGELOG
index 2241b855..4587940a 100644
--- a/CHANGELOG
+++ b/CHANGELOG
@@ -1,3 +1,6 @@
+Development (next version)
+- Implemented single-kernel version of convolution as GEMM
+- Various minor fixes and enhancements
 
 Version 1.5.0
 - Added support for shuffle instructions for NVIDIA GPUs (thanks to 'tyler-utah')
diff --git a/CMakeLists.txt b/CMakeLists.txt
index fc3101b7..1356a509 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -213,6 +213,7 @@ endif()
 # Sets the supported routines and the used kernels. New routines and kernels should be added here.
 set(KERNELS copy_fast copy_pad transpose_fast transpose_pad xaxpy xdot xger
             xgemm xgemm_direct xgemv invert)
+set(KERNELS_EXTRA xconvgemm)  # kernels for which not to include a tuner in 'all tuners' target
 set(DATABASES copy pad padtranspose transpose xaxpy xdot
               xgemm xgemm_direct xgemv xgemv_fast xgemv_fast_rot xger invert
               gemm_routine trsv_routine)
@@ -434,7 +435,8 @@ if(TUNERS)
   endif()
 
   # Adds tuning executables
-  foreach(KERNEL ${KERNELS})
+  set(ALLKERNELS ${KERNELS} ${KERNELS_EXTRA})
+  foreach(KERNEL ${ALLKERNELS})
     add_executable(clblast_tuner_${KERNEL} ${TUNERS_COMMON} src/tuning/kernels/${KERNEL}.cpp)
     target_link_libraries(clblast_tuner_${KERNEL} ${API_LIBRARIES})
     target_include_directories(clblast_tuner_${KERNEL} PUBLIC $<TARGET_PROPERTY:clblast,INTERFACE_INCLUDE_DIRECTORIES> ${API_INCLUDE_DIRS})
diff --git a/doc/routines.md b/doc/routines.md
index a4cb5e57..3ba8283e 100644
--- a/doc/routines.md
+++ b/doc/routines.md
@@ -94,7 +94,7 @@ In addition, some extra non-BLAS routines are also supported by CLBlast, classif
 | xOMATCOPY  | ✔ | ✔ | ✔ | ✔ | ✔ | (Out-of-place copying/transposing/scaling of matrices)
 | xIM2COL    | ✔ | ✔ | ✔ | ✔ | ✔ | (Image to column transform as used to express convolution as GEMM)
 | xCOL2IM    | ✔ | ✔ | ✔ | ✔ | ✔ | (Column to image transform as used in machine learning)
-| xCONVGEMM  | ✔ | ✔ | - | - | ✔ | (Experimental, implemented as im2col followed by batched GEMM)
+| xCONVGEMM  | ✔ | ✔ | - | - | ✔ | (Experimental, implemented as either im2col followed by batched GEMM or as a single kernel)
 
 Some less commonly used BLAS routines are not yet supported by CLBlast. They are xROTG, xROTMG, xROT, xROTM, xTBSV, and xTPSV.
 
diff --git a/src/kernels/levelx/xconvgemm_part1.opencl b/src/kernels/levelx/xconvgemm_part1.opencl
index abdb5324..25ccba51 100644
--- a/src/kernels/levelx/xconvgemm_part1.opencl
+++ b/src/kernels/levelx/xconvgemm_part1.opencl
@@ -11,7 +11,6 @@
 // uses parameters from the direct GEMM kernel. This is the part with the loads from memory (1/2).
 // This uses "CONVGEMM_WITH_IM2COL" as a switch to select between direct convgemm or first running
 // the im2col kernel to create a 'col' temporary matrix.
-// TODO: Currently only works with 'CONVGEMM_WITH_IM2COL' set
 //
 // =================================================================================================
 
@@ -30,12 +29,17 @@ INLINE_FUNC real GlobalToPrivateCheckedImage(const __global real* restrict image
                                              const int kernel_h, const int kernel_w,
                                              const int pad_h, const int pad_w,
                                              const int stride_h, const int stride_w,
-                                             const int dilation_h, const int dilation_w) {
+                                             const int dilation_h, const int dilation_w,
+                                             const bool kernel_flip) {
 
   // Im2col indices
   const int kernel_2d_index = kwg % (kernel_h * kernel_w);
-  const int kw_id = kernel_2d_index % kernel_w;
-  const int kh_id = kernel_2d_index / kernel_w;
+  const int kw_id = (kernel_flip)
+                  ? kernel_w - kernel_2d_index % kernel_w - 1
+                  : kernel_2d_index % kernel_w;
+  const int kh_id = (kernel_flip)
+                  ? kernel_h - kernel_2d_index / kernel_w - 1
+                  : kernel_2d_index / kernel_w;
   const int c_id = kwg / (kernel_h * kernel_w);
   const int h_index = -pad_h + kh_id * dilation_h + stride_h * h_id;
   const int w_index = -pad_w + kw_id * dilation_w + stride_w * w_id;
@@ -55,14 +59,15 @@ INLINE_FUNC real GlobalToPrivateCheckedImage(const __global real* restrict image
 
 // Loads global off-chip memory into local (shared) memory on-chip. This function is specific for
 // loading the image input tensor. This includes a bounds check.
-INLINE_FUNC real GlobalToLocalCheckedImage(const __global realMD* restrict imagegm, LOCAL_PTR real* alm,
+INLINE_FUNC real GlobalToLocalCheckedImage(const __global real* restrict imagegm, LOCAL_PTR real* alm,
                                            const int image_offset_batch,
-                                           const int h_id, const int w_id, const int kwg,
+                                           const int output_w, const int kwg,
                                            const int input_h, const int input_w, const int channels,
                                            const int kernel_h, const int kernel_w,
                                            const int pad_h, const int pad_w,
                                            const int stride_h, const int stride_w,
-                                           const int dilation_h, const int dilation_w) {
+                                           const int dilation_h, const int dilation_w,
+                                           const bool kernel_flip) {
   #if MDIMCD == MDIMAD
     const int la0 = get_local_id(0);
     const int la1 = get_local_id(1);
@@ -82,10 +87,17 @@ INLINE_FUNC real GlobalToLocalCheckedImage(const __global realMD* restrict image
       int idm = mg + GetGroupID0()*WGD;
       int idk = kg + kwg;
 
+      const int w_id = idm % output_w;
+      const int h_id = idm / output_w;
+
       // Im2col indices
       const int kernel_2d_index = idk % (kernel_h * kernel_w);
-      const int kw_id = kernel_2d_index % kernel_w;
-      const int kh_id = kernel_2d_index / kernel_w;
+      const int kw_id = (kernel_flip)
+                      ? kernel_w - kernel_2d_index % kernel_w - 1
+                      : kernel_2d_index % kernel_w;
+      const int kh_id = (kernel_flip)
+                      ? kernel_h - kernel_2d_index / kernel_w - 1
+                      : kernel_2d_index / kernel_w;
       const int c_id = idk / (kernel_h * kernel_w);
       const int h_index = -pad_h + kh_id * dilation_h + stride_h * h_id;
       const int w_index = -pad_w + kw_id * dilation_w + stride_w * w_id;
@@ -104,7 +116,8 @@ INLINE_FUNC real GlobalToLocalCheckedImage(const __global realMD* restrict image
   }
 }
 
-#endif
+#endif  // defined(ROUTINE_CONVGEMM) && !defined(CONVGEMM_WITH_IM2COL)
+
 // =================================================================================================
 
 // End of the C++11 raw string literal
diff --git a/src/kernels/levelx/xconvgemm_part2.opencl b/src/kernels/levelx/xconvgemm_part2.opencl
index e0ac24a0..693cb120 100644
--- a/src/kernels/levelx/xconvgemm_part2.opencl
+++ b/src/kernels/levelx/xconvgemm_part2.opencl
@@ -11,7 +11,6 @@
 // uses parameters from the direct GEMM kernel. This part contains the main kernel (2/2).
 // This uses "CONVGEMM_WITH_IM2COL" as a switch to select between direct convgemm or first running
 // the im2col kernel to create a 'col' temporary matrix.
-// TODO: Currently only works with 'CONVGEMM_WITH_IM2COL' set
 //
 // =================================================================================================
 
@@ -23,20 +22,25 @@ R"(
 #if defined(ROUTINE_CONVGEMM)
 
 // ConvGEMM kernel
+#if defined(CONVGEMM_WITH_IM2COL)
 __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
 void Xconvgemm(const int num_patches, const int num_kernels, const int patch_size,
                const __global realND* restrict kernelgm, const int kernel_offset,
                __global real* resultgm, const int result_offset, const int result_stride,
-#if defined(CONVGEMM_WITH_IM2COL)
                const __global realMD* restrict colgm, const int col_offset, const int col_stride)
 #else
-               const __global realMD* restrict imagegm, const int image_offset,
-               const int input_h, const int input_w, const int channels,
-               const int kernel_h, const int kernel_w,
-               const int pad_h, const int pad_w,
-               const int stride_h, const int stride_w,
-               const int dilation_h, const int dilation_w,
-               const int output_h, const int output_w)
+INLINE_FUNC void Xconvgemm(const int num_patches, const int num_kernels, const int patch_size,
+                           const __global realND* restrict kernelgm, const int kernel_offset,
+                           __global real* resultgm, const int result_offset, const int result_stride,
+                           const __global realMD* restrict imagegm, const int image_offset,
+                           const int input_h, const int input_w, const int channels,
+                           const int kernel_h, const int kernel_w,
+                           const int pad_h, const int pad_w,
+                           const int stride_h, const int stride_w,
+                           const int dilation_h, const int dilation_w,
+                           const int output_h, const int output_w,
+                           LOCAL_PTR real* alm, LOCAL_PTR real* blm,
+                           const bool kernel_flip)
 #endif
 {
 
@@ -49,12 +53,16 @@ void Xconvgemm(const int num_patches, const int num_kernels, const int patch_siz
   #endif
   const int result_offset_batch = result_offset + result_stride * batch;
 
+#if defined(CONVGEMM_WITH_IM2COL)
   __local real alm[WGD * (WGD + PADA)];
   __local real blm[WGD * (WGD + PADB)];
+#endif
 
   // Extra pointers to scalar versions of global memory
   #if defined(CONVGEMM_WITH_IM2COL)
     const __global real* restrict colgms = (const __global real* restrict) colgm;
+  #else
+    const __global real* restrict imagegms = (const __global real* restrict) imagegm;
   #endif
   const __global real* restrict kernelgms = (const __global real* restrict) kernelgm;
 
@@ -100,10 +108,10 @@ void Xconvgemm(const int num_patches, const int num_kernels, const int patch_siz
           GlobalToLocalScalarA(colgms, alm, num_patches, col_offset_batch, kwg, false, false);
         }
       #else
-        GlobalToLocalCheckedImage(imagegm, alm, image_offset_batch, h_id, w_id, kwg,
+        GlobalToLocalCheckedImage(imagegms, alm, image_offset_batch, output_w, kwg,
                                   input_h, input_w, channels, kernel_h, kernel_w,
                                   pad_h, pad_w, stride_h, stride_w,
-                                  dilation_h, dilation_w);
+                                  dilation_h, dilation_w, kernel_flip);
       #endif
       if (patch_size % VWND == 0 && kernel_offset % VWND == 0) {
         GlobalToLocalDirectB(kernelgm, blm, patch_size, kernel_offset, kwg, true, false);
@@ -151,10 +159,12 @@ void Xconvgemm(const int num_patches, const int num_kernels, const int patch_siz
         #if defined(CONVGEMM_WITH_IM2COL)
           apd[_mi] = GlobalToPrivateDirectA(colgms, _mi, num_patches, col_offset_batch, idm, kwg, false, false);
         #else
-          apd[_mi] = GlobalToPrivateCheckedImage(imagegm, image_offset_batch, h_id, w_id, kwg,
+          const int w_id = (idm + _mi) % output_w;
+          const int h_id = (idm + _mi) / output_w;
+          apd[_mi] = GlobalToPrivateCheckedImage(imagegms, image_offset_batch, h_id, w_id, kwg,
                                                  input_h, input_w, channels, kernel_h, kernel_w,
                                                  pad_h, pad_w, stride_h, stride_w,
-                                                 dilation_h, dilation_w);
+                                                 dilation_h, dilation_w, kernel_flip);
         #endif
       }
       #pragma unroll
@@ -193,10 +203,10 @@ void Xconvgemm(const int num_patches, const int num_kernels, const int patch_siz
       #if defined(CONVGEMM_WITH_IM2COL)
         GlobalToLocalCheckedA(colgms, alm, num_patches, col_offset_batch, kwg, false, false, num_patches, patch_size);
       #else
-        GlobalToLocalCheckedImage(imagegm, alm, image_offset_batch, h_id, w_id, kwg,
+        GlobalToLocalCheckedImage(imagegms, alm, image_offset_batch, output_w, kwg,
                                   input_h, input_w, channels, kernel_h, kernel_w,
                                   pad_h, pad_w, stride_h, stride_w,
-                                  dilation_h, dilation_w);
+                                  dilation_h, dilation_w, kernel_flip);
       #endif
       GlobalToLocalCheckedB(kernelgms, blm, patch_size, kernel_offset, kwg, true, false, num_kernels, patch_size);
       barrier(CLK_LOCAL_MEM_FENCE);
@@ -239,10 +249,12 @@ void Xconvgemm(const int num_patches, const int num_kernels, const int patch_siz
       #if defined(CONVGEMM_WITH_IM2COL)
         apd[_mi] = GlobalToPrivateCheckedA(colgms, _mi, num_patches, col_offset_batch, idm, kwg, false, false, num_patches);
       #else
-        apd[_mi] = GlobalToPrivateCheckedImage(imagegm, image_offset_batch, h_id, w_id, kwg,
+        const int w_id = (idm + _mi) % output_w;
+        const int h_id = (idm + _mi) / output_w;
+        apd[_mi] = GlobalToPrivateCheckedImage(imagegms, image_offset_batch, h_id, w_id, kwg,
                                                input_h, input_w, channels, kernel_h, kernel_w,
                                                pad_h, pad_w, stride_h, stride_w,
-                                               dilation_h, dilation_w);
+                                               dilation_h, dilation_w, kernel_flip);
       #endif
       }
       #pragma unroll
@@ -272,7 +284,53 @@ void Xconvgemm(const int num_patches, const int num_kernels, const int patch_siz
   }
 }
 
-#endif
+#if !defined(CONVGEMM_WITH_IM2COL)
+__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
+void XconvgemmFlip(const int num_patches, const int num_kernels, const int patch_size,
+                   const __global realND* restrict kernelgm, const int kernel_offset,
+                   __global real* resultgm, const int result_offset, const int result_stride,
+                   const __global realMD* restrict imagegm, const int image_offset,
+                   const int input_h, const int input_w, const int channels,
+                   const int kernel_h, const int kernel_w,
+                   const int pad_h, const int pad_w,
+                   const int stride_h, const int stride_w,
+                   const int dilation_h, const int dilation_w,
+                   const int output_h, const int output_w) {
+  const bool kernel_flip = true;
+  __local real alm[WGD * (WGD + PADA)];
+  __local real blm[WGD * (WGD + PADB)];
+  Xconvgemm(num_patches, num_kernels, patch_size,
+            kernelgm, kernel_offset, resultgm, result_offset, result_stride,
+            imagegm, image_offset, input_h, input_w, channels, kernel_h, kernel_w,
+            pad_h, pad_w, stride_h, stride_w, dilation_h, dilation_w,
+            output_h, output_w, alm, blm, kernel_flip);
+}
+
+__kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
+void XconvgemmNormal(const int num_patches, const int num_kernels, const int patch_size,
+                     const __global realND* restrict kernelgm, const int kernel_offset,
+                     __global real* resultgm, const int result_offset, const int result_stride,
+                     const __global realMD* restrict imagegm, const int image_offset,
+                     const int input_h, const int input_w, const int channels,
+                     const int kernel_h, const int kernel_w,
+                     const int pad_h, const int pad_w,
+                     const int stride_h, const int stride_w,
+                     const int dilation_h, const int dilation_w,
+                     const int output_h, const int output_w) {
+  const bool kernel_flip = false;
+  __local real alm[WGD * (WGD + PADA)];
+  __local real blm[WGD * (WGD + PADB)];
+  Xconvgemm(num_patches, num_kernels, patch_size,
+            kernelgm, kernel_offset, resultgm, result_offset, result_stride,
+            imagegm, image_offset, input_h, input_w, channels, kernel_h, kernel_w,
+            pad_h, pad_w, stride_h, stride_w, dilation_h, dilation_w,
+            output_h, output_w, alm, blm, kernel_flip);
+}
+
+#endif  // !defined(CONVGEMM_WITH_IM2COL)
+
+#endif  // defined(ROUTINE_CONVGEMM)
+
 // =================================================================================================
 
 // End of the C++11 raw string literal
diff --git a/src/routines/levelx/xconvgemm.cpp b/src/routines/levelx/xconvgemm.cpp
index 88127b0f..8bd24f15 100644
--- a/src/routines/levelx/xconvgemm.cpp
+++ b/src/routines/levelx/xconvgemm.cpp
@@ -53,9 +53,6 @@ void Xconvgemm<T>::DoConvgemm(const KernelMode kernel_mode,
                               const Buffer<T> &kernel_buffer, const size_t kernel_offset,
                               const Buffer<T> &result_buffer, const size_t result_offset) {
 
-  // TODO: Implement single-kernel approach
-  assert(method_ == ConvGemmMethod::kWithIm2Col);
-
   // Tests for a valid batch count
   if (batch_count == 0) {
     throw BLASError(StatusCode::kInvalidBatchCount);
@@ -121,7 +118,12 @@ void Xconvgemm<T>::DoConvgemm(const KernelMode kernel_mode,
   }
 
   // Retrieves the proper XgemmDirect kernel from the compiled binary
-  auto kernel = Kernel(program_, "Xconvgemm");
+  const std::string kernel_name = (method_ == ConvGemmMethod::kWithIm2Col)
+                                ? "Xconvgemm"
+                                : (kernel_mode == KernelMode::kConvolution)
+                                ? "XconvgemmFlip"
+                                : "XconvgemmNormal";
+  auto kernel = Kernel(program_, kernel_name);
 
   // Sets the kernel arguments
   kernel.SetArgument(0, static_cast<int>(num_patches));
diff --git a/src/routines/levelx/xconvgemm.hpp b/src/routines/levelx/xconvgemm.hpp
index 20cfff60..16082fc6 100644
--- a/src/routines/levelx/xconvgemm.hpp
+++ b/src/routines/levelx/xconvgemm.hpp
@@ -29,7 +29,7 @@ class Xconvgemm: public Routine {
   // Constructor
   enum class ConvGemmMethod {kWithIm2Col, kSingleKernel};
   Xconvgemm(Queue &queue, EventPointer event, const std::string &name = "CONVGEMM",
-            const ConvGemmMethod method = ConvGemmMethod::kWithIm2Col);
+            const ConvGemmMethod method = ConvGemmMethod::kSingleKernel);
 
   // Templated-precision implementation of the routine
   void DoConvgemm(const KernelMode kernel_mode,
diff --git a/src/tuning/kernels/xconvgemm.cpp b/src/tuning/kernels/xconvgemm.cpp
new file mode 100644
index 00000000..15dfe829
--- /dev/null
+++ b/src/tuning/kernels/xconvgemm.cpp
@@ -0,0 +1,38 @@
+
+// =================================================================================================
+// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This
+// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max-
+// width of 100 characters per line.
+//
+// Author(s):
+//   Cedric Nugteren <www.cedricnugteren.nl>
+//
+// This file uses the auto-tuner to tune the convgemm kernels.
+//
+// =================================================================================================
+
+#include "tuning/kernels/xconvgemm.hpp"
+
+// Shortcuts to the clblast namespace
+using half = clblast::half;
+using float2 = clblast::float2;
+using double2 = clblast::double2;
+
+// Function to tune a specific variation V (not within the clblast namespace)
+template <int V>
+void StartVariation(int argc, char *argv[]) {
+  const auto command_line_args = clblast::RetrieveCommandLineArguments(argc, argv);
+  switch(clblast::GetPrecision(command_line_args)) {
+    case clblast::Precision::kHalf: clblast::Tuner<half>(argc, argv, V, clblast::XConvGemmGetTunerDefaults, clblast::XConvGemmGetTunerSettings<half>, clblast::XConvGemmTestValidArguments<half>, clblast::XConvGemmSetConstraints, clblast::XConvGemmComputeLocalMemSize<half>, clblast::XConvGemmSetArguments<half>); break;
+    case clblast::Precision::kSingle: clblast::Tuner<float>(argc, argv, V, clblast::XConvGemmGetTunerDefaults, clblast::XConvGemmGetTunerSettings<float>, clblast::XConvGemmTestValidArguments<float>, clblast::XConvGemmSetConstraints, clblast::XConvGemmComputeLocalMemSize<float>, clblast::XConvGemmSetArguments<float>); break;
+    case clblast::Precision::kDouble: clblast::Tuner<double>(argc, argv, V, clblast::XConvGemmGetTunerDefaults, clblast::XConvGemmGetTunerSettings<double>, clblast::XConvGemmTestValidArguments<double>, clblast::XConvGemmSetConstraints, clblast::XConvGemmComputeLocalMemSize<double>, clblast::XConvGemmSetArguments<double>); break;
+  }
+}
+
+// Main function (not within the clblast namespace)
+int main(int argc, char *argv[]) {
+  StartVariation<1>(argc, argv);
+  return 0;
+}
+
+// =================================================================================================
diff --git a/src/tuning/kernels/xconvgemm.hpp b/src/tuning/kernels/xconvgemm.hpp
new file mode 100644
index 00000000..9ba70f5e
--- /dev/null
+++ b/src/tuning/kernels/xconvgemm.hpp
@@ -0,0 +1,186 @@
+
+// =================================================================================================
+// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This
+// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max-
+// width of 100 characters per line.
+//
+// Author(s):
+//   Cedric Nugteren <www.cedricnugteren.nl>
+//
+// This file uses the auto-tuner to tune the ConvGemm kernels. These kernels are based on the GEMM
+// direct kernel and will use those parameters, this tuner is just optional to use for advanced
+// users.
+//
+// =================================================================================================
+
+#include <string>
+#include <vector>
+
+#include "utilities/utilities.hpp"
+#include "tuning/tuning.hpp"
+
+namespace clblast {
+// =================================================================================================
+
+// Helper functions
+template <typename T>
+size_t OutputHeight(const Arguments<T> &args) {
+  const auto size = args.height + 2 * args.pad_h;
+  const auto padding = args.dilation_h * (args.kernel_h - 1) + 1;
+  if (size >= padding) { return (size - padding) / args.stride_h + 1; }
+  return 1;
+}
+template <typename T>
+size_t OutputWidth(const Arguments<T> &args) {
+  const auto size = args.width + 2 * args.pad_w;
+  const auto padding = args.dilation_w * (args.kernel_w - 1) + 1;
+  if (size >= padding) { return (size - padding) / args.stride_w + 1; }
+  return 1;
+}
+
+// Settings for this kernel (default command-line arguments)
+TunerDefaults XConvGemmGetTunerDefaults(const int) {
+  auto settings = TunerDefaults();
+  settings.options = {kArgChannels, kArgHeight, kArgWidth, kArgKernelH, kArgKernelW,
+                      kArgNumKernels, kArgBatchCount, kArgFraction};
+  settings.channels = 32;
+  settings.height = 66;
+  settings.width = 66;  // num_patches = 64x64 = 4096
+  settings.kernel_h = 3;
+  settings.kernel_w = 3;
+  settings.num_kernels = 32;
+  settings.default_batch_count = 16;
+  settings.default_fraction = 1.0;
+  settings.default_num_runs = 2;
+  return settings;
+}
+
+// Settings for this kernel (general)
+template <typename T>
+TunerSettings XConvGemmGetTunerSettings(const int, const Arguments<T> &args) {
+  auto settings = TunerSettings();
+
+  // Identification of the kernel
+  settings.kernel_family = "xconvgemm";
+  settings.kernel_name = "XconvgemmNormal";
+  settings.sources =
+"#define ROUTINE_CONVGEMM"
+#include "../src/kernels/level3/xgemm_direct_part1.opencl"
+#include "../src/kernels/level3/xgemm_direct_part2.opencl"
+#include "../src/kernels/level3/xgemm_direct_part3.opencl"
+#include "../src/kernels/levelx/xconvgemm_part1.opencl"
+#include "../src/kernels/levelx/xconvgemm_part2.opencl"
+  ;
+
+  // Helper variables
+  const auto patch_size = args.kernel_h * args.kernel_w * args.channels;
+  const auto num_patches = OutputHeight(args) * OutputWidth(args);
+
+  // Buffer sizes
+  settings.size_a = args.batch_count * args.channels * args.height * args.width;
+  settings.size_b = args.num_kernels * args.channels * args.kernel_h * args.kernel_w;
+  settings.size_c = args.batch_count * args.num_kernels * OutputHeight(args) * OutputWidth(args);
+
+  // Inputs and outputs IDs (X:0, Y:1, A:2, B:3, C:4, temp:5)
+  settings.inputs = {2, 3, 4};
+  settings.outputs = {4};
+
+  // Sets the base thread configuration
+  settings.global_size = {num_patches, args.num_kernels};
+  settings.global_size_ref = settings.global_size;
+  settings.local_size = {1, 1};
+  settings.local_size_ref = {8, 8};
+
+  // Transforms the thread configuration based on the parameters
+  settings.mul_local = {{"MDIMCD", "NDIMCD"}};
+  settings.mul_global = {{"MDIMCD", "NDIMCD"}};
+  settings.div_global = {{"WGD", "WGD"}};
+
+  // Sets the tuning parameters and their possible values
+  settings.parameters = {
+    {"WGD", {8, 16, 32}},
+    {"MDIMCD", {8, 16, 32}},
+    {"NDIMCD", {8, 16, 32}},
+    {"MDIMAD", {8, 16, 32}},
+    {"NDIMBD", {8, 16, 32}},
+    {"KWID", {1}},
+    {"VWMD", {1, 2, 4, 8}},
+    {"VWND", {1, 2, 4, 8}},
+    {"PADA", {0}},
+    {"PADB", {0}},
+  };
+
+  // Describes how to compute the performance metrics
+  settings.metric_amount = args.batch_count * 2 * num_patches * args.num_kernels * patch_size;
+  settings.performance_unit = "GFLOPS";
+
+  return settings;
+}
+
+// Tests for valid arguments
+template <typename T>
+void XConvGemmTestValidArguments(const int, const Arguments<T> &) { }
+std::vector<Constraint> XConvGemmSetConstraints(const int) {
+  auto constraints = std::vector<Constraint>();
+  auto MultipleOfX = [] (std::vector<size_t> v) { return IsMultiple(v[0], v[1]); };
+  auto MultipleOfXMulY = [] (std::vector<size_t> v) { return IsMultiple(v[0], v[1]*v[2]); };
+  auto MultipleOfXMulYDivZ = [] (std::vector<size_t> v) { return IsMultiple(v[0], (v[1]*v[2])/v[3]); };
+  // Requirement for unrolling the WGD loop
+  constraints.push_back({MultipleOfX, {"WGD", "KWID"}});
+  // Required for integer MWID and NWID
+  constraints.push_back({MultipleOfXMulY, {"WGD", "MDIMCD", "VWMD"}});
+  constraints.push_back({MultipleOfXMulY, {"WGD", "NDIMCD", "VWND"}});
+  // Required for integer MWIAD and NWIBD
+  constraints.push_back({MultipleOfXMulY, {"WGD", "MDIMAD", "VWMD"}});
+  constraints.push_back({MultipleOfXMulY, {"WGD", "NDIMBD", "VWND"}});
+  // WGD has to be a multiple of KDIMAD = ((MDIMCD*NDIMCD)/(MDIMAD)) and KDIMBD = (...)
+  constraints.push_back({MultipleOfXMulYDivZ, {"WGD", "MDIMCD", "NDIMCD", "MDIMAD"}});
+  constraints.push_back({MultipleOfXMulYDivZ, {"WGD", "MDIMCD", "NDIMCD", "NDIMBD"}});
+
+  return constraints;
+}
+template <typename T>
+LocalMemSizeInfo XConvGemmComputeLocalMemSize(const int) {
+  return {
+      [] (std::vector<size_t> v) -> size_t {
+          return GetBytes(PrecisionValue<T>()) * ((v[0]*(v[0] + v[1]) + v[0]*(v[0] + v[2])));
+      },
+      {"WGD", "PADA", "PADB"}
+  };
+}
+
+// Sets the kernel's arguments
+template <typename T>
+void XConvGemmSetArguments(const int, Kernel &kernel, const Arguments<T> &args, std::vector<Buffer<T>>& buffers) {
+  const auto output_h = OutputHeight(args);
+  const auto output_w = OutputWidth(args);
+  const auto patch_size = args.kernel_h * args.kernel_w * args.channels;
+  const auto num_patches = output_h * output_w;
+  const auto result_stride = args.num_kernels * output_h * output_w;
+  kernel.SetArgument(0, static_cast<int>(num_patches));
+  kernel.SetArgument(1, static_cast<int>(args.num_kernels));
+  kernel.SetArgument(2, static_cast<int>(patch_size));
+  kernel.SetArgument(3, buffers[3]()); // 3 == B matrix ==> kernel buffer
+  kernel.SetArgument(4, 0); // c_offset
+  kernel.SetArgument(5, buffers[4]()); // 4 == C matrix ==> result buffer
+  kernel.SetArgument(6, 0); // c_offset
+  kernel.SetArgument(7, static_cast<int>(result_stride));
+  kernel.SetArgument(8, buffers[2]()); // 2 == A matrix ==> image buffer
+  kernel.SetArgument(9, 0); // c_offset
+  kernel.SetArgument(10, static_cast<int>(args.height));
+  kernel.SetArgument(11, static_cast<int>(args.width));
+  kernel.SetArgument(12, static_cast<int>(args.channels));
+  kernel.SetArgument(13, static_cast<int>(args.kernel_h));
+  kernel.SetArgument(14, static_cast<int>(args.kernel_w));
+  kernel.SetArgument(15, 0); // pad_h
+  kernel.SetArgument(16, 0); // pad_w
+  kernel.SetArgument(17, 1); // stride_h
+  kernel.SetArgument(18, 1); // stride_w
+  kernel.SetArgument(19, 1); // dilation_h
+  kernel.SetArgument(20, 1); // dilation_w
+  kernel.SetArgument(21, static_cast<int>(output_h));
+  kernel.SetArgument(22, static_cast<int>(output_w));
+}
+
+// =================================================================================================
+} // namespace clblast
diff --git a/src/tuning/tuning.cpp b/src/tuning/tuning.cpp
index d382fb18..f76af774 100644
--- a/src/tuning/tuning.cpp
+++ b/src/tuning/tuning.cpp
@@ -122,8 +122,14 @@ void Tuner(int argc, char* argv[], const int V,
     if (o == kArgM)        { args.m        = GetArgument(command_line_args, help, kArgM, defaults.default_m); }
     if (o == kArgN)        { args.n        = GetArgument(command_line_args, help, kArgN, defaults.default_n); }
     if (o == kArgK)        { args.k        = GetArgument(command_line_args, help, kArgK, defaults.default_k); }
-    if (o == kArgAlpha)    { args.alpha    = GetArgument(command_line_args, help, kArgAlpha, GetScalar<T>()); }
-    if (o == kArgBeta)     { args.beta     = GetArgument(command_line_args, help, kArgBeta, GetScalar<T>()); }
+    if (o == kArgChannels)   { args.channels    = GetArgument(command_line_args, help, kArgChannels, defaults.channels); }
+    if (o == kArgHeight)     { args.height      = GetArgument(command_line_args, help, kArgHeight, defaults.height); }
+    if (o == kArgWidth)      { args.width       = GetArgument(command_line_args, help, kArgWidth, defaults.width); }
+    if (o == kArgKernelH)    { args.kernel_h    = GetArgument(command_line_args, help, kArgKernelH, defaults.kernel_h); }
+    if (o == kArgKernelW)    { args.kernel_w    = GetArgument(command_line_args, help, kArgKernelW, defaults.kernel_w); }
+    if (o == kArgNumKernels) { args.num_kernels = GetArgument(command_line_args, help, kArgNumKernels, defaults.num_kernels); }
+    if (o == kArgAlpha)      { args.alpha       = GetArgument(command_line_args, help, kArgAlpha, GetScalar<T>()); }
+    if (o == kArgBeta)       { args.beta        = GetArgument(command_line_args, help, kArgBeta, GetScalar<T>()); }
     if (o == kArgBatchCount) { args.batch_count = GetArgument(command_line_args, help, kArgBatchCount, defaults.default_batch_count); }
   }
   args.fraction = GetArgument(command_line_args, help, kArgFraction, defaults.default_fraction);
diff --git a/src/tuning/tuning.hpp b/src/tuning/tuning.hpp
index 37a042ff..61cc3bda 100644
--- a/src/tuning/tuning.hpp
+++ b/src/tuning/tuning.hpp
@@ -41,6 +41,13 @@ struct TunerDefaults {
   size_t default_m = 1;
   size_t default_n = 1;
   size_t default_k = 1;
+  size_t channels = 1;
+  size_t height = 1;
+  size_t width = 1;
+  size_t kernel_h = 3;
+  size_t kernel_w = 3;
+  size_t num_kernels = 1;
+  size_t batch_count = 1;
 
   // Other defaults
   size_t default_batch_count = 1;