Added first naive version of the batched AXPY routine

4 files changed, 139 insertions, 34 deletions

diff --git a/src/clblast.cpp b/src/clblast.cpp
index 55562419..e9cac664 100644
--- a/src/clblast.cpp
+++ b/src/clblast.cpp
@@ -2178,8 +2178,8 @@ template StatusCode PUBLIC_API Omatcopy<half>(const Layout, const Transpose,
 template <typename T>
 StatusCode AxpyBatched(const size_t n,
                        const T *alphas,
-                       const cl_mem *x_buffers, const size_t x_offset, const size_t x_inc,
-                       cl_mem *y_buffers, const size_t y_offset, const size_t y_inc,
+                       const cl_mem *x_buffers, const size_t x_inc,
+                       cl_mem *y_buffers, const size_t y_inc,
                        const size_t batch_count,
                        cl_command_queue* queue, cl_event* event) {
   try {
@@ -2195,40 +2195,40 @@ StatusCode AxpyBatched(const size_t n,
     }
     routine.DoAxpyBatched(n,
                           alphas_cpp,
-                          x_buffers_cpp, x_offset, x_inc,
-                          y_buffers_cpp, y_offset, y_inc,
+                          x_buffers_cpp, x_inc,
+                          y_buffers_cpp, y_inc,
                           batch_count);
     return StatusCode::kSuccess;
   } catch (...) { return DispatchException(); }
 }
 template StatusCode PUBLIC_API AxpyBatched<float>(const size_t,
                                                   const float*,
-                                                  const cl_mem*, const size_t, const size_t,
-                                                  cl_mem*, const size_t, const size_t,
+                                                  const cl_mem*, const size_t,
+                                                  cl_mem*, const size_t,
                                                   const size_t,
                                                   cl_command_queue*, cl_event*);
 template StatusCode PUBLIC_API AxpyBatched<double>(const size_t,
                                                    const double*,
-                                                   const cl_mem*, const size_t, const size_t,
-                                                   cl_mem*, const size_t, const size_t,
+                                                   const cl_mem*, const size_t,
+                                                   cl_mem*, const size_t,
                                                    const size_t,
                                                    cl_command_queue*, cl_event*);
 template StatusCode PUBLIC_API AxpyBatched<float2>(const size_t,
                                                    const float2*,
-                                                   const cl_mem*, const size_t, const size_t,
-                                                   cl_mem*, const size_t, const size_t,
+                                                   const cl_mem*, const size_t,
+                                                   cl_mem*, const size_t,
                                                    const size_t,
                                                    cl_command_queue*, cl_event*);
 template StatusCode PUBLIC_API AxpyBatched<double2>(const size_t,
                                                     const double2*,
-                                                    const cl_mem*, const size_t, const size_t,
-                                                    cl_mem*, const size_t, const size_t,
+                                                    const cl_mem*, const size_t,
+                                                    cl_mem*, const size_t,
                                                     const size_t,
                                                     cl_command_queue*, cl_event*);
 template StatusCode PUBLIC_API AxpyBatched<half>(const size_t,
                                                  const half*,
-                                                 const cl_mem*, const size_t, const size_t,
-                                                 cl_mem*, const size_t, const size_t,
+                                                 const cl_mem*, const size_t,
+                                                 cl_mem*, const size_t,
                                                  const size_t,
                                                  cl_command_queue*, cl_event*);
 // =================================================================================================
diff --git a/src/clblast_c.cpp b/src/clblast_c.cpp
index 83450e6f..bd8ea51a 100644
--- a/src/clblast_c.cpp
+++ b/src/clblast_c.cpp
@@ -3450,8 +3450,8 @@ CLBlastStatusCode CLBlastHomatcopy(const CLBlastLayout layout, const CLBlastTran
 // AXPY
 CLBlastStatusCode CLBlastSaxpyBatched(const size_t n,
                                       const float *alphas,
-                                      const cl_mem *x_buffers, const size_t x_offset, const size_t x_inc,
-                                      cl_mem *y_buffers, const size_t y_offset, const size_t y_inc,
+                                      const cl_mem *x_buffers, const size_t x_inc,
+                                      cl_mem *y_buffers, const size_t y_inc,
                                       const size_t batch_count,
                                       cl_command_queue* queue, cl_event* event) {
   auto alphas_cpp = std::vector<float>();
@@ -3462,8 +3462,8 @@ CLBlastStatusCode CLBlastSaxpyBatched(const size_t n,
     return static_cast<CLBlastStatusCode>(
       clblast::AxpyBatched(n,
                            alphas_cpp.data(),
-                           x_buffers, x_offset, x_inc,
-                           y_buffers, y_offset, y_inc,
+                           x_buffers, x_inc,
+                           y_buffers, y_inc,
                            batch_count,
                            queue, event)
     );
@@ -3471,8 +3471,8 @@ CLBlastStatusCode CLBlastSaxpyBatched(const size_t n,
 }
 CLBlastStatusCode CLBlastDaxpyBatched(const size_t n,
                                       const double *alphas,
-                                      const cl_mem *x_buffers, const size_t x_offset, const size_t x_inc,
-                                      cl_mem *y_buffers, const size_t y_offset, const size_t y_inc,
+                                      const cl_mem *x_buffers, const size_t x_inc,
+                                      cl_mem *y_buffers, const size_t y_inc,
                                       const size_t batch_count,
                                       cl_command_queue* queue, cl_event* event) {
   auto alphas_cpp = std::vector<double>();
@@ -3483,8 +3483,8 @@ CLBlastStatusCode CLBlastDaxpyBatched(const size_t n,
     return static_cast<CLBlastStatusCode>(
       clblast::AxpyBatched(n,
                            alphas_cpp.data(),
-                           x_buffers, x_offset, x_inc,
-                           y_buffers, y_offset, y_inc,
+                           x_buffers, x_inc,
+                           y_buffers, y_inc,
                            batch_count,
                            queue, event)
     );
@@ -3492,8 +3492,8 @@ CLBlastStatusCode CLBlastDaxpyBatched(const size_t n,
 }
 CLBlastStatusCode CLBlastCaxpyBatched(const size_t n,
                                       const cl_float2 *alphas,
-                                      const cl_mem *x_buffers, const size_t x_offset, const size_t x_inc,
-                                      cl_mem *y_buffers, const size_t y_offset, const size_t y_inc,
+                                      const cl_mem *x_buffers, const size_t x_inc,
+                                      cl_mem *y_buffers, const size_t y_inc,
                                       const size_t batch_count,
                                       cl_command_queue* queue, cl_event* event) {
   auto alphas_cpp = std::vector<float2>();
@@ -3504,8 +3504,8 @@ CLBlastStatusCode CLBlastCaxpyBatched(const size_t n,
     return static_cast<CLBlastStatusCode>(
       clblast::AxpyBatched(n,
                            alphas_cpp.data(),
-                           x_buffers, x_offset, x_inc,
-                           y_buffers, y_offset, y_inc,
+                           x_buffers, x_inc,
+                           y_buffers, y_inc,
                            batch_count,
                            queue, event)
     );
@@ -3513,8 +3513,8 @@ CLBlastStatusCode CLBlastCaxpyBatched(const size_t n,
 }
 CLBlastStatusCode CLBlastZaxpyBatched(const size_t n,
                                       const cl_double2 *alphas,
-                                      const cl_mem *x_buffers, const size_t x_offset, const size_t x_inc,
-                                      cl_mem *y_buffers, const size_t y_offset, const size_t y_inc,
+                                      const cl_mem *x_buffers, const size_t x_inc,
+                                      cl_mem *y_buffers, const size_t y_inc,
                                       const size_t batch_count,
                                       cl_command_queue* queue, cl_event* event) {
   auto alphas_cpp = std::vector<double2>();
@@ -3525,8 +3525,8 @@ CLBlastStatusCode CLBlastZaxpyBatched(const size_t n,
     return static_cast<CLBlastStatusCode>(
       clblast::AxpyBatched(n,
                            alphas_cpp.data(),
-                           x_buffers, x_offset, x_inc,
-                           y_buffers, y_offset, y_inc,
+                           x_buffers, x_inc,
+                           y_buffers, y_inc,
                            batch_count,
                            queue, event)
     );
@@ -3534,8 +3534,8 @@ CLBlastStatusCode CLBlastZaxpyBatched(const size_t n,
 }
 CLBlastStatusCode CLBlastHaxpyBatched(const size_t n,
                                       const cl_half *alphas,
-                                      const cl_mem *x_buffers, const size_t x_offset, const size_t x_inc,
-                                      cl_mem *y_buffers, const size_t y_offset, const size_t y_inc,
+                                      const cl_mem *x_buffers, const size_t x_inc,
+                                      cl_mem *y_buffers, const size_t y_inc,
                                       const size_t batch_count,
                                       cl_command_queue* queue, cl_event* event) {
   auto alphas_cpp = std::vector<half>();
@@ -3546,8 +3546,8 @@ CLBlastStatusCode CLBlastHaxpyBatched(const size_t n,
     return static_cast<CLBlastStatusCode>(
       clblast::AxpyBatched(n,
                            alphas_cpp.data(),
-                           x_buffers, x_offset, x_inc,
-                           y_buffers, y_offset, y_inc,
+                           x_buffers, x_inc,
+                           y_buffers, y_inc,
                            batch_count,
                            queue, event)
     );
diff --git a/src/routines/levelx/xaxpybatched.cpp b/src/routines/levelx/xaxpybatched.cpp
new file mode 100644
index 00000000..55458f43
--- /dev/null
+++ b/src/routines/levelx/xaxpybatched.cpp
@@ -0,0 +1,59 @@
+
+// =================================================================================================
+// 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 implements the XaxpyBatched class (see the header for information about the class).
+//
+// =================================================================================================
+
+#include "routines/levelx/xaxpybatched.hpp"
+
+#include <string>
+#include <vector>
+
+namespace clblast {
+// =================================================================================================
+
+// Constructor: forwards to base class constructor
+template <typename T>
+XaxpyBatched<T>::XaxpyBatched(Queue &queue, EventPointer event, const std::string &name):
+    Xaxpy<T>(queue, event, name) {
+}
+
+// =================================================================================================
+
+// The main routine
+template <typename T>
+void XaxpyBatched<T>::DoAxpyBatched(const size_t n, const std::vector<T> &alphas,
+                                    const std::vector<Buffer<T>> &x_buffers, const size_t x_inc,
+                                    const std::vector<Buffer<T>> &y_buffers, const size_t y_inc,
+                                    const size_t batch_count) {
+  if (batch_count < 1) { throw BLASError(StatusCode::kInvalidBatchCount); }
+  if (alphas.size() != batch_count) { throw BLASError(StatusCode::kInvalidBatchCount); }
+  if (x_buffers.size() != batch_count) { throw BLASError(StatusCode::kInvalidBatchCount); }
+  if (y_buffers.size() != batch_count) { throw BLASError(StatusCode::kInvalidBatchCount); }
+
+  // Naive implementation: calls regular Axpy multiple times
+  for (auto batch = size_t{0}; batch < batch_count; ++batch) {
+    DoAxpy(n, alphas[batch],
+           x_buffers[batch], 0, x_inc,
+           y_buffers[batch], 0, y_inc);
+  }
+}
+
+// =================================================================================================
+
+// Compiles the templated class
+template class XaxpyBatched<half>;
+template class XaxpyBatched<float>;
+template class XaxpyBatched<double>;
+template class XaxpyBatched<float2>;
+template class XaxpyBatched<double2>;
+
+// =================================================================================================
+} // namespace clblast
diff --git a/src/routines/levelx/xaxpybatched.hpp b/src/routines/levelx/xaxpybatched.hpp
new file mode 100644
index 00000000..7fd14a74
--- /dev/null
+++ b/src/routines/levelx/xaxpybatched.hpp
@@ -0,0 +1,46 @@
+
+// =================================================================================================
+// 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 implements the XaxpyBatched routine. This is a non-blas batched version of AXPY.
+//
+// =================================================================================================
+
+#ifndef CLBLAST_ROUTINES_XAXPYBATCHED_H_
+#define CLBLAST_ROUTINES_XAXPYBATCHED_H_
+
+#include <vector>
+
+#include "routines/level1/xaxpy.hpp"
+
+namespace clblast {
+// =================================================================================================
+
+// See comment at top of file for a description of the class
+template <typename T>
+class XaxpyBatched: public Xaxpy<T> {
+ public:
+
+  // Uses the regular Xaxpy routine
+  using Xaxpy<T>::DoAxpy;
+
+  // Constructor
+  XaxpyBatched(Queue &queue, EventPointer event, const std::string &name = "AXPYBATCHED");
+
+  // Templated-precision implementation of the routine
+  void DoAxpyBatched(const size_t n, const std::vector<T> &alphas,
+                     const std::vector<Buffer<T>> &x_buffers, const size_t x_inc,
+                     const std::vector<Buffer<T>> &y_buffers, const size_t y_inc,
+                     const size_t batch_count);
+};
+
+// =================================================================================================
+} // namespace clblast
+
+// CLBLAST_ROUTINES_XAXPYBATCHED_H_
+#endif