Added proper testing of the alpha parameter; finalized the batched AXPY implementation

4 files changed, 16 insertions, 18 deletions

diff --git a/CHANGELOG b/CHANGELOG
index d5e0a2ba..254d6b7b 100644
--- a/CHANGELOG
+++ b/CHANGELOG
@@ -7,13 +7,14 @@ Development version (next release)
 - Fixed bugs in the half-precision routines HTBMV/HTPMV/HTRMV/HSYR2K/HTRMM
 - Tests now also exit with an error code when OpenCL errors or compilation errors occur
 - Tests now also check for the L2 error in case of half-precision
-- Added the OverrideParameters function to the API to be able to supply custom tuning parmeters
 - Various minor fixes and enhancements
 - Added tuned parameters for various devices (see README)
-- Added level-2 routines:
+- Added the OverrideParameters function to the API to be able to supply custom tuning parmeters
+- Added triangular solver (level-2 & level-3) routines:
   * STRSV/DTRSV/CTRSV/ZTRSV (experimental, un-optimized)
-- Added level-3 routines:
   * STRSM/DTRSM/CTRSM/ZTRSM (experimental, un-optimized)
+- Added batched (non-BLAS) routines:
+  * SAXPYBATCHED/DAXPYBATCHED/CAXPYBATCHED/ZAXPYBATCHED/HAXPYBATCHED (batched version of AXPY)
 
 Version 0.10.0
 - Updated to version 8.0 of the CLCudaAPI C++11 OpenCL header
diff --git a/src/kernels/level1/xaxpy.opencl b/src/kernels/level1/xaxpy.opencl
index 3f5ab2b5..f44bbce0 100644
--- a/src/kernels/level1/xaxpy.opencl
+++ b/src/kernels/level1/xaxpy.opencl
@@ -57,9 +57,9 @@ void XaxpyFast(const int n, const real_arg arg_alpha,
 
 // Full version of the kernel with offsets and strided accesses: batched version
 __kernel __attribute__((reqd_work_group_size(WGS, 1, 1)))
-void XaxpyBatched(const int n, const __global real_arg* arg_alphas,
-                  const __global real* restrict xgm, const __global int* restrict x_offsets, const int x_inc,
-                  __global real* ygm, const __global int* restrict y_offsets, const int y_inc) {
+void XaxpyBatched(const int n, const __constant real_arg* arg_alphas,
+                  const __global real* restrict xgm, const __constant int* x_offsets, const int x_inc,
+                  __global real* ygm, const __constant int* y_offsets, const int y_inc) {
   const int batch = get_group_id(1);
   const real alpha = GetRealArg(arg_alphas[batch]);
 
diff --git a/src/utilities/utilities.hpp b/src/utilities/utilities.hpp
index d271ffee..b3db8c22 100644
--- a/src/utilities/utilities.hpp
+++ b/src/utilities/utilities.hpp
@@ -164,6 +164,8 @@ struct Arguments {
   std::vector<size_t> a_offsets = {0};
   std::vector<size_t> b_offsets = {0};
   std::vector<size_t> c_offsets = {0};
+  std::vector<T> alphas = {ConstantOne<T>()};
+  std::vector<T> betas = {ConstantOne<T>()};
   // Sizes
   size_t x_size = 1;
   size_t y_size = 1;
diff --git a/test/routines/levelx/xaxpybatched.hpp b/test/routines/levelx/xaxpybatched.hpp
index 8f6a5985..ee15ff92 100644
--- a/test/routines/levelx/xaxpybatched.hpp
+++ b/test/routines/levelx/xaxpybatched.hpp
@@ -46,11 +46,6 @@ class TestXaxpyBatched {
             kArgBatchCount, kArgAlpha};
   }
 
-  // Helper to determine a different alpha value per batch
-  static T GetAlpha(const T alpha_base, const size_t batch_id) {
-    return alpha_base + Constant<T>(batch_id);
-  }
-
   // Helper for the sizes per batch
   static size_t PerBatchSizeX(const Arguments<T> &args) { return args.n * args.x_inc; }
   static size_t PerBatchSizeY(const Arguments<T> &args) { return args.n * args.y_inc; }
@@ -67,11 +62,15 @@ class TestXaxpyBatched {
   static void SetSizes(Arguments<T> &args) {
     args.x_size = GetSizeX(args);
     args.y_size = GetSizeY(args);
+
+    // Also sets the batch-related variables
     args.x_offsets = std::vector<size_t>(args.batch_count);
     args.y_offsets = std::vector<size_t>(args.batch_count);
+    args.alphas = std::vector<T>(args.batch_count);
     for (auto batch = size_t{0}; batch < args.batch_count; ++batch) {
       args.x_offsets[batch] = batch * PerBatchSizeX(args) + args.x_offset;
       args.y_offsets[batch] = batch * PerBatchSizeY(args) + args.y_offset;
+      args.alphas[batch] = args.alpha + Constant<T>(batch);
     }
   }
 
@@ -94,11 +93,7 @@ class TestXaxpyBatched {
   static StatusCode RunRoutine(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) {
     auto queue_plain = queue();
     auto event = cl_event{};
-    auto alphas = std::vector<T>();
-    for (auto batch = size_t{0}; batch < args.batch_count; ++batch) {
-      alphas.push_back(GetAlpha(args.alpha, batch));
-    }
-    auto status = AxpyBatched(args.n, alphas.data(),
+    auto status = AxpyBatched(args.n, args.alphas.data(),
                               buffers.x_vec(), args.x_offsets.data(), args.x_inc,
                               buffers.y_vec(), args.y_offsets.data(), args.y_inc,
                               args.batch_count,
@@ -113,7 +108,7 @@ class TestXaxpyBatched {
       auto queue_plain = queue();
       for (auto batch = size_t{0}; batch < args.batch_count; ++batch) {
         auto event = cl_event{};
-        auto status = clblasXaxpy(args.n, GetAlpha(args.alpha, batch),
+        auto status = clblasXaxpy(args.n, args.alphas[batch],
                                   buffers.x_vec, args.x_offsets[batch], args.x_inc,
                                   buffers.y_vec, args.y_offsets[batch], args.y_inc,
                                   1, &queue_plain, 0, nullptr, &event);
@@ -134,7 +129,7 @@ class TestXaxpyBatched {
       buffers.x_vec.Read(queue, args.x_size, x_vec_cpu);
       buffers.y_vec.Read(queue, args.y_size, y_vec_cpu);
       for (auto batch = size_t{0}; batch < args.batch_count; ++batch) {
-        cblasXaxpy(args.n, GetAlpha(args.alpha, batch),
+        cblasXaxpy(args.n, args.alphas[batch],
                    x_vec_cpu, args.x_offsets[batch], args.x_inc,
                    y_vec_cpu, args.y_offsets[batch], args.y_inc);
       }