Fixed half-precision bugs in HTBMV/HTPMV/HTRMV/HSYR2K/HTRMM related to incorrect constants

CHANGELOG

@@ -4,6 +4,7 @@ Development version (next release)
 - Fixed a bug having to re-create the binary even if it was in the cache
 - Fixed a bug when using offsets in the direct version of the GEMM kernels
 - Fixed a missing cl_khr_fp64 when running double-precision on Intel CPUs
+- 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
 - Added the OverrideParameters function to the API to be able to supply custom tuning parmeters
 - Various minor fixes and enhancements

src/routines/level2/xtbmv.cpp

@@ -52,9 +52,9 @@ void Xtbmv<T>::DoTbmv(const Layout layout, const Triangle triangle,
   auto fast_kernels = false;
   try {
     MatVec(layout, a_transpose,
-           n, n, static_cast<T>(1),
+           n, n, ConstantOne<T>(),
            a_buffer, a_offset, a_ld,
-           scratch_buffer, x_offset, x_inc, static_cast<T>(0),
+           scratch_buffer, x_offset, x_inc, ConstantZero<T>(),
            x_buffer, x_offset, x_inc,
            fast_kernels, fast_kernels,
            parameter, false, k, 0);

src/routines/level2/xtpmv.cpp

@@ -52,9 +52,9 @@ void Xtpmv<T>::DoTpmv(const Layout layout, const Triangle triangle,
   auto fast_kernels = false;
   try {
     MatVec(layout, a_transpose,
-           n, n, static_cast<T>(1),
+           n, n, ConstantOne<T>(),
            ap_buffer, ap_offset, n,
-           scratch_buffer, x_offset, x_inc, static_cast<T>(0),
+           scratch_buffer, x_offset, x_inc, ConstantZero<T>(),
            x_buffer, x_offset, x_inc,
            fast_kernels, fast_kernels,
            parameter, true, 0, 0);

src/routines/level2/xtrmv.cpp

@@ -52,9 +52,9 @@ void Xtrmv<T>::DoTrmv(const Layout layout, const Triangle triangle,
   auto fast_kernels = false;
   try {
     MatVec(layout, a_transpose,
-           n, n, static_cast<T>(1),
+           n, n, ConstantOne<T>(),
            a_buffer, a_offset, a_ld,
-           scratch_buffer, x_offset, x_inc, static_cast<T>(0),
+           scratch_buffer, x_offset, x_inc, ConstantZero<T>(),
            x_buffer, x_offset, x_inc,
            fast_kernels, fast_kernels,
            parameter, false, 0, 0);

src/routines/level3/xsyr2k.cpp

@@ -149,7 +149,7 @@ void Xsyr2k<T>::DoSyr2k(const Layout layout, const Triangle triangle, const Tran
   eventWaitList.push_back(eventKernel1);
 
   // Swaps the arguments for matrices A and B, and sets 'beta' to 1
-  auto one = static_cast<T>(1);
+  auto one = ConstantOne<T>();
   kernel.SetArgument(3, GetRealArg(one));
   kernel.SetArgument(4, b_temp());
   kernel.SetArgument(5, a_temp());

src/routines/level3/xtrmm.cpp

@@ -101,7 +101,7 @@ void Xtrmm<T>::DoTrmm(const Layout layout, const Side side, const Triangle trian
            alpha,
            temp_triangular, 0, k,
            b_buffer_copy, b_offset, b_ld,
-           static_cast<T>(0.0),
+           ConstantZero<T>(),
            b_buffer, b_offset, b_ld);
   }
 
@@ -113,7 +113,7 @@ void Xtrmm<T>::DoTrmm(const Layout layout, const Side side, const Triangle trian
              alpha,
              b_buffer_copy, b_offset, b_ld,
              temp_triangular, 0, k,
-             static_cast<T>(0.0),
+             ConstantZero<T>(),
              b_buffer, b_offset, b_ld);
     } catch (BLASError &e) {
       // A and B are now reversed, so also reverse the error codes returned from the Xgemm routine