Added CUDA interface to get temporary-buffer size for GEMM routine

1 files changed, 12 insertions, 0 deletions

diff --git a/include/clblast_cuda.h b/include/clblast_cuda.h
index 0f510981..e0d1d638 100644
--- a/include/clblast_cuda.h
+++ b/include/clblast_cuda.h
@@ -69,6 +69,7 @@ enum class StatusCode {
   kInsufficientMemoryY       = -1007, // Vector Y's OpenCL buffer is too small
 
   // Custom additional status codes for CLBlast
+  kInsufficientMemoryTemp    = -2050, // Temporary buffer provided to GEMM routine is too small
   kInvalidBatchCount         = -2049, // The batch count needs to be positive
   kInvalidOverrideKernel     = -2048, // Trying to override parameters for an invalid kernel
   kMissingOverrideParameter  = -2047, // Missing override parameter(s) for the target kernel
@@ -620,6 +621,17 @@ StatusCode GemmBatched(const Layout layout, const Transpose a_transpose, const T
 
 // =================================================================================================
 
+// Retrieves the required size of the temporary buffer for the GEMM kernel (optional)
+template <typename T>
+StatusCode GemmTempBufferSize(const Layout layout, const Transpose a_transpose, const Transpose b_transpose,
+                              const size_t m, const size_t n, const size_t k,
+                              const size_t a_offset, const size_t a_ld,
+                              const size_t b_offset, const size_t b_ld,
+                              const size_t c_offset, const size_t c_ld,
+                              const CUdevice device, size_t& temp_buffer_size);
+
+// =================================================================================================
+
 // CLBlast stores binaries of compiled kernels into a cache in case the same kernel is used later on
 // for the same device. This cache can be cleared to free up system memory or in case of debugging.
 StatusCode PUBLIC_API ClearCache();