From 772d21a8b656bc9c6ad5dbacb394517a28afe36a Mon Sep 17 00:00:00 2001
From: Michael Fabian 'Xaymar' Dirks <info@xaymar.com>
Date: Tue, 28 Sep 2021 03:46:50 +0200
Subject: [PATCH] nvidia/cuda: Add cuMemsetD# functions

---
 source/nvidia/cuda/nvidia-cuda.cpp | 6 ++++++
 source/nvidia/cuda/nvidia-cuda.hpp | 6 ++++++
 2 files changed, 12 insertions(+)

diff --git a/source/nvidia/cuda/nvidia-cuda.cpp b/source/nvidia/cuda/nvidia-cuda.cpp
index 8c83b828..da2f44b9 100644
--- a/source/nvidia/cuda/nvidia-cuda.cpp
+++ b/source/nvidia/cuda/nvidia-cuda.cpp
@@ -157,6 +157,12 @@ streamfx::nvidia::cuda::cuda::cuda() : _library()
 		P_CUDA_LOAD_SYMBOL_OPT_V2(cuMemcpyHtoD);
 		P_CUDA_LOAD_SYMBOL_OPT_V2(cuMemcpyHtoDAsync);
 		P_CUDA_LOAD_SYMBOL_OPT_V2(cuMemHostGetDevicePointer);
+		P_CUDA_LOAD_SYMBOL_V2(cuMemsetD8);
+		P_CUDA_LOAD_SYMBOL(cuMemsetD8Async);
+		P_CUDA_LOAD_SYMBOL_OPT_V2(cuMemsetD16);
+		P_CUDA_LOAD_SYMBOL_OPT(cuMemsetD16Async);
+		P_CUDA_LOAD_SYMBOL_OPT_V2(cuMemsetD32);
+		P_CUDA_LOAD_SYMBOL_OPT(cuMemsetD32Async);
 
 		// Virtual Memory Management
 		// - Not yet needed.
diff --git a/source/nvidia/cuda/nvidia-cuda.hpp b/source/nvidia/cuda/nvidia-cuda.hpp
index c43a4440..5929a50b 100644
--- a/source/nvidia/cuda/nvidia-cuda.hpp
+++ b/source/nvidia/cuda/nvidia-cuda.hpp
@@ -265,6 +265,12 @@ namespace streamfx::nvidia::cuda {
 		P_CUDA_DEFINE_FUNCTION(cuMemcpyHtoAAsync, array_t dst, std::size_t dstOffset, void* src, std::size_t byteCount);
 		P_CUDA_DEFINE_FUNCTION(cuMemcpyHtoD, device_ptr_t dst, void* src, std::size_t byteCount);
 		P_CUDA_DEFINE_FUNCTION(cuMemcpyHtoDAsync, device_ptr_t dst, void* src, std::size_t byteCount);
+		P_CUDA_DEFINE_FUNCTION(cuMemsetD8, device_ptr_t dst, uint8_t d, size_t byteCount);
+		P_CUDA_DEFINE_FUNCTION(cuMemsetD8Async, device_ptr_t dst, uint8_t d, size_t byteCount, stream_t stream);
+		P_CUDA_DEFINE_FUNCTION(cuMemsetD16, device_ptr_t dst, uint16_t d, size_t byteCount);
+		P_CUDA_DEFINE_FUNCTION(cuMemsetD16Async, device_ptr_t dst, uint16_t d, size_t byteCount, stream_t stream);
+		P_CUDA_DEFINE_FUNCTION(cuMemsetD32, device_ptr_t dst, uint32_t d, size_t byteCount);
+		P_CUDA_DEFINE_FUNCTION(cuMemsetD32Async, device_ptr_t dst, uint32_t d, size_t byteCount, stream_t stream);
 
 		// Virtual Memory Management
 		// - Not yet needed.