From 99b49c2b6ca0c32ccad96e2d4ec7ae4bc4f99bda Mon Sep 17 00:00:00 2001
From: Lixfel <git@lixfel.de>
Date: Sat, 18 Jul 2026 12:22:33 +0200
Subject: [PATCH] CL_MEM_USE_HOST_PTR for CLArrays instead of
 CL_MEM_ALLOC_HOST_PTR for alignment reasons

---
 src/opencl.cpp | 25 ++++++++++++++++++++++---
 src/opencl.h   |  5 +++++
 2 files changed, 27 insertions(+), 3 deletions(-)

diff --git a/src/opencl.cpp b/src/opencl.cpp
index 980d407..512231c 100644
--- a/src/opencl.cpp
+++ b/src/opencl.cpp
@@ -14,7 +14,8 @@
      limitations under the License.
  */
 #include "opencl.h"
-#include "cl_kernels.h"
+
+#include <sys/mman.h>
 
 #include <iostream>
 #include <utility>
@@ -142,6 +143,11 @@ std::shared_ptr<RawImage> OpenCL::acquireNV12(int width, int height) {
 }
 
 static inline cl::Buffer clAlloc(cl_mem_flags type, cl::size_type size, void* data) {
+	if(type &= CL_MEM_USE_HOST_PTR) {
+		if(!mlock(data, size)) // Some OpenCL drivers (nvidia, AMD) require pinned memory for zero copy operations
+			std:: cerr << "[OpenCL] Could not lock memory, potential performance degradation" << std::endl;
+	}
+
 	int error;
 	cl::Buffer buffer(type | CL_MEM_READ_WRITE, size, data, &error);
 	if(error != CL_SUCCESS) {
@@ -151,8 +157,21 @@ static inline cl::Buffer clAlloc(cl_mem_flags type, cl::size_type size, void* da
 	return buffer;
 }
 
-CLArray::CLArray(int size): buffer(clAlloc((cl_mem_flags) CL_MEM_ALLOC_HOST_PTR, (cl::size_type) size, nullptr)), size(size) {}
-CLArray::CLArray(void* data, const int size): buffer(clAlloc((cl_mem_flags) CL_MEM_COPY_HOST_PTR, (cl::size_type) size, data)), size(size) {}
+/* Manually align to 4KiB instead of using CL_MEM_ALLOC_HOST_PTR due to possible device constraints.
+- Spinnaker USB requires 1024 byte alignment
+- Nvidia, AMD, Intel require usually page boundary (4KiB on x86 and AMD64) alignment
+  https://www.intel.com/content/dam/develop/external/us/en/documents/opencl-zero-copy-in-opencl-1-2.pdf
+*/
+#define MINIMUM_ALIGNMENT 4096
+
+CLArray::CLArray(int size): ptr(aligned_alloc(MINIMUM_ALIGNMENT, size)), buffer(clAlloc((cl_mem_flags) CL_MEM_USE_HOST_PTR, (cl::size_type) size, (void*) ptr)), size(size) {}
+CLArray::CLArray(void* data, const int size): ptr(nullptr), buffer(clAlloc((cl_mem_flags) CL_MEM_COPY_HOST_PTR, (cl::size_type) size, data)), size(size) {}
+CLArray::~CLArray() {
+	if(ptr) {
+		munlock(ptr, size);
+		free((void*) ptr);
+	}
+}
 
 static inline cl::Image2D allocImage(int width, int height, const PixelFormat* format) {
 	int error;
diff --git a/src/opencl.h b/src/opencl.h
index 5fe9506..b075ed0 100644
--- a/src/opencl.h
+++ b/src/opencl.h
@@ -158,11 +158,16 @@ class CLArray {
 public:
 	explicit CLArray(int size);
 	CLArray(void* data, int size);
+	~CLArray();
 
 	template<typename T> CLMap<T> read() const { return CLMap<T>(buffer, size, CL_MAP_READ); }
 	template<typename T> CLMap<T> write() { return CLMap<T>(buffer, size, CL_MAP_WRITE_INVALIDATE_REGION); }
 	template<typename T> CLMap<T> readWrite() { return CLMap<T>(buffer, size, CL_MAP_WRITE); }
 
+private:
+	const void* ptr;
+
+public:
 	const cl::Buffer buffer;
 	const int size;
 };