feat(gpu): add more mem api

Author: fumiama

gpu/mem.go

@@ -16,6 +16,16 @@ func MemAllocHostDevicePair(size uintptr, alignment uintptr) (h, d unsafe.Pointe
 	return
 }
 
+// MemAllocHost allocs host mem.
+func MemAllocHost(size uintptr, alignment uintptr) (unsafe.Pointer, error) {
+	return g().ctx.MemAllocHost(size, alignment)
+}
+
+// MemAllocDevice allocs device mem.
+func MemAllocDevice(size uintptr, alignment uintptr) (unsafe.Pointer, error) {
+	return g().ctx.MemAllocDevice(g().dev, size, alignment)
+}
+
 // MemCopyGo2Host copies go array to host mem and returns the go array repr of dst.
 func MemCopyGo2Host[T any](dst unsafe.Pointer, src []T) []T {
 	dstSlice := unsafe.Slice((*T)(dst), len(src))

internal/build/darkimg_ocl.cl

@@ -19,8 +19,8 @@ kernel void isdark(
 
     if (x < outW && y < outH) {
         float2 normCoord = (float2)(
-            (float)x / (float)outW,
-            (float)y / (float)outH
+            ((float)x + 0.5f) / (float)outW,
+            ((float)y + 0.5f) / (float)outH
         );
 
         float4 pixel = read_imagef(inputImg, smp, normCoord);

internal/build/kmeans_ocl.cl

@@ -25,8 +25,8 @@ kernel void assign_first_iter(
         pixel = read_imagef(inputImg, (int2)(x, y));
     } else {
         float2 normCoord = (float2)(
-            (float)x / (float)dstW,
-            (float)y / (float)dstH
+            ((float)x + 0.5f) / (float)dstW,
+            ((float)y + 0.5f) / (float)dstH
         );
         pixel = read_imagef(inputImg, smp, normCoord);
     }

internal/build/resize_ocl.cl

@@ -13,8 +13,8 @@ kernel void scale(
     }
 
     float2 normCoord = (float2)(
-        (float)x / (float)outW,
-        (float)y / (float)outH
+        ((float)x + 0.5f) / (float)outW,
+        ((float)y + 0.5f) / (float)outH
     );
 
     float4 pixel = read_imagef(inputImg, smp, normCoord);