Reworked SPD atomic counter to use a compute buffer (which has atomic operations available on Metal) instead of a one-pixel read/write texture.

Runtime/FSR2/Fsr2Context.cs

@@ -213,12 +213,11 @@ namespace FidelityFX.FSR2
                 // Auto exposure always used to track luma changes in locking logic
                 commandBuffer.SetRenderTarget(_resources.AutoExposure[frameIndex ^ 1]);
                 commandBuffer.ClearRenderTarget(false, true, new Color(0f, 1e8f, 0f, 0f));
-
-                // Reset atomic counter to 0
-                commandBuffer.SetRenderTarget(_resources.SpdAtomicCounter);
-                commandBuffer.ClearRenderTarget(false, true, Color.clear);
             }
             
+            // Reset atomic counter to 0
+            commandBuffer.SetBufferCounterValue(_resources.SpdAtomicCounter, 0);
+            
             // FSR3: need to clear here since we need the content of this surface for frame interpolation, so clearing in the lock pass is not an option
             bool depthInverted = (_contextDescription.Flags & Fsr2.InitializationFlags.EnableDepthInverted) == Fsr2.InitializationFlags.EnableDepthInverted;
             commandBuffer.SetRenderTarget(Fsr2ShaderIDs.UavReconstructedPrevNearestDepth);

Runtime/FSR2/Fsr2Pass.cs

@@ -129,7 +129,7 @@ namespace FidelityFX.FSR2
             commandBuffer.SetComputeTextureParam(ComputeShader, KernelIndex, Fsr2ShaderIDs.SrvInputColor, color.RenderTarget, color.MipLevel, color.SubElement);
             commandBuffer.SetComputeTextureParam(ComputeShader, KernelIndex, Fsr2ShaderIDs.SrvAutoExposure, Resources.AutoExposure[frameIndex ^ 1]);
 
-            commandBuffer.SetComputeTextureParam(ComputeShader, KernelIndex, Fsr2ShaderIDs.UavSpdAtomicCount, Resources.SpdAtomicCounter);
+            commandBuffer.SetComputeBufferParam(ComputeShader, KernelIndex, Fsr2ShaderIDs.UavSpdAtomicCount, Resources.SpdAtomicCounter);
             commandBuffer.SetComputeTextureParam(ComputeShader, KernelIndex, Fsr2ShaderIDs.UavExposureMipLumaChange, Resources.SceneLuminance, ShadingChangeMipLevel);
             commandBuffer.SetComputeTextureParam(ComputeShader, KernelIndex, Fsr2ShaderIDs.UavExposureMip5, Resources.SceneLuminance, 5);
             commandBuffer.SetComputeTextureParam(ComputeShader, KernelIndex, Fsr2ShaderIDs.UavAutoExposure, Resources.AutoExposure[frameIndex]);

Runtime/FSR2/Fsr2Resources.cs

@@ -35,7 +35,7 @@ namespace FidelityFX.FSR2
         public Texture2D DefaultReactive;
         public Texture2D LanczosLut;
         public Texture2D MaximumBiasLut;
-        public RenderTexture SpdAtomicCounter;
+        public ComputeBuffer SpdAtomicCounter;
         public RenderTexture SceneLuminance;
         public RenderTexture AutoReactive;
         public RenderTexture AutoComposition;
@@ -87,9 +87,8 @@ namespace FidelityFX.FSR2
             DefaultReactive.Apply();
             
             // Resource FSR2_SpdAtomicCounter: FFX_RESOURCE_USAGE_UAV, FFX_SURFACE_FORMAT_R32_UINT, FFX_RESOURCE_FLAGS_ALIASABLE
-            // Despite what the original FSR2 codebase says, this resource really isn't aliasable. Resetting this counter to 0 every frame breaks auto-exposure on MacOS Metal.
-            SpdAtomicCounter = new RenderTexture(1, 1, 0, GraphicsFormat.R32_UInt) { name = "FSR2_SpdAtomicCounter", enableRandomWrite = true };
-            SpdAtomicCounter.Create();
+            // Despite what the original FSR2 codebase says, this resource really isn't aliasable.
+            SpdAtomicCounter = new ComputeBuffer(1, sizeof(uint), ComputeBufferType.Counter, ComputeBufferMode.Dynamic);
 
             // Resource FSR2_ExposureMips: FFX_RESOURCE_USAGE_UAV, FFX_SURFACE_FORMAT_R16_FLOAT, FFX_RESOURCE_FLAGS_ALIASABLE
             // This is a rather special case: it's an aliasable resource, but because we require a mipmap chain and bind specific mip levels per shader, we can't easily use temporary RTs for this.
@@ -191,6 +190,7 @@ namespace FidelityFX.FSR2
             DestroyResource(ref DefaultExposure);
             DestroyResource(ref MaximumBiasLut);
             DestroyResource(ref LanczosLut);
+            DestroyResource(ref SpdAtomicCounter);
         }
 
         public void DestroyTcrAutogenResources()
@@ -226,6 +226,15 @@ namespace FidelityFX.FSR2
             resource = null;
         }
         
+        private static void DestroyResource(ref ComputeBuffer resource)
+        {
+            if (resource == null)
+                return;
+            
+            resource.Release();
+            resource = null;
+        }
+
         private static void DestroyResource(RenderTexture[] resource)
         {
             for (int i = 0; i < resource.Length; ++i)

Runtime/FSR3/Fsr3UpscalerContext.cs

@@ -232,12 +232,11 @@ namespace FidelityFX.FSR3
                 // Auto exposure always used to track luma changes in locking logic
                 commandBuffer.SetRenderTarget(_resources.FrameInfo[frameIndex ^ 1]);
                 commandBuffer.ClearRenderTarget(false, true, new Color(0f, 1e8f, 0f, 0f));
-
-                // Reset atomic counter to 0
-                commandBuffer.SetRenderTarget(_resources.SpdAtomicCounter);
-                commandBuffer.ClearRenderTarget(false, true, Color.clear);
             }
             
+            // Reset atomic counter to 0
+            commandBuffer.SetBufferCounterValue(_resources.SpdAtomicCounter, 0);
+            
             // FSR3: need to clear here since we need the content of this surface for frame interpolation, so clearing in the lock pass is not an option
             bool depthInverted = (_contextDescription.Flags & Fsr3Upscaler.InitializationFlags.EnableDepthInverted) == Fsr3Upscaler.InitializationFlags.EnableDepthInverted;
             commandBuffer.SetRenderTarget(_resources.ReconstructedPrevNearestDepth);

Runtime/FSR3/Fsr3UpscalerPass.cs

@@ -157,7 +157,7 @@ namespace FidelityFX.FSR3
             commandBuffer.SetComputeTextureParam(ComputeShader, KernelIndex, Fsr3ShaderIDs.SrvFarthestDepth, Fsr3ShaderIDs.UavIntermediate);
             commandBuffer.SetComputeTextureParam(ComputeShader, KernelIndex, Fsr3ShaderIDs.SrvFrameInfo, Resources.FrameInfo[frameIndex ^ 1]);
 
-            commandBuffer.SetComputeTextureParam(ComputeShader, KernelIndex, Fsr3ShaderIDs.UavSpdAtomicCount, Resources.SpdAtomicCounter);
+            commandBuffer.SetComputeBufferParam(ComputeShader, KernelIndex, Fsr3ShaderIDs.UavSpdAtomicCount, Resources.SpdAtomicCounter);
             commandBuffer.SetComputeTextureParam(ComputeShader, KernelIndex, Fsr3ShaderIDs.UavFrameInfo, Resources.FrameInfo[frameIndex]);
             commandBuffer.SetComputeTextureParam(ComputeShader, KernelIndex, Fsr3ShaderIDs.UavSpdMip0, Resources.SpdMips, 0);
             commandBuffer.SetComputeTextureParam(ComputeShader, KernelIndex, Fsr3ShaderIDs.UavSpdMip1, Resources.SpdMips, 1);
@@ -194,7 +194,7 @@ namespace FidelityFX.FSR3
             commandBuffer.SetComputeTextureParam(ComputeShader, KernelIndex, Fsr3ShaderIDs.SrvDilatedMotionVectors, Resources.DilatedVelocity);
             commandBuffer.SetComputeTextureParam(ComputeShader, KernelIndex, Fsr3ShaderIDs.SrvInputExposure, exposure.RenderTarget, exposure.MipLevel, exposure.SubElement);
             
-            commandBuffer.SetComputeTextureParam(ComputeShader, KernelIndex, Fsr3ShaderIDs.UavSpdAtomicCount, Resources.SpdAtomicCounter);
+            commandBuffer.SetComputeBufferParam(ComputeShader, KernelIndex, Fsr3ShaderIDs.UavSpdAtomicCount, Resources.SpdAtomicCounter);
             commandBuffer.SetComputeTextureParam(ComputeShader, KernelIndex, Fsr3ShaderIDs.UavSpdMip0, Resources.SpdMips, 0);
             commandBuffer.SetComputeTextureParam(ComputeShader, KernelIndex, Fsr3ShaderIDs.UavSpdMip1, Resources.SpdMips, 1);
             commandBuffer.SetComputeTextureParam(ComputeShader, KernelIndex, Fsr3ShaderIDs.UavSpdMip2, Resources.SpdMips, 2);

Runtime/FSR3/Fsr3UpscalerResources.cs

@@ -35,7 +35,7 @@ namespace FidelityFX.FSR3
         public Texture2D DefaultExposure;
         public Texture2D DefaultReactive;
         
-        public RenderTexture SpdAtomicCounter;
+        public ComputeBuffer SpdAtomicCounter;
         public RenderTexture SpdMips;
         public RenderTexture DilatedVelocity;
         public RenderTexture DilatedDepth;
@@ -83,9 +83,8 @@ namespace FidelityFX.FSR3
             DefaultExposure.Apply();
             
             // Resource FSR3UPSCALER_SpdAtomicCounter: FFX_RESOURCE_USAGE_UAV, FFX_SURFACE_FORMAT_R32_UINT, FFX_RESOURCE_FLAGS_ALIASABLE
-            // Despite what the original FSR3 codebase says, this resource really isn't aliasable. Resetting this counter to 0 every frame breaks auto-exposure on MacOS Metal.
-            SpdAtomicCounter = new RenderTexture(1, 1, 0, GraphicsFormat.R32_UInt) { name = "FSR3UPSCALER_SpdAtomicCounter", enableRandomWrite = true };
-            SpdAtomicCounter.Create();
+            // Despite what the original FSR3 codebase says, this resource really isn't aliasable.
+            SpdAtomicCounter = new ComputeBuffer(1, sizeof(uint), ComputeBufferType.Counter, ComputeBufferMode.Dynamic);
             
             // Resource FSR3UPSCALER_SpdMips: FFX_RESOURCE_USAGE_UAV, FFX_SURFACE_FORMAT_R16G16_FLOAT, FFX_RESOURCE_FLAGS_ALIASABLE
             // This is a rather special case: it's an aliasable resource, but because we require a mipmap chain and bind specific mip levels per shader, we can't easily use temporary RTs for this.
@@ -235,6 +234,15 @@ namespace FidelityFX.FSR3
             resource = null;
         }
 
+        private static void DestroyResource(ref ComputeBuffer resource)
+        {
+            if (resource == null)
+                return;
+            
+            resource.Release();
+            resource = null;
+        }
+
         private static void DestroyResource(RenderTexture[] resource)
         {
             for (int i = 0; i < resource.Length; ++i)

Shaders/ffx_fsr_unity_common.cginc

@@ -51,9 +51,7 @@
 // #endif
 
 // Hack to work around the lack of texture atomics on Metal
-// TODO: replace texture atomics with a compute buffer setup
 #if defined(SHADER_API_METAL)
-#define InterlockedAdd(dest, val, orig)     { (orig) = (dest); (dest) += (val); }
 #define InterlockedMin(dest, val)           { (dest) = min((dest), (val)); }
 #define InterlockedMax(dest, val)           { (dest) = max((dest), (val)); }
 #endif

Shaders/shaders/fsr2/ffx_fsr2_callbacks_hlsl.h

@@ -467,7 +467,7 @@ SamplerState s_LinearClamp : register(s1);
         RWTexture2D<FfxFloat32x2>                 rw_auto_exposure                          : FFX_FSR2_DECLARE_UAV(FSR2_BIND_UAV_AUTO_EXPOSURE);
     #endif
     #if defined FSR2_BIND_UAV_SPD_GLOBAL_ATOMIC
-        globallycoherent RWTexture2D<FfxUInt32>   rw_spd_global_atomic                      : FFX_FSR2_DECLARE_UAV(FSR2_BIND_UAV_SPD_GLOBAL_ATOMIC);
+        globallycoherent RWStructuredBuffer<FfxUInt32> rw_spd_global_atomic                 : FFX_FSR2_DECLARE_UAV(FSR2_BIND_UAV_SPD_GLOBAL_ATOMIC);
     #endif
 
     #if defined FSR2_BIND_UAV_AUTOREACTIVE
@@ -948,14 +948,14 @@ void SPD_SetMipmap(FfxInt32x2 iPxPos, FfxUInt32 slice, FfxFloat32 value)
 void SPD_IncreaseAtomicCounter(inout FfxUInt32 spdCounter)
 {
 #if defined FSR2_BIND_UAV_SPD_GLOBAL_ATOMIC
-    InterlockedAdd(rw_spd_global_atomic[FfxInt32x2(0, 0)], 1, spdCounter);
+    InterlockedAdd(rw_spd_global_atomic[0], 1, spdCounter);
 #endif // #if defined FSR2_BIND_UAV_SPD_GLOBAL_ATOMIC
 }
 
 void SPD_ResetAtomicCounter()
 {
 #if defined FSR2_BIND_UAV_SPD_GLOBAL_ATOMIC
-    rw_spd_global_atomic[FfxInt32x2(0, 0)] = 0;
+    rw_spd_global_atomic[0] = 0;
 #endif // #if defined FSR2_BIND_UAV_SPD_GLOBAL_ATOMIC
 }
 

Shaders/shaders/fsr3upscaler/ffx_fsr3upscaler_callbacks_hlsl.h

@@ -982,16 +982,16 @@ void StorePyramid(FFX_PARAMETER_IN FfxInt32x2 iPxPos, FFX_PARAMETER_IN FfxFloat3
 #endif
 
 #if defined FSR3UPSCALER_BIND_UAV_SPD_GLOBAL_ATOMIC
-globallycoherent RWTexture2D<FfxUInt32>   rw_spd_global_atomic : FFX_FSR3UPSCALER_DECLARE_UAV(FSR3UPSCALER_BIND_UAV_SPD_GLOBAL_ATOMIC);
+globallycoherent RWStructuredBuffer<FfxUInt32>   rw_spd_global_atomic : FFX_FSR3UPSCALER_DECLARE_UAV(FSR3UPSCALER_BIND_UAV_SPD_GLOBAL_ATOMIC);
 
 void SPD_IncreaseAtomicCounter(inout FfxUInt32 spdCounter)
 {
-    InterlockedAdd(rw_spd_global_atomic[FfxInt32x2(0, 0)], 1, spdCounter);
+    InterlockedAdd(rw_spd_global_atomic[0], 1, spdCounter);
 }
 
 void SPD_ResetAtomicCounter()
 {
-    rw_spd_global_atomic[FfxInt32x2(0, 0)] = 0;
+    rw_spd_global_atomic[0] = 0;
 }
 #endif