Metal Shaders

Signed-off-by: Isaac Marovitz <isaacryu@icloud.com>

Author: IsaacMarovitz

MarathonRecomp/gpu/shader/hlsl/enhanced_burnout_blur_ps.hlsl

@@ -2,16 +2,20 @@
 
 #ifdef __spirv__
 
-#define g_SampleCount vk::RawBufferLoad<int>(g_PushConstants.SharedConstants + 4)
+#define g_SampleCount vk::RawBufferLoad<int>(g_PushConstants.PixelShaderConstants + 0)
 
 #define s0_TextureDescriptorIndex vk::RawBufferLoad<int>(g_PushConstants.SharedConstants + 0)
 #define s0_SamplerDescriptorIndex vk::RawBufferLoad<uint>(g_PushConstants.SharedConstants + 192)
 
 #else
 
+cbuffer PixelShaderConstants : register(b1, space1)
+{
+    int g_SampleCount;
+};
+
 cbuffer SharedConstants : register(b2, space4)
 {
-    int g_SampleCount : packoffset(c0.y);
     uint s0_Texture2DDescriptorIndex : packoffset(c0.x);
     uint s0_SamplerDescriptorIndex : packoffset(c12.x);
     DEFINE_SHARED_CONSTANTS();
@@ -25,19 +29,20 @@ float4 shaderMain(
     in float2 oVelocity : TEXCOORD1,
     in float2 oVelScale : TEXCOORD2) : SV_Target
 {
+    Texture2D<float4> texture = g_Texture2DDescriptorHeap[s0_TextureDescriptorIndex];
+    SamplerState samplerState = g_SamplerDescriptorHeap[s0_SamplerDescriptorIndex];
+
     if (g_SampleCount == 1)
-        return texture.SampleLevel(SAMPLER, oTexCoord0, 0);
+        return texture.SampleLevel(samplerState, oTexCoord0, 0);
 
     float velocityMag = sqrt(dot(oVelocity, oVelocity));
     float blurAmount = saturate(velocityMag - 0.25);
 
     if (blurAmount == 0.0)
-        return texture.SampleLevel(SAMPLER, oTexCoord0, 0);
+        return texture.SampleLevel(samplerState, oTexCoord0, 0);
 
     blurAmount = min(blurAmount * blurAmount, 0.25);
-    float result = float4(0.0);
-
-    Texture2D<float> texture = g_Texture2DDescriptorHeap[s0_TextureDescriptorIndex];
+    float4 result = float4(0.0, 0.0, 0.0, 0.0);
 
     UNROLL
     for (int i = 0; i < g_SampleCount; i++)
@@ -47,7 +52,7 @@ float4 shaderMain(
         float2 offset = blurAmount * oVelScale * t * g_SampleCount;
         float2 samplePos = oTexCoord0 + offset;
 
-        result += texture.SampleLevel(s0_SamplerDescriptorIndex, samplePos, 0);
+        result += texture.SampleLevel(samplerState, samplePos, 0);
     }
 
     return result * (1.0 / (float)g_SampleCount);

MarathonRecomp/gpu/shader/hlsl/enhanced_burnout_blur_vs.hlsl

@@ -2,21 +2,17 @@
 
 #ifdef __spirv__
 
-#define g_Velocity = vk::RawBufferLoad<float4>(g_PushConstants.VertexShaderConstants + 3360, 0x10);
-#define g_SampleCount vk::RawBufferLoad<int>(g_PushConstants.SharedConstants + 4)
+#define g_SampleCount vk::RawBufferLoad<int>(g_PushConstants.VertexShaderConstants + 0)
+#define g_Velocity vk::RawBufferLoad<float4>(g_PushConstants.VertexShaderConstants + 3360, 0x10)
 
 #else
 
 cbuffer VertexShaderConstants : register(b1, space4)
 {
+    int g_SampleCount;
     float4 g_Velocity : packoffset(c210);
 };
 
-cbuffer SharedConstants : register(b2, space4)
-{
-    int g_SampleCount : packoffset(c0.y);
-};
-
 #endif
 
 void shaderMain(
@@ -32,20 +28,20 @@ void shaderMain(
 
     if (g_SampleCount == 1)
     {
-        oVelocity = float2(0.0);
-        oVelScale = float2(0.0);
+        oVelocity = float2(0.0, 0.0);
+        oVelScale = float2(0.0, 0.0);
         return;
     }
 
     float2 centeredUV = iTexCoord0 * 2.0 - 1.0;
 
     oVelocity.x = -g_Velocity.x - centeredUV.y;
-    oVelocity.y = centeredUV.x - centeredUV.y;
+    oVelocity.y = centeredUV.x - g_Velocity.y;
 
     float sampleScaleX = 0.00002 / (float)g_SampleCount;
     float sampleScaleY = 0.00001 / (float)g_SampleCount;
 
     float2 scaledVec = oVelocity * g_Velocity.w;
-    output oVelScale.x = scaledVec * sampleScaleX;
-    output oVelScale.y = scaledVec * sampleScaleY;
+    oVelScale.x = scaledVec.x * sampleScaleX;
+    oVelScale.y = scaledVec.y * sampleScaleY;
 }

MarathonRecomp/gpu/shader/msl/enhanced_burnout_blur_ps.metal

@@ -0,0 +1,48 @@
+#include "../../../../tools/XenosRecomp/XenosRecomp/shader_common.h"
+
+#define g_SampleCount (*(reinterpret_cast<device int*>(g_PushConstants.PixelShaderConstants + 0)))
+
+#define s0_TextureDescriptorIndex (*(reinterpret_cast<device int*>(g_PushConstants.SharedConstants + 0)))
+#define s0_SamplerDescriptorIndex (*(reinterpret_cast<device uint*>(g_PushConstants.SharedConstants + 192)))
+
+struct Interpolators
+{
+    float4 oPos [[position]];
+    float2 oTexCoord0 [[user(TEXCOORD0)]];
+    float2 oVelocity [[user(TEXCOORD1)]];
+    float2 oVelScale [[user(TEXCOORD2)]];
+};
+
+[[fragment]]
+float4 shaderMain(Interpolators input [[stage_in]],
+                  constant Texture2DDescriptorHeap* g_Texture2DDescriptorHeap [[buffer(0)]],
+                  constant SamplerDescriptorHeap* g_SamplerDescriptorHeap [[buffer(3)]],
+                  constant PushConstants& g_PushConstants [[buffer(8)]])
+{
+    texture2d<float> texture = g_Texture2DDescriptorHeap[s0_TextureDescriptorIndex].tex;
+    sampler samp = g_SamplerDescriptorHeap[s0_SamplerDescriptorIndex].samp;
+
+    if (g_SampleCount == 1)
+        return texture.sample(samp, input.oTexCoord0, level(0));
+
+    float velocityMag = sqrt(dot(input.oVelocity, input.oVelocity));
+    float blurAmount = saturate(velocityMag - 0.25);
+
+    if (blurAmount == 0.0)
+        return texture.sample(samp, input.oTexCoord0, level(0));
+
+    blurAmount = min(blurAmount * blurAmount, 0.25);
+    float4 result = float4(0.0);
+
+    for (int i = 0; i < g_SampleCount; i++)
+    {
+        float t = ((float)i / (float)(g_SampleCount - 1)) * 2.0 - 1.0;
+
+        float2 offset = blurAmount * input.oVelScale * t * g_SampleCount;
+        float2 samplePos = input.oTexCoord0 + offset;
+
+        result += texture.sample(samp, samplePos, level(0));
+    }
+
+    return result * (1.0 / (float)g_SampleCount);
+}

MarathonRecomp/gpu/shader/msl/enhanced_burnout_blur_vs.metal

@@ -0,0 +1,49 @@
+#include "../../../../tools/XenosRecomp/XenosRecomp/shader_common.h"
+
+#define g_SampleCount (*(reinterpret_cast<device int*>(g_PushConstants.VertexShaderConstants + 0)))
+#define g_Velocity (*(reinterpret_cast<device float4*>(g_PushConstants.VertexShaderConstants + 3360)))
+
+struct VertexShaderInput
+{
+    float4 iPosition0 [[attribute(0)]];
+    float2 iTexCoord0 [[attribute(4)]];
+};
+
+struct Interpolators
+{
+    float4 oPos [[position]];
+    float2 oTexCoord0 [[user(TEXCOORD0)]];
+    float2 oVelocity [[user(TEXCOORD1)]];
+    float2 oVelScale [[user(TEXCOORD2)]];
+};
+
+[[vertex]]
+Interpolators shaderMain(VertexShaderInput input [[stage_in]],
+                         constant PushConstants& g_PushConstants [[buffer(8)]])
+{
+    Interpolators output;
+
+    output.oPos = input.iPosition0;
+    output.oTexCoord0 = input.iTexCoord0;
+
+    if (g_SampleCount == 1)
+    {
+        output.oVelocity = float2(0.0);
+        output.oVelScale = float2(0.0);
+        return output;
+    }
+
+    float2 centeredUV = input.iTexCoord0 * 2.0 - 1.0;
+
+    output.oVelocity.x = -g_Velocity.x - centeredUV.y;
+    output.oVelocity.y = centeredUV.x - g_Velocity.y;
+
+    float sampleScaleX = 0.00002 / (float)g_SampleCount;
+    float sampleScaleY = 0.00001 / (float)g_SampleCount;
+
+    float2 scaledVec = output.oVelocity * g_Velocity.w;
+    output.oVelScale.x = scaledVec.x * sampleScaleX;
+    output.oVelScale.y = scaledVec.y * sampleScaleY;
+
+    return output;
+}