MSL: Add missing interlock handling to atomic image buffers.

2 files changed, 19 insertions, 7 deletions

diff --git a/reference/opt/shaders-msl/frag/pixel-interlock-ordered.msl2.argument.frag b/reference/opt/shaders-msl/frag/pixel-interlock-ordered.msl2.argument.frag
index f77b8ed9..dfef91c8 100644
--- a/reference/opt/shaders-msl/frag/pixel-interlock-ordered.msl2.argument.frag
+++ b/reference/opt/shaders-msl/frag/pixel-interlock-ordered.msl2.argument.frag
@@ -1,3 +1,4 @@
+#pragma clang diagnostic ignored "-Wmissing-prototypes"
 #pragma clang diagnostic ignored "-Wunused-variable"
 
 #include <metal_stdlib>
@@ -28,16 +29,22 @@ struct spvDescriptorSetBuffer0
     texture2d<float, access::write> img4 [[id(1)]];
     texture2d<float, access::write> img [[id(2), raster_order_group(0)]];
     texture2d<float> img3 [[id(3), raster_order_group(0)]];
-    volatile device Buffer* m_34 [[id(4), raster_order_group(0)]];
-    device Buffer2* m_44 [[id(5), raster_order_group(0)]];
+    texture2d<uint> img2 [[id(4), raster_order_group(0)]];
+    device atomic_uint* img2_atomic [[id(5), raster_order_group(0)]];
+    volatile device Buffer* m_42 [[id(6), raster_order_group(0)]];
+    device Buffer2* m_52 [[id(7), raster_order_group(0)]];
 };
 
+// Returns buffer coords corresponding to 2D texture coords for emulating 2D texture atomics
+#define spvImage2DAtomicCoord(tc, tex) (((tex).get_width() * (tc).x) + (tc).y)
+
 fragment void main0(constant spvDescriptorSetBuffer0& spvDescriptorSet0 [[buffer(0)]])
 {
     (*spvDescriptorSet0.m_9).baz = 0;
     spvDescriptorSet0.img4.write(float4(1.0, 0.0, 0.0, 1.0), uint2(int2(1)));
     spvDescriptorSet0.img.write(spvDescriptorSet0.img3.read(uint2(int2(0))), uint2(int2(0)));
-    (*spvDescriptorSet0.m_34).foo += 42;
-    uint _49 = atomic_fetch_and_explicit((volatile device atomic_uint*)&(*spvDescriptorSet0.m_34).bar, (*spvDescriptorSet0.m_44).quux, memory_order_relaxed);
+    uint _39 = atomic_fetch_add_explicit((device atomic_uint*)&spvDescriptorSet0.img2_atomic[spvImage2DAtomicCoord(int2(0), spvDescriptorSet0.img2)], 1u, memory_order_relaxed);
+    (*spvDescriptorSet0.m_42).foo += 42;
+    uint _55 = atomic_fetch_and_explicit((volatile device atomic_uint*)&(*spvDescriptorSet0.m_42).bar, (*spvDescriptorSet0.m_52).quux, memory_order_relaxed);
 }
 
diff --git a/reference/opt/shaders-msl/frag/pixel-interlock-ordered.msl2.frag b/reference/opt/shaders-msl/frag/pixel-interlock-ordered.msl2.frag
index 803416c6..67b79d9d 100644
--- a/reference/opt/shaders-msl/frag/pixel-interlock-ordered.msl2.frag
+++ b/reference/opt/shaders-msl/frag/pixel-interlock-ordered.msl2.frag
@@ -1,3 +1,4 @@
+#pragma clang diagnostic ignored "-Wmissing-prototypes"
 #pragma clang diagnostic ignored "-Wunused-variable"
 
 #include <metal_stdlib>
@@ -22,12 +23,16 @@ struct Buffer2
     uint quux;
 };
 
-fragment void main0(device Buffer3& _9 [[buffer(0)]], volatile device Buffer& _34 [[buffer(1), raster_order_group(0)]], device Buffer2& _44 [[buffer(2), raster_order_group(0)]], texture2d<float, access::write> img4 [[texture(0)]], texture2d<float, access::write> img [[texture(1), raster_order_group(0)]], texture2d<float> img3 [[texture(2), raster_order_group(0)]])
+// Returns buffer coords corresponding to 2D texture coords for emulating 2D texture atomics
+#define spvImage2DAtomicCoord(tc, tex) (((tex).get_width() * (tc).x) + (tc).y)
+
+fragment void main0(device Buffer3& _9 [[buffer(0)]], volatile device Buffer& _42 [[buffer(2), raster_order_group(0)]], device Buffer2& _52 [[buffer(3), raster_order_group(0)]], texture2d<float, access::write> img4 [[texture(0)]], texture2d<float, access::write> img [[texture(1), raster_order_group(0)]], texture2d<float> img3 [[texture(2), raster_order_group(0)]], texture2d<uint> img2 [[texture(3), raster_order_group(0)]], device atomic_uint* img2_atomic [[buffer(1), raster_order_group(0)]])
 {
     _9.baz = 0;
     img4.write(float4(1.0, 0.0, 0.0, 1.0), uint2(int2(1)));
     img.write(img3.read(uint2(int2(0))), uint2(int2(0)));
-    _34.foo += 42;
-    uint _49 = atomic_fetch_and_explicit((volatile device atomic_uint*)&_34.bar, _44.quux, memory_order_relaxed);
+    uint _39 = atomic_fetch_add_explicit((device atomic_uint*)&img2_atomic[spvImage2DAtomicCoord(int2(0), img2)], 1u, memory_order_relaxed);
+    _42.foo += 42;
+    uint _55 = atomic_fetch_and_explicit((volatile device atomic_uint*)&_42.bar, _52.quux, memory_order_relaxed);
 }