Add tests for concurrent writes to matrices (#654)

Adds new offload tests as mentioned should be done in
llvm/llvm-project#174629

These tests are concurrently writing to a matrix in groupshared memory
instead of directly in UAVs because the writing of matrices to UAVs is
not currently implemented in Clang.

The XFAIL on Clang with bug
llvm/llvm-project#174629 can be removed after
llvm/llvm-project#176216 is merged.

Author: Icohedron

test/WaveOps/GroupSharedMatrixEltComponentDataRace.test

@@ -0,0 +1,74 @@
+#--- source.hlsl
+
+// This tests that we can concurrently write to the components of a float4x4
+// as though it is an array. The large group size is used to force the write
+// to occur across multiple waves so that the datarace can be observed.
+
+RWStructuredBuffer<float> In : register(u0);
+RWStructuredBuffer<float> Out : register(u1);
+
+groupshared float4x4 SharedMat;
+
+[numthreads(32, 4, 4)]
+void main(uint3 ThreadID : SV_GroupThreadID) {
+  uint FlatIdx = ThreadID.y + ThreadID.z * 4;
+  if (ThreadID.x == 0) {
+    SharedMat[ThreadID.y][ThreadID.z] = In[FlatIdx];
+  }
+  GroupMemoryBarrierWithGroupSync();
+  if (ThreadID.x == 0) {
+    Out[FlatIdx] = SharedMat[ThreadID.y][ThreadID.z];
+  }
+}
+//--- pipeline.yaml
+
+---
+Shaders:
+  - Stage: Compute
+    Entry: main
+    DispatchSize: [1, 1, 1]
+Buffers:
+  - Name: In
+    Format: Float32
+    Data: [ 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0, 16.0 ]
+  - Name: Out
+    Format: Float32
+    FillSize: 64
+  - Name: ExpectedOut
+    Format: Float32
+    Data: [ 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0, 16.0 ]
+Results:
+  - Result: Out
+    Rule: BufferFloatULP
+    ULPT: 0
+    Actual: Out
+    Expected: ExpectedOut
+DescriptorSets:
+  - Resources:
+    - Name: In
+      Kind: RWStructuredBuffer
+      DirectXBinding:
+        Register: 0
+        Space: 0
+      VulkanBinding:
+        Binding: 0
+    - Name: Out
+      Kind: RWStructuredBuffer
+      DirectXBinding:
+        Register: 1
+        Space: 0
+      VulkanBinding:
+        Binding: 1
+...
+#--- end
+
+# Bug https://github.com/llvm/llvm-project/issues/172577
+# XFAIL: Clang && Vulkan
+
+# Bug https://github.com/llvm/offload-test-suite/issues/452
+# XFAIL: Metal
+
+# RUN: split-file %s %t
+# RUN: %dxc_target -T cs_6_0 -Fo %t.o %t/source.hlsl
+# RUN: %offloader %t/pipeline.yaml %t.o
+

test/WaveOps/GroupSharedMatrixRowComponentDataRace.test

@@ -0,0 +1,76 @@
+#--- source.hlsl
+
+// This tests that we can concurrently write to the rows of a float4x4 as
+// though it is an array of four float4s. The large group size is used to force the
+// write to occur across multiple waves so that the datarace can be observed.
+
+RWStructuredBuffer<float4> In : register(u0);
+RWStructuredBuffer<float4> Out : register(u1);
+
+groupshared float4x4 SharedMat;
+
+[numthreads(128, 4, 1)]
+void main(uint3 ThreadID : SV_GroupThreadID, uint FlatThreadID : SV_GroupIndex) {
+  if (ThreadID.x == 0) {
+    SharedMat[ThreadID.y] = In[ThreadID.y];
+  }
+  GroupMemoryBarrierWithGroupSync();
+  if (ThreadID.x == 0) {
+    Out[ThreadID.y] = SharedMat[ThreadID.y];
+  }
+}
+//--- pipeline.yaml
+
+---
+Shaders:
+  - Stage: Compute
+    Entry: main
+    DispatchSize: [1, 1, 1]
+Buffers:
+  - Name: In
+    Format: Float32
+    Stride: 16
+    Data: [ 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0, 16.0 ]
+  - Name: Out
+    Format: Float32
+    Stride: 16
+    FillSize: 64
+  - Name: ExpectedOut
+    Format: Float32
+    Stride: 16
+    Data: [ 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0, 16.0 ]
+Results:
+  - Result: Out
+    Rule: BufferFloatULP
+    ULPT: 0
+    Actual: Out
+    Expected: ExpectedOut
+DescriptorSets:
+  - Resources:
+    - Name: In
+      Kind: RWStructuredBuffer
+      DirectXBinding:
+        Register: 0
+        Space: 0
+      VulkanBinding:
+        Binding: 0
+    - Name: Out
+      Kind: RWStructuredBuffer
+      DirectXBinding:
+        Register: 1
+        Space: 0
+      VulkanBinding:
+        Binding: 1
+...
+#--- end
+
+# Bug https://github.com/llvm/llvm-project/issues/172577
+# XFAIL: Clang && Vulkan
+
+# Bug https://github.com/llvm/offload-test-suite/issues/452
+# XFAIL: Metal
+
+# RUN: split-file %s %t
+# RUN: %dxc_target -T cs_6_0 -Fo %t.o %t/source.hlsl
+# RUN: %offloader %t/pipeline.yaml %t.o
+