[Vulkan][Nvidia] Fence wait failure for boolean matrix `and()` intrinsic in compute shader

[farzonl]

Summary

offloader.exe fails with gpu-exec: error: Failed waiting for fence. when running matrix_bool_and_operator.test on NVIDIA Vulkan devices. The same test passes on AMD Vulkan devices.

Test case

#--- source.hlsl

RWStructuredBuffer<bool> In : register(u0);
RWStructuredBuffer<bool> AndTrueTrueOut : register(u1);
RWStructuredBuffer<bool> AndTrueFalseOut : register(u2);
RWStructuredBuffer<bool> AndTrueMixOut : register(u3);
RWStructuredBuffer<bool> AndTrueMix2Out : register(u4);

[numthreads(1,1,1)]
void main() {
  bool2x3 MatFalse = bool2x3(In[0], In[1], In[2],
                            In[0], In[1], In[2]);
  bool2x3 MatTrue = bool2x3(In[3], In[4], In[5],
                            In[3], In[4], In[5]); 
  bool2x3 MatMix  = bool2x3(In[0], In[1], In[2],
                            In[3], In[4], In[5]); 
  bool2x3 MatMix2  = bool2x3(In[3], In[4], In[5],
                             In[0], In[1], In[2]); 

  bool2x3 MatAndTrueTrue = and(MatTrue,MatTrue);
  bool2x3 MatAndTrueFalse = and(MatTrue,MatFalse);
  bool2x3 MatAndTrueMix1 = and(MatTrue,MatMix);
  bool2x3 MatAndTrueMix2 = and(MatTrue,MatMix2);

  const uint COLS = 3;          // bool2x3 => 2 rows, 3 columns
  for(uint i = 0; i < 6; i++) {
    uint row = i / COLS;
    uint col = i % COLS;
    AndTrueTrueOut[i]  = MatAndTrueTrue[row][col];
    AndTrueFalseOut[i] = MatAndTrueFalse[row][col];
    AndTrueMixOut[i]   = MatAndTrueMix1[row][col];
    AndTrueMix2Out[i]  = MatAndTrueMix2[row][col];
  }
}

//--- pipeline.yaml

---
Shaders:
  - Stage: Compute
    Entry: main
    DispatchSize: [1, 1, 1]
Buffers:
  - Name: In
    Format: Int32
    Data: [ 0, 0, 0, 1, 1, 1 ]
  - Name: AndTrueTrueOut
    Format: Int32
    FillSize: 24
  - Name: AndTrueTrueExpectedOut
    Format: Int32
    Data: [ 1, 1, 1, 1, 1, 1 ]
  - Name: AndTrueFalseOut
    Format: Int32
    FillSize: 24
  - Name: AndTrueFalseExpectedOut
    Format: Int32
    Data: [ 0, 0, 0, 0, 0, 0 ]
  - Name: AndTrueMixOut
    Format: Int32
    FillSize: 24
  - Name: AndTrueMixExpectedOut
    Format: Int32
    Data: [ 0, 0, 0, 1, 1, 1 ]
  - Name: AndTrueMix2Out
    Format: Int32
    FillSize: 24
  - Name: AndTrueMix2ExpectedOut
    Format: Int32
    Data: [ 1, 1, 1, 0, 0, 0 ]
Results:
  - Result: AndTrueTrueOut
    Rule: BufferExact
    Actual: AndTrueTrueOut
    Expected: AndTrueTrueExpectedOut
  - Result: AndTrueFalseOut
    Rule: BufferExact
    Actual: AndTrueFalseOut
    Expected: AndTrueFalseExpectedOut
  - Result: AndTrueMixOut
    Rule: BufferExact
    Actual: AndTrueMixOut
    Expected: AndTrueMixExpectedOut
  - Result: AndTrueMix2Out
    Rule: BufferExact
    Actual: AndTrueMix2Out
    Expected: AndTrueMix2ExpectedOut
DescriptorSets:
  - Resources:
    - Name: In
      Kind: RWBuffer
      DirectXBinding:
        Register: 0
        Space: 0
      VulkanBinding:
        Binding: 0
    - Name: AndTrueTrueOut
      Kind: RWBuffer
      DirectXBinding:
        Register: 1
        Space: 0
      VulkanBinding:
        Binding: 1
    - Name: AndTrueFalseOut
      Kind: RWBuffer
      DirectXBinding:
        Register: 2
        Space: 0
      VulkanBinding:
        Binding: 2
    - Name: AndTrueMixOut
      Kind: RWBuffer
      DirectXBinding:
        Register: 3
        Space: 0
      VulkanBinding:
        Binding: 3
    - Name: AndTrueMix2Out
      Kind: RWBuffer
      DirectXBinding:
        Register: 4
        Space: 0
      VulkanBinding:
        Binding: 4
...
#--- end

# Bug https://github.com/microsoft/DirectXShaderCompiler/issues/8129
# XFAIL: DXC && Vulkan

# XFAIL: Clang && NVIDIA

# Note: Metal does not support boolean matrix types. For
# Vulkan KosmicKrisp and MoltenVK the output is always False.
# UNSUPPORTED: Darwin

# 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

Error output

d:\repos\llvm-build\bin\offloader.exe -debug-layer -adapter-regex="nvidia" D:\repos\llvm-build\tools\OffloadTest\test\clang-vk\Basic\Matrix\Output\matrix_bool_and_operator.test.tmp/pipeline.yaml D:\repos\llvm-build\tools\OffloadTest\test\clang-vk\Basic\Matrix\Output\matrix_bool_and_operator.test.tmp.o
# executed command: 'd:\repos\llvm-build\bin\offloader.exe' -debug-layer -adapter-regex=nvidia 'D:\repos\llvm-build\tools\OffloadTest\test\clang-vk\Basic\Matrix\Output\matrix_bool_and_operator.test.tmp/pipeline.yaml' 'D:\repos\llvm-build\tools\OffloadTest\test\clang-vk\Basic\Matrix\Output\matrix_bool_and_operator.test.tmp.o'
# .---command stdout------------
# | Using Vulkan API
# | Physical device created.
# | Shader module created.
# | Copy command buffer created.
# | Memory buffers created.
# | Executed copy command buffer.
# | Execute command buffer created.
# | Descriptors: { type = 5, count = 5 }
# | Descriptor pool created.
# | Binding 5 descriptors.
# | Num Descriptor sets: 1
# | Updating Descriptor [0] { 0, 0 }
# | Updating Descriptor [1] { 0, 1 }
# | Updating Descriptor [2] { 0, 2 }
# | Updating Descriptor [3] { 0, 3 }
# | Updating Descriptor [4] { 0, 4 }
# | WriteDescriptors: 5
# | Descriptor sets created.
# | Compute pipeline created.
# | Dispatched compute shader: { 1, 1, 1 }
# | Commands created.
# `-----------------------------
# .---command stderr------------
# | gpu-exec: error: Failed waiting for fence.
# `-----------------------------
# error: command failed with exit status: 1

--

********************
Testing:  0.. 10.. 20.. 30.. 40.. 50.. 60.. 70.. 80.. 90..
********************
Failed Tests (2):
  OffloadTest-clang-vk :: Basic/Matrix/matrix_bool_and_operator.test