[Intel][Vulkan][DXC] fma tests fail on check-hlsl-vk on intel

[NeKon69]

Test for fma (test/Feature/HLSLLib/fma.test) fails on Intel Vulkan DXC at high precision (windows-intel, check-hlsl-vk).

First seen in this run

Log:

******************** TEST 'OffloadTest-vk :: Feature/HLSLLib/fma.test' FAILED ********************
Exit Code: 1

Command Output (stdout):
--
# RUN: at line 102
e:\actions-runner\_work\offload-test-suite\offload-test-suite\llvm-project\build\bin\split-file.exe E:\actions-runner\_work\offload-test-suite\offload-test-suite\OffloadTest\test\Feature\HLSLLib\fma.test E:\actions-runner\_work\offload-test-suite\offload-test-suite\llvm-project\build\tools\OffloadTest\test\vk\Feature\HLSLLib\Output\fma.test.tmp
# executed command: 'e:\actions-runner\_work\offload-test-suite\offload-test-suite\llvm-project\build\bin\split-file.exe' 'E:\actions-runner\_work\offload-test-suite\offload-test-suite\OffloadTest\test\Feature\HLSLLib\fma.test' 'E:\actions-runner\_work\offload-test-suite\offload-test-suite\llvm-project\build\tools\OffloadTest\test\vk\Feature\HLSLLib\Output\fma.test.tmp'
# RUN: at line 103
"E:\actions-runner\_work\offload-test-suite\offload-test-suite\DXC\build\bin\dxc.exe" -spirv -fspv-target-env=vulkan1.3 -Gis -HV 202x -T cs_6_5 -Fo E:\actions-runner\_work\offload-test-suite\offload-test-suite\llvm-project\build\tools\OffloadTest\test\vk\Feature\HLSLLib\Output\fma.test.tmp.o E:\actions-runner\_work\offload-test-suite\offload-test-suite\llvm-project\build\tools\OffloadTest\test\vk\Feature\HLSLLib\Output\fma.test.tmp/source.hlsl
# executed command: 'E:\actions-runner\_work\offload-test-suite\offload-test-suite\DXC\build\bin\dxc.exe' -spirv -fspv-target-env=vulkan1.3 -Gis -HV 202x -T cs_6_5 -Fo 'E:\actions-runner\_work\offload-test-suite\offload-test-suite\llvm-project\build\tools\OffloadTest\test\vk\Feature\HLSLLib\Output\fma.test.tmp.o' 'E:\actions-runner\_work\offload-test-suite\offload-test-suite\llvm-project\build\tools\OffloadTest\test\vk\Feature\HLSLLib\Output\fma.test.tmp/source.hlsl'
# RUN: at line 104
e:\actions-runner\_work\offload-test-suite\offload-test-suite\llvm-project\build\bin\offloader.exe -debug-layer E:\actions-runner\_work\offload-test-suite\offload-test-suite\llvm-project\build\tools\OffloadTest\test\vk\Feature\HLSLLib\Output\fma.test.tmp/pipeline.yaml E:\actions-runner\_work\offload-test-suite\offload-test-suite\llvm-project\build\tools\OffloadTest\test\vk\Feature\HLSLLib\Output\fma.test.tmp.o
# executed command: 'e:\actions-runner\_work\offload-test-suite\offload-test-suite\llvm-project\build\bin\offloader.exe' -debug-layer 'E:\actions-runner\_work\offload-test-suite\offload-test-suite\llvm-project\build\tools\OffloadTest\test\vk\Feature\HLSLLib\Output\fma.test.tmp/pipeline.yaml' 'E:\actions-runner\_work\offload-test-suite\offload-test-suite\llvm-project\build\tools\OffloadTest\test\vk\Feature\HLSLLib\Output\fma.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 = 7, count = 4 }
# | Descriptor pool created.
# | Binding 4 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 }
# | WriteDescriptors: 4
# | Descriptor sets created.
# | Compute pipeline created.
# | Dispatched compute shader: { 1, 1, 1 }
# | Commands created.
# | Executed compute command buffer.
# | Compute pipeline created.
# | Cleanup complete.
# `-----------------------------
# .---command stderr------------
# | Test failed: Result
# | Comparison Rule: BufferExact
# | Expected:
# | ---
# | Name:            Expected
# | Format:          Float64
# | Stride:          32
# | Data:            [ 3.5, -7.75, -3.5, 6, -3.5, 6.5, 9.25, -6.5, 1.49012e-08, 
# |                    -1.49012e-08, 4.93038e-32, -3, 1.49012e-08, -1.49012e-08, 
# |                    4.93038e-32, -3 ]
# | OutputProps:
# |   Height:          0
# |   Width:           0
# |   Depth:           0
# | ...
# | Got:
# | ---
# | Name:            Out
# | Format:          Float64
# | Stride:          32
# | Data:            [ 3.5, -7.75, -3.5, 6, -3.5, 6.5, 9.25, -6.5, 1.49012e-08, 
# |                    -1.49012e-08, 4.93038e-32, -3, 1.49012e-08, -1.49012e-08, 
# |                    0, -3 ]
# | OutputProps:
# |   Height:          0
# |   Width:           0
# |   Depth:           0
# | ...
# | Full Hex 64bit representation of Expected Buffer Values:
# | [ 0x3.5, 0x-7.75, 0x-3.5, 0x6, 0x-3.5, 0x6.5, 0x9.25, 0x-6.5, 0x1.49012e-08, 0x-1.49012e-08, 0x4.93038e-32, 0x-3, 0x1.49012e-08, 0x-1.49012e-08, 0x4.93038e-32, 0x-3 ]
# | Full Hex 64bit representation of Actual Buffer Values:
# | [ 0x3.5, 0x-7.75, 0x-3.5, 0x6, 0x-3.5, 0x6.5, 0x9.25, 0x-6.5, 0x1.49012e-08, 0x-1.49012e-08, 0x4.93038e-32, 0x-3, 0x1.49012e-08, 0x-1.49012e-08, 0x0, 0x-3 ]
# | 
# `-----------------------------
# error: command failed with exit status: 1

--

[damyanp]

@kmpeng - if confirmed this is a driver issue, get the appropriate internal bug filed.

[kmpeng]

The root cause appears to be DXC emitting GLSL.std.450 Fma instead of OpFmaKHR for the intrinsic (Godbolt). The GLSL instruction doesn't guarantee fused semantics, so the Intel driver is within spec when it produces this different result. Filed microsoft/DirectXShaderCompiler#8330 to track this on the DXC side.

[kmpeng]

Closing as this is now being tracked by the DXC issue microsoft/DirectXShaderCompiler#8330.