[HLSL] AMD-specific `QuantizedGemm` DML operator test failures

[Icohedron]

The following HLSL shaders and their corresponding DML operator tests are failing to pass on AMD (Radeon RX 6750 XT).

Shaders and corresponding failing DML operator tests

• QuantizedGemm_4096_64_16_uint4_packed32_float32_native_0
  • OperatorTests::QuantizedGemm#44
  • OperatorTests::QuantizedGemm#80
• QuantizedGemm_8192_64_16_uint4_packed32_float32_native_0
  • OperatorTests::QuantizedGemm#50
  • OperatorTests::QuantizedGemm#71

Sample Reproduction

AMD-only failure. This machine has a Radeon RX 6750 XT.

> .\TE.exe DirectML.Test.OperatorTests.dll /name:"OperatorTests::QuantizedGemm#44" /p:DisableMetacommands=1 /logOutput:low
Test Authoring and Execution Framework v10.72 for x64

StartGroup: OperatorTests::QuantizedGemm#44
Error: Output Tensor #0:
Error: Tensor Sizes: 1,1,1,10240
Error: Tensor Data Type: float32
Error: Index: 0000 @00000000 [0,0,0,0].  Ref: -4514.3652343750 (0xC58D12EC).  DML: -2227.7182617188 (0xC50B3B7E).  Abs: 2286.646973.  Rel: 50.652680%.  Ulp: 8509294
Error: Index: 0001 @00000001 [0,0,0,1].  Ref: -4617.8159179688 (0xC5904E87).  DML: -2282.3461914062 (0xC50EA58A).  Abs: 2335.469727.  Rel: 50.575202%.  Ulp: 8497405
Error: Index: 0002 @00000002 [0,0,0,2].  Ref: -4444.2119140625 (0xC58AE1B2).  DML: -2196.8825683594 (0xC5094E1F).  Abs: 2247.329346.  Rel: 50.567556%.  Ulp: 8491923
Error: Index: 0003 @00000003 [0,0,0,3].  Ref: -4426.8925781250 (0xC58A5724).  DML: -2132.1977539062 (0xC505432A).  Abs: 2294.694824.  Rel: 51.835340%.  Ulp: 8721402
Error: Index: 0004 @00000004 [0,0,0,4].  Ref: -4591.8095703125 (0xC58F7E7A).  DML: -2281.8234863281 (0xC50E9D2D).  Abs: 2309.986084.  Rel: 50.306661%.  Ulp: 8446285
Error: Index: 0005 @00000005 [0,0,0,5].  Ref: -4383.8754882812 (0xC588FF01).  DML: -2182.3945312500 (0xC5086650).  Abs: 2201.480957.  Rel: 50.217689%.  Ulp: 8427697
Error: Index: 0006 @00000006 [0,0,0,6].  Ref: -4476.1918945312 (0xC58BE189).  DML: -2244.4165039062 (0xC50C46AA).  Abs: 2231.775391.  Rel: 49.858796%.  Ulp: 8362719
Error: Index: 0007 @00000007 [0,0,0,7].  Ref: -4563.1181640625 (0xC58E98F2).  DML: -2235.4443359375 (0xC50BB71C).  Abs: 2327.673828.  Rel: 51.010597%.  Ulp: 8577494
Error: Index: 0008 @00000008 [0,0,0,8].  Ref: -4525.8417968750 (0xC58D6EBC).  DML: -2221.6650390625 (0xC50ADAA4).  Abs: 2304.176758.  Rel: 50.911562%.  Ulp: 8557592
Error: 5120 / 10240 (50.000000%) of elements were found to be above tolerance.
Error: Max absolute delta: 2478.656494.  Allowed absolute tolerance: 0.003000.
Error: Max ULP delta: 9057861.  Allowed tolerance: 10241 ULPs (float32).
Error: Verify: Fail [File: C:\__w\1\s\DirectML\SharedToolingLib\External\Test\TaefHelper\TaefHelper.cpp, Function: TaefHelper::Fail, Line: 133]
EndGroup: OperatorTests::QuantizedGemm#44 [Failed]

Summary of Non-passing Tests:
    OperatorTests::QuantizedGemm#44 [Failed]

Summary: Total=1, Passed=0, Failed=1, Blocked=0, Not Run=0, Skipped=0

[Icohedron]

This may not be a driver bug. The shader being used for AMD by these tests differs from the shaders used by NVIDIA and Intel.
These DML operator tests use a wave size of 64 threads on AMD instead of 32 on NVIDIA and Intel.

[Icohedron]

It would appear the issue is that Clang does not write the WaveSize to the DXIL, so the shader intended to be ran with a wave size of 64 is being executed with a wave size of 32. This error manifested as a wave active sum being roughly half the value it was supposed to be.
It also explains why an NVIDIA GPU could analyze the PIX capture of the Clang version of these DML operator tests shaders.

[Icohedron]

Keeping this issue open as I have not confirmed if the missing WaveSize attribute is the only reason these tests are failing.

[Icohedron]

Confirmed the missing WaveSize attribute was the only thing causing these tests to fail. Closing this issue as a duplicate of #70118