[SM6.9] Generate native vector variants of elementwise operators #7123
Assignees
Labels
Comments
pow2clk
changed the title
pow2clk
changed the title
pow2clk
added a commit
to pow2clk/DirectXShaderCompiler
that referenced
this issue
Feb 20, 2025
Disables various forms of scalarization and vector elimination to permit vectors to pass through to final DXIL when used in native LLVM operations and loading/storing. Introduces a few vector manipulation llvm instructions to DXIL allowing for them to appear in output DXIL. Skips passes for 6.9 that scalarize, convert to arrays, or otherwise eliminate vectors. This eliminates the element-by-element loading of the vectors In many cases, this required plumbing the shader model information to passes that didn't have it before. Many changes were needed for the MatrixBitcastLower pass related to linking to avoid converting matrix vectors, but also to perform the conversion if a shader was compiled for 6.9+, but then linked to a earlier target. This now adapts to the linker target to either preserve vectors for 6.9 or arrays for previous versions. This requires running the DynamicIndexing VectorToArray pass during linking since 6_x and 6_9+ will fail to run this in the initial compile, but will still need to lower vectors to arrays. Ternary conditional/select operators were element extracted in codegen. Removing this allows 6.9 to preserve the vectors, but also maintains behavior for previous shader models because the operations get scalarized later anyway. Keep groupshared variables as vectors for 6.9. They are no longer represented as indivual groupshared scalars. Adds extensive tests for these operations using different types and sizes and testing them appropriately. Booleans produce significantly different code, so they get their own test. Fixes microsoft#7123
pow2clk
added a commit
to pow2clk/DirectXShaderCompiler
that referenced
this issue
Mar 10, 2025
Disables various forms of scalarization and vector elimination to permit vectors to pass through to final DXIL when used in native LLVM operations and loading/storing. Introduces a few vector manipulation llvm instructions to DXIL allowing for them to appear in output DXIL. Skips passes for 6.9 that scalarize, convert to arrays, or otherwise eliminate vectors. This eliminates the element-by-element loading of the vectors In many cases, this required plumbing the shader model information to passes that didn't have it before. Many changes were needed for the MatrixBitcastLower pass related to linking to avoid converting matrix vectors, but also to perform the conversion if a shader was compiled for 6.9+, but then linked to a earlier target. This now adapts to the linker target to either preserve vectors for 6.9 or arrays for previous versions. This requires running the DynamicIndexing VectorToArray pass during linking since 6_x and 6_9+ will fail to run this in the initial compile, but will still need to lower vectors to arrays. Ternary conditional/select operators were element extracted in codegen. Removing this allows 6.9 to preserve the vectors, but also maintains behavior for previous shader models because the operations get scalarized later anyway. Keep groupshared variables as vectors for 6.9. They are no longer represented as indivual groupshared scalars. Adds extensive tests for these operations using different types and sizes and testing them appropriately. Booleans produce significantly different code, so they get their own test. Fixes microsoft#7123
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
In 6.9, elementwise vector operators should produce operators that work on the native vectors instead of scalarizing them. These operations include:
+,-(unary and binary),*,/,%,+=,-=,*=,/=|,&,^,~,|=,&=,^===,!=,<,>,<=,>=||,&&,?:Note that logic operations are disallowed for vectors as of HLSL2021 in favor of
and,or, andselectThis will largely be accomplished by disabling the scalarization passes that previously converted vectors to scalar variants. Testing should verify that the DXIL output preserves the vectors as arguments for all of these operators.
The text was updated successfully, but these errors were encountered: