Retain alignment when lowering groupshared matrices #6589
Conversation
When flattening the global for a groupshared matrix, the alignment information was getting lost. As a result, the alignments of the loads and stores were calculating their own alignment based on preferred alignment and trailing zeros of the index. The preferred alignment switched to 16 when the type size was over 128 bits due to a heuristic whose rationale is lost to time. When the global has its own alignment, that gets used, so by retaining it through lowering, the alignments are consistent and reliable. Includes testing for a few matrix variants fixes microsoft#6416
...LSLFileCheck/hlsl/types/modifiers/groupshared/groupshared-member-matrix-subscript-align.hlsl
Show resolved
Hide resolved
...LSLFileCheck/hlsl/types/modifiers/groupshared/groupshared-member-matrix-subscript-align.hlsl
Outdated
Show resolved
Hide resolved
Added verification that dimensions and variables are consistent
It's certainly an option. What I have here preserves the alignment that is present from the start. I'm a bit more nervous about interjecting the alignment that the zero index happened to get when loads and stores were generated from the beginning. I fear that might have more unforeseen consequences than this does, but I can experiment with it. |
To keep consistent with previous behavior, use the preferred alignment calculation when assigning alignment to the matrix groupshared instead of preserving the original value.
Add half and double variants for test cases Add pass test for hlmatrixlower
In the updated PR we are using the datalayout to compute the preferred alignment. This looks like the correct way to do it to me. |
…rosoft#6589) When flattening the global for a groupshared matrix, the alignment information was getting lost. As a result, the alignments of the loads and stores were calculating their own alignment based on preferred alignment and trailing zeros of the index. The preferred alignment switched to 16 when the type size was over 128 bits due to a heuristic whose rationale is lost to time. When the global has its own alignment, that gets used, so by calculating it at lowering, the alignments are consistent and reliable. Includes testing for a few matrix variants and a pass test. fixes microsoft#6416 (cherry picked from commit a6f4025)
| // RUN: %dxc -DTYPE=float1x4 /Tcs_6_0 %s | FileCheck %s -check-prefixes=CHECK,CHECK4 | ||
| // RUN: %dxc -DTYPE=float2x2 /Tcs_6_0 %s | FileCheck %s -check-prefixes=CHECK,CHECK4 | ||
| // RUN: %dxc -DTYPE=double4x4 /Tcs_6_0 %s | FileCheck %s -check-prefixes=CHECK,LCHECK,CHECK8,LCHECKF | ||
| // RUN: %dxc -DTYPE=float16_t4x4 /Tcs_6_2 %s -enable-16bit-types | FileCheck %s -check-prefixes=CHECK,LCHECK,CHECK2,LCHECKF |
There was a problem hiding this comment.
I think it might be important to also test with a matrix array, non-zero index, and odd matrix sizes, like (3x3, 1x3, 3x1). Since the alignment of the matrix is considered to be the alignment of the element, different array elements for a matrix whose size isn't an even multiple of 16 bytes won't be aligned by that amount.
…rosoft#6589) When flattening the global for a groupshared matrix, the alignment information was getting lost. As a result, the alignments of the loads and stores were calculating their own alignment based on preferred alignment and trailing zeros of the index. The preferred alignment switched to 16 when the type size was over 128 bits due to a heuristic whose rationale is lost to time. When the global has its own alignment, that gets used, so by calculating it at lowering, the alignments are consistent and reliable. Includes testing for a few matrix variants and a pass test. fixes microsoft#6416
When flattening the global for a groupshared matrix, the alignment information was getting lost. As a result, the alignments of the loads and stores were calculating their own alignment based on preferred alignment and trailing zeros of the index. The preferred alignment switched to 16 when the type size was over 128 bits due to a heuristic whose rationale is lost to time. When the global has its own alignment, that gets used, so by retaining it through lowering, the alignments are consistent and reliable.
Includes testing for a few matrix variants
fixes #6416