dxc generates invalid alignment on groupshared matrix load/store instructions #6416

dmpots · 2024-03-14T18:38:35Z

Description
When loading or storing a matrix to groupshared dxc is using an alignment on the instruction that is not guaranteed by the underlying global variable.

Steps to Reproduce
https://godbolt.org/z/9c9os31MW

// dxc /Tcs_6_0 alignbug.hlsl
struct Data {
   float4x4 m;
};

groupshared Data GData;
StructuredBuffer<Data> input : register(t0);
RWStructuredBuffer<Data> output : register(u0);

[RootSignature("SRV(t0), UAV(u0)")]
[numthreads(128,1,1)]
void main(uint Id : SV_DispatchThreadId, uint g : SV_GroupID)
{
   GData = input[0];
   GroupMemoryBarrierWithGroupSync();
   output[Id] = GData;

}

Actual Behavior
Compiling this shader produces loads and stores that do not match the alignment of the underlying global.

The global does not specify an alignment so it will get the default alignment of 4 for this type

@"\01?GData@@3UData@@A.0.v.v" = addrspace(3) global [16 x float] undef

But the load

%22 = load float, float addrspace(3)* getelementptr inbounds ([16 x float], [16 x float] addrspace(3)* @"\01?GData@@3UData@@A.0.v.v", i32 0, i32 0), align 16

and store

store float %3, float addrspace(3)* getelementptr inbounds ([16 x float], [16 x float] addrspace(3)* @"\01?GData@@3UData@@A.0.v.v", i32 0, i32 0), align 16

both specify an alignment of align 16 for the first element (other elements are overaligned as well).

From the llvm language reference manual, the aligment on the instruction should be guaranteed by somehow, but the alignment of the global does not do it.

The optional constant align argument specifies the alignment of the operation (that is, the alignment of the memory address). It is the responsibility of the code emitter to ensure that the alignment information is correct.

Environment

DXC version dxc main as of 3/14/2024
Host Operating System windows

The text was updated successfully, but these errors were encountered:

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

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 #6416

…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)

dmpots added bug Bug, regression, crash needs-triage Awaiting triage labels Mar 14, 2024

damyanp changed the title ~~dxc generates invalid alignment on groupshared matric load/store instructions~~ dxc generates invalid alignment on groupshared matrix load/store instructions Mar 15, 2024

damyanp removed the needs-triage Awaiting triage label Mar 15, 2024

damyanp added this to the Next Release milestone Mar 15, 2024

damyanp added matrix-bug Bugs relating to matrix types correctness Bugs that impact shader correctness labels Mar 15, 2024

damyanp assigned pow2clk Apr 22, 2024

pow2clk mentioned this issue May 6, 2024

Retain alignment when lowering groupshared matrices #6589

Merged

pow2clk closed this as completed in #6589 May 22, 2024

pow2clk mentioned this issue May 22, 2024

Calculate preferred alignment when lowering groupshared matrices #6645

Closed

pow2clk removed this from the May 2024 Release milestone May 22, 2024

damyanp added this to the Next 2024 Release milestone May 22, 2024

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

dxc generates invalid alignment on groupshared matrix load/store instructions #6416

dxc generates invalid alignment on groupshared matrix load/store instructions #6416

dmpots commented Mar 14, 2024

dxc generates invalid alignment on groupshared matrix load/store instructions #6416

dxc generates invalid alignment on groupshared matrix load/store instructions #6416

Comments

dmpots commented Mar 14, 2024