Skip to content

[SPIR-V] Fix precision for dot2add #7861

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
s-perron merged 2 commits into from
Nov 10, 2025
Merged

[SPIR-V] Fix precision for dot2add #7861

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
e5ec194
Fix precision for .
luciechoi Oct 31, 2025
6a495e2
Multiplication happens with half precision
luciechoi Nov 1, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
21 changes: 14 additions & 7 deletions tools/clang/lib/SPIRV/SpirvEmitter.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13165,17 +13165,24 @@ SpirvInstruction *SpirvEmitter::processIntrinsicDP2a(const CallExpr *callExpr) {
SpirvInstruction *arg1Instr = doExpr(arg1);
SpirvInstruction *arg2Instr = doExpr(arg2);

// Create the dot product of the half2 vectors.
SpirvInstruction *dotInstr = spvBuilder.createBinaryOp(
spv::Op::OpDot, componentType, arg0Instr, arg1Instr, loc, range);
// Convert the half vectors into float vectors to preserve precision
// as specified in the HLSL documentation.
arg0Instr = spvBuilder.createUnaryOp(
spv::Op::OpFConvert,
astContext.getExtVectorType(astContext.FloatTy, vecSize), arg0Instr, loc,
range);
arg1Instr = spvBuilder.createUnaryOp(
spv::Op::OpFConvert,
astContext.getExtVectorType(astContext.FloatTy, vecSize), arg1Instr, loc,
range);

// Convert dot product (half type) to result type (float).
// Create the dot product of the half2 vectors.
QualType resultType = callExpr->getType();
SpirvInstruction *floatDotInstr = spvBuilder.createUnaryOp(
spv::Op::OpFConvert, resultType, dotInstr, loc, range);
SpirvInstruction *dotInstr = spvBuilder.createBinaryOp(
spv::Op::OpDot, resultType, arg0Instr, arg1Instr, loc, range);

// Sum the dot product result and accumulator and return.
return spvBuilder.createBinaryOp(spv::Op::OpFAdd, resultType, floatDotInstr,
return spvBuilder.createBinaryOp(spv::Op::OpFAdd, resultType, dotInstr,
arg2Instr, loc, range);
}

Expand Down
15 changes: 9 additions & 6 deletions tools/clang/test/CodeGenSPIRV/intrinsics.dot2add.hlsl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,8 +12,9 @@ float main() : SV_Target {
// CHECK: [[input1A:%[0-9]+]] = OpLoad %v2half %input1A
// CHECK: [[input1B:%[0-9]+]] = OpLoad %v2half %input1B
// CHECK: [[acc1:%[0-9]+]] = OpLoad %float %acc1
// CHECK: [[dot1_0:%[0-9]+]] = OpDot %half [[input1A]] [[input1B]]
// CHECK: [[dot1:%[0-9]+]] = OpFConvert %float [[dot1_0]]
// CHECK: [[input1A_0:%[0-9]+]] = OpFConvert %v2float [[input1A]]
// CHECK: [[input1B_0:%[0-9]+]] = OpFConvert %v2float [[input1B]]
// CHECK: [[dot1:%[0-9]+]] = OpDot %float [[input1A_0]] [[input1B_0]]
// CHECK: [[res1:%[0-9]+]] = OpFAdd %float [[dot1]] [[acc1]]
res += dot2add(input1A, input1B, acc1);

Expand All @@ -25,8 +26,9 @@ float main() : SV_Target {
// CHECK: [[input2B:%[0-9]+]] = OpVectorShuffle %v2half [[input2_1]] [[input2_1]] 2 3
// CHECK: [[acc2_0:%[0-9]+]] = OpLoad %int %acc2
// CHECK: [[acc2:%[0-9]+]] = OpConvertSToF %float [[acc2_0]]
// CHECK: [[dot2_0:%[0-9]+]] = OpDot %half [[input2A]] [[input2B]]
// CHECK: [[dot2:%[0-9]+]] = OpFConvert %float [[dot2_0]]
// CHECK: [[input2A_0:%[0-9]+]] = OpFConvert %v2float [[input2A]]
// CHECK: [[input2B_0:%[0-9]+]] = OpFConvert %v2float [[input2B]]
// CHECK: [[dot2:%[0-9]+]] = OpDot %float [[input2A_0]] [[input2B_0]]
// CHECK: [[res2:%[0-9]+]] = OpFAdd %float [[dot2]] [[acc2]]
res += dot2add(input2.xy, input2.zw, acc2);

Expand All @@ -44,8 +46,9 @@ float main() : SV_Target {
// CHECK: [[input3B:%[0-9]+]] = OpConvertSToF %v2half [[input3B_5]]
// CHECK: [[acc3_1:%[0-9]+]] = OpLoad %half %acc3
// CHECK: [[acc3:%[0-9]+]] = OpFConvert %float [[acc3_1]]
// CHECK: [[dot3_1:%[0-9]+]] = OpDot %half [[input3A]] [[input3B]]
// CHECK: [[dot3:%[0-9]+]] = OpFConvert %float [[dot3_1]]
// CHECK: [[input3A_0:%[0-9]+]] = OpFConvert %v2float [[input3A]]
// CHECK: [[input3B_0:%[0-9]+]] = OpFConvert %v2float [[input3B]]
// CHECK: [[dot3:%[0-9]+]] = OpDot %float [[input3A_0]] [[input3B_0]]
// CHECK: [[res3:%[0-9]+]] = OpFAdd %float [[dot3]] [[acc3]]
res += dot2add(input3A, input3B, acc3);

Expand Down