shader: encode 11/10-bit float formatted buffer stores as truncated unsigned floats

This commit is contained in:
nmzik committed 2026-10-25 02:20:44 +02:00
1 parent 9a20c6f738
commit 59d8256b36
4 files changed
+49 -25

No files matched your search

@@ -54,9 +54,30 @@ uint32_t EmitExt(EmitterState& state, uint32_t type, uint32_t opcode,
return result;
}
uint32_t EmitF32ToF16RtzBits(EmitterState& state, uint32_t value) {
uint32_t EmitF32ToU32(EmitterState& state, uint32_t src, bool signed_value) {
// Clamping first keeps the truncating conversion defined; NaN and values past the largest
// float below the integer limit are selected afterwards.
const auto clamped =
EmitExt(state, TypeF32(state), GLSLstd450NClamp,
{src, ConstantF32(state, signed_value ? 0xcf000000u : 0u),
ConstantF32(state, signed_value ? 0x4effffffu : 0x4f7fffffu)});
const auto converted = signed_value
? Unary(state, spv::OpBitcast, TypeU32(state),
Unary(state, spv::OpConvertFToS, TypeI32(state), clamped))
: Unary(state, spv::OpConvertFToU, TypeU32(state), clamped);
const auto above = Binary(state, spv::OpFOrdGreaterThanEqual, TypeBool(state), src,
ConstantF32(state, signed_value ? 0x4f000000u : 0x4f800000u));
const auto high =
Select(state, TypeU32(state), above,
ConstantU32(state, signed_value ? 0x7fffffffu : 0xffffffffu), converted);
return Select(state, TypeU32(state), Unary(state, spv::OpIsNan, TypeBool(state), src),
ConstantU32(state, 0), high);
}
} // namespace
uint32_t EmitF32BitsToF16RtzBits(EmitterState& state, uint32_t bits) {
// Native half conversion causes regressions; use compact integer RTZ conversion.
const auto bits = EmitBitcastF32ToU32(state, value);
const auto abs = EmitAndConstant(state, bits, 0x7fffffffu);
const auto sign = EmitAndConstant(
state, Binary(state, spv::OpShiftRightLogical, TypeU32(state), bits, ConstantU32(state, 16)),
@@ -92,28 +113,6 @@ uint32_t EmitF32ToF16RtzBits(EmitterState& state, uint32_t value) {
return EmitOrU32(state, sign, magnitude);
}
uint32_t EmitF32ToU32(EmitterState& state, uint32_t src, bool signed_value) {
// Clamping first keeps the truncating conversion defined; NaN and values past the largest
// float below the integer limit are selected afterwards.
const auto clamped =
EmitExt(state, TypeF32(state), GLSLstd450NClamp,
{src, ConstantF32(state, signed_value ? 0xcf000000u : 0u),
ConstantF32(state, signed_value ? 0x4effffffu : 0x4f7fffffu)});
const auto converted = signed_value
? Unary(state, spv::OpBitcast, TypeU32(state),
Unary(state, spv::OpConvertFToS, TypeI32(state), clamped))
: Unary(state, spv::OpConvertFToU, TypeU32(state), clamped);
const auto above = Binary(state, spv::OpFOrdGreaterThanEqual, TypeBool(state), src,
ConstantF32(state, signed_value ? 0x4f000000u : 0x4f800000u));
const auto high =
Select(state, TypeU32(state), above,
ConstantU32(state, signed_value ? 0x7fffffffu : 0xffffffffu), converted);
return Select(state, TypeU32(state), Unary(state, spv::OpIsNan, TypeBool(state), src),
ConstantU32(state, 0), high);
}
} // namespace
uint32_t EmitFPFma32(EmitterState& state, uint32_t a, uint32_t b, uint32_t c) {
const auto result = EmitGlsl<GLSLstd450Fma, IR::Type::F32>(state, a, b, c);
state.builder.AddAnnotation(spv::OpDecorate, result, spv::DecorationNoContraction);
@@ -199,7 +198,8 @@ uint32_t EmitCompositeExtractU32x2(EmitterState& state, uint32_t arg0, IR::Value
uint32_t EmitPackFloat2x16Rtz(EmitterState& state, uint32_t arg0, uint32_t arg1) {
return EmitNative<spv::OpBitFieldInsert, IR::Type::U32>(
state, EmitF32ToF16RtzBits(state, arg0), EmitF32ToF16RtzBits(state, arg1),
state, EmitF32BitsToF16RtzBits(state, EmitBitcastF32ToU32(state, arg0)),
EmitF32BitsToF16RtzBits(state, EmitBitcastF32ToU32(state, arg1)),
ConstantU32(state, 16), ConstantU32(state, 16));
}