Share median classifications and preserve RDNA2 signed-zero selection

This commit is contained in:
nmzik
2026-10-19 00:59:31 +02:00
parent 75d7a4311d
commit defb2f11d3
3 changed files with 88 additions and 40 deletions
@@ -100,10 +100,9 @@ uint32_t EmitClassMaskF16(EmitterState& state, uint32_t bits, uint32_t mask) {
return EmitClassMaskBits(state, bits, mask, true);
}
uint32_t EmitMinMaxF32Value(EmitterState& state, uint32_t lhs, uint32_t rhs, bool max_value) {
const auto lhs_class = EmitClassifyF32(state, lhs);
const auto rhs_class = EmitClassifyF32(state, rhs);
static uint32_t EmitMinMaxF32(EmitterState& state, uint32_t lhs, uint32_t rhs,
const F32Class& lhs_class, const F32Class& rhs_class,
bool max_value) {
const auto numeric_cond = state.builder.AllocateId();
state.builder.AddFunction(max_value ? spv::OpFOrdGreaterThanEqual : spv::OpFOrdLessThan,
TypeBool(state), numeric_cond, lhs, rhs);
@@ -121,6 +120,34 @@ uint32_t EmitMinMaxF32Value(EmitterState& state, uint32_t lhs, uint32_t rhs, boo
return EmitBitcastU32ToF32(state, result_bits);
}
uint32_t EmitMinMaxF32Value(EmitterState& state, uint32_t lhs, uint32_t rhs, bool max_value) {
const auto lhs_class = EmitClassifyF32(state, lhs);
const auto rhs_class = EmitClassifyF32(state, rhs);
return EmitMinMaxF32(state, lhs, rhs, lhs_class, rhs_class, max_value);
}
uint32_t EmitFPMedTri32(EmitterState& state, uint32_t a, uint32_t b, uint32_t c) {
const auto a_class = EmitClassifyF32(state, a);
const auto b_class = EmitClassifyF32(state, b);
const auto c_class = EmitClassifyF32(state, c);
const auto min_ab = EmitMinMaxF32(state, a, b, a_class, b_class, false);
const auto min_ab_class = EmitClassifyF32(state, min_ab);
const auto min3 = EmitMinMaxF32(state, min_ab, c, min_ab_class, c_class, false);
const auto max_ab = EmitMinMaxF32(state, a, b, a_class, b_class, true);
const auto max_ab_class = EmitClassifyF32(state, max_ab);
const auto max3 = EmitMinMaxF32(state, max_ab, c, max_ab_class, c_class, true);
const auto max_bc = EmitMinMaxF32(state, b, c, b_class, c_class, true);
const auto max_ac = EmitMinMaxF32(state, a, c, a_class, c_class, true);
// RDNA2 compares numerically against MAX3, including ties between signed zeros.
const auto a_is_max = Binary(state, spv::OpFOrdEqual, TypeBool(state), max3, a);
const auto b_is_max = Binary(state, spv::OpFOrdEqual, TypeBool(state), max3, b);
const auto median = Select(state, TypeF32(state), a_is_max, max_bc,
Select(state, TypeF32(state), b_is_max, max_ac, max_ab));
const auto nan_ab = EmitLogicalOrBool(state, a_class.nan, b_class.nan);
const auto any_nan = EmitLogicalOrBool(state, nan_ab, c_class.nan);
return Select(state, TypeF32(state), any_nan, min3, median);
}
uint32_t EmitFlushF32DenormToSignedZero(EmitterState& state, uint32_t value) {
const auto bits = state.builder.AllocateId();
const auto abs_bits = state.builder.AllocateId();