shader: support high products of 24-bit integers

This commit is contained in:
nmzik committed 2026-10-28 22:49:21 +01:00
1 parent 22610cb93e
commit 54b937c85f
7 files changed
+128 -4

No files matched your search

@@ -249,7 +249,9 @@ enum class Opcode {
V_MIN_I32,
V_MAX_I32,
V_MUL_I32_I24,
V_MUL_HI_I32_I24,
V_MUL_U32_U24,
V_MUL_HI_U32_U24,
V_MUL_LO_U32,
V_MUL_HI_U32,
V_MUL_LO_I32,
@@ -35,7 +35,9 @@ constexpr Vop2OpcodeInfo VOP2_OPCODE_LIST[] = {
{0x05u, Opcode::V_SUBREV_F32},
{0x08u, Opcode::V_MUL_F32, Vop2SdwaProfile::Float32},
{0x09u, Opcode::V_MUL_I32_I24, Vop2SdwaProfile::IntegerPartialDestination},
{0x0au, Opcode::V_MUL_HI_I32_I24, Vop2SdwaProfile::IntegerPartialDestination},
{0x0bu, Opcode::V_MUL_U32_U24, Vop2SdwaProfile::IntegerPartialDestination},
{0x0cu, Opcode::V_MUL_HI_U32_U24, Vop2SdwaProfile::IntegerPartialDestination},
{0x0fu, Opcode::V_MIN_F32},
{0x10u, Opcode::V_MAX_F32},
{0x11u, Opcode::V_MIN_I32, Vop2SdwaProfile::IntegerFullDestination},
@@ -1539,7 +1541,8 @@ bool SupportsNativeVop3Clamp(Opcode opcode) {
return SupportsNativeVop3ResultModifiers(opcode) || UsesInexactClampControl(opcode) ||
opcode == Opcode::V_MAD_I16 || opcode == Opcode::V_MAD_U16 ||
opcode == Opcode::V_MIN3_I16 || opcode == Opcode::V_MAX3_I16 ||
opcode == Opcode::V_MAX3_U16 || opcode == Opcode::V_MED3_U16;
opcode == Opcode::V_MAX3_U16 || opcode == Opcode::V_MED3_U16 ||
opcode == Opcode::V_MUL_HI_I32_I24 || opcode == Opcode::V_MUL_HI_U32_U24;
}
bool HasUnsupportedNativeVop3Modifiers(Opcode opcode, bool permlane, bool mad_mix,
@@ -1570,6 +1573,10 @@ bool HasUnsupportedNativeVop3Modifiers(Opcode opcode, bool permlane, bool mad_mi
return clamp != 0u || omod != 0u || neg != 0u;
}
switch (opcode) {
case Opcode::V_MUL_HI_I32_I24:
case Opcode::V_MUL_HI_U32_U24:
// The high product fits in 32 bits, so clamp has no effect.
return abs != 0u || op_sel != 0u || omod != 0u || neg != 0u;
case Opcode::V_LDEXP_F32: return (abs & ~1u) != 0u || op_sel != 0u || (neg & ~1u) != 0u;
case Opcode::V_CVT_F32_UBYTE0:
case Opcode::V_CVT_F32_UBYTE1: