mirror of
https://github.com/KytyPS5/KytyPS5.git
synced 2026-10-20 02:47:02 +02:00
211 lines
11 KiB
C++
211 lines
11 KiB
C++
#include "graphics/shader/recompiler/backend/spirv/spirvEmitterInternal.h"
|
|
|
|
namespace Libs::Graphics::ShaderRecompiler::Spirv::Emitter {
|
|
|
|
uint32_t EmitAndConstant(EmitterState& state, uint32_t value, uint32_t mask) {
|
|
const auto ret = state.builder.AllocateId();
|
|
state.builder.AddFunction(spv::OpBitwiseAnd, TypeU32(state), ret, value,
|
|
ConstantU32(state, mask));
|
|
return ret;
|
|
}
|
|
|
|
uint32_t EmitCompareU32Constant(EmitterState& state, spv::Op opcode, uint32_t value,
|
|
uint32_t constant) {
|
|
const auto ret = state.builder.AllocateId();
|
|
state.builder.AddFunction(opcode, TypeBool(state), ret, value, ConstantU32(state, constant));
|
|
return ret;
|
|
}
|
|
|
|
F32Class EmitClassifyF32Bits(EmitterState& state, uint32_t bits) {
|
|
F32Class cls;
|
|
cls.bits = bits;
|
|
const auto abs_bits = EmitAndConstant(state, cls.bits, 0x7fffffffu);
|
|
cls.nan = EmitCompareU32Constant(state, spv::OpUGreaterThan, abs_bits, 0x7f800000u);
|
|
cls.zero = EmitCompareU32Constant(state, spv::OpIEqual, abs_bits, 0);
|
|
return cls;
|
|
}
|
|
|
|
F32Class EmitClassifyF32(EmitterState& state, uint32_t value) {
|
|
return EmitClassifyF32Bits(state, EmitBitcastF32ToU32(state, value));
|
|
}
|
|
|
|
uint32_t EmitClassMaskBitMatch(EmitterState& state, uint32_t mask, uint32_t bit,
|
|
uint32_t class_match) {
|
|
const auto selected =
|
|
EmitCompareU32Constant(state, spv::OpINotEqual, EmitAndConstant(state, mask, 1u << bit), 0);
|
|
return EmitLogicalAndBool(state, selected, class_match);
|
|
}
|
|
|
|
static uint32_t EmitClassMaskBits(EmitterState& state, uint32_t bits, uint32_t mask, bool half) {
|
|
const uint32_t sign_mask = half ? 0x8000u : 0x80000000u;
|
|
const uint32_t exponent_mask = half ? 0x7c00u : 0x7f800000u;
|
|
const uint32_t mantissa_mask = half ? 0x03ffu : 0x007fffffu;
|
|
const uint32_t quiet_mask = half ? 0x0200u : 0x00400000u;
|
|
const auto sign_bits = EmitAndConstant(state, bits, sign_mask);
|
|
const auto abs_bits = EmitAndConstant(state, bits, sign_mask - 1u);
|
|
const auto exponent_bits = EmitAndConstant(state, abs_bits, exponent_mask);
|
|
const auto mantissa_bits = EmitAndConstant(state, abs_bits, mantissa_mask);
|
|
const auto quiet_bits = EmitAndConstant(state, mantissa_bits, quiet_mask);
|
|
|
|
const auto sign = EmitCompareU32Constant(state, spv::OpINotEqual, sign_bits, 0);
|
|
const auto positive = EmitLogicalNotBool(state, sign);
|
|
const auto exponent_zero = EmitCompareU32Constant(state, spv::OpIEqual, exponent_bits, 0);
|
|
const auto exponent_nonzero = EmitLogicalNotBool(state, exponent_zero);
|
|
const auto exponent_inf =
|
|
EmitCompareU32Constant(state, spv::OpIEqual, exponent_bits, exponent_mask);
|
|
const auto finite_exponent = EmitLogicalNotBool(state, exponent_inf);
|
|
const auto mantissa_zero = EmitCompareU32Constant(state, spv::OpIEqual, mantissa_bits, 0);
|
|
const auto mantissa_nonzero = EmitLogicalNotBool(state, mantissa_zero);
|
|
const auto quiet = EmitCompareU32Constant(state, spv::OpINotEqual, quiet_bits, 0);
|
|
|
|
const auto nan_common = EmitLogicalAndBool(state, exponent_inf, mantissa_nonzero);
|
|
const auto snan = EmitLogicalAndBool(state, nan_common, EmitLogicalNotBool(state, quiet));
|
|
const auto qnan = EmitLogicalAndBool(state, nan_common, quiet);
|
|
const auto inf = EmitLogicalAndBool(state, exponent_inf, mantissa_zero);
|
|
const auto normal = EmitLogicalAndBool(state, exponent_nonzero, finite_exponent);
|
|
const auto denorm = EmitLogicalAndBool(state, exponent_zero, mantissa_nonzero);
|
|
const auto zero = EmitCompareU32Constant(state, spv::OpIEqual, abs_bits, 0);
|
|
|
|
uint32_t match = EmitClassMaskBitMatch(state, mask, 0, snan);
|
|
match = EmitLogicalOrBool(state, match, EmitClassMaskBitMatch(state, mask, 1, qnan));
|
|
match = EmitLogicalOrBool(
|
|
state, match, EmitClassMaskBitMatch(state, mask, 2, EmitLogicalAndBool(state, inf, sign)));
|
|
match = EmitLogicalOrBool(
|
|
state, match,
|
|
EmitClassMaskBitMatch(state, mask, 3, EmitLogicalAndBool(state, normal, sign)));
|
|
match = EmitLogicalOrBool(
|
|
state, match,
|
|
EmitClassMaskBitMatch(state, mask, 4, EmitLogicalAndBool(state, denorm, sign)));
|
|
match = EmitLogicalOrBool(
|
|
state, match, EmitClassMaskBitMatch(state, mask, 5, EmitLogicalAndBool(state, zero, sign)));
|
|
match = EmitLogicalOrBool(
|
|
state, match,
|
|
EmitClassMaskBitMatch(state, mask, 6, EmitLogicalAndBool(state, zero, positive)));
|
|
match = EmitLogicalOrBool(
|
|
state, match,
|
|
EmitClassMaskBitMatch(state, mask, 7, EmitLogicalAndBool(state, denorm, positive)));
|
|
match = EmitLogicalOrBool(
|
|
state, match,
|
|
EmitClassMaskBitMatch(state, mask, 8, EmitLogicalAndBool(state, normal, positive)));
|
|
return EmitLogicalOrBool(
|
|
state, match,
|
|
EmitClassMaskBitMatch(state, mask, 9, EmitLogicalAndBool(state, inf, positive)));
|
|
}
|
|
|
|
uint32_t EmitClassMaskF32(EmitterState& state, uint32_t value, uint32_t mask) {
|
|
return EmitClassMaskBits(state, EmitBitcastF32ToU32(state, value), mask, false);
|
|
}
|
|
|
|
uint32_t EmitClassMaskF16(EmitterState& state, uint32_t bits, uint32_t mask) {
|
|
return EmitClassMaskBits(state, bits, mask, true);
|
|
}
|
|
|
|
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);
|
|
const auto ordered_bits =
|
|
EmitSelectValueU32(state, numeric_cond, lhs_class.bits, rhs_class.bits);
|
|
|
|
const auto both_zero = EmitLogicalAndBool(state, lhs_class.zero, rhs_class.zero);
|
|
const auto zero_bits = max_value ? EmitAndU32(state, lhs_class.bits, rhs_class.bits)
|
|
: EmitOrU32(state, lhs_class.bits, rhs_class.bits);
|
|
const auto numeric_bits = EmitSelectValueU32(state, both_zero, zero_bits, ordered_bits);
|
|
|
|
const auto rhs_nan_bits =
|
|
EmitSelectValueU32(state, rhs_class.nan, lhs_class.bits, numeric_bits);
|
|
const auto result_bits = EmitSelectValueU32(state, lhs_class.nan, rhs_class.bits, rhs_nan_bits);
|
|
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();
|
|
const auto sign_bits = state.builder.AllocateId();
|
|
const auto subnormal = state.builder.AllocateId();
|
|
const auto selected = state.builder.AllocateId();
|
|
const auto ret = state.builder.AllocateId();
|
|
state.builder.AddFunction(spv::OpBitcast, TypeU32(state), bits, value);
|
|
state.builder.AddFunction(spv::OpBitwiseAnd, TypeU32(state), abs_bits, bits,
|
|
ConstantU32(state, 0x7fffffffu));
|
|
state.builder.AddFunction(spv::OpBitwiseAnd, TypeU32(state), sign_bits, bits,
|
|
ConstantU32(state, 0x80000000u));
|
|
state.builder.AddFunction(spv::OpULessThan, TypeBool(state), subnormal, abs_bits,
|
|
ConstantU32(state, 0x00800000u));
|
|
state.builder.AddFunction(spv::OpSelect, TypeU32(state), selected, subnormal, sign_bits, bits);
|
|
state.builder.AddFunction(spv::OpBitcast, TypeF32(state), ret, selected);
|
|
return ret;
|
|
}
|
|
|
|
uint32_t EmitTrigCycleF32(EmitterState& state, uint32_t src, bool preserve_signed_zero) {
|
|
const auto fract = state.builder.AllocateId();
|
|
const auto bits = state.builder.AllocateId();
|
|
const auto abs_bits = state.builder.AllocateId();
|
|
const auto large = state.builder.AllocateId();
|
|
const auto finite = state.builder.AllocateId();
|
|
const auto large_finite = state.builder.AllocateId();
|
|
const auto reduced = state.builder.AllocateId();
|
|
state.builder.AddFunction(spv::OpExtInst, TypeF32(state), fract, GlslStd450(state),
|
|
GLSLstd450Fract, src);
|
|
state.builder.AddFunction(spv::OpBitcast, TypeU32(state), bits, src);
|
|
state.builder.AddFunction(spv::OpBitwiseAnd, TypeU32(state), abs_bits, bits,
|
|
ConstantU32(state, 0x7fffffffu));
|
|
state.builder.AddFunction(spv::OpUGreaterThanEqual, TypeBool(state), large, abs_bits,
|
|
ConstantU32(state, 0x4b000000u));
|
|
state.builder.AddFunction(spv::OpULessThan, TypeBool(state), finite, abs_bits,
|
|
ConstantU32(state, 0x7f800000u));
|
|
state.builder.AddFunction(spv::OpLogicalAnd, TypeBool(state), large_finite, large, finite);
|
|
state.builder.AddFunction(spv::OpSelect, TypeF32(state), reduced, large_finite,
|
|
ConstantF32(state, 0), fract);
|
|
if (!preserve_signed_zero) {
|
|
return reduced;
|
|
}
|
|
const auto zero = state.builder.AllocateId();
|
|
const auto ret = state.builder.AllocateId();
|
|
state.builder.AddFunction(spv::OpIEqual, TypeBool(state), zero, abs_bits,
|
|
ConstantU32(state, 0));
|
|
state.builder.AddFunction(spv::OpSelect, TypeF32(state), ret, zero, src, reduced);
|
|
return ret;
|
|
}
|
|
|
|
uint32_t EmitF16BitsToF32(EmitterState& state, uint32_t bits) {
|
|
const auto unpacked = state.builder.AllocateId();
|
|
const auto ret = state.builder.AllocateId();
|
|
state.builder.AddFunction(spv::OpExtInst, TypeF32Vector(state, 2), unpacked, GlslStd450(state),
|
|
GLSLstd450UnpackHalf2x16, bits);
|
|
state.builder.AddFunction(spv::OpCompositeExtract, TypeF32(state), ret, unpacked, 0);
|
|
return ret;
|
|
}
|
|
|
|
} // namespace Libs::Graphics::ShaderRecompiler::Spirv::Emitter
|