raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-20 19:07:01 +02:00
Code Issues Packages Projects Releases Wiki Activity

shader: support RDNA2 BVH64 ray intersections with shared traversal

Browse Source
This commit is contained in:
nmzik
2026-10-04 11:59:42 +04:00
parent 33e74cab2c
commit dbf66b7f4f
6 changed files with 77 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/shader/recompiler/backend/spirv/spirvEmitterBvh.cpp
+15 -9
View File
@@ -226,13 +226,13 @@ void DefineBvhIntersect(EmitterState& s) {
if (!s.requirements.bvh) return;
const auto u = TypeU32(s), f = TypeF32(s), b = TypeBool(s), wide = TypeScalarU64(s);
const auto vec4 = TypeU32Vector(s, 4), vec3 = TypeF32Vector(s, 3);
const auto signature = s.builder.Type(spv::OpTypeFunction, vec4, vec4, u, f, vec3, vec3, vec3);
const auto signature = s.builder.Type(spv::OpTypeFunction, vec4, vec4, wide, f, vec3, vec3, vec3);
s.bvh_intersect_function = s.builder.AllocateId();
s.builder.AddName(s.bvh_intersect_function, "bvh_intersect");
s.builder.AddFunction(spv::OpFunction, vec4, s.bvh_intersect_function,
spv::FunctionControlMaskNone, signature);
std::array<uint32_t, 6> args;
const std::array types {vec4, u, f, vec3, vec3, vec3};
const std::array types {vec4, wide, f, vec3, vec3, vec3};
for (uint32_t i = 0; i < args.size(); ++i) {
args[i] = s.builder.AllocateId();
s.builder.AddFunction(spv::OpFunctionParameter, types[i], args[i]);
@@ -253,7 +253,7 @@ void DefineBvhIntersect(EmitterState& s) {
const auto shr = [&](uint32_t value, uint32_t shift) {
return Binary(s, spv::OpShiftRightLogical, u, value, ConstantU32(s, shift));
};
const auto kind = and_bits(node, 7);
const auto kind = and_bits(Unary(s, spv::OpUConvert, u, node), 7);
const auto triangle = Binary(s, spv::OpULessThan, b, kind, ConstantU32(s, 4));
const auto full = Binary(s, spv::OpIEqual, b, kind, ConstantU32(s, 5));
const auto half = Binary(s, spv::OpIEqual, b, kind, ConstantU32(s, 4));
@@ -262,7 +262,7 @@ void DefineBvhIntersect(EmitterState& s) {
const auto grow = and_bits(shr(words[1], 23), 0xff);
const auto base = Binary(s, spv::OpShiftLeftLogical, wide,
DeviceAddressFromWords(s, words[0], and_bits(words[1], 0xff)), ConstantDeviceAddress(s, 8));
const auto index = Unary(s, spv::OpUConvert, wide, shr(node, 3));
const auto index = Binary(s, spv::OpShiftRightLogical, wide, node, ConstantDeviceAddress(s, 3));
const auto last_index = Binary(s, spv::OpIAdd, wide, index,
Select(s, wide, full, ConstantDeviceAddress(s, 1), ConstantDeviceAddress(s, 0)));
const auto address = Binary(s, spv::OpIAdd, wide, base,
@@ -325,17 +325,23 @@ void DefineBvhIntersect(EmitterState& s) {
uint32_t EmitBvhIntersect(ValueEmitContext& ctx, const IR::Inst& inst) {
auto& s = ctx.state;
const auto* ray = ctx.ImageAddress(inst.Arg(1));
if (ray == nullptr || ray->NumArgs() < 11 || s.bvh_intersect_function == 0)
ctx.Fail(inst, "BVH intersection requires eleven address components and its shared function");
const auto node_words = inst.Flags<uint32_t>();
if (ray == nullptr || (node_words != 1u && node_words != 2u) ||
ray->NumArgs() < node_words + 10u || s.bvh_intersect_function == 0)
ctx.Fail(inst, "BVH intersection requires its node pointer, ray components and shared function");
const auto vector = [&](uint32_t first) {
Vec3 values;
for (uint32_t i = 0; i < 3; ++i)
values[i] = Unary(s, spv::OpBitcast, TypeF32(s), ctx.Arg(*ray, first + i));
return Vector(s, TypeF32Vector(s, 3), values);
};
const auto descriptor = ctx.Arg(inst, 0), node = ctx.Arg(*ray, 0);
const auto extent = Unary(s, spv::OpBitcast, TypeF32(s), ctx.Arg(*ray, 1));
const auto origin = vector(2), direction = vector(5), inverse = vector(8);
const auto descriptor = ctx.Arg(inst, 0);
const auto node = node_words == 2u
? DeviceAddressFromWords(s, ctx.Arg(*ray, 0), ctx.Arg(*ray, 1))
: Unary(s, spv::OpUConvert, TypeScalarU64(s), ctx.Arg(*ray, 0));
const auto extent = Unary(s, spv::OpBitcast, TypeF32(s), ctx.Arg(*ray, node_words));
const auto origin = vector(node_words + 1u), direction = vector(node_words + 4u),
inverse = vector(node_words + 7u);
return EmitValueOrDefaultIfCondition(s, ctx.Arg(inst, 2), TypeU32Vector(s, 4),
ConstantU32CompositeZero(s, 4), [&] {
const auto result = s.builder.AllocateId();
src/graphics/shader/recompiler/frontend/decode/ImageOps.cpp
+5 -3
View File
@@ -228,7 +228,8 @@ Opcode DecodeMimgOpcode(uint32_t opcode, const MimgSampleInfo* sample, const Mim
case 0x09u: return Opcode::IMAGE_STORE_MIP;
case 0x0eu: return Opcode::IMAGE_GET_RESINFO;
case 0x60u: return Opcode::IMAGE_GET_LOD;
case 0xe6u: return Opcode::IMAGE_BVH_INTERSECT_RAY;
case 0xe6u:
case 0xe7u: return Opcode::IMAGE_BVH_INTERSECT_RAY;
default: return Opcode::UNSUPPORTED;
}
}
@@ -258,6 +259,7 @@ uint32_t DecodeMimgAddressComponents(uint32_t opcode, ImageDimension dimension,
switch (opcode) {
case 0xe6u: return 11u;
case 0xe7u: return 12u;
case 0x0eu: return 1u;
case 0x01u:
case 0x09u: return ImageCoordComponents(dimension) + 1u;
@@ -379,10 +381,10 @@ void DecodeMimg(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
if (d16 && !supports_d16) {
SetUnsupported(inst, Family::MIMG, opcode, "MIMG opcode does not support D16 data");
}
if (opcode == 0xe6u &&
if (inst.opcode == Opcode::IMAGE_BVH_INTERSECT_RAY &&
(a16 || !r128 || inst.dmask != 0xfu || (nsa_dwords != 0u && nsa_dwords != 3u))) {
SetUnsupported(inst, Family::MIMG, opcode,
"BVH intersection requires eleven full-float ray DWORDs and R128/dmask:0xf");
"BVH intersection requires full-float ray DWORDs and R128/dmask:0xf");
}
DecodeVectorGpr(vdata, inst.dst);
src/graphics/shader/recompiler/frontend/translate/Memory.cpp
+2 -1
View File
@@ -905,7 +905,8 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
switch (inst.opcode) {
case Decoder::Opcode::IMAGE_BVH_INTERSECT_RAY: {
const auto result = ir.Emit(IR::ValueOpcode::BvhIntersect,
{ConstructU32x4(inst.src1, 4), MakeImageAddress(inst, inst.src0), ir.GetExec()});
{ConstructU32x4(inst.src1, 4), MakeImageAddress(inst, inst.src0), ir.GetExec()},
inst.image_address_components - 10u);
for (uint32_t component = 0; component < 4; ++component) {
WriteOperand(OffsetOperand(inst.dst, component),
ir.Emit(IR::ValueOpcode::CompositeExtractU32x4, {result, IR::Value(component)}));