shader: add BVH8 intersections and instance transforms

src/graphics/shader/recompiler/backend/spirv/spirvEmitterBvh.cpp

@@ -30,16 +30,7 @@ uint32_t NodeWord(EmitterState& s, uint32_t block, uint32_t word) {
 uint32_t Triangle(EmitterState& s, uint32_t block, uint32_t kind, uint32_t bary,
                   Vec3 origin, Vec3 direction) {
 	const auto f = TypeF32(s), u = TypeU32(s), b = TypeBool(s);
-	const auto zero = ConstantF32Value(s, 0.0f);
-	const auto op = [&](spv::Op code, uint32_t x, uint32_t y) {
-		const auto value = Binary(s, code, f, x, y);
-		s.builder.AddAnnotation(spv::OpDecorate, value, spv::DecorationNoContraction);
-		return value;
-	};
 	const auto cmp = [&](spv::Op code, uint32_t x, uint32_t y) { return Binary(s, code, b, x, y); };
-	const auto either = [&](uint32_t x, uint32_t y) { return Binary(s, spv::OpLogicalOr, b, x, y); };
-	const auto both = [&](uint32_t x, uint32_t y) { return Binary(s, spv::OpLogicalAnd, b, x, y); };
-	const auto abs = [&](uint32_t x) { return EmitGlsl<GLSLstd450FAbs, IR::Type::F32>(s, x); };
 	const auto is_kind = [&](uint32_t n) { return cmp(spv::OpIEqual, kind, ConstantU32(s, n)); };
 	// The four triangles are (0,1,2), (1,3,2), (2,3,4), (2,4,0).
 	const std::array vertices {
@@ -56,6 +47,28 @@ uint32_t Triangle(EmitterState& s, uint32_t block, uint32_t kind, uint32_t bary,
 			    NodeWord(s, block, Binary(s, spv::OpIAdd, u, first, ConstantU32(s, axis))));
 	}
 	const auto flag = NodeWord(s, block, ConstantU32(s, 15));
+	const auto remap = Binary(s, spv::OpShiftRightLogical, u, flag,
+	    Binary(s, spv::OpIMul, u, kind, ConstantU32(s, 8)));
+	return EmitBvhTriangle(s, positions, origin, direction, remap,
+	    Binary(s, spv::OpIAdd, u, flag, kind), bary);
+}
+
+} // namespace
+
+uint32_t EmitBvhTriangle(EmitterState& s, const std::array<Vec3, 3>& positions,
+                         Vec3 origin, Vec3 direction, uint32_t remap,
+                         uint32_t triangle_id, uint32_t bary) {
+	const auto f = TypeF32(s), u = TypeU32(s), b = TypeBool(s);
+	const auto zero = ConstantF32Value(s, 0.0f);
+	const auto op = [&](spv::Op code, uint32_t x, uint32_t y) {
+		const auto value = Binary(s, code, f, x, y);
+		s.builder.AddAnnotation(spv::OpDecorate, value, spv::DecorationNoContraction);
+		return value;
+	};
+	const auto cmp = [&](spv::Op code, uint32_t x, uint32_t y) { return Binary(s, code, b, x, y); };
+	const auto either = [&](uint32_t x, uint32_t y) { return Binary(s, spv::OpLogicalOr, b, x, y); };
+	const auto both = [&](uint32_t x, uint32_t y) { return Binary(s, spv::OpLogicalAnd, b, x, y); };
+	const auto abs = [&](uint32_t x) { return EmitGlsl<GLSLstd450FAbs, IR::Type::F32>(s, x); };
 	const auto y_largest = cmp(spv::OpFOrdLessThan, abs(direction[0]), abs(direction[1]));
 	const auto z_largest = cmp(spv::OpFOrdLessThan,
 	    Select(s, f, y_largest, abs(direction[1]), abs(direction[0])), abs(direction[2]));
@@ -112,8 +125,6 @@ uint32_t Triangle(EmitterState& s, uint32_t block, uint32_t kind, uint32_t bary,
 		    cmp(spv::OpFOrdGreaterThan, projected[i][1], zero), right);
 		missed = either(missed, both(cmp(spv::OpFOrdEqual, edge[i], zero), excluded));
 	}
-	const auto remap = Binary(s, spv::OpShiftRightLogical, u, flag,
-	    Binary(s, spv::OpIMul, u, kind, ConstantU32(s, 8)));
 	const auto coordinate = [&](uint32_t shift) {
 		const auto index = Binary(s, spv::OpBitwiseAnd, u,
 		    Binary(s, spv::OpShiftRightLogical, u, remap, ConstantU32(s, shift)), ConstantU32(s, 3));
@@ -124,11 +135,13 @@ uint32_t Triangle(EmitterState& s, uint32_t block, uint32_t kind, uint32_t bary,
 	const std::array result {
 		Unary(s, spv::OpBitcast, u, Select(s, f, missed, ConstantF32(s, 0x7f800000), numerator)),
 		Unary(s, spv::OpBitcast, u, Select(s, f, missed, ConstantF32Value(s, 1.0f), denominator)),
-		Select(s, u, bary, coordinate(0), Binary(s, spv::OpIAdd, u, flag, kind)),
+		Select(s, u, bary, coordinate(0), triangle_id),
 		Select(s, u, bary, coordinate(2), Select(s, u, missed, ConstantU32(s, 0), ConstantU32(s, 1))) };
 	return Vector(s, TypeU32Vector(s, 4), result);
 }
 
+namespace {
+
 std::array<uint32_t, 6> Box(EmitterState& s, uint32_t first, uint32_t second,
                            uint32_t half, uint32_t child) {
 	const auto f = TypeF32(s);

src/graphics/shader/recompiler/backend/spirv/spirvEmitterInstructions.h

@@ -338,6 +338,8 @@ inline constexpr auto EmitDataConsume           = EmitAppendConsume;
 uint32_t              EmitSwizzleU32(ValueEmitContext& ctx, const IR::Inst& inst);
 void                  EmitImage(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitBvhIntersect(ValueEmitContext& ctx, const IR::Inst& inst);
+uint32_t              EmitBvh8Intersect(ValueEmitContext& ctx, const IR::Inst& inst);
+uint32_t              EmitBvh8Extract(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitBvhStack(ValueEmitContext& ctx, const IR::Inst& inst);
 inline constexpr auto EmitImageQueryDimensions = EmitImage;
 inline constexpr auto EmitImageQueryLod        = EmitImage;

src/graphics/shader/recompiler/backend/spirv/spirvEmitterInternal.h

@@ -120,6 +120,8 @@ struct EmitterState {
 	uint32_t                                         indirect_format_geometry_function = 0;
 	std::map<std::array<uint32_t, 4>, uint32_t>        indirect_buffer_load_functions;
 	uint32_t                                         bvh_intersect_function  = 0;
+	uint32_t                                         bvh8_result_type       = 0;
+	uint32_t                                         bvh8_intersect_function = 0;
 	uint32_t                                         bvh_stack_function      = 0;
 	uint32_t                                         gds_variable            = 0;
 	uint32_t                                         gds_length              = 0;
@@ -170,6 +172,7 @@ uint32_t TypeBoolVector(EmitterState& state, uint32_t components);
 uint32_t TypeU32(EmitterState& state);
 uint32_t TypeU64(EmitterState& state);
 uint32_t TypeU32Pair(EmitterState& state);
+uint32_t TypeBvh8Result(EmitterState& state);
 uint32_t TypeI32(EmitterState& state);
 uint32_t TypeI32Pair(EmitterState& state);
 uint32_t TypeF32(EmitterState& state);
@@ -543,6 +546,10 @@ void EmitProgram(EmitterState& state);
 void DefineGetBdaPointer(EmitterState& state);
 void DefineIndirectBufferLoads(ValueEmitContext& ctx);
 void DefineBvhIntersect(EmitterState& state);
+void DefineTrinityBvhIntersect(EmitterState& state);
+uint32_t EmitBvhTriangle(EmitterState& state, const std::array<std::array<uint32_t, 3>, 3>& positions,
+                         std::array<uint32_t, 3> origin, std::array<uint32_t, 3> direction,
+                         uint32_t remap, uint32_t triangle_id, uint32_t bary);
 void DefineTrinityFunctions(EmitterState& state);
 uint32_t GetBdaPointer(EmitterState& state, uint32_t address);
 

src/graphics/shader/recompiler/backend/spirv/spirvEmitterProgram.cpp

@@ -414,6 +414,7 @@ uint32_t TypeId(EmitterState& state, IR::Type type) {
 		case IR::Type::U32:
 		case IR::Type::F16: return TypeU32(state);
 		case IR::Type::U64: return TypeU64(state);
+		case IR::Type::Bvh8Result: return TypeBvh8Result(state);
 		case IR::Type::U32x2: return TypeU32Pair(state);
 		case IR::Type::F32: return TypeF32(state);
 		case IR::Type::F64: return TypeF64(state);

src/graphics/shader/recompiler/backend/spirv/spirvEmitterTrinity.cpp

@@ -184,6 +184,7 @@ void DefineBvhStack(EmitterState& s) {
 
 void DefineTrinityFunctions(EmitterState& s) {
 	DefineBvhStack(s);
+	DefineTrinityBvhIntersect(s);
 }
 
 uint32_t EmitBvhStack(ValueEmitContext& ctx, const IR::Inst& inst) {

src/graphics/shader/recompiler/backend/spirv/spirvEmitterTrinityBvh.cpp

@@ -0,0 +1,511 @@
+#include "graphics/shader/recompiler/backend/spirv/spirvEmitterInstructions.h"
+
+namespace Libs::Graphics::ShaderRecompiler::Spirv::Emitter {
+
+uint32_t TypeBvh8Result(EmitterState& s) {
+	if (s.bvh8_result_type == 0) {
+		const auto v = TypeU32Vector(s, 4);
+		s.bvh8_result_type = s.builder.Type(spv::OpTypeStruct, v, v, v, v, TypeU32(s));
+	}
+	return s.bvh8_result_type;
+}
+
+namespace {
+using Vec3 = std::array<uint32_t, 3>;
+
+struct Bvh8Emitter {
+	EmitterState& s;
+	const uint32_t u = TypeU32(s), f = TypeF32(s), b = TypeBool(s), w = TypeU64(s);
+	const uint32_t v4 = TypeU32Vector(s, 4), v3 = TypeF32Vector(s, 3);
+	uint32_t triangle_function = 0, boxes_function = 0, decompress_function = 0;
+
+	uint32_t C(uint32_t x) { return ConstantU32(s, x); }
+	uint32_t And(uint32_t x, uint32_t y) { return Binary(s, spv::OpBitwiseAnd, u, x, y); }
+	uint32_t Or(uint32_t x, uint32_t y) { return Binary(s, spv::OpBitwiseOr, u, x, y); }
+	uint32_t Add(uint32_t x, uint32_t y) { return EmitAddU32(s, x, y); }
+	uint32_t Sub(uint32_t x, uint32_t y) { return Binary(s, spv::OpISub, u, x, y); }
+	uint32_t Shl(uint32_t x, uint32_t n) { return Binary(s, spv::OpShiftLeftLogical, u, x, C(n)); }
+	uint32_t Shr(uint32_t x, uint32_t n) { return Binary(s, spv::OpShiftRightLogical, u, x, C(n)); }
+	uint32_t Eq(uint32_t x, uint32_t y) { return Binary(s, spv::OpIEqual, b, x, y); }
+	uint32_t Both(uint32_t x, uint32_t y) { return Binary(s, spv::OpLogicalAnd, b, x, y); }
+	uint32_t Either(uint32_t x, uint32_t y) { return Binary(s, spv::OpLogicalOr, b, x, y); }
+	uint32_t Not(uint32_t x) { return Unary(s, spv::OpLogicalNot, b, x); }
+	uint32_t Bit(uint32_t x, uint32_t n) { return Not(Eq(And(x, C(1u << n)), C(0))); }
+	uint32_t Choose(uint32_t condition, uint32_t x, uint32_t y) { return Select(s, u, condition, x, y); }
+	uint32_t Float(uint32_t x) { return Unary(s, spv::OpBitcast, f, x); }
+	uint32_t Raw(uint32_t x) { return Unary(s, spv::OpBitcast, u, x); }
+	uint32_t Fop(spv::Op op, uint32_t x, uint32_t y) { return Binary(s, op, f, x, y); }
+	uint32_t Fcmp(spv::Op op, uint32_t x, uint32_t y) { return Binary(s, op, b, x, y); }
+	uint32_t Min(uint32_t x, uint32_t y) { return EmitGlsl<GLSLstd450FMin, IR::Type::F32>(s, x, y); }
+	uint32_t Max(uint32_t x, uint32_t y) { return EmitGlsl<GLSLstd450FMax, IR::Type::F32>(s, x, y); }
+	uint32_t Extract(uint32_t type, uint32_t value, uint32_t index) {
+		const auto id = s.builder.AllocateId();
+		s.builder.AddFunction(spv::OpCompositeExtract, type, id, value, index);
+		return id;
+	}
+	uint32_t Make(uint32_t type, std::span<const uint32_t> values) {
+		const auto id = s.builder.AllocateId();
+		s.builder.AddFunction(spv::OpCompositeConstruct, type, id, values);
+		return id;
+	}
+	uint32_t Make4(uint32_t x, uint32_t y, uint32_t z, uint32_t t) {
+		return Make(v4, std::array{x, y, z, t});
+	}
+	Vec3 Unpack3(uint32_t value) {
+		return {Extract(f, value, 0), Extract(f, value, 1), Extract(f, value, 2)};
+	}
+	uint32_t Word(uint32_t block, uint32_t index) {
+		const auto offset = Unary(s, spv::OpUConvert, w, Shl(index, 2));
+		const auto address = Binary(s, spv::OpIAdd, w, block, offset);
+		const auto pointer = Unary(s, spv::OpConvertUToPtr, TypePhysicalU32Pointer(s), address);
+		const auto id = s.builder.AllocateId();
+		s.builder.AddFunction(spv::OpLoad, u, id, pointer, spv::MemoryAccessAlignedMask, 4u);
+		return id;
+	}
+	uint32_t Call(uint32_t type, uint32_t function, std::span<const uint32_t> args) {
+		const auto id = s.builder.AllocateId();
+		std::vector<uint32_t> call {spv::OpFunctionCall, type, id, function};
+		call.insert(call.end(), args.begin(), args.end());
+		s.builder.AddFunction(call);
+		return id;
+	}
+	std::vector<uint32_t> Begin(uint32_t& function, const char* name, uint32_t result,
+	                            std::span<const uint32_t> types) {
+		function = s.builder.AllocateId();
+		s.builder.AddName(function, name);
+		s.builder.AddFunction(spv::OpFunction, result, function, spv::FunctionControlMaskNone,
+		    s.builder.Type(spv::OpTypeFunction, result, types));
+		std::vector<uint32_t> args;
+		for (auto type: types) {
+			args.push_back(s.builder.AllocateId());
+			s.builder.AddFunction(spv::OpFunctionParameter, type, args.back());
+		}
+		EmitLabel(s, s.builder.AllocateId());
+		return args;
+	}
+	void End(uint32_t result) {
+		s.builder.AddFunction(spv::OpReturnValue, result);
+		s.builder.AddFunction(spv::OpFunctionEnd);
+	}
+	uint32_t Result(const std::array<uint32_t, 10>& dst, const Vec3& org, const Vec3& dir,
+	                uint32_t instance) {
+		const std::array groups {Make4(dst[0], dst[1], dst[2], dst[3]),
+		    Make4(dst[4], dst[5], dst[6], dst[7]), Make4(dst[8], dst[9], Raw(org[0]), Raw(org[1])),
+		    Make4(Raw(org[2]), Raw(dir[0]), Raw(dir[1]), Raw(dir[2])), instance};
+		return Make(TypeBvh8Result(s), groups);
+	}
+	uint32_t NormalizeFlags(uint32_t flags) {
+		flags = Choose(Bit(flags, 2), And(flags, C(~0x30u)), flags);
+		return Or(flags, Choose(Eq(And(flags, C(0x30)), C(0x30)), C(0x100), C(0)));
+	}
+
+	void DefineTriangle() {
+		const auto a = Begin(triangle_function, "trinity_triangle", v4,
+		    std::array{w, u, u, u, v3, v3});
+		const auto block = a[0], slot = a[1], flags = a[2], descriptor = a[3];
+		const auto origin = Unpack3(a[4]), direction = Unpack3(a[5]);
+		const auto flag_word = Word(block, C(15));
+		const auto remap = Binary(s, spv::OpShiftRightLogical, u, flag_word, Shl(slot, 3));
+		const auto bary = Bit(descriptor, 24);
+		const auto pointer_flags = Both(bary, Bit(descriptor, 23));
+		const auto tri_flags = Choose(pointer_flags, remap, C(0));
+		const auto procedural = Bit(tri_flags, 6);
+		const auto raw = EmitValueOrDefaultIfCondition(s, Not(procedural), v4,
+		    Make4(C(0), C(0x3f800000), C(0x3f800000), C(0x3f800000)), [&] {
+			std::array<Vec3, 3> positions;
+			const std::array vertices {slot, Choose(Eq(slot, C(0)), C(1), C(3)), C(2)};
+			for (uint32_t vertex = 0; vertex < 3; ++vertex)
+				for (uint32_t axis = 0; axis < 3; ++axis)
+					positions[vertex][axis] = Float(Word(block,
+					    Add(Binary(s, spv::OpIMul, u, vertices[vertex], C(3)), C(axis))));
+			return EmitBvhTriangle(s, positions, origin, direction, remap, Add(flag_word, slot), bary);
+		});
+		std::array<uint32_t, 4> result;
+		for (uint32_t i = 0; i < 4; ++i) result[i] = Extract(u, raw, i);
+		auto missed = Eq(result[0], C(0x7f800000));
+		const auto flip = Both(Both(Not(procedural), Not(missed)),
+		    Binary(s, spv::OpLogicalNotEqual, b, Bit(flags, 3), Bit(tri_flags, 5)));
+		for (uint32_t i = 0; i < 4; ++i)
+			result[i] = Binary(s, spv::OpBitwiseXor, u, result[i],
+			    Choose(i < 2 ? flip : Both(flip, bary), C(0x80000000), C(0)));
+		const auto front = Fcmp(spv::OpFOrdGreaterThan, Float(result[1]), ConstantF32Value(s, 0.0f));
+		const auto opaque = Both(Either(Bit(tri_flags, 7), Bit(flags, 0)), Not(Bit(flags, 1)));
+		missed = Either(missed, Either(Both(opaque, Bit(flags, 6)), Both(Not(opaque), Bit(flags, 7))));
+		missed = Either(missed, Both(Both(Not(Bit(tri_flags, 4)), Not(procedural)),
+		    Select(s, b, front, Bit(flags, 5), Bit(flags, 4))));
+		missed = Either(missed, Select(s, b, procedural, Bit(flags, 9), Bit(flags, 8)));
+		const auto sign = [&](uint32_t value, uint32_t enabled) {
+			return Choose(pointer_flags, Or(And(value, C(0x7fffffff)),
+			    Choose(enabled, C(0x80000000), C(0))), value);
+		};
+		result[2] = sign(result[2], Either(Not(opaque), procedural));
+		result[3] = sign(result[3], Not(opaque));
+		const std::array miss {C(0x7f800000), C(0x3f800000), Choose(bary, C(0), C(~0u)), C(0)};
+		for (uint32_t i = 0; i < 4; ++i) result[i] = Choose(missed, miss[i], result[i]);
+		End(Make(v4, result));
+	}
+
+	uint32_t Instance(uint32_t block, uint32_t flags, uint32_t mask, Vec3 org, Vec3 dir) {
+		const auto lo = Word(block, C(12)), hi = Word(block, C(13));
+		const auto inst_flags = Shr(hi, 22);
+		auto merged = Or(And(Or(flags, inst_flags), C(0x3f0)), And(inst_flags, C(4)));
+		merged = Or(merged, And(Binary(s, spv::OpBitwiseXor, u, flags, inst_flags), C(8)));
+		merged = Or(merged, Choose(Either(Bit(flags, 0), Both(Bit(inst_flags, 0), Not(Bit(flags, 1)))), C(1), C(0)));
+		merged = Or(merged, Choose(Either(Bit(flags, 1), Both(Bit(inst_flags, 1), Not(Bit(flags, 0)))), C(2), C(0)));
+		const auto user = Word(block, C(15));
+		std::array<uint32_t, 10> dst;
+		dst.fill(C(~0u));
+		dst[2] = lo;
+		dst[3] = Or(And(hi, C(0x3fffff)), Shl(merged, 22));
+		dst[6] = And(user, C(0xffffff));
+		dst[7] = Choose(Eq(And(mask, Shr(user, 24)), C(0)), C(~0u), Word(block, C(14)));
+		Vec3 transformed_org, transformed_dir;
+		for (uint32_t axis = 0; axis < 3; ++axis) {
+			Vec3 row;
+			for (uint32_t i = 0; i < 3; ++i) row[i] = Float(Word(block, C(axis * 4 + i)));
+			const auto dot = [&](const Vec3& value) {
+				return Fop(spv::OpFAdd, Fop(spv::OpFAdd, Fop(spv::OpFMul, row[0], value[0]),
+				    Fop(spv::OpFMul, row[1], value[1])), Fop(spv::OpFMul, row[2], value[2]));
+			};
+			transformed_org[axis] = Fop(spv::OpFAdd, dot(org), Float(Word(block, C(axis * 4 + 3))));
+			transformed_dir[axis] = dot(dir);
+		}
+		return Result(dst, transformed_org, transformed_dir, C(1));
+	}
+
+	void DefineDecompress();
+	void DefineBoxes();
+	uint32_t Boxes(uint32_t first, uint32_t second, uint32_t full, uint32_t compressed,
+	               uint32_t flags, uint32_t descriptor, uint32_t extent, uint32_t org, uint32_t dir);
+	void DefineIntersect();
+};
+
+void Bvh8Emitter::DefineDecompress() {
+	const auto a = Begin(decompress_function, "trinity_shared_exponent", f, std::array{u, u, b});
+	const auto exponent = a[0], packed = a[1], minimum = a[2];
+	const auto negative = Bit(packed, 17), mantissa = And(packed, C(0x1ffff));
+	const auto outward = Eq(Choose(negative, C(1), C(0)), Choose(minimum, C(1), C(0)));
+	const auto nonzero = Not(Eq(mantissa, C(0)));
+	const auto highest = EmitGlsl<GLSLstd450FindUMsb, IR::Type::U32>(s, mantissa);
+	const auto shift = EmitGlsl<GLSLstd450UMin, IR::Type::U32>(s, Sub(C(16), highest), exponent);
+	const auto decoded_exp = Choose(nonzero, Sub(exponent, shift),
+	    Choose(Both(outward, Binary(s, spv::OpUGreaterThan, b, exponent, C(17))), Sub(exponent, C(17)), C(0)));
+	// Zero mantissas take a separate result, but their unused shifts must still be defined.
+	const auto amount = Add(C(7), Choose(nonzero, shift, C(0)));
+	const auto trail = Choose(Both(outward, Not(Eq(decoded_exp, C(255)))),
+	    Sub(Binary(s, spv::OpShiftLeftLogical, u, C(1), amount), C(1)), C(0));
+	const auto fraction = Choose(nonzero,
+	    And(Or(Binary(s, spv::OpShiftLeftLogical, u, mantissa, amount), trail), C(0x7fffff)),
+	    Choose(outward, C(0x7fffff), C(0)));
+	End(Float(Or(Or(fraction, Shl(decoded_exp, 23)), Choose(negative, C(0x80000000), C(0)))));
+}
+
+void Bvh8Emitter::DefineBoxes() {
+	const auto result_type = s.builder.Type(spv::OpTypeStruct, v4, v4);
+	const auto a = Begin(boxes_function, "trinity_box_slot", result_type,
+	    std::array{w, w, b, b, u, u, f, v3, v3});
+	const auto first = a[0], second = a[1], full = a[2], compressed = a[3];
+	const auto flags = a[4], descriptor = a[5], extent = a[6];
+	const auto org = Unpack3(a[7]), dir = Unpack3(a[8]);
+	const auto zero = ConstantF32Value(s, 0.0f), infinity = ConstantF32(s, 0x7f800000);
+	Vec3 inverse;
+	for (uint32_t i = 0; i < 3; ++i) inverse[i] = Fop(spv::OpFDiv, ConstantF32Value(s, 1.0f), dir[i]);
+	const auto packed = [&](uint32_t offset, uint32_t width) {
+		const auto word = Shr(offset, 5), shift = And(offset, C(31));
+		const auto lo = Binary(s, spv::OpShiftRightLogical, u, Word(first, word), shift);
+		const auto next = EmitGlsl<GLSLstd450UMin, IR::Type::U32>(s, Add(word, C(1)), C(15));
+		const auto hi = Binary(s, spv::OpShiftLeftLogical, u, Word(first, next), And(Sub(C(32), shift), C(31)));
+		return And(Or(lo, Choose(Eq(shift, C(0)), C(0), hi)), C((1u << width) - 1));
+	};
+	const auto format = Choose(compressed, C(0), Choose(full, C(1), C(2)));
+	const auto entry_label = s.current_label;
+	const auto header = s.builder.AllocateId(), body = s.builder.AllocateId();
+	const auto cont = s.builder.AllocateId(), merge = s.builder.AllocateId();
+	const auto initial_nodes = Make4(C(~0u), C(~0u), C(~0u), C(~0u));
+	const auto initial_keys = Make4(C(0x7f800000), C(0x7f800000), C(0x7f800000), C(0x7f800000));
+	const auto base = And(Word(first, C(0)), C(0x1fffffff));
+	s.builder.AddFunction(spv::OpBranch, header);
+	EmitLabel(s, header);
+	const auto child = s.builder.AllocateId(), prefix = s.builder.AllocateId();
+	const auto nodes = s.builder.AllocateId(), keys = s.builder.AllocateId();
+	const auto child_phi = s.builder.AddDeferredPhi(u, child, 2);
+	const auto prefix_phi = s.builder.AddDeferredPhi(u, prefix, 2);
+	const auto nodes_phi = s.builder.AddDeferredPhi(v4, nodes, 2);
+	const auto keys_phi = s.builder.AddDeferredPhi(v4, keys, 2);
+	const auto more = Binary(s, spv::OpULessThan, b, child, C(4));
+	s.builder.AddFunction(spv::OpLoopMerge, merge, cont, spv::LoopControlDontUnrollMask);
+	s.builder.AddFunction(spv::OpBranchConditional, more, body, merge);
+	EmitLabel(s, body);
+	const auto kind = packed(Add(C(29), Binary(s, spv::OpIMul, u, child, C(3))), 3);
+	const auto direct_node = Word(first, child);
+	const auto node = Choose(compressed, Or(Shl(prefix, 3), kind), direct_node);
+	const auto child_flags = Choose(compressed,
+	    And(packed(Add(C(41), Shl(child, 2)), 4), Choose(Eq(child, C(3)), C(7), C(15))), C(0));
+	const auto step = Choose(Eq(kind, C(2)), C(4),
+	    Choose(Either(Either(Eq(kind, C(3)), Eq(kind, C(5))), Eq(kind, C(6))), C(2), C(1)));
+	const auto next_prefix = Add(prefix, step);
+	const auto bounds_type = s.builder.Type(spv::OpTypeStruct, v3, v3);
+	const auto box = EmitIndexSwitch(s, format, 3, bounds_type, [&](uint32_t encoding) {
+		std::array<uint32_t, 6> bounds;
+		if (encoding == 0) {
+			const auto yz = Word(first, C(2));
+			const std::array exponents {Shr(Word(first, C(1)), 24), And(yz, C(255)), And(Shr(yz, 8), C(255))};
+			for (uint32_t component = 0; component < 6; ++component) {
+				const auto offset = Add(C(80 + component * 18), Binary(s, spv::OpIMul, u, child, C(108)));
+				bounds[component] = Call(f, decompress_function,
+				    std::array{exponents[component % 3], packed(offset, 18), ConstantBool(s, component < 3)});
+			}
+		} else if (encoding == 1) {
+			for (uint32_t component = 0; component < 6; ++component) {
+				const auto word = Add(C(4 + component), Binary(s, spv::OpIMul, u, child, C(6)));
+				const auto block = Select(s, w, Binary(s, spv::OpULessThan, b, word, C(16)), first, second);
+				bounds[component] = Float(Word(block, And(word, C(15))));
+			}
+		} else {
+			for (uint32_t pair = 0; pair < 3; ++pair) {
+				const auto word = Add(C(4 + pair), Binary(s, spv::OpIMul, u, child, C(3)));
+				const auto value = EmitGlsl<GLSLstd450UnpackHalf2x16, IR::Type::F32x2>(s, Word(first, word));
+				bounds[pair * 2] = Extract(f, value, 0);
+				bounds[pair * 2 + 1] = Extract(f, value, 1);
+			}
+		}
+		return Make(bounds_type, std::array{Make(v3, std::span(bounds).first(3)),
+		    Make(v3, std::span(bounds).last(3))});
+	});
+	const auto lo_bound = Unpack3(Extract(v3, box, 0)), hi_bound = Unpack3(Extract(v3, box, 1));
+	Vec3 near, far;
+	auto valid = Not(Eq(And(child_flags, C(3)), C(3)));
+	for (uint32_t axis = 0; axis < 3; ++axis) {
+		valid = Both(valid, Fcmp(spv::OpFOrdLessThanEqual, lo_bound[axis], hi_bound[axis]));
+		const auto lo = Fop(spv::OpFMul, Fop(spv::OpFSub, lo_bound[axis], org[axis]), inverse[axis]);
+		const auto hi = Fop(spv::OpFMul, Fop(spv::OpFSub, hi_bound[axis], org[axis]), inverse[axis]);
+		const auto forward = Fcmp(spv::OpFOrdGreaterThanEqual, inverse[axis], zero);
+		near[axis] = Select(s, f, forward, lo, hi);
+		far[axis] = Select(s, f, forward, hi, lo);
+	}
+	const auto entry = Max(Max(near[0], near[1]), near[2]);
+	const auto exit = Min(Min(far[0], far[1]), far[2]);
+	const auto exit_bits = Raw(Select(s, f, Fcmp(spv::OpFOrdEqual, exit, zero), zero, exit));
+	const auto grown = Float(EmitGlsl<GLSLstd450UMin, IR::Type::U32>(s,
+	    Add(exit_bits, And(Shr(descriptor, 23), C(255))), C(0x7f800000)));
+	valid = Both(valid, Both(Fcmp(spv::OpFOrdGreaterThanEqual, exit, zero),
+	    Both(Fcmp(spv::OpFOrdLessThan, entry, extent), Fcmp(spv::OpFOrdLessThanEqual, entry, grown))));
+	const auto opaque = Both(Either(Bit(child_flags, 0), Bit(flags, 0)), Not(Bit(flags, 1)));
+	const auto nonopaque = Both(Either(Bit(child_flags, 1), Bit(flags, 1)), Not(Bit(flags, 0)));
+	const auto culled = Either(Either(Both(opaque, Bit(flags, 6)), Both(nonopaque, Bit(flags, 7))),
+	    Either(Both(Bit(child_flags, 2), Bit(flags, 8)), Both(Bit(child_flags, 3), Bit(flags, 9))));
+	valid = Both(valid, Not(culled));
+	const auto heuristic = And(Shr(descriptor, 21), C(3));
+	auto key = Max(entry, zero);
+	key = Select(s, f, Eq(heuristic, C(1)), Unary(s, spv::OpFNegate, f, Fop(spv::OpFSub, exit, entry)), key);
+	key = Select(s, f, Eq(heuristic, C(2)), Fop(spv::OpFAdd, entry, exit), key);
+	const auto next_nodes = s.builder.AllocateId(), next_keys = s.builder.AllocateId();
+	s.builder.AddFunction(spv::OpVectorInsertDynamic, v4, next_nodes, nodes, Choose(valid, node, C(~0u)), child);
+	s.builder.AddFunction(spv::OpVectorInsertDynamic, v4, next_keys, keys, Raw(Select(s, f, valid, key, infinity)), child);
+	s.builder.AddFunction(spv::OpBranch, cont);
+	EmitLabel(s, cont);
+	const auto next_child = Add(child, C(1));
+	s.builder.AddFunction(spv::OpBranch, header);
+	s.builder.PatchDeferredPhi(child_phi, 0, C(0), entry_label);
+	s.builder.PatchDeferredPhi(child_phi, 1, next_child, cont);
+	s.builder.PatchDeferredPhi(prefix_phi, 0, base, entry_label);
+	s.builder.PatchDeferredPhi(prefix_phi, 1, next_prefix, cont);
+	s.builder.PatchDeferredPhi(nodes_phi, 0, initial_nodes, entry_label);
+	s.builder.PatchDeferredPhi(nodes_phi, 1, next_nodes, cont);
+	s.builder.PatchDeferredPhi(keys_phi, 0, initial_keys, entry_label);
+	s.builder.PatchDeferredPhi(keys_phi, 1, next_keys, cont);
+	EmitLabel(s, merge);
+	End(Make(result_type, std::array{nodes, keys}));
+}
+
+uint32_t Bvh8Emitter::Boxes(uint32_t first, uint32_t second, uint32_t full, uint32_t compressed,
+                            uint32_t flags, uint32_t descriptor, uint32_t extent,
+                            uint32_t org, uint32_t dir) {
+	return Call(s.builder.Type(spv::OpTypeStruct, v4, v4), boxes_function,
+	    std::array{first, second, full, compressed, flags, descriptor, extent, org, dir});
+}
+
+void Bvh8Emitter::DefineIntersect() {
+	const auto result_type = TypeBvh8Result(s);
+	const auto a = Begin(s.bvh8_intersect_function, "trinity_bvh_intersect", result_type,
+	    std::array{v4, w, u, f, u, v3, v3});
+	const auto descriptor = a[0], encoded_base = a[1], node = a[2], extent = a[3], mask = a[4];
+	const auto org = Unpack3(a[5]), dir = Unpack3(a[6]);
+	std::array<uint32_t, 4> d;
+	for (uint32_t i = 0; i < 4; ++i) d[i] = Extract(u, descriptor, i);
+	const auto kind = And(node, C(7));
+	const auto triangle = Binary(s, spv::OpULessThan, b, kind, C(2));
+	const auto instance = Eq(kind, C(6));
+	const auto full = Either(Eq(kind, C(2)), Eq(kind, C(5)));
+	const auto big = Either(Eq(kind, C(2)), Eq(kind, C(3)));
+	const auto compressed = Both(Not(full), Bit(d[3], 20));
+	const auto flags = NormalizeFlags(Choose(Bit(d[3], 23),
+	    Unary(s, spv::OpUConvert, u, Binary(s, spv::OpShiftRightLogical, w, encoded_base, C(54))), C(0)));
+	const auto base = Binary(s, spv::OpShiftLeftLogical, w, PackU64(s, d[0], And(d[1], C(255))), C(8));
+	const auto lane_base = Binary(s, spv::OpShiftLeftLogical, w,
+	    Binary(s, spv::OpBitwiseAnd, w, encoded_base, ConstantU64(s, (uint64_t{1} << 54) - 1)), C(3));
+	const auto offset = Binary(s, spv::OpShiftLeftLogical, w, Unary(s, spv::OpUConvert, w, Shr(node, 3)), C(6));
+	const auto address = Binary(s, spv::OpIAdd, w, base, Binary(s, spv::OpIAdd, w, lane_base, offset));
+	const auto capacity = Binary(s, spv::OpIAdd, w,
+	    PackU64(s, d[2], And(d[3], C(0x3ff))), ConstantU64(s, 1));
+	const auto limit = Binary(s, spv::OpIAdd, w, base,
+	    Binary(s, spv::OpShiftLeftLogical, w, capacity, C(6)));
+	auto valid = Both(Eq(Shr(d[3], 28), C(8)),
+	    Eq(And(Unary(s, spv::OpUConvert, u, encoded_base), C(7)), C(0)));
+	valid = Both(valid, Both(Not(Eq(kind, C(7))), Either(Not(instance), Bit(d[3], 22))));
+	valid = Both(valid, Not(Unary(s, spv::OpIsNan, b, extent)));
+	for (uint32_t i = 0; i < 3; ++i) {
+		valid = Both(valid, Not(Either(Unary(s, spv::OpIsNan, b, org[i]), Unary(s, spv::OpIsInf, b, org[i]))));
+		valid = Both(valid, Not(Unary(s, spv::OpIsNan, b, dir[i])));
+	}
+	const auto opaque = Bit(flags, 0), nonopaque = Bit(flags, 1);
+	valid = Both(valid, Not(Either(Either(Both(opaque, Bit(flags, 6)), Both(nonopaque, Bit(flags, 7))),
+	    Both(Bit(flags, 6), Bit(flags, 7)))));
+	const auto fetch_size = Select(s, w, full, ConstantU64(s, 128), ConstantU64(s, 64));
+	const auto fetch = [&](uint32_t at, uint32_t enabled) {
+		const auto end = Binary(s, spv::OpIAdd, w, at, fetch_size);
+		const auto in_bounds = Both(Binary(s, spv::OpULessThanEqual, b, end, limit),
+		    Binary(s, spv::OpULessThanEqual, b, end, ConstantU64(s, (uint64_t{1} << 48) - 1)));
+		return EmitValueOrDefaultIfCondition(s, Both(Both(valid, enabled), in_bounds), w,
+		    ConstantU64(s, 0), [&] { return GetBdaPointer(s, at); });
+	};
+	const auto first = fetch(address, ConstantBool(s, true));
+	const auto present = [&](uint32_t pointer) { return Not(Binary(s, spv::OpIEqual, b, pointer, ConstantU64(s, 0))); };
+	const auto tail = [&](uint32_t at, uint32_t pointer) {
+		return EmitValueOrDefaultIfCondition(s, Both(full, present(pointer)), w, ConstantU64(s, 0), [&] {
+			return GetBdaPointer(s, Binary(s, spv::OpIAdd, w, at, ConstantU64(s, 64)));
+		});
+	};
+	const auto second = tail(address, first);
+	const auto available = Both(present(first), Either(Not(full), present(second)));
+	std::array<uint32_t, 10> invalid;
+	invalid.fill(C(~0u));
+	for (uint32_t slot = 0; slot < 2; ++slot) {
+		invalid[slot * 4] = Choose(triangle, C(0x7f800000), C(~0u));
+		invalid[slot * 4 + 1] = Choose(triangle, C(0x3f800000), C(~0u));
+		invalid[slot * 4 + 3] = Choose(triangle, C(0), C(~0u));
+	}
+	const auto default_result = Result(invalid, org, dir, C(0));
+	const auto value = EmitValueOrDefaultIfCondition(s, Both(instance, available), result_type, default_result, [&] {
+		return Instance(first, flags, mask, org, dir);
+	});
+	const auto result = EmitValueOrDefaultIfCondition(s, Both(Not(instance), valid), result_type, value, [&] {
+		const auto second_address = Binary(s, spv::OpIAdd, w, address, fetch_size);
+		const auto next_first = fetch(second_address, big);
+		const auto next_second = tail(second_address, next_first);
+		const auto next_available = Both(present(next_first), Either(Not(full), present(next_second)));
+		const auto triangles = EmitValueOrDefaultIfCondition(s, triangle, result_type, default_result, [&] {
+			auto dst = invalid;
+			for (uint32_t slot = 0; slot < 2; ++slot) {
+				const auto active = Both(available, slot == 0 ? ConstantBool(s, true) : Eq(kind, C(1)));
+				const auto absent = Both(available, Both(Eq(kind, C(0)), Bit(d[3], 24)));
+				const auto miss = Make4(C(0x7f800000), C(0x3f800000),
+				    slot == 1 ? Choose(absent, C(0), C(~0u)) : C(~0u), C(0));
+				const auto hit = EmitValueOrDefaultIfCondition(s, active, v4, miss, [&] {
+					return Call(v4, triangle_function, std::array{first, C(slot), flags, d[3], a[5], a[6]});
+				});
+				for (uint32_t i = 0; i < 4; ++i) dst[slot * 4 + i] = Extract(u, hit, i);
+				dst[8 + slot] = EmitValueOrDefaultIfCondition(s, active, u, C(~0u), [&] { return Word(first, C(14)); });
+			}
+			return Result(dst, org, dir, C(0));
+		});
+		return EmitValueOrDefaultIfCondition(s, Not(triangle), result_type, triangles, [&] {
+			auto dst = invalid;
+			const auto slot_type = s.builder.Type(spv::OpTypeStruct, v4, v4);
+			const auto miss = Make(slot_type, std::array{Make4(C(~0u), C(~0u), C(~0u), C(~0u)),
+			    Make4(C(0x7f800000), C(0x7f800000), C(0x7f800000), C(0x7f800000))});
+			std::array<uint32_t, 8> keys;
+			for (uint32_t slot = 0; slot < 2; ++slot) {
+				const auto hit = EmitValueOrDefaultIfCondition(s, slot == 0 ? available : next_available,
+				    slot_type, miss, [&] {
+					return Boxes(slot == 0 ? first : next_first, slot == 0 ? second : next_second,
+					    full, compressed, flags, d[1], extent, a[5], a[6]);
+				});
+				const auto indices = Extract(v4, hit, 0), depths = Extract(v4, hit, 1);
+				for (uint32_t i = 0; i < 4; ++i) {
+					dst[slot * 4 + i] = Extract(u, indices, i);
+					keys[slot * 4 + i] = Extract(u, depths, i);
+				}
+			}
+			const auto sort = Bit(d[1], 31), wide = Both(sort, Bit(d[3], 21)), tri_first = Bit(d[1], 20);
+			constexpr std::array<std::pair<uint32_t, uint32_t>, 19> network {{{0,1}, {2,3}, {0,2}, {1,3}, {1,2},
+			    {4,5}, {6,7}, {4,6}, {5,7}, {5,6}, {0,4}, {1,5}, {1,4}, {2,6}, {3,7}, {3,6}, {2,4}, {3,5}, {3,4}}};
+			const auto unsorted = Make(slot_type, std::array{Make(v4, std::span(dst).first(4)),
+			    Make(v4, std::span(dst).subspan(4, 4))});
+			const auto sorted_result = EmitValueOrDefaultIfCondition(s, sort, slot_type, unsorted, [&] {
+				auto sorted_dst = dst;
+				for (size_t step = 0; step < network.size(); ++step) {
+					const auto [x, y] = network[step];
+					const auto is_tri = [&](uint32_t index) { return Binary(s, spv::OpULessThan, b, And(index, C(7)), C(2)); };
+					const auto xt = is_tri(sorted_dst[x]), yt = is_tri(sorted_dst[y]);
+					auto swap = Fcmp(spv::OpFOrdGreaterThan, Float(keys[x]), Float(keys[y]));
+					swap = Select(s, b, Both(tri_first, Binary(s, spv::OpLogicalNotEqual, b, xt, yt)), yt, swap);
+					if (step >= 10) swap = Both(wide, swap);
+					const auto key = keys[x], index = sorted_dst[x];
+					const auto later = [&](uint32_t position) {
+						return std::ranges::any_of(std::span(network).subspan(step + 1), [=](auto pair) {
+							return pair.first == position || pair.second == position;
+						});
+					};
+					if (later(x)) keys[x] = Choose(swap, keys[y], key);
+					if (later(y)) keys[y] = Choose(swap, key, keys[y]);
+					sorted_dst[x] = Choose(swap, sorted_dst[y], index);
+					sorted_dst[y] = Choose(swap, index, sorted_dst[y]);
+				}
+				return Make(slot_type, std::array{Make(v4, std::span(sorted_dst).first(4)),
+				    Make(v4, std::span(sorted_dst).subspan(4, 4))});
+			});
+			for (uint32_t slot = 0; slot < 2; ++slot) {
+				const auto indices = Extract(v4, sorted_result, slot);
+				for (uint32_t i = 0; i < 4; ++i) dst[slot * 4 + i] = Extract(u, indices, i);
+			}
+			return Result(dst, org, dir, C(0));
+		});
+	});
+	End(result);
+}
+
+} // namespace
+
+void DefineTrinityBvhIntersect(EmitterState& s) {
+	if (!s.program.info.bvh8) return;
+	Bvh8Emitter e {s};
+	e.DefineDecompress();
+	e.DefineTriangle();
+	e.DefineBoxes();
+	e.DefineIntersect();
+}
+
+uint32_t EmitBvh8Extract(ValueEmitContext& ctx, const IR::Inst& inst) {
+	const auto index = inst.Arg(1).U32();
+	if (index > 16) ctx.Fail(inst, "invalid BVH8 result component");
+	const auto result = ctx.state.builder.AllocateId();
+	if (index == 16)
+		ctx.state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(ctx.state), result,
+		    ctx.Arg(inst, 0), 4u);
+	else
+		ctx.state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(ctx.state), result,
+		    ctx.Arg(inst, 0), index / 4, index % 4);
+	return result;
+}
+
+uint32_t EmitBvh8Intersect(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.bvh8_intersect_function == 0)
+		ctx.Fail(inst, "BVH8 intersection requires the complete ray and shared function");
+	Bvh8Emitter e {s};
+	const auto vector = [&](uint32_t first) {
+		return e.Make(e.v3, Vec3{e.Float(ctx.Arg(*ray, first)), e.Float(ctx.Arg(*ray, first + 1)),
+		    e.Float(ctx.Arg(*ray, first + 2))});
+	};
+	const std::array args {ctx.Arg(inst, 0), PackU64(s, ctx.Arg(*ray, 0), ctx.Arg(*ray, 1)),
+	    ctx.Arg(*ray, 10), e.Float(ctx.Arg(*ray, 2)), ctx.Arg(*ray, 3), vector(4), vector(7)};
+	const auto type = TypeBvh8Result(s);
+	return EmitValueOrDefaultIfCondition(s, ctx.Arg(inst, 2), type,
+	    s.builder.Constant(spv::OpConstantNull, type), [&] {
+		return e.Call(type, s.bvh8_intersect_function, args);
+	});
+}
+
+} // namespace Libs::Graphics::ShaderRecompiler::Spirv::Emitter

src/graphics/shader/recompiler/frontend/decode/ImageOps.cpp

@@ -281,6 +281,7 @@ 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 0xe5u: return Opcode::IMAGE_BVH8_INTERSECT_RAY;
 		case 0xe6u:
 		case 0xe7u: return Opcode::IMAGE_BVH_INTERSECT_RAY;
 		default: return Opcode::UNSUPPORTED;
@@ -297,6 +298,7 @@ uint32_t DecodeMimgAddressComponents(uint32_t opcode, ImageDimension dimension,
 	}
 
 	switch (opcode) {
+		case 0xe5u:
 		case 0xe6u: return 11u;
 		case 0xe7u: return 12u;
 		case 0x0eu: return 1u;
@@ -427,6 +429,36 @@ void DecodeMimg(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 		               "BVH intersection requires full-float ray DWORDs and R128/dmask:0xf");
 	}
 
+	if (opcode == 0xe5u) {
+		inst.data_components = inst.data_dwords = 10;
+		if (a16 || d16 || !r128 || inst.dmask != 0xfu ||
+		    (nsa_dwords != 0u && nsa_dwords != 2u) || vdata > 246u ||
+		    (nsa_dwords == 0u && vaddr > 245u)) {
+			SetUnsupported(inst, Family::MIMG, opcode, "invalid BVH8 ray or destination encoding");
+		}
+		if (nsa_dwords == 2u) {
+			const auto origin = inst.image_nsa_addr[3], direction = inst.image_nsa_addr[4];
+			const auto node = inst.image_nsa_addr[5];
+			if (origin > 253u || direction > 253u)
+				SetUnsupported(inst, Family::MIMG, opcode, "BVH8 ray registers exceed VGPR range");
+			for (uint32_t i = 0; i < 3; ++i) {
+				inst.image_nsa_addr[3 + i] = origin + i;
+				inst.image_nsa_addr[6 + i] = direction + i;
+			}
+			inst.image_nsa_addr[9] = node;
+		}
+		const auto origin = nsa_dwords != 0 ? inst.image_nsa_addr[3] : vaddr + 4;
+		const auto direction = nsa_dwords != 0 ? inst.image_nsa_addr[6] : vaddr + 7;
+		bool overlapping = origin < direction + 3 && direction < origin + 3;
+		for (const auto output : {2u, 3u, 6u, 7u}) {
+			const auto reg = vdata + output;
+			overlapping |= (reg >= origin && reg < origin + 3) ||
+			               (reg >= direction && reg < direction + 3);
+		}
+		if (overlapping)
+			SetUnsupported(inst, Family::MIMG, opcode, "overlapping BVH8 instance outputs are unsupported");
+	}
+
 	DecodeVectorGpr(vdata, inst.dst);
 	DecodeVectorGpr(vaddr, inst.src0);
 	DecodeScalarSource(srsrc * 4u, pc, inst.src1);

src/graphics/shader/recompiler/frontend/decode/ShaderDecoder.h

@@ -701,6 +701,7 @@ enum class Opcode {
 	DS_WRITE_ADDTID_B32,
 	DS_READ_ADDTID_B32,
 	IMAGE_BVH_INTERSECT_RAY,
+	IMAGE_BVH8_INTERSECT_RAY,
 	IMAGE_GET_RESINFO,
 	IMAGE_GET_LOD,
 	IMAGE_LOAD,

src/graphics/shader/recompiler/frontend/translate/Memory.cpp

@@ -326,7 +326,8 @@ IR::Value Translator::MakeImageAddress(const Decoder::Instruction& inst,
 	    inst.image_sample_flags, inst.image_dimension, inst.image_address_components);
 	EXIT_IF(count > components.size());
 	const auto nsa_components =
-	    std::min(inst.image_nsa_dwords * 4u, Decoder::MaxImageNsaAddressComponents);
+	    inst.opcode == Decoder::Opcode::IMAGE_BVH8_INTERSECT_RAY && inst.image_nsa_dwords != 0u
+	        ? 10u : std::min(inst.image_nsa_dwords * 4u, Decoder::MaxImageNsaAddressComponents);
 	for (uint32_t index = 0; index < count; index++) {
 		if (index != 0u && index - 1u < nsa_components) {
 			components[index] =
@@ -948,6 +949,31 @@ void Translator::DS_PERMUTE(const Decoder::Instruction& inst, bool backward) {
 	                               {ReadU32(inst.src1), address, ir.GetExec()}));
 }
 
+void Translator::IMAGE_BVH8(const Decoder::Instruction& inst) {
+	std::array<IR::U32, 10> previous;
+	for (uint32_t i = 0; i < previous.size(); ++i)
+		previous[i] = ReadU32(OffsetOperand(inst.dst, i));
+	const auto result = ir.Emit(IR::ValueOpcode::Bvh8Intersect,
+	    {ConstructU32x4(inst.src1, 4), MakeImageAddress(inst, inst.src0), ir.GetExec()});
+	const auto get = [&](uint32_t i) {
+		return IR::U32(ir.Emit(IR::ValueOpcode::Bvh8Extract, {result, IR::Value(i)}));
+	};
+	const auto instance = ir.INotEqual(get(16), IR::U32(IR::Value(0u)));
+	for (uint32_t i = 0; i < 10; ++i) {
+		auto value = get(i);
+		if (i == 0 || i == 1 || i == 4 || i == 5 || i == 8 || i == 9)
+			value = ir.Select(instance, previous[i], value);
+		WriteOperand(OffsetOperand(inst.dst, i), value);
+	}
+	for (uint32_t i = 0; i < 6; ++i) {
+		auto reg = inst.image_nsa_dwords != 0 ? inst.image_nsa_addr[3 + i] : inst.src0.reg + 4 + i;
+		auto target = inst.src0;
+		target.reg = reg;
+		// Non-instance results may alias the ray inputs.
+		WriteOperand(target, ir.Select(instance, get(10 + i), ReadU32(target)));
+	}
+}
+
 void Translator::EmitMemory(const Decoder::Instruction& inst) {
 	switch (inst.opcode) {
 		// Translated scalar loads have no cache of their own to invalidate.
@@ -956,6 +982,7 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 		case Decoder::Opcode::BUFFER_GL1_INV:
 			ir.Emit(IR::ValueOpcode::DataCacheInvalidate);
 			return;
+		case Decoder::Opcode::IMAGE_BVH8_INTERSECT_RAY: return IMAGE_BVH8(inst);
 		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()},

src/graphics/shader/recompiler/frontend/translate/Translator.h

@@ -105,6 +105,7 @@ private:
 	void WriteSharedU32(uint32_t width, IR::U32 address, const std::array<IR::Value, 4>& values,
 	                    const IR::MemoryInfo& memory, uint32_t pc);
 	void DS_READ(const Decoder::Instruction& inst);
+	void IMAGE_BVH8(const Decoder::Instruction& inst);
 	void DS_BVH_STACK(const Decoder::Instruction& inst);
 	void DS_READ2(const Decoder::Instruction& inst);
 	void DS_WRITE(const Decoder::Instruction& inst);

src/graphics/shader/recompiler/ir/Serialize.h

@@ -155,7 +155,7 @@ KYTY_IO_FIELDS(SampledResourcePair, a, b, c)
 KYTY_IO_FIELDS(StageInput, a, b, c, d, e)
 KYTY_IO_FIELDS(StageOutput, a, b, c, d)
 KYTY_IO_FIELDS(ShaderInfo, a, b, c, d, e, f, g, h, i, j, k, l, m, n, o, p, q, r, s, t, u, v, w, x, y, z,
-               A, B, C, D, E, F, G, H, I, J, K, L)
+               A, B, C, D, E, F, G, H, I, J, K, L, M)
 KYTY_IO_FIELDS(DescriptorSource::IndirectDescriptor::SelectorRead::CountedRows, a, b)
 KYTY_IO_FIELDS(DescriptorSource::IndirectDescriptor::SelectorRead::MaskRanks, a, b, c)
 KYTY_IO_FIELDS(DescriptorSource::IndirectDescriptor::SelectorRead, a, b, c, d, e, f, g, h, i, j, k,

src/graphics/shader/recompiler/ir/ShaderIR.h

@@ -543,6 +543,7 @@ struct ShaderInfo {
 	bool                            uses_image_heap               = false;
 	bool                            uses_dispatch_threads         = false;
 	bool                            bvh                           = false;
+	bool                            bvh8                          = false;
 	bool                            bvh_stack                     = false;
 	bool                            subgroup_ballot               = false;
 	bool                            subgroup_vote                 = false;

src/graphics/shader/recompiler/ir/Type.cpp

@@ -31,6 +31,7 @@ std::string TypeName(Type type) {
 	    "SamplerResource",
 	    "ImageAddress",
 	    "F64",
+	    "Bvh8Result",
 	};
 	const auto bits = static_cast<uint32_t>(type);
 	if (bits == 0) {

src/graphics/shader/recompiler/ir/Type.h

@@ -28,6 +28,7 @@ enum class Type : uint32_t {
 	SamplerResource = 1u << 18u,
 	ImageAddress    = 1u << 19u,
 	F64             = 1u << 20u,
+	Bvh8Result      = 1u << 21u,
 };
 
 constexpr Type operator|(Type lhs, Type rhs) {

src/graphics/shader/recompiler/ir/opcodes/ValueOpcodes.cpp

@@ -44,6 +44,7 @@ constexpr Type AddressResource = Type::AddressResource;
 constexpr Type ImageResource   = Type::ImageResource;
 constexpr Type SamplerResource = Type::SamplerResource;
 constexpr Type ImageAddress    = Type::ImageAddress;
+constexpr Type Bvh8Result      = Type::Bvh8Result;
 
 constexpr std::array<OpcodeMeta, static_cast<size_t>(ValueOpcode::Count)> MetaTable = {{
 #define VALUE_OPCODE(name, ...) MakeMeta(#name __VA_OPT__(, ) __VA_ARGS__),

src/graphics/shader/recompiler/ir/opcodes/ValueOpcodes.inc

@@ -318,6 +318,8 @@ VALUE_OPCODE(SwizzleU32, U32, U32, U32, U1)
 
 VALUE_OPCODE(ImageQueryDimensions, U32x4, ImageResource, ImageAddress)
 VALUE_OPCODE(BvhIntersect, U32x4, U32x4, ImageAddress, U1)
+VALUE_OPCODE(Bvh8Intersect, Bvh8Result, U32x4, ImageAddress, U1)
+VALUE_OPCODE(Bvh8Extract, U32, Bvh8Result, U32)
 VALUE_OPCODE(BvhStack, U32x2, U32, U32, U32x4, U32x4, U1)
 VALUE_OPCODE(ImageQueryLod, U32x4, ImageResource, SamplerResource, ImageAddress)
 VALUE_OPCODE(ImageRead, U32x4, ImageResource, ImageAddress, U1)

src/graphics/shader/recompiler/ir/passes/ResourceTracking.cpp

@@ -4182,7 +4182,7 @@ private:
 
 	void Collect(Inst& inst) {
 		const auto op           = inst.GetOpcode();
-		if (op == ValueOpcode::BvhIntersect) {
+		if (op == ValueOpcode::BvhIntersect || op == ValueOpcode::Bvh8Intersect) {
 			m_info.uses_dma = true;
 			return;
 		}

src/graphics/shader/recompiler/ir/passes/ShaderInfoCollection.cpp

@@ -515,6 +515,7 @@ void Visit(Program& program, ShaderStageInputInfo input_info, InputUsage& inputs
 		case ValueOpcode::StoreCompletion: info.subgroup_barrier = true; break;
 		case ValueOpcode::DataCacheInvalidate: info.coherent_buffers = true; break;
 		case ValueOpcode::BvhIntersect: info.bvh = true; break;
+		case ValueOpcode::Bvh8Intersect: info.bvh8 = true; break;
 		case ValueOpcode::BvhStack: info.bvh_stack = true; break;
 		case ValueOpcode::ConditionRef:
 			info.subgroup_vote |=

tests/ShaderRayTracingGpuTests.inc

@@ -1,3 +1,306 @@
+TestCase TrinityBvhIntersections(bool compressed, bool bary, bool sorted,
+                                 bool wide_sort, u32 heuristic = 0, bool triangles_first = false,
+                                 u32 wave_size = 32, bool edges = false) {
+  TestCase test;
+  test.name = compressed ? (sorted ? "TrinityCompressedSorted" : "TrinityCompressedUnsorted")
+                         : (sorted ? "TrinityFp32Sorted" : "TrinityFp32Unsorted");
+  if (!bary) test.name = "TrinityTriangleIds";
+  if (heuristic == 1) test.name = "TrinityLargestInterval";
+  if (heuristic == 2) test.name = "TrinityClosestMidpoint";
+  if (triangles_first) test.name = "TrinityTrianglesFirst";
+  if (wave_size == 64) test.name = "TrinityBvhWave64";
+  if (edges) test.name = compressed ? "TrinityCompressedEdges" : "TrinityFp32Edges";
+  constexpr u32 NodeStart = 4096, OutputStart = 2048, GuestBase = 0x10000;
+  constexpr u32 Invalid = ~0u, Infinity = 0x7f800000, Flag = 0x8484, Shape = 0x12345678;
+  const auto bits = [](float x) { return std::bit_cast<u32>(x); };
+  test.initial.resize(NodeStart + 256 * 16);
+  auto& data = test.initial;
+  const std::array<float, 12> vertices {-1,-1,2, 1,-1,2, 0,1,2, 2,1,2};
+  for (u32 i = 0; i < vertices.size(); ++i) data[NodeStart + i] = bits(vertices[i]);
+  data[NodeStart + 14] = Shape;
+  data[NodeStart + 15] = Flag;
+  const std::array<float, 12> matrix {2,0,0,1, 0,3,0,-2, 0,0,4,3};
+  for (u32 i = 0; i < matrix.size(); ++i) data[NodeStart + 16 + i] = bits(matrix[i]);
+  data[NodeStart + 28] = 0x1000;
+  data[NodeStart + 29] = 9u << 22; // Force opaque and flip facing.
+  data[NodeStart + 30] = 0x98;
+  data[NodeStart + 31] = 0x04123456;
+  const std::array<float, 8> near {5,1,3,7,4,2,6,8};
+  const std::array<float, 8> length {1,8,3,5,7,2,4,6};
+  const std::array<u32, 8> children {128,138,171,190,256,266,299,318};
+  const auto set_bits = [&](u32 base, u32 offset, u32 count, u32 value) {
+    for (u32 bit = 0; bit < count; ++bit)
+      data[base + (offset + bit) / 32] |= ((value >> bit) & 1) << ((offset + bit) % 32);
+  };
+  for (u32 slot = 0; slot < 2; ++slot) {
+    const u32 base = NodeStart + (3 + slot * (compressed ? 1 : 2)) * 16;
+    if (compressed) {
+      set_bits(base, 0, 29, slot ? 32 : 16);
+      for (u32 i = 0; i < 4; ++i) set_bits(base, 29 + i * 3, 3, children[slot * 4 + i] & 7);
+      set_bits(base, 56, 8, 127); set_bits(base, 64, 8, 127); set_bits(base, 72, 8, 130);
+    }
+    for (u32 i = 0; i < 4; ++i) {
+      const u32 index = slot * 4 + i;
+      if (compressed) {
+        // Exact integer minima; positive maxima round outward within one quantization step.
+        const std::array<u32, 6> planes {0x30000, 0x30000, u32(near[index]) * 8192,
+            0x10000, 0x10000, u32(near[index] + length[index]) * 8192};
+        for (u32 axis = 0; axis < 6; ++axis) set_bits(base, 80 + i * 108 + axis * 18, 18, planes[axis]);
+      } else {
+        data[base + i] = children[index];
+        const std::array<float, 6> bounds {-1,-1,near[index], 1,1,near[index] + length[index]};
+        for (u32 axis = 0; axis < 6; ++axis) data[base + 4 + i * 6 + axis] = bits(bounds[axis]);
+      }
+    }
+  }
+  if (compressed) {
+    std::copy_n(data.begin() + NodeStart + 3 * 16, 16, data.begin() + NodeStart + 10 * 16);
+  } else {
+    constexpr std::array<u32, 4> lo {0x4500, 0x3c00, 0x4200, 0x4700};
+    constexpr std::array<u32, 4> hi {0x4600, 0x4880, 0x4600, 0x4a00};
+    for (u32 i = 0; i < 4; ++i) {
+      data[NodeStart + 160 + i] = children[i];
+      data[NodeStart + 164 + i * 3] = 0xbc00bc00;
+      data[NodeStart + 165 + i * 3] = 0x3c000000 | lo[i];
+      data[NodeStart + 166 + i * 3] = (hi[i] << 16) | 0x3c00;
+    }
+  }
+  if (edges) {
+    // Slot zero ends exactly at the descriptor limit; slot one lies beyond it.
+    std::copy_n(data.begin() + NodeStart + 3 * 16, compressed ? 16 : 32,
+                data.begin() + NodeStart + (compressed ? 253 : 252) * 16);
+    for (u32 index : {12u, 13u})
+      std::copy_n(data.begin() + NodeStart, 16, data.begin() + NodeStart + index * 16);
+    data[NodeStart + 12 * 16 + 15] = 0xc4c4; // Opaque procedural pair.
+    data[NodeStart + 13 * 16 + 15] = 0x9494; // Opaque, double-sided pair.
+    if (compressed) {
+      std::copy_n(data.begin() + NodeStart + 3 * 16, 16, data.begin() + NodeStart + 11 * 16);
+      set_bits(NodeStart + 11 * 16, 41, 4, 1);
+      set_bits(NodeStart + 11 * 16, 45, 4, 2);
+      set_bits(NodeStart + 11 * 16, 49, 4, 3);
+    }
+  }
+  std::vector<std::array<u32, 16>> expected(wave_size);
+  for (u32 lane = 0; lane < wave_size; ++lane) {
+    const u32 which = lane % 16, input = lane * 16;
+    auto& out = expected[lane];
+    out.fill(Invalid);
+    data[input + 2] = bits(20); data[input + 3] = 4; data[input + 9] = bits(1);
+    const auto miss = [&](u32 slot, bool invalid = false) {
+      out[slot * 4] = Infinity; out[slot * 4 + 1] = bits(1);
+      out[slot * 4 + 2] = !invalid && bary ? 0 : Invalid; out[slot * 4 + 3] = 0;
+    };
+    const auto hit = [&](u32 slot, float u, float v) {
+      out[slot * 4] = bits(-8); out[slot * 4 + 1] = bits(-4);
+      out[slot * 4 + 2] = bary ? bits(u) : Flag + slot;
+      out[slot * 4 + 3] = bary ? bits(v) : 1;
+    };
+    if (which <= 2 || which >= 9 && which <= 14) {
+      data[input + 10] = which == 1 || which == 2 ? 1 : 0;
+      miss(0); miss(1);
+      out[8] = Shape;
+      if (which == 1) {
+        data[input + 4] = bits(1); hit(1, 2, 1); out[9] = Shape;
+      } else {
+        hit(0, 1, 1);
+        if (which == 2) { miss(1); out[9] = Shape; }
+      }
+      if (which == 9 || which == 10 || which == 11) {
+        if (which == 9) data[input + 4] = 0x7fc00000;
+        if (which == 10) data[input + 10] = 0xfff8;
+        if (which == 11) data[input + 1] = 0x41u << 22;
+        miss(0, true); miss(1, true); out[8] = out[9] = Invalid;
+      } else if (which == 12) {
+        data[input + 1] = 0x100u << 22; miss(0);
+      } else if (which == 13) {
+        data[input + 1] = 8u << 22; out[0] = bits(8); out[1] = bits(4);
+      } else if (which == 14) {
+        data[input + 1] = 2u << 22;
+        if (bary) { out[2] = bits(-1); out[3] = bits(-1); }
+      }
+    } else if (which >= 3 && which <= 5) {
+      data[input + 10] = which == 3 ? (3 * 8 + (compressed ? 3 : 2))
+                           : which == 4 ? (3 * 8 + (compressed ? 4 : 5)) : (10 * 8 + 4);
+      std::array<u32, 8> order {0,1,2,3,4,5,6,7};
+      const auto compare = [&](u32 x, u32 y) {
+        const bool xt = (children[x] & 7) < 2, yt = (children[y] & 7) < 2;
+        if (triangles_first && xt != yt) return xt;
+        if (heuristic == 1) return length[x] > length[y];
+        if (heuristic == 2) return 2 * near[x] + length[x] < 2 * near[y] + length[y];
+        return near[x] < near[y];
+      };
+      if (sorted) {
+        if (wide_sort && which == 3) std::stable_sort(order.begin(), order.end(), compare);
+        else {
+          std::stable_sort(order.begin(), order.begin() + 4, compare);
+          std::stable_sort(order.begin() + 4, order.end(), compare);
+        }
+      }
+      for (u32 i = 0; i < (which == 3 ? 8u : 4u); ++i) out[i] = children[order[i]];
+    } else if (which == 6 || which == 7 || which == 15) {
+      data[input + 10] = 14;
+      if (which == 15) { data[input] = 8; data[input + 10] = 6; }
+      data[input + 4] = bits(1); data[input + 5] = bits(2); data[input + 6] = bits(3);
+      if (which == 7) data[input + 3] = 2;
+      for (u32 i = 0; i < 10; ++i) out[i] = 0xabc00000 + i;
+      out[2] = 0x1000; out[3] = 9u << 22; out[6] = 0x123456; out[7] = which == 7 ? Invalid : 0x98;
+    } else data[input + 10] = 7;
+    if (edges) {
+      switch (which) {
+        case 0:
+        case 1:
+        case 2:
+        case 5:
+          if (which == 0) data[input] = 1; // Misaligned encoded base must not read across a block.
+          if (which == 1) data[input + 4] = Infinity;
+          if (which == 2) data[input + 2] = 0x7fc00000;
+          if (which == 5) { data[input + 1] = 0x2000; data[input + 10] = 0; }
+          miss(0, true); miss(1, true); out[8] = out[9] = Invalid;
+          break;
+        case 3:
+          data[input + 10] = compressed ? 253 * 8 + 3 : 252 * 8 + 2;
+          for (u32 i = 4; i < 8; ++i) out[i] = Invalid;
+          break;
+        case 4:
+          if (compressed) {
+            data[input + 10] = 11 * 8 + 4; data[input + 1] = 0x40u << 22;
+            out[0] = out[2] = Invalid;
+          } else {
+            data[input + 2] = bits(1);
+            for (u32 i = 0; i < 4; ++i) out[i] = Invalid;
+          }
+          break;
+        case 6:
+          data[input + 1] = 8u << 22; out[3] = 1u << 22;
+          break;
+        case 7:
+          data[input + 1] = 2u << 22; out[3] = 10u << 22;
+          break;
+        case 9:
+        case 10:
+          data[input + 10] = 12 * 8;
+          data[input + 4] = bits(100);
+          out[8] = Shape; miss(0); miss(1);
+          if (which == 9) { out[0] = bits(0); out[1] = bits(1); out[2] = bits(-1); out[3] = bits(1); }
+          else data[input + 1] = 0x200u << 22;
+          break;
+        case 11:
+          data[input + 1] = 0x10u << 22; miss(0); miss(1); out[8] = Shape;
+          break;
+        case 12:
+        case 13:
+          data[input + 1] = (which == 12 ? 0x14u : 0x10u) << 22;
+          if (which == 13) data[input + 10] = 13 * 8;
+          hit(0, 1, 1);
+          break;
+        case 14:
+          // The compressed +X bound rounds outward beyond exactly one.
+          if (compressed) {
+            data[input + 1] = 0; data[input + 4] = bits(1.0f + 0x1p-17f);
+            data[input + 10] = 3 * 8 + 4;
+            out.fill(Invalid);
+            for (u32 i = 0; i < 4; ++i) out[i] = children[i];
+          }
+          break;
+      }
+    }
+    for (u32 i = 0; i < 6; ++i) out[10 + i] = data[input + 4 + i];
+    if (which == 6 || which == 7 || which == 15) {
+      out[10] = bits(3); out[11] = bits(4); out[12] = bits(15); out[15] = bits(4);
+    }
+    if (lane % 32 >= 16) {
+      for (u32 i = 0; i < 10; ++i) out[i] = 0xabc00000 + i;
+      for (u32 i = 0; i < 6; ++i) out[10 + i] = data[input + 4 + i];
+    }
+  }
+  test.expected = data;
+  for (u32 lane = 0; lane < wave_size; ++lane)
+    std::copy(expected[lane].begin(), expected[lane].end(), test.expected.begin() + OutputStart + lane * 16);
+  auto& code = test.code;
+  AppendSMovLiteral(&code, 8, GuestBase >> 8);
+  AppendSMovLiteral(&code, 9, (sorted ? 1u << 31 : 0) | (triangles_first ? 1u << 20 : 0) | (heuristic << 21));
+  AppendSMovLiteral(&code, 10, edges ? 253 : 255);
+  AppendSMovLiteral(&code, 11, 0x80c00000 | (bary ? 1u << 24 : 0) |
+      (compressed ? 1u << 20 : 0) | (wide_sort ? 1u << 21 : 0));
+  code.push_back(EncodeVop2(0x1a, 1, InlineU32(6), 0));
+  for (u32 offset : {0u,4u,8u})
+    code.insert(code.end(), {EncodeMubuf0(offset == 8 ? 0x0f : 0x0e, offset * 4), EncodeMubuf1(4 + offset, 0, 1)});
+  for (u32 i = 0; i < 10; ++i) AppendVMovLiteral(&code, 32 + i, 0xabc00000 + i);
+  AppendSMovLiteral(&code, 126, 0x0000ffff);
+  if (wave_size == 64) AppendSMovLiteral(&code, 127, 0x0000ffff);
+  code.insert(code.end(), {EncodeMimg0(0xe5, 15, 2, false, 0, true), EncodeMimg1(32, 4, 2),
+      0x08070605, 0x00000e0b});
+  code.push_back(EncodeSop1(0x04, 126, 193));
+  code.push_back(EncodeVop2(0x1a, 2, InlineU32(4), 0));
+  for (u32 i = 0; i < 10; ++i) AppendStoreVgprAtLaneDwordOffset(&code, 32 + i, 2, OutputStart + i);
+  for (u32 i = 0; i < 6; ++i) AppendStoreVgprAtLaneDwordOffset(&code, 8 + i, 2, OutputStart + 10 + i);
+  AppendEnd(&code);
+  test.bda_mappings = {{GuestBase, NodeStart * sizeof(u32)}};
+  test.compute_info.wave_size = wave_size;
+  test.compute_info.threads_num[0] = wave_size;
+  test.compute_info.threads_num[1] = test.compute_info.threads_num[2] = 1;
+  test.compute_info.thread_ids_num = 1;
+  test.has_compute_info = true;
+  test.decoded_counts = {{"IMAGE_BVH8_INTERSECT_RAY", 1}};
+  test.required_spirv = {"trinity_bvh_intersect"};
+  return test;
+}
+
+TestCase TrinityBvhAliases() {
+  auto test = TrinityBvhIntersections(false, true, false, false);
+  test.name = "TrinityBvhRegisterAliases";
+  constexpr u32 OutputStart = 2048;
+  for (u32 lane = 0; lane < 32; ++lane) {
+    auto* out = test.expected.data() + OutputStart + lane * 16;
+    const auto* input = test.initial.data() + lane * 16;
+    const bool active = lane < 16, instance = active && (lane == 6 || lane == 7 || lane == 15);
+    if (!active) {
+      for (u32 i = 0; i < 6; ++i) out[4 + i] = input[i];
+    } else if (instance) {
+      out[4] = input[0]; out[5] = input[1];
+      out[8] = out[10]; out[9] = out[11];
+    } else {
+      // A non-instance writes its shape IDs into registers formerly holding ray XY.
+      out[10] = out[8]; out[11] = out[9];
+    }
+  }
+  auto& code = test.code;
+  code.clear();
+  AppendSMovLiteral(&code, 8, 0x100);
+  AppendSMovLiteral(&code, 9, 0);
+  AppendSMovLiteral(&code, 10, 255);
+  AppendSMovLiteral(&code, 11, 0x81c00000);
+  code.push_back(EncodeVop2(0x1a, 50, InlineU32(6), 0));
+  code.push_back(EncodeVop2(0x1a, 51, InlineU32(4), 0));
+  for (u32 i = 0; i < 10; ++i) AppendVMovLiteral(&code, i, 0xabc00000 + i);
+  for (u32 offset : {0u, 4u, 8u})
+    code.insert(code.end(), {EncodeMubuf0(offset == 8 ? 0x0f : 0x0e, offset * 4),
+        EncodeMubuf1(4 + offset, 0, 50)});
+  AppendSMovLiteral(&code, 126, 0x0000ffff);
+  code.insert(code.end(), {EncodeMimg0(0xe5, 15, 2, false, 0, true), EncodeMimg1(0, 4, 2),
+      0x08070605, 0x00000e0b});
+  code.push_back(EncodeSop1(0x04, 126, 193));
+  for (u32 i = 0; i < 10; ++i) AppendStoreVgprAtLaneDwordOffset(&code, i, 51, OutputStart + i);
+  for (u32 i = 0; i < 6; ++i) AppendStoreVgprAtLaneDwordOffset(&code, 8 + i, 51, OutputStart + 10 + i);
+  AppendEnd(&code);
+  return test;
+}
+
+std::vector<TestCase> TrinityBvhCases() {
+  return {TrinityBvhIntersections(false, true, false, false),
+          TrinityBvhIntersections(false, true, true, true),
+          TrinityBvhIntersections(true, true, false, false),
+          TrinityBvhIntersections(true, true, true, false),
+          TrinityBvhIntersections(false, false, true, true),
+          TrinityBvhIntersections(false, true, true, true, 1),
+          TrinityBvhIntersections(false, true, true, true, 2),
+          TrinityBvhIntersections(true, true, true, true, 0, true),
+          TrinityBvhIntersections(true, true, true, true, 0, false, 64),
+          TrinityBvhIntersections(false, true, false, false, 0, false, 32, true),
+          TrinityBvhIntersections(true, true, false, false, 0, false, 32, true),
+          TrinityBvhAliases()};
+}
+
 TestCase TrinityStack(u32 wave_size, u32 capacity, bool storage = false) {
   TestCase test;
   static constexpr const char* names[2][3] = {

tests/ShaderRayTracingTests.inc

@@ -38,7 +38,48 @@ void TestTrinityStackInstruction() {
   }
 }
 
+void TestTrinityBvhInstruction() {
+  namespace D = ShaderRecompiler::Decoder;
+  const std::array code{0xf1949f05u, 0x00010000u, 0x0a030201u, 0x0000040du,
+      EncodeMubuf0(0x1e, 0, false), EncodeMubuf1(0, 0, 0),
+      EncodeMubuf0(0x1e, 16, false), EncodeMubuf1(4, 0, 0),
+      EncodeMubuf0(0x1d, 32, false), EncodeMubuf1(8, 0, 0),
+      EncodeMubuf0(0x1e, 40, false), EncodeMubuf1(10, 0, 0),
+      EncodeMubuf0(0x1d, 56, false), EncodeMubuf1(14, 0, 0), EncodeSopp(0x01)};
+  D::Program decoded;
+  D::DecodeProgram(code, decoded);
+  const auto& inst = decoded.instructions.front();
+  Check(inst.opcode == D::Opcode::IMAGE_BVH8_INTERSECT_RAY && inst.word_count == 4 &&
+            inst.data_dwords == 10 && inst.image_address_components == 11 &&
+            inst.image_nsa_addr[3] == 10 && inst.image_nsa_addr[5] == 12 &&
+            inst.image_nsa_addr[6] == 13 && inst.image_nsa_addr[8] == 15 &&
+            inst.image_nsa_addr[9] == 4 && inst.src1.reg == 4,
+        "Trinity BVH8 grouped ray registers were not decoded");
+  const auto result = RecompileForTest(code, MakeCompileOptions(ShaderType::Compute));
+  Check(result.program.info.bvh8 && !result.program.info.bvh &&
+            result.program.info.uses_dma && result.program.info.images.empty(),
+        "Trinity BVH8 did not use its separate raw-memory helper");
+  Check(CountSourceOccurrences(result.ir_dump, " = Bvh8Intersect ") == 1,
+        "Trinity BVH8 result and ray writes duplicated the intersection");
+  CheckSpirvBinaryValidates(result.spirv);
+  Check(SpirvInstructionOpcodeCount(result.spirv, 54u) == 6 &&
+            SpirvInstructionOpcodeCount(result.spirv, 246u) == 1,
+        "Trinity BVH8 duplicated helpers or unrolled its four-child loop");
+  CheckSpirvBinaryValidates(RecompileForTest(code, MakeCompileOptions(ShaderType::Pixel)).spirv);
+  for (const auto words : {std::array{0xf1949f05u, 0x40010000u, 0x0a030201u, 0x0000040du},
+                           std::array{0xf1949f05u, 0x0001f700u, 0x0a030201u, 0x0000040du},
+                           std::array{0xf1949f05u, 0x00010000u, 0xfe030201u, 0x0000040du},
+                           std::array{0xf1949f05u, 0x00010700u, 0x0a030201u, 0x0000040du},
+                           std::array{0xf1949f05u, 0x00010000u, 0x0a030201u, 0x0000040bu}}) {
+    const std::array invalid{words[0], words[1], words[2], words[3], EncodeSopp(0x01)};
+    D::DecodeProgram(invalid, decoded);
+    Check(decoded.instructions[0].opcode == D::Opcode::UNSUPPORTED,
+          "Trinity BVH8 accepted truncated ray or destination registers");
+  }
+}
+
 void TestRayTracingInstructions() {
+  TestTrinityBvhInstruction();
   TestTrinityStackInstruction();
   namespace D = ShaderRecompiler::Decoder;
   const std::array code{

tests/ShaderRecompilerComputeTests.cpp

@@ -46807,6 +46807,7 @@ std::vector<TestCase> MakeCases() {
   AddCase(ScratchIsPrivatePerInvocation);
   cases.push_back(FlatStackApertures(32));
   cases.push_back(FlatStackApertures(64));
+  for (auto& test : TrinityBvhCases()) cases.push_back(std::move(test));
   for (const u32 wave : {32u, 64u}) {
     for (const u32 capacity : {8u, 16u, 31u}) cases.push_back(TrinityStack(wave, capacity));
     cases.push_back(TrinityStack(wave, 8, true));
@@ -53350,6 +53351,7 @@ int main(int argc, char **argv) {
 #endif
   if (argc == 2 && std::strcmp(argv[1], "--ray-tracing-only") == 0) {
     VulkanHarness vulkan;
+    for (const auto& test : TrinityBvhCases()) RunCase(&vulkan, test);
     for (const u32 wave : {32u, 64u}) {
       for (const u32 capacity : {8u, 16u, 31u}) RunCase(&vulkan, TrinityStack(wave, capacity));
       RunCase(&vulkan, TrinityStack(wave, 8, true));