raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-25 20:36:54 +01:00
Code Issues Packages Projects Releases Wiki Activity

shader: clamp V_FRACT_F16/F32/F64 results below one

Browse Source
This commit is contained in:
TheCruZ
2026-10-25 00:36:03 +02:00
committed by nmzik
parent 40aaab482e
commit e864cb20bd
6 changed files with 73 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/ShaderRecompilerComputeTests.cpp
+39 -5
View File
@@ -27670,16 +27670,16 @@ TestCase VectorSpecialF16Ops() {
TestCase VectorFractF16CapturedAndEdges() {
using O = ShaderOpcode;
// RDNA2 section 12.8: x - floor(x), rounded to half precision.
// x - floor(x) rounded to half precision, clamped to 0x3bff below one.
// Load at runtime to exercise the emitted operation, including half subnormals.
const std::array<u32, 18> inputs{
0x3d00u, 0xbd00u, 0x0000u, 0x8000u, 0x0001u, 0x8001u,
0x03ffu, 0x83ffu, 0x7bffu, 0xfbffu, 0x3800u, 0xb800u,
0x8c00u, 0x8c01u, 0x7c00u, 0xfc00u, 0x7e55u, 0x7d01u};
const std::array<u32, 18> fractions{
0x3400u, 0x3a00u, 0x0000u, 0x0000u, 0x0001u, 0x3c00u,
0x03ffu, 0x3c00u, 0x0000u, 0x0000u, 0x3800u, 0x3800u,
0x3c00u, 0x3bffu, 0x7e00u, 0x7e00u, 0x7e00u, 0x7e00u};
0x3400u, 0x3a00u, 0x0000u, 0x0000u, 0x0001u, 0x3bffu,
0x03ffu, 0x3bffu, 0x0000u, 0x0000u, 0x3800u, 0x3800u,
0x3bffu, 0x3bffu, 0x7e00u, 0x7e00u, 0x7e00u, 0x7e00u};
TestCase test;
test.name = "VectorFractF16CapturedAndEdges";
for (u32 bits : inputs) {
@@ -27710,6 +27710,37 @@ TestCase VectorFractF16CapturedAndEdges() {
return test;
}
// V_FRACT_F32 clamps to 0x3f7fffff where x - floor(x) rounds to 1.0, before output modifiers.
TestCase VectorFractF32Edges() {
using O = ShaderOpcode;
constexpr std::array<std::array<u32, 3>, 4> cases{{
{0xb3800000u, 0, 0x3f7fffffu}, // -2^-24: exact
{0xb3000000u, 0, 0x3f7fffffu}, // -2^-25: the tie rounds to 1.0
{0xb3000000u, 1, 0x3fffffffu}, // omod *2 applies after the clamp
{0x7fc00001u, 0, 0x7fc00000u}}}; // Quiet NaN class
TestCase test;
test.name = "VectorFractF32Edges";
for (const auto &[source, omod, expected] : cases) test.initial.push_back(source);
test.expected = test.initial;
auto &code = test.code;
for (u32 i = 0; i < cases.size(); i++) {
AppendVMovU32(&code, 30, i * 4u);
AppendBufferLoadDword(&code, 7, 30);
AppendVop3(&code, 0x1a0, 9, Vgpr(7), 0, 0, 0, 0, false, cases[i][1]);
if (i + 1u == cases.size()) {
code.push_back(EncodeVop2(0x1b, 9, 255u, 9));
code.push_back(0x7fc00000u);
}
AppendStoreVgpr(&code, 9, static_cast<u32>(cases.size()) + i);
test.expected.push_back(cases[i][2]);
}
AppendEnd(&code);
test.opcodes = {O::V_MOV_B32, O::BUFFER_LOAD_DWORD, O::V_FRACT_F32,
O::V_AND_B32, O::BUFFER_STORE_DWORD, O::S_ENDPGM};
return test;
}
TestCase VectorFractF16Modifiers() {
using O = ShaderOpcode;
@@ -29628,7 +29659,7 @@ TestCase VectorFractF64CapturedAndEdges() {
{0x4320000000000001ull, 0x3fe0000000000000ull}, // 2^51 + .5
{0, 0}, {0x8000000000000000ull, 0},
{0x401c000000000000ull, 0}, {0xc01c000000000000ull, 0},
{1, 1}, {0x8000000000000001ull, 0x3ff0000000000000ull},
{1, 1}, {0x8000000000000001ull, 0x3fefffffffffffffull}, // Clamped below one.
{0x000fffffffffffffull, 0x000fffffffffffffull},
{0x7fefffffffffffffull, 0},
{0x7ff0000000000000ull, 0x7ff8000000000000ull},
@@ -44123,6 +44154,7 @@ std::vector<TestCase> MakeCases() {
AddCase(VectorMinMaxMed3F16Ops);
AddCase(VectorSpecialF16Ops);
AddCase(VectorFractF16CapturedAndEdges);
AddCase(VectorFractF32Edges);
AddCase(VectorFractF16Modifiers);
AddCase(VectorCosF16CapturedSdwaAndEdges);
AddCase(VectorSinF16SdwaAndEdges);
@@ -50749,6 +50781,8 @@ int main(int argc, char **argv) {
if (argc == 2 && std::strcmp(argv[1], "--fract-f16-only") == 0) {
VulkanHarness vulkan;
RunCase(&vulkan, VectorFractF16CapturedAndEdges());
RunCase(&vulkan, VectorFractF32Edges());
RunCase(&vulkan, VectorFractF64CapturedAndEdges());
RunCase(&vulkan, VectorFractF16Modifiers());
RunCase(&vulkan, NativeAndSdwa16BitDestinationWrites());
RunCase(&vulkan, VectorSpecialF16Ops());