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: lower complete wave minima for sparse fragment subgroups

Browse Source
This commit is contained in:
nmzik
2026-10-14 18:03:48 +02:00
parent 892e70c8a1
commit 302240da70
12 changed files with 330 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/ShaderRecompilerComputeTests.cpp
+63
View File
@@ -39536,6 +39536,63 @@ TestCase ImageAtomicGlc0DoesNotReturnOldValue() {
return test;
}
void AppendWave64Minimum(std::vector<u32> &code) {
code.insert(code.end(), {EncodeSop1(0x04, 24, 126), // Save contributing lanes.
EncodeSop1(0x28, 106, 126)});
AppendVop3(&code, 0x101, 13, 193, Vgpr(13), 24); // Inactive lanes contribute UINT_MAX.
for (u32 shift : {1u, 2u, 4u, 8u}) {
code.push_back(EncodeVop2(0x13, 13, 250, 13));
code.push_back(EncodeVop2Dpp(13, 0x110u + shift));
}
// Exact RDNA2 wave64 reduction from PS 964747887d898821, PCs 0x224..0x244.
code.insert(code.end(), {0xd778100cu, 0x0305830du,
EncodeVop2(0x13, 13, Vgpr(13), 12), EncodeSop1(0x04, 126, 106),
0xd7600006u, 0x00013f0du, 0xd7600007u, 0x00017f0du,
EncodeSop2(0x07, 8, 6, 7), EncodeVop1(0x01, 0, 8)});
}
GraphicsCase GraphicsSparseWaveMinimum() {
GraphicsCase test;
test.name = "GraphicsSparseWaveMinimum";
auto &code = test.fragment_code;
code = {EncodeVintrp(0x02, 13, 0, 0, 2)};
AppendWave64Minimum(code);
code.insert(code.end(), {EncodeExp0(0x00, 0xf), EncodeExp1(0, 0, 0, 0)});
AppendEnd(&code);
test.pixel_interpolator_settings = {0x00000400u};
test.expected_pixel = {0x3e800000u, 0x3e800000u, 0x3e800000u, 0x3e800000u};
test.opcodes = {ShaderOpcode::V_MIN_U32, ShaderOpcode::V_PERMLANEX16_B32,
ShaderOpcode::V_READLANE_B32, ShaderOpcode::S_MIN_U32};
return test;
}
TestCase ComputeWave64Minimum(bool masked) {
TestCase test;
test.name = masked ? "ComputeWave64MinimumMasked" : "ComputeWave64Minimum";
auto &code = test.code;
code = {EncodeVop2(0x1a, 20, InlineU32(2), 0)};
AppendVMovU32(&code, 13, 100);
code.push_back(EncodeVop2(0x26, 13, Vgpr(13), 0)); // 100 - lane: minimum is in the upper half.
if (masked) {
AppendSMovLiteral(&code, 126, 0xfffffffeu);
AppendSMovLiteral(&code, 127, 0x7fffffffu);
}
AppendWave64Minimum(code);
AppendBufferStoreDword(&code, 0, 20);
AppendEnd(&code);
test.initial.assign(64, 0xa5a5a5a5u);
test.expected.assign(64, masked ? 38u : 37u);
if (masked) test.expected.front() = test.expected.back() = 0xa5a5a5a5u;
test.compute_info.threads_num[0] = 64;
test.compute_info.threads_num[1] = test.compute_info.threads_num[2] = 1;
test.compute_info.thread_ids_num = 1;
test.compute_info.wave_size = 64;
test.has_compute_info = true;
test.opcodes = {ShaderOpcode::V_MIN_U32, ShaderOpcode::V_PERMLANEX16_B32,
ShaderOpcode::V_READLANE_B32, ShaderOpcode::S_MIN_U32};
return test;
}
GraphicsCase GraphicsInterpolationExport() {
using O = ShaderOpcode;
@@ -40193,6 +40250,8 @@ std::vector<TestCase> MakeCases() {
AddCase(VectorDppRowXmask);
cases.push_back(VectorDppRowShare(32));
cases.push_back(VectorDppRowShare(64));
cases.push_back(ComputeWave64Minimum(false));
cases.push_back(ComputeWave64Minimum(true));
AddCase(VectorDppBankMaskPreservesDestination);
AddCase(VectorDppBoundsControlZeroPreservesDestination);
AddCase(Vop3FmacF32NegatedSourceAccumulates);
@@ -40586,6 +40645,7 @@ std::vector<TestCase> MakeCases() {
std::vector<GraphicsCase> MakeGraphicsCases() {
return {
GraphicsSparseWaveMinimum(),
GraphicsInterpolationExport(),
GraphicsPositionWExport(),
GraphicsPackedHalfCentroid(),
@@ -45974,6 +46034,9 @@ int main(int argc, char **argv) {
}
if (argc == 2 && std::strcmp(argv[1], "--dpp-only") == 0) {
VulkanHarness vulkan;
RunGraphicsCase(&vulkan, GraphicsSparseWaveMinimum());
RunCase(&vulkan, ComputeWave64Minimum(false));
RunCase(&vulkan, ComputeWave64Minimum(true));
RunCase(&vulkan, VectorDpp8Captured(false));
for (const u32 wave_size : {32u, 64u})
for (const bool fi : {false, true})