raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-20 10:56:58 +02:00
Code Issues Packages Projects Releases Wiki Activity

shader: share scalar condition results with direct subgroup votes

Browse Source 
Co-authored-by: fxpw <fxpw@outlook.com>
This commit is contained in:
nmzik
2026-10-15 00:07:35 +02:00
parent 50844badbf
commit bb2d6dbeb3
7 changed files with 109 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/ShaderRecompilerComputeTests.cpp
+25 -8
View File
@@ -2017,6 +2017,9 @@ public:
vk::PhysicalDeviceProperties2 properties{};
properties.pNext = &subgroup;
m_physical_device.getProperties2(&properties);
Require("VulkanHarness", "subgroup vote support",
bool(subgroup.supportedOperations & vk::SubgroupFeatureFlagBits::eVote),
"device does not support subgroup All/Any votes");
return subgroup.subgroupSize;
}
// A consumer GPU may lack the production rasterization features the fragment
@@ -30000,10 +30003,13 @@ TestCase BranchVccnzUsesWaveMask() {
return test;
}
TestCase ScalarBranchRestoresInactiveLanes(u32 wave_size, u32 threads) {
TestCase ScalarBranchRestoresInactiveLanes(u32 wave_size, u32 threads, bool upper_nonzero = false) {
using O = ShaderOpcode;
TestCase test;
if (threads < wave_size) {
if (upper_nonzero) {
test.name = wave_size == 32 ? "ScalarBranchRestoresLastLaneWave32"
: "ScalarBranchRestoresUpperHalfWave64";
} else if (threads < wave_size) {
test.name = wave_size == 32 ? "ScalarBranchRestoresPartialWave32"
: "ScalarBranchRestoresPartialWave64";
} else {
@@ -30014,10 +30020,16 @@ TestCase ScalarBranchRestoresInactiveLanes(u32 wave_size, u32 threads) {
code = {EncodeVop1(0x01, 1, InlineU32(32)),
EncodeVop2(0x25, 1, Vgpr(0), 1),
EncodeSop1(0x04, 8, 126), // Save all participating lanes.
EncodeSMovB32(20, InlineU32(0)),
EncodeSMovB32(126, InlineU32(1)), EncodeSMovB32(127, InlineU32(0)),
EncodeSopp(0x08, 3), // EXEC is nonzero for the whole wave.
EncodeSop1(0x04, 126, 8)};
EncodeSMovB32(20, InlineU32(0))};
AppendSMovLiteral(&code, 126, upper_nonzero ? (wave_size == 32 ? 0x80000000u : 0u) : 1u);
AppendSMovLiteral(&code, 127, upper_nonzero && wave_size == 64 ? 0x80000000u : 0u);
if (upper_nonzero) {
code.push_back(EncodeSopp(0x09, 1)); // Last lane makes the whole-wave EXEC nonzero.
code.push_back(EncodeSopp(0x02, 3));
} else {
code.push_back(EncodeSopp(0x08, 3)); // EXEC is nonzero for the whole wave.
}
code.push_back(EncodeSop1(0x04, 126, 8));
// The branch restores inactive lanes before reading the last lane's value.
AppendVop3(&code, 0x360, 20, Vgpr(1), InlineU32(threads - 1));
code.push_back(EncodeSop1(0x04, 126, 8));
@@ -30026,9 +30038,10 @@ TestCase ScalarBranchRestoresInactiveLanes(u32 wave_size, u32 threads) {
test.initial.assign(threads, 0);
test.expected.assign(threads, 32 + threads - 1);
test.opcodes = {O::V_MOV_B32, O::V_ADD_NC_U32, O::S_MOV_B32, O::S_MOV_B64,
O::S_CBRANCH_EXECZ, O::V_READLANE_B32, O::V_LSHLREV_B32,
upper_nonzero ? O::S_CBRANCH_EXECNZ : O::S_CBRANCH_EXECZ,
O::V_READLANE_B32, O::V_LSHLREV_B32,
O::BUFFER_STORE_DWORD, O::S_ENDPGM};
test.required_spirv = {"OpGroupNonUniformBallot", "OpAll"};
test.required_spirv = {upper_nonzero ? "OpGroupNonUniformAny" : "OpGroupNonUniformAll"};
test.compute_info.wave_size = wave_size;
test.compute_info.threads_num[0] = threads;
test.compute_info.thread_ids_num = 1;
@@ -41148,6 +41161,8 @@ std::vector<TestCase> MakeCases() {
AddCase([] { return ScalarBranchRestoresInactiveLanes(64, 64); });
AddCase([] { return ScalarBranchRestoresInactiveLanes(32, 4); });
AddCase([] { return ScalarBranchRestoresInactiveLanes(64, 4); });
AddCase([] { return ScalarBranchRestoresInactiveLanes(32, 32, true); });
AddCase([] { return ScalarBranchRestoresInactiveLanes(64, 64, true); });
AddCase(BranchVccnzUsesCarryProducedWaveMask);
AddCase(ScalarMemRealtimeCapturedPlaceholder);
AddCase(ScalarMemoryLoadVariants);
@@ -47204,6 +47219,8 @@ int main(int argc, char **argv) {
RunCase(&vulkan, ScalarBranchRestoresInactiveLanes(64, 64));
RunCase(&vulkan, ScalarBranchRestoresInactiveLanes(32, 4));
RunCase(&vulkan, ScalarBranchRestoresInactiveLanes(64, 4));
RunCase(&vulkan, ScalarBranchRestoresInactiveLanes(32, 32, true));
RunCase(&vulkan, ScalarBranchRestoresInactiveLanes(64, 64, true));
RunCase(&vulkan, BranchVccnzUsesCarryProducedWaveMask());
RunCase(&vulkan, ImageSampleAndGather());
RunCase(&vulkan, ImageGatherExplicitLod());