raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-24 21:06:55 +02:00
Code Issues Packages Projects Releases Wiki Activity

shader: fix LDS counter bounds

Browse Source
This commit is contained in:
nmzik
2026-09-27 21:42:59 +04:00
parent e68b9b8d37
commit ab4b10cf48
4 changed files with 129 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/ShaderRecompilerComputeTests.cpp
+101 -5
View File
@@ -27279,17 +27279,17 @@ TestCase DsAppendConsumeUsesEncodedLdsSelector() {
using O = ShaderOpcode;
std::vector<u32> code;
AppendSMovLiteral(&code, 124, 0x0000ffffu);
AppendSMovLiteral(&code, 124, 0xffff0000u);
AppendVMovU32(&code, 1, 0);
AppendVMovU32(&code, 2, 10);
code.push_back(EncodeDs0(0x0d, 0));
code.push_back(EncodeDs1(0, 2, 1));
code.push_back(EncodeDs0(0x3e, 0));
code.push_back(EncodeDs0(0x3e, 1));
code.push_back(EncodeDs1(3, 0, 0));
code.push_back(EncodeDs0(0x3d, 0));
code.push_back(EncodeDs0(0x3d, 2));
code.push_back(EncodeDs1(4, 0, 0));
AppendSMovLiteral(&code, 124, 0);
code.push_back(EncodeDs0(0x3e, 0));
code.push_back(EncodeDs0(0x3e, 3));
code.push_back(EncodeDs1(5, 0, 0));
code.push_back(EncodeDs0(0x36, 0));
code.push_back(EncodeDs1(6, 0, 1));
@@ -27302,11 +27302,64 @@ TestCase DsAppendConsumeUsesEncodedLdsSelector() {
return {"DsAppendConsumeLdsSelector",
code,
{},
{10, 11, 0, 10},
{10, 11, 10, 11},
{O::S_MOV_B32, O::V_MOV_B32, O::DS_WRITE_B32, O::DS_READ_B32,
O::DS_APPEND, O::DS_CONSUME, O::BUFFER_STORE_DWORD, O::S_ENDPGM}};
}
TestCase DsAppendAllocatesAcrossWaves(u32 wave_size) {
using O = ShaderOpcode;
TestCase test;
test.name = wave_size == 64 ? "DsAppendAllocatesAcrossWaves64"
: "DsAppendAllocatesAcrossWaves32";
auto &code = test.code;
AppendVMovU32(&code, 2, 0);
AppendVMovU32(&code, 5, 1);
// Only workgroup lane zero initializes the shared allocation counter.
code.push_back(EncodeSop1(0x04, 10, 126));
code.push_back(EncodeVopc(0xc2, InlineU32(0), 0));
code.push_back(EncodeSop1(0x04, 126, 106));
code.push_back(EncodeDs0(0x0d, 0));
code.push_back(EncodeDs1(0, 2, 2));
code.push_back(EncodeSop1(0x04, 126, 10));
code.push_back(EncodeSopp(0x0a, 0));
for (u32 batch = 0; batch < 2; ++batch) {
// Captured CS ea0aceac sets M0=0 and uses OFFSET=1. LDS ignores M0
// and aligns OFFSET before allocating once for every active wave.
AppendSMovLiteral(&code, 124, batch == 0 ? 0 : 0xffff0000u);
code.push_back(EncodeDs0(0x3e, batch == 0 ? 1 : 3));
code.push_back(EncodeDs1(3, 0, 0));
code.push_back(EncodeVop2(0x24, 6, 127, 3));
code.push_back(EncodeVop2(0x23, 6, 126, 6));
code.push_back(EncodeVop2(0x1a, 7, InlineU32(2), 6));
// A histogram detects duplicate allocations without assuming wave order.
AppendBufferStoreOpcode(&code, 0x32, 5, 7);
}
code.push_back(EncodeSopp(0x0a, 0));
code.push_back(EncodeDs0(0x36, 0));
code.push_back(EncodeDs1(8, 0, 2));
AppendStoreVgprAtLaneDwordOffset(&code, 8, 0, 512);
AppendEnd(&code);
test.expected.assign(512, 1);
test.expected.resize(768, 512);
test.initial.resize(test.expected.size());
test.compute_info.threads_num[0] = 256;
test.compute_info.threads_num[1] = 1;
test.compute_info.threads_num[2] = 1;
test.compute_info.thread_ids_num = 1;
test.compute_info.wave_size = wave_size;
test.compute_info.lds_size_dwords = 1;
test.has_compute_info = true;
test.forbidden_spirv = {"\"gds\""};
test.opcodes = {O::S_MOV_B32, O::S_MOV_B64, O::S_BARRIER,
O::V_MOV_B32, O::V_CMP_EQ_U32, O::DS_WRITE_B32,
O::DS_APPEND, O::V_MBCNT_HI_U32_B32,
O::V_MBCNT_LO_U32_B32, O::V_LSHLREV_B32,
O::BUFFER_ATOMIC_ADD, O::DS_READ_B32,
O::BUFFER_STORE_DWORD, O::S_ENDPGM};
return test;
}
TestCase DsAppendUsesEncodedGdsSelector() {
using O = ShaderOpcode;
@@ -27344,6 +27397,46 @@ TestCase DsAppendUsesEncodedGdsSelector() {
return test;
}
TestCase DsAppendConsumeGdsRegionBounds() {
using O = ShaderOpcode;
struct Access {
u32 m0, offset;
bool valid;
};
constexpr Access accesses[] = {
{0x00000004u, 3, true},
{0x00000004u, 4, false},
{0x00000003u, 0, false},
{0xbffc0004u, 3, true},
{0xbffc0008u, 0, false},
{0xc0000004u, 0, false},
{0x0000c000u, 0xbfff, true},
{0x0000c000u, 0xc000, false},
};
TestCase test;
test.name = "DsAppendConsumeGdsRegionBounds";
auto &code = test.code;
for (const auto &access : accesses) {
AppendSMovLiteral(&code, 124, access.m0);
code.push_back(EncodeDs0(0x3e, access.offset, true));
code.push_back(EncodeDs1(0, 0, 0));
code.push_back(EncodeDs0(0x3d, access.offset, true));
code.push_back(EncodeDs1(1, 0, 0));
AppendStoreVgpr(&code, 0, static_cast<u32>(test.expected.size()));
test.expected.push_back(access.valid ? 10 : 0);
AppendStoreVgpr(&code, 1, static_cast<u32>(test.expected.size()));
test.expected.push_back(access.valid ? 11 : 0);
}
AppendEnd(&code);
test.initial.resize(test.expected.size());
// Backing beyond the guest partition must not make an invalid region valid.
test.gds_initial.assign(0x10000u / sizeof(u32), 10);
test.expected_gds = test.gds_initial;
test.opcodes = {O::S_MOV_B32, O::DS_APPEND, O::DS_CONSUME,
O::BUFFER_STORE_DWORD, O::S_ENDPGM};
return test;
}
TestCase DsGdsSubdwordAndAtomicWrites() {
using O = ShaderOpcode;
@@ -29686,7 +29779,10 @@ std::vector<TestCase> MakeCases() {
AddCase(DsReadU16D16CapturedPreservesHighHalf);
AddCase(DsReadU16D16HiCapturedPreservesLowHalf);
AddCase(DsAppendConsumeUsesEncodedLdsSelector);
cases.push_back(DsAppendAllocatesAcrossWaves(32));
cases.push_back(DsAppendAllocatesAcrossWaves(64));
AddCase(DsAppendUsesEncodedGdsSelector);
AddCase(DsAppendConsumeGdsRegionBounds);
AddCase(DsGdsSubdwordAndAtomicWrites);
AddCase(DsReadWrite2Variants);
AddCase(DsWideReadSnapshotsOverlappingAddress);