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

Fix scalar buffer reads across sparse guest mappings

Browse Source
This commit is contained in:
nmzik
2026-10-16 18:19:09 +02:00
parent b899717900
commit d7974dc763
13 changed files with 250 additions and 290 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/ShaderRecompilerComputeTests.cpp
+64 -9
View File
@@ -30517,7 +30517,7 @@ TestCase ScalarMemoryLoadVariants() {
std::vector<u32> expected = initial;
expected.insert(expected.end(), initial.begin(), initial.end());
return {"ScalarMemoryLoadVariants",
TestCase test{"ScalarMemoryLoadVariants",
code,
initial,
expected,
@@ -30526,6 +30526,8 @@ TestCase ScalarMemoryLoadVariants() {
O::S_BUFFER_LOAD_DWORDX2, O::S_BUFFER_LOAD_DWORDX4,
O::S_BUFFER_LOAD_DWORDX8, O::S_BUFFER_LOAD_DWORDX16, O::V_MOV_B32,
O::BUFFER_STORE_DWORD, O::S_ENDPGM}};
test.bda_mappings = {{0u, 0u}};
return test;
}
TestCase ScalarBufferOffsetAlignmentAndCarry() {
@@ -30548,6 +30550,7 @@ TestCase ScalarBufferOffsetAlignmentAndCarry() {
{O::S_MOV_B32, O::S_ADD_U32, O::S_BUFFER_LOAD_DWORD,
O::V_MOV_B32, O::BUFFER_STORE_DWORD, O::S_ENDPGM}};
test.forbidden_spirv = {"flattened_srt"};
test.bda_mappings = {{0u, 0u}};
test.compute_info.threads_num[0] = 1;
test.compute_info.threads_num[1] = 1;
test.compute_info.threads_num[2] = 1;
@@ -31181,7 +31184,7 @@ TestCase BufferLoadsGpuSelectedDescriptors(u32 width, u32 wave = 32u) {
}
}
AppendEnd(&test.code);
test.bda_mappings = {{GuestBase, 0}};
test.bda_mappings = {{0u, 0u}, {GuestBase, 0}};
test.opcodes = {O::V_MOV_B32, O::S_MOV_B32, O::BUFFER_LOAD_DWORD,
O::V_READFIRSTLANE_B32, O::S_MUL_I32, O::S_BUFFER_LOAD_DWORDX4,
O::BUFFER_STORE_DWORD, O::S_ENDPGM};
@@ -31365,7 +31368,7 @@ TestCase BufferLoadGpuSelectedDescriptors(u32 wave, ShaderOpcode load) {
code.push_back(EncodeSopc(0x0a, 28, 29));
code.push_back(EncodeSopp(0x05, loop - static_cast<u32>(code.size()) - 1u));
AppendEnd(&code);
test.bda_mappings = {{GuestBase, 0}};
test.bda_mappings = {{0u, 0u}, {GuestBase, 0}};
test.fault_free_address = GuestBase + 0x100000u;
test.compute_info.threads_num[0] = 1;
test.compute_info.threads_num[1] = test.compute_info.threads_num[2] = 1;
@@ -31417,11 +31420,17 @@ void CheckSharedIndirectBufferLoads(VulkanHarness &vulkan) {
auto compiled = CompileCase(test, vulkan.SubgroupSize());
struct Profile { u32 components = 0, uses = 0, definitions = 0, calls = 0, reads = 0; };
std::map<std::string, Profile> profiles;
const u32 halves = wave == 64u && vulkan.SubgroupSize() == 32u ? 2u : 1u;
u32 scalar_reads = 0;
for (const auto *block : compiled.program.blocks)
for (const auto &inst : *block) {
const auto components = BufferComponentCount(inst.GetOpcode());
if (components == 0u || inst.GetOpcode() == ValueOpcode::ReadConstBuffer) continue;
if (components == 0u) continue;
const auto &memory = compiled.program.memory_info[inst.Flags<MemoryFlags>().index];
if (inst.GetOpcode() == ValueOpcode::ReadConstBuffer) {
scalar_reads += !memory.planning_only;
continue;
}
if (memory.kind != ResourceKind::IndirectBuffer || memory.formatted) continue;
const auto name = "load_indirect_buffer_" + std::to_string(memory.data_bits) + "x" +
std::to_string(components) + (memory.coherent ? "_coherent" : "");
@@ -31452,11 +31461,11 @@ void CheckSharedIndirectBufferLoads(VulkanHarness &vulkan) {
}
offset += words.size();
}
Require(test.name, "raw load function scope", functions.size() == profiles.size() && inline_reads == 0u,
Require(test.name, "raw load function scope", functions.size() == profiles.size() && inline_reads == scalar_reads * halves,
"a raw load profile was unused, missing or still emitted inline");
for (const auto &[name, profile] : profiles)
Require(test.name, name.c_str(), profile.uses != 0u && profile.definitions == 1u &&
profile.calls == profile.uses * (wave == 64u && vulkan.SubgroupSize() == 32u ? 2u : 1u) &&
profile.calls == profile.uses * halves &&
profile.reads == profile.components,
"raw load helpers duplicated definitions, calls or component reads");
auto buffer = vulkan.CreateStorageBuffer(test.name, test.initial, test.initial.size(), true);
@@ -33760,6 +33769,48 @@ TestCase ScalarBufferFromLoopReadlane(u32 wave_size) {
return test;
}
TestCase ScalarBufferLoadsDisjointPages() {
using O = ShaderOpcode;
constexpr u32 GuestBase = 0x10000, PageSize = 0x10000, PageStride = 0x400000;
TestCase test;
test.name = "ScalarBufferLoadsDisjointPages";
auto &code = test.code;
code.push_back(EncodeSop2(0x1e, 12, 4, InlineU32(22)));
code.push_back(EncodeSmem0(0x0a, 20, 4));
code.push_back(EncodeSmem1(256, 12));
code.push_back(EncodeVop1(0x01, 0, 4));
code.push_back(EncodeVop2(0x1a, 0, InlineU32(2), 0));
for (u32 component = 0; component < 4; ++component)
AppendStoreSgprAtLaneDwordOffset(&code, 20 + component, 0, component);
AppendEnd(&code);
test.initial.resize(2 * PageSize / sizeof(u32));
for (u32 page = 0; page < 2; ++page) {
for (u32 component = 0; component < 4; ++component) {
const u32 value = 0x12340000u + page * 16 + component;
test.initial[(page * PageSize + 256) / sizeof(u32) + component] = value;
test.expected.push_back(value);
}
}
test.user_data = MakeNativeUserData(nullptr);
test.user_data[8] = GuestBase;
test.user_data[10] = 32 * 1024 * 1024;
test.has_user_data = true;
// A large guest descriptor covers separate allocations, while its SSBO view ends at the first.
test.storage_buffer_range_bytes = PageSize;
test.bda_mappings = {{GuestBase, 0}, {GuestBase + PageStride, PageSize}};
test.opcodes = {O::S_LSHL_B32, O::S_BUFFER_LOAD_DWORDX4, O::V_MOV_B32,
O::V_LSHLREV_B32, O::V_ADD_NC_U32, O::BUFFER_STORE_DWORD, O::S_ENDPGM};
test.compute_info.wave_size = 32;
test.compute_info.threads_num[0] = 1;
test.compute_info.threads_num[1] = test.compute_info.threads_num[2] = 1;
test.compute_info.group_id[0] = true;
test.compute_info.workgroup_register = 4;
test.has_compute_info = true;
test.dispatch_x = 2;
return test;
}
TestCase FlatVirtualAddressRebasesGuestAllocation() {
using O = ShaderOpcode;
@@ -38031,7 +38082,7 @@ void CheckGpuSelectedRecordBufferAccess(VulkanHarness &vulkan) {
}
// Compile against poison selectors. A separate GPU dispatch writes the actual keys.
auto compiled = CompileCase(test, vulkan.SubgroupSize());
if (table_bytes == 0u || components != 0u)
if (components != 0u)
test.forbidden_spirv = {"convert_indirect_buffer_format", "load_indirect_buffer_",
"indirect_buffer_format_geometry"};
else
@@ -38047,8 +38098,8 @@ void CheckGpuSelectedRecordBufferAccess(VulkanHarness &vulkan) {
if (opcode == spv::OpFunctionCall && words[3] == bda_function) ++bda_calls;
offset += words.size();
}
Require(test.name, "empty descriptor table specialization", (bda_calls == 0u) == (table_bytes == 0u || components != 0u),
"empty table retained BDA reads or nonempty table lost GPU descriptor loads");
Require(test.name, "GPU descriptor reads", (bda_calls == 0u) == (components != 0u),
"CPU-planned descriptors retained BDA reads or runtime descriptors lost them");
for (const auto &descriptor : compiled.resources.buffers) {
ShaderBufferResource resource{};
std::copy_n(descriptor.dwords.begin(), 4u, resource.fields);
@@ -42023,6 +42074,7 @@ std::vector<TestCase> MakeCases() {
AddCase(ScalarLoadAlignsDynamicBase);
cases.push_back(ScalarBufferFromLoopReadlane(32));
cases.push_back(ScalarBufferFromLoopReadlane(64));
AddCase(ScalarBufferLoadsDisjointPages);
AddCase(BufferLoadStore);
AddCase(BufferLoadDwordOffenIdxenUsesVaddrPlusOneOffset);
AddCase(BufferStoreDwordOffenIdxenUsesVaddrPlusOneOffset);
@@ -48020,6 +48072,9 @@ int main(int argc, char **argv) {
}
if (argc == 2 && std::strcmp(argv[1], "--indirect-buffer-only") == 0) {
VulkanHarness vulkan;
RunCase(&vulkan, ScalarBufferLoadsDisjointPages());
RunCase(&vulkan, ScalarMemoryLoadVariants());
RunCase(&vulkan, ScalarBufferOffsetAlignmentAndCarry());
RunCase(&vulkan, ScalarBufferFromLoopReadlane(32));
RunCase(&vulkan, ScalarBufferFromLoopReadlane(64));
CheckBoundedFormattedBufferCopy(vulkan);