raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-25 12:26:51 +01:00
Code Issues Packages Projects Releases Wiki Activity

gpu: decode RELEASE_MEM directly from the driver's native packet

Browse Source
This commit is contained in:
nmzik
2026-10-23 21:11:07 +02:00
parent 227f12ee4b
commit bf5dd62eea
10 changed files with 121 additions and 356 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/ShaderRecompilerComputeTests.cpp
+52 -13
View File
@@ -2943,7 +2943,7 @@ public:
std::array<uint32_t, 8> packet{};
const auto address = reinterpret_cast<uint64_t>(destination);
const auto event_index = (action >= 0x2fu ? 6u : 5u);
packet[0] = KYTY_PM4(8, Pm4::IT_NOP, Pm4::R_RELEASE_MEM);
packet[0] = KYTY_PM4(8, Pm4::IT_RELEASE_MEM, 0);
packet[1] = action | (event_index << 8u) | (gcr_cntl << 12u);
packet[2] = (interrupt_selector << 24u) | (data_sel << 29u);
packet[3] = static_cast<uint32_t>(address);
@@ -3075,16 +3075,34 @@ public:
gpu.SendCommandSync([&] {
processor->BufferInit();
const auto clock_before = Sync::ReadReferenceClock();
auto clock_release =
make_release_mem(3, 2, &release_label, UINT64_MAX, 0x04u);
Pm4Execution clock_execution;
Require("GpuCommandLane", "release reference clock with interrupt",
processor->Process(clock_execution, clock_release) ==
Pm4ProcessResult::Complete &&
release_label >= clock_before &&
release_label <= Sync::ReadReferenceClock(),
"clock write with writeback and interrupt lost its data");
// Every end-of-pipe and end-of-shader action, with or without GL2
// writeback, writes each label size and splits for its interrupt.
for (const uint32_t action : {0x04u, 0x14u, 0x28u, 0x2bu, 0x2du, 0x2fu, 0x30u}) {
for (const uint32_t gcr : {0u, 1u << 9u}) {
for (const uint32_t data_sel : {1u, 2u, 3u}) {
release_label = 0;
const auto clock_before = Sync::ReadReferenceClock();
auto release = make_release_mem(data_sel, 2, &release_label,
0x1122334455667788ull, action, gcr);
release[3] |= 1u; // ADDRESS_LO starts at bit 2.
Pm4Execution release_execution;
const auto release_tick = gpu_scheduler.CurrentTick();
const auto release_result =
processor->Process(release_execution, release);
bool label_written =
release_label == (data_sel == 1 ? 0x55667788ull : 0x1122334455667788ull);
if (data_sel == 3) {
label_written = release_label >= clock_before &&
release_label <= Sync::ReadReferenceClock();
}
Require("GpuCommandLane", "release with write-confirm interrupt",
release_result == Pm4ProcessResult::Complete &&
label_written &&
gpu_scheduler.CurrentTick() == release_tick + 1,
"an end-of-pipe action lost its label or interrupt split");
}
}
}
for (const auto interrupt : {0u, 3u, 1u, 2u}) {
release_label = 0;
@@ -3213,7 +3231,7 @@ public:
void *destination, uint64_t value, uint32_t context_id) {
std::array<uint32_t, 8> packet{};
const auto address = reinterpret_cast<uint64_t>(destination);
packet[0] = KYTY_PM4(8, Pm4::IT_NOP, Pm4::R_RELEASE_MEM);
packet[0] = KYTY_PM4(8, Pm4::IT_RELEASE_MEM, 0);
packet[1] = 0x28u | (5u << 8u);
packet[2] = (interrupt_selector << 24u) | (data_selector << 29u);
packet[3] = static_cast<uint32_t>(address);
@@ -49632,7 +49650,7 @@ void CheckPm4IndirectControlFlow(RenderContext &renderer) {
disabled_commands[5] == KYTY_PM4(2, Pm4::IT_NOP, Pm4::R_ZERO) &&
disabled_commands[6] == 0x68750123u &&
release[0] == KYTY_PM4(8, Pm4::IT_NOP, Pm4::R_ZERO) &&
unpatched_commands[7] == KYTY_PM4(8, Pm4::IT_NOP, Pm4::R_RELEASE_MEM) &&
unpatched_commands[7] == 0xc0064900u &&
Gen5::AgcGetPacketSize(disabled_commands.data()) == disabled_write.size() &&
Gen5::AgcGetPacketSize(disabled_commands.data() + 5) == 2 &&
Gen5::AgcGetPacketSize(release) == 8 &&
@@ -49647,6 +49665,27 @@ void CheckPm4IndirectControlFlow(RenderContext &renderer) {
processor.Process(disabled_execution, disabled_commands) == Pm4ProcessResult::Complete &&
returned == 0 && released == 0 && selected == 88,
"a disabled packet still executed or hid the packet after it");
// Like the driver, the end-of-pipe patch helpers take only a native
// RELEASE_MEM, keep the address alignment bits and skip kSendIntCtx.
std::array<uint32_t, 8> patched{};
std::copy_n(unpatched_commands.begin() + 7, patched.size(), patched.begin());
patched[3] |= 1u;
const auto* label = reinterpret_cast<const volatile Gen5::Label*>(0x123456789abcdef7ull);
bool patch_helpers_match =
Gen5::AgcQueueEndOfPipeActionPatchAddress(patched.data(), label) == 0 &&
patched[3] == 0x9abcdef5u && patched[4] == 0x12345678u &&
Gen5::AgcQueueEndOfPipeActionPatchData(patched.data(), 0x1122334455667788ull) == 0 &&
patched[5] == 0x55667788u && patched[6] == 0x11223344u;
patched[2] |= 4u << 24u;
patch_helpers_match &= Gen5::AgcQueueEndOfPipeActionPatchAddress(patched.data(), nullptr) == 0 &&
patched[3] == 0x9abcdef5u && patched[4] == 0x12345678u;
patched[0] = KYTY_PM4(8, Pm4::IT_NOP, 0x18);
patch_helpers_match &=
Gen5::AgcQueueEndOfPipeActionPatchAddress(patched.data(), label) ==
static_cast<int>(0x8a6c000cu) &&
Gen5::AgcQueueEndOfPipeActionPatchData(patched.data(), 0) == static_cast<int>(0x8a6c000cu);
Require(name, "end-of-pipe patch helpers", patch_helpers_match,
"RELEASE_MEM or its patch helpers diverged from the driver");
const auto then_commands = write(&selected, 11);
const auto else_commands = write(&selected, 33);
const auto branch_suffix = write(&selected, 44);