3
0
Fork 0
forked from suyu/suyu

Maxwell3D: Correct query reporting.

This commit is contained in:
Fernando Sahmkow 2020-01-27 22:48:15 -04:00 committed by FernandoS27
parent 84ea9c2b42
commit 0cb3bcfbb7
2 changed files with 61 additions and 54 deletions

View file

@ -519,61 +519,66 @@ void Maxwell3D::ProcessFirmwareCall4() {
regs.reg_array[0xd00] = 1; regs.reg_array[0xd00] = 1;
} }
void Maxwell3D::ProcessQueryGet() { void Maxwell3D::StampQueryResult(u64 payload, bool long_query) {
struct LongQueryResult {
u64_le value;
u64_le timestamp;
};
static_assert(sizeof(LongQueryResult) == 16, "LongQueryResult has wrong size");
const GPUVAddr sequence_address{regs.query.QueryAddress()}; const GPUVAddr sequence_address{regs.query.QueryAddress()};
// Since the sequence address is given as a GPU VAddr, we have to convert it to an application if (long_query) {
// VAddr before writing. // Write the 128-bit result structure in long mode. Note: We emulate an infinitely fast
// GPU, this command may actually take a while to complete in real hardware due to GPU
// wait queues.
LongQueryResult query_result{};
query_result.value = payload;
// TODO(Subv): Generate a real GPU timestamp and write it here instead of CoreTiming
query_result.timestamp = system.CoreTiming().GetTicks();
memory_manager.WriteBlock(sequence_address, &query_result, sizeof(query_result));
} else {
memory_manager.Write<u32>(sequence_address, static_cast<u32>(payload));
}
}
void Maxwell3D::ProcessQueryGet() {
// TODO(Subv): Support the other query units. // TODO(Subv): Support the other query units.
ASSERT_MSG(regs.query.query_get.unit == Regs::QueryUnit::Crop, ASSERT_MSG(regs.query.query_get.unit == Regs::QueryUnit::Crop,
"Units other than CROP are unimplemented"); "Units other than CROP are unimplemented");
u64 result = 0; switch (regs.query.query_get.operation) {
case Regs::QueryOperation::Release: {
// TODO(Subv): Support the other query variables u64 result = regs.query.query_sequence;
StampQueryResult(result, regs.query.query_get.short_query == 0);
break;
}
case Regs::QueryOperation::Acquire: {
// Todo(Blinkhawk): Under this operation, the GPU waits for the CPU
// to write a value that matches the current payload.
UNIMPLEMENTED_MSG("Unimplemented query operation ACQUIRE");
break;
}
case Regs::QueryOperation::Counter: {
u64 result{};
switch (regs.query.query_get.select) { switch (regs.query.query_get.select) {
case Regs::QuerySelect::Zero: case Regs::QuerySelect::Zero:
// This seems to actually write the query sequence to the query address. result = 0;
result = regs.query.query_sequence;
break; break;
default: default:
result = 1; result = 1;
UNIMPLEMENTED_MSG("Unimplemented query select type {}", UNIMPLEMENTED_MSG("Unimplemented query select type {}",
static_cast<u32>(regs.query.query_get.select.Value())); static_cast<u32>(regs.query.query_get.select.Value()));
} }
StampQueryResult(result, regs.query.query_get.short_query == 0);
// TODO(Subv): Research and implement how query sync conditions work. break;
}
struct LongQueryResult { case Regs::QueryOperation::Trap: {
u64_le value; UNIMPLEMENTED_MSG("Unimplemented query operation TRAP");
u64_le timestamp; break;
}; }
static_assert(sizeof(LongQueryResult) == 16, "LongQueryResult has wrong size"); default: {
UNIMPLEMENTED_MSG("Unknown query operation");
switch (regs.query.query_get.mode) {
case Regs::QueryMode::Write:
case Regs::QueryMode::Write2: {
u32 sequence = regs.query.query_sequence;
if (regs.query.query_get.short_query) {
// Write the current query sequence to the sequence address.
// TODO(Subv): Find out what happens if you use a long query type but mark it as a short
// query.
memory_manager.Write<u32>(sequence_address, sequence);
} else {
// Write the 128-bit result structure in long mode. Note: We emulate an infinitely fast
// GPU, this command may actually take a while to complete in real hardware due to GPU
// wait queues.
LongQueryResult query_result{};
query_result.value = result;
// TODO(Subv): Generate a real GPU timestamp and write it here instead of CoreTiming
query_result.timestamp = system.CoreTiming().GetTicks();
memory_manager.WriteBlock(sequence_address, &query_result, sizeof(query_result));
}
break; break;
} }
default:
UNIMPLEMENTED_MSG("Query mode {} not implemented",
static_cast<u32>(regs.query.query_get.mode.Value()));
} }
} }

View file

@ -71,12 +71,11 @@ public:
static constexpr std::size_t MaxConstBuffers = 18; static constexpr std::size_t MaxConstBuffers = 18;
static constexpr std::size_t MaxConstBufferSize = 0x10000; static constexpr std::size_t MaxConstBufferSize = 0x10000;
enum class QueryMode : u32 { enum class QueryOperation : u32 {
Write = 0, Release = 0,
Sync = 1, Acquire = 1,
// TODO(Subv): It is currently unknown what the difference between method 2 and method 0 Counter = 2,
// is. Trap = 3,
Write2 = 2,
}; };
enum class QueryUnit : u32 { enum class QueryUnit : u32 {
@ -1077,7 +1076,7 @@ public:
u32 query_sequence; u32 query_sequence;
union { union {
u32 raw; u32 raw;
BitField<0, 2, QueryMode> mode; BitField<0, 2, QueryOperation> operation;
BitField<4, 1, u32> fence; BitField<4, 1, u32> fence;
BitField<12, 4, QueryUnit> unit; BitField<12, 4, QueryUnit> unit;
BitField<16, 1, QuerySyncCondition> sync_cond; BitField<16, 1, QuerySyncCondition> sync_cond;
@ -1409,6 +1408,9 @@ private:
/// Handles a write to the QUERY_GET register. /// Handles a write to the QUERY_GET register.
void ProcessQueryGet(); void ProcessQueryGet();
// Writes the query result accordingly
void StampQueryResult(u64 payload, bool long_query);
// Handles Conditional Rendering // Handles Conditional Rendering
void ProcessQueryCondition(); void ProcessQueryCondition();