VideoCore: implement channels on gpu caches.

This commit is contained in:
Fernando Sahmkow 2021-11-05 15:52:31 +01:00
parent 5ca7d7118f
commit 8847b6645c
50 changed files with 1469 additions and 817 deletions

View file

@ -18,6 +18,8 @@
#include "core/hle/service/nvdrv/nvdata.h"
#include "core/perf_stats.h"
#include "video_core/cdma_pusher.h"
#include "video_core/control/channel_state.h"
#include "video_core/control/scheduler.h"
#include "video_core/dma_pusher.h"
#include "video_core/engines/fermi_2d.h"
#include "video_core/engines/kepler_compute.h"
@ -36,65 +38,58 @@ MICROPROFILE_DEFINE(GPU_wait, "GPU", "Wait for the GPU", MP_RGB(128, 128, 192));
struct GPU::Impl {
explicit Impl(GPU& gpu_, Core::System& system_, bool is_async_, bool use_nvdec_)
: gpu{gpu_}, system{system_}, memory_manager{std::make_unique<Tegra::MemoryManager>(
system)},
dma_pusher{std::make_unique<Tegra::DmaPusher>(system, gpu)}, use_nvdec{use_nvdec_},
maxwell_3d{std::make_unique<Engines::Maxwell3D>(system, *memory_manager)},
fermi_2d{std::make_unique<Engines::Fermi2D>()},
kepler_compute{std::make_unique<Engines::KeplerCompute>(system, *memory_manager)},
maxwell_dma{std::make_unique<Engines::MaxwellDMA>(system, *memory_manager)},
kepler_memory{std::make_unique<Engines::KeplerMemory>(system, *memory_manager)},
: gpu{gpu_}, system{system_}, use_nvdec{use_nvdec_},
shader_notify{std::make_unique<VideoCore::ShaderNotify>()}, is_async{is_async_},
gpu_thread{system_, is_async_} {}
gpu_thread{system_, is_async_}, scheduler{std::make_unique<Control::Scheduler>(gpu)} {}
~Impl() = default;
std::shared_ptr<Control::ChannelState> CreateChannel(s32 channel_id) {
auto channel_state = std::make_shared<Tegra::Control::ChannelState>(channel_id);
channels.emplace(channel_id, channel_state);
scheduler->DeclareChannel(channel_state);
return channel_state;
}
void BindChannel(s32 channel_id) {
if (bound_channel == channel_id) {
return;
}
auto it = channels.find(channel_id);
ASSERT(it != channels.end());
bound_channel = channel_id;
current_channel = it->second.get();
rasterizer->BindChannel(*current_channel);
}
std::shared_ptr<Control::ChannelState> AllocateChannel() {
return CreateChannel(new_channel_id++);
}
void InitChannel(Control::ChannelState& to_init) {
to_init.Init(system, gpu);
to_init.BindRasterizer(rasterizer);
rasterizer->InitializeChannel(to_init);
}
void ReleaseChannel(Control::ChannelState& to_release) {
UNIMPLEMENTED();
}
void CreateHost1xChannel() {
if (host1x_channel) {
return;
}
host1x_channel = CreateChannel(0);
host1x_channel->memory_manager = std::make_shared<Tegra::MemoryManager>(system);
InitChannel(*host1x_channel);
}
/// Binds a renderer to the GPU.
void BindRenderer(std::unique_ptr<VideoCore::RendererBase> renderer_) {
renderer = std::move(renderer_);
rasterizer = renderer->ReadRasterizer();
memory_manager->BindRasterizer(rasterizer);
maxwell_3d->BindRasterizer(rasterizer);
fermi_2d->BindRasterizer(rasterizer);
kepler_compute->BindRasterizer(rasterizer);
kepler_memory->BindRasterizer(rasterizer);
maxwell_dma->BindRasterizer(rasterizer);
}
/// Calls a GPU method.
void CallMethod(const GPU::MethodCall& method_call) {
LOG_TRACE(HW_GPU, "Processing method {:08X} on subchannel {}", method_call.method,
method_call.subchannel);
ASSERT(method_call.subchannel < bound_engines.size());
if (ExecuteMethodOnEngine(method_call.method)) {
CallEngineMethod(method_call);
} else {
CallPullerMethod(method_call);
}
}
/// Calls a GPU multivalue method.
void CallMultiMethod(u32 method, u32 subchannel, const u32* base_start, u32 amount,
u32 methods_pending) {
LOG_TRACE(HW_GPU, "Processing method {:08X} on subchannel {}", method, subchannel);
ASSERT(subchannel < bound_engines.size());
if (ExecuteMethodOnEngine(method)) {
CallEngineMultiMethod(method, subchannel, base_start, amount, methods_pending);
} else {
for (std::size_t i = 0; i < amount; i++) {
CallPullerMethod(GPU::MethodCall{
method,
base_start[i],
subchannel,
methods_pending - static_cast<u32>(i),
});
}
}
}
/// Flush all current written commands into the host GPU for execution.
@ -146,42 +141,44 @@ struct GPU::Impl {
/// Returns a reference to the Maxwell3D GPU engine.
[[nodiscard]] Engines::Maxwell3D& Maxwell3D() {
return *maxwell_3d;
ASSERT(current_channel);
return *current_channel->maxwell_3d;
}
/// Returns a const reference to the Maxwell3D GPU engine.
[[nodiscard]] const Engines::Maxwell3D& Maxwell3D() const {
return *maxwell_3d;
ASSERT(current_channel);
return *current_channel->maxwell_3d;
}
/// Returns a reference to the KeplerCompute GPU engine.
[[nodiscard]] Engines::KeplerCompute& KeplerCompute() {
return *kepler_compute;
ASSERT(current_channel);
return *current_channel->kepler_compute;
}
/// Returns a reference to the KeplerCompute GPU engine.
[[nodiscard]] const Engines::KeplerCompute& KeplerCompute() const {
return *kepler_compute;
ASSERT(current_channel);
return *current_channel->kepler_compute;
}
/// Returns a reference to the GPU memory manager.
[[nodiscard]] Tegra::MemoryManager& MemoryManager() {
return *memory_manager;
}
/// Returns a const reference to the GPU memory manager.
[[nodiscard]] const Tegra::MemoryManager& MemoryManager() const {
return *memory_manager;
CreateHost1xChannel();
return *host1x_channel->memory_manager;
}
/// Returns a reference to the GPU DMA pusher.
[[nodiscard]] Tegra::DmaPusher& DmaPusher() {
return *dma_pusher;
ASSERT(current_channel);
return *current_channel->dma_pusher;
}
/// Returns a const reference to the GPU DMA pusher.
[[nodiscard]] const Tegra::DmaPusher& DmaPusher() const {
return *dma_pusher;
ASSERT(current_channel);
return *current_channel->dma_pusher;
}
/// Returns a reference to the underlying renderer.
@ -306,7 +303,7 @@ struct GPU::Impl {
/// This can be used to launch any necessary threads and register any necessary
/// core timing events.
void Start() {
gpu_thread.StartThread(*renderer, renderer->Context(), *dma_pusher);
gpu_thread.StartThread(*renderer, renderer->Context(), *scheduler);
cpu_context = renderer->GetRenderWindow().CreateSharedContext();
cpu_context->MakeCurrent();
}
@ -328,8 +325,8 @@ struct GPU::Impl {
}
/// Push GPU command entries to be processed
void PushGPUEntries(Tegra::CommandList&& entries) {
gpu_thread.SubmitList(std::move(entries));
void PushGPUEntries(s32 channel, Tegra::CommandList&& entries) {
gpu_thread.SubmitList(channel, std::move(entries));
}
/// Push GPU command buffer entries to be processed
@ -381,303 +378,16 @@ struct GPU::Impl {
interrupt_manager.GPUInterruptSyncpt(syncpoint_id, value);
}
void ProcessBindMethod(const GPU::MethodCall& method_call) {
// Bind the current subchannel to the desired engine id.
LOG_DEBUG(HW_GPU, "Binding subchannel {} to engine {}", method_call.subchannel,
method_call.argument);
const auto engine_id = static_cast<EngineID>(method_call.argument);
bound_engines[method_call.subchannel] = static_cast<EngineID>(engine_id);
switch (engine_id) {
case EngineID::FERMI_TWOD_A:
dma_pusher->BindSubchannel(fermi_2d.get(), method_call.subchannel);
break;
case EngineID::MAXWELL_B:
dma_pusher->BindSubchannel(maxwell_3d.get(), method_call.subchannel);
break;
case EngineID::KEPLER_COMPUTE_B:
dma_pusher->BindSubchannel(kepler_compute.get(), method_call.subchannel);
break;
case EngineID::MAXWELL_DMA_COPY_A:
dma_pusher->BindSubchannel(maxwell_dma.get(), method_call.subchannel);
break;
case EngineID::KEPLER_INLINE_TO_MEMORY_B:
dma_pusher->BindSubchannel(kepler_memory.get(), method_call.subchannel);
break;
default:
UNIMPLEMENTED_MSG("Unimplemented engine {:04X}", engine_id);
}
}
void ProcessFenceActionMethod() {
switch (regs.fence_action.op) {
case GPU::FenceOperation::Acquire:
WaitFence(regs.fence_action.syncpoint_id, regs.fence_value);
break;
case GPU::FenceOperation::Increment:
IncrementSyncPoint(regs.fence_action.syncpoint_id);
break;
default:
UNIMPLEMENTED_MSG("Unimplemented operation {}", regs.fence_action.op.Value());
}
}
void ProcessWaitForInterruptMethod() {
// TODO(bunnei) ImplementMe
LOG_WARNING(HW_GPU, "(STUBBED) called");
}
void ProcessSemaphoreTriggerMethod() {
const auto semaphoreOperationMask = 0xF;
const auto op =
static_cast<GpuSemaphoreOperation>(regs.semaphore_trigger & semaphoreOperationMask);
if (op == GpuSemaphoreOperation::WriteLong) {
struct Block {
u32 sequence;
u32 zeros = 0;
u64 timestamp;
};
Block block{};
block.sequence = regs.semaphore_sequence;
// TODO(Kmather73): Generate a real GPU timestamp and write it here instead of
// CoreTiming
block.timestamp = GetTicks();
memory_manager->WriteBlock(regs.semaphore_address.SemaphoreAddress(), &block,
sizeof(block));
} else {
const u32 word{memory_manager->Read<u32>(regs.semaphore_address.SemaphoreAddress())};
if ((op == GpuSemaphoreOperation::AcquireEqual && word == regs.semaphore_sequence) ||
(op == GpuSemaphoreOperation::AcquireGequal &&
static_cast<s32>(word - regs.semaphore_sequence) > 0) ||
(op == GpuSemaphoreOperation::AcquireMask && (word & regs.semaphore_sequence))) {
// Nothing to do in this case
} else {
regs.acquire_source = true;
regs.acquire_value = regs.semaphore_sequence;
if (op == GpuSemaphoreOperation::AcquireEqual) {
regs.acquire_active = true;
regs.acquire_mode = false;
} else if (op == GpuSemaphoreOperation::AcquireGequal) {
regs.acquire_active = true;
regs.acquire_mode = true;
} else if (op == GpuSemaphoreOperation::AcquireMask) {
// TODO(kemathe) The acquire mask operation waits for a value that, ANDed with
// semaphore_sequence, gives a non-0 result
LOG_ERROR(HW_GPU, "Invalid semaphore operation AcquireMask not implemented");
} else {
LOG_ERROR(HW_GPU, "Invalid semaphore operation");
}
}
}
}
void ProcessSemaphoreRelease() {
memory_manager->Write<u32>(regs.semaphore_address.SemaphoreAddress(),
regs.semaphore_release);
}
void ProcessSemaphoreAcquire() {
const u32 word = memory_manager->Read<u32>(regs.semaphore_address.SemaphoreAddress());
const auto value = regs.semaphore_acquire;
if (word != value) {
regs.acquire_active = true;
regs.acquire_value = value;
// TODO(kemathe73) figure out how to do the acquire_timeout
regs.acquire_mode = false;
regs.acquire_source = false;
}
}
/// Calls a GPU puller method.
void CallPullerMethod(const GPU::MethodCall& method_call) {
regs.reg_array[method_call.method] = method_call.argument;
const auto method = static_cast<BufferMethods>(method_call.method);
switch (method) {
case BufferMethods::BindObject: {
ProcessBindMethod(method_call);
break;
}
case BufferMethods::Nop:
case BufferMethods::SemaphoreAddressHigh:
case BufferMethods::SemaphoreAddressLow:
case BufferMethods::SemaphoreSequence:
break;
case BufferMethods::UnkCacheFlush:
rasterizer->SyncGuestHost();
break;
case BufferMethods::WrcacheFlush:
rasterizer->SignalReference();
break;
case BufferMethods::FenceValue:
break;
case BufferMethods::RefCnt:
rasterizer->SignalReference();
break;
case BufferMethods::FenceAction:
ProcessFenceActionMethod();
break;
case BufferMethods::WaitForInterrupt:
rasterizer->WaitForIdle();
break;
case BufferMethods::SemaphoreTrigger: {
ProcessSemaphoreTriggerMethod();
break;
}
case BufferMethods::NotifyIntr: {
// TODO(Kmather73): Research and implement this method.
LOG_ERROR(HW_GPU, "Special puller engine method NotifyIntr not implemented");
break;
}
case BufferMethods::Unk28: {
// TODO(Kmather73): Research and implement this method.
LOG_ERROR(HW_GPU, "Special puller engine method Unk28 not implemented");
break;
}
case BufferMethods::SemaphoreAcquire: {
ProcessSemaphoreAcquire();
break;
}
case BufferMethods::SemaphoreRelease: {
ProcessSemaphoreRelease();
break;
}
case BufferMethods::Yield: {
// TODO(Kmather73): Research and implement this method.
LOG_ERROR(HW_GPU, "Special puller engine method Yield not implemented");
break;
}
default:
LOG_ERROR(HW_GPU, "Special puller engine method {:X} not implemented", method);
break;
}
}
/// Calls a GPU engine method.
void CallEngineMethod(const GPU::MethodCall& method_call) {
const EngineID engine = bound_engines[method_call.subchannel];
switch (engine) {
case EngineID::FERMI_TWOD_A:
fermi_2d->CallMethod(method_call.method, method_call.argument,
method_call.IsLastCall());
break;
case EngineID::MAXWELL_B:
maxwell_3d->CallMethod(method_call.method, method_call.argument,
method_call.IsLastCall());
break;
case EngineID::KEPLER_COMPUTE_B:
kepler_compute->CallMethod(method_call.method, method_call.argument,
method_call.IsLastCall());
break;
case EngineID::MAXWELL_DMA_COPY_A:
maxwell_dma->CallMethod(method_call.method, method_call.argument,
method_call.IsLastCall());
break;
case EngineID::KEPLER_INLINE_TO_MEMORY_B:
kepler_memory->CallMethod(method_call.method, method_call.argument,
method_call.IsLastCall());
break;
default:
UNIMPLEMENTED_MSG("Unimplemented engine");
}
}
/// Calls a GPU engine multivalue method.
void CallEngineMultiMethod(u32 method, u32 subchannel, const u32* base_start, u32 amount,
u32 methods_pending) {
const EngineID engine = bound_engines[subchannel];
switch (engine) {
case EngineID::FERMI_TWOD_A:
fermi_2d->CallMultiMethod(method, base_start, amount, methods_pending);
break;
case EngineID::MAXWELL_B:
maxwell_3d->CallMultiMethod(method, base_start, amount, methods_pending);
break;
case EngineID::KEPLER_COMPUTE_B:
kepler_compute->CallMultiMethod(method, base_start, amount, methods_pending);
break;
case EngineID::MAXWELL_DMA_COPY_A:
maxwell_dma->CallMultiMethod(method, base_start, amount, methods_pending);
break;
case EngineID::KEPLER_INLINE_TO_MEMORY_B:
kepler_memory->CallMultiMethod(method, base_start, amount, methods_pending);
break;
default:
UNIMPLEMENTED_MSG("Unimplemented engine");
}
}
/// Determines where the method should be executed.
[[nodiscard]] bool ExecuteMethodOnEngine(u32 method) {
const auto buffer_method = static_cast<BufferMethods>(method);
return buffer_method >= BufferMethods::NonPullerMethods;
}
struct Regs {
static constexpr size_t NUM_REGS = 0x40;
union {
struct {
INSERT_PADDING_WORDS_NOINIT(0x4);
struct {
u32 address_high;
u32 address_low;
[[nodiscard]] GPUVAddr SemaphoreAddress() const {
return static_cast<GPUVAddr>((static_cast<GPUVAddr>(address_high) << 32) |
address_low);
}
} semaphore_address;
u32 semaphore_sequence;
u32 semaphore_trigger;
INSERT_PADDING_WORDS_NOINIT(0xC);
// The pusher and the puller share the reference counter, the pusher only has read
// access
u32 reference_count;
INSERT_PADDING_WORDS_NOINIT(0x5);
u32 semaphore_acquire;
u32 semaphore_release;
u32 fence_value;
GPU::FenceAction fence_action;
INSERT_PADDING_WORDS_NOINIT(0xE2);
// Puller state
u32 acquire_mode;
u32 acquire_source;
u32 acquire_active;
u32 acquire_timeout;
u32 acquire_value;
};
std::array<u32, NUM_REGS> reg_array;
};
} regs{};
GPU& gpu;
Core::System& system;
std::unique_ptr<Tegra::MemoryManager> memory_manager;
std::unique_ptr<Tegra::DmaPusher> dma_pusher;
std::map<u32, std::unique_ptr<Tegra::CDmaPusher>> cdma_pushers;
std::unique_ptr<VideoCore::RendererBase> renderer;
VideoCore::RasterizerInterface* rasterizer = nullptr;
const bool use_nvdec;
/// Mapping of command subchannels to their bound engine ids
std::array<EngineID, 8> bound_engines{};
/// 3D engine
std::unique_ptr<Engines::Maxwell3D> maxwell_3d;
/// 2D engine
std::unique_ptr<Engines::Fermi2D> fermi_2d;
/// Compute engine
std::unique_ptr<Engines::KeplerCompute> kepler_compute;
/// DMA engine
std::unique_ptr<Engines::MaxwellDMA> maxwell_dma;
/// Inline memory engine
std::unique_ptr<Engines::KeplerMemory> kepler_memory;
std::shared_ptr<Control::ChannelState> host1x_channel;
s32 new_channel_id{1};
/// Shader build notifier
std::unique_ptr<VideoCore::ShaderNotify> shader_notify;
/// When true, we are about to shut down emulation session, so terminate outstanding tasks
@ -710,33 +420,10 @@ struct GPU::Impl {
VideoCommon::GPUThread::ThreadManager gpu_thread;
std::unique_ptr<Core::Frontend::GraphicsContext> cpu_context;
#define ASSERT_REG_POSITION(field_name, position) \
static_assert(offsetof(Regs, field_name) == position * 4, \
"Field " #field_name " has invalid position")
ASSERT_REG_POSITION(semaphore_address, 0x4);
ASSERT_REG_POSITION(semaphore_sequence, 0x6);
ASSERT_REG_POSITION(semaphore_trigger, 0x7);
ASSERT_REG_POSITION(reference_count, 0x14);
ASSERT_REG_POSITION(semaphore_acquire, 0x1A);
ASSERT_REG_POSITION(semaphore_release, 0x1B);
ASSERT_REG_POSITION(fence_value, 0x1C);
ASSERT_REG_POSITION(fence_action, 0x1D);
ASSERT_REG_POSITION(acquire_mode, 0x100);
ASSERT_REG_POSITION(acquire_source, 0x101);
ASSERT_REG_POSITION(acquire_active, 0x102);
ASSERT_REG_POSITION(acquire_timeout, 0x103);
ASSERT_REG_POSITION(acquire_value, 0x104);
#undef ASSERT_REG_POSITION
enum class GpuSemaphoreOperation {
AcquireEqual = 0x1,
WriteLong = 0x2,
AcquireGequal = 0x4,
AcquireMask = 0x8,
};
std::unique_ptr<Tegra::Control::Scheduler> scheduler;
std::unordered_map<s32, std::shared_ptr<Tegra::Control::ChannelState>> channels;
Tegra::Control::ChannelState* current_channel;
s32 bound_channel{-1};
};
GPU::GPU(Core::System& system, bool is_async, bool use_nvdec)
@ -744,19 +431,26 @@ GPU::GPU(Core::System& system, bool is_async, bool use_nvdec)
GPU::~GPU() = default;
std::shared_ptr<Control::ChannelState> GPU::AllocateChannel() {
return impl->AllocateChannel();
}
void GPU::InitChannel(Control::ChannelState& to_init) {
impl->InitChannel(to_init);
}
void GPU::BindChannel(s32 channel_id) {
impl->BindChannel(channel_id);
}
void GPU::ReleaseChannel(Control::ChannelState& to_release) {
impl->ReleaseChannel(to_release);
}
void GPU::BindRenderer(std::unique_ptr<VideoCore::RendererBase> renderer) {
impl->BindRenderer(std::move(renderer));
}
void GPU::CallMethod(const MethodCall& method_call) {
impl->CallMethod(method_call);
}
void GPU::CallMultiMethod(u32 method, u32 subchannel, const u32* base_start, u32 amount,
u32 methods_pending) {
impl->CallMultiMethod(method, subchannel, base_start, amount, methods_pending);
}
void GPU::FlushCommands() {
impl->FlushCommands();
}
@ -881,8 +575,8 @@ void GPU::ReleaseContext() {
impl->ReleaseContext();
}
void GPU::PushGPUEntries(Tegra::CommandList&& entries) {
impl->PushGPUEntries(std::move(entries));
void GPU::PushGPUEntries(s32 channel, Tegra::CommandList&& entries) {
impl->PushGPUEntries(channel, std::move(entries));
}
void GPU::PushCommandBuffer(u32 id, Tegra::ChCommandHeaderList& entries) {