VideoCore: implement channels on gpu caches.
This commit is contained in:
Fernando Sahmkow
parent
c77b8df12e
commit
139ea93512
50 changed files with 1469 additions and 817 deletions
297
src/video_core/engines/puller.cpp
Normal file
297
src/video_core/engines/puller.cpp
Normal file
|
|
@ -0,0 +1,297 @@
|
|||
// Copyright 2021 yuzu Emulator Project
|
||||
// Licensed under GPLv2 or any later version
|
||||
// Refer to the license.txt file included.
|
||||
|
||||
#include "common/assert.h"
|
||||
#include "common/logging/log.h"
|
||||
#include "common/settings.h"
|
||||
#include "core/core.h"
|
||||
#include "video_core/control/channel_state.h"
|
||||
#include "video_core/dma_pusher.h"
|
||||
#include "video_core/engines/fermi_2d.h"
|
||||
#include "video_core/engines/kepler_compute.h"
|
||||
#include "video_core/engines/kepler_memory.h"
|
||||
#include "video_core/engines/maxwell_3d.h"
|
||||
#include "video_core/engines/maxwell_dma.h"
|
||||
#include "video_core/engines/puller.h"
|
||||
#include "video_core/gpu.h"
|
||||
#include "video_core/memory_manager.h"
|
||||
#include "video_core/rasterizer_interface.h"
|
||||
|
||||
namespace Tegra::Engines {
|
||||
|
||||
Puller::Puller(GPU& gpu_, MemoryManager& memory_manager_, DmaPusher& dma_pusher_,
|
||||
Control::ChannelState& channel_state_)
|
||||
: gpu{gpu_}, memory_manager{memory_manager_}, dma_pusher{dma_pusher_}, channel_state{
|
||||
channel_state_} {}
|
||||
|
||||
Puller::~Puller() = default;
|
||||
|
||||
void Puller::ProcessBindMethod(const 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(channel_state.fermi_2d.get(), method_call.subchannel);
|
||||
break;
|
||||
case EngineID::MAXWELL_B:
|
||||
dma_pusher.BindSubchannel(channel_state.maxwell_3d.get(), method_call.subchannel);
|
||||
break;
|
||||
case EngineID::KEPLER_COMPUTE_B:
|
||||
dma_pusher.BindSubchannel(channel_state.kepler_compute.get(), method_call.subchannel);
|
||||
break;
|
||||
case EngineID::MAXWELL_DMA_COPY_A:
|
||||
dma_pusher.BindSubchannel(channel_state.maxwell_dma.get(), method_call.subchannel);
|
||||
break;
|
||||
case EngineID::KEPLER_INLINE_TO_MEMORY_B:
|
||||
dma_pusher.BindSubchannel(channel_state.kepler_memory.get(), method_call.subchannel);
|
||||
break;
|
||||
default:
|
||||
UNIMPLEMENTED_MSG("Unimplemented engine {:04X}", engine_id);
|
||||
}
|
||||
}
|
||||
|
||||
void Puller::ProcessFenceActionMethod() {
|
||||
switch (regs.fence_action.op) {
|
||||
case Puller::FenceOperation::Acquire:
|
||||
// UNIMPLEMENTED_MSG("Channel Scheduling pending.");
|
||||
// WaitFence(regs.fence_action.syncpoint_id, regs.fence_value);
|
||||
break;
|
||||
case Puller::FenceOperation::Increment:
|
||||
rasterizer->SignalSyncPoint(regs.fence_action.syncpoint_id);
|
||||
break;
|
||||
default:
|
||||
UNIMPLEMENTED_MSG("Unimplemented operation {}", regs.fence_action.op.Value());
|
||||
}
|
||||
}
|
||||
|
||||
void Puller::ProcessWaitForInterruptMethod() {
|
||||
// TODO(bunnei) ImplementMe
|
||||
LOG_WARNING(HW_GPU, "(STUBBED) called");
|
||||
}
|
||||
|
||||
void Puller::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 = gpu.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 Puller::ProcessSemaphoreRelease() {
|
||||
memory_manager.Write<u32>(regs.semaphore_address.SemaphoreAddress(), regs.semaphore_release);
|
||||
}
|
||||
|
||||
void Puller::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 Puller::CallPullerMethod(const 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:
|
||||
case BufferMethods::UnkCacheFlush:
|
||||
case BufferMethods::WrcacheFlush:
|
||||
case BufferMethods::FenceValue:
|
||||
break;
|
||||
case BufferMethods::RefCnt:
|
||||
rasterizer->SignalReference();
|
||||
break;
|
||||
case BufferMethods::FenceAction:
|
||||
ProcessFenceActionMethod();
|
||||
break;
|
||||
case BufferMethods::WaitForInterrupt:
|
||||
ProcessWaitForInterruptMethod();
|
||||
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 Puller::CallEngineMethod(const MethodCall& method_call) {
|
||||
const EngineID engine = bound_engines[method_call.subchannel];
|
||||
|
||||
switch (engine) {
|
||||
case EngineID::FERMI_TWOD_A:
|
||||
channel_state.fermi_2d->CallMethod(method_call.method, method_call.argument,
|
||||
method_call.IsLastCall());
|
||||
break;
|
||||
case EngineID::MAXWELL_B:
|
||||
channel_state.maxwell_3d->CallMethod(method_call.method, method_call.argument,
|
||||
method_call.IsLastCall());
|
||||
break;
|
||||
case EngineID::KEPLER_COMPUTE_B:
|
||||
channel_state.kepler_compute->CallMethod(method_call.method, method_call.argument,
|
||||
method_call.IsLastCall());
|
||||
break;
|
||||
case EngineID::MAXWELL_DMA_COPY_A:
|
||||
channel_state.maxwell_dma->CallMethod(method_call.method, method_call.argument,
|
||||
method_call.IsLastCall());
|
||||
break;
|
||||
case EngineID::KEPLER_INLINE_TO_MEMORY_B:
|
||||
channel_state.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 Puller::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:
|
||||
channel_state.fermi_2d->CallMultiMethod(method, base_start, amount, methods_pending);
|
||||
break;
|
||||
case EngineID::MAXWELL_B:
|
||||
channel_state.maxwell_3d->CallMultiMethod(method, base_start, amount, methods_pending);
|
||||
break;
|
||||
case EngineID::KEPLER_COMPUTE_B:
|
||||
channel_state.kepler_compute->CallMultiMethod(method, base_start, amount, methods_pending);
|
||||
break;
|
||||
case EngineID::MAXWELL_DMA_COPY_A:
|
||||
channel_state.maxwell_dma->CallMultiMethod(method, base_start, amount, methods_pending);
|
||||
break;
|
||||
case EngineID::KEPLER_INLINE_TO_MEMORY_B:
|
||||
channel_state.kepler_memory->CallMultiMethod(method, base_start, amount, methods_pending);
|
||||
break;
|
||||
default:
|
||||
UNIMPLEMENTED_MSG("Unimplemented engine");
|
||||
}
|
||||
}
|
||||
|
||||
/// Calls a GPU method.
|
||||
void Puller::CallMethod(const 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 Puller::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(MethodCall{
|
||||
method,
|
||||
base_start[i],
|
||||
subchannel,
|
||||
methods_pending - static_cast<u32>(i),
|
||||
});
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
void Puller::BindRasterizer(VideoCore::RasterizerInterface* rasterizer_) {
|
||||
rasterizer = rasterizer_;
|
||||
}
|
||||
|
||||
/// Determines where the method should be executed.
|
||||
[[nodiscard]] bool Puller::ExecuteMethodOnEngine(u32 method) {
|
||||
const auto buffer_method = static_cast<BufferMethods>(method);
|
||||
return buffer_method >= BufferMethods::NonPullerMethods;
|
||||
}
|
||||
|
||||
} // namespace Tegra::Engines
|
||||
179
src/video_core/engines/puller.h
Normal file
179
src/video_core/engines/puller.h
Normal file
|
|
@ -0,0 +1,179 @@
|
|||
// Copyright 2021 yuzu Emulator Project
|
||||
// Licensed under GPLv2 or any later version
|
||||
// Refer to the license.txt file included.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <array>
|
||||
#include <cstddef>
|
||||
#include <vector>
|
||||
#include "common/bit_field.h"
|
||||
#include "common/common_funcs.h"
|
||||
#include "common/common_types.h"
|
||||
#include "video_core/engines/engine_interface.h"
|
||||
|
||||
namespace Core {
|
||||
class System;
|
||||
}
|
||||
|
||||
namespace Tegra {
|
||||
class MemoryManager;
|
||||
class DmaPusher;
|
||||
|
||||
enum class EngineID {
|
||||
FERMI_TWOD_A = 0x902D, // 2D Engine
|
||||
MAXWELL_B = 0xB197, // 3D Engine
|
||||
KEPLER_COMPUTE_B = 0xB1C0,
|
||||
KEPLER_INLINE_TO_MEMORY_B = 0xA140,
|
||||
MAXWELL_DMA_COPY_A = 0xB0B5,
|
||||
};
|
||||
|
||||
namespace Control {
|
||||
struct ChannelState;
|
||||
}
|
||||
} // namespace Tegra
|
||||
|
||||
namespace VideoCore {
|
||||
class RasterizerInterface;
|
||||
}
|
||||
|
||||
namespace Tegra::Engines {
|
||||
|
||||
class Puller final {
|
||||
public:
|
||||
struct MethodCall {
|
||||
u32 method{};
|
||||
u32 argument{};
|
||||
u32 subchannel{};
|
||||
u32 method_count{};
|
||||
|
||||
explicit MethodCall(u32 method_, u32 argument_, u32 subchannel_ = 0, u32 method_count_ = 0)
|
||||
: method(method_), argument(argument_), subchannel(subchannel_),
|
||||
method_count(method_count_) {}
|
||||
|
||||
[[nodiscard]] bool IsLastCall() const {
|
||||
return method_count <= 1;
|
||||
}
|
||||
};
|
||||
|
||||
enum class FenceOperation : u32 {
|
||||
Acquire = 0,
|
||||
Increment = 1,
|
||||
};
|
||||
|
||||
union FenceAction {
|
||||
u32 raw;
|
||||
BitField<0, 1, FenceOperation> op;
|
||||
BitField<8, 24, u32> syncpoint_id;
|
||||
};
|
||||
|
||||
explicit Puller(GPU& gpu_, MemoryManager& memory_manager_, DmaPusher& dma_pusher,
|
||||
Control::ChannelState& channel_state);
|
||||
~Puller();
|
||||
|
||||
void CallMethod(const MethodCall& method_call);
|
||||
|
||||
void CallMultiMethod(u32 method, u32 subchannel, const u32* base_start, u32 amount,
|
||||
u32 methods_pending);
|
||||
|
||||
void BindRasterizer(VideoCore::RasterizerInterface* rasterizer);
|
||||
|
||||
void CallPullerMethod(const MethodCall& method_call);
|
||||
|
||||
void CallEngineMethod(const MethodCall& method_call);
|
||||
|
||||
void CallEngineMultiMethod(u32 method, u32 subchannel, const u32* base_start, u32 amount,
|
||||
u32 methods_pending);
|
||||
|
||||
private:
|
||||
Tegra::GPU& gpu;
|
||||
|
||||
MemoryManager& memory_manager;
|
||||
DmaPusher& dma_pusher;
|
||||
Control::ChannelState& channel_state;
|
||||
VideoCore::RasterizerInterface* rasterizer = nullptr;
|
||||
|
||||
static constexpr std::size_t NUM_REGS = 0x800;
|
||||
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;
|
||||
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{};
|
||||
|
||||
void ProcessBindMethod(const MethodCall& method_call);
|
||||
void ProcessFenceActionMethod();
|
||||
void ProcessSemaphoreAcquire();
|
||||
void ProcessSemaphoreRelease();
|
||||
void ProcessSemaphoreTriggerMethod();
|
||||
void ProcessWaitForInterruptMethod();
|
||||
[[nodiscard]] bool ExecuteMethodOnEngine(u32 method);
|
||||
|
||||
/// Mapping of command subchannels to their bound engine ids
|
||||
std::array<EngineID, 8> bound_engines{};
|
||||
|
||||
enum class GpuSemaphoreOperation {
|
||||
AcquireEqual = 0x1,
|
||||
WriteLong = 0x2,
|
||||
AcquireGequal = 0x4,
|
||||
AcquireMask = 0x8,
|
||||
};
|
||||
|
||||
#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
|
||||
};
|
||||
|
||||
} // namespace Tegra::Engines
|
||||
Loading…
Add table
Add a link
Reference in a new issue