WIP tilemanager

This commit is contained in:
georgemoralis 2023-09-28 18:17:39 +03:00
parent 56a7398fa6
commit 23eeec09d0
4 changed files with 195 additions and 3 deletions

View File

@ -54,7 +54,8 @@ add_executable(shadps4
src/Core/PS4/HLE/Kernel/event_queues.h
src/Core/PS4/HLE/Kernel/cpu_management.cpp
src/Core/PS4/HLE/Kernel/cpu_management.h
"src/Util/Singleton.h" "src/Util/Disassembler.cpp" "src/Util/Disassembler.h" "src/Core/PS4/Util/aerolib.h" "src/Core/PS4/Loader/SymbolsResolver.h" "src/Core/PS4/Loader/SymbolsResolver.cpp" "src/Core/PS4/HLE/Libs.cpp" "src/Core/PS4/HLE/Libs.h" "src/Core/PS4/HLE/LibC.cpp" "src/Core/PS4/HLE/LibC.h" "src/Lib/Timer.cpp" "src/Lib/Timer.h" "src/Core/PS4/HLE/LibKernel.cpp" "src/Core/PS4/HLE/LibKernel.h" "src/Core/PS4/HLE/LibSceGnmDriver.cpp" "src/Core/PS4/HLE/LibSceGnmDriver.h" "src/Core/PS4/HLE/Kernel/ThreadManagement.cpp" "src/Core/PS4/HLE/Kernel/ThreadManagement.h" "src/Core/PS4/HLE/ErrorCodes.h" "src/debug.h" "src/Core/PS4/HLE/Kernel/memory_management.cpp" "src/Core/PS4/HLE/Kernel/memory_management.h" "src/Core/PS4/GPU/gpu_memory.cpp" "src/Core/PS4/GPU/gpu_memory.h" "src/emulator.cpp" "src/emulator.h" "src/Core/PS4/HLE/Kernel/Objects/event_queue.h" "src/Core/PS4/HLE/Kernel/Objects/event_queue.cpp" "src/Core/PS4/HLE/Graphics/Objects/video_out_ctx.cpp" "src/Core/PS4/HLE/Graphics/Objects/video_out_ctx.h" "src/Core/PS4/HLE/Graphics/graphics_ctx.h" "src/vulkan_util.cpp" "src/vulkan_util.h" "src/Core/PS4/GPU/video_out_buffer.cpp" "src/Core/PS4/GPU/video_out_buffer.h" "src/Core/PS4/HLE/Graphics/graphics_render.cpp" "src/Core/PS4/HLE/Graphics/graphics_render.h")
"src/Util/Singleton.h" "src/Util/Disassembler.cpp" "src/Util/Disassembler.h" "src/Core/PS4/Util/aerolib.h" "src/Core/PS4/Loader/SymbolsResolver.h" "src/Core/PS4/Loader/SymbolsResolver.cpp" "src/Core/PS4/HLE/Libs.cpp" "src/Core/PS4/HLE/Libs.h" "src/Core/PS4/HLE/LibC.cpp" "src/Core/PS4/HLE/LibC.h" "src/Lib/Timer.cpp" "src/Lib/Timer.h" "src/Core/PS4/HLE/LibKernel.cpp" "src/Core/PS4/HLE/LibKernel.h" "src/Core/PS4/HLE/LibSceGnmDriver.cpp" "src/Core/PS4/HLE/LibSceGnmDriver.h" "src/Core/PS4/HLE/Kernel/ThreadManagement.cpp" "src/Core/PS4/HLE/Kernel/ThreadManagement.h" "src/Core/PS4/HLE/ErrorCodes.h" "src/debug.h" "src/Core/PS4/HLE/Kernel/memory_management.cpp" "src/Core/PS4/HLE/Kernel/memory_management.h" "src/Core/PS4/GPU/gpu_memory.cpp" "src/Core/PS4/GPU/gpu_memory.h" "src/emulator.cpp" "src/emulator.h" "src/Core/PS4/HLE/Kernel/Objects/event_queue.h" "src/Core/PS4/HLE/Kernel/Objects/event_queue.cpp" "src/Core/PS4/HLE/Graphics/Objects/video_out_ctx.cpp" "src/Core/PS4/HLE/Graphics/Objects/video_out_ctx.h" "src/Core/PS4/HLE/Graphics/graphics_ctx.h" "src/vulkan_util.cpp" "src/vulkan_util.h" "src/Core/PS4/GPU/video_out_buffer.cpp" "src/Core/PS4/GPU/video_out_buffer.h" "src/Core/PS4/HLE/Graphics/graphics_render.cpp" "src/Core/PS4/HLE/Graphics/graphics_render.h" "src/Core/PS4/GPU/tile_manager.cpp" "src/Core/PS4/GPU/tile_manager.h")
find_package(OpenGL REQUIRED)
target_link_libraries(shadps4 PUBLIC fmt mincore spdlog IMGUI SDL3-shared ${OPENGL_LIBRARY} vulkan-1 spirv-tools-opt spirv-tools)

View File

@ -0,0 +1,169 @@
#include "tile_manager.h"
#include "Lib/Threads.h"
#include "Util/Singleton.h"
namespace GPU {
static u32 IntLog2(u32 i) { return 31 - __builtin_clz(i | 1u); }
class TileManager {
public:
TileManager(){}
virtual ~TileManager() { }
Lib::Mutex m_mutex;
};
class TileManager32 {
public:
u32 m_macro_tile_height = 0;
u32 m_bank_height = 0;
u32 m_num_banks = 0;
u32 m_num_pipes = 0;
u32 m_padded_width = 0;
u32 m_padded_height = 0;
u32 m_pipe_bits = 0;
u32 m_bank_bits = 0;
void Init(u32 width, u32 height, bool is_neo) {
m_macro_tile_height = (is_neo ? 128 : 64);
m_bank_height = is_neo ? 2 : 1;
m_num_banks = is_neo ? 8 : 16;
m_num_pipes = is_neo ? 16 : 8;
m_padded_width = width;
if (height == 1080) {
m_padded_height = is_neo ? 1152 : 1088;
}
if (height == 720) {
m_padded_height = 768;
}
m_pipe_bits = is_neo ? 4 : 3;
m_bank_bits = is_neo ? 3 : 4;
}
static u32 getElementIdx(u32 x, u32 y) {
u32 elem = 0;
elem |= ((x >> 0u) & 0x1u) << 0u;
elem |= ((x >> 1u) & 0x1u) << 1u;
elem |= ((y >> 0u) & 0x1u) << 2u;
elem |= ((x >> 2u) & 0x1u) << 3u;
elem |= ((y >> 1u) & 0x1u) << 4u;
elem |= ((y >> 2u) & 0x1u) << 5u;
return elem;
}
static u32 getPipeIdx(u32 x, u32 y, bool is_neo) {
u32 pipe = 0;
if (!is_neo) {
pipe |= (((x >> 3u) ^ (y >> 3u) ^ (x >> 4u)) & 0x1u) << 0u;
pipe |= (((x >> 4u) ^ (y >> 4u)) & 0x1u) << 1u;
pipe |= (((x >> 5u) ^ (y >> 5u)) & 0x1u) << 2u;
} else {
pipe |= (((x >> 3u) ^ (y >> 3u) ^ (x >> 4u)) & 0x1u) << 0u;
pipe |= (((x >> 4u) ^ (y >> 4u)) & 0x1u) << 1u;
pipe |= (((x >> 5u) ^ (y >> 5u)) & 0x1u) << 2u;
pipe |= (((x >> 6u) ^ (y >> 5u)) & 0x1u) << 3u;
}
return pipe;
}
static u32 getBankIdx(u32 x, u32 y, u32 bank_width, u32 bank_height, u32 num_banks, u32 num_pipes) {
const u32 x_shift_offset = IntLog2(bank_width * num_pipes);
const u32 y_shift_offset = IntLog2(bank_height);
const u32 xs = x >> x_shift_offset;
const u32 ys = y >> y_shift_offset;
u32 bank = 0;
switch (num_banks) {
case 8:
bank |= (((xs >> 3u) ^ (ys >> 5u)) & 0x1u) << 0u;
bank |= (((xs >> 4u) ^ (ys >> 4u) ^ (ys >> 5u)) & 0x1u) << 1u;
bank |= (((xs >> 5u) ^ (ys >> 3u)) & 0x1u) << 2u;
break;
case 16:
bank |= (((xs >> 3u) ^ (ys >> 6u)) & 0x1u) << 0u;
bank |= (((xs >> 4u) ^ (ys >> 5u) ^ (ys >> 6u)) & 0x1u) << 1u;
bank |= (((xs >> 5u) ^ (ys >> 4u)) & 0x1u) << 2u;
bank |= (((xs >> 6u) ^ (ys >> 3u)) & 0x1u) << 3u;
break;
default:;
}
return bank;
}
u64 getTiledOffs(u32 x, u32 y, bool is_neo) const {
u64 element_index = getElementIdx(x, y);
u32 xh = x;
u32 yh = y;
u64 pipe = getPipeIdx(xh, yh, is_neo);
u64 bank = getBankIdx(xh, yh, 1, m_bank_height, m_num_banks, m_num_pipes);
u32 tile_bytes = (8 * 8 * 32 + 7) / 8;
u64 element_offset = (element_index * 32);
u64 tile_split_slice = 0;
if (tile_bytes > 512) {
tile_split_slice = element_offset / (static_cast<u64>(512) * 8);
element_offset %= (static_cast<u64>(512) * 8);
tile_bytes = 512;
}
u64 macro_tile_bytes = (128 / 8) * (m_macro_tile_height / 8) * tile_bytes / (m_num_pipes * m_num_banks);
u64 macro_tiles_per_row = m_padded_width / 128;
u64 macro_tile_row_index = y / m_macro_tile_height;
u64 macro_tile_column_index = x / 128;
u64 macro_tile_index = (macro_tile_row_index * macro_tiles_per_row) + macro_tile_column_index;
u64 macro_tile_offset = macro_tile_index * macro_tile_bytes;
u64 macro_tiles_per_slice = macro_tiles_per_row * (m_padded_height / m_macro_tile_height);
u64 slice_bytes = macro_tiles_per_slice * macro_tile_bytes;
u64 slice_offset = tile_split_slice * slice_bytes;
u64 tile_row_index = (y / 8) % m_bank_height;
u64 tile_index = tile_row_index;
u64 tile_offset = tile_index * tile_bytes;
u64 tile_split_slice_rotation = ((m_num_banks / 2) + 1) * tile_split_slice;
bank ^= tile_split_slice_rotation;
bank &= (m_num_banks - 1);
u64 total_offset = (slice_offset + macro_tile_offset + tile_offset) * 8 + element_offset;
u64 bit_offset = total_offset & 0x7u;
total_offset /= 8;
u64 pipe_interleave_offset = total_offset & 0xffu;
u64 offset = total_offset >> 8u;
u64 byte_offset = pipe_interleave_offset | (pipe << (8u)) | (bank << (8u + m_pipe_bits)) | (offset << (8u + m_pipe_bits + m_bank_bits));
return ((byte_offset << 3u) | bit_offset) / 8;
}
};
void convertTileToLinear(void* dst, const void* src,u32 width, u32 height, bool is_neo) {
TileManager32 t;
t.Init(width, height, is_neo);
auto* g_TileManager = Singleton<TileManager>::Instance();
Lib::LockMutexGuard lock(g_TileManager->m_mutex);
for (u32 y = 0; y < height; y++) {
u32 x = 0;
u64 linear_offset = y * width * 4;
for (; x + 1 < width; x += 2) {
auto tiled_offset = t.getTiledOffs(x, y, is_neo);
*reinterpret_cast<u64*>(static_cast<u08*>(dst) + linear_offset) =
*reinterpret_cast<const u64*>(static_cast<const u08*>(src) + tiled_offset);
linear_offset += 8;
}
if (x < width) {
auto tiled_offset = t.getTiledOffs(x, y, is_neo);
*reinterpret_cast<u32*>(static_cast<u08*>(dst) + linear_offset) =
*reinterpret_cast<const u32*>(static_cast<const u08*>(src) + tiled_offset);
}
}
}
} // namespace GPU

View File

@ -0,0 +1,8 @@
#pragma once
#include "types.h"
namespace GPU {
void convertTileToLinear(void* dst, const void* src, u32 width, u32 height, bool neo);
}

View File

@ -4,6 +4,7 @@
#include "debug.h"
#include <vulkan_util.h>
#include "tile_manager.h"
constexpr bool log_file_videoOutBuffer = true; // disable it to disable logging
@ -11,13 +12,26 @@ static void update_func(HLE::Libs::Graphics::GraphicCtx* ctx, const u64* params,
int virtual_addr_num) {
auto pitch = params[GPU::VideoOutBufferObj::PITCH_PARAM];
bool tiled = (params[GPU::VideoOutBufferObj::IS_TILE_PARAM] != 0);
bool neo = (params[GPU::VideoOutBufferObj::IS_NEO_PARAM] != 0);
auto width = params[GPU::VideoOutBufferObj::WIDTH_PARAM];
auto height = params[GPU::VideoOutBufferObj::HEIGHT_PARAM];
auto* vk_obj = static_cast<HLE::Libs::Graphics::VideoOutVulkanImage*>(obj);
vk_obj->layout = VK_IMAGE_LAYOUT_UNDEFINED;
Graphics::Vulkan::vulkanFillImage(ctx, vk_obj, reinterpret_cast<void*>(*virtual_addr), *size, pitch,
static_cast<uint64_t>(VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL));
if (tiled)
{
auto* tempbuff = new u08[*size];
GPU::convertTileToLinear(tempbuff, reinterpret_cast<void*>(*virtual_addr), width, height, neo);
Graphics::Vulkan::vulkanFillImage(ctx, vk_obj, tempbuff, *size, pitch, static_cast<uint64_t>(VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL));
delete[] tempbuff;
} else {
Graphics::Vulkan::vulkanFillImage(ctx, vk_obj, reinterpret_cast<void*>(*virtual_addr), *size, pitch,
static_cast<uint64_t>(VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL));
}
}
static void* create_func(HLE::Libs::Graphics::GraphicCtx* ctx, const u64* params, const u64* virtual_addr, const u64* size, int virtual_addr_num,