Surface management rework (1/3) (#307)

* amdgpu: proper CB and DB sizes calculation; minor refactoring

* texture_cache: separate file for image_info

* texture_cache: image guest address moved into image info

* texture_cache: surface size calculation

* shader_recompiler: fixed sin/cos

Thanks to red_pring and gandalfthewhite0173

* initial preparations for subresources upload

* review comments
This commit is contained in:
psucien 2024-07-20 11:51:21 +02:00 committed by GitHub
parent 2b52a17845
commit 64459f1a76
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
21 changed files with 467 additions and 233 deletions

View File

@ -453,6 +453,8 @@ set(VIDEO_CORE src/video_core/amdgpu/liverpool.cpp
src/video_core/renderer_vulkan/vk_swapchain.h
src/video_core/texture_cache/image.cpp
src/video_core/texture_cache/image.h
src/video_core/texture_cache/image_info.cpp
src/video_core/texture_cache/image_info.h
src/video_core/texture_cache/image_view.cpp
src/video_core/texture_cache/image_view.h
src/video_core/texture_cache/sampler.cpp

View File

@ -62,7 +62,6 @@ struct BufferAttribute {
struct BufferAttributeGroup {
bool is_occupied;
BufferAttribute attrib;
u32 size_in_bytes;
};
struct VideoOutBuffer {

View File

@ -122,8 +122,6 @@ int VideoOutDriver::RegisterBuffers(VideoOutPort* port, s32 startIndex, void* co
auto& group = port->groups[group_index];
std::memcpy(&group.attrib, attribute, sizeof(BufferAttribute));
group.size_in_bytes =
attribute->height * attribute->pitch_in_pixel * PixelFormatBpp(attribute->pixel_format);
group.is_occupied = true;
for (u32 i = 0; i < bufferNum; i++) {

View File

@ -100,11 +100,11 @@ Id EmitFPNeg64(EmitContext& ctx, Id value) {
}
Id EmitFPSin(EmitContext& ctx, Id value) {
return ctx.OpSin(ctx.F32[1], value);
return ctx.OpSin(ctx.F32[1], ctx.OpFMul(ctx.F32[1], ctx.pi_x2, value));
}
Id EmitFPCos(EmitContext& ctx, Id value) {
return ctx.OpCos(ctx.F32[1], value);
return ctx.OpCos(ctx.F32[1], ctx.OpFMul(ctx.F32[1], ctx.pi_x2, value));
}
Id EmitFPExp2(EmitContext& ctx, Id value) {

View File

@ -1,11 +1,14 @@
// SPDX-FileCopyrightText: Copyright 2024 shadPS4 Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later
#include <boost/container/static_vector.hpp>
#include <fmt/format.h>
#include "common/div_ceil.h"
#include "shader_recompiler/backend/spirv/spirv_emit_context.h"
#include <boost/container/static_vector.hpp>
#include <fmt/format.h>
#include <numbers>
namespace Shader::Backend::SPIRV {
namespace {
@ -100,6 +103,8 @@ void EmitContext::DefineArithmeticTypes() {
u32_zero_value = ConstU32(0U);
f32_zero_value = ConstF32(0.0f);
pi_x2 = ConstF32(2.0f * float{std::numbers::pi});
input_f32 = Name(TypePointer(spv::StorageClass::Input, F32[1]), "input_f32");
input_u32 = Name(TypePointer(spv::StorageClass::Input, U32[1]), "input_u32");
input_s32 = Name(TypePointer(spv::StorageClass::Input, S32[1]), "input_s32");

View File

@ -143,6 +143,8 @@ public:
Id full_result_i32x2;
Id full_result_u32x2;
Id pi_x2;
Id true_value{};
Id false_value{};
Id u32_one_value{};

View File

@ -321,7 +321,7 @@ struct Liverpool {
struct DepthBuffer {
enum class ZFormat : u32 {
Invald = 0,
Invalid = 0,
Z16 = 1,
Z32Float = 3,
};
@ -367,8 +367,14 @@ struct Liverpool {
return u64(z_read_base) << 8;
}
size_t GetSizeAligned() const {
return depth_slice.tile_max * 8;
u32 NumSamples() const {
return 1u << z_info.num_samples; // spec doesn't say it is a log2
}
size_t GetDepthSliceSize() const {
ASSERT(z_info.format != ZFormat::Invalid);
const auto bpe = z_info.format == ZFormat::Z32Float ? 4 : 2;
return (depth_slice.tile_max + 1) * 64 * bpe * NumSamples();
}
};
@ -733,12 +739,19 @@ struct Liverpool {
return VAddr(fmask_base_address) << 8;
}
size_t GetSizeAligned() const {
u32 NumSamples() const {
return 1 << attrib.num_fragments_log2;
}
u32 NumSlices() const {
return view.slice_max + 1;
}
size_t GetColorSliceSize() const {
const auto num_bytes_per_element = NumBits(info.format) / 8u;
const auto slice_size = (slice.tile_max + 1) * 64u;
const auto total_size = slice_size * (view.slice_max + 1) * num_bytes_per_element;
ASSERT(total_size > 0);
return total_size;
const auto slice_size =
num_bytes_per_element * (slice.tile_max + 1) * 64u * NumSamples();
return slice_size;
}
TilingMode GetTilingMode() const {
@ -819,6 +832,17 @@ struct Liverpool {
BitField<6, 1, u32> depth_compress_disable;
};
union DepthView {
BitField<0, 11, u32> slice_start;
BitField<13, 11, u32> slice_max;
BitField<24, 1, u32> z_read_only;
BitField<25, 1, u32> stencil_read_only;
u32 NumSlices() const {
return slice_max + 1u;
}
};
union AaConfig {
BitField<0, 3, u32> msaa_num_samples;
BitField<4, 1, u32> aa_mask_centroid_dtmn;
@ -849,7 +873,9 @@ struct Liverpool {
ComputeProgram cs_program;
INSERT_PADDING_WORDS(0xA008 - 0x2E00 - 80 - 3 - 5);
DepthRenderControl depth_render_control;
INSERT_PADDING_WORDS(4);
INSERT_PADDING_WORDS(1);
DepthView depth_view;
INSERT_PADDING_WORDS(2);
Address depth_htile_data_base;
INSERT_PADDING_WORDS(2);
float depth_bounds_min;
@ -1050,6 +1076,7 @@ static_assert(GFX6_3D_REG_INDEX(cs_program.dim_z) == 0x2E03);
static_assert(GFX6_3D_REG_INDEX(cs_program.address_lo) == 0x2E0C);
static_assert(GFX6_3D_REG_INDEX(cs_program.user_data) == 0x2E40);
static_assert(GFX6_3D_REG_INDEX(depth_render_control) == 0xA000);
static_assert(GFX6_3D_REG_INDEX(depth_view) == 0xA002);
static_assert(GFX6_3D_REG_INDEX(depth_htile_data_base) == 0xA005);
static_assert(GFX6_3D_REG_INDEX(screen_scissor) == 0xA00C);
static_assert(GFX6_3D_REG_INDEX(depth_buffer.z_info) == 0xA010);

View File

@ -36,6 +36,12 @@ struct Buffer {
u32 element_size : 2;
u32 index_stride : 2;
u32 add_tid_enable : 1;
u32 : 6;
u32 type : 2; // overlaps with T# type, so should be 0 for buffer
bool Valid() const {
return type == 0u;
}
operator bool() const noexcept {
return base_address != 0;
@ -106,22 +112,25 @@ constexpr std::string_view NameOf(ImageType type) {
}
enum class TilingMode : u32 {
Depth_MicroTiled = 0x5u,
Depth_MacroTiled = 0u,
Display_Linear = 0x8u,
Display_MacroTiled = 0xAu,
Texture_MicroTiled = 0xDu,
Texture_MacroTiled = 0xEu,
};
constexpr std::string_view NameOf(TilingMode type) {
switch (type) {
case TilingMode::Depth_MicroTiled:
return "Depth_MicroTiled";
case TilingMode::Depth_MacroTiled:
return "Depth_MacroTiled";
case TilingMode::Display_Linear:
return "Display_Linear";
case TilingMode::Display_MacroTiled:
return "Display_MacroTiled";
case TilingMode::Texture_MicroTiled:
return "Texture_MicroTiled";
case TilingMode::Texture_MacroTiled:
return "Texture_MacroTiled";
default:
return "Unknown";
}
@ -149,7 +158,7 @@ struct Image {
u64 pow2pad : 1;
u64 mtype2 : 1;
u64 atc : 1;
u64 type : 4;
u64 type : 4; // overlaps with V# type, so shouldn't be 0 for buffer
u64 depth : 13;
u64 pitch : 14;
@ -162,6 +171,10 @@ struct Image {
u64 lod_hw_cnt_en : 1;
u64 : 43;
bool Valid() const {
return (type & 0x8u) != 0;
}
VAddr Address() const {
return base_address << 8;
}
@ -201,17 +214,19 @@ struct Image {
}
TilingMode GetTilingMode() const {
if (tiling_index >= 0 && tiling_index <= 7) {
return tiling_index == 5 ? TilingMode::Texture_MicroTiled
: TilingMode::Depth_MacroTiled;
}
if (tiling_index == 0x13) {
return TilingMode::Texture_MicroTiled;
}
return static_cast<TilingMode>(tiling_index);
}
bool IsTiled() const {
return GetTilingMode() != TilingMode::Display_Linear;
}
size_t GetSizeAligned() const {
// TODO: Derive this properly from tiling params
return Pitch() * (height + 1) * NumComponents(GetDataFmt());
}
};
static_assert(sizeof(Image) == 32); // 256bits

View File

@ -505,7 +505,7 @@ vk::Format DepthFormat(DepthBuffer::ZFormat z_format, DepthBuffer::StencilFormat
stencil_format == DepthBuffer::StencilFormat::Stencil8) {
return vk::Format::eD16UnormS8Uint;
}
if (z_format == DepthBuffer::ZFormat::Invald &&
if (z_format == DepthBuffer::ZFormat::Invalid &&
stencil_format == DepthBuffer::StencilFormat::Invalid) {
return vk::Format::eUndefined;
}

View File

@ -182,8 +182,9 @@ bool RendererVulkan::ShowSplash(Frame* frame /*= nullptr*/) {
info.size =
VideoCore::Extent3D{splash->GetImageInfo().width, splash->GetImageInfo().height, 1};
info.pitch = splash->GetImageInfo().width;
info.guest_address = VAddr(splash->GetImageData().data());
info.guest_size_bytes = splash->GetImageData().size();
splash_img.emplace(instance, scheduler, info, VAddr(splash->GetImageData().data()));
splash_img.emplace(instance, scheduler, info);
texture_cache.RefreshImage(*splash_img);
}
frame = PrepareFrameInternal(*splash_img);

View File

@ -40,7 +40,7 @@ public:
Frame* PrepareFrame(const Libraries::VideoOut::BufferAttributeGroup& attribute,
VAddr cpu_address) {
const auto info = VideoCore::ImageInfo{attribute};
const auto info = VideoCore::ImageInfo{attribute, cpu_address};
const auto image_id = texture_cache.FindImage(info, cpu_address);
auto& image = texture_cache.GetImage(image_id);
return PrepareFrameInternal(image);
@ -54,7 +54,7 @@ public:
VideoCore::Image& RegisterVideoOutSurface(
const Libraries::VideoOut::BufferAttributeGroup& attribute, VAddr cpu_address) {
vo_buffers_addr.emplace_back(cpu_address);
const auto info = VideoCore::ImageInfo{attribute};
const auto info = VideoCore::ImageInfo{attribute, cpu_address};
const auto image_id = texture_cache.FindImage(info, cpu_address);
return texture_cache.GetImage(image_id);
}

View File

@ -128,7 +128,7 @@ bool ComputePipeline::BindResources(Core::MemoryManager* memory, StreamBuffer& s
for (const auto& image_desc : info.images) {
const auto tsharp =
info.ReadUd<AmdGpu::Image>(image_desc.sgpr_base, image_desc.dword_offset);
const auto& image_view = texture_cache.FindImageView(tsharp, image_desc.is_storage);
const auto& image_view = texture_cache.FindTexture(tsharp, image_desc.is_storage);
const auto& image = texture_cache.GetImage(image_view.image_id);
image_infos.emplace_back(VK_NULL_HANDLE, *image_view.image_view, image.layout);
set_writes.push_back({

View File

@ -366,7 +366,7 @@ void GraphicsPipeline::BindResources(Core::MemoryManager* memory, StreamBuffer&
for (const auto& image_desc : stage.images) {
const auto& tsharp = tsharps.emplace_back(
stage.ReadUd<AmdGpu::Image>(image_desc.sgpr_base, image_desc.dword_offset));
const auto& image_view = texture_cache.FindImageView(tsharp, image_desc.is_storage);
const auto& image_view = texture_cache.FindTexture(tsharp, image_desc.is_storage);
const auto& image = texture_cache.GetImage(image_view.image_id);
image_infos.emplace_back(VK_NULL_HANDLE, *image_view.image_view, image.layout);
set_writes.push_back({

View File

@ -113,7 +113,7 @@ void Rasterizer::BeginRendering() {
}
const auto& hint = liverpool->last_cb_extent[col_buf_id];
const auto& image_view = texture_cache.RenderTarget(col_buf, hint);
const auto& image_view = texture_cache.FindRenderTarget(col_buf, hint);
const auto& image = texture_cache.GetImage(image_view.image_id);
state.width = std::min<u32>(state.width, image.info.size.width);
state.height = std::min<u32>(state.height, image.info.size.height);
@ -130,13 +130,14 @@ void Rasterizer::BeginRendering() {
texture_cache.TouchMeta(col_buf.CmaskAddress(), false);
}
if (regs.depth_buffer.z_info.format != Liverpool::DepthBuffer::ZFormat::Invald &&
if (regs.depth_buffer.z_info.format != Liverpool::DepthBuffer::ZFormat::Invalid &&
regs.depth_buffer.Address() != 0) {
const auto htile_address = regs.depth_htile_data_base.GetAddress();
const bool is_clear = regs.depth_render_control.depth_clear_enable ||
texture_cache.IsMetaCleared(htile_address);
const auto& hint = liverpool->last_db_extent;
const auto& image_view = texture_cache.DepthTarget(regs.depth_buffer, htile_address, hint,
const auto& image_view = texture_cache.FindDepthTarget(
regs.depth_buffer, regs.depth_view.NumSlices(), htile_address, hint,
regs.depth_control.depth_write_enable);
const auto& image = texture_cache.GetImage(image_view.image_id);
state.width = std::min<u32>(state.width, image.info.size.width);

View File

@ -2,7 +2,6 @@
// SPDX-License-Identifier: GPL-2.0-or-later
#include "common/assert.h"
#include "common/config.h"
#include "video_core/renderer_vulkan/liverpool_to_vk.h"
#include "video_core/renderer_vulkan/vk_instance.h"
#include "video_core/renderer_vulkan/vk_scheduler.h"
@ -14,25 +13,8 @@
namespace VideoCore {
using namespace Vulkan;
using VideoOutFormat = Libraries::VideoOut::PixelFormat;
using Libraries::VideoOut::TilingMode;
static vk::Format ConvertPixelFormat(const VideoOutFormat format) {
switch (format) {
case VideoOutFormat::A8R8G8B8Srgb:
return vk::Format::eB8G8R8A8Srgb;
case VideoOutFormat::A8B8G8R8Srgb:
return vk::Format::eR8G8B8A8Srgb;
case VideoOutFormat::A2R10G10B10:
case VideoOutFormat::A2R10G10B10Srgb:
return vk::Format::eA2R10G10B10UnormPack32;
default:
break;
}
UNREACHABLE_MSG("Unknown format={}", static_cast<u32>(format));
return {};
}
bool ImageInfo::IsBlockCoded() const {
switch (pixel_format) {
case vk::Format::eBc1RgbaSrgbBlock:
@ -101,93 +83,6 @@ static vk::ImageUsageFlags ImageUsageFlags(const ImageInfo& info) {
return usage;
}
static vk::ImageType ConvertImageType(AmdGpu::ImageType type) noexcept {
switch (type) {
case AmdGpu::ImageType::Color1D:
case AmdGpu::ImageType::Color1DArray:
return vk::ImageType::e1D;
case AmdGpu::ImageType::Color2D:
case AmdGpu::ImageType::Cube:
case AmdGpu::ImageType::Color2DArray:
return vk::ImageType::e2D;
case AmdGpu::ImageType::Color3D:
return vk::ImageType::e3D;
default:
UNREACHABLE();
}
}
ImageInfo::ImageInfo(const Libraries::VideoOut::BufferAttributeGroup& group) noexcept {
const auto& attrib = group.attrib;
is_tiled = attrib.tiling_mode == TilingMode::Tile;
tiling_mode =
is_tiled ? AmdGpu::TilingMode::Display_MacroTiled : AmdGpu::TilingMode::Display_Linear;
pixel_format = ConvertPixelFormat(attrib.pixel_format);
type = vk::ImageType::e2D;
size.width = attrib.width;
size.height = attrib.height;
pitch = attrib.tiling_mode == TilingMode::Linear ? size.width : (size.width + 127) & (~127);
const bool is_32bpp = attrib.pixel_format != VideoOutFormat::A16R16G16B16Float;
ASSERT(is_32bpp);
if (!is_tiled) {
guest_size_bytes = pitch * size.height * 4;
return;
}
if (Config::isNeoMode()) {
guest_size_bytes = pitch * ((size.height + 127) & (~127)) * 4;
} else {
guest_size_bytes = pitch * ((size.height + 63) & (~63)) * 4;
}
usage.vo_buffer = true;
}
ImageInfo::ImageInfo(const AmdGpu::Liverpool::ColorBuffer& buffer,
const AmdGpu::Liverpool::CbDbExtent& hint /*= {}*/) noexcept {
is_tiled = buffer.IsTiled();
tiling_mode = buffer.GetTilingMode();
pixel_format = LiverpoolToVK::SurfaceFormat(buffer.info.format, buffer.NumFormat());
num_samples = 1 << buffer.attrib.num_fragments_log2;
type = vk::ImageType::e2D;
size.width = hint.Valid() ? hint.width : buffer.Pitch();
size.height = hint.Valid() ? hint.height : buffer.Height();
size.depth = 1;
pitch = size.width;
guest_size_bytes = buffer.GetSizeAligned();
meta_info.cmask_addr = buffer.info.fast_clear ? buffer.CmaskAddress() : 0;
meta_info.fmask_addr = buffer.info.compression ? buffer.FmaskAddress() : 0;
usage.render_target = true;
}
ImageInfo::ImageInfo(const AmdGpu::Liverpool::DepthBuffer& buffer, VAddr htile_address,
const AmdGpu::Liverpool::CbDbExtent& hint) noexcept {
is_tiled = false;
pixel_format = LiverpoolToVK::DepthFormat(buffer.z_info.format, buffer.stencil_info.format);
type = vk::ImageType::e2D;
num_samples = 1 << buffer.z_info.num_samples; // spec doesn't say it is a log2
size.width = hint.Valid() ? hint.width : buffer.Pitch();
size.height = hint.Valid() ? hint.height : buffer.Height();
size.depth = 1;
pitch = size.width;
guest_size_bytes = buffer.GetSizeAligned();
meta_info.htile_addr = buffer.z_info.tile_surface_en ? htile_address : 0;
usage.depth_target = true;
}
ImageInfo::ImageInfo(const AmdGpu::Image& image) noexcept {
is_tiled = image.IsTiled();
tiling_mode = image.GetTilingMode();
pixel_format = LiverpoolToVK::SurfaceFormat(image.GetDataFmt(), image.GetNumberFmt());
type = ConvertImageType(image.GetType());
size.width = image.width + 1;
size.height = image.height + 1;
size.depth = 1;
pitch = image.Pitch();
resources.levels = image.NumLevels();
resources.layers = image.NumLayers();
guest_size_bytes = image.GetSizeAligned();
usage.texture = true;
}
UniqueImage::UniqueImage(vk::Device device_, VmaAllocator allocator_)
: device{device_}, allocator{allocator_} {}
@ -217,9 +112,9 @@ void UniqueImage::Create(const vk::ImageCreateInfo& image_ci) {
}
Image::Image(const Vulkan::Instance& instance_, Vulkan::Scheduler& scheduler_,
const ImageInfo& info_, VAddr cpu_addr)
const ImageInfo& info_)
: instance{&instance_}, scheduler{&scheduler_}, info{info_},
image{instance->GetDevice(), instance->GetAllocator()}, cpu_addr{cpu_addr},
image{instance->GetDevice(), instance->GetAllocator()}, cpu_addr{info.guest_address},
cpu_addr_end{cpu_addr + info.guest_size_bytes} {
ASSERT(info.pixel_format != vk::Format::eUndefined);
vk::ImageCreateFlags flags{vk::ImageCreateFlagBits::eMutableFormat |

View File

@ -9,6 +9,7 @@
#include "video_core/amdgpu/liverpool.h"
#include "video_core/amdgpu/resource.h"
#include "video_core/renderer_vulkan/vk_common.h"
#include "video_core/texture_cache/image_info.h"
#include "video_core/texture_cache/image_view.h"
#include "video_core/texture_cache/types.h"
@ -34,47 +35,6 @@ enum ImageFlagBits : u32 {
};
DECLARE_ENUM_FLAG_OPERATORS(ImageFlagBits)
struct ImageInfo {
ImageInfo() = default;
explicit ImageInfo(const Libraries::VideoOut::BufferAttributeGroup& group) noexcept;
explicit ImageInfo(const AmdGpu::Liverpool::ColorBuffer& buffer,
const AmdGpu::Liverpool::CbDbExtent& hint = {}) noexcept;
explicit ImageInfo(const AmdGpu::Liverpool::DepthBuffer& buffer, VAddr htile_address,
const AmdGpu::Liverpool::CbDbExtent& hint = {}) noexcept;
explicit ImageInfo(const AmdGpu::Image& image) noexcept;
bool IsTiled() const {
return tiling_mode != AmdGpu::TilingMode::Display_Linear;
}
bool IsBlockCoded() const;
bool IsPacked() const;
bool IsDepthStencil() const;
struct {
VAddr cmask_addr;
VAddr fmask_addr;
VAddr htile_addr;
} meta_info{};
struct {
u32 texture : 1;
u32 storage : 1;
u32 render_target : 1;
u32 depth_target : 1;
u32 vo_buffer : 1;
} usage{}; // Usage data tracked during image lifetime
bool is_tiled = false;
vk::Format pixel_format = vk::Format::eUndefined;
vk::ImageType type = vk::ImageType::e1D;
SubresourceExtent resources;
Extent3D size{1, 1, 1};
u32 num_samples = 1;
u32 pitch = 0;
u32 guest_size_bytes = 0;
AmdGpu::TilingMode tiling_mode{AmdGpu::TilingMode::Display_Linear};
};
struct UniqueImage {
explicit UniqueImage(vk::Device device, VmaAllocator allocator);
~UniqueImage();
@ -109,8 +69,7 @@ private:
constexpr Common::SlotId NULL_IMAGE_ID{0};
struct Image {
explicit Image(const Vulkan::Instance& instance, Vulkan::Scheduler& scheduler,
const ImageInfo& info, VAddr cpu_addr);
Image(const Vulkan::Instance& instance, Vulkan::Scheduler& scheduler, const ImageInfo& info);
~Image();
Image(const Image&) = delete;

View File

@ -0,0 +1,268 @@
// SPDX-FileCopyrightText: Copyright 2024 shadPS4 Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later
#include "common/assert.h"
#include "common/config.h"
#include "video_core/renderer_vulkan/liverpool_to_vk.h"
#include "video_core/texture_cache/image_info.h"
namespace VideoCore {
using namespace Vulkan;
using Libraries::VideoOut::TilingMode;
using VideoOutFormat = Libraries::VideoOut::PixelFormat;
static vk::Format ConvertPixelFormat(const VideoOutFormat format) {
switch (format) {
case VideoOutFormat::A8R8G8B8Srgb:
return vk::Format::eB8G8R8A8Srgb;
case VideoOutFormat::A8B8G8R8Srgb:
return vk::Format::eR8G8B8A8Srgb;
case VideoOutFormat::A2R10G10B10:
case VideoOutFormat::A2R10G10B10Srgb:
return vk::Format::eA2R10G10B10UnormPack32;
default:
break;
}
UNREACHABLE_MSG("Unknown format={}", static_cast<u32>(format));
return {};
}
static vk::ImageType ConvertImageType(AmdGpu::ImageType type) noexcept {
switch (type) {
case AmdGpu::ImageType::Color1D:
case AmdGpu::ImageType::Color1DArray:
return vk::ImageType::e1D;
case AmdGpu::ImageType::Color2D:
case AmdGpu::ImageType::Cube:
case AmdGpu::ImageType::Color2DArray:
return vk::ImageType::e2D;
case AmdGpu::ImageType::Color3D:
return vk::ImageType::e3D;
default:
UNREACHABLE();
}
}
// clang-format off
// The table of macro tiles parameters for given tiling index (row) and bpp (column)
static constexpr std::array macro_tile_extents{
std::pair{256u, 128u}, std::pair{256u, 128u}, std::pair{256u, 128u}, std::pair{256u, 128u}, // 00
std::pair{256u, 128u}, std::pair{128u, 128u}, std::pair{128u, 128u}, std::pair{128u, 128u}, // 01
std::pair{256u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u}, std::pair{128u, 64u}, // 02
std::pair{256u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u}, std::pair{128u, 64u}, // 03
std::pair{256u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u}, std::pair{128u, 64u}, // 04
std::pair{0u, 0u}, std::pair{0u, 0u}, std::pair{0u, 0u}, std::pair{0u, 0u}, // 05
std::pair{256u, 256u}, std::pair{256u, 128u}, std::pair{128u, 128u}, std::pair{128u, 128u}, // 06
std::pair{256u, 256u}, std::pair{256u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u}, // 07
std::pair{0u, 0u}, std::pair{0u, 0u}, std::pair{0u, 0u}, std::pair{0u, 0u}, // 08
std::pair{0u, 0u}, std::pair{0u, 0u}, std::pair{0u, 0u}, std::pair{0u, 0u}, // 09
std::pair{256u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u}, std::pair{128u, 64u}, // 0A
std::pair{256u, 256u}, std::pair{256u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u}, // 0B
std::pair{256u, 256u}, std::pair{256u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u}, // 0C
std::pair{0u, 0u}, std::pair{0u, 0u}, std::pair{0u, 0u}, std::pair{0u, 0u}, // 0D
std::pair{256u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u}, std::pair{128u, 64u}, // 0E
std::pair{256u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u}, std::pair{128u, 64u}, // 0F
std::pair{256u, 256u}, std::pair{256u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u}, // 10
std::pair{256u, 256u}, std::pair{256u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u}, // 11
std::pair{256u, 256u}, std::pair{256u, 128u}, std::pair{128u, 128u}, std::pair{128u, 64u}, // 12
std::pair{0u, 0u}, std::pair{0u, 0u}, std::pair{0u, 0u}, std::pair{0u, 0u}, // 13
std::pair{128u, 64u}, std::pair{128u, 64u}, std::pair{64u, 64u}, std::pair{64u, 64u}, // 14
std::pair{128u, 64u}, std::pair{128u, 64u}, std::pair{64u, 64u}, std::pair{64u, 64u}, // 15
std::pair{128u, 128u}, std::pair{128u, 64u}, std::pair{64u, 64u}, std::pair{64u, 64u}, // 16
std::pair{128u, 128u}, std::pair{128u, 64u}, std::pair{64u, 64u}, std::pair{64u, 64u}, // 17
std::pair{128u, 128u}, std::pair{128u, 64u}, std::pair{64u, 64u}, std::pair{64u, 64u}, // 18
std::pair{128u, 64u}, std::pair{64u, 64u}, std::pair{64u, 64u}, std::pair{64u, 64u}, // 19
std::pair{128u, 64u}, std::pair{64u, 64u}, std::pair{64u, 64u}, std::pair{64u, 64u}, // 1A
};
// clang-format on
static constexpr std::pair micro_tile_extent{8u, 8u};
static constexpr auto hw_pipe_interleave = 256u;
static constexpr std::pair<u32, u32> GetMacroTileExtents(u32 tiling_idx, u32 bpp, u32 num_samples) {
ASSERT(num_samples == 1);
const auto row = tiling_idx * 4;
const auto column = std::bit_width(bpp) - 4; // bpps are 8, 16, 32, 64
return macro_tile_extents[row + column];
}
static constexpr size_t ImageSizeLinearAligned(u32 pitch, u32 height, u32 bpp, u32 num_samples) {
const auto pitch_align = std::max(8u, 64u / ((bpp + 7) / 8));
auto pitch_aligned = (pitch + pitch_align - 1) & ~(pitch_align - 1);
const auto height_aligned = height;
size_t log_sz = 1;
const auto slice_align = std::max(64u, hw_pipe_interleave / (bpp + 7) / 8);
while (log_sz % slice_align) {
log_sz = pitch_aligned * height_aligned * num_samples;
pitch_aligned += pitch_align;
}
return (log_sz * bpp + 7) / 8;
}
static constexpr size_t ImageSizeMicroTiled(u32 pitch, u32 height, u32 bpp, u32 num_samples) {
const auto& [pitch_align, height_align] = micro_tile_extent;
auto pitch_aligned = (pitch + pitch_align - 1) & ~(pitch_align - 1);
const auto height_aligned = (height + height_align - 1) & ~(height_align - 1);
size_t log_sz = 1;
while (log_sz % 256) {
log_sz = (pitch_aligned * height_aligned * bpp * num_samples + 7) / 8;
pitch_aligned += 8;
}
return log_sz;
}
static constexpr size_t ImageSizeMacroTiled(u32 pitch, u32 height, u32 bpp, u32 num_samples,
u32 tiling_idx) {
const auto& [pitch_align, height_align] = GetMacroTileExtents(tiling_idx, bpp, num_samples);
ASSERT(pitch_align != 0 && height_align != 0);
const auto pitch_aligned = (pitch + pitch_align - 1) & ~(pitch_align - 1);
const auto height_aligned = (height + height_align - 1) & ~(height_align - 1);
return (pitch_aligned * height_aligned * bpp * num_samples + 7) / 8;
}
ImageInfo::ImageInfo(const Libraries::VideoOut::BufferAttributeGroup& group,
VAddr cpu_address) noexcept {
const auto& attrib = group.attrib;
is_tiled = attrib.tiling_mode == TilingMode::Tile;
tiling_mode =
is_tiled ? AmdGpu::TilingMode::Display_MacroTiled : AmdGpu::TilingMode::Display_Linear;
pixel_format = ConvertPixelFormat(attrib.pixel_format);
type = vk::ImageType::e2D;
size.width = attrib.width;
size.height = attrib.height;
pitch = attrib.tiling_mode == TilingMode::Linear ? size.width : (size.width + 127) & (~127);
usage.vo_buffer = true;
const bool is_32bpp = attrib.pixel_format != VideoOutFormat::A16R16G16B16Float;
ASSERT(is_32bpp);
guest_address = cpu_address;
if (!is_tiled) {
guest_size_bytes = pitch * size.height * 4;
} else {
if (Config::isNeoMode()) {
guest_size_bytes = pitch * ((size.height + 127) & (~127)) * 4;
} else {
guest_size_bytes = pitch * ((size.height + 63) & (~63)) * 4;
}
}
mips_layout.emplace_back(0, guest_size_bytes);
}
ImageInfo::ImageInfo(const AmdGpu::Liverpool::ColorBuffer& buffer,
const AmdGpu::Liverpool::CbDbExtent& hint /*= {}*/) noexcept {
is_tiled = buffer.IsTiled();
tiling_mode = buffer.GetTilingMode();
pixel_format = LiverpoolToVK::SurfaceFormat(buffer.info.format, buffer.NumFormat());
num_samples = 1 << buffer.attrib.num_fragments_log2;
type = vk::ImageType::e2D;
size.width = hint.Valid() ? hint.width : buffer.Pitch();
size.height = hint.Valid() ? hint.height : buffer.Height();
size.depth = 1;
pitch = buffer.Pitch();
resources.layers = buffer.NumSlices();
meta_info.cmask_addr = buffer.info.fast_clear ? buffer.CmaskAddress() : 0;
meta_info.fmask_addr = buffer.info.compression ? buffer.FmaskAddress() : 0;
usage.render_target = true;
guest_address = buffer.Address();
const auto color_slice_sz = buffer.GetColorSliceSize();
guest_size_bytes = color_slice_sz * buffer.NumSlices();
mips_layout.emplace_back(0, color_slice_sz);
}
ImageInfo::ImageInfo(const AmdGpu::Liverpool::DepthBuffer& buffer, u32 num_slices,
VAddr htile_address, const AmdGpu::Liverpool::CbDbExtent& hint) noexcept {
is_tiled = false;
pixel_format = LiverpoolToVK::DepthFormat(buffer.z_info.format, buffer.stencil_info.format);
type = vk::ImageType::e2D;
num_samples = 1 << buffer.z_info.num_samples; // spec doesn't say it is a log2
size.width = hint.Valid() ? hint.width : buffer.Pitch();
size.height = hint.Valid() ? hint.height : buffer.Height();
size.depth = 1;
pitch = size.width;
resources.layers = num_slices;
meta_info.htile_addr = buffer.z_info.tile_surface_en ? htile_address : 0;
usage.depth_target = true;
guest_address = buffer.Address();
const auto depth_slice_sz = buffer.GetDepthSliceSize();
guest_size_bytes = depth_slice_sz * num_slices;
mips_layout.emplace_back(0, depth_slice_sz);
}
ImageInfo::ImageInfo(const AmdGpu::Image& image) noexcept {
is_tiled = image.IsTiled();
tiling_mode = image.GetTilingMode();
pixel_format = LiverpoolToVK::SurfaceFormat(image.GetDataFmt(), image.GetNumberFmt());
type = ConvertImageType(image.GetType());
is_cube = image.GetType() == AmdGpu::ImageType::Cube;
is_volume = image.GetType() == AmdGpu::ImageType::Color3D;
size.width = image.width + 1;
size.height = image.height + 1;
size.depth = is_volume ? image.depth + 1 : 1;
pitch = image.Pitch();
resources.levels = image.NumLevels();
resources.layers = image.NumLayers();
usage.texture = true;
guest_address = image.Address();
mips_layout.reserve(resources.levels);
const auto num_bits = NumBits(image.GetDataFmt());
const auto is_block = IsBlockCoded();
const auto is_pow2 = image.pow2pad;
guest_size_bytes = 0;
for (auto mip = 0u; mip < resources.levels; ++mip) {
auto bpp = num_bits;
auto mip_w = pitch >> mip;
auto mip_h = size.height >> mip;
if (is_block) {
mip_w = (mip_w + 3) / 4;
mip_h = (mip_h + 3) / 4;
bpp *= 16;
}
mip_w = std::max(mip_w, 1u);
mip_h = std::max(mip_h, 1u);
auto mip_d = std::max(size.depth >> mip, 1u);
if (is_pow2) {
mip_w = std::bit_ceil(mip_w);
mip_h = std::bit_ceil(mip_h);
mip_d = std::bit_ceil(mip_d);
}
size_t mip_size = 0;
switch (tiling_mode) {
case AmdGpu::TilingMode::Display_Linear: {
ASSERT(!is_cube);
mip_size = ImageSizeLinearAligned(mip_w, mip_h, bpp, num_samples);
break;
}
case AmdGpu::TilingMode::Texture_MicroTiled: {
mip_size = ImageSizeMicroTiled(mip_w, mip_h, bpp, num_samples);
break;
}
case AmdGpu::TilingMode::Display_MacroTiled:
case AmdGpu::TilingMode::Texture_MacroTiled:
case AmdGpu::TilingMode::Depth_MacroTiled: {
ASSERT(!is_cube && !is_block);
ASSERT(num_samples == 1);
ASSERT(num_bits <= 64);
mip_size = ImageSizeMacroTiled(mip_w, mip_h, bpp, num_samples, image.tiling_index);
break;
}
default: {
UNREACHABLE();
}
}
mip_size *= mip_d;
mips_layout.emplace_back(guest_size_bytes, mip_size);
guest_size_bytes += mip_size;
}
guest_size_bytes *= resources.layers;
}
} // namespace VideoCore

View File

@ -0,0 +1,61 @@
// SPDX-FileCopyrightText: Copyright 2024 shadPS4 Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later
#pragma once
#include "common/enum.h"
#include "common/types.h"
#include "core/libraries/videoout/buffer.h"
#include "video_core/amdgpu/liverpool.h"
#include "video_core/texture_cache/types.h"
namespace VideoCore {
struct ImageInfo {
ImageInfo() = default;
ImageInfo(const Libraries::VideoOut::BufferAttributeGroup& group, VAddr cpu_address) noexcept;
ImageInfo(const AmdGpu::Liverpool::ColorBuffer& buffer,
const AmdGpu::Liverpool::CbDbExtent& hint = {}) noexcept;
ImageInfo(const AmdGpu::Liverpool::DepthBuffer& buffer, u32 num_slices, VAddr htile_address,
const AmdGpu::Liverpool::CbDbExtent& hint = {}) noexcept;
ImageInfo(const AmdGpu::Image& image) noexcept;
bool IsTiled() const {
return tiling_mode != AmdGpu::TilingMode::Display_Linear;
}
bool IsBlockCoded() const;
bool IsPacked() const;
bool IsDepthStencil() const;
struct {
VAddr cmask_addr;
VAddr fmask_addr;
VAddr htile_addr;
} meta_info{};
struct {
u32 texture : 1;
u32 storage : 1;
u32 render_target : 1;
u32 depth_target : 1;
u32 stencil : 1;
u32 vo_buffer : 1;
} usage{}; // Usage data tracked during image lifetime
bool is_cube = false;
bool is_volume = false;
bool is_tiled = false;
bool is_read_only = false;
vk::Format pixel_format = vk::Format::eUndefined;
vk::ImageType type = vk::ImageType::e1D;
SubresourceExtent resources;
Extent3D size{1, 1, 1};
u32 num_samples = 1;
u32 pitch = 0;
AmdGpu::TilingMode tiling_mode{AmdGpu::TilingMode::Display_Linear};
std::vector<std::pair<u32, u32>> mips_layout;
VAddr guest_address{0};
u32 guest_size_bytes{0};
};
} // namespace VideoCore

View File

@ -89,7 +89,7 @@ TextureCache::TextureCache(const Vulkan::Instance& instance_, Vulkan::Scheduler&
ImageInfo info;
info.pixel_format = vk::Format::eR8G8B8A8Unorm;
info.type = vk::ImageType::e2D;
const ImageId null_id = slot_images.insert(instance, scheduler, info, 0);
const ImageId null_id = slot_images.insert(instance, scheduler, info);
ASSERT(null_id.index == 0);
ImageViewInfo view_info;
@ -112,12 +112,13 @@ void TextureCache::OnCpuWrite(VAddr address) {
});
}
ImageId TextureCache::FindImage(const ImageInfo& info, VAddr cpu_address, bool refresh_on_create) {
ImageId TextureCache::FindImage(const ImageInfo& info, bool refresh_on_create) {
std::unique_lock lock{m_page_table};
boost::container::small_vector<ImageId, 2> image_ids;
ForEachImageInRegion(cpu_address, info.guest_size_bytes, [&](ImageId image_id, Image& image) {
ForEachImageInRegion(
info.guest_address, info.guest_size_bytes, [&](ImageId image_id, Image& image) {
// Address and width must match.
if (image.cpu_addr != cpu_address || image.info.size.width != info.size.width) {
if (image.cpu_addr != info.guest_address || image.info.size.width != info.size.width) {
return;
}
if (info.IsDepthStencil() != image.info.IsDepthStencil() &&
@ -131,7 +132,7 @@ ImageId TextureCache::FindImage(const ImageInfo& info, VAddr cpu_address, bool r
ImageId image_id{};
if (image_ids.empty()) {
image_id = slot_images.insert(instance, scheduler, info, cpu_address);
image_id = slot_images.insert(instance, scheduler, info);
RegisterImage(image_id);
} else {
image_id = image_ids[0];
@ -169,9 +170,9 @@ ImageView& TextureCache::RegisterImageView(ImageId image_id, const ImageViewInfo
return slot_image_views[view_id];
}
ImageView& TextureCache::FindImageView(const AmdGpu::Image& desc, bool is_storage) {
ImageView& TextureCache::FindTexture(const AmdGpu::Image& desc, bool is_storage) {
const ImageInfo info{desc};
const ImageId image_id = FindImage(info, desc.Address());
const ImageId image_id = FindImage(info);
Image& image = slot_images[image_id];
auto& usage = image.info.usage;
@ -190,10 +191,10 @@ ImageView& TextureCache::FindImageView(const AmdGpu::Image& desc, bool is_storag
return RegisterImageView(image_id, view_info);
}
ImageView& TextureCache::RenderTarget(const AmdGpu::Liverpool::ColorBuffer& buffer,
ImageView& TextureCache::FindRenderTarget(const AmdGpu::Liverpool::ColorBuffer& buffer,
const AmdGpu::Liverpool::CbDbExtent& hint) {
const ImageInfo info{buffer, hint};
const ImageId image_id = FindImage(info, buffer.Address());
const ImageId image_id = FindImage(info);
Image& image = slot_images[image_id];
image.flags &= ~ImageFlagBits::CpuModified;
@ -207,11 +208,12 @@ ImageView& TextureCache::RenderTarget(const AmdGpu::Liverpool::ColorBuffer& buff
return RegisterImageView(image_id, view_info);
}
ImageView& TextureCache::DepthTarget(const AmdGpu::Liverpool::DepthBuffer& buffer,
VAddr htile_address, const AmdGpu::Liverpool::CbDbExtent& hint,
ImageView& TextureCache::FindDepthTarget(const AmdGpu::Liverpool::DepthBuffer& buffer,
u32 num_slices, VAddr htile_address,
const AmdGpu::Liverpool::CbDbExtent& hint,
bool write_enabled) {
const ImageInfo info{buffer, htile_address, hint};
const ImageId image_id = FindImage(info, buffer.Address(), false);
const ImageInfo info{buffer, num_slices, htile_address, hint};
const ImageId image_id = FindImage(info, false);
Image& image = slot_images[image_id];
image.flags &= ~ImageFlagBits::CpuModified;
@ -244,21 +246,24 @@ void TextureCache::RefreshImage(Image& image) {
return;
}
ASSERT(image.info.resources.levels == image.info.mips_layout.size());
const u8* image_data = reinterpret_cast<const u8*>(image.cpu_addr);
for (u32 m = 0; m < image.info.resources.levels; m++) {
const u32 width = image.info.size.width >> m;
const u32 height = image.info.size.height >> m;
const u32 map_size = width * height * image.info.resources.layers;
const u32 width = std::max(image.info.size.width >> m, 1u);
const u32 height = std::max(image.info.size.height >> m, 1u);
const u32 depth = image.info.is_volume ? std::max(image.info.size.depth >> m, 1u) : 1u;
const u32 map_size = image.info.mips_layout[m].second * image.info.resources.layers;
// Upload data to the staging buffer.
const auto [data, offset, _] = staging.Map(map_size, 16);
if (image.info.is_tiled) {
ConvertTileToLinear(data, image_data, width, height, Config::isNeoMode());
} else {
std::memcpy(data, image_data, map_size);
std::memcpy(data,
image_data + image.info.mips_layout[m].first * image.info.resources.layers,
map_size);
}
staging.Commit(map_size);
image_data += map_size;
// Copy to the image.
const vk::BufferImageCopy image_copy = {
@ -272,7 +277,7 @@ void TextureCache::RefreshImage(Image& image) {
.layerCount = u32(image.info.resources.layers),
},
.imageOffset = {0, 0, 0},
.imageExtent = {width, height, 1},
.imageExtent = {width, height, depth},
};
scheduler.EndRendering();

View File

@ -47,18 +47,19 @@ public:
/// Invalidates any image in the logical page range.
void OnCpuWrite(VAddr address);
/// Retrieves the image handle of the image with the provided attributes and address.
[[nodiscard]] ImageId FindImage(const ImageInfo& info, VAddr cpu_address,
bool refresh_on_create = true);
/// Retrieves the image handle of the image with the provided attributes.
[[nodiscard]] ImageId FindImage(const ImageInfo& info, bool refresh_on_create = true);
/// Retrieves an image view with the properties of the specified image descriptor.
[[nodiscard]] ImageView& FindImageView(const AmdGpu::Image& image, bool is_storage);
[[nodiscard]] ImageView& FindTexture(const AmdGpu::Image& image, bool is_storage);
/// Retrieves the render target with specified properties
[[nodiscard]] ImageView& RenderTarget(const AmdGpu::Liverpool::ColorBuffer& buffer,
[[nodiscard]] ImageView& FindRenderTarget(const AmdGpu::Liverpool::ColorBuffer& buffer,
const AmdGpu::Liverpool::CbDbExtent& hint);
[[nodiscard]] ImageView& DepthTarget(const AmdGpu::Liverpool::DepthBuffer& buffer,
VAddr htile_address,
/// Retrieves the depth target with specified properties
[[nodiscard]] ImageView& FindDepthTarget(const AmdGpu::Liverpool::DepthBuffer& buffer,
u32 num_slices, VAddr htile_address,
const AmdGpu::Liverpool::CbDbExtent& hint,
bool write_enabled);

View File

@ -19,10 +19,6 @@
namespace VideoCore {
static u32 IntLog2(u32 i) {
return 31 - __builtin_clz(i | 1u);
}
class TileManager32 {
public:
u32 m_macro_tile_height = 0;
@ -81,8 +77,8 @@ public:
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 x_shift_offset = std::bit_width(bank_width * num_pipes) - 1;
const u32 y_shift_offset = std::bit_width(bank_height) - 1;
const u32 xs = x >> x_shift_offset;
const u32 ys = y >> y_shift_offset;
u32 bank = 0;
@ -210,8 +206,7 @@ vk::Format DemoteImageFormatForDetiling(vk::Format format) {
const DetilerContext* TileManager::GetDetiler(const Image& image) const {
const auto format = DemoteImageFormatForDetiling(image.info.pixel_format);
if (image.info.tiling_mode == AmdGpu::TilingMode::Texture_MicroTiled ||
image.info.tiling_mode == AmdGpu::TilingMode::Depth_MicroTiled) {
if (image.info.tiling_mode == AmdGpu::TilingMode::Texture_MicroTiled) {
switch (format) {
case vk::Format::eR8Uint:
return &detilers[DetilerType::Micro8x1];