mirrors/shadPS4
1
0
Fork 0
mirror of https://github.com/shadps4-emu/shadPS4.git synced 2025-01-30 01:58:33 +00:00
Code Issues Actions 3 Packages Projects Releases Wiki Activity

renderer_vulkan: Parse fetch shader per-pipeline (#1656)
Some checks are pending
Build and Release / reuse (push) Waiting to run
Details
Build and Release / clang-format (push) Waiting to run
Details
Build and Release / get-info (push) Waiting to run
Details
Build and Release / windows-sdl (push) Blocked by required conditions
Details
Build and Release / windows-qt (push) Blocked by required conditions
Details
Build and Release / macos-sdl (push) Blocked by required conditions
Details
Build and Release / macos-qt (push) Blocked by required conditions
Details
Build and Release / linux-sdl (push) Blocked by required conditions
Details
Build and Release / linux-qt (push) Blocked by required conditions
Details
Build and Release / pre-release (push) Blocked by required conditions
Details

Browse source 
* shader_recompiler: Read image format info directly from sharps instead of storing in shader info.

* renderer_vulkan: Parse fetch shader per-pipeline

* Few minor fixes.

* shader_recompiler: Specialize on vertex attribute number types.

* shader_recompiler: Move GetDrawOffsets to fetch shader
This commit is contained in:
squidbus 2024-12-04 03:03:47 -08:00 committed by GitHub
parent 74b091fd08
commit 920acb8d8b
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
21 changed files with 286 additions and 182 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

3
src/shader_recompiler/backend/spirv/emit_spirv_image.cpp
View file

@ -187,7 +187,8 @@ Id EmitImageFetch(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, const
Id EmitImageQueryDimensions(EmitContext& ctx, IR::Inst* inst, u32 handle, Id lod, bool has_mips) {
const auto& texture = ctx.images[handle & 0xFFFF];
const Id image = ctx.OpLoad(texture.image_type, texture.id);
const auto type = ctx.info.images[handle & 0xFFFF].type;
const auto sharp = ctx.info.images[handle & 0xFFFF].GetSharp(ctx.info);
const auto type = sharp.GetBoundType();
const Id zero = ctx.u32_zero_value;
const auto mips{[&] { return has_mips ? ctx.OpImageQueryLevels(ctx.U32[1], image) : zero; }};
const bool uses_lod{type != AmdGpu::ImageType::Color2DMsaa && !texture.is_storage};

91
src/shader_recompiler/backend/spirv/spirv_emit_context.cpp
View file

@ -4,6 +4,7 @@
#include "common/assert.h"
#include "common/div_ceil.h"
#include "shader_recompiler/backend/spirv/spirv_emit_context.h"
#include "shader_recompiler/frontend/fetch_shader.h"
#include "shader_recompiler/ir/passes/srt.h"
#include "video_core/amdgpu/types.h"
@ -155,18 +156,12 @@ void EmitContext::DefineInterfaces() {
}
const VectorIds& GetAttributeType(EmitContext& ctx, AmdGpu::NumberFormat fmt) {
switch (fmt) {
case AmdGpu::NumberFormat::Float:
case AmdGpu::NumberFormat::Unorm:
case AmdGpu::NumberFormat::Snorm:
case AmdGpu::NumberFormat::SnormNz:
case AmdGpu::NumberFormat::Sscaled:
case AmdGpu::NumberFormat::Uscaled:
case AmdGpu::NumberFormat::Srgb:
switch (GetNumberClass(fmt)) {
case AmdGpu::NumberClass::Float:
return ctx.F32;
case AmdGpu::NumberFormat::Sint:
case AmdGpu::NumberClass::Sint:
return ctx.S32;
case AmdGpu::NumberFormat::Uint:
case AmdGpu::NumberClass::Uint:
return ctx.U32;
default:
break;
@ -176,18 +171,12 @@ const VectorIds& GetAttributeType(EmitContext& ctx, AmdGpu::NumberFormat fmt) {
EmitContext::SpirvAttribute EmitContext::GetAttributeInfo(AmdGpu::NumberFormat fmt, Id id,
u32 num_components, bool output) {
switch (fmt) {
case AmdGpu::NumberFormat::Float:
case AmdGpu::NumberFormat::Unorm:
case AmdGpu::NumberFormat::Snorm:
case AmdGpu::NumberFormat::SnormNz:
case AmdGpu::NumberFormat::Sscaled:
case AmdGpu::NumberFormat::Uscaled:
case AmdGpu::NumberFormat::Srgb:
switch (GetNumberClass(fmt)) {
case AmdGpu::NumberClass::Float:
return {id, output ? output_f32 : input_f32, F32[1], num_components, false};
case AmdGpu::NumberFormat::Uint:
case AmdGpu::NumberClass::Uint:
return {id, output ? output_u32 : input_u32, U32[1], num_components, true};
case AmdGpu::NumberFormat::Sint:
case AmdGpu::NumberClass::Sint:
return {id, output ? output_s32 : input_s32, S32[1], num_components, true};
default:
break;
@ -280,33 +269,42 @@ void EmitContext::DefineInputs() {
base_vertex = DefineVariable(U32[1], spv::BuiltIn::BaseVertex, spv::StorageClass::Input);
instance_id = DefineVariable(U32[1], spv::BuiltIn::InstanceIndex, spv::StorageClass::Input);
for (const auto& input : info.vs_inputs) {
ASSERT(input.binding < IR::NumParams);
const Id type{GetAttributeType(*this, input.fmt)[4]};
if (input.instance_step_rate == Info::VsInput::InstanceIdType::OverStepRate0 ||
input.instance_step_rate == Info::VsInput::InstanceIdType::OverStepRate1) {
const auto fetch_shader = Gcn::ParseFetchShader(info);
if (!fetch_shader) {
break;
}
for (const auto& attrib : fetch_shader->attributes) {
ASSERT(attrib.semantic < IR::NumParams);
const auto sharp = attrib.GetSharp(info);
const Id type{GetAttributeType(*this, sharp.GetNumberFmt())[4]};
if (attrib.UsesStepRates()) {
const u32 rate_idx =
input.instance_step_rate == Info::VsInput::InstanceIdType::OverStepRate0 ? 0
: 1;
attrib.GetStepRate() == Gcn::VertexAttribute::InstanceIdType::OverStepRate0 ? 0
: 1;
const u32 num_components = AmdGpu::NumComponents(sharp.GetDataFmt());
const auto buffer =
std::ranges::find_if(info.buffers, [&attrib](const auto& buffer) {
return buffer.instance_attrib == attrib.semantic;
});
// Note that we pass index rather than Id
input_params[input.binding] = SpirvAttribute{
input_params[attrib.semantic] = SpirvAttribute{
.id = rate_idx,
.pointer_type = input_u32,
.component_type = U32[1],
.num_components = input.num_components,
.num_components = std::min<u16>(attrib.num_elements, num_components),
.is_integer = true,
.is_loaded = false,
.buffer_handle = input.instance_data_buf,
.buffer_handle = int(buffer - info.buffers.begin()),
};
} else {
Id id{DefineInput(type, input.binding)};
if (input.instance_step_rate == Info::VsInput::InstanceIdType::Plain) {
Name(id, fmt::format("vs_instance_attr{}", input.binding));
Id id{DefineInput(type, attrib.semantic)};
if (attrib.GetStepRate() == Gcn::VertexAttribute::InstanceIdType::Plain) {
Name(id, fmt::format("vs_instance_attr{}", attrib.semantic));
} else {
Name(id, fmt::format("vs_in_attr{}", input.binding));
Name(id, fmt::format("vs_in_attr{}", attrib.semantic));
}
input_params[input.binding] = GetAttributeInfo(input.fmt, id, 4, false);
input_params[attrib.semantic] =
GetAttributeInfo(sharp.GetNumberFmt(), id, 4, false);
interfaces.push_back(id);
}
}
@ -553,9 +551,10 @@ void EmitContext::DefineBuffers() {
void EmitContext::DefineTextureBuffers() {
for (const auto& desc : info.texture_buffers) {
const bool is_integer =
desc.nfmt == AmdGpu::NumberFormat::Uint || desc.nfmt == AmdGpu::NumberFormat::Sint;
const VectorIds& sampled_type{GetAttributeType(*this, desc.nfmt)};
const auto sharp = desc.GetSharp(info);
const auto nfmt = sharp.GetNumberFmt();
const bool is_integer = AmdGpu::IsInteger(nfmt);
const VectorIds& sampled_type{GetAttributeType(*this, nfmt)};
const u32 sampled = desc.is_written ? 2 : 1;
const Id image_type{TypeImage(sampled_type[1], spv::Dim::Buffer, false, false, false,
sampled, spv::ImageFormat::Unknown)};
@ -650,10 +649,11 @@ spv::ImageFormat GetFormat(const AmdGpu::Image& image) {
}
Id ImageType(EmitContext& ctx, const ImageResource& desc, Id sampled_type) {
const auto image = ctx.info.ReadUdSharp<AmdGpu::Image>(desc.sharp_idx);
const auto image = desc.GetSharp(ctx.info);
const auto format = desc.is_atomic ? GetFormat(image) : spv::ImageFormat::Unknown;
const auto type = image.GetBoundType();
const u32 sampled = desc.is_storage ? 2 : 1;
switch (desc.type) {
switch (type) {
case AmdGpu::ImageType::Color1D:
return ctx.TypeImage(sampled_type, spv::Dim::Dim1D, false, false, false, sampled, format);
case AmdGpu::ImageType::Color1DArray:
@ -672,14 +672,15 @@ Id ImageType(EmitContext& ctx, const ImageResource& desc, Id sampled_type) {
default:
break;
}
throw InvalidArgument("Invalid texture type {}", desc.type);
throw InvalidArgument("Invalid texture type {}", type);
}
void EmitContext::DefineImagesAndSamplers() {
for (const auto& image_desc : info.images) {
const bool is_integer = image_desc.nfmt == AmdGpu::NumberFormat::Uint ||
image_desc.nfmt == AmdGpu::NumberFormat::Sint;
const VectorIds& data_types = GetAttributeType(*this, image_desc.nfmt);
const auto sharp = image_desc.GetSharp(info);
const auto nfmt = sharp.GetNumberFmt();
const bool is_integer = AmdGpu::IsInteger(nfmt);
const VectorIds& data_types = GetAttributeType(*this, nfmt);
const Id sampled_type = data_types[1];
const Id image_type{ImageType(*this, image_desc, sampled_type)};
const Id pointer_type{TypePointer(spv::StorageClass::UniformConstant, image_type)};

12
src/shader_recompiler/frontend/fetch_shader.cpp
View file

@ -34,8 +34,14 @@ namespace Shader::Gcn {
* We take the reverse way, extract the original input semantics from these instructions.
**/
FetchShaderData ParseFetchShader(const u32* code, u32* out_size) {
FetchShaderData data{};
std::optional<FetchShaderData> ParseFetchShader(const Shader::Info& info) {
if (!info.has_fetch_shader) {
return std::nullopt;
}
const u32* code;
std::memcpy(&code, &info.user_data[info.fetch_shader_sgpr_base], sizeof(code));
FetchShaderData data{.code = code};
GcnCodeSlice code_slice(code, code + std::numeric_limits<u32>::max());
GcnDecodeContext decoder;
@ -49,7 +55,7 @@ FetchShaderData ParseFetchShader(const u32* code, u32* out_size) {
u32 semantic_index = 0;
while (!code_slice.atEnd()) {
const auto inst = decoder.decodeInstruction(code_slice);
*out_size += inst.length;
data.size += inst.length;
if (inst.opcode == Opcode::S_SETPC_B64) {
break;

56
src/shader_recompiler/frontend/fetch_shader.h
View file

@ -3,26 +3,80 @@
#pragma once
#include <ranges>
#include <vector>
#include "common/types.h"
#include "shader_recompiler/info.h"
namespace Shader::Gcn {
struct VertexAttribute {
enum InstanceIdType : u8 {
None = 0,
OverStepRate0 = 1,
OverStepRate1 = 2,
Plain = 3,
};
u8 semantic; ///< Semantic index of the attribute
u8 dest_vgpr; ///< Destination VGPR to load first component.
u8 num_elements; ///< Number of components to load
u8 sgpr_base; ///< SGPR that contains the pointer to the list of vertex V#
u8 dword_offset; ///< The dword offset of the V# that describes this attribute.
u8 instance_data; ///< Indicates that the buffer will be accessed in instance rate
[[nodiscard]] InstanceIdType GetStepRate() const {
return static_cast<InstanceIdType>(instance_data);
}
[[nodiscard]] bool UsesStepRates() const {
const auto step_rate = GetStepRate();
return step_rate == OverStepRate0 || step_rate == OverStepRate1;
}
[[nodiscard]] constexpr AmdGpu::Buffer GetSharp(const Shader::Info& info) const noexcept {
return info.ReadUdReg<AmdGpu::Buffer>(sgpr_base, dword_offset);
}
bool operator==(const VertexAttribute& other) const {
return semantic == other.semantic && dest_vgpr == other.dest_vgpr &&
num_elements == other.num_elements && sgpr_base == other.sgpr_base &&
dword_offset == other.dword_offset && instance_data == other.instance_data;
}
};
struct FetchShaderData {
const u32* code;
u32 size = 0;
std::vector<VertexAttribute> attributes;
s8 vertex_offset_sgpr = -1; ///< SGPR of vertex offset from VADDR
s8 instance_offset_sgpr = -1; ///< SGPR of instance offset from VADDR
[[nodiscard]] bool UsesStepRates() const {
return std::ranges::find_if(attributes, [](const VertexAttribute& attribute) {
return attribute.UsesStepRates();
}) != attributes.end();
}
[[nodiscard]] std::pair<u32, u32> GetDrawOffsets(const AmdGpu::Liverpool::Regs& regs,
const Info& info) const {
u32 vertex_offset = regs.index_offset;
u32 instance_offset = 0;
if (vertex_offset == 0 && vertex_offset_sgpr != -1) {
vertex_offset = info.user_data[vertex_offset_sgpr];
}
if (instance_offset_sgpr != -1) {
instance_offset = info.user_data[instance_offset_sgpr];
}
return {vertex_offset, instance_offset};
}
bool operator==(const FetchShaderData& other) const {
return attributes == other.attributes && vertex_offset_sgpr == other.vertex_offset_sgpr &&
instance_offset_sgpr == other.instance_offset_sgpr;
}
};
FetchShaderData ParseFetchShader(const u32* code, u32* out_size);
std::optional<FetchShaderData> ParseFetchShader(const Shader::Info& info);
} // namespace Shader::Gcn

36
src/shader_recompiler/frontend/translate/translate.cpp
View file

@ -368,13 +368,11 @@ void Translator::SetDst64(const InstOperand& operand, const IR::U64F64& value_ra
void Translator::EmitFetch(const GcnInst& inst) {
// Read the pointer to the fetch shader assembly.
const u32 sgpr_base = inst.src[0].code;
const u32* code;
std::memcpy(&code, &info.user_data[sgpr_base], sizeof(code));
info.has_fetch_shader = true;
info.fetch_shader_sgpr_base = inst.src[0].code;
// Parse the assembly to generate a list of attributes.
u32 fetch_size{};
const auto fetch_data = ParseFetchShader(code, &fetch_size);
const auto fetch_data = ParseFetchShader(info);
ASSERT(fetch_data.has_value());
if (Config::dumpShaders()) {
using namespace Common::FS;
@ -384,13 +382,10 @@ void Translator::EmitFetch(const GcnInst& inst) {
}
const auto filename = fmt::format("vs_{:#018x}.fetch.bin", info.pgm_hash);
const auto file = IOFile{dump_dir / filename, FileAccessMode::Write};
file.WriteRaw<u8>(code, fetch_size);
file.WriteRaw<u8>(fetch_data->code, fetch_data->size);
}
info.vertex_offset_sgpr = fetch_data.vertex_offset_sgpr;
info.instance_offset_sgpr = fetch_data.instance_offset_sgpr;
for (const auto& attrib : fetch_data.attributes) {
for (const auto& attrib : fetch_data->attributes) {
const IR::Attribute attr{IR::Attribute::Param0 + attrib.semantic};
IR::VectorReg dst_reg{attrib.dest_vgpr};
@ -420,29 +415,14 @@ void Translator::EmitFetch(const GcnInst& inst) {
// In case of programmable step rates we need to fallback to instance data pulling in
// shader, so VBs should be bound as regular data buffers
s32 instance_buf_handle = -1;
const auto step_rate = static_cast<Info::VsInput::InstanceIdType>(attrib.instance_data);
if (step_rate == Info::VsInput::OverStepRate0 ||
step_rate == Info::VsInput::OverStepRate1) {
if (attrib.UsesStepRates()) {
info.buffers.push_back({
.sharp_idx = info.srt_info.ReserveSharp(attrib.sgpr_base, attrib.dword_offset, 4),
.used_types = IR::Type::F32,
.is_instance_data = true,
.instance_attrib = attrib.semantic,
});
instance_buf_handle = s32(info.buffers.size() - 1);
info.uses_step_rates = true;
}
const u32 num_components = AmdGpu::NumComponents(buffer.GetDataFmt());
info.vs_inputs.push_back({
.fmt = buffer.GetNumberFmt(),
.binding = attrib.semantic,
.num_components = std::min<u16>(attrib.num_elements, num_components),
.sgpr_base = attrib.sgpr_base,
.dword_offset = attrib.dword_offset,
.instance_step_rate = step_rate,
.instance_data_buf = instance_buf_handle,
});
}
}

47
src/shader_recompiler/info.h
View file

@ -45,6 +45,7 @@ struct BufferResource {
AmdGpu::Buffer inline_cbuf;
bool is_gds_buffer{};
bool is_instance_data{};
u8 instance_attrib{};
bool is_written{};
bool IsStorage(AmdGpu::Buffer buffer) const noexcept {
@ -57,7 +58,6 @@ using BufferResourceList = boost::container::small_vector<BufferResource, 16>;
struct TextureBufferResource {
u32 sharp_idx;
AmdGpu::NumberFormat nfmt;
bool is_written{};
constexpr AmdGpu::Buffer GetSharp(const Info& info) const noexcept;
@ -66,8 +66,6 @@ using TextureBufferResourceList = boost::container::small_vector<TextureBufferRe
struct ImageResource {
u32 sharp_idx;
AmdGpu::ImageType type;
AmdGpu::NumberFormat nfmt;
bool is_storage{};
bool is_depth{};
bool is_atomic{};
@ -115,24 +113,6 @@ static_assert(sizeof(PushData) <= 128,
* Contains general information generated by the shader recompiler for an input program.
*/
struct Info {
struct VsInput {
enum InstanceIdType : u8 {
None = 0,
OverStepRate0 = 1,
OverStepRate1 = 2,
Plain = 3,
};
AmdGpu::NumberFormat fmt;
u16 binding;
u16 num_components;
u8 sgpr_base;
u8 dword_offset;
InstanceIdType instance_step_rate;
s32 instance_data_buf;
};
boost::container::static_vector<VsInput, 32> vs_inputs{};
struct AttributeFlags {
bool Get(IR::Attribute attrib, u32 comp = 0) const {
return flags[Index(attrib)] & (1 << comp);
@ -179,9 +159,6 @@ struct Info {
CopyShaderData gs_copy_data;
s8 vertex_offset_sgpr = -1;
s8 instance_offset_sgpr = -1;
BufferResourceList buffers;
TextureBufferResourceList texture_buffers;
ImageResourceList images;
@ -208,10 +185,11 @@ struct Info {
bool uses_shared{};
bool uses_fp16{};
bool uses_fp64{};
bool uses_step_rates{};
bool translation_failed{}; // indicates that shader has unsupported instructions
bool has_readconst{};
u8 mrt_mask{0u};
bool has_fetch_shader{false};
u32 fetch_shader_sgpr_base{0u};
explicit Info(Stage stage_, ShaderParams params)
: stage{stage_}, pgm_hash{params.hash}, pgm_base{params.Base()},
@ -252,18 +230,6 @@ struct Info {
bnd.user_data += ud_mask.NumRegs();
}
[[nodiscard]] std::pair<u32, u32> GetDrawOffsets(const AmdGpu::Liverpool::Regs& regs) const {
u32 vertex_offset = regs.index_offset;
u32 instance_offset = 0;
if (vertex_offset == 0 && vertex_offset_sgpr != -1) {
vertex_offset = user_data[vertex_offset_sgpr];
}
if (instance_offset_sgpr != -1) {
instance_offset = user_data[instance_offset_sgpr];
}
return {vertex_offset, instance_offset};
}
void RefreshFlatBuf() {
flattened_ud_buf.resize(srt_info.flattened_bufsize_dw);
ASSERT(user_data.size() <= NumUserDataRegs);
@ -284,7 +250,12 @@ constexpr AmdGpu::Buffer TextureBufferResource::GetSharp(const Info& info) const
}
constexpr AmdGpu::Image ImageResource::GetSharp(const Info& info) const noexcept {
return info.ReadUdSharp<AmdGpu::Image>(sharp_idx);
const auto image = info.ReadUdSharp<AmdGpu::Image>(sharp_idx);
if (!image.Valid()) {
// Fall back to null image if unbound.
return AmdGpu::Image::Null();
}
return image;
}
constexpr AmdGpu::Sampler SamplerResource::GetSharp(const Info& info) const noexcept {

4
src/shader_recompiler/ir/passes/resource_tracking_pass.cpp
View file

@ -381,7 +381,6 @@ void PatchTextureBufferInstruction(IR::Block& block, IR::Inst& inst, Info& info,
const auto buffer = info.ReadUdSharp<AmdGpu::Buffer>(sharp);
const s32 binding = descriptors.Add(TextureBufferResource{
.sharp_idx = sharp,
.nfmt = buffer.GetNumberFmt(),
.is_written = inst.GetOpcode() == IR::Opcode::StoreBufferFormatF32,
});
@ -660,11 +659,8 @@ void PatchImageInstruction(IR::Block& block, IR::Inst& inst, Info& info, Descrip
}
}
const auto type = image.IsPartialCubemap() ? AmdGpu::ImageType::Color2DArray : image.GetType();
u32 image_binding = descriptors.Add(ImageResource{
.sharp_idx = tsharp,
.type = type,
.nfmt = image.GetNumberFmt(),
.is_storage = is_storage,
.is_depth = bool(inst_info.is_depth),
.is_atomic = IsImageAtomicInstruction(inst),

1
src/shader_recompiler/profile.h
View file

@ -22,6 +22,7 @@ struct Profile {
bool support_fp32_denorm_preserve{};
bool support_fp32_denorm_flush{};
bool support_explicit_workgroup_layout{};
bool support_legacy_vertex_attributes{};
bool has_broken_spirv_clamp{};
bool lower_left_origin_mode{};
bool needs_manual_interpolation{};

45
src/shader_recompiler/specialization.h
View file

@ -6,12 +6,19 @@
#include <bitset>
#include "common/types.h"
#include "frontend/fetch_shader.h"
#include "shader_recompiler/backend/bindings.h"
#include "shader_recompiler/info.h"
#include "shader_recompiler/ir/passes/srt.h"
namespace Shader {
struct VsAttribSpecialization {
AmdGpu::NumberClass num_class{};
auto operator<=>(const VsAttribSpecialization&) const = default;
};
struct BufferSpecialization {
u16 stride : 14;
u16 is_storage : 1;
@ -50,6 +57,8 @@ struct StageSpecialization {
const Shader::Info* info;
RuntimeInfo runtime_info;
Gcn::FetchShaderData fetch_shader_data{};
boost::container::small_vector<VsAttribSpecialization, 32> vs_attribs;
std::bitset<MaxStageResources> bitset{};
boost::container::small_vector<BufferSpecialization, 16> buffers;
boost::container::small_vector<TextureBufferSpecialization, 8> tex_buffers;
@ -57,9 +66,19 @@ struct StageSpecialization {
boost::container::small_vector<FMaskSpecialization, 8> fmasks;
Backend::Bindings start{};
explicit StageSpecialization(const Shader::Info& info_, RuntimeInfo runtime_info_,
Backend::Bindings start_)
explicit StageSpecialization(const Info& info_, RuntimeInfo runtime_info_,
const Profile& profile_, Backend::Bindings start_)
: info{&info_}, runtime_info{runtime_info_}, start{start_} {
if (const auto fetch_shader = Gcn::ParseFetchShader(info_)) {
fetch_shader_data = *fetch_shader;
if (info_.stage == Stage::Vertex && !profile_.support_legacy_vertex_attributes) {
// Specialize shader on VS input number types to follow spec.
ForEachSharp(vs_attribs, fetch_shader_data.attributes,
[](auto& spec, const auto& desc, AmdGpu::Buffer sharp) {
spec.num_class = AmdGpu::GetNumberClass(sharp.GetNumberFmt());
});
}
}
u32 binding{};
if (info->has_readconst) {
binding++;
@ -75,8 +94,7 @@ struct StageSpecialization {
});
ForEachSharp(binding, images, info->images,
[](auto& spec, const auto& desc, AmdGpu::Image sharp) {
spec.type = sharp.IsPartialCubemap() ? AmdGpu::ImageType::Color2DArray
: sharp.GetType();
spec.type = sharp.GetBoundType();
spec.is_integer = AmdGpu::IsInteger(sharp.GetNumberFmt());
});
ForEachSharp(binding, fmasks, info->fmasks,
@ -86,6 +104,17 @@ struct StageSpecialization {
});
}
void ForEachSharp(auto& spec_list, auto& desc_list, auto&& func) {
for (const auto& desc : desc_list) {
auto& spec = spec_list.emplace_back();
const auto sharp = desc.GetSharp(*info);
if (!sharp) {
continue;
}
func(spec, desc, sharp);
}
}
void ForEachSharp(u32& binding, auto& spec_list, auto& desc_list, auto&& func) {
for (const auto& desc : desc_list) {
auto& spec = spec_list.emplace_back();
@ -106,6 +135,14 @@ struct StageSpecialization {
if (runtime_info != other.runtime_info) {
return false;
}
if (fetch_shader_data != other.fetch_shader_data) {
return false;
}
for (u32 i = 0; i < vs_attribs.size(); i++) {
if (vs_attribs[i] != other.vs_attribs[i]) {
return false;
}
}
u32 binding{};
if (info->has_readconst != other.info->has_readconst) {
return false;

19
src/video_core/amdgpu/pixel_format.h
View file

@ -10,7 +10,24 @@
namespace AmdGpu {
[[nodiscard]] constexpr bool IsInteger(NumberFormat nfmt) {
enum NumberClass {
Float,
Sint,
Uint,
};
[[nodiscard]] constexpr NumberClass GetNumberClass(const NumberFormat nfmt) {
switch (nfmt) {
case NumberFormat::Sint:
return Sint;
case NumberFormat::Uint:
return Uint;
default:
return Float;
}
}
[[nodiscard]] constexpr bool IsInteger(const NumberFormat nfmt) {
return nfmt == AmdGpu::NumberFormat::Sint || nfmt == AmdGpu::NumberFormat::Uint;
}

4
src/video_core/amdgpu/resource.h
View file

@ -304,6 +304,10 @@ struct Image {
const auto viewed_slice = last_array - base_array + 1;
return GetType() == ImageType::Cube && viewed_slice < 6;
}
ImageType GetBoundType() const noexcept {
return IsPartialCubemap() ? ImageType::Color2DArray : GetType();
}
};
static_assert(sizeof(Image) == 32); // 256bits

21
src/video_core/buffer_cache/buffer_cache.cpp
View file

@ -5,6 +5,7 @@
#include "common/alignment.h"
#include "common/scope_exit.h"
#include "common/types.h"
#include "shader_recompiler/frontend/fetch_shader.h"
#include "shader_recompiler/info.h"
#include "video_core/amdgpu/liverpool.h"
#include "video_core/buffer_cache/buffer_cache.h"
@ -107,7 +108,8 @@ void BufferCache::DownloadBufferMemory(Buffer& buffer, VAddr device_addr, u64 si
}
}
bool BufferCache::BindVertexBuffers(const Shader::Info& vs_info) {
bool BufferCache::BindVertexBuffers(
const Shader::Info& vs_info, const std::optional<Shader::Gcn::FetchShaderData>& fetch_shader) {
boost::container::small_vector<vk::VertexInputAttributeDescription2EXT, 16> attributes;
boost::container::small_vector<vk::VertexInputBindingDescription2EXT, 16> bindings;
SCOPE_EXIT {
@ -126,7 +128,7 @@ bool BufferCache::BindVertexBuffers(const Shader::Info& vs_info) {
}
};
if (vs_info.vs_inputs.empty()) {
if (!fetch_shader || fetch_shader->attributes.empty()) {
return false;
}
@ -150,30 +152,29 @@ bool BufferCache::BindVertexBuffers(const Shader::Info& vs_info) {
// Calculate buffers memory overlaps
bool has_step_rate = false;
boost::container::static_vector<BufferRange, NumVertexBuffers> ranges{};
for (const auto& input : vs_info.vs_inputs) {
if (input.instance_step_rate == Shader::Info::VsInput::InstanceIdType::OverStepRate0 ||
input.instance_step_rate == Shader::Info::VsInput::InstanceIdType::OverStepRate1) {
for (const auto& attrib : fetch_shader->attributes) {
if (attrib.UsesStepRates()) {
has_step_rate = true;
continue;
}
const auto& buffer = vs_info.ReadUdReg<AmdGpu::Buffer>(input.sgpr_base, input.dword_offset);
const auto& buffer = attrib.GetSharp(vs_info);
if (buffer.GetSize() == 0) {
continue;
}
guest_buffers.emplace_back(buffer);
ranges.emplace_back(buffer.base_address, buffer.base_address + buffer.GetSize());
attributes.push_back({
.location = input.binding,
.binding = input.binding,
.location = attrib.semantic,
.binding = attrib.semantic,
.format =
Vulkan::LiverpoolToVK::SurfaceFormat(buffer.GetDataFmt(), buffer.GetNumberFmt()),
.offset = 0,
});
bindings.push_back({
.binding = input.binding,
.binding = attrib.semantic,
.stride = buffer.GetStride(),
.inputRate = input.instance_step_rate == Shader::Info::VsInput::None
.inputRate = attrib.GetStepRate() == Shader::Gcn::VertexAttribute::InstanceIdType::None
? vk::VertexInputRate::eVertex
: vk::VertexInputRate::eInstance,
.divisor = 1,

8
src/video_core/buffer_cache/buffer_cache.h
View file

@ -20,8 +20,11 @@ struct Liverpool;
}
namespace Shader {
struct Info;
namespace Gcn {
struct FetchShaderData;
}
struct Info;
} // namespace Shader
namespace VideoCore {
@ -76,7 +79,8 @@ public:
void InvalidateMemory(VAddr device_addr, u64 size);
/// Binds host vertex buffers for the current draw.
bool BindVertexBuffers(const Shader::Info& vs_info);
bool BindVertexBuffers(const Shader::Info& vs_info,
const std::optional<Shader::Gcn::FetchShaderData>& fetch_shader);
/// Bind host index buffer for the current draw.
u32 BindIndexBuffer(bool& is_indexed, u32 index_offset);

32
src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp
View file

@ -2,6 +2,7 @@
// SPDX-License-Identifier: GPL-2.0-or-later
#include <algorithm>
#include <utility>
#include <boost/container/small_vector.hpp>
#include <boost/container/static_vector.hpp>
@ -10,6 +11,8 @@
#include "video_core/amdgpu/resource.h"
#include "video_core/buffer_cache/buffer_cache.h"
#include "video_core/renderer_vulkan/vk_graphics_pipeline.h"
#include "shader_recompiler/frontend/fetch_shader.h"
#include "video_core/renderer_vulkan/vk_instance.h"
#include "video_core/renderer_vulkan/vk_scheduler.h"
#include "video_core/texture_cache/texture_cache.h"
@ -20,8 +23,10 @@ GraphicsPipeline::GraphicsPipeline(const Instance& instance_, Scheduler& schedul
DescriptorHeap& desc_heap_, const GraphicsPipelineKey& key_,
vk::PipelineCache pipeline_cache,
std::span<const Shader::Info*, MaxShaderStages> infos,
std::optional<const Shader::Gcn::FetchShaderData> fetch_shader_,
std::span<const vk::ShaderModule> modules)
: Pipeline{instance_, scheduler_, desc_heap_, pipeline_cache}, key{key_} {
: Pipeline{instance_, scheduler_, desc_heap_, pipeline_cache}, key{key_},
fetch_shader{std::move(fetch_shader_)} {
const vk::Device device = instance.GetDevice();
std::ranges::copy(infos, stages.begin());
BuildDescSetLayout();
@ -46,32 +51,31 @@ GraphicsPipeline::GraphicsPipeline(const Instance& instance_, Scheduler& schedul
boost::container::static_vector<vk::VertexInputBindingDescription, 32> vertex_bindings;
boost::container::static_vector<vk::VertexInputAttributeDescription, 32> vertex_attributes;
if (!instance.IsVertexInputDynamicState()) {
const auto& vs_info = stages[u32(Shader::Stage::Vertex)];
for (const auto& input : vs_info->vs_inputs) {
if (input.instance_step_rate == Shader::Info::VsInput::InstanceIdType::OverStepRate0 ||
input.instance_step_rate == Shader::Info::VsInput::InstanceIdType::OverStepRate1) {
if (fetch_shader && !instance.IsVertexInputDynamicState()) {
const auto& vs_info = GetStage(Shader::Stage::Vertex);
for (const auto& attrib : fetch_shader->attributes) {
if (attrib.UsesStepRates()) {
// Skip attribute binding as the data will be pulled by shader
continue;
}
const auto buffer =
vs_info->ReadUdReg<AmdGpu::Buffer>(input.sgpr_base, input.dword_offset);
const auto buffer = attrib.GetSharp(vs_info);
if (buffer.GetSize() == 0) {
continue;
}
vertex_attributes.push_back({
.location = input.binding,
.binding = input.binding,
.location = attrib.semantic,
.binding = attrib.semantic,
.format = LiverpoolToVK::SurfaceFormat(buffer.GetDataFmt(), buffer.GetNumberFmt()),
.offset = 0,
});
vertex_bindings.push_back({
.binding = input.binding,
.binding = attrib.semantic,
.stride = buffer.GetStride(),
.inputRate = input.instance_step_rate == Shader::Info::VsInput::None
? vk::VertexInputRate::eVertex
: vk::VertexInputRate::eInstance,
.inputRate =
attrib.GetStepRate() == Shader::Gcn::VertexAttribute::InstanceIdType::None
? vk::VertexInputRate::eVertex
: vk::VertexInputRate::eInstance,
});
}
}

7
src/video_core/renderer_vulkan/vk_graphics_pipeline.h
View file

@ -4,6 +4,7 @@
#include <xxhash.h>
#include "common/types.h"
#include "shader_recompiler/frontend/fetch_shader.h"
#include "video_core/renderer_vulkan/liverpool_to_vk.h"
#include "video_core/renderer_vulkan/vk_common.h"
#include "video_core/renderer_vulkan/vk_pipeline_common.h"
@ -59,9 +60,14 @@ public:
GraphicsPipeline(const Instance& instance, Scheduler& scheduler, DescriptorHeap& desc_heap,
const GraphicsPipelineKey& key, vk::PipelineCache pipeline_cache,
std::span<const Shader::Info*, MaxShaderStages> stages,
std::optional<const Shader::Gcn::FetchShaderData> fetch_shader,
std::span<const vk::ShaderModule> modules);
~GraphicsPipeline();
const std::optional<const Shader::Gcn::FetchShaderData>& GetFetchShader() const noexcept {
return fetch_shader;
}
bool IsEmbeddedVs() const noexcept {
static constexpr size_t EmbeddedVsHash = 0x9b2da5cf47f8c29f;
return key.stage_hashes[u32(Shader::Stage::Vertex)] == EmbeddedVsHash;
@ -94,6 +100,7 @@ private:
private:
GraphicsPipelineKey key;
std::optional<const Shader::Gcn::FetchShaderData> fetch_shader{};
};
} // namespace Vulkan

7
src/video_core/renderer_vulkan/vk_instance.cpp
View file

@ -265,6 +265,7 @@ bool Instance::CreateDevice() {
const bool robustness = add_extension(VK_EXT_ROBUSTNESS_2_EXTENSION_NAME);
list_restart = add_extension(VK_EXT_PRIMITIVE_TOPOLOGY_LIST_RESTART_EXTENSION_NAME);
maintenance5 = add_extension(VK_KHR_MAINTENANCE_5_EXTENSION_NAME);
legacy_vertex_attributes = add_extension(VK_EXT_LEGACY_VERTEX_ATTRIBUTES_EXTENSION_NAME);
// These extensions are promoted by Vulkan 1.3, but for greater compatibility we use Vulkan 1.2
// with extensions.
@ -403,6 +404,9 @@ bool Instance::CreateDevice() {
vk::PhysicalDeviceFragmentShaderBarycentricFeaturesKHR{
.fragmentShaderBarycentric = true,
},
vk::PhysicalDeviceLegacyVertexAttributesFeaturesEXT{
.legacyVertexAttributes = true,
},
#ifdef __APPLE__
feature_chain.get<vk::PhysicalDevicePortabilitySubsetFeaturesKHR>(),
#endif
@ -445,6 +449,9 @@ bool Instance::CreateDevice() {
if (!fragment_shader_barycentric) {
device_chain.unlink<vk::PhysicalDeviceFragmentShaderBarycentricFeaturesKHR>();
}
if (!legacy_vertex_attributes) {
device_chain.unlink<vk::PhysicalDeviceLegacyVertexAttributesFeaturesEXT>();
}
auto [device_result, dev] = physical_device.createDeviceUnique(device_chain.get());
if (device_result != vk::Result::eSuccess) {

7
src/video_core/renderer_vulkan/vk_instance.h
View file

@ -148,10 +148,16 @@ public:
return fragment_shader_barycentric;
}
/// Returns true when VK_EXT_primitive_topology_list_restart is supported.
bool IsListRestartSupported() const {
return list_restart;
}
/// Returns true when VK_EXT_legacy_vertex_attributes is supported.
bool IsLegacyVertexAttributesSupported() const {
return legacy_vertex_attributes;
}
/// Returns true when geometry shaders are supported by the device
bool IsGeometryStageSupported() const {
return features.geometryShader;
@ -320,6 +326,7 @@ private:
bool null_descriptor{};
bool maintenance5{};
bool list_restart{};
bool legacy_vertex_attributes{};
u64 min_imported_host_pointer_alignment{};
u32 subgroup_size{};
bool tooling_info{};

40
src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
View file

@ -169,6 +169,7 @@ PipelineCache::PipelineCache(const Instance& instance_, Scheduler& scheduler_,
.support_fp32_denorm_preserve = bool(vk12_props.shaderDenormPreserveFloat32),
.support_fp32_denorm_flush = bool(vk12_props.shaderDenormFlushToZeroFloat32),
.support_explicit_workgroup_layout = true,
.support_legacy_vertex_attributes = instance_.IsLegacyVertexAttributesSupported(),
.needs_manual_interpolation = instance.IsFragmentShaderBarycentricSupported() &&
instance.GetDriverID() == vk::DriverId::eNvidiaProprietary,
};
@ -187,7 +188,7 @@ const GraphicsPipeline* PipelineCache::GetGraphicsPipeline() {
const auto [it, is_new] = graphics_pipelines.try_emplace(graphics_key);
if (is_new) {
it.value() = graphics_pipeline_pool.Create(instance, scheduler, desc_heap, graphics_key,
*pipeline_cache, infos, modules);
*pipeline_cache, infos, fetch_shader, modules);
}
return it->second;
}
@ -304,8 +305,12 @@ bool PipelineCache::RefreshGraphicsKey() {
}
auto params = Liverpool::GetParams(*pgm);
std::tie(infos[stage_out_idx], modules[stage_out_idx], key.stage_hashes[stage_out_idx]) =
GetProgram(stage_in, params, binding);
std::optional<Shader::Gcn::FetchShaderData> fetch_shader_;
std::tie(infos[stage_out_idx], modules[stage_out_idx], fetch_shader_,
key.stage_hashes[stage_out_idx]) = GetProgram(stage_in, params, binding);
if (fetch_shader_) {
fetch_shader = fetch_shader_;
}
return true;
};
@ -341,16 +346,14 @@ bool PipelineCache::RefreshGraphicsKey() {
}
}
const auto* vs_info = infos[static_cast<u32>(Shader::Stage::Vertex)];
if (vs_info && !instance.IsVertexInputDynamicState()) {
const auto vs_info = infos[static_cast<u32>(Shader::Stage::Vertex)];
if (vs_info && fetch_shader && !instance.IsVertexInputDynamicState()) {
u32 vertex_binding = 0;
for (const auto& input : vs_info->vs_inputs) {
if (input.instance_step_rate == Shader::Info::VsInput::InstanceIdType::OverStepRate0 ||
input.instance_step_rate == Shader::Info::VsInput::InstanceIdType::OverStepRate1) {
for (const auto& attrib : fetch_shader->attributes) {
if (attrib.UsesStepRates()) {
continue;
}
const auto& buffer =
vs_info->ReadUdReg<AmdGpu::Buffer>(input.sgpr_base, input.dword_offset);
const auto& buffer = attrib.GetSharp(*vs_info);
if (buffer.GetSize() == 0) {
continue;
}
@ -394,7 +397,7 @@ bool PipelineCache::RefreshComputeKey() {
Shader::Backend::Bindings binding{};
const auto* cs_pgm = &liverpool->regs.cs_program;
const auto cs_params = Liverpool::GetParams(*cs_pgm);
std::tie(infos[0], modules[0], compute_key) =
std::tie(infos[0], modules[0], fetch_shader, compute_key) =
GetProgram(Shader::Stage::Compute, cs_params, binding);
return true;
}
@ -425,24 +428,26 @@ vk::ShaderModule PipelineCache::CompileModule(Shader::Info& info,
return module;
}
std::tuple<const Shader::Info*, vk::ShaderModule, u64> PipelineCache::GetProgram(
Shader::Stage stage, Shader::ShaderParams params, Shader::Backend::Bindings& binding) {
std::tuple<const Shader::Info*, vk::ShaderModule, std::optional<Shader::Gcn::FetchShaderData>, u64>
PipelineCache::GetProgram(Shader::Stage stage, Shader::ShaderParams params,
Shader::Backend::Bindings& binding) {
const auto runtime_info = BuildRuntimeInfo(stage);
auto [it_pgm, new_program] = program_cache.try_emplace(params.hash);
if (new_program) {
Program* program = program_pool.Create(stage, params);
auto start = binding;
const auto module = CompileModule(program->info, runtime_info, params.code, 0, binding);
const auto spec = Shader::StageSpecialization(program->info, runtime_info, start);
const auto spec = Shader::StageSpecialization(program->info, runtime_info, profile, start);
program->AddPermut(module, std::move(spec));
it_pgm.value() = program;
return std::make_tuple(&program->info, module, HashCombine(params.hash, 0));
return std::make_tuple(&program->info, module, spec.fetch_shader_data,
HashCombine(params.hash, 0));
}
Program* program = it_pgm->second;
auto& info = program->info;
info.RefreshFlatBuf();
const auto spec = Shader::StageSpecialization(info, runtime_info, binding);
const auto spec = Shader::StageSpecialization(info, runtime_info, profile, binding);
size_t perm_idx = program->modules.size();
vk::ShaderModule module{};
@ -456,7 +461,8 @@ std::tuple<const Shader::Info*, vk::ShaderModule, u64> PipelineCache::GetProgram
module = it->module;
perm_idx = std::distance(program->modules.begin(), it);
}
return std::make_tuple(&info, module, HashCombine(params.hash, perm_idx));
return std::make_tuple(&info, module, spec.fetch_shader_data,
HashCombine(params.hash, perm_idx));
}
void PipelineCache::DumpShader(std::span<const u32> code, u64 hash, Shader::Stage stage,

7
src/video_core/renderer_vulkan/vk_pipeline_cache.h
View file

@ -47,8 +47,10 @@ public:
const ComputePipeline* GetComputePipeline();
std::tuple<const Shader::Info*, vk::ShaderModule, u64> GetProgram(
Shader::Stage stage, Shader::ShaderParams params, Shader::Backend::Bindings& binding);
std::tuple<const Shader::Info*, vk::ShaderModule, std::optional<Shader::Gcn::FetchShaderData>,
u64>
GetProgram(Shader::Stage stage, Shader::ShaderParams params,
Shader::Backend::Bindings& binding);
private:
bool RefreshGraphicsKey();
@ -80,6 +82,7 @@ private:
tsl::robin_map<GraphicsPipelineKey, GraphicsPipeline*> graphics_pipelines;
std::array<const Shader::Info*, MaxShaderStages> infos{};
std::array<vk::ShaderModule, MaxShaderStages> modules{};
std::optional<Shader::Gcn::FetchShaderData> fetch_shader{};
GraphicsPipelineKey graphics_key{};
u64 compute_key{};
};

14
src/video_core/renderer_vulkan/vk_rasterizer.cpp
View file

@ -187,13 +187,14 @@ void Rasterizer::Draw(bool is_indexed, u32 index_offset) {
}
const auto& vs_info = pipeline->GetStage(Shader::Stage::Vertex);
buffer_cache.BindVertexBuffers(vs_info);
const auto& fetch_shader = pipeline->GetFetchShader();
buffer_cache.BindVertexBuffers(vs_info, fetch_shader);
const u32 num_indices = buffer_cache.BindIndexBuffer(is_indexed, index_offset);
BeginRendering(*pipeline, state);
UpdateDynamicState(*pipeline);
const auto [vertex_offset, instance_offset] = vs_info.GetDrawOffsets(regs);
const auto [vertex_offset, instance_offset] = fetch_shader->GetDrawOffsets(regs, vs_info);
const auto cmdbuf = scheduler.CommandBuffer();
cmdbuf.bindPipeline(vk::PipelineBindPoint::eGraphics, pipeline->Handle());
@ -243,7 +244,8 @@ void Rasterizer::DrawIndirect(bool is_indexed, VAddr arg_address, u32 offset, u3
}
const auto& vs_info = pipeline->GetStage(Shader::Stage::Vertex);
buffer_cache.BindVertexBuffers(vs_info);
const auto& fetch_shader = pipeline->GetFetchShader();
buffer_cache.BindVertexBuffers(vs_info, fetch_shader);
buffer_cache.BindIndexBuffer(is_indexed, 0);
const auto& [buffer, base] =
@ -397,10 +399,8 @@ bool Rasterizer::BindResources(const Pipeline* pipeline) {
if (!stage) {
continue;
}
if (stage->uses_step_rates) {
push_data.step0 = regs.vgt_instance_step_rate_0;
push_data.step1 = regs.vgt_instance_step_rate_1;
}
push_data.step0 = regs.vgt_instance_step_rate_0;
push_data.step1 = regs.vgt_instance_step_rate_1;
stage->PushUd(binding, push_data);
BindBuffers(*stage, binding, push_data, set_writes, buffer_barriers);

7
src/video_core/texture_cache/image_view.cpp
View file

@ -87,12 +87,9 @@ ImageViewInfo::ImageViewInfo(const AmdGpu::Image& image, const Shader::ImageReso
range.extent.levels = image.last_level - image.base_level + 1;
}
range.extent.layers = image.last_array - image.base_array + 1;
type = ConvertImageViewType(image.GetType());
type = ConvertImageViewType(image.GetBoundType());
// Adjust view type for partial cubemaps and arrays
if (image.IsPartialCubemap()) {
type = vk::ImageViewType::e2DArray;
}
// Adjust view type for arrays
if (type == vk::ImageViewType::eCube) {
if (desc.is_array) {
type = vk::ImageViewType::eCubeArray;