sneedmca/Lime3DS
1
0
Fork 0
mirror of https://github.com/Lime3DS/Lime3DS synced 2024-10-30 19:57:52 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 7

spv_fs_shader_gen: Implement quaternion correction with barycentric extension (#7152)

Browse source
This commit is contained in:
GPUCode 2023-11-13 08:40:21 +02:00 committed by GitHub
parent 312068eebf
commit 168f168c33
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
3 changed files with 79 additions and 12 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
View file

@ -472,7 +472,7 @@ void PipelineCache::UseFragmentShader(const Pica::Regs& regs,
if (new_shader) {
const bool use_spirv = Settings::values.spirv_shader_gen.GetValue();
if (use_spirv && !fs_config.UsesShadowPipeline()) {
const std::vector code = SPIRV::GenerateFragmentShader(fs_config);
const std::vector code = SPIRV::GenerateFragmentShader(fs_config, profile);
shader.module = CompileSPV(code, instance.GetDevice());
shader.MarkDone();
} else {

80
src/video_core/shader/generator/spv_fs_shader_gen.cpp
View file

@ -2,6 +2,7 @@
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#include <boost/container/small_vector.hpp>
#include "video_core/shader/generator/spv_fs_shader_gen.h"
namespace Pica::Shader::Generator::SPIRV {
@ -14,8 +15,10 @@ using TevStageConfig = TexturingRegs::TevStageConfig;
constexpr u32 SPIRV_VERSION_1_3 = 0x00010300;
FragmentModule::FragmentModule(const FSConfig& config_)
: Sirit::Module{SPIRV_VERSION_1_3}, config{config_} {
FragmentModule::FragmentModule(const FSConfig& config_, const Profile& profile_)
: Sirit::Module{SPIRV_VERSION_1_3}, config{config_}, profile{profile_},
use_fragment_shader_barycentric{profile.has_fragment_shader_barycentric &&
config.lighting.enable} {
DefineArithmeticTypes();
DefineUniformStructs();
DefineInterface();
@ -270,9 +273,48 @@ void FragmentModule::WriteLighting() {
return OpFAdd(vec_ids.Get(3), v, val2);
};
// Perform quaternion correction in the fragment shader if fragment_shader_barycentric is
// supported.
Id normquat{};
if (use_fragment_shader_barycentric) {
const auto are_quaternions_opposite = [&](Id qa, Id qb) {
const Id dot_q{OpDot(f32_id, qa, qb)};
const Id dot_le_zero{OpFOrdLessThan(bool_id, dot_q, ConstF32(0.f))};
return OpCompositeConstruct(bvec_ids.Get(4), dot_le_zero, dot_le_zero, dot_le_zero,
dot_le_zero);
};
const Id input_pointer_id{TypePointer(spv::StorageClass::Input, vec_ids.Get(4))};
const Id normquat_0{
OpLoad(vec_ids.Get(4), OpAccessChain(input_pointer_id, normquat_id, ConstS32(0)))};
const Id normquat_1{
OpLoad(vec_ids.Get(4), OpAccessChain(input_pointer_id, normquat_id, ConstS32(1)))};
const Id normquat_1_correct{OpSelect(vec_ids.Get(4),
are_quaternions_opposite(normquat_0, normquat_1),
OpFNegate(vec_ids.Get(4), normquat_1), normquat_1)};
const Id normquat_2{
OpLoad(vec_ids.Get(4), OpAccessChain(input_pointer_id, normquat_id, ConstS32(2)))};
const Id normquat_2_correct{OpSelect(vec_ids.Get(4),
are_quaternions_opposite(normquat_0, normquat_2),
OpFNegate(vec_ids.Get(4), normquat_2), normquat_2)};
const Id bary_coord{OpLoad(vec_ids.Get(3), gl_bary_coord_id)};
const Id bary_coord_x{OpCompositeExtract(f32_id, bary_coord, 0)};
const Id bary_coord_y{OpCompositeExtract(f32_id, bary_coord, 1)};
const Id bary_coord_z{OpCompositeExtract(f32_id, bary_coord, 2)};
const Id normquat_0_final{OpVectorTimesScalar(vec_ids.Get(4), normquat_0, bary_coord_x)};
const Id normquat_1_final{
OpVectorTimesScalar(vec_ids.Get(4), normquat_1_correct, bary_coord_y)};
const Id normquat_2_final{
OpVectorTimesScalar(vec_ids.Get(4), normquat_2_correct, bary_coord_z)};
normquat = OpFAdd(vec_ids.Get(4), normquat_0_final,
OpFAdd(vec_ids.Get(4), normquat_1_final, normquat_2_final));
} else {
normquat = OpLoad(vec_ids.Get(4), normquat_id);
}
// Rotate the surface-local normal by the interpolated normal quaternion to convert it to
// eyespace.
const Id normalized_normquat{OpNormalize(vec_ids.Get(4), OpLoad(vec_ids.Get(4), normquat_id))};
const Id normalized_normquat{OpNormalize(vec_ids.Get(4), normquat)};
const Id normal{quaternion_rotate(normalized_normquat, surface_normal)};
const Id tangent{quaternion_rotate(normalized_normquat, surface_tangent)};
@ -1452,13 +1494,24 @@ void FragmentModule::DefineEntryPoint() {
AddCapability(spv::Capability::Shader);
AddCapability(spv::Capability::SampledBuffer);
AddCapability(spv::Capability::ImageQuery);
if (use_fragment_shader_barycentric) {
AddCapability(spv::Capability::FragmentBarycentricKHR);
AddExtension("SPV_KHR_fragment_shader_barycentric");
}
SetMemoryModel(spv::AddressingModel::Logical, spv::MemoryModel::GLSL450);
const Id main_type{TypeFunction(TypeVoid())};
const Id main_func{OpFunction(TypeVoid(), spv::FunctionControlMask::MaskNone, main_type)};
AddEntryPoint(spv::ExecutionModel::Fragment, main_func, "main", primary_color_id,
texcoord_id[0], texcoord_id[1], texcoord_id[2], texcoord0_w_id, normquat_id,
view_id, color_id, gl_frag_coord_id, gl_frag_depth_id);
boost::container::small_vector<Id, 11> interface_ids{
primary_color_id, texcoord_id[0], texcoord_id[1], texcoord_id[2], texcoord0_w_id,
normquat_id, view_id, color_id, gl_frag_coord_id, gl_frag_depth_id,
};
if (use_fragment_shader_barycentric) {
interface_ids.push_back(gl_bary_coord_id);
}
AddEntryPoint(spv::ExecutionModel::Fragment, main_func, "main", interface_ids);
AddExecutionMode(main_func, spv::ExecutionMode::OriginUpperLeft);
AddExecutionMode(main_func, spv::ExecutionMode::DepthReplacing);
}
@ -1510,7 +1563,12 @@ void FragmentModule::DefineInterface() {
texcoord_id[1] = DefineInput(vec_ids.Get(2), 3);
texcoord_id[2] = DefineInput(vec_ids.Get(2), 4);
texcoord0_w_id = DefineInput(f32_id, 5);
normquat_id = DefineInput(vec_ids.Get(4), 6);
if (use_fragment_shader_barycentric) {
normquat_id = DefineInput(TypeArray(vec_ids.Get(4), ConstU32(3U)), 6);
Decorate(normquat_id, spv::Decoration::PerVertexKHR);
} else {
normquat_id = DefineInput(vec_ids.Get(4), 6);
}
view_id = DefineInput(vec_ids.Get(3), 7);
color_id = DefineOutput(vec_ids.Get(4), 0);
@ -1537,10 +1595,14 @@ void FragmentModule::DefineInterface() {
gl_frag_depth_id = DefineVar(f32_id, spv::StorageClass::Output);
Decorate(gl_frag_coord_id, spv::Decoration::BuiltIn, spv::BuiltIn::FragCoord);
Decorate(gl_frag_depth_id, spv::Decoration::BuiltIn, spv::BuiltIn::FragDepth);
if (use_fragment_shader_barycentric) {
gl_bary_coord_id = DefineVar(vec_ids.Get(3), spv::StorageClass::Input);
Decorate(gl_bary_coord_id, spv::Decoration::BuiltIn, spv::BuiltIn::BaryCoordKHR);
}
}
std::vector<u32> GenerateFragmentShader(const FSConfig& config) {
FragmentModule module{config};
std::vector<u32> GenerateFragmentShader(const FSConfig& config, const Profile& profile) {
FragmentModule module{config, profile};
module.Generate();
return module.Assemble();
}

9
src/video_core/shader/generator/spv_fs_shader_gen.h
View file

@ -30,7 +30,7 @@ class FragmentModule : public Sirit::Module {
static constexpr u32 NUM_NON_PROC_TEX_UNITS = 3;
public:
explicit FragmentModule(const FSConfig& config);
explicit FragmentModule(const FSConfig& config, const Profile& profile);
~FragmentModule();
/// Emits SPIR-V bytecode corresponding to the provided pica fragment configuration
@ -215,6 +215,10 @@ private:
private:
const FSConfig& config;
const Profile& profile;
bool use_fragment_shader_barycentric{};
Id void_id{};
Id bool_id{};
Id f32_id{};
@ -242,6 +246,7 @@ private:
Id gl_frag_coord_id{};
Id gl_frag_depth_id{};
Id gl_bary_coord_id{};
Id depth{};
Id tex0_id{};
@ -284,6 +289,6 @@ private:
* @param separable_shader generates shader that can be used for separate shader object
* @returns String of the shader source code
*/
std::vector<u32> GenerateFragmentShader(const FSConfig& config);
std::vector<u32> GenerateFragmentShader(const FSConfig& config, const Profile& profile);
} // namespace Pica::Shader::Generator::SPIRV