ShuriZma
/
xenia
mirror of https://github.com/xenia-project/xenia.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

[Vulkan] PsParamGen

This commit is contained in:
Triang3l 2022-06-26 15:01:27 +03:00
parent a99a1be880
commit 6688b13773
5 changed files with 212 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

166
src/xenia/gpu/spirv_shader_translator.cc
View File

@ -106,6 +106,9 @@ void SpirvShaderTranslator::Reset() {
uniform_float_constants_ = spv::NoResult;
input_fragment_coord_ = spv::NoResult;
input_front_facing_ = spv::NoResult;
sampler_bindings_.clear();
texture_bindings_.clear();
@ -1011,6 +1014,17 @@ spv::Id SpirvShaderTranslator::SpirvSmearScalarResultOrConstant(
is_spec_constant);
}
uint32_t SpirvShaderTranslator::GetPsParamGenInterpolator() const {
assert_true(is_pixel_shader());
Modification modification = GetSpirvShaderModification();
// param_gen_interpolator is already 4 bits, no need for an interpolator count
// safety check.
return (modification.pixel.param_gen_enable &&
modification.pixel.param_gen_interpolator < register_count())
? modification.pixel.param_gen_interpolator
: UINT32_MAX;
}
void SpirvShaderTranslator::EnsureBuildPointAvailable() {
if (!builder_->getBuildPoint()->isTerminated()) {
return;
@ -1261,6 +1275,31 @@ void SpirvShaderTranslator::StartFragmentShaderBeforeMain() {
main_interface_.push_back(interpolator);
}
bool param_gen_needed = GetPsParamGenInterpolator() != UINT32_MAX;
// Fragment coordinates.
// TODO(Triang3l): More conditions - fragment shader interlock render backend,
// alpha to coverage (if RT 0 is written, and there's no early depth /
// stencil), depth writing in the fragment shader (per-sample if supported).
if (param_gen_needed) {
input_fragment_coord_ = builder_->createVariable(
spv::NoPrecision, spv::StorageClassInput, type_float4_, "gl_FragCoord");
builder_->addDecoration(input_fragment_coord_, spv::DecorationBuiltIn,
spv::BuiltInFragCoord);
main_interface_.push_back(input_fragment_coord_);
}
// Is front facing.
// TODO(Triang3l): Needed for stencil in the fragment shader interlock render
// backend.
if (param_gen_needed && !GetSpirvShaderModification().pixel.param_gen_point) {
input_front_facing_ = builder_->createVariable(
spv::NoPrecision, spv::StorageClassInput, type_bool_, "gl_FrontFacing");
builder_->addDecoration(input_front_facing_, spv::DecorationBuiltIn,
spv::BuiltInFrontFacing);
main_interface_.push_back(input_front_facing_);
}
// Framebuffer attachment outputs.
std::fill(output_fragment_data_.begin(), output_fragment_data_.end(),
spv::NoResult);
@ -1288,12 +1327,16 @@ void SpirvShaderTranslator::StartFragmentShaderBeforeMain() {
}
void SpirvShaderTranslator::StartFragmentShaderInMain() {
uint32_t param_gen_interpolator = GetPsParamGenInterpolator();
// Copy the interpolators to general-purpose registers.
// TODO(Triang3l): Centroid.
// TODO(Triang3l): ps_param_gen.
uint32_t interpolator_count =
std::min(xenos::kMaxInterpolators, register_count());
for (uint32_t i = 0; i < interpolator_count; ++i) {
if (i == param_gen_interpolator) {
continue;
}
id_vector_temp_.clear();
// Register array element.
id_vector_temp_.push_back(builder_->makeIntConstant(int(i)));
@ -1303,6 +1346,127 @@ void SpirvShaderTranslator::StartFragmentShaderInMain() {
var_main_registers_, id_vector_temp_));
}
// Pixel parameters.
if (param_gen_interpolator != UINT32_MAX) {
Modification modification = GetSpirvShaderModification();
// Rounding the position down, and taking the absolute value, so in case the
// host GPU for some reason has quads used for derivative calculation at odd
// locations, the left and top edges will have correct derivative magnitude
// and LODs.
// Assuming that if PsParamGen is needed at all, param_gen_point is always
// set for point primitives, and is always disabled for other primitive
// types.
// OpFNegate requires sign bit flipping even for 0.0 (in this case, the
// first column or row of pixels) only since SPIR-V 1.5 revision 2 (not the
// base 1.5).
// TODO(Triang3l): When SPIR-V 1.6 is used in Xenia, see if OpFNegate can be
// used there, should be cheaper because it may be implemented as a hardware
// instruction modifier, though it respects the rule for subnormal numbers -
// see the actual hardware instructions in both OpBitwiseXor and OpFNegate
// cases.
spv::Id const_sign_bit = builder_->makeUintConstant(UINT32_C(1) << 31);
// TODO(Triang3l): Resolution scale inversion.
// X - pixel X .0 in the magnitude, is back-facing in the sign bit.
assert_true(input_fragment_coord_ != spv::NoResult);
id_vector_temp_.clear();
id_vector_temp_.push_back(const_int_0_);
spv::Id param_gen_x = builder_->createLoad(
builder_->createAccessChain(spv::StorageClassInput,
input_fragment_coord_, id_vector_temp_),
spv::NoPrecision);
id_vector_temp_.clear();
id_vector_temp_.push_back(param_gen_x);
param_gen_x = builder_->createBuiltinCall(
type_float_, ext_inst_glsl_std_450_, GLSLstd450Floor, id_vector_temp_);
id_vector_temp_.clear();
id_vector_temp_.push_back(param_gen_x);
param_gen_x = builder_->createBuiltinCall(
type_float_, ext_inst_glsl_std_450_, GLSLstd450FAbs, id_vector_temp_);
if (!modification.pixel.param_gen_point) {
assert_true(input_front_facing_ != spv::NoResult);
param_gen_x = builder_->createTriOp(
spv::OpSelect, type_float_,
builder_->createBinOp(
spv::OpLogicalOr, type_bool_,
builder_->createBinOp(
spv::OpIEqual, type_bool_,
builder_->createBinOp(
spv::OpBitwiseAnd, type_uint_,
main_system_constant_flags_,
builder_->makeUintConstant(kSysFlag_PrimitivePolygonal)),
const_uint_0_),
builder_->createLoad(input_front_facing_, spv::NoPrecision)),
param_gen_x,
builder_->createUnaryOp(
spv::OpBitcast, type_float_,
builder_->createBinOp(
spv::OpBitwiseXor, type_uint_,
builder_->createUnaryOp(spv::OpBitcast, type_uint_,
param_gen_x),
const_sign_bit)));
}
// Y - pixel Y .0 in the magnitude, is point in the sign bit.
id_vector_temp_.clear();
id_vector_temp_.push_back(builder_->makeIntConstant(1));
spv::Id param_gen_y = builder_->createLoad(
builder_->createAccessChain(spv::StorageClassInput,
input_fragment_coord_, id_vector_temp_),
spv::NoPrecision);
id_vector_temp_.clear();
id_vector_temp_.push_back(param_gen_y);
param_gen_y = builder_->createBuiltinCall(
type_float_, ext_inst_glsl_std_450_, GLSLstd450Floor, id_vector_temp_);
id_vector_temp_.clear();
id_vector_temp_.push_back(param_gen_y);
param_gen_y = builder_->createBuiltinCall(
type_float_, ext_inst_glsl_std_450_, GLSLstd450FAbs, id_vector_temp_);
if (modification.pixel.param_gen_point) {
param_gen_y = builder_->createUnaryOp(
spv::OpBitcast, type_float_,
builder_->createBinOp(
spv::OpBitwiseXor, type_uint_,
builder_->createUnaryOp(spv::OpBitcast, type_uint_, param_gen_y),
const_sign_bit));
}
// Z - point S in the magnitude, is line in the sign bit.
spv::Id param_gen_z;
if (modification.pixel.param_gen_point) {
// TODO(Triang3l): Point coordinates.
param_gen_z = const_float_0_;
} else {
param_gen_z = builder_->createUnaryOp(
spv::OpBitcast, type_float_,
builder_->createTriOp(
spv::OpSelect, type_uint_,
builder_->createBinOp(
spv::OpINotEqual, type_bool_,
builder_->createBinOp(
spv::OpBitwiseAnd, type_uint_,
main_system_constant_flags_,
builder_->makeUintConstant(kSysFlag_PrimitiveLine)),
const_uint_0_),
const_sign_bit, const_uint_0_));
}
// W - point T in the magnitude.
// TODO(Triang3l): Point coordinates.
spv::Id param_gen_w = const_float_0_;
// Store the pixel parameters.
id_vector_temp_.clear();
id_vector_temp_.reserve(4);
id_vector_temp_.push_back(param_gen_x);
id_vector_temp_.push_back(param_gen_y);
id_vector_temp_.push_back(param_gen_z);
id_vector_temp_.push_back(param_gen_w);
spv::Id param_gen =
builder_->createCompositeConstruct(type_float4_, id_vector_temp_);
id_vector_temp_.clear();
id_vector_temp_.push_back(
builder_->makeIntConstant(int(param_gen_interpolator)));
builder_->createStore(param_gen, builder_->createAccessChain(
spv::StorageClassFunction,
var_main_registers_, id_vector_temp_));
}
// Initialize the colors for safety.
for (uint32_t i = 0; i < xenos::kMaxColorRenderTargets; ++i) {
spv::Id output_fragment_data_rt = output_fragment_data_[i];

17
src/xenia/gpu/spirv_shader_translator.h
View File

@ -46,6 +46,12 @@ class SpirvShaderTranslator : public ShaderTranslator {
struct PixelShaderModification {
// Dynamically indexable register count from SQ_PROGRAM_CNTL.
uint32_t dynamic_addressable_register_count : 8;
uint32_t param_gen_enable : 1;
uint32_t param_gen_interpolator : 4;
// If param_gen_enable is set, this must be set for point primitives, and
// must not be set for other primitive types - enables the point sprite
// coordinates input, and also effects the flag bits in PsParamGen.
uint32_t param_gen_point : 1;
} pixel;
uint64_t value = 0;
@ -56,6 +62,8 @@ class SpirvShaderTranslator : public ShaderTranslator {
kSysFlag_XYDividedByW_Shift,
kSysFlag_ZDividedByW_Shift,
kSysFlag_WNotReciprocal_Shift,
kSysFlag_PrimitivePolygonal_Shift,
kSysFlag_PrimitiveLine_Shift,
kSysFlag_ConvertColor0ToGamma_Shift,
kSysFlag_ConvertColor1ToGamma_Shift,
kSysFlag_ConvertColor2ToGamma_Shift,
@ -66,6 +74,8 @@ class SpirvShaderTranslator : public ShaderTranslator {
kSysFlag_XYDividedByW = 1u << kSysFlag_XYDividedByW_Shift,
kSysFlag_ZDividedByW = 1u << kSysFlag_ZDividedByW_Shift,
kSysFlag_WNotReciprocal = 1u << kSysFlag_WNotReciprocal_Shift,
kSysFlag_PrimitivePolygonal = 1u << kSysFlag_PrimitivePolygonal_Shift,
kSysFlag_PrimitiveLine = 1u << kSysFlag_PrimitiveLine_Shift,
kSysFlag_ConvertColor0ToGamma = 1u << kSysFlag_ConvertColor0ToGamma_Shift,
kSysFlag_ConvertColor1ToGamma = 1u << kSysFlag_ConvertColor1ToGamma_Shift,
kSysFlag_ConvertColor2ToGamma = 1u << kSysFlag_ConvertColor2ToGamma_Shift,
@ -301,6 +311,9 @@ class SpirvShaderTranslator : public ShaderTranslator {
GetSpirvShaderModification().vertex.host_vertex_shader_type);
}
// Returns UINT32_MAX if PsParamGen doesn't need to be written.
uint32_t GetPsParamGenInterpolator() const;
// Must be called before emitting any SPIR-V operations that must be in a
// block in translator callbacks to ensure that if the last instruction added
// was something like OpBranch - in this case, an unreachable block is
@ -535,6 +548,10 @@ class SpirvShaderTranslator : public ShaderTranslator {
spv::Id input_vertex_index_;
// VS as TES only - int.
spv::Id input_primitive_id_;
// PS, only when needed - float4.
spv::Id input_fragment_coord_;
// PS, only when needed - bool.
spv::Id input_front_facing_;
// In vertex or tessellation evaluation shaders - outputs, always
// xenos::kMaxInterpolators.

16
src/xenia/gpu/vulkan/vulkan_command_processor.cc
View File

@ -2336,7 +2336,8 @@ bool VulkanCommandProcessor::IssueDraw(xenos::PrimitiveType prim_type,
normalized_depth_control);
// Update system constants before uploading them.
UpdateSystemConstantValues(primitive_processing_result.host_index_endian,
UpdateSystemConstantValues(primitive_polygonal,
primitive_processing_result.host_index_endian,
viewport_info, used_texture_mask);
// Update uniform buffers and descriptor sets after binding the pipeline with
@ -3257,14 +3258,15 @@ void VulkanCommandProcessor::UpdateDynamicState(
}
void VulkanCommandProcessor::UpdateSystemConstantValues(
xenos::Endian index_endian, const draw_util::ViewportInfo& viewport_info,
uint32_t used_texture_mask) {
bool primitive_polygonal, xenos::Endian index_endian,
const draw_util::ViewportInfo& viewport_info, uint32_t used_texture_mask) {
#if XE_UI_VULKAN_FINE_GRAINED_DRAW_SCOPES
SCOPE_profile_cpu_f("gpu");
#endif // XE_UI_VULKAN_FINE_GRAINED_DRAW_SCOPES
const RegisterFile& regs = *register_file_;
auto pa_cl_vte_cntl = regs.Get<reg::PA_CL_VTE_CNTL>();
auto vgt_draw_initiator = regs.Get<reg::VGT_DRAW_INITIATOR>();
int32_t vgt_indx_offset = int32_t(regs[XE_GPU_REG_VGT_INDX_OFFSET].u32);
// Get the color info register values for each render target.
@ -3295,6 +3297,14 @@ void VulkanCommandProcessor::UpdateSystemConstantValues(
if (pa_cl_vte_cntl.vtx_w0_fmt) {
flags |= SpirvShaderTranslator::kSysFlag_WNotReciprocal;
}
// Whether the primitive is polygonal, and gl_FrontFacing matters.
if (primitive_polygonal) {
flags |= SpirvShaderTranslator::kSysFlag_PrimitivePolygonal;
}
// Primitive type.
if (draw_util::IsPrimitiveLine(regs)) {
flags |= SpirvShaderTranslator::kSysFlag_PrimitiveLine;
}
// Gamma writing.
// TODO(Triang3l): Gamma as sRGB check.
for (uint32_t i = 0; i < xenos::kMaxColorRenderTargets; ++i) {

3
src/xenia/gpu/vulkan/vulkan_command_processor.h
View File

@ -449,7 +449,8 @@ class VulkanCommandProcessor : public CommandProcessor {
void UpdateDynamicState(const draw_util::ViewportInfo& viewport_info,
bool primitive_polygonal,
reg::RB_DEPTHCONTROL normalized_depth_control);
void UpdateSystemConstantValues(xenos::Endian index_endian,
void UpdateSystemConstantValues(bool primitive_polygonal,
xenos::Endian index_endian,
const draw_util::ViewportInfo& viewport_info,
uint32_t used_texture_mask);
bool UpdateBindings(const VulkanShader* vertex_shader,

16
src/xenia/gpu/vulkan/vulkan_pipeline_cache.cc
View File

@ -134,13 +134,27 @@ VulkanPipelineCache::GetCurrentPixelShaderModification(
assert_true(shader.type() == xenos::ShaderType::kPixel);
assert_true(shader.is_ucode_analyzed());
const auto& regs = register_file_;
auto sq_program_cntl = regs.Get<reg::SQ_PROGRAM_CNTL>();
SpirvShaderTranslator::Modification modification(
shader_translator_->GetDefaultPixelShaderModification(
shader.GetDynamicAddressableRegisterCount(
sq_program_cntl.ps_num_reg)));
if (sq_program_cntl.param_gen) {
auto sq_context_misc = regs.Get<reg::SQ_CONTEXT_MISC>();
if (sq_context_misc.param_gen_pos <
std::min(std::max(modification.pixel.dynamic_addressable_register_count,
shader.register_static_address_bound()),
xenos::kMaxInterpolators)) {
modification.pixel.param_gen_enable = 1;
modification.pixel.param_gen_interpolator = sq_context_misc.param_gen_pos;
auto vgt_draw_initiator = regs.Get<reg::VGT_DRAW_INITIATOR>();
modification.pixel.param_gen_point = uint32_t(
vgt_draw_initiator.prim_type == xenos::PrimitiveType::kPointList);
}
}
return modification;
}