miguelr/suyu
1
0
Fork 0
forked from suyu/suyu
Code Pull requests Activity

vk_shader_decompiler: Misc fixes

Browse source 
Fix missing OpSelectionMerge instruction. This caused devices loses on
most hardware, Intel didn't care.

Fix [-1;1] -> [0;1] depth conversions.

Conditionally use VK_EXT_scalar_block_layout. This allows us to use
non-std140 layouts on UBOs.

Update external Vulkan headers.
This commit is contained in:
ReinUsesLisp 2019-05-26 01:43:07 -03:00
parent dec3c981d0
commit a4c5e3e339
3 changed files with 68 additions and 46 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
externals/Vulkan-Headers vendored

@ -1 +1 @@
Subproject commit 15e5c4db7500b936ae758236f2e72fc1aec22020 Subproject commit d05c8df88da98ec1ab3bc600d7f5783b4060895b

104
src/video_core/renderer_vulkan/vk_shader_decompiler.cpp
View file

@ -17,6 +17,7 @@
#include "video_core/engines/maxwell_3d.h" #include "video_core/engines/maxwell_3d.h"
#include "video_core/engines/shader_bytecode.h" #include "video_core/engines/shader_bytecode.h"
#include "video_core/engines/shader_header.h" #include "video_core/engines/shader_header.h"
#include "video_core/renderer_vulkan/vk_device.h"
#include "video_core/renderer_vulkan/vk_shader_decompiler.h" #include "video_core/renderer_vulkan/vk_shader_decompiler.h"
#include "video_core/shader/shader_ir.h" #include "video_core/shader/shader_ir.h"
@ -33,7 +34,8 @@ using ShaderStage = Tegra::Engines::Maxwell3D::Regs::ShaderStage;
using Operation = const OperationNode&; using Operation = const OperationNode&;
// TODO(Rodrigo): Use rasterizer's value // TODO(Rodrigo): Use rasterizer's value
constexpr u32 MAX_CONSTBUFFER_ELEMENTS = 0x1000; constexpr u32 MAX_CONSTBUFFER_FLOATS = 0x4000;
constexpr u32 MAX_CONSTBUFFER_ELEMENTS = MAX_CONSTBUFFER_FLOATS / 4;
constexpr u32 STAGE_BINDING_STRIDE = 0x100; constexpr u32 STAGE_BINDING_STRIDE = 0x100;
enum class Type { Bool, Bool2, Float, Int, Uint, HalfFloat }; enum class Type { Bool, Bool2, Float, Int, Uint, HalfFloat };
@ -87,8 +89,8 @@ bool IsPrecise(Operation operand) {
class SPIRVDecompiler : public Sirit::Module { class SPIRVDecompiler : public Sirit::Module {
public: public:
explicit SPIRVDecompiler(const ShaderIR& ir, ShaderStage stage) explicit SPIRVDecompiler(const VKDevice& device, const ShaderIR& ir, ShaderStage stage)
: Module(0x00010300), ir{ir}, stage{stage}, header{ir.GetHeader()} { : Module(0x00010300), device{device}, ir{ir}, stage{stage}, header{ir.GetHeader()} {
AddCapability(spv::Capability::Shader); AddCapability(spv::Capability::Shader);
AddExtension("SPV_KHR_storage_buffer_storage_class"); AddExtension("SPV_KHR_storage_buffer_storage_class");
AddExtension("SPV_KHR_variable_pointers"); AddExtension("SPV_KHR_variable_pointers");
@ -195,7 +197,9 @@ public:
entries.samplers.emplace_back(sampler); entries.samplers.emplace_back(sampler);
} }
for (const auto& attribute : ir.GetInputAttributes()) { for (const auto& attribute : ir.GetInputAttributes()) {
entries.attributes.insert(GetGenericAttributeLocation(attribute)); if (IsGenericAttribute(attribute)) {
entries.attributes.insert(GetGenericAttributeLocation(attribute));
}
} }
entries.clip_distances = ir.GetClipDistances(); entries.clip_distances = ir.GetClipDistances();
entries.shader_length = ir.GetLength(); entries.shader_length = ir.GetLength();
@ -210,7 +214,6 @@ private:
std::array<OperationDecompilerFn, static_cast<std::size_t>(OperationCode::Amount)>; std::array<OperationDecompilerFn, static_cast<std::size_t>(OperationCode::Amount)>;
static constexpr auto INTERNAL_FLAGS_COUNT = static_cast<std::size_t>(InternalFlag::Amount); static constexpr auto INTERNAL_FLAGS_COUNT = static_cast<std::size_t>(InternalFlag::Amount);
static constexpr u32 CBUF_STRIDE = 16;
void AllocateBindings() { void AllocateBindings() {
const u32 binding_base = static_cast<u32>(stage) * STAGE_BINDING_STRIDE; const u32 binding_base = static_cast<u32>(stage) * STAGE_BINDING_STRIDE;
@ -315,6 +318,7 @@ private:
constexpr std::array<const char*, INTERNAL_FLAGS_COUNT> names = {"zero", "sign", "carry", constexpr std::array<const char*, INTERNAL_FLAGS_COUNT> names = {"zero", "sign", "carry",
"overflow"}; "overflow"};
for (std::size_t flag = 0; flag < INTERNAL_FLAGS_COUNT; ++flag) { for (std::size_t flag = 0; flag < INTERNAL_FLAGS_COUNT; ++flag) {
const auto flag_code = static_cast<InternalFlag>(flag);
const Id id = OpVariable(t_prv_bool, spv::StorageClass::Private, v_false); const Id id = OpVariable(t_prv_bool, spv::StorageClass::Private, v_false);
internal_flags[flag] = AddGlobalVariable(Name(id, names[flag])); internal_flags[flag] = AddGlobalVariable(Name(id, names[flag]));
} }
@ -374,7 +378,9 @@ private:
u32 binding = const_buffers_base_binding; u32 binding = const_buffers_base_binding;
for (const auto& entry : ir.GetConstantBuffers()) { for (const auto& entry : ir.GetConstantBuffers()) {
const auto [index, size] = entry; const auto [index, size] = entry;
const Id id = OpVariable(t_cbuf_ubo, spv::StorageClass::Uniform); const Id type =
device.IsExtScalarBlockLayoutSupported() ? t_cbuf_scalar_ubo : t_cbuf_std140_ubo;
const Id id = OpVariable(type, spv::StorageClass::Uniform);
AddGlobalVariable(Name(id, fmt::format("cbuf_{}", index))); AddGlobalVariable(Name(id, fmt::format("cbuf_{}", index)));
Decorate(id, spv::Decoration::Binding, binding++); Decorate(id, spv::Decoration::Binding, binding++);
@ -569,33 +575,35 @@ private:
const Node offset = cbuf->GetOffset(); const Node offset = cbuf->GetOffset();
const Id buffer_id = constant_buffers.at(cbuf->GetIndex()); const Id buffer_id = constant_buffers.at(cbuf->GetIndex());
Id buffer_index{}; Id pointer{};
Id buffer_element{}; if (device.IsExtScalarBlockLayoutSupported()) {
const Id buffer_offset = Emit(OpShiftRightLogical(
if (const auto immediate = std::get_if<ImmediateNode>(offset)) { t_uint, BitcastTo<Type::Uint>(Visit(offset)), Constant(t_uint, 2u)));
// Direct access pointer = Emit(
const u32 offset_imm = immediate->GetValue(); OpAccessChain(t_cbuf_float, buffer_id, Constant(t_uint, 0u), buffer_offset));
ASSERT(offset_imm % 4 == 0);
buffer_index = Constant(t_uint, offset_imm / 16);
buffer_element = Constant(t_uint, (offset_imm / 4) % 4);
} else if (std::holds_alternative<OperationNode>(*offset)) {
// Indirect access
// TODO(Rodrigo): Use a uniform buffer stride of 4 and drop this slow math (which
// emits sub-optimal code on GLSL from my testing).
const Id offset_id = BitcastTo<Type::Uint>(Visit(offset));
const Id unsafe_offset = Emit(OpUDiv(t_uint, offset_id, Constant(t_uint, 4)));
const Id final_offset = Emit(
OpUMod(t_uint, unsafe_offset, Constant(t_uint, MAX_CONSTBUFFER_ELEMENTS - 1)));
buffer_index = Emit(OpUDiv(t_uint, final_offset, Constant(t_uint, 4)));
buffer_element = Emit(OpUMod(t_uint, final_offset, Constant(t_uint, 4)));
} else { } else {
UNREACHABLE_MSG("Unmanaged offset node type"); Id buffer_index{};
Id buffer_element{};
if (const auto immediate = std::get_if<ImmediateNode>(offset)) {
// Direct access
const u32 offset_imm = immediate->GetValue();
ASSERT(offset_imm % 4 == 0);
buffer_index = Constant(t_uint, offset_imm / 16);
buffer_element = Constant(t_uint, (offset_imm / 4) % 4);
} else if (std::holds_alternative<OperationNode>(*offset)) {
// Indirect access
const Id offset_id = BitcastTo<Type::Uint>(Visit(offset));
const Id unsafe_offset = Emit(OpUDiv(t_uint, offset_id, Constant(t_uint, 4)));
const Id final_offset = Emit(OpUMod(
t_uint, unsafe_offset, Constant(t_uint, MAX_CONSTBUFFER_ELEMENTS - 1)));
buffer_index = Emit(OpUDiv(t_uint, final_offset, Constant(t_uint, 4)));
buffer_element = Emit(OpUMod(t_uint, final_offset, Constant(t_uint, 4)));
} else {
UNREACHABLE_MSG("Unmanaged offset node type");
}
pointer = Emit(OpAccessChain(t_cbuf_float, buffer_id, Constant(t_uint, 0),
buffer_index, buffer_element));
} }
const Id pointer = Emit(OpAccessChain(t_cbuf_float, buffer_id, Constant(t_uint, 0),
buffer_index, buffer_element));
return Emit(OpLoad(t_float, pointer)); return Emit(OpLoad(t_float, pointer));
} else if (const auto gmem = std::get_if<GmemNode>(node)) { } else if (const auto gmem = std::get_if<GmemNode>(node)) {
@ -612,7 +620,9 @@ private:
// It's invalid to call conditional on nested nodes, use an operation instead // It's invalid to call conditional on nested nodes, use an operation instead
const Id true_label = OpLabel(); const Id true_label = OpLabel();
const Id skip_label = OpLabel(); const Id skip_label = OpLabel();
Emit(OpBranchConditional(Visit(conditional->GetCondition()), true_label, skip_label)); const Id condition = Visit(conditional->GetCondition());
Emit(OpSelectionMerge(skip_label, spv::SelectionControlMask::MaskNone));
Emit(OpBranchConditional(condition, true_label, skip_label));
Emit(true_label); Emit(true_label);
VisitBasicBlock(conditional->GetCode()); VisitBasicBlock(conditional->GetCode());
@ -968,11 +978,11 @@ private:
case ShaderStage::Vertex: { case ShaderStage::Vertex: {
// TODO(Rodrigo): We should use VK_EXT_depth_range_unrestricted instead, but it doesn't // TODO(Rodrigo): We should use VK_EXT_depth_range_unrestricted instead, but it doesn't
// seem to be working on Nvidia's drivers and Intel (mesa and blob) doesn't support it. // seem to be working on Nvidia's drivers and Intel (mesa and blob) doesn't support it.
const Id position = AccessElement(t_float4, per_vertex, position_index); const Id z_pointer = AccessElement(t_out_float, per_vertex, position_index, 2u);
Id depth = Emit(OpLoad(t_float, AccessElement(t_out_float, position, 2))); Id depth = Emit(OpLoad(t_float, z_pointer));
depth = Emit(OpFAdd(t_float, depth, Constant(t_float, 1.0f))); depth = Emit(OpFAdd(t_float, depth, Constant(t_float, 1.0f)));
depth = Emit(OpFMul(t_float, depth, Constant(t_float, 0.5f))); depth = Emit(OpFMul(t_float, depth, Constant(t_float, 0.5f)));
Emit(OpStore(AccessElement(t_out_float, position, 2), depth)); Emit(OpStore(z_pointer, depth));
break; break;
} }
case ShaderStage::Fragment: { case ShaderStage::Fragment: {
@ -1293,6 +1303,7 @@ private:
&SPIRVDecompiler::YNegate, &SPIRVDecompiler::YNegate,
}; };
const VKDevice& device;
const ShaderIR& ir; const ShaderIR& ir;
const ShaderStage stage; const ShaderStage stage;
const Tegra::Shader::Header header; const Tegra::Shader::Header header;
@ -1331,12 +1342,18 @@ private:
const Id t_out_float4 = Name(TypePointer(spv::StorageClass::Output, t_float4), "out_float4"); const Id t_out_float4 = Name(TypePointer(spv::StorageClass::Output, t_float4), "out_float4");
const Id t_cbuf_float = TypePointer(spv::StorageClass::Uniform, t_float); const Id t_cbuf_float = TypePointer(spv::StorageClass::Uniform, t_float);
const Id t_cbuf_array = const Id t_cbuf_std140 = Decorate(
Decorate(Name(TypeArray(t_float4, Constant(t_uint, MAX_CONSTBUFFER_ELEMENTS)), "CbufArray"), Name(TypeArray(t_float4, Constant(t_uint, MAX_CONSTBUFFER_ELEMENTS)), "CbufStd140Array"),
spv::Decoration::ArrayStride, CBUF_STRIDE); spv::Decoration::ArrayStride, 16u);
const Id t_cbuf_struct = MemberDecorate( const Id t_cbuf_scalar = Decorate(
Decorate(TypeStruct(t_cbuf_array), spv::Decoration::Block), 0, spv::Decoration::Offset, 0); Name(TypeArray(t_float, Constant(t_uint, MAX_CONSTBUFFER_FLOATS)), "CbufScalarArray"),
const Id t_cbuf_ubo = TypePointer(spv::StorageClass::Uniform, t_cbuf_struct); spv::Decoration::ArrayStride, 4u);
const Id t_cbuf_std140_struct = MemberDecorate(
Decorate(TypeStruct(t_cbuf_std140), spv::Decoration::Block), 0, spv::Decoration::Offset, 0);
const Id t_cbuf_scalar_struct = MemberDecorate(
Decorate(TypeStruct(t_cbuf_scalar), spv::Decoration::Block), 0, spv::Decoration::Offset, 0);
const Id t_cbuf_std140_ubo = TypePointer(spv::StorageClass::Uniform, t_cbuf_std140_struct);
const Id t_cbuf_scalar_ubo = TypePointer(spv::StorageClass::Uniform, t_cbuf_scalar_struct);
const Id t_gmem_float = TypePointer(spv::StorageClass::StorageBuffer, t_float); const Id t_gmem_float = TypePointer(spv::StorageClass::StorageBuffer, t_float);
const Id t_gmem_array = const Id t_gmem_array =
@ -1385,8 +1402,9 @@ private:
std::map<u32, Id> labels; std::map<u32, Id> labels;
}; };
DecompilerResult Decompile(const VideoCommon::Shader::ShaderIR& ir, Maxwell::ShaderStage stage) { DecompilerResult Decompile(const VKDevice& device, const VideoCommon::Shader::ShaderIR& ir,
auto decompiler = std::make_unique<SPIRVDecompiler>(ir, stage); Maxwell::ShaderStage stage) {
auto decompiler = std::make_unique<SPIRVDecompiler>(device, ir, stage);
decompiler->Decompile(); decompiler->Decompile();
return {std::move(decompiler), decompiler->GetShaderEntries()}; return {std::move(decompiler), decompiler->GetShaderEntries()};
} }

8
src/video_core/renderer_vulkan/vk_shader_decompiler.h
View file

@ -20,10 +20,13 @@ namespace VideoCommon::Shader {
class ShaderIR; class ShaderIR;
} }
namespace Vulkan {
class VKDevice;
}
namespace Vulkan::VKShader { namespace Vulkan::VKShader {
using Maxwell = Tegra::Engines::Maxwell3D::Regs; using Maxwell = Tegra::Engines::Maxwell3D::Regs;
using SamplerEntry = VideoCommon::Shader::Sampler; using SamplerEntry = VideoCommon::Shader::Sampler;
constexpr u32 DESCRIPTOR_SET = 0; constexpr u32 DESCRIPTOR_SET = 0;
@ -75,6 +78,7 @@ struct ShaderEntries {
using DecompilerResult = std::pair<std::unique_ptr<Sirit::Module>, ShaderEntries>; using DecompilerResult = std::pair<std::unique_ptr<Sirit::Module>, ShaderEntries>;
DecompilerResult Decompile(const VideoCommon::Shader::ShaderIR& ir, Maxwell::ShaderStage stage); DecompilerResult Decompile(const VKDevice& device, const VideoCommon::Shader::ShaderIR& ir,
Maxwell::ShaderStage stage);
} // namespace Vulkan::VKShader } // namespace Vulkan::VKShader