Skip to content

S
swiftshader
Commit 24466043 by Chris Forbes
Options

Partial support for input attachments

- Collect input attachment formats from the current subpass - Include InputAttachmentIndex in descriptor decorations - Make window-space integer coords available to fragment shader - Rework EmitImageLoad to use VkFormat rather than spirv format - Use window-space-coord-relative addressing for GetTexelPointer on spv::DimSubpassData images - Use input attachment format rather than OpTypeImage baked-in format for spv::DimSubpassData images. Bug: b/131171141 Test: dEQP-VK.renderpass*.input.* Change-Id: I15412e66516ea907d4b2ffd913a9eb1a3f8a9bb9 Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/29689Tested-by: 's avatarChris Forbes <chrisforbes@google.com> Presubmit-Ready: Chris Forbes <chrisforbes@google.com> Reviewed-by: 's avatarNicolas Capens <nicolascapens@google.com>
parent 72631d34
Showing with 99 additions and 70 deletions
src/Pipeline/PixelProgram.cpp
......@@ -24,6 +24,9 @@ namespace sw
void PixelProgram::setBuiltins(Int &x, Int &y, Float4(&z)[4], Float4 &w)
{
routine.windowSpacePosition[0] = x + SIMD::Int(0,1,0,1);
routine.windowSpacePosition[1] = y + SIMD::Int(0,0,1,1);
auto it = spirvShader->inputBuiltins.find(spv::BuiltInFragCoord);
if (it != spirvShader->inputBuiltins.end())
{
......
src/Pipeline/SpirvShader.cpp
......@@ -22,6 +22,7 @@
#include "Vulkan/VkDescriptorSet.hpp"
#include "Vulkan/VkPipelineLayout.hpp"
#include "Vulkan/VkDescriptorSetLayout.hpp"
#include "Vulkan/VkRenderPass.hpp"
#include "Device/Config.hpp"
#include <spirv/unified1/spirv.hpp>
......@@ -218,6 +219,7 @@ namespace
}};
}
sw::SIMD::Pointer interleaveByLane(sw::SIMD::Pointer p)
{
p *= sw::SIMD::Width;
......@@ -228,6 +230,30 @@ namespace
return p;
}
VkFormat SpirvFormatToVulkanFormat(spv::ImageFormat format)
{
switch (format)
{
case spv::ImageFormatRgba32f: return VK_FORMAT_R32G32B32A32_SFLOAT;
case spv::ImageFormatRgba32i: return VK_FORMAT_R32G32B32A32_SINT;
case spv::ImageFormatRgba32ui: return VK_FORMAT_R32G32B32A32_UINT;
case spv::ImageFormatR32f: return VK_FORMAT_R32_SFLOAT;
case spv::ImageFormatR32i: return VK_FORMAT_R32_SINT;
case spv::ImageFormatR32ui: return VK_FORMAT_R32_UINT;
case spv::ImageFormatRgba8: return VK_FORMAT_R8G8B8A8_UNORM;
case spv::ImageFormatRgba8Snorm: return VK_FORMAT_R8G8B8A8_SNORM;
case spv::ImageFormatRgba8i: return VK_FORMAT_R8G8B8A8_SINT;
case spv::ImageFormatRgba8ui: return VK_FORMAT_R8G8B8A8_UINT;
case spv::ImageFormatRgba16f: return VK_FORMAT_R16G16B16A16_SFLOAT;
case spv::ImageFormatRgba16i: return VK_FORMAT_R16G16B16A16_SINT;
case spv::ImageFormatRgba16ui: return VK_FORMAT_R16G16B16A16_UINT;
default:
UNIMPLEMENTED("SPIR-V ImageFormat %u", format);
return VK_FORMAT_UNDEFINED;
}
}
} // anonymous namespace
namespace sw
......@@ -317,13 +343,26 @@ namespace sw
volatile int SpirvShader::serialCounter = 1; // Start at 1, 0 is invalid shader.
SpirvShader::SpirvShader(InsnStore const &insns)
SpirvShader::SpirvShader(InsnStore const &insns, vk::RenderPass *renderPass, uint32_t subpassIndex)
: insns{insns}, inputs{MAX_INTERFACE_COMPONENTS},
outputs{MAX_INTERFACE_COMPONENTS},
serialID{serialCounter++}, modes{}
{
ASSERT(insns.size() > 0);
if (renderPass != VK_NULL_HANDLE)
{
// capture formats of any input attachments present
auto subpass = renderPass->getSubpass(subpassIndex);
inputAttachmentFormats.reserve(subpass.inputAttachmentCount);
for (auto i = 0u; i < subpass.inputAttachmentCount; i++)
{
auto attachmentIndex = subpass.pInputAttachments[i].attachment;
inputAttachmentFormats.push_back(attachmentIndex != VK_ATTACHMENT_UNUSED
? renderPass->getAttachment(attachmentIndex).format : VK_FORMAT_UNDEFINED);
}
}
// Simplifying assumptions (to be satisfied by earlier transformations)
// - There is exactly one entrypoint in the module, and it's the one we want
// - The only input/output OpVariables present are those used by the entrypoint
......@@ -357,6 +396,9 @@ namespace sw
case spv::DecorationBinding:
descriptorDecorations[targetId].Binding = value;
break;
case spv::DecorationInputAttachmentIndex:
descriptorDecorations[targetId].InputAttachmentIndex = value;
break;
default:
// Only handling descriptor decorations here.
break;
......@@ -1684,6 +1726,11 @@ namespace sw
{
Binding = src.Binding;
}
if (src.InputAttachmentIndex >= 0)
{
InputAttachmentIndex = src.InputAttachmentIndex;
}
}
void SpirvShader::ApplyDecorationsForId(Decorations *d, TypeOrObjectID id) const
......@@ -4524,15 +4571,24 @@ namespace sw
return EmitResult::Continue;
}
SIMD::Pointer SpirvShader::GetTexelAddress(SIMD::Pointer ptr, GenericValue const & coordinate, Type const & imageType, Pointer<Byte> descriptor, int texelSize) const
SIMD::Pointer SpirvShader::GetTexelAddress(SpirvRoutine const *routine, SIMD::Pointer ptr, GenericValue const & coordinate, Type const & imageType, Pointer<Byte> descriptor, int texelSize) const
{
bool isArrayed = imageType.definition.word(5) != 0;
auto dim = static_cast<spv::Dim>(imageType.definition.word(3));
int dims = getType(coordinate.type).sizeInComponents - (isArrayed ? 1 : 0);
ptr += coordinate.Int(0) * SIMD::Int(texelSize);
SIMD::Int u = coordinate.Int(0);
SIMD::Int v = (getType(coordinate.type).sizeInComponents > 1) ? coordinate.Int(1) : RValue<SIMD::Int>(0);
if (dim == spv::DimSubpassData)
{
u += routine->windowSpacePosition[0];
v += routine->windowSpacePosition[1];
}
ptr += u * SIMD::Int(texelSize);
if (dims > 1)
{
ptr += coordinate.Int(1) * SIMD::Int(
ptr += v * SIMD::Int(
*Pointer<Int>(descriptor + OFFSET(vk::StorageImageDescriptor, rowPitchBytes)));
}
if (dims > 2)
......@@ -4561,131 +4617,96 @@ namespace sw
ASSERT(insn.wordCount() == 5);
ASSERT(imageType.definition.opcode() == spv::OpTypeImage);
auto dim = static_cast<spv::Dim>(imageType.definition.word(3));
auto coordinate = GenericValue(this, state->routine, insn.word(4));
auto pointer = state->routine->getPointer(imageId);
Pointer<Byte> binding = pointer.base;
Pointer<Byte> imageBase = *Pointer<Pointer<Byte>>(binding + OFFSET(vk::StorageImageDescriptor, ptr));
const DescriptorDecorations &d = descriptorDecorations.at(imageId);
auto imageSizeInBytes = *Pointer<Int>(binding + OFFSET(vk::StorageImageDescriptor, sizeInBytes));
auto &dst = state->routine->createIntermediate(resultId, resultType.sizeInComponents);
SIMD::Int packed[4];
auto numPackedElements = 0u;
int texelSize = 0;
auto format = static_cast<spv::ImageFormat>(imageType.definition.word(8));
switch (format)
{
case spv::ImageFormatRgba32f:
case spv::ImageFormatRgba32i:
case spv::ImageFormatRgba32ui:
texelSize = 16;
numPackedElements = 4;
break;
case spv::ImageFormatR32f:
case spv::ImageFormatR32i:
case spv::ImageFormatR32ui:
texelSize = 4;
numPackedElements = 1;
break;
case spv::ImageFormatRgba8:
texelSize = 4;
numPackedElements = 1;
break;
case spv::ImageFormatRgba8Snorm:
texelSize = 4;
numPackedElements = 1;
break;
case spv::ImageFormatRgba8i:
case spv::ImageFormatRgba8ui:
texelSize = 4;
numPackedElements = 1;
break;
case spv::ImageFormatRgba16f:
texelSize = 8;
numPackedElements = 2;
break;
case spv::ImageFormatRgba16i:
case spv::ImageFormatRgba16ui:
texelSize = 8;
numPackedElements = 2;
break;
default:
UNIMPLEMENTED("spv::ImageFormat %u", format);
}
// For subpass data, format in the instruction is spv::ImageFormatUnknown. Get it from
// the renderpass data instead. In all other cases, we can use the format in the instruction.
auto vkFormat = (dim == spv::DimSubpassData)
? inputAttachmentFormats[d.InputAttachmentIndex]
: SpirvFormatToVulkanFormat(static_cast<spv::ImageFormat>(imageType.definition.word(8)));
auto texelSize = vk::Format(vkFormat).bytes();
auto basePtr = SIMD::Pointer(imageBase, imageSizeInBytes);
auto texelPtr = GetTexelAddress(basePtr, coordinate, imageType, binding, texelSize);
auto texelPtr = GetTexelAddress(state->routine, basePtr, coordinate, imageType, binding, texelSize);
for (auto i = 0u; i < numPackedElements; i++)
SIMD::Int packed[4];
for (auto i = 0; i < texelSize/4; i++)
{
packed[i] = SIMD::Load<SIMD::Int>(texelPtr, state->activeLaneMask());
texelPtr += sizeof(float);
}
switch(format)
switch(vkFormat)
{
case spv::ImageFormatRgba32f:
case spv::ImageFormatRgba32i:
case spv::ImageFormatRgba32ui:
case VK_FORMAT_R32G32B32A32_SFLOAT:
case VK_FORMAT_R32G32B32A32_SINT:
case VK_FORMAT_R32G32B32A32_UINT:
dst.move(0, packed[0]);
dst.move(1, packed[1]);
dst.move(2, packed[2]);
dst.move(3, packed[3]);
break;
case spv::ImageFormatR32i:
case spv::ImageFormatR32ui:
case VK_FORMAT_R32_SINT:
case VK_FORMAT_R32_UINT:
dst.move(0, packed[0]);
// Fill remaining channels with 0,0,1 (of the correct type)
dst.move(1, SIMD::Int(0));
dst.move(2, SIMD::Int(0));
dst.move(3, SIMD::Int(1));
break;
case spv::ImageFormatR32f:
case VK_FORMAT_R32_SFLOAT:
dst.move(0, packed[0]);
// Fill remaining channels with 0,0,1 (of the correct type)
dst.move(1, SIMD::Float(0));
dst.move(2, SIMD::Float(0));
dst.move(3, SIMD::Float(1));
break;
case spv::ImageFormatRgba16i:
case VK_FORMAT_R16G16B16A16_SINT:
dst.move(0, (packed[0] << 16) >> 16);
dst.move(1, (packed[0]) >> 16);
dst.move(2, (packed[1] << 16) >> 16);
dst.move(3, (packed[1]) >> 16);
break;
case spv::ImageFormatRgba16ui:
case VK_FORMAT_R16G16B16A16_UINT:
dst.move(0, packed[0] & SIMD::Int(0xffff));
dst.move(1, (packed[0] >> 16) & SIMD::Int(0xffff));
dst.move(2, packed[1] & SIMD::Int(0xffff));
dst.move(3, (packed[1] >> 16) & SIMD::Int(0xffff));
break;
case spv::ImageFormatRgba16f:
case VK_FORMAT_R16G16B16A16_SFLOAT:
dst.move(0, HalfToFloatBits(As<SIMD::UInt>(packed[0]) & SIMD::UInt(0x0000FFFF)));
dst.move(1, HalfToFloatBits((As<SIMD::UInt>(packed[0]) & SIMD::UInt(0xFFFF0000)) >> 16));
dst.move(2, HalfToFloatBits(As<SIMD::UInt>(packed[1]) & SIMD::UInt(0x0000FFFF)));
dst.move(3, HalfToFloatBits((As<SIMD::UInt>(packed[1]) & SIMD::UInt(0xFFFF0000)) >> 16));
break;
case spv::ImageFormatRgba8Snorm:
case VK_FORMAT_R8G8B8A8_SNORM:
dst.move(0, Min(Max(SIMD::Float(((packed[0]<<24) & SIMD::Int(0xFF000000))) * SIMD::Float(1.0f / float(0x7f000000)), SIMD::Float(-1.0f)), SIMD::Float(1.0f)));
dst.move(1, Min(Max(SIMD::Float(((packed[0]<<16) & SIMD::Int(0xFF000000))) * SIMD::Float(1.0f / float(0x7f000000)), SIMD::Float(-1.0f)), SIMD::Float(1.0f)));
dst.move(2, Min(Max(SIMD::Float(((packed[0]<<8) & SIMD::Int(0xFF000000))) * SIMD::Float(1.0f / float(0x7f000000)), SIMD::Float(-1.0f)), SIMD::Float(1.0f)));
dst.move(3, Min(Max(SIMD::Float(((packed[0]) & SIMD::Int(0xFF000000))) * SIMD::Float(1.0f / float(0x7f000000)), SIMD::Float(-1.0f)), SIMD::Float(1.0f)));
break;
case spv::ImageFormatRgba8:
case VK_FORMAT_R8G8B8A8_UNORM:
dst.move(0, SIMD::Float((packed[0] & SIMD::Int(0xFF))) * SIMD::Float(1.0f / 255.f));
dst.move(1, SIMD::Float(((packed[0]>>8) & SIMD::Int(0xFF))) * SIMD::Float(1.0f / 255.f));
dst.move(2, SIMD::Float(((packed[0]>>16) & SIMD::Int(0xFF))) * SIMD::Float(1.0f / 255.f));
dst.move(3, SIMD::Float(((packed[0]>>24) & SIMD::Int(0xFF))) * SIMD::Float(1.0f / 255.f));
break;
case spv::ImageFormatRgba8ui:
case VK_FORMAT_R8G8B8A8_UINT:
dst.move(0, (As<SIMD::UInt>(packed[0]) & SIMD::UInt(0xFF)));
dst.move(1, ((As<SIMD::UInt>(packed[0])>>8) & SIMD::UInt(0xFF)));
dst.move(2, ((As<SIMD::UInt>(packed[0])>>16) & SIMD::UInt(0xFF)));
dst.move(3, ((As<SIMD::UInt>(packed[0])>>24) & SIMD::UInt(0xFF)));
break;
case spv::ImageFormatRgba8i:
case VK_FORMAT_R8G8B8A8_SINT:
dst.move(0, (packed[0] << 24) >> 24);
dst.move(1, (packed[0] << 16) >> 24);
dst.move(2, (packed[0] << 8) >> 24);
......@@ -4786,7 +4807,7 @@ namespace sw
}
auto basePtr = SIMD::Pointer(imageBase, imageSizeInBytes);
auto texelPtr = GetTexelAddress(basePtr, coordinate, imageType, binding, texelSize);
auto texelPtr = GetTexelAddress(state->routine, basePtr, coordinate, imageType, binding, texelSize);
for (auto i = 0u; i < numPackedElements; i++)
{
......@@ -4818,7 +4839,7 @@ namespace sw
auto imageSizeInBytes = *Pointer<Int>(binding + OFFSET(vk::StorageImageDescriptor, sizeInBytes));
auto basePtr = SIMD::Pointer(imageBase, imageSizeInBytes);
auto ptr = GetTexelAddress(basePtr, coordinate, imageType, binding, sizeof(uint32_t));
auto ptr = GetTexelAddress(state->routine, basePtr, coordinate, imageType, binding, sizeof(uint32_t));
state->routine->createPointer(resultId, ptr);
......
src/Pipeline/SpirvShader.hpp
......@@ -45,6 +45,7 @@ namespace vk
class PipelineLayout;
class ImageView;
class Sampler;
class RenderPass;
} // namespace vk
namespace sw
......@@ -474,7 +475,7 @@ namespace sw
return serialID;
}
explicit SpirvShader(InsnStore const &insns);
SpirvShader(InsnStore const &insns, vk::RenderPass *renderPass, uint32_t subpassIndex);
struct Modes
{
......@@ -565,11 +566,13 @@ namespace sw
{
int32_t DescriptorSet = -1;
int32_t Binding = -1;
int32_t InputAttachmentIndex = -1;
void Apply(DescriptorDecorations const &src);
};
std::unordered_map<Object::ID, DescriptorDecorations> descriptorDecorations;
std::vector<VkFormat> inputAttachmentFormats;
struct InterfaceComponent
{
......@@ -840,7 +843,7 @@ namespace sw
EmitResult EmitAtomicOp(InsnIterator insn, EmitState *state) const;
EmitResult EmitAtomicCompareExchange(InsnIterator insn, EmitState *state) const;
SIMD::Pointer GetTexelAddress(SIMD::Pointer base, GenericValue const & coordinate, Type const & imageType, Pointer<Byte> descriptor, int texelSize) const;
SIMD::Pointer GetTexelAddress(SpirvRoutine const * routine, SIMD::Pointer base, GenericValue const & coordinate, Type const & imageType, Pointer<Byte> descriptor, int texelSize) const;
// OpcodeName() returns the name of the opcode op.
// If NDEBUG is defined, then OpcodeName() will only return the numerical code.
......@@ -896,6 +899,7 @@ namespace sw
Pointer<Int> descriptorDynamicOffsets;
Pointer<Byte> pushConstants;
Int killMask = Int{0};
SIMD::Int windowSpacePosition[2];
void createVariable(SpirvShader::Object::ID id, uint32_t size)
{
......
src/Vulkan/VkPipeline.cpp
......@@ -15,6 +15,7 @@
#include "VkPipeline.hpp"
#include "VkPipelineLayout.hpp"
#include "VkShaderModule.hpp"
#include "VkRenderPass.hpp"
#include "Pipeline/ComputeProgram.hpp"
#include "Pipeline/SpirvShader.hpp"
......@@ -452,7 +453,7 @@ void GraphicsPipeline::compileShaders(const VkAllocationCallbacks* pAllocator, c
// FIXME (b/119409619): use an allocator here so we can control all memory allocations
// TODO: also pass in any pipeline state which will affect shader compilation
auto spirvShader = new sw::SpirvShader{code};
auto spirvShader = new sw::SpirvShader{code, Cast(pCreateInfo->renderPass), pCreateInfo->subpass};
switch (pStage->stage)
{
......@@ -544,7 +545,7 @@ void ComputePipeline::compileShaders(const VkAllocationCallbacks* pAllocator, co
ASSERT(shader == nullptr);
// FIXME(b/119409619): use allocator.
shader = new sw::SpirvShader(code);
shader = new sw::SpirvShader(code, nullptr, 0);
vk::DescriptorSet::Bindings descriptorSets; // FIXME(b/129523279): Delay code generation until invoke time.
sw::ComputeProgram program(shader, layout, descriptorSets);
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment