Skip to content

S
swiftshader
Commit 51e9e565 by Nicolas Capens Committed by Nicolas Capens
Options

Implement Y′CbCr conversion

As specified in https://www.khronos.org/registry/DataFormat/specs/1.2/dataformat.1.2.html Bug: b/132437008 Change-Id: I730a9eb74796c625d3e635a66b18661c73f7c01f Tests: dEQP-VK.*ycbcr* Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/31618 Presubmit-Ready: Nicolas Capens <nicolascapens@google.com> Tested-by: 's avatarNicolas Capens <nicolascapens@google.com> Kokoro-Presubmit: kokoro <noreply+kokoro@google.com> Reviewed-by: 's avatarBen Clayton <bclayton@google.com>
parent 8ce8707b
Showing with 184 additions and 107 deletions
src/Device/Sampler.hpp
...@@ -153,6 +153,10 @@ namespace sw ...@@ -153,6 +153,10 @@ namespace sw
bool unnormalizedCoordinates; bool unnormalizedCoordinates;
bool largeTexture; bool largeTexture;
VkSamplerYcbcrModelConversion ycbcrModel;
bool studioSwing; // Narrow range
bool swappedChroma; // Cb/Cr components in reverse order
#if PERF_PROFILE #if PERF_PROFILE
bool compressedFormat; bool compressedFormat;
#endif #endif
......
src/Pipeline/SamplerCore.cpp
...@@ -1603,100 +1603,126 @@ namespace sw ...@@ -1603,100 +1603,126 @@ namespace sw
if(isYcbcrFormat()) if(isYcbcrFormat())
{ {
// Generic YPbPr to RGB transformation // Luminance
// R = Y + 2 * (1 - Kr) * Pr Int c0 = Int(buffer[0][index[0]]);
// G = Y - 2 * Kb * (1 - Kb) / Kg * Pb - 2 * Kr * (1 - Kr) / Kg * Pr Int c1 = Int(buffer[0][index[1]]);
// B = Y + 2 * (1 - Kb) * Pb Int c2 = Int(buffer[0][index[2]]);
Int c3 = Int(buffer[0][index[3]]);
c0 = c0 | (c1 << 8) | (c2 << 16) | (c3 << 24);
UShort4 Y = As<UShort4>(Unpack(As<Byte4>(c0)));
float Kb = 0.114f; UShort4 Cb, Cr;
float Kr = 0.299f;
int studioSwing = 1;
switch(state.textureFormat) // Chroma
{ {
case VK_FORMAT_G8_B8R8_2PLANE_420_UNORM: computeIndices(index, uuuu, vvvv, wwww, offset, mipmap + sizeof(Mipmap), function);
// VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_601 UShort4 U, V;
Kb = 0.114f;
Kr = 0.299f;
studioSwing = 1;
break;
/*
// VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_709
Kb = 0.0722f;
Kr = 0.2126f;
studioSwing = 1;
break;
// VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_2020
Kb = 0.114f;
Kr = 0.299f;
studioSwing = 0;
break;
*/
default:
ASSERT(false);
}
const float Kg = 1.0f - Kr - Kb; if(state.textureFormat == VK_FORMAT_G8_B8_R8_3PLANE_420_UNORM)
{
const float Rr = 2 * (1 - Kr); c0 = Int(buffer[1][index[0]]);
const float Gb = -2 * Kb * (1 - Kb) / Kg; c1 = Int(buffer[1][index[1]]);
const float Gr = -2 * Kr * (1 - Kr) / Kg; c2 = Int(buffer[1][index[2]]);
const float Bb = 2 * (1 - Kb); c3 = Int(buffer[1][index[3]]);
c0 = c0 | (c1 << 8) | (c2 << 16) | (c3 << 24);
U = As<UShort4>(Unpack(As<Byte4>(c0)));
// Scaling and bias for studio-swing range: Y = [16 .. 235], U/V = [16 .. 240] c0 = Int(buffer[2][index[0]]);
const float Yy = studioSwing ? 255.0f / (235 - 16) : 1.0f; c1 = Int(buffer[2][index[1]]);
const float Uu = studioSwing ? 255.0f / (240 - 16) : 1.0f; c2 = Int(buffer[2][index[2]]);
const float Vv = studioSwing ? 255.0f / (240 - 16) : 1.0f; c3 = Int(buffer[2][index[3]]);
c0 = c0 | (c1 << 8) | (c2 << 16) | (c3 << 24);
V = As<UShort4>(Unpack(As<Byte4>(c0)));
}
else if(state.textureFormat == VK_FORMAT_G8_B8R8_2PLANE_420_UNORM)
{
Short4 UV;
UV = Insert(UV, Pointer<Short>(buffer[1])[index[0]], 0); // TODO: Insert(UShort4, UShort)
UV = Insert(UV, Pointer<Short>(buffer[1])[index[1]], 1);
UV = Insert(UV, Pointer<Short>(buffer[1])[index[2]], 2);
UV = Insert(UV, Pointer<Short>(buffer[1])[index[3]], 3);
U = (UV & Short4(0x00FFu)) | (UV << 8);
V = (UV & Short4(0xFF00u)) | As<Short4>(As<UShort4>(UV) >> 8);
}
else UNSUPPORTED("state.textureFormat %d", (int)state.textureFormat);
const float Rv = Vv * Rr; if(!state.swappedChroma)
const float Gu = Uu * Gb; {
const float Gv = Vv * Gr; Cb = U;
const float Bu = Uu * Bb; Cr = V;
}
else
{
Cb = V;
Cr = U;
}
}
const float R0 = (studioSwing * -16 * Yy - 128 * Rv) / 255; if(state.ycbcrModel == VK_SAMPLER_YCBCR_MODEL_CONVERSION_RGB_IDENTITY)
const float G0 = (studioSwing * -16 * Yy - 128 * Gu - 128 * Gv) / 255; {
const float B0 = (studioSwing * -16 * Yy - 128 * Bu) / 255; // YCbCr formats are treated as signed 15-bit.
c.x = Cr >> 1;
c.y = Y >> 1;
c.z = Cb >> 1;
}
else
{
// Scaling and bias for studio-swing range: Y = [16 .. 235], U/V = [16 .. 240]
// Scale down by 0x0101 to normalize the 8.8 samples, and up by 0x7FFF for signed 15-bit output.
Float4 y = (Float4(Y) - Float4(state.studioSwing ? 16 * 0x0101 : 0)) * Float4(float(0x7FFF) / (state.studioSwing ? 219 * 0x0101 : 255 * 0x0101));
Float4 u = (Float4(Cb) - Float4(128 * 0x0101)) * Float4(float(0x7FFF) / (state.studioSwing ? 224 * 0x0101 : 255 * 0x0101));
Float4 v = (Float4(Cr) - Float4(128 * 0x0101)) * Float4(float(0x7FFF) / (state.studioSwing ? 224 * 0x0101 : 255 * 0x0101));
Int c0 = Int(buffer[0][index[0]]); if(state.ycbcrModel == VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_IDENTITY)
Int c1 = Int(buffer[0][index[1]]); {
Int c2 = Int(buffer[0][index[2]]); c.x = Short4(v);
Int c3 = Int(buffer[0][index[3]]); c.y = Short4(y);
c0 = c0 | (c1 << 8) | (c2 << 16) | (c3 << 24); c.z = Short4(u);
UShort4 Y = As<UShort4>(Unpack(As<Byte4>(c0))); }
else
{
// Generic YCbCr to RGB transformation:
// R = Y + 2 * (1 - Kr) * Cr
// G = Y - 2 * Kb * (1 - Kb) / Kg * Cb - 2 * Kr * (1 - Kr) / Kg * Cr
// B = Y + 2 * (1 - Kb) * Cb
computeIndices(index, uuuu, vvvv, wwww, offset, mipmap + sizeof(Mipmap), function); float Kb = 0.114f;
c0 = Int(buffer[1][index[0]]); float Kr = 0.299f;
c1 = Int(buffer[1][index[1]]);
c2 = Int(buffer[1][index[2]]);
c3 = Int(buffer[1][index[3]]);
c0 = c0 | (c1 << 8) | (c2 << 16) | (c3 << 24);
UShort4 V = As<UShort4>(Unpack(As<Byte4>(c0)));
c0 = Int(buffer[2][index[0]]); switch(state.ycbcrModel)
c1 = Int(buffer[2][index[1]]); {
c2 = Int(buffer[2][index[2]]); case VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_709:
c3 = Int(buffer[2][index[3]]); Kb = 0.0722f;
c0 = c0 | (c1 << 8) | (c2 << 16) | (c3 << 24); Kr = 0.2126f;
UShort4 U = As<UShort4>(Unpack(As<Byte4>(c0))); break;
case VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_601:
Kb = 0.114f;
Kr = 0.299f;
break;
case VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_2020:
Kb = 0.0593f;
Kr = 0.2627f;
break;
default:
UNSUPPORTED("ycbcrModel %d", int(state.ycbcrModel));
}
const UShort4 yY = UShort4(iround(Yy * 0x4000)); const float Kg = 1.0f - Kr - Kb;
const UShort4 rV = UShort4(iround(Rv * 0x4000));
const UShort4 gU = UShort4(iround(-Gu * 0x4000));
const UShort4 gV = UShort4(iround(-Gv * 0x4000));
const UShort4 bU = UShort4(iround(Bu * 0x4000));
const UShort4 r0 = UShort4(iround(-R0 * 0x4000)); const float Rr = 2 * (1 - Kr);
const UShort4 g0 = UShort4(iround(G0 * 0x4000)); const float Gb = -2 * Kb * (1 - Kb) / Kg;
const UShort4 b0 = UShort4(iround(-B0 * 0x4000)); const float Gr = -2 * Kr * (1 - Kr) / Kg;
const float Bb = 2 * (1 - Kb);
UShort4 y = MulHigh(Y, yY); Float4 r = y + Float4(Rr) * v;
UShort4 r = SubSat(y + MulHigh(V, rV), r0); Float4 g = y + Float4(Gb) * u + Float4(Gr) * v;
UShort4 g = SubSat(y + g0, MulHigh(U, gU) + MulHigh(V, gV)); Float4 b = y + Float4(Bb) * u ;
UShort4 b = SubSat(y + MulHigh(U, bU), b0);
c.x = Min(r, UShort4(0x3FFF)) << 2; c.x = Short4(r);
c.y = Min(g, UShort4(0x3FFF)) << 2; c.y = Short4(g);
c.z = Min(b, UShort4(0x3FFF)) << 2; c.z = Short4(b);
}
}
} }
else else
{ {
...@@ -1991,8 +2017,8 @@ namespace sw ...@@ -1991,8 +2017,8 @@ namespace sw
if(isYcbcrFormat()) if(isYcbcrFormat())
{ {
buffer[1] = *Pointer<Pointer<Byte>>(mipmap + OFFSET(Mipmap,buffer[1])); buffer[1] = *Pointer<Pointer<Byte>>(mipmap + sizeof(Mipmap) * 1 + OFFSET(Mipmap,buffer[0]));
buffer[2] = *Pointer<Pointer<Byte>>(mipmap + OFFSET(Mipmap,buffer[2])); buffer[2] = *Pointer<Pointer<Byte>>(mipmap + sizeof(Mipmap) * 2 + OFFSET(Mipmap,buffer[0]));
} }
} }
else else
......
src/Pipeline/SpirvShaderSampling.cpp
...@@ -96,6 +96,13 @@ SpirvShader::ImageSampler *SpirvShader::getImageSampler(uint32_t inst, vk::Sampl ...@@ -96,6 +96,13 @@ SpirvShader::ImageSampler *SpirvShader::getImageSampler(uint32_t inst, vk::Sampl
(imageDescriptor->extent.height > SHRT_MAX) || (imageDescriptor->extent.height > SHRT_MAX) ||
(imageDescriptor->extent.depth > SHRT_MAX); (imageDescriptor->extent.depth > SHRT_MAX);
if(sampler->ycbcrConversion)
{
samplerState.ycbcrModel = sampler->ycbcrConversion->ycbcrModel;
samplerState.studioSwing = (sampler->ycbcrConversion->ycbcrRange == VK_SAMPLER_YCBCR_RANGE_ITU_NARROW);
samplerState.swappedChroma = (sampler->ycbcrConversion->components.r != VK_COMPONENT_SWIZZLE_R);
}
if(sampler->anisotropyEnable != VK_FALSE) if(sampler->anisotropyEnable != VK_FALSE)
{ {
UNSUPPORTED("anisotropyEnable"); UNSUPPORTED("anisotropyEnable");
...@@ -238,6 +245,11 @@ sw::FilterType SpirvShader::convertFilterMode(const vk::Sampler *sampler) ...@@ -238,6 +245,11 @@ sw::FilterType SpirvShader::convertFilterMode(const vk::Sampler *sampler)
sw::MipmapType SpirvShader::convertMipmapMode(const vk::Sampler *sampler) sw::MipmapType SpirvShader::convertMipmapMode(const vk::Sampler *sampler)
{ {
if(sampler->ycbcrConversion)
{
return MIPMAP_NONE; // YCbCr images can only have one mipmap level.
}
switch(sampler->mipmapMode) switch(sampler->mipmapMode)
{ {
case VK_SAMPLER_MIPMAP_MODE_NEAREST: return MIPMAP_POINT; case VK_SAMPLER_MIPMAP_MODE_NEAREST: return MIPMAP_POINT;
......
src/Vulkan/VkDescriptorSetLayout.cpp
...@@ -364,41 +364,76 @@ void DescriptorSetLayout::WriteDescriptorSet(DescriptorSet *dstSet, VkDescriptor ...@@ -364,41 +364,76 @@ void DescriptorSetLayout::WriteDescriptorSet(DescriptorSet *dstSet, VkDescriptor
auto &subresourceRange = imageView->getSubresourceRange(); auto &subresourceRange = imageView->getSubresourceRange();
for(int mipmapLevel = 0; mipmapLevel < sw::MIPMAP_LEVELS; mipmapLevel++) if(format.isYcbcrFormat())
{ {
int level = sw::clamp(mipmapLevel, 0, (int)subresourceRange.levelCount - 1); // Level within the image view ASSERT(subresourceRange.levelCount == 1);
VkImageAspectFlagBits aspect = static_cast<VkImageAspectFlagBits>(imageView->getSubresourceRange().aspectMask); // YCbCr images can only have one level, so we can store parameters for the
sw::Mipmap &mipmap = texture->mipmap[mipmapLevel]; // different planes in the descriptor's mipmap levels instead.
if(imageView->getType() == VK_IMAGE_VIEW_TYPE_CUBE) const int level = 0;
VkOffset3D offset = {0, 0, 0};
texture->mipmap[0].buffer[0] = imageView->getOffsetPointer(offset, VK_IMAGE_ASPECT_PLANE_0_BIT, level, 0, ImageView::SAMPLING);
texture->mipmap[1].buffer[0] = imageView->getOffsetPointer(offset, VK_IMAGE_ASPECT_PLANE_1_BIT, level, 0, ImageView::SAMPLING);
if(format.getAspects() & VK_IMAGE_ASPECT_PLANE_2_BIT)
{ {
for(int face = 0; face < 6; face++) texture->mipmap[2].buffer[0] = imageView->getOffsetPointer(offset, VK_IMAGE_ASPECT_PLANE_2_BIT, level, 0, ImageView::SAMPLING);
{
// Obtain the pointer to the corner of the level including the border, for seamless sampling.
// This is taken into account in the sampling routine, which can't handle negative texel coordinates.
VkOffset3D offset = {-1, -1, 0};
// TODO(b/129523279): Implement as 6 consecutive layers instead of separate pointers.
mipmap.buffer[face] = imageView->getOffsetPointer(offset, aspect, level, face, ImageView::SAMPLING);
}
}
else
{
VkOffset3D offset = {0, 0, 0};
mipmap.buffer[0] = imageView->getOffsetPointer(offset, aspect, level, 0, ImageView::SAMPLING);
} }
VkExtent3D extent = imageView->getMipLevelExtent(level); VkExtent3D extent = imageView->getMipLevelExtent(0);
int width = extent.width; int width = extent.width;
int height = extent.height; int height = extent.height;
int layers = imageView->getSubresourceRange().layerCount; // TODO(b/129523279): Untangle depth vs layers throughout the sampler int pitchP0 = imageView->rowPitchBytes(VK_IMAGE_ASPECT_PLANE_0_BIT, level, ImageView::SAMPLING) /
int depth = layers > 1 ? layers : extent.depth; imageView->getFormat(VK_IMAGE_ASPECT_PLANE_0_BIT).bytes();
int pitchP = imageView->rowPitchBytes(aspect, level, ImageView::SAMPLING) / format.bytes();
int sliceP = (layers > 1 ? imageView->layerPitchBytes(aspect, ImageView::SAMPLING) : imageView->slicePitchBytes(aspect, level, ImageView::SAMPLING)) / format.bytes(); // Write plane 0 parameters to mipmap level 0.
WriteTextureLevelInfo(texture, 0, width, height, 1, pitchP0, 0);
// Plane 2, if present, has equal parameters to plane 1, so we use mipmap level 1 for both.
int pitchP1 = imageView->rowPitchBytes(VK_IMAGE_ASPECT_PLANE_1_BIT, level, ImageView::SAMPLING) /
imageView->getFormat(VK_IMAGE_ASPECT_PLANE_1_BIT).bytes();
WriteTextureLevelInfo(texture, 1, width / 2, height / 2, 1, pitchP1, 0);
}
else
{
for(int mipmapLevel = 0; mipmapLevel < sw::MIPMAP_LEVELS; mipmapLevel++)
{
int level = sw::clamp(mipmapLevel, 0, (int)subresourceRange.levelCount - 1); // Level within the image view
VkImageAspectFlagBits aspect = static_cast<VkImageAspectFlagBits>(imageView->getSubresourceRange().aspectMask);
sw::Mipmap &mipmap = texture->mipmap[mipmapLevel];
if(imageView->getType() == VK_IMAGE_VIEW_TYPE_CUBE)
{
for(int face = 0; face < 6; face++)
{
// Obtain the pointer to the corner of the level including the border, for seamless sampling.
// This is taken into account in the sampling routine, which can't handle negative texel coordinates.
VkOffset3D offset = {-1, -1, 0};
// TODO(b/129523279): Implement as 6 consecutive layers instead of separate pointers.
mipmap.buffer[face] = imageView->getOffsetPointer(offset, aspect, level, face, ImageView::SAMPLING);
}
}
else
{
VkOffset3D offset = {0, 0, 0};
mipmap.buffer[0] = imageView->getOffsetPointer(offset, aspect, level, 0, ImageView::SAMPLING);
}
VkExtent3D extent = imageView->getMipLevelExtent(level);
WriteTextureLevelInfo(texture, mipmapLevel, width, height, depth, pitchP, sliceP); int width = extent.width;
int height = extent.height;
int layers = imageView->getSubresourceRange().layerCount; // TODO(b/129523279): Untangle depth vs layers throughout the sampler
int depth = layers > 1 ? layers : extent.depth;
int pitchP = imageView->rowPitchBytes(aspect, level, ImageView::SAMPLING) / format.bytes();
int sliceP = (layers > 1 ? imageView->layerPitchBytes(aspect, ImageView::SAMPLING) : imageView->slicePitchBytes(aspect, level, ImageView::SAMPLING)) / format.bytes();
WriteTextureLevelInfo(texture, mipmapLevel, width, height, depth, pitchP, sliceP);
}
} }
} }
} }
......
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