blob: d9845ad92985408b9e234f42b4a549f7d8c11c82 [file] [log] [blame]
// Copyright 2016 The SwiftShader Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "SamplerCore.hpp"
#include "Constants.hpp"
#include "PixelRoutine.hpp"
#include "System/Debug.hpp"
#include "Vulkan/VkSampler.hpp"
namespace sw {
SamplerCore::SamplerCore(Pointer<Byte> &constants, const Sampler &state, SamplerFunction function)
: constants(constants)
, state(state)
, function(function)
{
}
SIMD::Float4 SamplerCore::sampleTexture(Pointer<Byte> &texture, SIMD::Float uvwa[4], const SIMD::Float &dRef, const Float &lodOrBias, const SIMD::Float &dsx, const SIMD::Float &dsy, SIMD::Int offset[4], const SIMD::Int &sample)
{
SIMD::Float4 c;
for(int i = 0; i < SIMD::Width / 4; i++)
{
Float4 uvwa128[4];
uvwa128[0] = Extract128(uvwa[0], i);
uvwa128[1] = Extract128(uvwa[1], i);
uvwa128[2] = Extract128(uvwa[2], i);
uvwa128[3] = Extract128(uvwa[3], i);
Vector4i offset128;
offset128[0] = Extract128(offset[0], i);
offset128[1] = Extract128(offset[1], i);
offset128[2] = Extract128(offset[2], i);
offset128[3] = Extract128(offset[3], i);
Vector4f c128 = sampleTexture128(texture, uvwa128, Extract128(dRef, i), lodOrBias, Extract128(dsx, i), Extract128(dsy, i), offset128, Extract128(sample, i));
c.x = Insert128(c.x, c128.x, i);
c.y = Insert128(c.y, c128.y, i);
c.z = Insert128(c.z, c128.z, i);
c.w = Insert128(c.w, c128.w, i);
}
return c;
}
Vector4f SamplerCore::sampleTexture128(Pointer<Byte> &texture, Float4 uvwa[4], const Float4 &dRef, const Float &lodOrBias, const Float4 &dsx, const Float4 &dsy, Vector4i &offset, const Int4 &sample)
{
Vector4f c;
Float4 u = uvwa[0];
Float4 v = uvwa[1];
Float4 w = uvwa[2];
Float4 a; // Array layer coordinate
switch(state.textureType)
{
case VK_IMAGE_VIEW_TYPE_1D_ARRAY: a = uvwa[1]; break;
case VK_IMAGE_VIEW_TYPE_2D_ARRAY: a = uvwa[2]; break;
case VK_IMAGE_VIEW_TYPE_CUBE_ARRAY: a = uvwa[3]; break;
default: break;
}
Float lod;
Float anisotropy;
Float4 uDelta;
Float4 vDelta;
Float4 M; // Major axis
if(state.isCube())
{
Int4 face = cubeFace(u, v, uvwa[0], uvwa[1], uvwa[2], M);
w = As<Float4>(face);
}
// Determine if we can skip the LOD computation. This is the case when the mipmap has only one level, except for LOD query,
// where we have to return the computed value. Anisotropic filtering requires computing the anisotropy factor even for a single mipmap level.
bool singleMipLevel = (state.minLod == state.maxLod);
bool requiresLodComputation = (function == Query) || (state.textureFilter == FILTER_ANISOTROPIC);
bool skipLodComputation = singleMipLevel && !requiresLodComputation;
if(skipLodComputation)
{
lod = state.minLod;
}
else if(function == Implicit || function == Bias || function == Grad || function == Query)
{
if(state.is1D())
{
computeLod1D(texture, lod, u, dsx, dsy);
}
else if(state.is2D())
{
computeLod2D(texture, lod, anisotropy, uDelta, vDelta, u, v, dsx, dsy);
}
else if(state.isCube())
{
computeLodCube(texture, lod, uvwa[0], uvwa[1], uvwa[2], dsx, dsy, M);
}
else
{
computeLod3D(texture, lod, u, v, w, dsx, dsy);
}
Float bias = state.mipLodBias;
if(function == Bias)
{
// Add SPIR-V Bias operand to the sampler provided bias and clamp to maxSamplerLodBias limit.
bias = Min(Max(bias + lodOrBias, -vk::MAX_SAMPLER_LOD_BIAS), vk::MAX_SAMPLER_LOD_BIAS);
}
lod += bias;
}
else if(function == Lod)
{
// Vulkan 1.1: "The absolute value of mipLodBias must be less than or equal to VkPhysicalDeviceLimits::maxSamplerLodBias"
// Hence no explicit clamping to maxSamplerLodBias is required in this case.
lod = lodOrBias + state.mipLodBias;
}
else if(function == Fetch)
{
// TODO: Eliminate int-float-int conversion.
lod = Float(As<Int>(lodOrBias));
}
else if(function == Base || function == Gather)
{
lod = Float(0);
}
else
UNREACHABLE("Sampler function %d", int(function));
if(function != Base && function != Fetch && function != Gather)
{
if(function == Query)
{
c.y = Float4(lod); // Unclamped LOD.
}
if(!skipLodComputation)
{
lod = Max(lod, state.minLod);
lod = Min(lod, state.maxLod);
}
if(function == Query)
{
if(state.mipmapFilter == MIPMAP_POINT)
{
lod = Round(lod); // TODO: Preferred formula is ceil(lod + 0.5) - 1
}
c.x = lod;
// c.y contains unclamped LOD.
return c;
}
}
bool force32BitFiltering = state.highPrecisionFiltering && !isYcbcrFormat() && (state.textureFilter != FILTER_POINT);
bool use32BitFiltering = hasFloatTexture() || hasUnnormalizedIntegerTexture() || force32BitFiltering ||
state.isCube() || state.unnormalizedCoordinates || state.compareEnable ||
borderModeActive() || (function == Gather) || (function == Fetch);
int numComponents = (function == Gather) ? 4 : textureComponentCount();
if(use32BitFiltering)
{
c = sampleFloatFilter(texture, u, v, w, a, dRef, offset, sample, lod, anisotropy, uDelta, vDelta);
}
else // 16-bit filtering.
{
Vector4s cs = sampleFilter(texture, u, v, w, a, offset, sample, lod, anisotropy, uDelta, vDelta);
for(int component = 0; component < numComponents; component++)
{
if(hasUnsignedTextureComponent(component))
{
c[component] = Float4(As<UShort4>(cs[component]));
}
else
{
c[component] = Float4(cs[component]);
}
}
}
if(hasNormalizedFormat() && !state.compareEnable)
{
sw::float4 scale = getComponentScale();
for(int component = 0; component < numComponents; component++)
{
int texelComponent = (function == Gather) ? getGatherComponent() : component;
c[component] *= Float4(1.0f / scale[texelComponent]);
}
}
if(state.textureFormat.isSignedNormalized())
{
for(int component = 0; component < numComponents; component++)
{
c[component] = Max(c[component], Float4(-1.0f));
}
}
if(state.textureFilter != FILTER_GATHER)
{
if((state.swizzle.r != VK_COMPONENT_SWIZZLE_R) ||
(state.swizzle.g != VK_COMPONENT_SWIZZLE_G) ||
(state.swizzle.b != VK_COMPONENT_SWIZZLE_B) ||
(state.swizzle.a != VK_COMPONENT_SWIZZLE_A))
{
const Vector4f col = c;
bool integer = hasUnnormalizedIntegerTexture();
c.x = applySwizzle(col, state.swizzle.r, integer);
c.y = applySwizzle(col, state.swizzle.g, integer);
c.z = applySwizzle(col, state.swizzle.b, integer);
c.w = applySwizzle(col, state.swizzle.a, integer);
}
}
else // Gather
{
VkComponentSwizzle swizzle = gatherSwizzle();
// R/G/B/A swizzles affect the component collected from each texel earlier.
// Handle the ZERO and ONE cases here because we don't need to know the format.
if(swizzle == VK_COMPONENT_SWIZZLE_ZERO)
{
c.x = c.y = c.z = c.w = Float4(0);
}
else if(swizzle == VK_COMPONENT_SWIZZLE_ONE)
{
bool integer = hasUnnormalizedIntegerTexture();
c.x = c.y = c.z = c.w = integer ? As<Float4>(Int4(1)) : RValue<Float4>(Float4(1.0f));
}
}
return c;
}
Float4 SamplerCore::applySwizzle(const Vector4f &c, VkComponentSwizzle swizzle, bool integer)
{
switch(swizzle)
{
default: UNSUPPORTED("VkComponentSwizzle %d", (int)swizzle);
case VK_COMPONENT_SWIZZLE_R: return c.x;
case VK_COMPONENT_SWIZZLE_G: return c.y;
case VK_COMPONENT_SWIZZLE_B: return c.z;
case VK_COMPONENT_SWIZZLE_A: return c.w;
case VK_COMPONENT_SWIZZLE_ZERO: return Float4(0.0f, 0.0f, 0.0f, 0.0f);
case VK_COMPONENT_SWIZZLE_ONE:
if(integer)
{
return Float4(As<Float4>(sw::Int4(1, 1, 1, 1)));
}
else
{
return Float4(1.0f, 1.0f, 1.0f, 1.0f);
}
break;
}
};
Short4 SamplerCore::offsetSample(Short4 &uvw, Pointer<Byte> &mipmap, int halfOffset, bool wrap, int count, Float &lod)
{
Short4 offset = *Pointer<Short4>(mipmap + halfOffset);
if(state.textureFilter == FILTER_MIN_LINEAR_MAG_POINT)
{
offset &= Short4(CmpNLE(Float4(lod), Float4(0.0f)));
}
else if(state.textureFilter == FILTER_MIN_POINT_MAG_LINEAR)
{
offset &= Short4(CmpLE(Float4(lod), Float4(0.0f)));
}
if(wrap)
{
switch(count)
{
case -1: return uvw - offset;
case 0: return uvw;
case +1: return uvw + offset;
case 2: return uvw + offset + offset;
}
}
else // Clamp or mirror
{
switch(count)
{
case -1: return SubSat(As<UShort4>(uvw), As<UShort4>(offset));
case 0: return uvw;
case +1: return AddSat(As<UShort4>(uvw), As<UShort4>(offset));
case 2: return AddSat(AddSat(As<UShort4>(uvw), As<UShort4>(offset)), As<UShort4>(offset));
}
}
return uvw;
}
Vector4s SamplerCore::sampleFilter(Pointer<Byte> &texture, Float4 &u, Float4 &v, Float4 &w, const Float4 &a, Vector4i &offset, const Int4 &sample, Float &lod, Float &anisotropy, Float4 &uDelta, Float4 &vDelta)
{
Vector4s c = sampleAniso(texture, u, v, w, a, offset, sample, lod, anisotropy, uDelta, vDelta, false);
if(function == Fetch)
{
return c;
}
if(state.mipmapFilter == MIPMAP_LINEAR)
{
Vector4s cc = sampleAniso(texture, u, v, w, a, offset, sample, lod, anisotropy, uDelta, vDelta, true);
lod *= Float(1 << 16);
UShort4 utri = UShort4(Float4(lod)); // TODO: Optimize
Short4 stri = utri >> 1; // TODO: Optimize
if(hasUnsignedTextureComponent(0))
cc.x = MulHigh(As<UShort4>(cc.x), utri);
else
cc.x = MulHigh(cc.x, stri);
if(hasUnsignedTextureComponent(1))
cc.y = MulHigh(As<UShort4>(cc.y), utri);
else
cc.y = MulHigh(cc.y, stri);
if(hasUnsignedTextureComponent(2))
cc.z = MulHigh(As<UShort4>(cc.z), utri);
else
cc.z = MulHigh(cc.z, stri);
if(hasUnsignedTextureComponent(3))
cc.w = MulHigh(As<UShort4>(cc.w), utri);
else
cc.w = MulHigh(cc.w, stri);
utri = ~utri;
stri = Short4(0x7FFF) - stri;
if(hasUnsignedTextureComponent(0))
c.x = MulHigh(As<UShort4>(c.x), utri);
else
c.x = MulHigh(c.x, stri);
if(hasUnsignedTextureComponent(1))
c.y = MulHigh(As<UShort4>(c.y), utri);
else
c.y = MulHigh(c.y, stri);
if(hasUnsignedTextureComponent(2))
c.z = MulHigh(As<UShort4>(c.z), utri);
else
c.z = MulHigh(c.z, stri);
if(hasUnsignedTextureComponent(3))
c.w = MulHigh(As<UShort4>(c.w), utri);
else
c.w = MulHigh(c.w, stri);
c.x += cc.x;
c.y += cc.y;
c.z += cc.z;
c.w += cc.w;
if(!hasUnsignedTextureComponent(0)) c.x += c.x;
if(!hasUnsignedTextureComponent(1)) c.y += c.y;
if(!hasUnsignedTextureComponent(2)) c.z += c.z;
if(!hasUnsignedTextureComponent(3)) c.w += c.w;
}
return c;
}
Vector4s SamplerCore::sampleAniso(Pointer<Byte> &texture, Float4 &u, Float4 &v, Float4 &w, const Float4 &a, Vector4i &offset, const Int4 &sample, Float &lod, Float &anisotropy, Float4 &uDelta, Float4 &vDelta, bool secondLOD)
{
Vector4s c;
if(state.textureFilter != FILTER_ANISOTROPIC)
{
c = sampleQuad(texture, u, v, w, a, offset, sample, lod, secondLOD);
}
else
{
Int N = RoundInt(anisotropy);
Vector4s cSum;
cSum.x = Short4(0);
cSum.y = Short4(0);
cSum.z = Short4(0);
cSum.w = Short4(0);
Float4 A = *Pointer<Float4>(constants + OFFSET(Constants, uvWeight) + 16 * N);
Float4 B = *Pointer<Float4>(constants + OFFSET(Constants, uvStart) + 16 * N);
UShort4 cw = *Pointer<UShort4>(constants + OFFSET(Constants, cWeight) + 8 * N);
Short4 sw = Short4(cw >> 1);
Float4 du = uDelta;
Float4 dv = vDelta;
Float4 u0 = u + B * du;
Float4 v0 = v + B * dv;
du *= A;
dv *= A;
Int i = 0;
Do
{
c = sampleQuad(texture, u0, v0, w, a, offset, sample, lod, secondLOD);
u0 += du;
v0 += dv;
if(hasUnsignedTextureComponent(0))
cSum.x += As<Short4>(MulHigh(As<UShort4>(c.x), cw));
else
cSum.x += MulHigh(c.x, sw);
if(hasUnsignedTextureComponent(1))
cSum.y += As<Short4>(MulHigh(As<UShort4>(c.y), cw));
else
cSum.y += MulHigh(c.y, sw);
if(hasUnsignedTextureComponent(2))
cSum.z += As<Short4>(MulHigh(As<UShort4>(c.z), cw));
else
cSum.z += MulHigh(c.z, sw);
if(hasUnsignedTextureComponent(3))
cSum.w += As<Short4>(MulHigh(As<UShort4>(c.w), cw));
else
cSum.w += MulHigh(c.w, sw);
i++;
}
Until(i >= N);
if(hasUnsignedTextureComponent(0))
c.x = cSum.x;
else
c.x = AddSat(cSum.x, cSum.x);
if(hasUnsignedTextureComponent(1))
c.y = cSum.y;
else
c.y = AddSat(cSum.y, cSum.y);
if(hasUnsignedTextureComponent(2))
c.z = cSum.z;
else
c.z = AddSat(cSum.z, cSum.z);
if(hasUnsignedTextureComponent(3))
c.w = cSum.w;
else
c.w = AddSat(cSum.w, cSum.w);
}
return c;
}
Vector4s SamplerCore::sampleQuad(Pointer<Byte> &texture, Float4 &u, Float4 &v, Float4 &w, const Float4 &a, Vector4i &offset, const Int4 &sample, Float &lod, bool secondLOD)
{
if(state.textureType != VK_IMAGE_VIEW_TYPE_3D)
{
return sampleQuad2D(texture, u, v, w, a, offset, sample, lod, secondLOD);
}
else
{
return sample3D(texture, u, v, w, offset, sample, lod, secondLOD);
}
}
Vector4s SamplerCore::sampleQuad2D(Pointer<Byte> &texture, Float4 &u, Float4 &v, Float4 &w, const Float4 &a, Vector4i &offset, const Int4 &sample, Float &lod, bool secondLOD)
{
Vector4s c;
int componentCount = textureComponentCount();
bool gather = (state.textureFilter == FILTER_GATHER);
Pointer<Byte> mipmap = selectMipmap(texture, lod, secondLOD);
Pointer<Byte> buffer = *Pointer<Pointer<Byte>>(mipmap + OFFSET(Mipmap, buffer));
Short4 uuuu = address(u, state.addressingModeU, mipmap);
Short4 vvvv = address(v, state.addressingModeV, mipmap);
Short4 wwww = address(w, state.addressingModeW, mipmap);
Short4 layerIndex = computeLayerIndex16(a, mipmap);
if(state.textureFilter == FILTER_POINT)
{
c = sampleTexel(uuuu, vvvv, wwww, layerIndex, offset, sample, mipmap, buffer);
}
else
{
Short4 uuuu0 = offsetSample(uuuu, mipmap, OFFSET(Mipmap, uHalf), state.addressingModeU == ADDRESSING_WRAP, -1, lod);
Short4 vvvv0 = offsetSample(vvvv, mipmap, OFFSET(Mipmap, vHalf), state.addressingModeV == ADDRESSING_WRAP, -1, lod);
Short4 uuuu1 = offsetSample(uuuu, mipmap, OFFSET(Mipmap, uHalf), state.addressingModeU == ADDRESSING_WRAP, +1, lod);
Short4 vvvv1 = offsetSample(vvvv, mipmap, OFFSET(Mipmap, vHalf), state.addressingModeV == ADDRESSING_WRAP, +1, lod);
Vector4s c00 = sampleTexel(uuuu0, vvvv0, wwww, layerIndex, offset, sample, mipmap, buffer);
Vector4s c10 = sampleTexel(uuuu1, vvvv0, wwww, layerIndex, offset, sample, mipmap, buffer);
Vector4s c01 = sampleTexel(uuuu0, vvvv1, wwww, layerIndex, offset, sample, mipmap, buffer);
Vector4s c11 = sampleTexel(uuuu1, vvvv1, wwww, layerIndex, offset, sample, mipmap, buffer);
if(!gather) // Blend
{
// Fractions
UShort4 f0u = As<UShort4>(uuuu0) * UShort4(*Pointer<UInt4>(mipmap + OFFSET(Mipmap, width)));
UShort4 f0v = As<UShort4>(vvvv0) * UShort4(*Pointer<UInt4>(mipmap + OFFSET(Mipmap, height)));
UShort4 f1u = ~f0u;
UShort4 f1v = ~f0v;
UShort4 f0u0v = MulHigh(f0u, f0v);
UShort4 f1u0v = MulHigh(f1u, f0v);
UShort4 f0u1v = MulHigh(f0u, f1v);
UShort4 f1u1v = MulHigh(f1u, f1v);
// Signed fractions
Short4 f1u1vs;
Short4 f0u1vs;
Short4 f1u0vs;
Short4 f0u0vs;
if(!hasUnsignedTextureComponent(0) || !hasUnsignedTextureComponent(1) || !hasUnsignedTextureComponent(2) || !hasUnsignedTextureComponent(3))
{
f1u1vs = f1u1v >> 1;
f0u1vs = f0u1v >> 1;
f1u0vs = f1u0v >> 1;
f0u0vs = f0u0v >> 1;
}
// Bilinear interpolation
if(componentCount >= 1)
{
if(has16bitTextureComponents() && hasUnsignedTextureComponent(0))
{
c00.x = As<UShort4>(c00.x) - MulHigh(As<UShort4>(c00.x), f0u) + MulHigh(As<UShort4>(c10.x), f0u);
c01.x = As<UShort4>(c01.x) - MulHigh(As<UShort4>(c01.x), f0u) + MulHigh(As<UShort4>(c11.x), f0u);
c.x = As<UShort4>(c00.x) - MulHigh(As<UShort4>(c00.x), f0v) + MulHigh(As<UShort4>(c01.x), f0v);
}
else
{
if(hasUnsignedTextureComponent(0))
{
c00.x = MulHigh(As<UShort4>(c00.x), f1u1v);
c10.x = MulHigh(As<UShort4>(c10.x), f0u1v);
c01.x = MulHigh(As<UShort4>(c01.x), f1u0v);
c11.x = MulHigh(As<UShort4>(c11.x), f0u0v);
}
else
{
c00.x = MulHigh(c00.x, f1u1vs);
c10.x = MulHigh(c10.x, f0u1vs);
c01.x = MulHigh(c01.x, f1u0vs);
c11.x = MulHigh(c11.x, f0u0vs);
}
c.x = (c00.x + c10.x) + (c01.x + c11.x);
if(!hasUnsignedTextureComponent(0)) c.x = AddSat(c.x, c.x); // Correct for signed fractions
}
}
if(componentCount >= 2)
{
if(has16bitTextureComponents() && hasUnsignedTextureComponent(1))
{
c00.y = As<UShort4>(c00.y) - MulHigh(As<UShort4>(c00.y), f0u) + MulHigh(As<UShort4>(c10.y), f0u);
c01.y = As<UShort4>(c01.y) - MulHigh(As<UShort4>(c01.y), f0u) + MulHigh(As<UShort4>(c11.y), f0u);
c.y = As<UShort4>(c00.y) - MulHigh(As<UShort4>(c00.y), f0v) + MulHigh(As<UShort4>(c01.y), f0v);
}
else
{
if(hasUnsignedTextureComponent(1))
{
c00.y = MulHigh(As<UShort4>(c00.y), f1u1v);
c10.y = MulHigh(As<UShort4>(c10.y), f0u1v);
c01.y = MulHigh(As<UShort4>(c01.y), f1u0v);
c11.y = MulHigh(As<UShort4>(c11.y), f0u0v);
}
else
{
c00.y = MulHigh(c00.y, f1u1vs);
c10.y = MulHigh(c10.y, f0u1vs);
c01.y = MulHigh(c01.y, f1u0vs);
c11.y = MulHigh(c11.y, f0u0vs);
}
c.y = (c00.y + c10.y) + (c01.y + c11.y);
if(!hasUnsignedTextureComponent(1)) c.y = AddSat(c.y, c.y); // Correct for signed fractions
}
}
if(componentCount >= 3)
{
if(has16bitTextureComponents() && hasUnsignedTextureComponent(2))
{
c00.z = As<UShort4>(c00.z) - MulHigh(As<UShort4>(c00.z), f0u) + MulHigh(As<UShort4>(c10.z), f0u);
c01.z = As<UShort4>(c01.z) - MulHigh(As<UShort4>(c01.z), f0u) + MulHigh(As<UShort4>(c11.z), f0u);
c.z = As<UShort4>(c00.z) - MulHigh(As<UShort4>(c00.z), f0v) + MulHigh(As<UShort4>(c01.z), f0v);
}
else
{
if(hasUnsignedTextureComponent(2))
{
c00.z = MulHigh(As<UShort4>(c00.z), f1u1v);
c10.z = MulHigh(As<UShort4>(c10.z), f0u1v);
c01.z = MulHigh(As<UShort4>(c01.z), f1u0v);
c11.z = MulHigh(As<UShort4>(c11.z), f0u0v);
}
else
{
c00.z = MulHigh(c00.z, f1u1vs);
c10.z = MulHigh(c10.z, f0u1vs);
c01.z = MulHigh(c01.z, f1u0vs);
c11.z = MulHigh(c11.z, f0u0vs);
}
c.z = (c00.z + c10.z) + (c01.z + c11.z);
if(!hasUnsignedTextureComponent(2)) c.z = AddSat(c.z, c.z); // Correct for signed fractions
}
}
if(componentCount >= 4)
{
if(has16bitTextureComponents() && hasUnsignedTextureComponent(3))
{
c00.w = As<UShort4>(c00.w) - MulHigh(As<UShort4>(c00.w), f0u) + MulHigh(As<UShort4>(c10.w), f0u);
c01.w = As<UShort4>(c01.w) - MulHigh(As<UShort4>(c01.w), f0u) + MulHigh(As<UShort4>(c11.w), f0u);
c.w = As<UShort4>(c00.w) - MulHigh(As<UShort4>(c00.w), f0v) + MulHigh(As<UShort4>(c01.w), f0v);
}
else
{
if(hasUnsignedTextureComponent(3))
{
c00.w = MulHigh(As<UShort4>(c00.w), f1u1v);
c10.w = MulHigh(As<UShort4>(c10.w), f0u1v);
c01.w = MulHigh(As<UShort4>(c01.w), f1u0v);
c11.w = MulHigh(As<UShort4>(c11.w), f0u0v);
}
else
{
c00.w = MulHigh(c00.w, f1u1vs);
c10.w = MulHigh(c10.w, f0u1vs);
c01.w = MulHigh(c01.w, f1u0vs);
c11.w = MulHigh(c11.w, f0u0vs);
}
c.w = (c00.w + c10.w) + (c01.w + c11.w);
if(!hasUnsignedTextureComponent(3)) c.w = AddSat(c.w, c.w); // Correct for signed fractions
}
}
}
else // Gather
{
VkComponentSwizzle swizzle = gatherSwizzle();
switch(swizzle)
{
case VK_COMPONENT_SWIZZLE_ZERO:
case VK_COMPONENT_SWIZZLE_ONE:
// Handled at the final component swizzle.
break;
default:
c.x = c01[swizzle - VK_COMPONENT_SWIZZLE_R];
c.y = c11[swizzle - VK_COMPONENT_SWIZZLE_R];
c.z = c10[swizzle - VK_COMPONENT_SWIZZLE_R];
c.w = c00[swizzle - VK_COMPONENT_SWIZZLE_R];
break;
}
}
}
return c;
}
Vector4s SamplerCore::sample3D(Pointer<Byte> &texture, Float4 &u_, Float4 &v_, Float4 &w_, Vector4i &offset, const Int4 &sample, Float &lod, bool secondLOD)
{
Vector4s c_;
int componentCount = textureComponentCount();
Pointer<Byte> mipmap = selectMipmap(texture, lod, secondLOD);
Pointer<Byte> buffer = *Pointer<Pointer<Byte>>(mipmap + OFFSET(Mipmap, buffer));
Short4 uuuu = address(u_, state.addressingModeU, mipmap);
Short4 vvvv = address(v_, state.addressingModeV, mipmap);
Short4 wwww = address(w_, state.addressingModeW, mipmap);
if(state.textureFilter == FILTER_POINT)
{
c_ = sampleTexel(uuuu, vvvv, wwww, 0, offset, sample, mipmap, buffer);
}
else
{
Vector4s c[2][2][2];
Short4 u[2][2][2];
Short4 v[2][2][2];
Short4 s[2][2][2];
for(int i = 0; i < 2; i++)
{
for(int j = 0; j < 2; j++)
{
for(int k = 0; k < 2; k++)
{
u[i][j][k] = offsetSample(uuuu, mipmap, OFFSET(Mipmap, uHalf), state.addressingModeU == ADDRESSING_WRAP, i * 2 - 1, lod);
v[i][j][k] = offsetSample(vvvv, mipmap, OFFSET(Mipmap, vHalf), state.addressingModeV == ADDRESSING_WRAP, j * 2 - 1, lod);
s[i][j][k] = offsetSample(wwww, mipmap, OFFSET(Mipmap, wHalf), state.addressingModeW == ADDRESSING_WRAP, k * 2 - 1, lod);
}
}
}
// Fractions
UShort4 f0u = As<UShort4>(u[0][0][0]) * UShort4(*Pointer<UInt4>(mipmap + OFFSET(Mipmap, width)));
UShort4 f0v = As<UShort4>(v[0][0][0]) * UShort4(*Pointer<UInt4>(mipmap + OFFSET(Mipmap, height)));
UShort4 f0s = As<UShort4>(s[0][0][0]) * UShort4(*Pointer<UInt4>(mipmap + OFFSET(Mipmap, depth)));
UShort4 f1u = ~f0u;
UShort4 f1v = ~f0v;
UShort4 f1s = ~f0s;
UShort4 f[2][2][2];
Short4 fs[2][2][2];
f[1][1][1] = MulHigh(f1u, f1v);
f[0][1][1] = MulHigh(f0u, f1v);
f[1][0][1] = MulHigh(f1u, f0v);
f[0][0][1] = MulHigh(f0u, f0v);
f[1][1][0] = MulHigh(f1u, f1v);
f[0][1][0] = MulHigh(f0u, f1v);
f[1][0][0] = MulHigh(f1u, f0v);
f[0][0][0] = MulHigh(f0u, f0v);
f[1][1][1] = MulHigh(f[1][1][1], f1s);
f[0][1][1] = MulHigh(f[0][1][1], f1s);
f[1][0][1] = MulHigh(f[1][0][1], f1s);
f[0][0][1] = MulHigh(f[0][0][1], f1s);
f[1][1][0] = MulHigh(f[1][1][0], f0s);
f[0][1][0] = MulHigh(f[0][1][0], f0s);
f[1][0][0] = MulHigh(f[1][0][0], f0s);
f[0][0][0] = MulHigh(f[0][0][0], f0s);
// Signed fractions
if(!hasUnsignedTextureComponent(0) || !hasUnsignedTextureComponent(1) || !hasUnsignedTextureComponent(2) || !hasUnsignedTextureComponent(3))
{
fs[0][0][0] = f[0][0][0] >> 1;
fs[0][0][1] = f[0][0][1] >> 1;
fs[0][1][0] = f[0][1][0] >> 1;
fs[0][1][1] = f[0][1][1] >> 1;
fs[1][0][0] = f[1][0][0] >> 1;
fs[1][0][1] = f[1][0][1] >> 1;
fs[1][1][0] = f[1][1][0] >> 1;
fs[1][1][1] = f[1][1][1] >> 1;
}
for(int i = 0; i < 2; i++)
{
for(int j = 0; j < 2; j++)
{
for(int k = 0; k < 2; k++)
{
c[i][j][k] = sampleTexel(u[i][j][k], v[i][j][k], s[i][j][k], 0, offset, sample, mipmap, buffer);
if(componentCount >= 1)
{
if(hasUnsignedTextureComponent(0))
c[i][j][k].x = MulHigh(As<UShort4>(c[i][j][k].x), f[1 - i][1 - j][1 - k]);
else
c[i][j][k].x = MulHigh(c[i][j][k].x, fs[1 - i][1 - j][1 - k]);
}
if(componentCount >= 2)
{
if(hasUnsignedTextureComponent(1))
c[i][j][k].y = MulHigh(As<UShort4>(c[i][j][k].y), f[1 - i][1 - j][1 - k]);
else
c[i][j][k].y = MulHigh(c[i][j][k].y, fs[1 - i][1 - j][1 - k]);
}
if(componentCount >= 3)
{
if(hasUnsignedTextureComponent(2))
c[i][j][k].z = MulHigh(As<UShort4>(c[i][j][k].z), f[1 - i][1 - j][1 - k]);
else
c[i][j][k].z = MulHigh(c[i][j][k].z, fs[1 - i][1 - j][1 - k]);
}
if(componentCount >= 4)
{
if(hasUnsignedTextureComponent(3))
c[i][j][k].w = MulHigh(As<UShort4>(c[i][j][k].w), f[1 - i][1 - j][1 - k]);
else
c[i][j][k].w = MulHigh(c[i][j][k].w, fs[1 - i][1 - j][1 - k]);
}
if(i != 0 || j != 0 || k != 0)
{
if(componentCount >= 1) c[0][0][0].x += c[i][j][k].x;
if(componentCount >= 2) c[0][0][0].y += c[i][j][k].y;
if(componentCount >= 3) c[0][0][0].z += c[i][j][k].z;
if(componentCount >= 4) c[0][0][0].w += c[i][j][k].w;
}
}
}
}
if(componentCount >= 1) c_.x = c[0][0][0].x;
if(componentCount >= 2) c_.y = c[0][0][0].y;
if(componentCount >= 3) c_.z = c[0][0][0].z;
if(componentCount >= 4) c_.w = c[0][0][0].w;
// Correct for signed fractions
if(componentCount >= 1)
if(!hasUnsignedTextureComponent(0)) c_.x = AddSat(c_.x, c_.x);
if(componentCount >= 2)
if(!hasUnsignedTextureComponent(1)) c_.y = AddSat(c_.y, c_.y);
if(componentCount >= 3)
if(!hasUnsignedTextureComponent(2)) c_.z = AddSat(c_.z, c_.z);
if(componentCount >= 4)
if(!hasUnsignedTextureComponent(3)) c_.w = AddSat(c_.w, c_.w);
}
return c_;
}
Vector4f SamplerCore::sampleFloatFilter(Pointer<Byte> &texture, Float4 &u, Float4 &v, Float4 &w, const Float4 &a, const Float4 &dRef, Vector4i &offset, const Int4 &sample, Float &lod, Float &anisotropy, Float4 &uDelta, Float4 &vDelta)
{
Vector4f c = sampleFloatAniso(texture, u, v, w, a, dRef, offset, sample, lod, anisotropy, uDelta, vDelta, false);
if(function == Fetch)
{
return c;
}
if(state.mipmapFilter == MIPMAP_LINEAR)
{
Vector4f cc = sampleFloatAniso(texture, u, v, w, a, dRef, offset, sample, lod, anisotropy, uDelta, vDelta, true);
Float4 lod4 = Float4(Frac(lod));
c.x = (cc.x - c.x) * lod4 + c.x;
c.y = (cc.y - c.y) * lod4 + c.y;
c.z = (cc.z - c.z) * lod4 + c.z;
c.w = (cc.w - c.w) * lod4 + c.w;
}
return c;
}
Vector4f SamplerCore::sampleFloatAniso(Pointer<Byte> &texture, Float4 &u, Float4 &v, Float4 &w, const Float4 &a, const Float4 &dRef, Vector4i &offset, const Int4 &sample, Float &lod, Float &anisotropy, Float4 &uDelta, Float4 &vDelta, bool secondLOD)
{
Vector4f c;
if(state.textureFilter != FILTER_ANISOTROPIC)
{
c = sampleFloat(texture, u, v, w, a, dRef, offset, sample, lod, secondLOD);
}
else
{
Int N = RoundInt(anisotropy);
Vector4f cSum;
cSum.x = Float4(0.0f);
cSum.y = Float4(0.0f);
cSum.z = Float4(0.0f);
cSum.w = Float4(0.0f);
Float4 A = *Pointer<Float4>(constants + OFFSET(Constants, uvWeight) + 16 * N);
Float4 B = *Pointer<Float4>(constants + OFFSET(Constants, uvStart) + 16 * N);
Float4 du = uDelta;
Float4 dv = vDelta;
Float4 u0 = u + B * du;
Float4 v0 = v + B * dv;
du *= A;
dv *= A;
Int i = 0;
Do
{
c = sampleFloat(texture, u0, v0, w, a, dRef, offset, sample, lod, secondLOD);
u0 += du;
v0 += dv;
cSum.x += c.x * A;
cSum.y += c.y * A;
cSum.z += c.z * A;
cSum.w += c.w * A;
i++;
}
Until(i >= N);
c.x = cSum.x;
c.y = cSum.y;
c.z = cSum.z;
c.w = cSum.w;
}
return c;
}
Vector4f SamplerCore::sampleFloat(Pointer<Byte> &texture, Float4 &u, Float4 &v, Float4 &w, const Float4 &a, const Float4 &dRef, Vector4i &offset, const Int4 &sample, Float &lod, bool secondLOD)
{
if(state.textureType != VK_IMAGE_VIEW_TYPE_3D)
{
return sampleFloat2D(texture, u, v, w, a, dRef, offset, sample, lod, secondLOD);
}
else
{
return sampleFloat3D(texture, u, v, w, dRef, offset, sample, lod, secondLOD);
}
}
Vector4f SamplerCore::sampleFloat2D(Pointer<Byte> &texture, Float4 &u, Float4 &v, Float4 &w, const Float4 &a, const Float4 &dRef, Vector4i &offset, const Int4 &sample, Float &lod, bool secondLOD)
{
Vector4f c;
int componentCount = textureComponentCount();
bool gather = (state.textureFilter == FILTER_GATHER);
Pointer<Byte> mipmap = selectMipmap(texture, lod, secondLOD);
Pointer<Byte> buffer = *Pointer<Pointer<Byte>>(mipmap + OFFSET(Mipmap, buffer));
Int4 x0, x1, y0, y1;
Float4 fu, fv;
Int4 filter = computeFilterOffset(lod);
address(u, x0, x1, fu, mipmap, offset.x, filter, OFFSET(Mipmap, width), state.addressingModeU);
address(v, y0, y1, fv, mipmap, offset.y, filter, OFFSET(Mipmap, height), state.addressingModeV);
Int4 pitchP = As<Int4>(*Pointer<UInt4>(mipmap + OFFSET(Mipmap, pitchP), 16));
y0 *= pitchP;
Int4 z;
if(state.isCube() || state.isArrayed())
{
Int4 face = As<Int4>(w);
Int4 layerIndex = computeLayerIndex(a, mipmap);
// For cube maps, the layer argument is per cube, each of which has 6 layers
if(state.textureType == VK_IMAGE_VIEW_TYPE_CUBE_ARRAY)
{
layerIndex *= Int4(6);
}
z = state.isCube() ? face : layerIndex;
if(state.textureType == VK_IMAGE_VIEW_TYPE_CUBE_ARRAY)
{
z += layerIndex;
}
z *= *Pointer<Int4>(mipmap + OFFSET(Mipmap, sliceP), 16);
}
if(state.textureFilter == FILTER_POINT || (function == Fetch))
{
c = sampleTexel(x0, y0, z, dRef, sample, mipmap, buffer);
}
else
{
y1 *= pitchP;
Vector4f c00 = sampleTexel(x0, y0, z, dRef, sample, mipmap, buffer);
Vector4f c10 = sampleTexel(x1, y0, z, dRef, sample, mipmap, buffer);
Vector4f c01 = sampleTexel(x0, y1, z, dRef, sample, mipmap, buffer);
Vector4f c11 = sampleTexel(x1, y1, z, dRef, sample, mipmap, buffer);
if(!gather) // Blend
{
if(componentCount >= 1) c00.x = c00.x + fu * (c10.x - c00.x);
if(componentCount >= 2) c00.y = c00.y + fu * (c10.y - c00.y);
if(componentCount >= 3) c00.z = c00.z + fu * (c10.z - c00.z);
if(componentCount >= 4) c00.w = c00.w + fu * (c10.w - c00.w);
if(componentCount >= 1) c01.x = c01.x + fu * (c11.x - c01.x);
if(componentCount >= 2) c01.y = c01.y + fu * (c11.y - c01.y);
if(componentCount >= 3) c01.z = c01.z + fu * (c11.z - c01.z);
if(componentCount >= 4) c01.w = c01.w + fu * (c11.w - c01.w);
if(componentCount >= 1) c.x = c00.x + fv * (c01.x - c00.x);
if(componentCount >= 2) c.y = c00.y + fv * (c01.y - c00.y);
if(componentCount >= 3) c.z = c00.z + fv * (c01.z - c00.z);
if(componentCount >= 4) c.w = c00.w + fv * (c01.w - c00.w);
}
else // Gather
{
VkComponentSwizzle swizzle = gatherSwizzle();
switch(swizzle)
{
case VK_COMPONENT_SWIZZLE_ZERO:
case VK_COMPONENT_SWIZZLE_ONE:
// Handled at the final component swizzle.
break;
default:
c.x = c01[swizzle - VK_COMPONENT_SWIZZLE_R];
c.y = c11[swizzle - VK_COMPONENT_SWIZZLE_R];
c.z = c10[swizzle - VK_COMPONENT_SWIZZLE_R];
c.w = c00[swizzle - VK_COMPONENT_SWIZZLE_R];
break;
}
}
}
return c;
}
Vector4f SamplerCore::sampleFloat3D(Pointer<Byte> &texture, Float4 &u, Float4 &v, Float4 &w, const Float4 &dRef, Vector4i &offset, const Int4 &sample, Float &lod, bool secondLOD)
{
Vector4f c;
int componentCount = textureComponentCount();
Pointer<Byte> mipmap = selectMipmap(texture, lod, secondLOD);
Pointer<Byte> buffer = *Pointer<Pointer<Byte>>(mipmap + OFFSET(Mipmap, buffer));
Int4 x0, x1, y0, y1, z0, z1;
Float4 fu, fv, fw;
Int4 filter = computeFilterOffset(lod);
address(u, x0, x1, fu, mipmap, offset.x, filter, OFFSET(Mipmap, width), state.addressingModeU);
address(v, y0, y1, fv, mipmap, offset.y, filter, OFFSET(Mipmap, height), state.addressingModeV);
address(w, z0, z1, fw, mipmap, offset.z, filter, OFFSET(Mipmap, depth), state.addressingModeW);
Int4 pitchP = As<Int4>(*Pointer<UInt4>(mipmap + OFFSET(Mipmap, pitchP), 16));
Int4 sliceP = As<Int4>(*Pointer<UInt4>(mipmap + OFFSET(Mipmap, sliceP), 16));
y0 *= pitchP;
z0 *= sliceP;
if(state.textureFilter == FILTER_POINT || (function == Fetch))
{
c = sampleTexel(x0, y0, z0, dRef, sample, mipmap, buffer);
}
else
{
y1 *= pitchP;
z1 *= sliceP;
Vector4f c000 = sampleTexel(x0, y0, z0, dRef, sample, mipmap, buffer);
Vector4f c100 = sampleTexel(x1, y0, z0, dRef, sample, mipmap, buffer);
Vector4f c010 = sampleTexel(x0, y1, z0, dRef, sample, mipmap, buffer);
Vector4f c110 = sampleTexel(x1, y1, z0, dRef, sample, mipmap, buffer);
Vector4f c001 = sampleTexel(x0, y0, z1, dRef, sample, mipmap, buffer);
Vector4f c101 = sampleTexel(x1, y0, z1, dRef, sample, mipmap, buffer);
Vector4f c011 = sampleTexel(x0, y1, z1, dRef, sample, mipmap, buffer);
Vector4f c111 = sampleTexel(x1, y1, z1, dRef, sample, mipmap, buffer);
// Blend first slice
if(componentCount >= 1) c000.x = c000.x + fu * (c100.x - c000.x);
if(componentCount >= 2) c000.y = c000.y + fu * (c100.y - c000.y);
if(componentCount >= 3) c000.z = c000.z + fu * (c100.z - c000.z);
if(componentCount >= 4) c000.w = c000.w + fu * (c100.w - c000.w);
if(componentCount >= 1) c010.x = c010.x + fu * (c110.x - c010.x);
if(componentCount >= 2) c010.y = c010.y + fu * (c110.y - c010.y);
if(componentCount >= 3) c010.z = c010.z + fu * (c110.z - c010.z);
if(componentCount >= 4) c010.w = c010.w + fu * (c110.w - c010.w);
if(componentCount >= 1) c000.x = c000.x + fv * (c010.x - c000.x);
if(componentCount >= 2) c000.y = c000.y + fv * (c010.y - c000.y);
if(componentCount >= 3) c000.z = c000.z + fv * (c010.z - c000.z);
if(componentCount >= 4) c000.w = c000.w + fv * (c010.w - c000.w);
// Blend second slice
if(componentCount >= 1) c001.x = c001.x + fu * (c101.x - c001.x);
if(componentCount >= 2) c001.y = c001.y + fu * (c101.y - c001.y);
if(componentCount >= 3) c001.z = c001.z + fu * (c101.z - c001.z);
if(componentCount >= 4) c001.w = c001.w + fu * (c101.w - c001.w);
if(componentCount >= 1) c011.x = c011.x + fu * (c111.x - c011.x);
if(componentCount >= 2) c011.y = c011.y + fu * (c111.y - c011.y);
if(componentCount >= 3) c011.z = c011.z + fu * (c111.z - c011.z);
if(componentCount >= 4) c011.w = c011.w + fu * (c111.w - c011.w);
if(componentCount >= 1) c001.x = c001.x + fv * (c011.x - c001.x);
if(componentCount >= 2) c001.y = c001.y + fv * (c011.y - c001.y);
if(componentCount >= 3) c001.z = c001.z + fv * (c011.z - c001.z);
if(componentCount >= 4) c001.w = c001.w + fv * (c011.w - c001.w);
// Blend slices
if(componentCount >= 1) c.x = c000.x + fw * (c001.x - c000.x);
if(componentCount >= 2) c.y = c000.y + fw * (c001.y - c000.y);
if(componentCount >= 3) c.z = c000.z + fw * (c001.z - c000.z);
if(componentCount >= 4) c.w = c000.w + fw * (c001.w - c000.w);
}
return c;
}
static Float log2sqrt(Float lod)
{
// log2(sqrt(lod)) // Equals 0.25 * log2(lod^2).
lod *= lod; // Squaring doubles the exponent and produces an extra bit of precision.
lod = Float(As<Int>(lod)) - Float(0x3F800000); // Interpret as integer and subtract the exponent bias.
lod *= As<Float>(Int(0x33000000)); // Scale by 0.25 * 2^-23 (mantissa length).
return lod;
}
static Float log2(Float lod)
{
lod *= lod; // Squaring doubles the exponent and produces an extra bit of precision.
lod = Float(As<Int>(lod)) - Float(0x3F800000); // Interpret as integer and subtract the exponent bias.
lod *= As<Float>(Int(0x33800000)); // Scale by 0.5 * 2^-23 (mantissa length).
return lod;
}
void SamplerCore::computeLod1D(Pointer<Byte> &texture, Float &lod, Float4 &uuuu, const Float4 &dsx, const Float4 &dsy)
{
Float4 dudxy;
if(function != Grad) // Implicit
{
dudxy = uuuu.yz - uuuu.xx;
}
else
{
dudxy = UnpackLow(dsx, dsy);
}
// Scale by texture dimensions.
Float4 dUdxy = dudxy * *Pointer<Float4>(texture + OFFSET(Texture, widthWidthHeightHeight));
// Note we could take the absolute value here and omit the square root below,
// but this is more consistent with the 2D calculation and still cheap.
Float4 dU2dxy = dUdxy * dUdxy;
lod = Max(Float(dU2dxy.x), Float(dU2dxy.y));
lod = log2sqrt(lod);
}
void SamplerCore::computeLod2D(Pointer<Byte> &texture, Float &lod, Float &anisotropy, Float4 &uDelta, Float4 &vDelta, Float4 &uuuu, Float4 &vvvv, const Float4 &dsx, const Float4 &dsy)
{
Float4 duvdxy;
if(function != Grad) // Implicit
{
duvdxy = Float4(uuuu.yz, vvvv.yz) - Float4(uuuu.xx, vvvv.xx);
}
else
{
Float4 dudxy = Float4(dsx.xx, dsy.xx);
Float4 dvdxy = Float4(dsx.yy, dsy.yy);
duvdxy = Float4(dudxy.xz, dvdxy.xz);
}
// Scale by texture dimensions.
Float4 dUVdxy = duvdxy * *Pointer<Float4>(texture + OFFSET(Texture, widthWidthHeightHeight));
Float4 dUV2dxy = dUVdxy * dUVdxy;
Float4 dUV2 = dUV2dxy.xy + dUV2dxy.zw;
lod = Max(Float(dUV2.x), Float(dUV2.y)); // Square length of major axis
if(state.textureFilter == FILTER_ANISOTROPIC)
{
Float det = Abs(Float(dUVdxy.x) * Float(dUVdxy.w) - Float(dUVdxy.y) * Float(dUVdxy.z));
Float4 dudx = duvdxy.xxxx;
Float4 dudy = duvdxy.yyyy;
Float4 dvdx = duvdxy.zzzz;
Float4 dvdy = duvdxy.wwww;
Int4 mask = As<Int4>(CmpNLT(dUV2.x, dUV2.y));
uDelta = As<Float4>((As<Int4>(dudx) & mask) | ((As<Int4>(dudy) & ~mask)));
vDelta = As<Float4>((As<Int4>(dvdx) & mask) | ((As<Int4>(dvdy) & ~mask)));
anisotropy = lod * Rcp(det, true /* relaxedPrecision */);
anisotropy = Min(anisotropy, state.maxAnisotropy);
// TODO(b/151263485): While we always need `lod` above, when there's only
// a single mipmap level the following calculations could be skipped.
lod *= Rcp(anisotropy * anisotropy, true /* relaxedPrecision */);
}
lod = log2sqrt(lod); // log2(sqrt(lod))
}
void SamplerCore::computeLodCube(Pointer<Byte> &texture, Float &lod, Float4 &u, Float4 &v, Float4 &w, const Float4 &dsx, const Float4 &dsy, Float4 &M)
{
Float4 dudxy, dvdxy, dsdxy;
if(function != Grad) // Implicit
{
Float4 U = u * M;
Float4 V = v * M;
Float4 W = w * M;
dudxy = Abs(U - U.xxxx);
dvdxy = Abs(V - V.xxxx);
dsdxy = Abs(W - W.xxxx);
}
else
{
dudxy = Float4(dsx.xx, dsy.xx);
dvdxy = Float4(dsx.yy, dsy.yy);
dsdxy = Float4(dsx.zz, dsy.zz);
dudxy = Abs(dudxy * Float4(M.x));
dvdxy = Abs(dvdxy * Float4(M.x));
dsdxy = Abs(dsdxy * Float4(M.x));
}
// Compute the largest Manhattan distance in two dimensions.
// This takes the footprint across adjacent faces into account.
Float4 duvdxy = dudxy + dvdxy;
Float4 dusdxy = dudxy + dsdxy;
Float4 dvsdxy = dvdxy + dsdxy;
dudxy = Max(Max(duvdxy, dusdxy), dvsdxy);
lod = Max(Float(dudxy.y), Float(dudxy.z)); // TODO: Max(dudxy.y, dudxy.z);
// Scale by texture dimension.
lod *= *Pointer<Float>(texture + OFFSET(Texture, width));
lod = log2(lod);
}
void SamplerCore::computeLod3D(Pointer<Byte> &texture, Float &lod, Float4 &uuuu, Float4 &vvvv, Float4 &wwww, const Float4 &dsx, const Float4 &dsy)
{
Float4 dudxy, dvdxy, dsdxy;
if(function != Grad) // Implicit
{
dudxy = uuuu - uuuu.xxxx;
dvdxy = vvvv - vvvv.xxxx;
dsdxy = wwww - wwww.xxxx;
}
else
{
dudxy = Float4(dsx.xx, dsy.xx);
dvdxy = Float4(dsx.yy, dsy.yy);
dsdxy = Float4(dsx.zz, dsy.zz);
}
// Scale by texture dimensions.
dudxy *= *Pointer<Float4>(texture + OFFSET(Texture, width));
dvdxy *= *Pointer<Float4>(texture + OFFSET(Texture, height));
dsdxy *= *Pointer<Float4>(texture + OFFSET(Texture, depth));
dudxy *= dudxy;
dvdxy *= dvdxy;
dsdxy *= dsdxy;
dudxy += dvdxy;
dudxy += dsdxy;
lod = Max(Float(dudxy.y), Float(dudxy.z)); // TODO: Max(dudxy.y, dudxy.z);
lod = log2sqrt(lod); // log2(sqrt(lod))
}
Int4 SamplerCore::cubeFace(Float4 &U, Float4 &V, Float4 &x, Float4 &y, Float4 &z, Float4 &M)
{
// TODO: Comply with Vulkan recommendation:
// Vulkan 1.1: "The rules should have as the first rule that rz wins over ry and rx, and the second rule that ry wins over rx."
Int4 xn = CmpLT(x, 0.0f); // x < 0
Int4 yn = CmpLT(y, 0.0f); // y < 0
Int4 zn = CmpLT(z, 0.0f); // z < 0
Float4 absX = Abs(x);
Float4 absY = Abs(y);
Float4 absZ = Abs(z);
Int4 xy = CmpNLE(absX, absY); // abs(x) > abs(y)
Int4 yz = CmpNLE(absY, absZ); // abs(y) > abs(z)
Int4 zx = CmpNLE(absZ, absX); // abs(z) > abs(x)
Int4 xMajor = xy & ~zx; // abs(x) > abs(y) && abs(x) > abs(z)
Int4 yMajor = yz & ~xy; // abs(y) > abs(z) && abs(y) > abs(x)
Int4 zMajor = zx & ~yz; // abs(z) > abs(x) && abs(z) > abs(y)
// FACE_POSITIVE_X = 000b
// FACE_NEGATIVE_X = 001b
// FACE_POSITIVE_Y = 010b
// FACE_NEGATIVE_Y = 011b
// FACE_POSITIVE_Z = 100b
// FACE_NEGATIVE_Z = 101b
Int yAxis = SignMask(yMajor);
Int zAxis = SignMask(zMajor);
Int4 n = ((xn & xMajor) | (yn & yMajor) | (zn & zMajor)) & Int4(0x80000000);
Int negative = SignMask(n);
Int faces = *Pointer<Int>(constants + OFFSET(Constants, transposeBit0) + negative * 4);
faces |= *Pointer<Int>(constants + OFFSET(Constants, transposeBit1) + yAxis * 4);
faces |= *Pointer<Int>(constants + OFFSET(Constants, transposeBit2) + zAxis * 4);
Int4 face;
face.x = faces & 0x7;
face.y = (faces >> 4) & 0x7;
face.z = (faces >> 8) & 0x7;
face.w = (faces >> 12) & 0x7;
M = Max(Max(absX, absY), absZ);
// U = xMajor ? (neg ^ -z) : ((zMajor & neg) ^ x)
U = As<Float4>((xMajor & (n ^ As<Int4>(-z))) | (~xMajor & ((zMajor & n) ^ As<Int4>(x))));
// V = !yMajor ? -y : (n ^ z)
V = As<Float4>((~yMajor & As<Int4>(-y)) | (yMajor & (n ^ As<Int4>(z))));
M = reciprocal(M) * 0.5f;
U = U * M + 0.5f;
V = V * M + 0.5f;
return face;
}
Short4 SamplerCore::applyOffset(Short4 &uvw, Int4 &offset, const Int4 &whd, AddressingMode mode)
{
Int4 tmp = Int4(As<UShort4>(uvw));
tmp = tmp + offset;
switch(mode)
{
case AddressingMode::ADDRESSING_WRAP:
tmp = (tmp + whd * Int4(-MIN_TEXEL_OFFSET)) % whd;
break;
case AddressingMode::ADDRESSING_CLAMP:
case AddressingMode::ADDRESSING_MIRROR:
case AddressingMode::ADDRESSING_MIRRORONCE:
case AddressingMode::ADDRESSING_BORDER: // TODO(b/29069044): Implement and test ADDRESSING_MIRROR, ADDRESSING_MIRRORONCE, ADDRESSING_BORDER
tmp = Min(Max(tmp, Int4(0)), whd - Int4(1));
break;
case AddressingMode::ADDRESSING_SEAMLESS:
ASSERT(false); // Cube sampling doesn't support offset.
default:
ASSERT(false);
}
return As<Short4>(UShort4(tmp));
}
void SamplerCore::computeIndices(UInt index[4], Short4 uuuu, Short4 vvvv, Short4 wwww, const Short4 &layerIndex, Vector4i &offset, const Int4 &sample, const Pointer<Byte> &mipmap)
{
uuuu = MulHigh(As<UShort4>(uuuu), UShort4(*Pointer<UInt4>(mipmap + OFFSET(Mipmap, width))));
if(function.offset)
{
uuuu = applyOffset(uuuu, offset.x, *Pointer<UInt4>(mipmap + OFFSET(Mipmap, width)), state.addressingModeU);
}
UInt4 indices = Int4(uuuu);
if(state.is2D() || state.is3D() || state.isCube())
{
vvvv = MulHigh(As<UShort4>(vvvv), UShort4(*Pointer<UInt4>(mipmap + OFFSET(Mipmap, height))));
if(function.offset)
{
vvvv = applyOffset(vvvv, offset.y, *Pointer<UInt4>(mipmap + OFFSET(Mipmap, height)), state.addressingModeV);
}
Short4 uv0uv1 = As<Short4>(UnpackLow(uuuu, vvvv));
Short4 uv2uv3 = As<Short4>(UnpackHigh(uuuu, vvvv));
Int2 i01 = MulAdd(uv0uv1, *Pointer<Short4>(mipmap + OFFSET(Mipmap, onePitchP)));
Int2 i23 = MulAdd(uv2uv3, *Pointer<Short4>(mipmap + OFFSET(Mipmap, onePitchP)));
indices = UInt4(As<UInt2>(i01), As<UInt2>(i23));
}
if(state.is3D())
{
wwww = MulHigh(As<UShort4>(wwww), UShort4(*Pointer<Int4>(mipmap + OFFSET(Mipmap, depth))));
if(function.offset)
{
wwww = applyOffset(wwww, offset.z, *Pointer<Int4>(mipmap + OFFSET(Mipmap, depth)), state.addressingModeW);
}
indices += As<UInt4>(Int4(As<UShort4>(wwww))) * *Pointer<UInt4>(mipmap + OFFSET(Mipmap, sliceP));
}
if(state.isArrayed())
{
Int4 layer = Int4(As<UShort4>(layerIndex));
if(state.textureType == VK_IMAGE_VIEW_TYPE_CUBE_ARRAY)
{
layer *= Int4(6);
}
UInt4 layerOffset = As<UInt4>(layer) * *Pointer<UInt4>(mipmap + OFFSET(Mipmap, sliceP));
indices += layerOffset;
}
if(function.sample)
{
UInt4 sampleOffset = Min(As<UInt4>(sample), *Pointer<UInt4>(mipmap + OFFSET(Mipmap, sampleMax), 16)) *
*Pointer<UInt4>(mipmap + OFFSET(Mipmap, samplePitchP), 16);
indices += sampleOffset;
}
index[0] = Extract(indices, 0);
index[1] = Extract(indices, 1);
index[2] = Extract(indices, 2);
index[3] = Extract(indices, 3);
}
void SamplerCore::computeIndices(UInt index[4], Int4 uuuu, Int4 vvvv, Int4 wwww, const Int4 &sample, Int4 valid, const Pointer<Byte> &mipmap)
{
UInt4 indices = uuuu;
if(state.is2D() || state.is3D() || state.isCube())
{
indices += As<UInt4>(vvvv);
}
if(state.is3D() || state.isCube() || state.isArrayed())
{
indices += As<UInt4>(wwww);
}
if(function.sample)
{
indices += Min(As<UInt4>(sample), *Pointer<UInt4>(mipmap + OFFSET(Mipmap, sampleMax), 16)) *
*Pointer<UInt4>(mipmap + OFFSET(Mipmap, samplePitchP), 16);
}
if(borderModeActive())
{
// Texels out of range are still sampled before being replaced
// with the border color, so sample them at linear index 0.
indices &= As<UInt4>(valid);
}
for(int i = 0; i < 4; i++)
{
index[i] = Extract(As<Int4>(indices), i);
}
}
Vector4s SamplerCore::sampleTexel(UInt index[4], Pointer<Byte> buffer)
{
Vector4s c;
if(has16bitPackedTextureFormat())
{
c.x = Insert(c.x, Pointer<Short>(buffer)[index[0]], 0);
c.x = Insert(c.x, Pointer<Short>(buffer)[index[1]], 1);
c.x = Insert(c.x, Pointer<Short>(buffer)[index[2]], 2);
c.x = Insert(c.x, Pointer<Short>(buffer)[index[3]], 3);
switch(state.textureFormat)
{
case VK_FORMAT_R5G6B5_UNORM_PACK16:
c.z = (c.x & Short4(0x001Fu)) << 11;
c.y = (c.x & Short4(0x07E0u)) << 5;
c.x = (c.x & Short4(0xF800u));
break;
case VK_FORMAT_B5G6R5_UNORM_PACK16:
c.z = (c.x & Short4(0xF800u));
c.y = (c.x & Short4(0x07E0u)) << 5;
c.x = (c.x & Short4(0x001Fu)) << 11;
break;
case VK_FORMAT_R4G4B4A4_UNORM_PACK16:
c.w = (c.x << 12) & Short4(0xF000u);
c.z = (c.x << 8) & Short4(0xF000u);
c.y = (c.x << 4) & Short4(0xF000u);
c.x = (c.x) & Short4(0xF000u);
break;
case VK_FORMAT_B4G4R4A4_UNORM_PACK16:
c.w = (c.x << 12) & Short4(0xF000u);
c.z = (c.x) & Short4(0xF000u);
c.y = (c.x << 4) & Short4(0xF000u);
c.x = (c.x << 8) & Short4(0xF000u);
break;
case VK_FORMAT_A4R4G4B4_UNORM_PACK16:
c.w = (c.x) & Short4(0xF000u);
c.z = (c.x << 12) & Short4(0xF000u);
c.y = (c.x << 8) & Short4(0xF000u);
c.x = (c.x << 4) & Short4(0xF000u);
break;
case VK_FORMAT_A4B4G4R4_UNORM_PACK16:
c.w = (c.x) & Short4(0xF000u);
c.z = (c.x << 4) & Short4(0xF000u);
c.y = (c.x << 8) & Short4(0xF000u);
c.x = (c.x << 12) & Short4(0xF000u);
break;
case VK_FORMAT_R5G5B5A1_UNORM_PACK16:
c.w = (c.x << 15) & Short4(0x8000u);
c.z = (c.x << 10) & Short4(0xF800u);
c.y = (c.x << 5) & Short4(0xF800u);
c.x = (c.x) & Short4(0xF800u);
break;
case VK_FORMAT_B5G5R5A1_UNORM_PACK16:
c.w = (c.x << 15) & Short4(0x8000u);
c.z = (c.x) & Short4(0xF800u);
c.y = (c.x << 5) & Short4(0xF800u);
c.x = (c.x << 10) & Short4(0xF800u);
break;
case VK_FORMAT_A1R5G5B5_UNORM_PACK16:
c.w = (c.x) & Short4(0x8000u);
c.z = (c.x << 11) & Short4(0xF800u);
c.y = (c.x << 6) & Short4(0xF800u);
c.x = (c.x << 1) & Short4(0xF800u);
break;
default:
ASSERT(false);
}
}
else if(has8bitTextureComponents())
{
switch(textureComponentCount())
{
case 4:
{
Byte4 c0 = Pointer<Byte4>(buffer)[index[0]];
Byte4 c1 = Pointer<Byte4>(buffer)[index[1]];
Byte4 c2 = Pointer<Byte4>(buffer)[index[2]];
Byte4 c3 = Pointer<Byte4>(buffer)[index[3]];
c.x = Unpack(c0, c1);
c.y = Unpack(c2, c3);
switch(state.textureFormat)
{
case VK_FORMAT_B8G8R8A8_UNORM:
case VK_FORMAT_B8G8R8A8_SRGB:
c.z = As<Short4>(UnpackLow(c.x, c.y));
c.x = As<Short4>(UnpackHigh(c.x, c.y));
c.y = c.z;
c.w = c.x;
c.z = UnpackLow(As<Byte8>(Short4(0)), As<Byte8>(c.z));
c.y = UnpackHigh(As<Byte8>(Short4(0)), As<Byte8>(c.y));
c.x = UnpackLow(As<Byte8>(Short4(0)), As<Byte8>(c.x));
c.w = UnpackHigh(As<Byte8>(Short4(0)), As<Byte8>(c.w));
break;
case VK_FORMAT_R8G8B8A8_UNORM:
case VK_FORMAT_R8G8B8A8_SNORM:
case VK_FORMAT_R8G8B8A8_SINT:
case VK_FORMAT_R8G8B8A8_SRGB:
case VK_FORMAT_A8B8G8R8_UNORM_PACK32:
case VK_FORMAT_A8B8G8R8_SNORM_PACK32:
case VK_FORMAT_A8B8G8R8_SINT_PACK32:
case VK_FORMAT_A8B8G8R8_SRGB_PACK32:
c.z = As<Short4>(UnpackHigh(c.x, c.y));
c.x = As<Short4>(UnpackLow(c.x, c.y));
c.y = c.x;
c.w = c.z;
c.x = UnpackLow(As<Byte8>(Short4(0)), As<Byte8>(c.x));
c.y = UnpackHigh(As<Byte8>(Short4(0)), As<Byte8>(c.y));
c.z = UnpackLow(As<Byte8>(Short4(0)), As<Byte8>(c.z));
c.w = UnpackHigh(As<Byte8>(Short4(0)), As<Byte8>(c.w));
// Propagate sign bit
if(state.textureFormat == VK_FORMAT_R8G8B8A8_SINT ||
state.textureFormat == VK_FORMAT_A8B8G8R8_SINT_PACK32)
{
c.x >>= 8;
c.y >>= 8;
c.z >>= 8;
c.w >>= 8;
}
break;
case VK_FORMAT_R8G8B8A8_UINT:
case VK_FORMAT_A8B8G8R8_UINT_PACK32:
c.z = As<Short4>(UnpackHigh(c.x, c.y));
c.x = As<Short4>(UnpackLow(c.x, c.y));
c.y = c.x;
c.w = c.z;
c.x = UnpackLow(As<Byte8>(c.x), As<Byte8>(Short4(0)));
c.y = UnpackHigh(As<Byte8>(c.y), As<Byte8>(Short4(0)));
c.z = UnpackLow(As<Byte8>(c.z), As<Byte8>(Short4(0)));
c.w = UnpackHigh(As<Byte8>(c.w), As<Byte8>(Short4(0)));
break;
default:
ASSERT(false);
}
}
break;
case 2:
c.x = Insert(c.x, Pointer<Short>(buffer)[index[0]], 0);
c.x = Insert(c.x, Pointer<Short>(buffer)[index[1]], 1);
c.x = Insert(c.x, Pointer<Short>(buffer)[index[2]], 2);
c.x = Insert(c.x, Pointer<Short>(buffer)[index[3]], 3);
switch(state.textureFormat)
{
case VK_FORMAT_R8G8_UNORM:
case VK_FORMAT_R8G8_SNORM:
case VK_FORMAT_R8G8_SRGB:
c.y = (c.x & Short4(0xFF00u));
c.x = (c.x << 8);
break;
case VK_FORMAT_R8G8_SINT:
c.y = c.x >> 8;
c.x = (c.x << 8) >> 8; // Propagate sign bit
break;
case VK_FORMAT_R8G8_UINT:
c.y = As<Short4>(As<UShort4>(c.x) >> 8);
c.x &= Short4(0x00FFu);
break;
default:
ASSERT(false);
}
break;
case 1:
{
Int c0 = Int(*Pointer<Byte>(buffer + index[0]));
Int c1 = Int(*Pointer<Byte>(buffer + index[1]));
Int c2 = Int(*Pointer<Byte>(buffer + index[2]));
Int c3 = Int(*Pointer<Byte>(buffer + index[3]));
c0 = c0 | (c1 << 8) | (c2 << 16) | (c3 << 24);
switch(state.textureFormat)
{
case VK_FORMAT_R8_SINT:
case VK_FORMAT_R8_UINT:
case VK_FORMAT_S8_UINT:
{
Int zero(0);
c.x = Unpack(As<Byte4>(c0), As<Byte4>(zero));
// Propagate sign bit
if(state.textureFormat == VK_FORMAT_R8_SINT)
{
c.x = (c.x << 8) >> 8;
}
}
break;
case VK_FORMAT_R8_SNORM:
case VK_FORMAT_R8_UNORM:
case VK_FORMAT_R8_SRGB:
// TODO: avoid populating the low bits at all.
c.x = Unpack(As<Byte4>(c0));
c.x &= Short4(0xFF00u);
break;
default:
c.x = Unpack(As<Byte4>(c0));
break;
}
}
break;
default:
ASSERT(false);
}
}
else if(has16bitTextureComponents())
{
switch(textureComponentCount())
{
case 4:
c.x = Pointer<Short4>(buffer)[index[0]];
c.y = Pointer<Short4>(buffer)[index[1]];
c.z = Pointer<Short4>(buffer)[index[2]];
c.w = Pointer<Short4>(buffer)[index[3]];
transpose4x4(c.x, c.y, c.z, c.w);
break;
case 2:
c.x = *Pointer<Short4>(buffer + 4 * index[0]);
c.x = As<Short4>(UnpackLow(c.x, *Pointer<Short4>(buffer + 4 * index[1])));
c.z = *Pointer<Short4>(buffer + 4 * index[2]);
c.z = As<Short4>(UnpackLow(c.z, *Pointer<Short4>(buffer + 4 * index[3])));
c.y = c.x;
c.x = UnpackLow(As<Int2>(c.x), As<Int2>(c.z));
c.y = UnpackHigh(As<Int2>(c.y), As<Int2>(c.z));
break;
case 1:
c.x = Insert(c.x, Pointer<Short>(buffer)[index[0]], 0);
c.x = Insert(c.x, Pointer<Short>(buffer)[index[1]], 1);
c.x = Insert(c.x, Pointer<Short>(buffer)[index[2]], 2);
c.x = Insert(c.x, Pointer<Short>(buffer)[index[3]], 3);
break;
default:
ASSERT(false);
}
}
else if(state.textureFormat == VK_FORMAT_A2B10G10R10_UNORM_PACK32)
{
Int4 cc;
cc = Insert(cc, Pointer<Int>(buffer)[index[0]], 0);
cc = Insert(cc, Pointer<Int>(buffer)[index[1]], 1);
cc = Insert(cc, Pointer<Int>(buffer)[index[2]], 2);
cc = Insert(cc, Pointer<Int>(buffer)[index[3]], 3);
c.x = Short4(cc << 6) & Short4(0xFFC0u);
c.y = Short4(cc >> 4) & Short4(0xFFC0u);
c.z = Short4(cc >> 14) & Short4(0xFFC0u);
c.w = Short4(cc >> 16) & Short4(0xC000u);
}
else if(state.textureFormat == VK_FORMAT_A2R10G10B10_UNORM_PACK32)
{
Int4 cc;
cc = Insert(cc, Pointer<Int>(buffer)[index[0]], 0);
cc = Insert(cc, Pointer<Int>(buffer)[index[1]], 1);
cc = Insert(cc, Pointer<Int>(buffer)[index[2]], 2);
cc = Insert(cc, Pointer<Int>(buffer)[index[3]], 3);
c.x = Short4(cc >> 14) & Short4(0xFFC0u);
c.y = Short4(cc >> 4) & Short4(0xFFC0u);
c.z = Short4(cc << 6) & Short4(0xFFC0u);
c.w = Short4(cc >> 16) & Short4(0xC000u);
}
else if(state.textureFormat == VK_FORMAT_A2B10G10R10_UINT_PACK32)
{
Int4 cc;
cc = Insert(cc, Pointer<Int>(buffer)[index[0]], 0);
cc = Insert(cc, Pointer<Int>(buffer)[index[1]], 1);
cc = Insert(cc, Pointer<Int>(buffer)[index[2]], 2);
cc = Insert(cc, Pointer<Int>(buffer)[index[3]], 3);
c.x = Short4(cc & Int4(0x3FF));
c.y = Short4((cc >> 10) & Int4(0x3FF));
c.z = Short4((cc >> 20) & Int4(0x3FF));
c.w = Short4((cc >> 30) & Int4(0x3));
}
else if(state.textureFormat == VK_FORMAT_A2R10G10B10_UINT_PACK32)
{
Int4 cc;
cc = Insert(cc, Pointer<Int>(buffer)[index[0]], 0);
cc = Insert(cc, Pointer<Int>(buffer)[index[1]], 1);
cc = Insert(cc, Pointer<Int>(buffer)[index[2]], 2);
cc = Insert(cc, Pointer<Int>(buffer)[index[3]], 3);
c.z = Short4((cc & Int4(0x3FF)));
c.y = Short4(((cc >> 10) & Int4(0x3FF)));
c.x = Short4(((cc >> 20) & Int4(0x3FF)));
c.w = Short4(((cc >> 30) & Int4(0x3)));
}
else
ASSERT(false);
if(state.textureFormat.isSRGBformat())
{
for(int i = 0; i < textureComponentCount(); i++)
{
if(isRGBComponent(i))
{
// The current table-based sRGB conversion requires 0xFF00 to represent 1.0.
ASSERT(state.textureFormat.has8bitTextureComponents());
sRGBtoLinearFF00(c[i]);
}
}
}
return c;
}
Vector4s SamplerCore::sampleTexel(Short4 &uuuu, Short4 &vvvv, Short4 &wwww, const Short4 &layerIndex, Vector4i &offset, const Int4 &sample, Pointer<Byte> &mipmap, Pointer<Byte> buffer)
{
Vector4s c;
UInt index[4];
computeIndices(index, uuuu, vvvv, wwww, layerIndex, offset, sample, mipmap);
if(isYcbcrFormat())
{
// Generates 15-bit output.
// Pointers to the planes of YCbCr images are stored in consecutive mipmap levels.
Pointer<Byte> bufferY = buffer; // *Pointer<Pointer<Byte>>(mipmap + 0 * sizeof(Mipmap) + OFFSET(Mipmap, buffer));
Pointer<Byte> bufferU = *Pointer<Pointer<Byte>>(mipmap + 1 * sizeof(Mipmap) + OFFSET(Mipmap, buffer)); // U/V for 2-plane interleaved formats.
Pointer<Byte> bufferV = *Pointer<Pointer<Byte>>(mipmap + 2 * sizeof(Mipmap) + OFFSET(Mipmap, buffer));
// Luminance (either 8-bit or 10-bit in bottom bits).
UShort4 Y;
{
switch(state.textureFormat)
{
case VK_FORMAT_G8_B8_R8_3PLANE_420_UNORM:
case VK_FORMAT_G8_B8R8_2PLANE_420_UNORM:
{
Y = Insert(Y, UShort(bufferY[index[0]]), 0);
Y = Insert(Y, UShort(bufferY[index[1]]), 1);
Y = Insert(Y, UShort(bufferY[index[2]]), 2);
Y = Insert(Y, UShort(bufferY[index[3]]), 3);
}
break;
case VK_FORMAT_G10X6_B10X6R10X6_2PLANE_420_UNORM_3PACK16:
{
Y = Insert(Y, Pointer<UShort>(bufferY)[index[0]], 0);
Y = Insert(Y, Pointer<UShort>(bufferY)[index[1]], 1);
Y = Insert(Y, Pointer<UShort>(bufferY)[index[2]], 2);
Y = Insert(Y, Pointer<UShort>(bufferY)[index[3]], 3);
// Top 10 bits of each 16 bits:
Y = (Y & UShort4(0xFFC0u)) >> 6;
}
break;
default:
UNSUPPORTED("state.textureFormat %d", (int)state.textureFormat);
break;
}
}
// Chroma (either 8-bit or 10-bit in bottom bits).
UShort4 Cb, Cr;
{
computeIndices(index, uuuu, vvvv, wwww, layerIndex, offset, sample, mipmap + sizeof(Mipmap));
UShort4 U, V;
switch(state.textureFormat)
{
case VK_FORMAT_G8_B8_R8_3PLANE_420_UNORM:
{
U = Insert(U, UShort(bufferU[index[0]]), 0);
U = Insert(U, UShort(bufferU[index[1]]), 1);
U = Insert(U, UShort(bufferU[index[2]]), 2);
U = Insert(U, UShort(bufferU[index[3]]), 3);
V = Insert(V, UShort(bufferV[index[0]]), 0);
V = Insert(V, UShort(bufferV[index[1]]), 1);
V = Insert(V, UShort(bufferV[index[2]]), 2);
V = Insert(V, UShort(bufferV[index[3]]), 3);
}
break;
case VK_FORMAT_G8_B8R8_2PLANE_420_UNORM:
{
UShort4 UV;
UV = Insert(UV, Pointer<UShort>(bufferU)[index[0]], 0);
UV = Insert(UV, Pointer<UShort>(bufferU)[index[1]], 1);
UV = Insert(UV, Pointer<UShort>(bufferU)[index[2]], 2);
UV = Insert(UV, Pointer<UShort>(bufferU)[index[3]], 3);
U = (UV & UShort4(0x00FFu));
V = (UV & UShort4(0xFF00u)) >> 8;
}
break;
case VK_FORMAT_G10X6_B10X6R10X6_2PLANE_420_UNORM_3PACK16:
{
UInt4 UV;
UV = Insert(UV, Pointer<UInt>(bufferU)[index[0]], 0);
UV = Insert(UV, Pointer<UInt>(bufferU)[index[1]], 1);
UV = Insert(UV, Pointer<UInt>(bufferU)[index[2]], 2);
UV = Insert(UV, Pointer<UInt>(bufferU)[index[3]], 3);
// Top 10 bits of first 16-bits:
U = UShort4((UV & UInt4(0x0000FFC0u)) >> 6);
// Top 10 bits of second 16-bits:
V = UShort4((UV & UInt4(0xFFC00000u)) >> 22);
}
break;
default:
UNSUPPORTED("state.textureFormat %d", (int)state.textureFormat);
break;
}
if(!state.swappedChroma)
{
Cb = U;
Cr = V;
}
else
{
Cb = V;
Cr = U;
}
}
uint8_t lumaBits = 8;
uint8_t chromaBits = 8;
switch(state.textureFormat)
{
case VK_FORMAT_G8_B8_R8_3PLANE_420_UNORM:
case VK_FORMAT_G8_B8R8_2PLANE_420_UNORM:
lumaBits = 8;
chromaBits = 8;
break;
case VK_FORMAT_G10X6_B10X6R10X6_2PLANE_420_UNORM_3PACK16:
lumaBits = 10;
chromaBits = 10;
break;
default:
UNSUPPORTED("state.textureFormat %d", (int)state.textureFormat);
break;
}
if(state.ycbcrModel == VK_SAMPLER_YCBCR_MODEL_CONVERSION_RGB_IDENTITY)
{
// Scale to the output 15-bit.
c.x = Cr << (15 - chromaBits);
c.y = Y << (15 - lumaBits);
c.z = Cb << (15 - chromaBits);
}
else
{
const float twoPowLumaBits = static_cast<float>(0x1u << lumaBits);
const float twoPowLumaBitsMinus8 = static_cast<float>(0x1u << (lumaBits - 8));
const float twoPowChromaBits = static_cast<float>(0x1u << chromaBits);
const float twoPowChromaBitsMinus1 = static_cast<float>(0x1u << (chromaBits - 1));
const float twoPowChromaBitsMinus8 = static_cast<float>(0x1u << (chromaBits - 8));
Float4 y = Float4(Y);
Float4 u = Float4(Cb);
Float4 v = Float4(Cr);
if(state.studioSwing)
{
// See https://www.khronos.org/registry/DataFormat/specs/1.3/dataformat.1.3.html#QUANTIZATION_NARROW
y = ((y / Float4(twoPowLumaBitsMinus8)) - Float4(16.0f)) / Float4(219.0f);
u = ((u / Float4(twoPowChromaBitsMinus8)) - Float4(128.0f)) / Float4(224.0f);
v = ((v / Float4(twoPowChromaBitsMinus8)) - Float4(128.0f)) / Float4(224.0f);
}
else
{
// See https://www.khronos.org/registry/DataFormat/specs/1.3/dataformat.1.3.html#QUANTIZATION_FULL
y = y / Float4(twoPowLumaBits - 1.0f);
u = (u - Float4(twoPowChromaBitsMinus1)) / Float4(twoPowChromaBits - 1.0f);
v = (v - Float4(twoPowChromaBitsMinus1)) / Float4(twoPowChromaBits - 1.0f);
}
// Now, `y` is in [0, 1] and `u` and `v` are in [-0.5, 0.5].
if(state.ycbcrModel == VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_IDENTITY)
{
c.x = Short4(v * static_cast<float>(0x7FFF));
c.y = Short4(y * static_cast<float>(0x7FFF));
c.z = Short4(u * static_cast<float>(0x7FFF));
}
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
float Kb = 0.114f;
float Kr = 0.299f;
switch(state.ycbcrModel)
{
case VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_709:
Kb = 0.0722f;
Kr = 0.2126f;
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 float Kg = 1.0f - Kr - Kb;
const float Rr = 2 * (1 - Kr);
const float Gb = -2 * Kb * (1 - Kb) / Kg;
const float Gr = -2 * Kr * (1 - Kr) / Kg;
const float Bb = 2 * (1 - Kb);
Float4 r = y + Float4(Rr) * v;
Float4 g = y + Float4(Gb) * u + Float4(Gr) * v;
Float4 b = y + Float4(Bb) * u;
c.x = Short4(r * static_cast<float>(0x7FFF));
c.y = Short4(g * static_cast<float>(0x7FFF));
c.z = Short4(b * static_cast<float>(0x7FFF));
}
}
}
else
{
return sampleTexel(index, buffer);
}
return c;
}
Vector4f SamplerCore::sampleTexel(Int4 &uuuu, Int4 &vvvv, Int4 &wwww, const Float4 &dRef, const Int4 &sample, Pointer<Byte> &mipmap, Pointer<Byte> buffer)
{
Int4 valid;
if(borderModeActive())
{
// Valid texels have positive coordinates.
Int4 negative = uuuu;
if(state.is2D() || state.is3D() || state.isCube()) negative |= vvvv;
if(state.is3D() || state.isCube() || state.isArrayed()) negative |= wwww;
valid = CmpNLT(negative, Int4(0));
}
UInt index[4];
computeIndices(index, uuuu, vvvv, wwww, sample, valid, mipmap);
Vector4f c;
if(hasFloatTexture() || has32bitIntegerTextureComponents())
{
UInt4 t0, t1, t2, t3;
switch(state.textureFormat)
{
case VK_FORMAT_R16_SFLOAT:
t0 = Int4(*Pointer<UShort4>(buffer + index[0] * 2));
t1 = Int4(*Pointer<UShort4>(buffer + index[1] * 2));
t2 = Int4(*Pointer<UShort4>(buffer + index[2] * 2));
t3 = Int4(*Pointer<UShort4>(buffer + index[3] * 2));
c.x.x = Extract(As<Float4>(halfToFloatBits(t0)), 0);
c.x.y = Extract(As<Float4>(halfToFloatBits(t1)), 0);
c.x.z = Extract(As<Float4>(halfToFloatBits(t2)), 0);
c.x.w = Extract(As<Float4>(halfToFloatBits(t3)), 0);
break;
case VK_FORMAT_R16G16_SFLOAT:
t0 = Int4(*Pointer<UShort4>(buffer + index[0] * 4));
t1 = Int4(*Pointer<UShort4>(buffer + index[1] * 4));
t2 = Int4(*Pointer<UShort4>(buffer + index[2] * 4));
t3 = Int4(*Pointer<UShort4>(buffer + index[3] * 4));
// TODO: shuffles
c.x = As<Float4>(halfToFloatBits(t0));
c.y = As<Float4>(halfToFloatBits(t1));
c.z = As<Float4>(halfToFloatBits(t2));
c.w = As<Float4>(halfToFloatBits(t3));
transpose4x4(c.x, c.y, c.z, c.w);
break;
case VK_FORMAT_R16G16B16A16_SFLOAT:
t0 = Int4(*Pointer<UShort4>(buffer + index[0] * 8));
t1 = Int4(*Pointer<UShort4>(buffer + index[1] * 8));
t2 = Int4(*Pointer<UShort4>(buffer + index[2] * 8));
t3 = Int4(*Pointer<UShort4>(buffer + index[3] * 8));
c.x = As<Float4>(halfToFloatBits(t0));
c.y = As<Float4>(halfToFloatBits(t1));
c.z = As<Float4>(halfToFloatBits(t2));
c.w = As<Float4>(halfToFloatBits(t3));
transpose4x4(c.x, c.y, c.z, c.w);
break;
case VK_FORMAT_R32_SFLOAT:
case VK_FORMAT_R32_SINT:
case VK_FORMAT_R32_UINT:
case VK_FORMAT_D32_SFLOAT:
// TODO: Optimal shuffling?
c.x.x = *Pointer<Float>(buffer + index[0] * 4);
c.x.y = *Pointer<Float>(buffer + index[1] * 4);
c.x.z = *Pointer<Float>(buffer + index[2] * 4);
c.x.w = *Pointer<Float>(buffer + index[3] * 4);
break;
case VK_FORMAT_R32G32_SFLOAT:
case VK_FORMAT_R32G32_SINT:
case VK_FORMAT_R32G32_UINT:
// TODO: Optimal shuffling?
c.x.xy = *Pointer<Float4>(buffer + index[0] * 8);
c.x.zw = *Pointer<Float4>(buffer + index[1] * 8 - 8);
c.z.xy = *Pointer<Float4>(buffer + index[2] * 8);
c.z.zw = *Pointer<Float4>(buffer + index[3] * 8 - 8);
c.y = c.x;
c.x = Float4(c.x.xz, c.z.xz);
c.y = Float4(c.y.yw, c.z.yw);
break;
case VK_FORMAT_R32G32B32A32_SFLOAT:
case VK_FORMAT_R32G32B32A32_SINT:
case VK_FORMAT_R32G32B32A32_UINT:
c.x = *Pointer<Float4>(buffer + index[0] * 16, 16);
c.y = *Pointer<Float4>(buffer + index[1] * 16, 16);
c.z = *Pointer<Float4>(buffer + index[2] * 16, 16);
c.w = *Pointer<Float4>(buffer + index[3] * 16, 16);
transpose4x4(c.x, c.y, c.z, c.w);
break;
case VK_FORMAT_E5B9G9R9_UFLOAT_PACK32:
{
Float4 t; // TODO: add Insert(UInt4, RValue<UInt>)
t.x = *Pointer<Float>(buffer + index[0] * 4);
t.y = *Pointer<Float>(buffer + index[1] * 4);
t.z = *Pointer<Float>(buffer + index[2] * 4);
t.w = *Pointer<Float>(buffer + index[3] * 4);
t0 = As<UInt4>(t);
c.w = Float4(UInt4(1) << ((t0 >> 27) & UInt4(0x1F))) * Float4(1.0f / (1 << 24));
c.x = Float4(t0 & UInt4(0x1FF)) * c.w;
c.y = Float4((t0 >> 9) & UInt4(0x1FF)) * c.w;
c.z = Float4((t0 >> 18) & UInt4(0x1FF)) * c.w;
}
break;
case VK_FORMAT_B10G11R11_UFLOAT_PACK32:
{
Float4 t; // TODO: add Insert(UInt4, RValue<UInt>)
t.x = *Pointer<Float>(buffer + index[0] * 4);
t.y = *Pointer<Float>(buffer + index[1] * 4);
t.z = *Pointer<Float>(buffer + index[2] * 4);
t.w = *Pointer<Float>(buffer + index[3] * 4);
t0 = As<UInt4>(t);
c.x = As<Float4>(halfToFloatBits((t0 << 4) & UInt4(0x7FF0)));
c.y = As<Float4>(halfToFloatBits((t0 >> 7) & UInt4(0x7FF0)));
c.z = As<Float4>(halfToFloatBits((t0 >> 17) & UInt4(0x7FE0)));
}
break;
default:
UNSUPPORTED("Format %d", VkFormat(state.textureFormat));
}
}
else
{
ASSERT(!isYcbcrFormat());
Vector4s cs = sampleTexel(index, buffer);
bool isInteger = state.textureFormat.isUnnormalizedInteger();
int componentCount = textureComponentCount();
for(int n = 0; n < componentCount; n++)
{