Skip to content

A
angle
Commit 41ee2024 by Minmin Gong Committed by Commit Bot
Options

D3D11: Improvements on ETC1 to BC1 transcode.

ETC1 data doesn't need to be fully decoded before encoding to BC1. Saves ~50% computation. BUG=angleproject:1285 Change-Id: I0cb479a08159ccd0472e5fdf40dd84323151883c Reviewed-on: https://chromium-review.googlesource.com/361553 Commit-Queue: Jamie Madill <jmadill@chromium.org> Reviewed-by: 's avatarJamie Madill <jmadill@chromium.org>
parent 8cf70d55
Showing with 209 additions and 138 deletions
src/image_util/loadimage_etc.cpp
......@@ -46,6 +46,8 @@ static const int intensityModifierNonOpaque[][4] =
};
// clang-format on
static const int kNumPixelsInBlock = 16;
struct ETC2Block
{
// Decodes unsigned single or dual channel block to bytes
......@@ -629,10 +631,13 @@ struct ETC2Block
(static_cast<uint16_t>(rgba.B >> 3) << 0);
}
uint32_t matchBC1Bits(const R8G8B8A8 *rgba,
uint32_t matchBC1Bits(const int *pixelIndices,
const int *pixelIndexCounts,
const R8G8B8A8 *subblockColors,
size_t numColors,
const R8G8B8A8 &minColor,
const R8G8B8A8 &maxColor,
bool opaque) const
bool nonOpaquePunchThroughAlpha) const
{
// Project each pixel on the (maxColor, minColor) line to decide which
// BC1 code to assign to it.
......@@ -653,102 +658,127 @@ struct ETC2Block
decodedColors[i][2] * direction[2];
}
uint32_t bits = 0;
if (opaque)
ASSERT(numColors <= kNumPixelsInBlock);
int encodedColors[kNumPixelsInBlock];
if (nonOpaquePunchThroughAlpha)
{
for (int i = 15; i >= 0; i--)
for (size_t i = 0; i < numColors; i++)
{
// In opaque mode, the code is from 0 to 3.
bits <<= 2;
const int dot =
rgba[i].R * direction[0] + rgba[i].G * direction[1] + rgba[i].B * direction[2];
const int factor = gl::clamp(
static_cast<int>(
(static_cast<float>(dot - stops[1]) / (stops[0] - stops[1])) * 3 + 0.5f),
0, 3);
switch (factor)
const int count = pixelIndexCounts[i];
if (count > 0)
{
case 0:
bits |= 1;
break;
case 1:
bits |= 3;
break;
case 2:
bits |= 2;
break;
case 3:
default:
bits |= 0;
break;
// In non-opaque mode, 3 is for tranparent pixels.
if (0 == subblockColors[i].A)
{
encodedColors[i] = 3;
}
else
{
const R8G8B8A8 &pixel = subblockColors[i];
const int dot = pixel.R * direction[0] + pixel.G * direction[1] +
pixel.B * direction[2];
const int factor = gl::clamp(
static_cast<int>(
(static_cast<float>(dot - stops[1]) / (stops[0] - stops[1])) * 2 +
0.5f),
0, 2);
switch (factor)
{
case 0:
encodedColors[i] = 0;
break;
case 1:
encodedColors[i] = 2;
break;
case 2:
default:
encodedColors[i] = 1;
break;
}
}
}
}
}
else
{
for (int i = 15; i >= 0; i--)
for (size_t i = 0; i < numColors; i++)
{
// In non-opaque mode, 3 is for tranparent pixels.
bits <<= 2;
if (0 == rgba[i].A)
const int count = pixelIndexCounts[i];
if (count > 0)
{
bits |= 3;
}
else
{
const int dot = rgba[i].R * direction[0] + rgba[i].G * direction[1] +
rgba[i].B * direction[2];
// In opaque mode, the code is from 0 to 3.
const R8G8B8A8 &pixel = subblockColors[i];
const int dot =
pixel.R * direction[0] + pixel.G * direction[1] + pixel.B * direction[2];
const int factor = gl::clamp(
static_cast<int>(
(static_cast<float>(dot - stops[1]) / (stops[0] - stops[1])) * 2 +
(static_cast<float>(dot - stops[1]) / (stops[0] - stops[1])) * 3 +
0.5f),
0, 2);
0, 3);
switch (factor)
{
case 0:
bits |= 0;
encodedColors[i] = 1;
break;
case 1:
bits |= 2;
encodedColors[i] = 3;
break;
case 2:
encodedColors[i] = 2;
break;
case 3:
default:
bits |= 1;
encodedColors[i] = 0;
break;
}
}
}
}
uint32_t bits = 0;
for (int i = kNumPixelsInBlock - 1; i >= 0; i--)
{
bits <<= 2;
bits |= encodedColors[pixelIndices[i]];
}
return bits;
}
void packBC1(void *bc1,
const R8G8B8A8 *rgba,
const R8G8B8A8 &minColor,
const R8G8B8A8 &maxColor,
bool opaque) const
const int *pixelIndices,
const int *pixelIndexCounts,
const R8G8B8A8 *subblockColors,
size_t numColors,
int minColorIndex,
int maxColorIndex,
bool nonOpaquePunchThroughAlpha) const
{
const R8G8B8A8 &minColor = subblockColors[minColorIndex];
const R8G8B8A8 &maxColor = subblockColors[maxColorIndex];
uint32_t bits;
uint16_t max16 = RGB8ToRGB565(maxColor);
uint16_t min16 = RGB8ToRGB565(minColor);
if (max16 != min16)
{
// Find the best BC1 code for each pixel
bits = matchBC1Bits(rgba, minColor, maxColor, opaque);
bits = matchBC1Bits(pixelIndices, pixelIndexCounts, subblockColors, numColors, minColor,
maxColor, nonOpaquePunchThroughAlpha);
}
else
{
// Same colors, BC1 index 0 is the color in both opaque and transparent mode
bits = 0;
// BC1 index 3 is transparent
if (!opaque)
if (nonOpaquePunchThroughAlpha)
{
for (int i = 0; i < 16; i++)
for (int i = 0; i < kNumPixelsInBlock; i++)
{
if (0 == rgba[i].A)
if (0 == subblockColors[pixelIndices[i]].A)
{
bits |= (3 << (i * 2));
}
......@@ -761,15 +791,7 @@ struct ETC2Block
std::swap(max16, min16);
uint32_t xorMask = 0;
if (opaque)
{
// In opaque mode switching the two colors is doing the
// following code swaps: 0 <-> 1 and 2 <-> 3. This is
// equivalent to flipping the first bit of each code
// (5 = 0b0101)
xorMask = 0x55555555;
}
else
if (nonOpaquePunchThroughAlpha)
{
// In transparent mode switching the colors is doing the
// following code swap: 0 <-> 1. 0xA selects the second bit of
......@@ -779,6 +801,14 @@ struct ETC2Block
// 0 or 1.
xorMask = ~((bits >> 1) | 0xAAAAAAAA);
}
else
{
// In opaque mode switching the two colors is doing the
// following code swaps: 0 <-> 1 and 2 <-> 3. This is
// equivalent to flipping the first bit of each code
// (5 = 0b0101)
xorMask = 0x55555555;
}
bits ^= xorMask;
}
......@@ -791,15 +821,15 @@ struct ETC2Block
// Encode the opaqueness in the order of the two BC1 colors
BC1Block *dest = reinterpret_cast<BC1Block *>(bc1);
if (opaque)
if (nonOpaquePunchThroughAlpha)
{
dest->color0 = max16;
dest->color1 = min16;
dest->color0 = min16;
dest->color1 = max16;
}
else
{
dest->color0 = min16;
dest->color1 = max16;
dest->color0 = max16;
dest->color1 = min16;
}
dest->bits = bits;
}
......@@ -842,15 +872,14 @@ struct ETC2Block
alphaValues, nonOpaquePunchThroughAlpha);
}
void decodeSubblock(R8G8B8A8 *rgbaBlock,
size_t x,
size_t y,
size_t w,
size_t h,
const uint8_t alphaValues[4][4],
bool flipbit,
size_t subblockIdx,
const R8G8B8A8 subblockColors[2][4]) const
void extractPixelIndices(int *pixelIndices,
int *pixelIndicesCounts,
size_t x,
size_t y,
size_t w,
size_t h,
bool flipbit,
size_t subblockIdx) const
{
size_t dxBegin = 0;
size_t dxEnd = 4;
......@@ -864,53 +893,72 @@ struct ETC2Block
for (size_t j = dyBegin; j < dyEnd && (y + j) < h; j++)
{
R8G8B8A8 *row = &rgbaBlock[j * 4];
int *row = &pixelIndices[j * 4];
for (size_t i = dxBegin; i < dxEnd && (x + i) < w; i++)
{
const size_t pixelIndex = getIndex(i, j);
row[i] = subblockColors[subblockIdx][pixelIndex];
row[i].A = alphaValues[j][i];
const size_t pixelIndex = subblockIdx * 4 + getIndex(i, j);
row[i] = static_cast<int>(pixelIndex);
pixelIndicesCounts[pixelIndex]++;
}
}
}
void selectEndPointPCA(const R8G8B8A8 *pixels, R8G8B8A8 *minColor, R8G8B8A8 *maxColor) const
void selectEndPointPCA(const int *pixelIndexCounts,
const R8G8B8A8 *subblockColors,
size_t numColors,
int *minColorIndex,
int *maxColorIndex) const
{
static const int kNumPixels = 16;
// determine color distribution
int mu[3], min[3], max[3];
for (int ch = 0; ch < 3; ch++)
{
int muv, minv, maxv;
muv = minv = maxv = (&pixels[0].R)[ch];
for (size_t i = 1; i < kNumPixels; i++)
int muv = 0;
int minv = 255;
int maxv = 0;
for (size_t i = 0; i < numColors; i++)
{
muv += (&pixels[i].R)[ch];
minv = std::min<int>(minv, (&pixels[i].R)[ch]);
maxv = std::max<int>(maxv, (&pixels[i].R)[ch]);
const int count = pixelIndexCounts[i];
if (count > 0)
{
const auto &pixel = subblockColors[i];
if (pixel.A > 0)
{
// Non-transparent pixels
muv += (&pixel.R)[ch] * count;
minv = std::min<int>(minv, (&pixel.R)[ch]);
maxv = std::max<int>(maxv, (&pixel.R)[ch]);
}
}
}
mu[ch] = (muv + kNumPixels / 2) / kNumPixels;
mu[ch] = (muv + kNumPixelsInBlock / 2) / kNumPixelsInBlock;
min[ch] = minv;
max[ch] = maxv;
}
// determine covariance matrix
int cov[6] = {0, 0, 0, 0, 0, 0};
for (size_t i = 0; i < kNumPixels; i++)
for (size_t i = 0; i < numColors; i++)
{
int r = pixels[i].R - mu[0];
int g = pixels[i].G - mu[1];
int b = pixels[i].B - mu[2];
cov[0] += r * r;
cov[1] += r * g;
cov[2] += r * b;
cov[3] += g * g;
cov[4] += g * b;
cov[5] += b * b;
const int count = pixelIndexCounts[i];
if (count > 0)
{
const auto &pixel = subblockColors[i];
if (pixel.A > 0)
{
int r = pixel.R - mu[0];
int g = pixel.G - mu[1];
int b = pixel.B - mu[2];
cov[0] += r * r * count;
cov[1] += r * g * count;
cov[2] += r * b * count;
cov[3] += g * g * count;
cov[4] += g * b * count;
cov[5] += b * b * count;
}
}
}
// Power iteration algorithm to get the eigenvalues and eigenvector
......@@ -966,27 +1014,35 @@ struct ETC2Block
}
// Pick colors at extreme points
int minD = pixels[0].R * vr + pixels[0].G * vg + pixels[0].B * vb;
int maxD = minD;
int minD = INT_MAX;
int maxD = 0;
size_t minIndex = 0;
size_t maxIndex = 0;
for (size_t i = 1; i < kNumPixels; i++)
for (size_t i = 0; i < numColors; i++)
{
int dot = pixels[i].R * vr + pixels[i].G * vg + pixels[i].B * vb;
if (dot < minD)
{
minD = dot;
minIndex = i;
}
if (dot > maxD)
const int count = pixelIndexCounts[i];
if (count > 0)
{
maxD = dot;
maxIndex = i;
const auto &pixel = subblockColors[i];
if (pixel.A > 0)
{
int dot = pixel.R * vr + pixel.G * vg + pixel.B * vb;
if (dot < minD)
{
minD = dot;
minIndex = i;
}
if (dot > maxD)
{
maxD = dot;
maxIndex = i;
}
}
}
}
*minColor = pixels[minIndex];
*maxColor = pixels[maxIndex];
*minColorIndex = static_cast<int>(minIndex);
*maxColorIndex = static_cast<int>(maxIndex);
}
void transcodeIndividualOrDifferentialBlockToBC1(uint8_t *dest,
......@@ -1009,45 +1065,60 @@ struct ETC2Block
// select axis by principal component analysis (PCA) to use as
// our two BC1 endpoints and then map pixels to BC1 by projecting on the
// line between the two endpoints and choosing the right fraction.
//
// In the future, we have a potential improvements to this algorithm.
// 1. We don't actually need to decode ETC blocks to RGBs. Instead,
// the subblock colors and pixel indices alreay contains enough
// information for transcode. A direct mapping would be more
// efficient here.
// The goal of this algorithm is make it faster than decode ETC to RGBs
// and then encode to BC. To achieve this, we only extract subblock
// colors, pixel indices, and counts of each pixel indices from ETC.
// With those information, we can only encode used subblock colors
// to BC1, and copy the bits to the right pixels.
// Fully decode and encode need to process 16 RGBA pixels. With this
// algorithm, it's 8 pixels at maximum for a individual or
// differential block. Saves us bandwidth and computations.
static const size_t kNumColors = 8;
const auto intensityModifier =
nonOpaquePunchThroughAlpha ? intensityModifierNonOpaque : intensityModifierDefault;
// Compute the colors that pixels can have in each subblock both for
// the decoding of the RGBA data and BC1 encoding
R8G8B8A8 subblockColors[2][4];
R8G8B8A8 subblockColors[kNumColors];
for (size_t modifierIdx = 0; modifierIdx < 4; modifierIdx++)
{
const int i1 = intensityModifier[u.idht.mode.idm.cw1][modifierIdx];
subblockColors[0][modifierIdx] = createRGBA(r1 + i1, g1 + i1, b1 + i1);
if (nonOpaquePunchThroughAlpha && (modifierIdx == 2))
{
// In ETC opaque punch through formats, individual and
// differential blocks take index 2 as transparent pixel.
// Thus we don't need to compute its color, just assign it
// as black.
subblockColors[modifierIdx] = createRGBA(0, 0, 0, 0);
subblockColors[4 + modifierIdx] = createRGBA(0, 0, 0, 0);
}
else
{
const int i1 = intensityModifier[u.idht.mode.idm.cw1][modifierIdx];
subblockColors[modifierIdx] = createRGBA(r1 + i1, g1 + i1, b1 + i1);
const int i2 = intensityModifier[u.idht.mode.idm.cw2][modifierIdx];
subblockColors[1][modifierIdx] = createRGBA(r2 + i2, g2 + i2, b2 + i2);
const int i2 = intensityModifier[u.idht.mode.idm.cw2][modifierIdx];
subblockColors[4 + modifierIdx] = createRGBA(r2 + i2, g2 + i2, b2 + i2);
}
}
R8G8B8A8 rgbaBlock[16];
// Decode the block in rgbaBlock.
int pixelIndices[kNumPixelsInBlock];
int pixelIndexCounts[kNumColors] = {0};
// Extract pixel indices from a ETC block.
for (size_t blockIdx = 0; blockIdx < 2; blockIdx++)
{
decodeSubblock(rgbaBlock, x, y, w, h, alphaValues, u.idht.mode.idm.flipbit, blockIdx,
subblockColors);
}
if (nonOpaquePunchThroughAlpha)
{
decodePunchThroughAlphaBlock(reinterpret_cast<uint8_t *>(rgbaBlock), x, y, w, h,
sizeof(R8G8B8A8) * 4);
extractPixelIndices(pixelIndices, pixelIndexCounts, x, y, w, h, u.idht.mode.idm.flipbit,
blockIdx);
}
R8G8B8A8 minColor, maxColor;
selectEndPointPCA(rgbaBlock, &minColor, &maxColor);
int minColorIndex, maxColorIndex;
selectEndPointPCA(pixelIndexCounts, subblockColors, kNumColors, &minColorIndex,
&maxColorIndex);
packBC1(dest, rgbaBlock, minColor, maxColor, !nonOpaquePunchThroughAlpha);
packBC1(dest, pixelIndices, pixelIndexCounts, subblockColors, kNumColors, minColorIndex,
maxColorIndex, nonOpaquePunchThroughAlpha);
}
void transcodeTBlockToBC1(uint8_t *dest,
......
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