Scarfski/filament
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Add async interface to Ktx2Reader.

Browse Source
This commit is contained in:
Philip Rideout
2022-04-28 15:08:26 -07:00
parent 92bc0752d1
commit d76551ab28
12 changed files with 389 additions and 111 deletions
Download Patch File Download Diff File Expand all files Collapse all files

325
libs/ktxreader/src/Ktx2Reader.cpp
View File

@@ -21,6 +21,8 @@
#include <utils/Log.h>
#include <vector>
#pragma clang diagnostic push
#pragma clang diagnostic ignored "-Warray-bounds"
#include <basisu_transcoder.h>
@@ -30,9 +32,13 @@ using namespace basist;
using namespace filament;
using TransferFunction = ktxreader::Ktx2Reader::TransferFunction;
using Result = ktxreader::Ktx2Reader::Result;
using Async = ktxreader::Ktx2Reader::Async;
using Buffer = std::vector<uint8_t>;
namespace {
struct FinalFormatInfo {
const char* name; // <-- for debug purposes only
bool isSupported;
bool isCompressed;
TransferFunction transferFunction;
@@ -73,35 +79,118 @@ static FinalFormatInfo getFinalFormatInfo(Texture::InternalFormat fmt) {
using tt = Texture::Type;
using tf = Texture::Format;
using ttf = transcoder_texture_format;
const auto sRGB = ktxreader::Ktx2Reader::TransferFunction::sRGB;
const auto LINEAR = ktxreader::Ktx2Reader::TransferFunction::LINEAR;
const auto sRGB = TransferFunction::sRGB;
const auto LINEAR = TransferFunction::LINEAR;
switch (fmt) {
case tif::ETC2_EAC_SRGBA8: return {true, true, sRGB, ttf::cTFETC2_RGBA, tct::ETC2_EAC_RGBA8};
case tif::ETC2_EAC_RGBA8: return {true, true, LINEAR, ttf::cTFETC2_RGBA, tct::ETC2_EAC_SRGBA8};
case tif::DXT1_SRGB: return {true, true, sRGB, ttf::cTFBC1_RGB, tct::DXT1_RGB};
case tif::DXT1_RGB: return {true, true, LINEAR, ttf::cTFBC1_RGB, tct::DXT1_SRGB};
case tif::DXT3_SRGBA: return {true, true, sRGB, ttf::cTFBC3_RGBA, tct::DXT3_RGBA};
case tif::DXT3_RGBA: return {true, true, LINEAR, ttf::cTFBC3_RGBA, tct::DXT3_SRGBA};
case tif::SRGB8_ALPHA8_ASTC_4x4: return {true, true, sRGB, ttf::cTFASTC_4x4_RGBA, tct::RGBA_ASTC_4x4};
case tif::RGBA_ASTC_4x4: return {true, true, LINEAR, ttf::cTFASTC_4x4_RGBA, tct::SRGB8_ALPHA8_ASTC_4x4};
case tif::EAC_R11: return {true, true, LINEAR, ttf::cTFETC2_EAC_R11, tct::EAC_R11};
case tif::ETC2_EAC_SRGBA8: return {"ETC2_EAC_SRGBA8", true, true, sRGB, ttf::cTFETC2_RGBA, tct::ETC2_EAC_RGBA8};
case tif::ETC2_EAC_RGBA8: return {"ETC2_EAC_RGBA8", true, true, LINEAR, ttf::cTFETC2_RGBA, tct::ETC2_EAC_SRGBA8};
case tif::DXT1_SRGB: return {"DXT1_SRGB", true, true, sRGB, ttf::cTFBC1_RGB, tct::DXT1_RGB};
case tif::DXT1_RGB: return {"DXT1_RGB", true, true, LINEAR, ttf::cTFBC1_RGB, tct::DXT1_SRGB};
case tif::DXT3_SRGBA: return {"DXT3_SRGBA", true, true, sRGB, ttf::cTFBC3_RGBA, tct::DXT3_RGBA};
case tif::DXT3_RGBA: return {"DXT3_RGBA", true, true, LINEAR, ttf::cTFBC3_RGBA, tct::DXT3_SRGBA};
case tif::SRGB8_ALPHA8_ASTC_4x4: return {"SRGB8_ALPHA8_ASTC_4x4", true, true, sRGB, ttf::cTFASTC_4x4_RGBA, tct::RGBA_ASTC_4x4};
case tif::RGBA_ASTC_4x4: return {"RGBA_ASTC_4x4", true, true, LINEAR, ttf::cTFASTC_4x4_RGBA, tct::SRGB8_ALPHA8_ASTC_4x4};
case tif::EAC_R11: return {"EAC_R11", true, true, LINEAR, ttf::cTFETC2_EAC_R11, tct::EAC_R11};
// The following format is useful for normal maps.
// Note that BasisU supports only the unsigned variant.
case tif::EAC_RG11: return {true, true, LINEAR, ttf::cTFETC2_EAC_RG11, tct::EAC_RG11};
case tif::EAC_RG11: return {"EAC_RG11", true, true, LINEAR, ttf::cTFETC2_EAC_RG11, tct::EAC_RG11};
// Uncompressed formats.
case tif::SRGB8_A8: return {true, false, sRGB, ttf::cTFRGBA32, {}, tt::UBYTE, tf::RGBA};
case tif::RGBA8: return {true, false, LINEAR, ttf::cTFRGBA32, {}, tt::UBYTE, tf::RGBA};
case tif::RGB565: return {true, false, LINEAR, ttf::cTFRGB565, {}, tt::USHORT_565, tf::RGB};
case tif::RGBA4: return {true, false, LINEAR, ttf::cTFRGBA4444, {}, tt::USHORT, tf::RGBA};
case tif::SRGB8_A8: return {"SRGB8_A8", true, false, sRGB, ttf::cTFRGBA32, {}, tt::UBYTE, tf::RGBA};
case tif::RGBA8: return {"RGBA8", true, false, LINEAR, ttf::cTFRGBA32, {}, tt::UBYTE, tf::RGBA};
case tif::RGB565: return {"RGB565", true, false, LINEAR, ttf::cTFRGB565, {}, tt::USHORT_565, tf::RGB};
case tif::RGBA4: return {"RGBA4", true, false, LINEAR, ttf::cTFRGBA4444, {}, tt::USHORT, tf::RGBA};
default: return {false};
default: return {};
}
}
// In theory we could pass "free" directly into the callback but doing so triggers ASAN warnings.
static void freeCallback(void* buf, size_t, void* userdata) {
free(buf);
}
// This helper is used by both the asynchronous and synchronous API's.
static Result transcodeImageLevel(ktx2_transcoder& transcoder,
ktx2_transcoder_state& transcoderState, Texture::InternalFormat format,
uint32_t levelIndex, Texture::PixelBufferDescriptor** pbd) {
using basisu::texture_format;
assert_invariant(levelIndex < KTX2_MAX_SUPPORTED_LEVEL_COUNT);
const FinalFormatInfo formatInfo = getFinalFormatInfo(format);
const texture_format destFormat = basis_get_basisu_texture_format(formatInfo.basisFormat);
const uint32_t layerIndex = 0;
const uint32_t faceIndex = 0;
const uint32_t decodeFlags = 0;
const uint32_t outputRowPitch = 0;
const uint32_t outputRowCount = 0;
const int channel0 = 0;
const int channel1 = 0;
basist::ktx2_image_level_info levelInfo;
transcoder.get_image_level_info(levelInfo, levelIndex, layerIndex, faceIndex);
if (formatInfo.isCompressed) {
const uint32_t qwordsPerBlock = basisu::get_qwords_per_block(destFormat);
const size_t byteCount = sizeof(uint64_t) * qwordsPerBlock * levelInfo.m_total_blocks;
uint64_t* const blocks = (uint64_t*) malloc(byteCount);
if (!transcoder.transcode_image_level(levelIndex, layerIndex, faceIndex, blocks,
levelInfo.m_total_blocks, formatInfo.basisFormat, decodeFlags,
outputRowPitch, outputRowCount, channel0,
channel1, &transcoderState)) {
return Result::COMPRESSED_TRANSCODE_FAILURE;
}
*pbd = new Texture::PixelBufferDescriptor(blocks,
byteCount, formatInfo.compressedPixelDataType, byteCount, freeCallback);
return Result::SUCCESS;
}
const uint32_t rowCount = levelInfo.m_orig_height;
const uint32_t bytesPerPix = basis_get_bytes_per_block_or_pixel(formatInfo.basisFormat);
const size_t byteCount = bytesPerPix * levelInfo.m_orig_width * rowCount;
uint64_t* const rows = (uint64_t*) malloc(byteCount);
if (!transcoder.transcode_image_level(levelIndex, layerIndex, faceIndex, rows,
byteCount / bytesPerPix, formatInfo.basisFormat, decodeFlags,
outputRowPitch, outputRowCount, channel0, channel1, &transcoderState)) {
return Result::UNCOMPRESSED_TRANSCODE_FAILURE;
}
*pbd = new Texture::PixelBufferDescriptor(rows, byteCount,
formatInfo.pixelDataFormat, formatInfo.pixelDataType, freeCallback);
return Result::SUCCESS;
}
namespace ktxreader {
class FAsync : public Async {
public:
FAsync(Texture* texture, Engine& engine, ktx2_transcoder* transcoder, Buffer&& buf) :
mTexture(texture), mEngine(engine), mTranscoder(transcoder),
mSourceBuffer(std::move(buf)) {}
Texture* getTexture() const noexcept { return mTexture; }
Result doTranscoding();
void uploadImages();
private:
using TranscoderResult = std::atomic<Texture::PixelBufferDescriptor*>;
// After each level is transcoded, the results are stashed in the following array until the
// foreground thread calls uploadImages(). Each slot in the array corresponds to a single
// miplevel in the texture.
TranscoderResult mTranscoderResults[KTX2_MAX_SUPPORTED_LEVEL_COUNT] = {};
Texture* const mTexture;
Engine& mEngine;
// We do not share the BasisU trancoder between Async objects. The BasisU transcoder
// allows parallelization at "level" granularity, but does not permit parallelization at
// "texture" granularity. i.e. the transcode_image_level() method is thread-safe but the
// start_transcoding() method is not.
std::unique_ptr<ktx2_transcoder> const mTranscoder;
// Storage for the content of the KTX2 file.
Buffer mSourceBuffer;
};
Ktx2Reader::Ktx2Reader(Engine& engine, bool quiet) :
mEngine(engine),
mQuiet(quiet),
@@ -114,20 +203,20 @@ Ktx2Reader::~Ktx2Reader() {
delete mTranscoder;
}
bool Ktx2Reader::requestFormat(Texture::InternalFormat format) {
Result Ktx2Reader::requestFormat(Texture::InternalFormat format) noexcept {
if (!getFinalFormatInfo(format).isSupported) {
return false;
return Result::FORMAT_UNSUPPORTED;
}
for (Texture::InternalFormat fmt : mRequestedFormats) {
if (fmt == format) {
return false;
return Result::FORMAT_ALREADY_REQUESTED;
}
}
mRequestedFormats.push_back(format);
return true;
return Result::SUCCESS;
}
void Ktx2Reader::unrequestFormat(Texture::InternalFormat format) {
void Ktx2Reader::unrequestFormat(Texture::InternalFormat format) noexcept {
for (auto iter = mRequestedFormats.begin(); iter != mRequestedFormats.end(); ++iter) {
if (*iter == format) {
mRequestedFormats.erase(iter);
@@ -136,15 +225,102 @@ void Ktx2Reader::unrequestFormat(Texture::InternalFormat format) {
}
}
Texture* Ktx2Reader::load(const uint8_t* data, size_t size, TransferFunction transfer) {
if (!mTranscoder->init(data, size)) {
Texture* Ktx2Reader::load(const void* data, size_t size, TransferFunction transfer) {
Texture* texture = createTexture(mTranscoder, data, size, transfer);
if (texture == nullptr) {
return nullptr;
}
if (!mTranscoder->start_transcoding()) {
mEngine.destroy(texture);
if (!mQuiet) {
utils::slog.e << "BasisU start_transcoding failed." << utils::io::endl;
}
return nullptr;
}
ktx2_transcoder_state basisThreadState;
basisThreadState.clear();
for (uint32_t levelIndex = 0, n = mTranscoder->get_levels(); levelIndex < n; levelIndex++) {
Texture::PixelBufferDescriptor* pbd;
Result result = transcodeImageLevel(*mTranscoder, basisThreadState, texture->getFormat(),
levelIndex, &pbd);
if (UTILS_UNLIKELY(result != Result::SUCCESS)) {
mEngine.destroy(texture);
if (!mQuiet) {
utils::slog.e << "Failed to transcode level " << levelIndex << utils::io::endl;
}
return nullptr;
}
texture->setImage(mEngine, levelIndex, std::move(*pbd));
}
return texture;
}
Result FAsync::doTranscoding() {
ktx2_transcoder_state basisThreadState;
basisThreadState.clear();
for (uint32_t levelIndex = 0, n = mTranscoder->get_levels(); levelIndex < n; levelIndex++) {
Texture::PixelBufferDescriptor* pbd;
Result result = transcodeImageLevel(*mTranscoder, basisThreadState, mTexture->getFormat(),
levelIndex, &pbd);
if (UTILS_UNLIKELY(result != Result::SUCCESS)) {
return result;
}
mTranscoderResults[levelIndex].store(pbd);
}
return Result::SUCCESS;
}
void FAsync::uploadImages() {
size_t levelIndex = 0;
UTILS_NOUNROLL
for (TranscoderResult& level : mTranscoderResults) {
Texture::PixelBufferDescriptor* pbd = level.load();
if (pbd) {
level.store(nullptr);
mTexture->setImage(mEngine, levelIndex, std::move(*pbd));
delete pbd;
}
++levelIndex;
}
}
Async* Ktx2Reader::asyncCreate(const void* data, size_t size, TransferFunction transfer) {
Buffer ktx2content((uint8_t*)data, (uint8_t*)data + size);
ktx2_transcoder* transcoder = new ktx2_transcoder();
Texture* texture = createTexture(transcoder, ktx2content.data(), ktx2content.size(), transfer);
if (texture == nullptr) {
delete transcoder;
return nullptr;
}
if (!transcoder->start_transcoding()) {
delete transcoder;
mEngine.destroy(texture);
return nullptr;
}
// There's no need to do any further work at this point but it should be noted that this is the
// point at which we first come to know the number of miplevels, dimensions, etc. If we had a
// dynamically sized array to store decoder results, we would reserve it here.
return new FAsync(texture, mEngine, transcoder, std::move(ktx2content));
}
void Ktx2Reader::asyncDestroy(Async** async) {
delete *async;
*async = nullptr;
}
Texture* Ktx2Reader::createTexture(ktx2_transcoder* transcoder, const void* data, size_t size,
TransferFunction transfer) {
if (!transcoder->init(data, size)) {
if (!mQuiet) {
utils::slog.e << "BasisU transcoder init failed." << utils::io::endl;
}
return nullptr;
}
if (mTranscoder->get_dfd_transfer_func() == KTX2_KHR_DF_TRANSFER_LINEAR &&
if (transcoder->get_dfd_transfer_func() == KTX2_KHR_DF_TRANSFER_LINEAR &&
transfer == TransferFunction::sRGB) {
if (!mQuiet) {
utils::slog.e << "Source texture is marked linear, but client is requesting sRGB."
@@ -153,7 +329,7 @@ Texture* Ktx2Reader::load(const uint8_t* data, size_t size, TransferFunction tra
return nullptr;
}
if (mTranscoder->get_dfd_transfer_func() == KTX2_KHR_DF_TRANSFER_SRGB &&
if (transcoder->get_dfd_transfer_func() == KTX2_KHR_DF_TRANSFER_SRGB &&
transfer == TransferFunction::LINEAR) {
if (!mQuiet) {
utils::slog.e << "Source texture is marked sRGB, but client is requesting linear."
@@ -163,7 +339,7 @@ Texture* Ktx2Reader::load(const uint8_t* data, size_t size, TransferFunction tra
}
// TODO: support cubemaps. For now we use KTX1 for cubemaps because basisu does not support HDR.
if (mTranscoder->get_faces() == 6) {
if (transcoder->get_faces() == 6) {
if (!mQuiet) {
utils::slog.e << "Cubemaps are not yet supported." << utils::io::endl;
}
@@ -171,32 +347,33 @@ Texture* Ktx2Reader::load(const uint8_t* data, size_t size, TransferFunction tra
}
// TODO: support texture arrays.
if (mTranscoder->get_layers() > 1) {
if (transcoder->get_layers() > 1) {
if (!mQuiet) {
utils::slog.e << "Texture arrays are not yet supported." << utils::io::endl;
}
return nullptr;
}
// Fierst pass through, just to make sure we can transcode it.
// First pass through, just to make sure we can transcode it.
bool found = false;
Texture::InternalFormat resolvedFormat;
FinalFormatInfo info;
for (Texture::InternalFormat requestedFormat : mRequestedFormats) {
if (!Texture::isTextureFormatSupported(mEngine, requestedFormat)) {
continue;
}
const auto info = getFinalFormatInfo(requestedFormat);
info = getFinalFormatInfo(requestedFormat);
if (!info.isSupported || info.transferFunction != transfer) {
continue;
}
if (!basis_is_format_supported(info.basisFormat, mTranscoder->get_format())) {
if (!basis_is_format_supported(info.basisFormat, transcoder->get_format())) {
continue;
}
const uint32_t layerIndex = 0;
const uint32_t faceIndex = 0;
for (uint32_t levelIndex = 0; levelIndex < mTranscoder->get_levels(); levelIndex++) {
for (uint32_t levelIndex = 0; levelIndex < transcoder->get_levels(); levelIndex++) {
basist::ktx2_image_level_info info;
if (!mTranscoder->get_image_level_info(info, levelIndex, layerIndex, faceIndex)) {
if (!transcoder->get_image_level_info(info, levelIndex, layerIndex, faceIndex)) {
continue;
}
}
@@ -212,75 +389,37 @@ Texture* Ktx2Reader::load(const uint8_t* data, size_t size, TransferFunction tra
return nullptr;
}
const auto formatInfo = getFinalFormatInfo(resolvedFormat);
Texture* texture = Texture::Builder()
.width(mTranscoder->get_width())
.height(mTranscoder->get_height())
.levels(mTranscoder->get_levels())
.width(transcoder->get_width())
.height(transcoder->get_height())
.levels(transcoder->get_levels())
.sampler(Texture::Sampler::SAMPLER_2D)
.format(resolvedFormat)
.build(mEngine);
if (texture == nullptr) {
if (!mQuiet) {
utils::slog.e << "Unable to construct texture using BasisU info." << utils::io::endl;
}
return nullptr;
if (texture == nullptr && !mQuiet) {
utils::slog.e << "Unable to construct texture using BasisU info." << utils::io::endl;
}
if (!mTranscoder->start_transcoding()) {
mEngine.destroy(texture);
if (!mQuiet) {
utils::slog.e << "BasisU start_transcoding failed." << utils::io::endl;
}
return nullptr;
}
// In theory we could pass "free" directly into the callback but that triggers ASAN warnings.
Texture::PixelBufferDescriptor::Callback cb = [](void* buf, size_t, void* userdata) {
free(buf);
};
const uint32_t layerIndex = 0;
const uint32_t faceIndex = 0;
for (uint32_t levelIndex = 0; levelIndex < mTranscoder->get_levels(); levelIndex++) {
basist::ktx2_image_level_info levelInfo;
mTranscoder->get_image_level_info(levelInfo, levelIndex, layerIndex, faceIndex);
const basisu::texture_format destFormat =
basis_get_basisu_texture_format(formatInfo.basisFormat);
if (formatInfo.isCompressed) {
const uint32_t qwordsPerBlock = basisu::get_qwords_per_block(destFormat);
const size_t byteCount = sizeof(uint64_t) * qwordsPerBlock * levelInfo.m_total_blocks;
uint64_t* const blocks = (uint64_t*) malloc(byteCount);
const uint32_t flags = 0;
if (!mTranscoder->transcode_image_level(levelIndex, layerIndex, faceIndex, blocks,
levelInfo.m_total_blocks, formatInfo.basisFormat, flags)) {
utils::slog.e << "Failed to transcode level " << levelIndex << utils::io::endl;
return nullptr;
}
Texture::PixelBufferDescriptor pbd(blocks, byteCount,
formatInfo.compressedPixelDataType, byteCount, cb, nullptr);
texture->setImage(mEngine, levelIndex, std::move(pbd));
} else {
// The transcoder still does work even for uncompressed formats, because of zstd.
const uint32_t rowCount = levelInfo.m_orig_height;
const uint32_t bytesPerPix = basis_get_bytes_per_block_or_pixel(formatInfo.basisFormat);
const size_t byteCount = bytesPerPix * levelInfo.m_orig_width * rowCount;
uint64_t* const rows = (uint64_t*) malloc(byteCount);
const uint32_t flags = 0;
if (!mTranscoder->transcode_image_level(levelIndex, layerIndex, faceIndex, rows,
byteCount / bytesPerPix, formatInfo.basisFormat, flags)) {
utils::slog.e << "Failed to transcode level " << levelIndex << utils::io::endl;
return nullptr;
}
Texture::PixelBufferDescriptor pbd(rows, byteCount, formatInfo.pixelDataFormat,
formatInfo.pixelDataType, cb, nullptr);
texture->setImage(mEngine, levelIndex, std::move(pbd));
}
}
#if BASISU_FORCE_DEVEL_MESSAGES
utils::slog.e << "Ktx2Reader created "
<< transcoder->get_width() << "x" << transcoder->get_height() << " texture with format "
<< info.name << utils::io::endl;
#endif
return texture;
}
Texture* Async::getTexture() const noexcept {
return static_cast<FAsync const*>(this)->getTexture();
}
Result Async::doTranscoding() {
return static_cast<FAsync*>(this)->doTranscoding();
}
void Async::uploadImages() {
return static_cast<FAsync*>(this)->uploadImages();
}
} // namespace ktxreader