Compressonator + bc7enc_rdo (BC7 CPU codec swap)

A fork of AMD Compressonator that replaces the CPU-side BC7 codec with the bc7e.ispc encoder from richgel999/bc7enc_rdo.

This README section describes the fork. Everything below the horizontal rule is AMD's original README, unmodified. The full investigation — every measurement, dead end, and correction — lives in NOTES.md.

1. What this fork is

Compressonator's DDS I/O, mip generation, batch processing and CLI are excellent and kept entirely intact. What changed is one thing: the BC7 block encoder underneath them.

The motivation is a documented endpoint p-bit rounding issue in DirectXTex/texconv — its endpoint quantization doesn't properly compensate rounding for the chosen parity bit. Microsoft partially fixed the correctness bug in 2018 but never adopted the fully optimal "search all p-bit combinations with compensated rounding" approach. bc7e.ispc implements the correct approach and vectorizes it through ISPC, which also sidesteps the second problem: there is no cross-vendor, cross-platform GPU-accelerated BC7 path that behaves identically on AMD, NVIDIA and machines with no usable GPU at all. A CPU SIMD encoder doesn't need one.

The p-bit issue is documented by Richard Geldreich in two April 2018 posts:

This is built on two existing projects, not written from scratch:

AMD Compressonator — the surrounding tool, all container/format handling, the CLI. MIT.
bc7enc_rdo / bc7e.ispc by Richard Geldreich — the BC7 encoder doing the actual block compression. Apache 2.0.

2. Usage

The codec is selected at build time, not at runtime. There is no CLI flag to switch encoders — the selection is a compile-time #ifdef (CMP_USE_BC7ENC_RDO / CMP_USE_BC7ENC_RDO_BATCH, see cmp_compressonatorlib/bc7/codec_bc7.cpp). Build the flavor you want:

CMake options	Encoder used
(neither set)	Stock Compressonator BC7 (unchanged upstream behavior)
`-DOPTION_CMP_USE_BC7ENC_RDO=ON`	bc7e.ispc, one block per call
`-DOPTION_CMP_USE_BC7ENC_RDO=ON -DOPTION_CMP_USE_BC7ENC_RDO_BATCH=ON`	bc7e.ispc, SIMD-batched — the recommended build

Once built, invocation is ordinary Compressonator:

compressonatorcli -fd BC7 -EncodeWith CPU -Quality 1.0 in.png out.dds

Existing flags still work — and were verified, not assumed

-Quality, -ModeMask, -ColourRestrict and -AlphaRestrict all map onto bc7e's parameters and were checked against the stock codec's behavior. This mattered because the first implementation got it wrong:

The Q→preset bucketing was verified by MD5-comparing encodes at interior quality values against the four bucket boundaries (Q = 0.25 / 0.45 / 0.65 / 0.85). 2 of 4 boundaries initially misresolved — Q=0.45 landed in fast instead of basic, Q=0.65 in basic instead of slow. See Phase 3 part 4b in NOTES.md.
Root cause turned out to be two float-precision truncations in series on the CLI path — std::stof at cmdline.cpp:411 and a (CODECFLOAT) cast at compress.cpp:223 (CODECFLOAT is typedef float). Phase 3 part 4b first blamed codec_bc7.cpp:131; that was traced and corrected in Phase 3 part 4c — that line sits on a dead string-overload path the CLI never reaches.
Post-fix the boundary sweep is 9/9 exact and the per-block and batched paths are bit-identical across 6 configurations, 100.0000% block match (14 976 or 30 000 blocks per config), including the -ColourRestrict / -AlphaRestrict / -ModeMask 255 combinations.

3. What's different from stock, and why

Omitted from this build

Feature	Why
Image analysis (`-Analysis`, SSIM)	Requires OpenCV, which drags shared-library dependencies back in. Not on the BC7 encode path.
KTX2	Requires Git LFS assets and a working host Python at configure time. Not needed for DDS output.
Brotli-G	Same — its configure step shells out to Python.
GPU / DirectCompute encoding	This fork's whole premise is a CPU encoder that behaves identically everywhere.
ETC / ETC2	License scope — see below.

On ETC/ETC2: Compressonator bundles ETCPack, which ships under an Ericsson SLA whose scope restricts use to Khronos-standard texture compression development. A general-purpose BC7 CLI that also happens to contain ETC codecs sits in ambiguous territory relative to that scope. Rather than interpret the SLA, this build removes the code: ETCPack is gated out at compile time behind OPTION_CMP_ETC=OFF. Verified by symbol and string audit on the built binaries — CCodec_ETC symbols 272 → 0, compressBlockETC 47 → 0, Ericsson SLA strings 4 → 0. Requesting an ETC format now behaves exactly like requesting an unknown one (exit 255, no file written), rather than crashing or emitting garbage.

(Full honesty, per the audit: 3 ETC2 symbols and 4 ETC format-name strings do survive. They are string literals from the format enum↔name lookup table at applications/_plugins/common/atiformats.cpp:111-114 — format identifiers only, no Ericsson-authored code, no SLA text.)

What you get in exchange

Fully static, zero-runtime-dependency binaries on both platforms:

Platform	Result
Linux	0 dynamic dependencies. `ldd` reports "not a dynamic executable". No libstdc++ version requirement, no OpenCV, no GL.
Windows	3 DLLs, all shipped with the OS: `KERNEL32.dll`, `ole32.dll`, `imagehlp.dll`. No Visual C++ redistributable, no UCRT — `MSVCP140`, `VCRUNTIME140`, `VCRUNTIME140_1` and all ten `api-ms-win-crt-*` entries are gone (16 → 3).

Install-nothing on any modern glibc or Windows target. That is the point: this binary is meant to be shipped inside another tool and shelled out to.

If you need the omitted features, run stock Compressonator alongside this build. They coexist fine — this fork is a specialized BC7 encoder, not a replacement for the full toolchain.

4. Building

Two canonical, tested build recipes. Prefer them over assembling flags by hand — they encode a long tail of platform-specific fixes.

Platform	Script
Linux	`tools/linux/build_flavor.sh [off\|unbatched\|batch]`
Windows	`tools/win/build_cli_batch.ps1 -Flavor <off\|unbatched\|batch>`

Both live in this repository, and both expect to be run from a checkout where compressonator/ and bc7enc_rdo/ sit side by side — they resolve the outer root by walking up from their own location.

The flavor parameter selects the encoder and the build directory:

Flavor	Encoder	Build dir
`off`	Stock Compressonator BC7	`build_cli_off`
`unbatched`	bc7e.ispc, per-block	`build_cli`
`batch`	bc7e.ispc, SIMD-batched	`build_cli_batch`

batch is the default and the recommended build.

Prerequisites

bc7enc_rdo checkout next to this repository — the build validates that bc7e.ispc is present and fails early if not.
ISPC v1.31.0, pinned. Linux expects it unpacked at tools/ispc/linux/bin/ispc; the Windows script expects the equivalent Windows package. The version is pinned deliberately — bc7e.ispc is compiled by it.
A C++ toolchain. Verified on GCC (Linux) and MSVC 17.14 / Visual Studio 2022 Build Tools with Windows SDK 10.0.26100 (Windows).

On Windows, build_cli_batch.ps1 deliberately does not bootstrap the Visual Studio environment; tools/win/build_flavor.bat wraps it with VsDevCmd.bat if you want a one-shot invocation.

5. Results

Headline benchmark: 9-file corpus of real game textures (hard-alpha UI cutouts plus a 2048×2048 normal map), best-of-3 wall clock, alpha-aware PSNR with the decoder held constant across every variant.

Variant	Corpus wall clock	vs. batched bc7e
texconv (DirectXTex CPU BC7, `-bc x`)	175.350 s	131× slower
Stock Compressonator `-Quality 1.0`	16.392 s	12.3× slower
Stock Compressonator, default quality	2.213 s	1.66× slower
Batched bc7e `-Quality 1.0`	1.336 s	—

Quality, RGB PSNR restricted to pixels with source alpha > 0:

vs. texconv: batched bc7e wins on every one of the 9 files, by +1.91 to +13.51 dB (median ≈ +2.6 dB). No dimension on which texconv wins. The +13.51 dB outlier is a hard-alpha cutout icon where texconv under-selects BC7 mode 7 (14.3% of blocks vs. 39.6% for both stock and this fork) — consistent with the p-bit issue in §1.
vs. stock at -Quality 1.0: matched. Better on 4 files, worse on 5, worst case 1.08 dB, no systematic direction. Same quality tier, 12.3× faster.

Independently reproduced on a second machine, OS and compiler. The above was measured on Linux/GCC (Ryzen 7 5800X); the full corpus was re-run on Windows/MSVC (Ryzen 7 5800H) and 26 of 27 recorded PSNR values reproduced digit-for-digit. The single exception is in the stock codec's float path, not this fork's — see §7. Absolute times aren't comparable across those machines; the qualitative result is.

Full tables, methodology and the investigation behind them: Phase 3 part 6 and Phase 4 part 4 in NOTES.md.

6. License and attribution

Two licenses apply to the shipped binary:

Component	License
AMD Compressonator (AMD/ATI code paths)	MIT
`bc7e.ispc` (richgel999/bc7enc_rdo)	Apache 2.0

Both license texts must be included in any release. The Apache 2.0 side additionally requires the release to identify the incorporated Apache-2.0 component, which is satisfied here:

This software incorporates bc7e.ispc from richgel999/bc7enc_rdo, © Richard Geldreich, licensed under the Apache License, Version 2.0.

ETCPack and its Ericsson SLA are not part of this build's license surface — see §3.

7. Known upstream issues found during this work

These are bugs in stock Compressonator, found while testing this fork's changes. They affect upstream regardless of this work.

#	Issue	Status
1	Zero-valid-modes crash. `-AlphaRestrict 1` with the default `-ModeMask` (0xCF) on mixed 0/255 alpha yields `validModeMask == 0` for affected blocks. Stock's search loop exits without setting `encodedBlock`, and the `if (!encodedBlock)` handler is a documented-as-error no-op — leaving the previous buffer contents as the encoded block. Non-deterministic garbage. A debug `assert` catches it, so it never trips in developer testing.	Not reported upstream
2	`CODECFLOAT` quality-precision truncation. `-Quality` loses precision twice on the CLI path (`std::stof` at `cmdline.cpp:411`, then a `(CODECFLOAT)` cast at `compress.cpp:223`; `CODECFLOAT` is `typedef float`). `-Quality 0.45` arrives as `0.44999998807907104`, landing on the wrong side of a threshold comparison. `CODECFLOAT` is used in 484 places, so the fix needs care.	Not reported upstream
3	MSVC-vs-GCC discrepancy in the stock codec. One file in the corpus (`ui_icon_mg36e`) reads 41.884 dB under MSVC vs. 41.89 dB recorded under GCC, in stock's float-heavy shaker refinement path. Checked at higher precision to rule out a rounding-boundary artifact — it's a real, tiny (≥0.005 dB) difference.	Not reported upstream — and not root-caused. No byte-level Linux reference exists for corpus outputs, so it couldn't be traced further.

8. Platform support

Platform	Status
Linux	Built, verified, fully static (0 dynamic dependencies).
Windows	Built, verified, statically linked (3 OS-shipped DLLs).
macOS	Out of scope. Not built, not tested. Contributions welcome.

Cross-platform correctness is backed by byte-identical output, not by assertion. Every verification MD5 produced by the Windows/MSVC build matches the Linux/GCC reference exactly:

Check	Result
Centibucket boundary sweep	9/9 exact MD5 match
Per-block vs. batched bit-identity	6/6 configs, 100.0000% block match
Zero-valid-modes guard output	payload MD5 `7f359e51bc5c` on both platforms

The same three harnesses were re-run after every relink and code-removal pass in this project — the static-linking change and the ETC removal both had to prove they were byte-level no-ops before being accepted.

Compressonator

Download the latest revision for changes that have been made since the last major release by clicking the CMake button above or the link here. Currently, only Compressonator Framework and Compressonator CLI are built every revision.

Compressonator is a set of tools to allow artists and developers to more easily create compressed texture assets or model mesh optimizations and easily visualize the quality impact of various compression and rendering technologies. It consists of a GUI application, a command line application and an SDK for easy integration into a developer tool chain.

Compressonator supports Microsoft Windows® and Linux builds.

For more details goto the online Compressonator Documents: http://compressonator.readthedocs.io/en/latest/

Build System Updates

The code is undergoing a build setup update. This notice will be removed when the changes are completed! Currently: To use the sln builds run build\fetch_dependencies.py to fetch required external lib dependencies into a Common folder above this repository.

Get Prebuilt Binaries and Installer here:

To build the source files follow the instructions in http://compressonator.readthedocs.io/en/latest/build_from_source/build_instructions.html

CMake Build Configuration

As of v4.2, The cmake command line options have settings to build specific libs and applications

Examples: Generating Visual Studio Solution File

Enable building all
    cmake -G "Visual Studio 15 2017 Win64"  
    
Disable all builds except external libs, minimal cmake base setup     
    cmake -DOPTION_ENABLE_ALL_APPS=OFF -G "Visual Studio 15 2017 Win64"
    
Enable only CLI app build    
    cmake -DOPTION_ENABLE_ALL_APPS=OFF -DOPTION_BUILD_APPS_CMP_CLI=ON -G "Visual Studio 15 2017 Win64"

For more details reference the CMakeList file on the root folder.

Style and Format Change

The source code of this product is being reformatted to follow the Google C++ Style Guide https://google.github.io/styleguide/cppguide.html

In the interim you may encounter a mix of both an older C++ coding style, as well as the newer Google C++ Style.

Please refer to the _clang-format file in the root directory of the product for additional style information.

Compressonator Core

Provides block level API access to updated performance and quality driven BCn codecs. The library is designed to be a small self-contained, cross-platform, and linkable library for user applications.

Example usage is shown as below to compress and decompress a single 4x4 image block using BC1 encoder

// To use Compressonator Core "C" interfaces, just include
// a single header file and CMP_Core lib into  your projects

#include "CMP_Core.h"

// Compress a sample image shape0 which is a 4x4 RGBA_8888 block.
// Users can use a pointer to any sized image buffers to reference
// a 4x4 block by supplying a stride offset for the next row.
// Optional BC1 settings is set to null in this example

unsigned char shape0_RGBA[64] = { filled with image source data as RGBA ...};

// cmpBuffer is a byte array of 8 byte to hold the compressed results.
unsigned char cmpBuffer[8]   = { 0 };

// Compress the source into cmpBuffer
CompressBlockBC1(shape0_RGBA, 16, cmpBuffer,null);

// Example to decompress comBuffer back to a RGBA_8888 4x4 image block
unsigned char imgBuffer[64] = { 0 };
DecompressBlockBC1(cmpBuffer,imgBuffer,null)

Compressonator Framework

Includes Compressonator core with interfaces for multi-threading, mipmap generation, file access of images and HPC pipeline interfaces.

Example Mip Level Processing using CPU

// To use Compressonator Framework "C" interfaces, just include
// a single header file and CMP_Framework lib into  your projects

#include "compressonator.h"

 //--------------------------
 // Init frameworks
 // plugin and IO interfaces
 //--------------------------
 CMP_InitFramework();

//---------------
// Load the image
//---------------
CMP_MipSet MipSetIn;
memset(&MipSetIn, 0, sizeof(CMP_MipSet));
cmp_status = CMP_LoadTexture(pszSourceFile, &MipSetIn);
if (cmp_status != CMP_OK) {
    std::printf("Error %d: Loading source file!\n",cmp_status);
    return -1;
}

//----------------------------------------------------------------------
// generate mipmap level for the source image, if not already generated
//----------------------------------------------------------------------

if (MipSetIn.m_nMipLevels <= 1)
{
    CMP_INT requestLevel = 10; // Request 10 miplevels for the source image

    //------------------------------------------------------------------------
    // Checks what the minimum image size will be for the requested mip levels
    // if the request is too large, a adjusted minimum size will be returned
    //------------------------------------------------------------------------
    CMP_INT nMinSize = CMP_CalcMinMipSize(MipSetIn.m_nHeight, MipSetIn.m_nWidth, 10);

    //--------------------------------------------------------------
    // now that the minimum size is known, generate the miplevels
    // users can set any requested minumum size to use. The correct
    // miplevels will be set acordingly.
    //--------------------------------------------------------------
    CMP_GenerateMIPLevels(&MipSetIn, nMinSize);
}

//==========================
// Set Compression Options
//==========================
KernelOptions   kernel_options;
memset(&kernel_options, 0, sizeof(KernelOptions));

kernel_options.format   = destFormat;   // Set the format to process
kernel_options.fquality = fQuality;     // Set the quality of the result
kernel_options.threads  = 0;            // Auto setting

//=====================================================
// example of using BC1 encoder options 
// kernel_options.bc15 is valid for BC1 to BC5 formats
//=====================================================
if (destFormat == CMP_FORMAT_BC1)
{
    // Enable punch through alpha setting
    kernel_options.bc15.useAlphaThreshold = true;
    kernel_options.bc15.alphaThreshold    = 128;

    // Enable setting channel weights
    kernel_options.bc15.useChannelWeights = true;
    kernel_options.bc15.channelWeights[0] = 0.3086f;
    kernel_options.bc15.channelWeights[1] = 0.6094f;
    kernel_options.bc15.channelWeights[2] = 0.0820f;
}

//--------------------------------------------------------------
// Setup a results buffer for the processed file,
// the content will be set after the source texture is processed
// in the call to CMP_ProcessTexture()
//--------------------------------------------------------------
CMP_MipSet MipSetCmp;
memset(&MipSetCmp, 0, sizeof(CMP_MipSet));

//===============================================
// Compress the texture using Framework Lib
//===============================================
cmp_status = CMP_ProcessTexture(&MipSetIn, &MipSetCmp, kernel_options, CompressionCallback);
if (cmp_status != CMP_OK) {
  ...
}

//----------------------------------------------------------------
// Save the result into a DDS file
//----------------------------------------------------------------
cmp_status = CMP_SaveTexture(pszDestFile, &MipSetCmp);

CMP_FreeMipSet(&MipSetIn);
CMP_FreeMipSet(&MipSetCmp);

Example GPU based processing using OpenCL

// Note: Only MD x64 build is used for GPU processing
// SDK files required for application:
//     compressonator.h
//     CMP_Framework_xx.lib  For static libs xx is either MD or MDd, 
//                      When using DLL's make sure the  CMP_Framework_xx_DLL.dll is in exe path
//
// File(s) required to run with the built application
//
// Using OpenCL (OCL) 
//     CMP_GPU_OCL_MD_DLL.dll    or CMP_GPU_OCL_MDd_DLL.dll
//     Encode Kernel files in plugins/compute folder
//     BC1_Encode_Kernel.cpp
//     BC1_Encode_Kernel.h
//     BCn_Common_kernel.h
//     Common_Def.h
//
// Using DirectX (DXC) 
//     CMP_GPU_DXC_MD_DLL.dll    or CMP_GPU_DXC_MDd_DLL.dll
//     Encode Kernel files in plugins/compute folder
//     BC1_Encode_Kernel.hlsl
//     BCn_Common_kernel.h
//     Common_Def.h

#include "compressonator.h"

CMP_FORMAT      destFormat = CMP_FORMAT_BC1;

//---------------
// Load the image
//---------------
CMP_MipSet MipSetIn;
memset(&MipSetIn, 0, sizeof(CMP_MipSet));
if (CMP_LoadTexture(pszSourceFile, &MipSetIn) != CMP_OK) {
    std::printf("Error: Loading source file!\n");
    return -1;
  } 

 //-----------------------------------------------------
 // when using GPU: The texture must have width and height as a multiple of 4
 // Check texture for width and height
 //-----------------------------------------------------
 if ((MipSetIn.m_nWidth % 4) > 0 || (MipSetIn.m_nHeight % 4) > 0) {
    std::printf("Error: Texture width and height must be multiple of 4\n");
    return -1;
 }
    
//----------------------------------------------------------------------------------------------------------
// Set the target compression format and the host framework to use
// For this example OpenCL is been used
//-----------------------------------------------------------------------------------------------------------
KernelOptions   kernel_options;
memset(&kernel_options, 0, sizeof(KernelOptions));

kernel_options.encodeWith = CMP_GPU_OCL;         // Using OpenCL GPU Encoder, can replace with DXC for DirectX
kernel_options.format     = destFormat;          // Set the format to process
kernel_options.fquality   = fQuality;            // Set the quality of the result

//--------------------------------------------------------------
// Setup a results buffer for the processed file,
// the content will be set after the source texture is processed
// in the call to CMP_ProcessTexture()
//--------------------------------------------------------------
CMP_MipSet MipSetCmp;
memset(&MipSetCmp, 0, sizeof(CMP_MipSet));

//===============================================
// Compress the texture using Framework Lib
//===============================================
cmp_status = CMP_ProcessTexture(&MipSetIn, &MipSetCmp, kernel_options, CompressionCallback);
if (cmp_status != CMP_OK) {
  ...
}

//----------------------------------------------------------------
// Save the result into a DDS file
//----------------------------------------------------------------
cmp_status = CMP_SaveTexture(pszDestFile, &MipSetCmp);

CMP_FreeMipSet(&MipSetIn);
CMP_FreeMipSet(&MipSetCmp);

Compressonator SDK

Compressonator SDK supported codecs includes BC1-BC7/DXTC, ETC1, ETC2, ASTC, ATC, ATI1N, ATI2N, all available in a single library.

With the new SDK installation, several example applications with source code are provided that demonstrate how easy it is to add texture compression to your own applications using either "High Level" or "Block Level" APIs.

A simple thread safe interface can compress, decompress and transcode any image as required

CMP_ConvertTexture(CMP_Texture* pSourceTexture, CMP_Texture* pDestTexture,...);

For Example:

// To use Compressonator's portable "C" interfaces, just include
// a single header file and Compresonator.lib into  your projects

#include "Compressonator.h"
...

//==========================
// Load Source Texture
//==========================
CMP_Texture srcTexture;
// note that LoadDDSFile function is a utils function to initialize the source CMP_Texture
// you can also initialize the source CMP_Texture the same way as initialize destination CMP_Texture
if (!LoadDDSFile(pszSourceFile, srcTexture))
{
  ...
}

//===================================
// Initialize Compressed Destination
//===================================
CMP_Texture destTexture;
destTexture.dwSize     = sizeof(destTexture);
destTexture.dwWidth    = srcTexture.dwWidth;
destTexture.dwHeight   = srcTexture.dwHeight;
destTexture.dwPitch    = 0;
destTexture.format     = CMP_FORMAT_BC6H;
destTexture.dwDataSize = CMP_CalculateBufferSize(&destTexture);
destTexture.pData      = (CMP_BYTE*)malloc(destTexture.dwDataSize);

//==========================
// Set Compression Options
//==========================
CMP_CompressOptions options = {0};
options.dwSize       = sizeof(options);
options.fquality     = 0.05f;
options.dwnumThreads = 8;

//==========================
// Compress Texture
//==========================
CMP_ERROR   cmp_status;
cmp_status = CMP_ConvertTexture(&srcTexture, &destTexture, &options, &CompressionCallback, NULL, NULL);
if (cmp_status != CMP_OK)
{
  ...
}

//==========================
// Save Compressed Testure
//==========================
SaveDDSFile(pszDestFile, destTexture))

free(srcTexture.pData);
free(destTexture.pData);

Compressonator CLI

Command line application that can be batch processed and supports:

Texture Compression, Decompression, Format Transcoding.
3D Model Optimization and Mesh Compression.
Performance and Analysis Logs such as SSIM, MSE, PSNR.
MIP Maps, Image Differences, etc. ...

C:\>CompressonatorCLI -fd BC7 .\images .results

C:\>CompressonatorCLI -log -fd BC7 .\images\ruby.png ruby_bc7.dds

CompressonatorCLI Performance Log v1.0

Source        : .\images\ruby.png, Height 416, Wideth 576, Size 0.936 MBytes
Destination   : ruby_bc7.dds
Using         : CPU
Quality       : 0.05
Processed to  : BC7        with  1 iteration(s) in 1.422 seconds
MSE           : 0.78
PSNR          : 49.2
SSIM          : 0.9978
Total time    : 1.432 seconds

--------------

Compressonator GUI

Comprehensive graphical application that can be used to visualize Images and 3D Models, with support for:

Texture Compression, Decompression, Format Transcoding.
3D Model Optimization and Mesh Compression.
Multiple Image and 3D Model Views.
MIP Maps, Differences, Analysis, etc. ...

glTF 2.0 Model Render View

Contributors

Compressonator's GitHub repository (http://github.com/GPUOpen-Tools/Compressonator) is moderated by Advanced Micro Devices, Inc. as part of the GPUOpen initiative.

AMD encourages any and all contributors to submit changes, features, and bug fixes via Git pull requests to this repository.

Users are also encouraged to submit issues and feature requests via the repository's issue tracker.

Name		Name	Last commit message	Last commit date
.github/workflows		.github/workflows
applications		applications
build		build
build_sdk		build_sdk
cmake		cmake
cmp_compressonatorlib		cmp_compressonatorlib
cmp_core		cmp_core
cmp_framework		cmp_framework
cmp_unittests		cmp_unittests
docs		docs
examples		examples
external		external
installer		installer
license		license
runtime		runtime
scripts		scripts
tools		tools
.gitattributes		.gitattributes
.gitignore		.gitignore
CMakeLists.txt		CMakeLists.txt
NOTES.md		NOTES.md
_clang-format		_clang-format
clang-tidy		clang-tidy
compressonator_root.props		compressonator_root.props
fetch_common.bat		fetch_common.bat
gpu_decodelib.def		gpu_decodelib.def
readme.md		readme.md
usedefinitions.h		usedefinitions.h
vsbuildALL.bat		vsbuildALL.bat
vsclean.bat		vsclean.bat
vsdev.bat		vsdev.bat

Navigation Menu

Repository files navigation

Compressonator + bc7enc_rdo (BC7 CPU codec swap)

1. What this fork is

2. Usage

Existing flags still work — and were verified, not assumed

3. What's different from stock, and why

Omitted from this build

What you get in exchange

4. Building

Prerequisites

5. Results

6. License and attribution

7. Known upstream issues found during this work

8. Platform support

Compressonator

Build System Updates

Get Prebuilt Binaries and Installer here:

CMake Build Configuration

Style and Format Change

Compressonator Core

Compressonator Framework

Compressonator SDK

Compressonator CLI

Compressonator GUI

Contributors

About

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

Compressonator + bc7enc_rdo (BC7 CPU codec swap)

1. What this fork is

2. Usage

Existing flags still work — and were verified, not assumed

3. What's different from stock, and why

Omitted from this build

What you get in exchange

4. Building

Prerequisites

5. Results

6. License and attribution

7. Known upstream issues found during this work

8. Platform support

Compressonator

Build System Updates

Get Prebuilt Binaries and Installer here:

CMake Build Configuration

Style and Format Change

Compressonator Core

Compressonator Framework

Compressonator SDK

Compressonator CLI

Compressonator GUI

Contributors

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages