diff options
| -rw-r--r-- | modules/cvtt/image_compress_cvtt.cpp | 23 | ||||
| -rw-r--r-- | thirdparty/README.md | 5 | ||||
| -rw-r--r-- | thirdparty/cvtt/ConvectionKernels_BC67.cpp | 339 | ||||
| -rw-r--r-- | thirdparty/cvtt/patches/revert_BC6H_reorg.patch | 393 |
4 files changed, 726 insertions, 34 deletions
diff --git a/modules/cvtt/image_compress_cvtt.cpp b/modules/cvtt/image_compress_cvtt.cpp index f19228cb18..ad70772270 100644 --- a/modules/cvtt/image_compress_cvtt.cpp +++ b/modules/cvtt/image_compress_cvtt.cpp @@ -30,8 +30,8 @@ #include "image_compress_cvtt.h" +#include "core/object/worker_thread_pool.h" #include "core/os/os.h" -#include "core/os/thread.h" #include "core/string/print_string.h" #include "core/templates/safe_refcount.h" @@ -129,6 +129,18 @@ static void _digest_row_task(const CVTTCompressionJobParams &p_job_params, const } } +static void _digest_job_queue(void *p_job_queue, uint32_t p_index) { + CVTTCompressionJobQueue *job_queue = static_cast<CVTTCompressionJobQueue *>(p_job_queue); + uint32_t num_tasks = job_queue->num_tasks; + uint32_t total_threads = WorkerThreadPool::get_singleton()->get_thread_count(); + uint32_t start = p_index * num_tasks / total_threads; + uint32_t end = (p_index + 1 == total_threads) ? num_tasks : ((p_index + 1) * num_tasks / total_threads); + + for (uint32_t i = start; i < end; i++) { + _digest_row_task(job_queue->job_params, job_queue->job_tasks[i]); + } +} + void image_compress_cvtt(Image *p_image, Image::UsedChannels p_channels) { if (p_image->get_format() >= Image::FORMAT_BPTC_RGBA) { return; //do not compress, already compressed @@ -220,7 +232,7 @@ void image_compress_cvtt(Image *p_image, Image::UsedChannels p_channels) { row_task.in_mm_bytes = in_bytes; row_task.out_mm_bytes = out_bytes; - _digest_row_task(job_queue.job_params, row_task); + tasks.push_back(row_task); out_bytes += 16 * (bw / 4); } @@ -230,6 +242,13 @@ void image_compress_cvtt(Image *p_image, Image::UsedChannels p_channels) { h = MAX(h / 2, 1); } + const CVTTCompressionRowTask *tasks_rb = tasks.ptr(); + + job_queue.job_tasks = &tasks_rb[0]; + job_queue.num_tasks = static_cast<uint32_t>(tasks.size()); + WorkerThreadPool::GroupID group_task = WorkerThreadPool::get_singleton()->add_native_group_task(&_digest_job_queue, &job_queue, WorkerThreadPool::get_singleton()->get_thread_count(), -1, true, SNAME("CVTT Compress")); + WorkerThreadPool::get_singleton()->wait_for_group_task_completion(group_task); + p_image->set_data(p_image->get_width(), p_image->get_height(), p_image->has_mipmaps(), target_format, data); } diff --git a/thirdparty/README.md b/thirdparty/README.md index f883a3a6da..92e0a91596 100644 --- a/thirdparty/README.md +++ b/thirdparty/README.md @@ -70,6 +70,11 @@ Files extracted from upstream source: - all .cpp, .h, and .txt files except the folders MakeTables and etc2packer. +Changes related to BC6H packing and unpacking made upstream in +https://github.com/elasota/cvtt/commit/2e4b6b2747aec11f4cc6dd09ef43fa8ce769f6e2 +have been removed as they caused massive quality regressions. Apply the patches +in the `patches/` folder when syncing on newer upstream commits. + ## doctest diff --git a/thirdparty/cvtt/ConvectionKernels_BC67.cpp b/thirdparty/cvtt/ConvectionKernels_BC67.cpp index 021d658c08..011f51570b 100644 --- a/thirdparty/cvtt/ConvectionKernels_BC67.cpp +++ b/thirdparty/cvtt/ConvectionKernels_BC67.cpp @@ -76,6 +76,205 @@ namespace cvtt }; } + namespace BC6HData + { + enum EField + { + NA, // N/A + M, // Mode + D, // Shape + RW, + RX, + RY, + RZ, + GW, + GX, + GY, + GZ, + BW, + BX, + BY, + BZ, + }; + + struct ModeDescriptor + { + EField m_eField; + uint8_t m_uBit; + }; + + const ModeDescriptor g_modeDescriptors[14][82] = + { + { // Mode 1 (0x00) - 10 5 5 5 + { M, 0 },{ M, 1 },{ GY, 4 },{ BY, 4 },{ BZ, 4 },{ RW, 0 },{ RW, 1 },{ RW, 2 },{ RW, 3 },{ RW, 4 }, + { RW, 5 },{ RW, 6 },{ RW, 7 },{ RW, 8 },{ RW, 9 },{ GW, 0 },{ GW, 1 },{ GW, 2 },{ GW, 3 },{ GW, 4 }, + { GW, 5 },{ GW, 6 },{ GW, 7 },{ GW, 8 },{ GW, 9 },{ BW, 0 },{ BW, 1 },{ BW, 2 },{ BW, 3 },{ BW, 4 }, + { BW, 5 },{ BW, 6 },{ BW, 7 },{ BW, 8 },{ BW, 9 },{ RX, 0 },{ RX, 1 },{ RX, 2 },{ RX, 3 },{ RX, 4 }, + { GZ, 4 },{ GY, 0 },{ GY, 1 },{ GY, 2 },{ GY, 3 },{ GX, 0 },{ GX, 1 },{ GX, 2 },{ GX, 3 },{ GX, 4 }, + { BZ, 0 },{ GZ, 0 },{ GZ, 1 },{ GZ, 2 },{ GZ, 3 },{ BX, 0 },{ BX, 1 },{ BX, 2 },{ BX, 3 },{ BX, 4 }, + { BZ, 1 },{ BY, 0 },{ BY, 1 },{ BY, 2 },{ BY, 3 },{ RY, 0 },{ RY, 1 },{ RY, 2 },{ RY, 3 },{ RY, 4 }, + { BZ, 2 },{ RZ, 0 },{ RZ, 1 },{ RZ, 2 },{ RZ, 3 },{ RZ, 4 },{ BZ, 3 },{ D, 0 },{ D, 1 },{ D, 2 }, + { D, 3 },{ D, 4 }, + }, + + { // Mode 2 (0x01) - 7 6 6 6 + { M, 0 },{ M, 1 },{ GY, 5 },{ GZ, 4 },{ GZ, 5 },{ RW, 0 },{ RW, 1 },{ RW, 2 },{ RW, 3 },{ RW, 4 }, + { RW, 5 },{ RW, 6 },{ BZ, 0 },{ BZ, 1 },{ BY, 4 },{ GW, 0 },{ GW, 1 },{ GW, 2 },{ GW, 3 },{ GW, 4 }, + { GW, 5 },{ GW, 6 },{ BY, 5 },{ BZ, 2 },{ GY, 4 },{ BW, 0 },{ BW, 1 },{ BW, 2 },{ BW, 3 },{ BW, 4 }, + { BW, 5 },{ BW, 6 },{ BZ, 3 },{ BZ, 5 },{ BZ, 4 },{ RX, 0 },{ RX, 1 },{ RX, 2 },{ RX, 3 },{ RX, 4 }, + { RX, 5 },{ GY, 0 },{ GY, 1 },{ GY, 2 },{ GY, 3 },{ GX, 0 },{ GX, 1 },{ GX, 2 },{ GX, 3 },{ GX, 4 }, + { GX, 5 },{ GZ, 0 },{ GZ, 1 },{ GZ, 2 },{ GZ, 3 },{ BX, 0 },{ BX, 1 },{ BX, 2 },{ BX, 3 },{ BX, 4 }, + { BX, 5 },{ BY, 0 },{ BY, 1 },{ BY, 2 },{ BY, 3 },{ RY, 0 },{ RY, 1 },{ RY, 2 },{ RY, 3 },{ RY, 4 }, + { RY, 5 },{ RZ, 0 },{ RZ, 1 },{ RZ, 2 },{ RZ, 3 },{ RZ, 4 },{ RZ, 5 },{ D, 0 },{ D, 1 },{ D, 2 }, + { D, 3 },{ D, 4 }, + }, + + { // Mode 3 (0x02) - 11 5 4 4 + { M, 0 },{ M, 1 },{ M, 2 },{ M, 3 },{ M, 4 },{ RW, 0 },{ RW, 1 },{ RW, 2 },{ RW, 3 },{ RW, 4 }, + { RW, 5 },{ RW, 6 },{ RW, 7 },{ RW, 8 },{ RW, 9 },{ GW, 0 },{ GW, 1 },{ GW, 2 },{ GW, 3 },{ GW, 4 }, + { GW, 5 },{ GW, 6 },{ GW, 7 },{ GW, 8 },{ GW, 9 },{ BW, 0 },{ BW, 1 },{ BW, 2 },{ BW, 3 },{ BW, 4 }, + { BW, 5 },{ BW, 6 },{ BW, 7 },{ BW, 8 },{ BW, 9 },{ RX, 0 },{ RX, 1 },{ RX, 2 },{ RX, 3 },{ RX, 4 }, + { RW,10 },{ GY, 0 },{ GY, 1 },{ GY, 2 },{ GY, 3 },{ GX, 0 },{ GX, 1 },{ GX, 2 },{ GX, 3 },{ GW,10 }, + { BZ, 0 },{ GZ, 0 },{ GZ, 1 },{ GZ, 2 },{ GZ, 3 },{ BX, 0 },{ BX, 1 },{ BX, 2 },{ BX, 3 },{ BW,10 }, + { BZ, 1 },{ BY, 0 },{ BY, 1 },{ BY, 2 },{ BY, 3 },{ RY, 0 },{ RY, 1 },{ RY, 2 },{ RY, 3 },{ RY, 4 }, + { BZ, 2 },{ RZ, 0 },{ RZ, 1 },{ RZ, 2 },{ RZ, 3 },{ RZ, 4 },{ BZ, 3 },{ D, 0 },{ D, 1 },{ D, 2 }, + { D, 3 },{ D, 4 }, + }, + + { // Mode 4 (0x06) - 11 4 5 4 + { M, 0 },{ M, 1 },{ M, 2 },{ M, 3 },{ M, 4 },{ RW, 0 },{ RW, 1 },{ RW, 2 },{ RW, 3 },{ RW, 4 }, + { RW, 5 },{ RW, 6 },{ RW, 7 },{ RW, 8 },{ RW, 9 },{ GW, 0 },{ GW, 1 },{ GW, 2 },{ GW, 3 },{ GW, 4 }, + { GW, 5 },{ GW, 6 },{ GW, 7 },{ GW, 8 },{ GW, 9 },{ BW, 0 },{ BW, 1 },{ BW, 2 },{ BW, 3 },{ BW, 4 }, + { BW, 5 },{ BW, 6 },{ BW, 7 },{ BW, 8 },{ BW, 9 },{ RX, 0 },{ RX, 1 },{ RX, 2 },{ RX, 3 },{ RW,10 }, + { GZ, 4 },{ GY, 0 },{ GY, 1 },{ GY, 2 },{ GY, 3 },{ GX, 0 },{ GX, 1 },{ GX, 2 },{ GX, 3 },{ GX, 4 }, + { GW,10 },{ GZ, 0 },{ GZ, 1 },{ GZ, 2 },{ GZ, 3 },{ BX, 0 },{ BX, 1 },{ BX, 2 },{ BX, 3 },{ BW,10 }, + { BZ, 1 },{ BY, 0 },{ BY, 1 },{ BY, 2 },{ BY, 3 },{ RY, 0 },{ RY, 1 },{ RY, 2 },{ RY, 3 },{ BZ, 0 }, + { BZ, 2 },{ RZ, 0 },{ RZ, 1 },{ RZ, 2 },{ RZ, 3 },{ GY, 4 },{ BZ, 3 },{ D, 0 },{ D, 1 },{ D, 2 }, + { D, 3 },{ D, 4 }, + }, + + { // Mode 5 (0x0a) - 11 4 4 5 + { M, 0 },{ M, 1 },{ M, 2 },{ M, 3 },{ M, 4 },{ RW, 0 },{ RW, 1 },{ RW, 2 },{ RW, 3 },{ RW, 4 }, + { RW, 5 },{ RW, 6 },{ RW, 7 },{ RW, 8 },{ RW, 9 },{ GW, 0 },{ GW, 1 },{ GW, 2 },{ GW, 3 },{ GW, 4 }, + { GW, 5 },{ GW, 6 },{ GW, 7 },{ GW, 8 },{ GW, 9 },{ BW, 0 },{ BW, 1 },{ BW, 2 },{ BW, 3 },{ BW, 4 }, + { BW, 5 },{ BW, 6 },{ BW, 7 },{ BW, 8 },{ BW, 9 },{ RX, 0 },{ RX, 1 },{ RX, 2 },{ RX, 3 },{ RW,10 }, + { BY, 4 },{ GY, 0 },{ GY, 1 },{ GY, 2 },{ GY, 3 },{ GX, 0 },{ GX, 1 },{ GX, 2 },{ GX, 3 },{ GW,10 }, + { BZ, 0 },{ GZ, 0 },{ GZ, 1 },{ GZ, 2 },{ GZ, 3 },{ BX, 0 },{ BX, 1 },{ BX, 2 },{ BX, 3 },{ BX, 4 }, + { BW,10 },{ BY, 0 },{ BY, 1 },{ BY, 2 },{ BY, 3 },{ RY, 0 },{ RY, 1 },{ RY, 2 },{ RY, 3 },{ BZ, 1 }, + { BZ, 2 },{ RZ, 0 },{ RZ, 1 },{ RZ, 2 },{ RZ, 3 },{ BZ, 4 },{ BZ, 3 },{ D, 0 },{ D, 1 },{ D, 2 }, + { D, 3 },{ D, 4 }, + }, + + { // Mode 6 (0x0e) - 9 5 5 5 + { M, 0 },{ M, 1 },{ M, 2 },{ M, 3 },{ M, 4 },{ RW, 0 },{ RW, 1 },{ RW, 2 },{ RW, 3 },{ RW, 4 }, + { RW, 5 },{ RW, 6 },{ RW, 7 },{ RW, 8 },{ BY, 4 },{ GW, 0 },{ GW, 1 },{ GW, 2 },{ GW, 3 },{ GW, 4 }, + { GW, 5 },{ GW, 6 },{ GW, 7 },{ GW, 8 },{ GY, 4 },{ BW, 0 },{ BW, 1 },{ BW, 2 },{ BW, 3 },{ BW, 4 }, + { BW, 5 },{ BW, 6 },{ BW, 7 },{ BW, 8 },{ BZ, 4 },{ RX, 0 },{ RX, 1 },{ RX, 2 },{ RX, 3 },{ RX, 4 }, + { GZ, 4 },{ GY, 0 },{ GY, 1 },{ GY, 2 },{ GY, 3 },{ GX, 0 },{ GX, 1 },{ GX, 2 },{ GX, 3 },{ GX, 4 }, + { BZ, 0 },{ GZ, 0 },{ GZ, 1 },{ GZ, 2 },{ GZ, 3 },{ BX, 0 },{ BX, 1 },{ BX, 2 },{ BX, 3 },{ BX, 4 }, + { BZ, 1 },{ BY, 0 },{ BY, 1 },{ BY, 2 },{ BY, 3 },{ RY, 0 },{ RY, 1 },{ RY, 2 },{ RY, 3 },{ RY, 4 }, + { BZ, 2 },{ RZ, 0 },{ RZ, 1 },{ RZ, 2 },{ RZ, 3 },{ RZ, 4 },{ BZ, 3 },{ D, 0 },{ D, 1 },{ D, 2 }, + { D, 3 },{ D, 4 }, + }, + + { // Mode 7 (0x12) - 8 6 5 5 + { M, 0 },{ M, 1 },{ M, 2 },{ M, 3 },{ M, 4 },{ RW, 0 },{ RW, 1 },{ RW, 2 },{ RW, 3 },{ RW, 4 }, + { RW, 5 },{ RW, 6 },{ RW, 7 },{ GZ, 4 },{ BY, 4 },{ GW, 0 },{ GW, 1 },{ GW, 2 },{ GW, 3 },{ GW, 4 }, + { GW, 5 },{ GW, 6 },{ GW, 7 },{ BZ, 2 },{ GY, 4 },{ BW, 0 },{ BW, 1 },{ BW, 2 },{ BW, 3 },{ BW, 4 }, + { BW, 5 },{ BW, 6 },{ BW, 7 },{ BZ, 3 },{ BZ, 4 },{ RX, 0 },{ RX, 1 },{ RX, 2 },{ RX, 3 },{ RX, 4 }, + { RX, 5 },{ GY, 0 },{ GY, 1 },{ GY, 2 },{ GY, 3 },{ GX, 0 },{ GX, 1 },{ GX, 2 },{ GX, 3 },{ GX, 4 }, + { BZ, 0 },{ GZ, 0 },{ GZ, 1 },{ GZ, 2 },{ GZ, 3 },{ BX, 0 },{ BX, 1 },{ BX, 2 },{ BX, 3 },{ BX, 4 }, + { BZ, 1 },{ BY, 0 },{ BY, 1 },{ BY, 2 },{ BY, 3 },{ RY, 0 },{ RY, 1 },{ RY, 2 },{ RY, 3 },{ RY, 4 }, + { RY, 5 },{ RZ, 0 },{ RZ, 1 },{ RZ, 2 },{ RZ, 3 },{ RZ, 4 },{ RZ, 5 },{ D, 0 },{ D, 1 },{ D, 2 }, + { D, 3 },{ D, 4 }, + }, + + { // Mode 8 (0x16) - 8 5 6 5 + { M, 0 },{ M, 1 },{ M, 2 },{ M, 3 },{ M, 4 },{ RW, 0 },{ RW, 1 },{ RW, 2 },{ RW, 3 },{ RW, 4 }, + { RW, 5 },{ RW, 6 },{ RW, 7 },{ BZ, 0 },{ BY, 4 },{ GW, 0 },{ GW, 1 },{ GW, 2 },{ GW, 3 },{ GW, 4 }, + { GW, 5 },{ GW, 6 },{ GW, 7 },{ GY, 5 },{ GY, 4 },{ BW, 0 },{ BW, 1 },{ BW, 2 },{ BW, 3 },{ BW, 4 }, + { BW, 5 },{ BW, 6 },{ BW, 7 },{ GZ, 5 },{ BZ, 4 },{ RX, 0 },{ RX, 1 },{ RX, 2 },{ RX, 3 },{ RX, 4 }, + { GZ, 4 },{ GY, 0 },{ GY, 1 },{ GY, 2 },{ GY, 3 },{ GX, 0 },{ GX, 1 },{ GX, 2 },{ GX, 3 },{ GX, 4 }, + { GX, 5 },{ GZ, 0 },{ GZ, 1 },{ GZ, 2 },{ GZ, 3 },{ BX, 0 },{ BX, 1 },{ BX, 2 },{ BX, 3 },{ BX, 4 }, + { BZ, 1 },{ BY, 0 },{ BY, 1 },{ BY, 2 },{ BY, 3 },{ RY, 0 },{ RY, 1 },{ RY, 2 },{ RY, 3 },{ RY, 4 }, + { BZ, 2 },{ RZ, 0 },{ RZ, 1 },{ RZ, 2 },{ RZ, 3 },{ RZ, 4 },{ BZ, 3 },{ D, 0 },{ D, 1 },{ D, 2 }, + { D, 3 },{ D, 4 }, + }, + + { // Mode 9 (0x1a) - 8 5 5 6 + { M, 0 },{ M, 1 },{ M, 2 },{ M, 3 },{ M, 4 },{ RW, 0 },{ RW, 1 },{ RW, 2 },{ RW, 3 },{ RW, 4 }, + { RW, 5 },{ RW, 6 },{ RW, 7 },{ BZ, 1 },{ BY, 4 },{ GW, 0 },{ GW, 1 },{ GW, 2 },{ GW, 3 },{ GW, 4 }, + { GW, 5 },{ GW, 6 },{ GW, 7 },{ BY, 5 },{ GY, 4 },{ BW, 0 },{ BW, 1 },{ BW, 2 },{ BW, 3 },{ BW, 4 }, + { BW, 5 },{ BW, 6 },{ BW, 7 },{ BZ, 5 },{ BZ, 4 },{ RX, 0 },{ RX, 1 },{ RX, 2 },{ RX, 3 },{ RX, 4 }, + { GZ, 4 },{ GY, 0 },{ GY, 1 },{ GY, 2 },{ GY, 3 },{ GX, 0 },{ GX, 1 },{ GX, 2 },{ GX, 3 },{ GX, 4 }, + { BZ, 0 },{ GZ, 0 },{ GZ, 1 },{ GZ, 2 },{ GZ, 3 },{ BX, 0 },{ BX, 1 },{ BX, 2 },{ BX, 3 },{ BX, 4 }, + { BX, 5 },{ BY, 0 },{ BY, 1 },{ BY, 2 },{ BY, 3 },{ RY, 0 },{ RY, 1 },{ RY, 2 },{ RY, 3 },{ RY, 4 }, + { BZ, 2 },{ RZ, 0 },{ RZ, 1 },{ RZ, 2 },{ RZ, 3 },{ RZ, 4 },{ BZ, 3 },{ D, 0 },{ D, 1 },{ D, 2 }, + { D, 3 },{ D, 4 }, + }, + + { // Mode 10 (0x1e) - 6 6 6 6 + { M, 0 },{ M, 1 },{ M, 2 },{ M, 3 },{ M, 4 },{ RW, 0 },{ RW, 1 },{ RW, 2 },{ RW, 3 },{ RW, 4 }, + { RW, 5 },{ GZ, 4 },{ BZ, 0 },{ BZ, 1 },{ BY, 4 },{ GW, 0 },{ GW, 1 },{ GW, 2 },{ GW, 3 },{ GW, 4 }, + { GW, 5 },{ GY, 5 },{ BY, 5 },{ BZ, 2 },{ GY, 4 },{ BW, 0 },{ BW, 1 },{ BW, 2 },{ BW, 3 },{ BW, 4 }, + { BW, 5 },{ GZ, 5 },{ BZ, 3 },{ BZ, 5 },{ BZ, 4 },{ RX, 0 },{ RX, 1 },{ RX, 2 },{ RX, 3 },{ RX, 4 }, + { RX, 5 },{ GY, 0 },{ GY, 1 },{ GY, 2 },{ GY, 3 },{ GX, 0 },{ GX, 1 },{ GX, 2 },{ GX, 3 },{ GX, 4 }, + { GX, 5 },{ GZ, 0 },{ GZ, 1 },{ GZ, 2 },{ GZ, 3 },{ BX, 0 },{ BX, 1 },{ BX, 2 },{ BX, 3 },{ BX, 4 }, + { BX, 5 },{ BY, 0 },{ BY, 1 },{ BY, 2 },{ BY, 3 },{ RY, 0 },{ RY, 1 },{ RY, 2 },{ RY, 3 },{ RY, 4 }, + { RY, 5 },{ RZ, 0 },{ RZ, 1 },{ RZ, 2 },{ RZ, 3 },{ RZ, 4 },{ RZ, 5 },{ D, 0 },{ D, 1 },{ D, 2 }, + { D, 3 },{ D, 4 }, + }, + + { // Mode 11 (0x03) - 10 10 + { M, 0 },{ M, 1 },{ M, 2 },{ M, 3 },{ M, 4 },{ RW, 0 },{ RW, 1 },{ RW, 2 },{ RW, 3 },{ RW, 4 }, + { RW, 5 },{ RW, 6 },{ RW, 7 },{ RW, 8 },{ RW, 9 },{ GW, 0 },{ GW, 1 },{ GW, 2 },{ GW, 3 },{ GW, 4 }, + { GW, 5 },{ GW, 6 },{ GW, 7 },{ GW, 8 },{ GW, 9 },{ BW, 0 },{ BW, 1 },{ BW, 2 },{ BW, 3 },{ BW, 4 }, + { BW, 5 },{ BW, 6 },{ BW, 7 },{ BW, 8 },{ BW, 9 },{ RX, 0 },{ RX, 1 },{ RX, 2 },{ RX, 3 },{ RX, 4 }, + { RX, 5 },{ RX, 6 },{ RX, 7 },{ RX, 8 },{ RX, 9 },{ GX, 0 },{ GX, 1 },{ GX, 2 },{ GX, 3 },{ GX, 4 }, + { GX, 5 },{ GX, 6 },{ GX, 7 },{ GX, 8 },{ GX, 9 },{ BX, 0 },{ BX, 1 },{ BX, 2 },{ BX, 3 },{ BX, 4 }, + { BX, 5 },{ BX, 6 },{ BX, 7 },{ BX, 8 },{ BX, 9 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 }, + { NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 }, + { NA, 0 },{ NA, 0 }, + }, + + { // Mode 12 (0x07) - 11 9 + { M, 0 },{ M, 1 },{ M, 2 },{ M, 3 },{ M, 4 },{ RW, 0 },{ RW, 1 },{ RW, 2 },{ RW, 3 },{ RW, 4 }, + { RW, 5 },{ RW, 6 },{ RW, 7 },{ RW, 8 },{ RW, 9 },{ GW, 0 },{ GW, 1 },{ GW, 2 },{ GW, 3 },{ GW, 4 }, + { GW, 5 },{ GW, 6 },{ GW, 7 },{ GW, 8 },{ GW, 9 },{ BW, 0 },{ BW, 1 },{ BW, 2 },{ BW, 3 },{ BW, 4 }, + { BW, 5 },{ BW, 6 },{ BW, 7 },{ BW, 8 },{ BW, 9 },{ RX, 0 },{ RX, 1 },{ RX, 2 },{ RX, 3 },{ RX, 4 }, + { RX, 5 },{ RX, 6 },{ RX, 7 },{ RX, 8 },{ RW,10 },{ GX, 0 },{ GX, 1 },{ GX, 2 },{ GX, 3 },{ GX, 4 }, + { GX, 5 },{ GX, 6 },{ GX, 7 },{ GX, 8 },{ GW,10 },{ BX, 0 },{ BX, 1 },{ BX, 2 },{ BX, 3 },{ BX, 4 }, + { BX, 5 },{ BX, 6 },{ BX, 7 },{ BX, 8 },{ BW,10 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 }, + { NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 }, + { NA, 0 },{ NA, 0 }, + }, + + { // Mode 13 (0x0b) - 12 8 + { M, 0 },{ M, 1 },{ M, 2 },{ M, 3 },{ M, 4 },{ RW, 0 },{ RW, 1 },{ RW, 2 },{ RW, 3 },{ RW, 4 }, + { RW, 5 },{ RW, 6 },{ RW, 7 },{ RW, 8 },{ RW, 9 },{ GW, 0 },{ GW, 1 },{ GW, 2 },{ GW, 3 },{ GW, 4 }, + { GW, 5 },{ GW, 6 },{ GW, 7 },{ GW, 8 },{ GW, 9 },{ BW, 0 },{ BW, 1 },{ BW, 2 },{ BW, 3 },{ BW, 4 }, + { BW, 5 },{ BW, 6 },{ BW, 7 },{ BW, 8 },{ BW, 9 },{ RX, 0 },{ RX, 1 },{ RX, 2 },{ RX, 3 },{ RX, 4 }, + { RX, 5 },{ RX, 6 },{ RX, 7 },{ RW,11 },{ RW,10 },{ GX, 0 },{ GX, 1 },{ GX, 2 },{ GX, 3 },{ GX, 4 }, + { GX, 5 },{ GX, 6 },{ GX, 7 },{ GW,11 },{ GW,10 },{ BX, 0 },{ BX, 1 },{ BX, 2 },{ BX, 3 },{ BX, 4 }, + { BX, 5 },{ BX, 6 },{ BX, 7 },{ BW,11 },{ BW,10 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 }, + { NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 }, + { NA, 0 },{ NA, 0 }, + }, + + { // Mode 14 (0x0f) - 16 4 + { M, 0 },{ M, 1 },{ M, 2 },{ M, 3 },{ M, 4 },{ RW, 0 },{ RW, 1 },{ RW, 2 },{ RW, 3 },{ RW, 4 }, + { RW, 5 },{ RW, 6 },{ RW, 7 },{ RW, 8 },{ RW, 9 },{ GW, 0 },{ GW, 1 },{ GW, 2 },{ GW, 3 },{ GW, 4 }, + { GW, 5 },{ GW, 6 },{ GW, 7 },{ GW, 8 },{ GW, 9 },{ BW, 0 },{ BW, 1 },{ BW, 2 },{ BW, 3 },{ BW, 4 }, + { BW, 5 },{ BW, 6 },{ BW, 7 },{ BW, 8 },{ BW, 9 },{ RX, 0 },{ RX, 1 },{ RX, 2 },{ RX, 3 },{ RW,15 }, + { RW,14 },{ RW,13 },{ RW,12 },{ RW,11 },{ RW,10 },{ GX, 0 },{ GX, 1 },{ GX, 2 },{ GX, 3 },{ GW,15 }, + { GW,14 },{ GW,13 },{ GW,12 },{ GW,11 },{ GW,10 },{ BX, 0 },{ BX, 1 },{ BX, 2 },{ BX, 3 },{ BW,15 }, + { BW,14 },{ BW,13 },{ BW,12 },{ BW,11 },{ BW,10 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 }, + { NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 }, + { NA, 0 },{ NA, 0 }, + }, + }; + } + namespace BC7Data { enum AlphaMode @@ -2998,9 +3197,9 @@ void cvtt::Internal::BC6HComputer::Pack(uint32_t flags, const PixelBlockF16* inp const BC7Data::BC6HModeInfo& modeInfo = BC7Data::g_hdrModes[mode]; - BC6H_IO::WriteFunc_t writeFunc = BC6H_IO::g_writeFuncs[mode]; + const BC6HData::ModeDescriptor *desc = BC6HData::g_modeDescriptors[mode]; - const int headerBits = modeInfo.m_partitioned ? 82 : 65; + const size_t headerBits = modeInfo.m_partitioned ? 82 : 65; for (int subset = 0; subset < 2; subset++) { @@ -3017,16 +3216,58 @@ void cvtt::Internal::BC6HComputer::Pack(uint32_t flags, const PixelBlockF16* inp uint16_t modeID = modeInfo.m_modeID; PackingVector pv; + pv.Init(); - { - uint32_t header[3]; - writeFunc(header, modeID, partition, - eps[0][0][0], eps[0][1][0], eps[1][0][0], eps[1][1][0], - eps[0][0][1], eps[0][1][1], eps[1][0][1], eps[1][1][1], - eps[0][0][2], eps[0][1][2], eps[1][0][2], eps[1][1][2] - ); - - pv.InitPacked(header, headerBits); + for (size_t i = 0; i < headerBits; i++) { + int32_t codedValue = 0; + switch (desc[i].m_eField) { + case BC6HData::M: + codedValue = modeID; + break; + case BC6HData::D: + codedValue = partition; + break; + case BC6HData::RW: + codedValue = eps[0][0][0]; + break; + case BC6HData::RX: + codedValue = eps[0][1][0]; + break; + case BC6HData::RY: + codedValue = eps[1][0][0]; + break; + case BC6HData::RZ: + codedValue = eps[1][1][0]; + break; + case BC6HData::GW: + codedValue = eps[0][0][1]; + break; + case BC6HData::GX: + codedValue = eps[0][1][1]; + break; + case BC6HData::GY: + codedValue = eps[1][0][1]; + break; + case BC6HData::GZ: + codedValue = eps[1][1][1]; + break; + case BC6HData::BW: + codedValue = eps[0][0][2]; + break; + case BC6HData::BX: + codedValue = eps[0][1][2]; + break; + case BC6HData::BY: + codedValue = eps[1][0][2]; + break; + case BC6HData::BZ: + codedValue = eps[1][1][2]; + break; + default: + assert(false); + break; + } + pv.Pack(static_cast<uint16_t>((codedValue >> desc[i].m_uBit) & 1), 1); } int fixupIndex1 = 0; @@ -3058,11 +3299,14 @@ void cvtt::Internal::BC6HComputer::SignExtendSingle(int &v, int bits) void cvtt::Internal::BC6HComputer::UnpackOne(PixelBlockF16 &output, const uint8_t *pBC, bool isSigned) { + UnpackingVector pv; + pv.Init(pBC); + int numModeBits = 2; - int modeBits = pBC[0] & 0x3; + int modeBits = pv.Unpack(2); if (modeBits != 0 && modeBits != 1) { - modeBits = pBC[0] & 0x1f; + modeBits |= pv.Unpack(3) << 2; numModeBits += 3; } @@ -3088,10 +3332,10 @@ void cvtt::Internal::BC6HComputer::UnpackOne(PixelBlockF16 &output, const uint8_ } const BC7Data::BC6HModeInfo& modeInfo = BC7Data::g_hdrModes[mode]; - const int headerBits = modeInfo.m_partitioned ? 82 : 65; - const BC6H_IO::ReadFunc_t readFunc = BC6H_IO::g_readFuncs[mode]; + const size_t headerBits = modeInfo.m_partitioned ? 82 : 65; + const BC6HData::ModeDescriptor *desc = BC6HData::g_modeDescriptors[mode]; - uint16_t partition = 0; + int32_t partition = 0; int32_t eps[2][2][3]; for (int subset = 0; subset < 2; subset++) @@ -3099,24 +3343,55 @@ void cvtt::Internal::BC6HComputer::UnpackOne(PixelBlockF16 &output, const uint8_ for (int ch = 0; ch < 3; ch++) eps[subset][epi][ch] = 0; - UnpackingVector pv; - pv.Init(pBC); - - { - uint32_t header[3]; - uint16_t codedEPs[2][2][3]; - pv.UnpackStart(header, headerBits); + for (size_t i = numModeBits; i < headerBits; i++) { + int32_t *pCodedValue = NULL; - readFunc(header, partition, - codedEPs[0][0][0], codedEPs[0][1][0], codedEPs[1][0][0], codedEPs[1][1][0], - codedEPs[0][0][1], codedEPs[0][1][1], codedEPs[1][0][1], codedEPs[1][1][1], - codedEPs[0][0][2], codedEPs[0][1][2], codedEPs[1][0][2], codedEPs[1][1][2] - ); + switch (desc[i].m_eField) { + case BC6HData::D: + pCodedValue = &partition; + break; + case BC6HData::RW: + pCodedValue = &eps[0][0][0]; + break; + case BC6HData::RX: + pCodedValue = &eps[0][1][0]; + break; + case BC6HData::RY: + pCodedValue = &eps[1][0][0]; + break; + case BC6HData::RZ: + pCodedValue = &eps[1][1][0]; + break; + case BC6HData::GW: + pCodedValue = &eps[0][0][1]; + break; + case BC6HData::GX: + pCodedValue = &eps[0][1][1]; + break; + case BC6HData::GY: + pCodedValue = &eps[1][0][1]; + break; + case BC6HData::GZ: + pCodedValue = &eps[1][1][1]; + break; + case BC6HData::BW: + pCodedValue = &eps[0][0][2]; + break; + case BC6HData::BX: + pCodedValue = &eps[0][1][2]; + break; + case BC6HData::BY: + pCodedValue = &eps[1][0][2]; + break; + case BC6HData::BZ: + pCodedValue = &eps[1][1][2]; + break; + default: + assert(false); + break; + } - for (int subset = 0; subset < 2; subset++) - for (int epi = 0; epi < 2; epi++) - for (int ch = 0; ch < 3; ch++) - eps[subset][epi][ch] = codedEPs[subset][epi][ch]; + (*pCodedValue) |= pv.Unpack(1) << desc[i].m_uBit; } uint16_t modeID = modeInfo.m_modeID; diff --git a/thirdparty/cvtt/patches/revert_BC6H_reorg.patch b/thirdparty/cvtt/patches/revert_BC6H_reorg.patch new file mode 100644 index 0000000000..11f9b9db46 --- /dev/null +++ b/thirdparty/cvtt/patches/revert_BC6H_reorg.patch @@ -0,0 +1,393 @@ +diff --git a/thirdparty/cvtt/ConvectionKernels_BC67.cpp b/thirdparty/cvtt/ConvectionKernels_BC67.cpp +index 021d658c08..011f51570b 100644 +--- a/thirdparty/cvtt/ConvectionKernels_BC67.cpp ++++ b/thirdparty/cvtt/ConvectionKernels_BC67.cpp +@@ -76,6 +76,205 @@ namespace cvtt + }; + } + ++ namespace BC6HData ++ { ++ enum EField ++ { ++ NA, // N/A ++ M, // Mode ++ D, // Shape ++ RW, ++ RX, ++ RY, ++ RZ, ++ GW, ++ GX, ++ GY, ++ GZ, ++ BW, ++ BX, ++ BY, ++ BZ, ++ }; ++ ++ struct ModeDescriptor ++ { ++ EField m_eField; ++ uint8_t m_uBit; ++ }; ++ ++ const ModeDescriptor g_modeDescriptors[14][82] = ++ { ++ { // Mode 1 (0x00) - 10 5 5 5 ++ { M, 0 },{ M, 1 },{ GY, 4 },{ BY, 4 },{ BZ, 4 },{ RW, 0 },{ RW, 1 },{ RW, 2 },{ RW, 3 },{ RW, 4 }, ++ { RW, 5 },{ RW, 6 },{ RW, 7 },{ RW, 8 },{ RW, 9 },{ GW, 0 },{ GW, 1 },{ GW, 2 },{ GW, 3 },{ GW, 4 }, ++ { GW, 5 },{ GW, 6 },{ GW, 7 },{ GW, 8 },{ GW, 9 },{ BW, 0 },{ BW, 1 },{ BW, 2 },{ BW, 3 },{ BW, 4 }, ++ { BW, 5 },{ BW, 6 },{ BW, 7 },{ BW, 8 },{ BW, 9 },{ RX, 0 },{ RX, 1 },{ RX, 2 },{ RX, 3 },{ RX, 4 }, ++ { GZ, 4 },{ GY, 0 },{ GY, 1 },{ GY, 2 },{ GY, 3 },{ GX, 0 },{ GX, 1 },{ GX, 2 },{ GX, 3 },{ GX, 4 }, ++ { BZ, 0 },{ GZ, 0 },{ GZ, 1 },{ GZ, 2 },{ GZ, 3 },{ BX, 0 },{ BX, 1 },{ BX, 2 },{ BX, 3 },{ BX, 4 }, ++ { BZ, 1 },{ BY, 0 },{ BY, 1 },{ BY, 2 },{ BY, 3 },{ RY, 0 },{ RY, 1 },{ RY, 2 },{ RY, 3 },{ RY, 4 }, ++ { BZ, 2 },{ RZ, 0 },{ RZ, 1 },{ RZ, 2 },{ RZ, 3 },{ RZ, 4 },{ BZ, 3 },{ D, 0 },{ D, 1 },{ D, 2 }, ++ { D, 3 },{ D, 4 }, ++ }, ++ ++ { // Mode 2 (0x01) - 7 6 6 6 ++ { M, 0 },{ M, 1 },{ GY, 5 },{ GZ, 4 },{ GZ, 5 },{ RW, 0 },{ RW, 1 },{ RW, 2 },{ RW, 3 },{ RW, 4 }, ++ { RW, 5 },{ RW, 6 },{ BZ, 0 },{ BZ, 1 },{ BY, 4 },{ GW, 0 },{ GW, 1 },{ GW, 2 },{ GW, 3 },{ GW, 4 }, ++ { GW, 5 },{ GW, 6 },{ BY, 5 },{ BZ, 2 },{ GY, 4 },{ BW, 0 },{ BW, 1 },{ BW, 2 },{ BW, 3 },{ BW, 4 }, ++ { BW, 5 },{ BW, 6 },{ BZ, 3 },{ BZ, 5 },{ BZ, 4 },{ RX, 0 },{ RX, 1 },{ RX, 2 },{ RX, 3 },{ RX, 4 }, ++ { RX, 5 },{ GY, 0 },{ GY, 1 },{ GY, 2 },{ GY, 3 },{ GX, 0 },{ GX, 1 },{ GX, 2 },{ GX, 3 },{ GX, 4 }, ++ { GX, 5 },{ GZ, 0 },{ GZ, 1 },{ GZ, 2 },{ GZ, 3 },{ BX, 0 },{ BX, 1 },{ BX, 2 },{ BX, 3 },{ BX, 4 }, ++ { BX, 5 },{ BY, 0 },{ BY, 1 },{ BY, 2 },{ BY, 3 },{ RY, 0 },{ RY, 1 },{ RY, 2 },{ RY, 3 },{ RY, 4 }, ++ { RY, 5 },{ RZ, 0 },{ RZ, 1 },{ RZ, 2 },{ RZ, 3 },{ RZ, 4 },{ RZ, 5 },{ D, 0 },{ D, 1 },{ D, 2 }, ++ { D, 3 },{ D, 4 }, ++ }, ++ ++ { // Mode 3 (0x02) - 11 5 4 4 ++ { M, 0 },{ M, 1 },{ M, 2 },{ M, 3 },{ M, 4 },{ RW, 0 },{ RW, 1 },{ RW, 2 },{ RW, 3 },{ RW, 4 }, ++ { RW, 5 },{ RW, 6 },{ RW, 7 },{ RW, 8 },{ RW, 9 },{ GW, 0 },{ GW, 1 },{ GW, 2 },{ GW, 3 },{ GW, 4 }, ++ { GW, 5 },{ GW, 6 },{ GW, 7 },{ GW, 8 },{ GW, 9 },{ BW, 0 },{ BW, 1 },{ BW, 2 },{ BW, 3 },{ BW, 4 }, ++ { BW, 5 },{ BW, 6 },{ BW, 7 },{ BW, 8 },{ BW, 9 },{ RX, 0 },{ RX, 1 },{ RX, 2 },{ RX, 3 },{ RX, 4 }, ++ { RW,10 },{ GY, 0 },{ GY, 1 },{ GY, 2 },{ GY, 3 },{ GX, 0 },{ GX, 1 },{ GX, 2 },{ GX, 3 },{ GW,10 }, ++ { BZ, 0 },{ GZ, 0 },{ GZ, 1 },{ GZ, 2 },{ GZ, 3 },{ BX, 0 },{ BX, 1 },{ BX, 2 },{ BX, 3 },{ BW,10 }, ++ { BZ, 1 },{ BY, 0 },{ BY, 1 },{ BY, 2 },{ BY, 3 },{ RY, 0 },{ RY, 1 },{ RY, 2 },{ RY, 3 },{ RY, 4 }, ++ { BZ, 2 },{ RZ, 0 },{ RZ, 1 },{ RZ, 2 },{ RZ, 3 },{ RZ, 4 },{ BZ, 3 },{ D, 0 },{ D, 1 },{ D, 2 }, ++ { D, 3 },{ D, 4 }, ++ }, ++ ++ { // Mode 4 (0x06) - 11 4 5 4 ++ { M, 0 },{ M, 1 },{ M, 2 },{ M, 3 },{ M, 4 },{ RW, 0 },{ RW, 1 },{ RW, 2 },{ RW, 3 },{ RW, 4 }, ++ { RW, 5 },{ RW, 6 },{ RW, 7 },{ RW, 8 },{ RW, 9 },{ GW, 0 },{ GW, 1 },{ GW, 2 },{ GW, 3 },{ GW, 4 }, ++ { GW, 5 },{ GW, 6 },{ GW, 7 },{ GW, 8 },{ GW, 9 },{ BW, 0 },{ BW, 1 },{ BW, 2 },{ BW, 3 },{ BW, 4 }, ++ { BW, 5 },{ BW, 6 },{ BW, 7 },{ BW, 8 },{ BW, 9 },{ RX, 0 },{ RX, 1 },{ RX, 2 },{ RX, 3 },{ RW,10 }, ++ { GZ, 4 },{ GY, 0 },{ GY, 1 },{ GY, 2 },{ GY, 3 },{ GX, 0 },{ GX, 1 },{ GX, 2 },{ GX, 3 },{ GX, 4 }, ++ { GW,10 },{ GZ, 0 },{ GZ, 1 },{ GZ, 2 },{ GZ, 3 },{ BX, 0 },{ BX, 1 },{ BX, 2 },{ BX, 3 },{ BW,10 }, ++ { BZ, 1 },{ BY, 0 },{ BY, 1 },{ BY, 2 },{ BY, 3 },{ RY, 0 },{ RY, 1 },{ RY, 2 },{ RY, 3 },{ BZ, 0 }, ++ { BZ, 2 },{ RZ, 0 },{ RZ, 1 },{ RZ, 2 },{ RZ, 3 },{ GY, 4 },{ BZ, 3 },{ D, 0 },{ D, 1 },{ D, 2 }, ++ { D, 3 },{ D, 4 }, ++ }, ++ ++ { // Mode 5 (0x0a) - 11 4 4 5 ++ { M, 0 },{ M, 1 },{ M, 2 },{ M, 3 },{ M, 4 },{ RW, 0 },{ RW, 1 },{ RW, 2 },{ RW, 3 },{ RW, 4 }, ++ { RW, 5 },{ RW, 6 },{ RW, 7 },{ RW, 8 },{ RW, 9 },{ GW, 0 },{ GW, 1 },{ GW, 2 },{ GW, 3 },{ GW, 4 }, ++ { GW, 5 },{ GW, 6 },{ GW, 7 },{ GW, 8 },{ GW, 9 },{ BW, 0 },{ BW, 1 },{ BW, 2 },{ BW, 3 },{ BW, 4 }, ++ { BW, 5 },{ BW, 6 },{ BW, 7 },{ BW, 8 },{ BW, 9 },{ RX, 0 },{ RX, 1 },{ RX, 2 },{ RX, 3 },{ RW,10 }, ++ { BY, 4 },{ GY, 0 },{ GY, 1 },{ GY, 2 },{ GY, 3 },{ GX, 0 },{ GX, 1 },{ GX, 2 },{ GX, 3 },{ GW,10 }, ++ { BZ, 0 },{ GZ, 0 },{ GZ, 1 },{ GZ, 2 },{ GZ, 3 },{ BX, 0 },{ BX, 1 },{ BX, 2 },{ BX, 3 },{ BX, 4 }, ++ { BW,10 },{ BY, 0 },{ BY, 1 },{ BY, 2 },{ BY, 3 },{ RY, 0 },{ RY, 1 },{ RY, 2 },{ RY, 3 },{ BZ, 1 }, ++ { BZ, 2 },{ RZ, 0 },{ RZ, 1 },{ RZ, 2 },{ RZ, 3 },{ BZ, 4 },{ BZ, 3 },{ D, 0 },{ D, 1 },{ D, 2 }, ++ { D, 3 },{ D, 4 }, ++ }, ++ ++ { // Mode 6 (0x0e) - 9 5 5 5 ++ { M, 0 },{ M, 1 },{ M, 2 },{ M, 3 },{ M, 4 },{ RW, 0 },{ RW, 1 },{ RW, 2 },{ RW, 3 },{ RW, 4 }, ++ { RW, 5 },{ RW, 6 },{ RW, 7 },{ RW, 8 },{ BY, 4 },{ GW, 0 },{ GW, 1 },{ GW, 2 },{ GW, 3 },{ GW, 4 }, ++ { GW, 5 },{ GW, 6 },{ GW, 7 },{ GW, 8 },{ GY, 4 },{ BW, 0 },{ BW, 1 },{ BW, 2 },{ BW, 3 },{ BW, 4 }, ++ { BW, 5 },{ BW, 6 },{ BW, 7 },{ BW, 8 },{ BZ, 4 },{ RX, 0 },{ RX, 1 },{ RX, 2 },{ RX, 3 },{ RX, 4 }, ++ { GZ, 4 },{ GY, 0 },{ GY, 1 },{ GY, 2 },{ GY, 3 },{ GX, 0 },{ GX, 1 },{ GX, 2 },{ GX, 3 },{ GX, 4 }, ++ { BZ, 0 },{ GZ, 0 },{ GZ, 1 },{ GZ, 2 },{ GZ, 3 },{ BX, 0 },{ BX, 1 },{ BX, 2 },{ BX, 3 },{ BX, 4 }, ++ { BZ, 1 },{ BY, 0 },{ BY, 1 },{ BY, 2 },{ BY, 3 },{ RY, 0 },{ RY, 1 },{ RY, 2 },{ RY, 3 },{ RY, 4 }, ++ { BZ, 2 },{ RZ, 0 },{ RZ, 1 },{ RZ, 2 },{ RZ, 3 },{ RZ, 4 },{ BZ, 3 },{ D, 0 },{ D, 1 },{ D, 2 }, ++ { D, 3 },{ D, 4 }, ++ }, ++ ++ { // Mode 7 (0x12) - 8 6 5 5 ++ { M, 0 },{ M, 1 },{ M, 2 },{ M, 3 },{ M, 4 },{ RW, 0 },{ RW, 1 },{ RW, 2 },{ RW, 3 },{ RW, 4 }, ++ { RW, 5 },{ RW, 6 },{ RW, 7 },{ GZ, 4 },{ BY, 4 },{ GW, 0 },{ GW, 1 },{ GW, 2 },{ GW, 3 },{ GW, 4 }, ++ { GW, 5 },{ GW, 6 },{ GW, 7 },{ BZ, 2 },{ GY, 4 },{ BW, 0 },{ BW, 1 },{ BW, 2 },{ BW, 3 },{ BW, 4 }, ++ { BW, 5 },{ BW, 6 },{ BW, 7 },{ BZ, 3 },{ BZ, 4 },{ RX, 0 },{ RX, 1 },{ RX, 2 },{ RX, 3 },{ RX, 4 }, ++ { RX, 5 },{ GY, 0 },{ GY, 1 },{ GY, 2 },{ GY, 3 },{ GX, 0 },{ GX, 1 },{ GX, 2 },{ GX, 3 },{ GX, 4 }, ++ { BZ, 0 },{ GZ, 0 },{ GZ, 1 },{ GZ, 2 },{ GZ, 3 },{ BX, 0 },{ BX, 1 },{ BX, 2 },{ BX, 3 },{ BX, 4 }, ++ { BZ, 1 },{ BY, 0 },{ BY, 1 },{ BY, 2 },{ BY, 3 },{ RY, 0 },{ RY, 1 },{ RY, 2 },{ RY, 3 },{ RY, 4 }, ++ { RY, 5 },{ RZ, 0 },{ RZ, 1 },{ RZ, 2 },{ RZ, 3 },{ RZ, 4 },{ RZ, 5 },{ D, 0 },{ D, 1 },{ D, 2 }, ++ { D, 3 },{ D, 4 }, ++ }, ++ ++ { // Mode 8 (0x16) - 8 5 6 5 ++ { M, 0 },{ M, 1 },{ M, 2 },{ M, 3 },{ M, 4 },{ RW, 0 },{ RW, 1 },{ RW, 2 },{ RW, 3 },{ RW, 4 }, ++ { RW, 5 },{ RW, 6 },{ RW, 7 },{ BZ, 0 },{ BY, 4 },{ GW, 0 },{ GW, 1 },{ GW, 2 },{ GW, 3 },{ GW, 4 }, ++ { GW, 5 },{ GW, 6 },{ GW, 7 },{ GY, 5 },{ GY, 4 },{ BW, 0 },{ BW, 1 },{ BW, 2 },{ BW, 3 },{ BW, 4 }, ++ { BW, 5 },{ BW, 6 },{ BW, 7 },{ GZ, 5 },{ BZ, 4 },{ RX, 0 },{ RX, 1 },{ RX, 2 },{ RX, 3 },{ RX, 4 }, ++ { GZ, 4 },{ GY, 0 },{ GY, 1 },{ GY, 2 },{ GY, 3 },{ GX, 0 },{ GX, 1 },{ GX, 2 },{ GX, 3 },{ GX, 4 }, ++ { GX, 5 },{ GZ, 0 },{ GZ, 1 },{ GZ, 2 },{ GZ, 3 },{ BX, 0 },{ BX, 1 },{ BX, 2 },{ BX, 3 },{ BX, 4 }, ++ { BZ, 1 },{ BY, 0 },{ BY, 1 },{ BY, 2 },{ BY, 3 },{ RY, 0 },{ RY, 1 },{ RY, 2 },{ RY, 3 },{ RY, 4 }, ++ { BZ, 2 },{ RZ, 0 },{ RZ, 1 },{ RZ, 2 },{ RZ, 3 },{ RZ, 4 },{ BZ, 3 },{ D, 0 },{ D, 1 },{ D, 2 }, ++ { D, 3 },{ D, 4 }, ++ }, ++ ++ { // Mode 9 (0x1a) - 8 5 5 6 ++ { M, 0 },{ M, 1 },{ M, 2 },{ M, 3 },{ M, 4 },{ RW, 0 },{ RW, 1 },{ RW, 2 },{ RW, 3 },{ RW, 4 }, ++ { RW, 5 },{ RW, 6 },{ RW, 7 },{ BZ, 1 },{ BY, 4 },{ GW, 0 },{ GW, 1 },{ GW, 2 },{ GW, 3 },{ GW, 4 }, ++ { GW, 5 },{ GW, 6 },{ GW, 7 },{ BY, 5 },{ GY, 4 },{ BW, 0 },{ BW, 1 },{ BW, 2 },{ BW, 3 },{ BW, 4 }, ++ { BW, 5 },{ BW, 6 },{ BW, 7 },{ BZ, 5 },{ BZ, 4 },{ RX, 0 },{ RX, 1 },{ RX, 2 },{ RX, 3 },{ RX, 4 }, ++ { GZ, 4 },{ GY, 0 },{ GY, 1 },{ GY, 2 },{ GY, 3 },{ GX, 0 },{ GX, 1 },{ GX, 2 },{ GX, 3 },{ GX, 4 }, ++ { BZ, 0 },{ GZ, 0 },{ GZ, 1 },{ GZ, 2 },{ GZ, 3 },{ BX, 0 },{ BX, 1 },{ BX, 2 },{ BX, 3 },{ BX, 4 }, ++ { BX, 5 },{ BY, 0 },{ BY, 1 },{ BY, 2 },{ BY, 3 },{ RY, 0 },{ RY, 1 },{ RY, 2 },{ RY, 3 },{ RY, 4 }, ++ { BZ, 2 },{ RZ, 0 },{ RZ, 1 },{ RZ, 2 },{ RZ, 3 },{ RZ, 4 },{ BZ, 3 },{ D, 0 },{ D, 1 },{ D, 2 }, ++ { D, 3 },{ D, 4 }, ++ }, ++ ++ { // Mode 10 (0x1e) - 6 6 6 6 ++ { M, 0 },{ M, 1 },{ M, 2 },{ M, 3 },{ M, 4 },{ RW, 0 },{ RW, 1 },{ RW, 2 },{ RW, 3 },{ RW, 4 }, ++ { RW, 5 },{ GZ, 4 },{ BZ, 0 },{ BZ, 1 },{ BY, 4 },{ GW, 0 },{ GW, 1 },{ GW, 2 },{ GW, 3 },{ GW, 4 }, ++ { GW, 5 },{ GY, 5 },{ BY, 5 },{ BZ, 2 },{ GY, 4 },{ BW, 0 },{ BW, 1 },{ BW, 2 },{ BW, 3 },{ BW, 4 }, ++ { BW, 5 },{ GZ, 5 },{ BZ, 3 },{ BZ, 5 },{ BZ, 4 },{ RX, 0 },{ RX, 1 },{ RX, 2 },{ RX, 3 },{ RX, 4 }, ++ { RX, 5 },{ GY, 0 },{ GY, 1 },{ GY, 2 },{ GY, 3 },{ GX, 0 },{ GX, 1 },{ GX, 2 },{ GX, 3 },{ GX, 4 }, ++ { GX, 5 },{ GZ, 0 },{ GZ, 1 },{ GZ, 2 },{ GZ, 3 },{ BX, 0 },{ BX, 1 },{ BX, 2 },{ BX, 3 },{ BX, 4 }, ++ { BX, 5 },{ BY, 0 },{ BY, 1 },{ BY, 2 },{ BY, 3 },{ RY, 0 },{ RY, 1 },{ RY, 2 },{ RY, 3 },{ RY, 4 }, ++ { RY, 5 },{ RZ, 0 },{ RZ, 1 },{ RZ, 2 },{ RZ, 3 },{ RZ, 4 },{ RZ, 5 },{ D, 0 },{ D, 1 },{ D, 2 }, ++ { D, 3 },{ D, 4 }, ++ }, ++ ++ { // Mode 11 (0x03) - 10 10 ++ { M, 0 },{ M, 1 },{ M, 2 },{ M, 3 },{ M, 4 },{ RW, 0 },{ RW, 1 },{ RW, 2 },{ RW, 3 },{ RW, 4 }, ++ { RW, 5 },{ RW, 6 },{ RW, 7 },{ RW, 8 },{ RW, 9 },{ GW, 0 },{ GW, 1 },{ GW, 2 },{ GW, 3 },{ GW, 4 }, ++ { GW, 5 },{ GW, 6 },{ GW, 7 },{ GW, 8 },{ GW, 9 },{ BW, 0 },{ BW, 1 },{ BW, 2 },{ BW, 3 },{ BW, 4 }, ++ { BW, 5 },{ BW, 6 },{ BW, 7 },{ BW, 8 },{ BW, 9 },{ RX, 0 },{ RX, 1 },{ RX, 2 },{ RX, 3 },{ RX, 4 }, ++ { RX, 5 },{ RX, 6 },{ RX, 7 },{ RX, 8 },{ RX, 9 },{ GX, 0 },{ GX, 1 },{ GX, 2 },{ GX, 3 },{ GX, 4 }, ++ { GX, 5 },{ GX, 6 },{ GX, 7 },{ GX, 8 },{ GX, 9 },{ BX, 0 },{ BX, 1 },{ BX, 2 },{ BX, 3 },{ BX, 4 }, ++ { BX, 5 },{ BX, 6 },{ BX, 7 },{ BX, 8 },{ BX, 9 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 }, ++ { NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 }, ++ { NA, 0 },{ NA, 0 }, ++ }, ++ ++ { // Mode 12 (0x07) - 11 9 ++ { M, 0 },{ M, 1 },{ M, 2 },{ M, 3 },{ M, 4 },{ RW, 0 },{ RW, 1 },{ RW, 2 },{ RW, 3 },{ RW, 4 }, ++ { RW, 5 },{ RW, 6 },{ RW, 7 },{ RW, 8 },{ RW, 9 },{ GW, 0 },{ GW, 1 },{ GW, 2 },{ GW, 3 },{ GW, 4 }, ++ { GW, 5 },{ GW, 6 },{ GW, 7 },{ GW, 8 },{ GW, 9 },{ BW, 0 },{ BW, 1 },{ BW, 2 },{ BW, 3 },{ BW, 4 }, ++ { BW, 5 },{ BW, 6 },{ BW, 7 },{ BW, 8 },{ BW, 9 },{ RX, 0 },{ RX, 1 },{ RX, 2 },{ RX, 3 },{ RX, 4 }, ++ { RX, 5 },{ RX, 6 },{ RX, 7 },{ RX, 8 },{ RW,10 },{ GX, 0 },{ GX, 1 },{ GX, 2 },{ GX, 3 },{ GX, 4 }, ++ { GX, 5 },{ GX, 6 },{ GX, 7 },{ GX, 8 },{ GW,10 },{ BX, 0 },{ BX, 1 },{ BX, 2 },{ BX, 3 },{ BX, 4 }, ++ { BX, 5 },{ BX, 6 },{ BX, 7 },{ BX, 8 },{ BW,10 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 }, ++ { NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 }, ++ { NA, 0 },{ NA, 0 }, ++ }, ++ ++ { // Mode 13 (0x0b) - 12 8 ++ { M, 0 },{ M, 1 },{ M, 2 },{ M, 3 },{ M, 4 },{ RW, 0 },{ RW, 1 },{ RW, 2 },{ RW, 3 },{ RW, 4 }, ++ { RW, 5 },{ RW, 6 },{ RW, 7 },{ RW, 8 },{ RW, 9 },{ GW, 0 },{ GW, 1 },{ GW, 2 },{ GW, 3 },{ GW, 4 }, ++ { GW, 5 },{ GW, 6 },{ GW, 7 },{ GW, 8 },{ GW, 9 },{ BW, 0 },{ BW, 1 },{ BW, 2 },{ BW, 3 },{ BW, 4 }, ++ { BW, 5 },{ BW, 6 },{ BW, 7 },{ BW, 8 },{ BW, 9 },{ RX, 0 },{ RX, 1 },{ RX, 2 },{ RX, 3 },{ RX, 4 }, ++ { RX, 5 },{ RX, 6 },{ RX, 7 },{ RW,11 },{ RW,10 },{ GX, 0 },{ GX, 1 },{ GX, 2 },{ GX, 3 },{ GX, 4 }, ++ { GX, 5 },{ GX, 6 },{ GX, 7 },{ GW,11 },{ GW,10 },{ BX, 0 },{ BX, 1 },{ BX, 2 },{ BX, 3 },{ BX, 4 }, ++ { BX, 5 },{ BX, 6 },{ BX, 7 },{ BW,11 },{ BW,10 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 }, ++ { NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 }, ++ { NA, 0 },{ NA, 0 }, ++ }, ++ ++ { // Mode 14 (0x0f) - 16 4 ++ { M, 0 },{ M, 1 },{ M, 2 },{ M, 3 },{ M, 4 },{ RW, 0 },{ RW, 1 },{ RW, 2 },{ RW, 3 },{ RW, 4 }, ++ { RW, 5 },{ RW, 6 },{ RW, 7 },{ RW, 8 },{ RW, 9 },{ GW, 0 },{ GW, 1 },{ GW, 2 },{ GW, 3 },{ GW, 4 }, ++ { GW, 5 },{ GW, 6 },{ GW, 7 },{ GW, 8 },{ GW, 9 },{ BW, 0 },{ BW, 1 },{ BW, 2 },{ BW, 3 },{ BW, 4 }, ++ { BW, 5 },{ BW, 6 },{ BW, 7 },{ BW, 8 },{ BW, 9 },{ RX, 0 },{ RX, 1 },{ RX, 2 },{ RX, 3 },{ RW,15 }, ++ { RW,14 },{ RW,13 },{ RW,12 },{ RW,11 },{ RW,10 },{ GX, 0 },{ GX, 1 },{ GX, 2 },{ GX, 3 },{ GW,15 }, ++ { GW,14 },{ GW,13 },{ GW,12 },{ GW,11 },{ GW,10 },{ BX, 0 },{ BX, 1 },{ BX, 2 },{ BX, 3 },{ BW,15 }, ++ { BW,14 },{ BW,13 },{ BW,12 },{ BW,11 },{ BW,10 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 }, ++ { NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 },{ NA, 0 }, ++ { NA, 0 },{ NA, 0 }, ++ }, ++ }; ++ } ++ + namespace BC7Data + { + enum AlphaMode +@@ -2998,9 +3197,9 @@ void cvtt::Internal::BC6HComputer::Pack(uint32_t flags, const PixelBlockF16* inp + + const BC7Data::BC6HModeInfo& modeInfo = BC7Data::g_hdrModes[mode]; + +- BC6H_IO::WriteFunc_t writeFunc = BC6H_IO::g_writeFuncs[mode]; ++ const BC6HData::ModeDescriptor *desc = BC6HData::g_modeDescriptors[mode]; + +- const int headerBits = modeInfo.m_partitioned ? 82 : 65; ++ const size_t headerBits = modeInfo.m_partitioned ? 82 : 65; + + for (int subset = 0; subset < 2; subset++) + { +@@ -3017,16 +3216,58 @@ void cvtt::Internal::BC6HComputer::Pack(uint32_t flags, const PixelBlockF16* inp + uint16_t modeID = modeInfo.m_modeID; + + PackingVector pv; ++ pv.Init(); + +- { +- uint32_t header[3]; +- writeFunc(header, modeID, partition, +- eps[0][0][0], eps[0][1][0], eps[1][0][0], eps[1][1][0], +- eps[0][0][1], eps[0][1][1], eps[1][0][1], eps[1][1][1], +- eps[0][0][2], eps[0][1][2], eps[1][0][2], eps[1][1][2] +- ); +- +- pv.InitPacked(header, headerBits); ++ for (size_t i = 0; i < headerBits; i++) { ++ int32_t codedValue = 0; ++ switch (desc[i].m_eField) { ++ case BC6HData::M: ++ codedValue = modeID; ++ break; ++ case BC6HData::D: ++ codedValue = partition; ++ break; ++ case BC6HData::RW: ++ codedValue = eps[0][0][0]; ++ break; ++ case BC6HData::RX: ++ codedValue = eps[0][1][0]; ++ break; ++ case BC6HData::RY: ++ codedValue = eps[1][0][0]; ++ break; ++ case BC6HData::RZ: ++ codedValue = eps[1][1][0]; ++ break; ++ case BC6HData::GW: ++ codedValue = eps[0][0][1]; ++ break; ++ case BC6HData::GX: ++ codedValue = eps[0][1][1]; ++ break; ++ case BC6HData::GY: ++ codedValue = eps[1][0][1]; ++ break; ++ case BC6HData::GZ: ++ codedValue = eps[1][1][1]; ++ break; ++ case BC6HData::BW: ++ codedValue = eps[0][0][2]; ++ break; ++ case BC6HData::BX: ++ codedValue = eps[0][1][2]; ++ break; ++ case BC6HData::BY: ++ codedValue = eps[1][0][2]; ++ break; ++ case BC6HData::BZ: ++ codedValue = eps[1][1][2]; ++ break; ++ default: ++ assert(false); ++ break; ++ } ++ pv.Pack(static_cast<uint16_t>((codedValue >> desc[i].m_uBit) & 1), 1); + } + + int fixupIndex1 = 0; +@@ -3058,11 +3299,14 @@ void cvtt::Internal::BC6HComputer::SignExtendSingle(int &v, int bits) + + void cvtt::Internal::BC6HComputer::UnpackOne(PixelBlockF16 &output, const uint8_t *pBC, bool isSigned) + { ++ UnpackingVector pv; ++ pv.Init(pBC); ++ + int numModeBits = 2; +- int modeBits = pBC[0] & 0x3; ++ int modeBits = pv.Unpack(2); + if (modeBits != 0 && modeBits != 1) + { +- modeBits = pBC[0] & 0x1f; ++ modeBits |= pv.Unpack(3) << 2; + numModeBits += 3; + } + +@@ -3088,10 +3332,10 @@ void cvtt::Internal::BC6HComputer::UnpackOne(PixelBlockF16 &output, const uint8_ + } + + const BC7Data::BC6HModeInfo& modeInfo = BC7Data::g_hdrModes[mode]; +- const int headerBits = modeInfo.m_partitioned ? 82 : 65; +- const BC6H_IO::ReadFunc_t readFunc = BC6H_IO::g_readFuncs[mode]; ++ const size_t headerBits = modeInfo.m_partitioned ? 82 : 65; ++ const BC6HData::ModeDescriptor *desc = BC6HData::g_modeDescriptors[mode]; + +- uint16_t partition = 0; ++ int32_t partition = 0; + int32_t eps[2][2][3]; + + for (int subset = 0; subset < 2; subset++) +@@ -3099,24 +3343,55 @@ void cvtt::Internal::BC6HComputer::UnpackOne(PixelBlockF16 &output, const uint8_ + for (int ch = 0; ch < 3; ch++) + eps[subset][epi][ch] = 0; + +- UnpackingVector pv; +- pv.Init(pBC); +- +- { +- uint32_t header[3]; +- uint16_t codedEPs[2][2][3]; +- pv.UnpackStart(header, headerBits); ++ for (size_t i = numModeBits; i < headerBits; i++) { ++ int32_t *pCodedValue = NULL; + +- readFunc(header, partition, +- codedEPs[0][0][0], codedEPs[0][1][0], codedEPs[1][0][0], codedEPs[1][1][0], +- codedEPs[0][0][1], codedEPs[0][1][1], codedEPs[1][0][1], codedEPs[1][1][1], +- codedEPs[0][0][2], codedEPs[0][1][2], codedEPs[1][0][2], codedEPs[1][1][2] +- ); ++ switch (desc[i].m_eField) { ++ case BC6HData::D: ++ pCodedValue = &partition; ++ break; ++ case BC6HData::RW: ++ pCodedValue = &eps[0][0][0]; ++ break; ++ case BC6HData::RX: ++ pCodedValue = &eps[0][1][0]; ++ break; ++ case BC6HData::RY: ++ pCodedValue = &eps[1][0][0]; ++ break; ++ case BC6HData::RZ: ++ pCodedValue = &eps[1][1][0]; ++ break; ++ case BC6HData::GW: ++ pCodedValue = &eps[0][0][1]; ++ break; ++ case BC6HData::GX: ++ pCodedValue = &eps[0][1][1]; ++ break; ++ case BC6HData::GY: ++ pCodedValue = &eps[1][0][1]; ++ break; ++ case BC6HData::GZ: ++ pCodedValue = &eps[1][1][1]; ++ break; ++ case BC6HData::BW: ++ pCodedValue = &eps[0][0][2]; ++ break; ++ case BC6HData::BX: ++ pCodedValue = &eps[0][1][2]; ++ break; ++ case BC6HData::BY: ++ pCodedValue = &eps[1][0][2]; ++ break; ++ case BC6HData::BZ: ++ pCodedValue = &eps[1][1][2]; ++ break; ++ default: ++ assert(false); ++ break; ++ } + +- for (int subset = 0; subset < 2; subset++) +- for (int epi = 0; epi < 2; epi++) +- for (int ch = 0; ch < 3; ch++) +- eps[subset][epi][ch] = codedEPs[subset][epi][ch]; ++ (*pCodedValue) |= pv.Unpack(1) << desc[i].m_uBit; + } + + uint16_t modeID = modeInfo.m_modeID; |