branchsbox/gpu-compute-dxtcancel
9 Commits over 0 Days - ∞cph!
Remove timers & debug logs
Remove reference shader
Backport Valve GPU texture compression dispatch code
Compiled DXT shaders (optimized)
Force DXT5 for now, since that's all we have compute support for
Dispatch compute and read back
Texture barriers for Vulkan
Output full timing info
Move behind a convar so I can compare CPU/GPU easily
Staging buffer, barriers, etc. for Vulkan
Time everything (temp, will remove before merge)
Switch compression levels for a big compression speed boost (at the cost of very slightly larger files)
Test case - 4096x4096:
Old: 955ms, on disk 6948kb
New: 151ms, on disk 7253kb
Add cached flag to OnReadPixels, speeds up memory access
Speeds up 4096x4096 texture read from ~92ms to ~5ms in testing
Re-add check for compressed block texel views
Fallback gracefully to CPU on platforms with no compute support
Get rid of some redundant timing info
Update stb_dxt.h from v1.04 to v1.12
Bail from trying to compress using CPU DXT compressor if non-power-of-two (rather than crashing)
Remove forced DXT5 compression
Get rid of some redundant timing info
Update stb_dxt.h from v1.04 to v1.12
Bail from trying to compress using CPU DXT compressor if non-power-of-two (rather than crashing)
Remove forced DXT5 compression
Re-add check for compressed block texel views
Fallback gracefully to CPU on platforms with no compute support
Add cached flag to OnReadPixels, speeds up memory access
Speeds up 4096x4096 texture read from ~92ms to ~5ms in testing
Time everything (temp, will remove before merge)
Switch compression levels for a big compression speed boost (at the cost of very slightly larger files)
Test case - 4096x4096:
Old: 955ms, on disk 6948kb
New: 151ms, on disk 7253kb
Texture barriers for Vulkan
Output full timing info
Move behind a convar so I can compare CPU/GPU easily
Staging buffer, barriers, etc. for Vulkan
Force DXT5 for now, since that's all we have compute support for
Dispatch compute and read back
Backport Valve GPU texture compression dispatch code
Compiled DXT shaders (optimized)
