-
-
Notifications
You must be signed in to change notification settings - Fork 95
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
VkFFT v1.3.2 #141
Merged
VkFFT v1.3.2 #141
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
… ..). Current default is main, which changes often and can lead to non-reproducible builds.
Allow specifying glslang version (e.g. cmake -DGLSLANG_GIT_TAG=13.0.0…
-Added double-double support in VkFFT. Requires cpu initialization in full quad precision, so only supports gcc for now. Potentially possible to add full FP128 support or some other FP128 library (like mpir) in the future. -Data has to be stored in double-double before VkFFT kernels calls (no fp128<->double-double conversion on the GPU yet). -Full 1e-32 precision, but same range as FP64. See Library for Double-Double and Quad-Double Arithmetic by Y Hida for more information on double-double. -Reuqires FMA contraction to be disabled (due to ab-cd contraction rounding mismatch). Doesn't work on Vulkan as I haven't found how to do that yet. -Fixed warnings (#138) -Added proper check for app to be zero before initializeVkFFT call and zeroing on deletion (#134) -Added an option to provide staging buffer in application and VkGPU handle (#129) -Added guards for build type (#128) -Fixed missing deallocation calls for the inverse Bluestein axes. Fixed the buffer layout size in Vulkan in some cases. -Refactored the code generator and container struct layout for better handling complex numbers (-5k loc). -Added more precision tests and benchmarks. -Will be merged in the main branch after more testing and update to the documentation.
…upload (+30% performance)
… from size[0]+2 to size[0] -ref: #139
…and quad double-double for computations
… also present in v1.3.1)
…ry - helps in quad double double precision
…uad double double -Recalculated radix-7 11 and 13 coefficients in higher precision, multithreaded Rader 11 and 13 is better than single threaded in quad double double -Fixed mistake in external FFT library tests not running correctly in multidimensional cases
-Updated documentation
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
-Added double-double support in VkFFT. Requires cpu initialization in full quad precision, so only supports gcc with quadmath dependency for now. Potentially possible to add full FP128 support or some other FP128 library (like mpir) in the future.
-Data has to be stored in double-double before VkFFT kernels calls (no fp128<->double-double conversion on the GPU yet).
-Full 1e-32 precision, but same range as FP64. See Library for Double-Double and Quad-Double Arithmetic by Y Hida for more information on double-double.
-Double-double requires FMA contraction to be disabled (due to ab-cd contraction rounding mismatch). Doesn't work on Vulkan as I haven't found how to do that yet.
-Added DST I-IV support.
-Fixed warnings (#138)
-Added proper check for app to be zero before initializeVkFFT call and zeroing on deletion (#134)
-Added an option to provide a staging buffer in the application and VkGPU handle (#129)
-Added guards for build type (#128)
-Changed default innermost stride for real buffers in out-of-place R2C from size[0]+2 to size[0] (#139)
-Allow specifying glslang version (#135)
-Improved instruction count and accuracy for radix-7.
-Fixed missing deallocation calls for the inverse Bluestein axes. Fixed the buffer layout size in Vulkan in some cases.
-Refactored the code generator and container struct layout for better handling complex numbers (-5k loc).
-Added more precision tests and benchmarks.