signed 8-bit integer support #3507
Conversation
|
Hey, things look good but I am getting the same failures you are. Do you need us to fix these or do you think you can handle this? |
|
I am looking into it, but if you could look into the remaining failing tests, that would be very much appreciated! The last 3 are all related to bilinear interpolation, but I'm not sure about the RichardsonLucyOnGrayscale test. |
include/af/defines.h
Outdated
| @@ -227,6 +227,7 @@ typedef enum { | |||
| #if AF_API_VERSION >= 37 | |||
| , f16 ///< 16-bit floating point value | |||
| #endif | |||
| , s8 ///< 8-bit signed integral value /// TODO AF_API_VERSION | |||
There was a problem hiding this comment.
| , s8 ///< 8-bit signed integral value /// TODO AF_API_VERSION | |
| #if AF_APU_VERSION >= 310 | |
| , s8 ///< 8-bit signed integral value /// TODO AF_API_VERSION | |
| #endif |
There was a problem hiding this comment.
With the current system, if I use 310 API versions 4.0+ will need to be set to 400, not 40. This is not really a problem currently, but maybe needs to be looked at in the future?
include/af/traits.hpp
Outdated
|
|
||
| template<> | ||
| struct dtype_traits<signed char> { | ||
| enum { | ||
| af_type = s8 , | ||
| ctype = f32 | ||
| }; | ||
| typedef signed char base_type; | ||
| static const char* getName() { return "schar"; } | ||
| }; |
There was a problem hiding this comment.
| template<> | |
| struct dtype_traits<signed char> { | |
| enum { | |
| af_type = s8 , | |
| ctype = f32 | |
| }; | |
| typedef signed char base_type; | |
| static const char* getName() { return "schar"; } | |
| }; | |
| #if AF_API_VERSION >= 310 | |
| template<> | |
| struct dtype_traits<signed char> { | |
| enum { | |
| af_type = s8 , | |
| ctype = f32 | |
| }; | |
| typedef signed char base_type; | |
| static const char* getName() { return "schar"; } | |
| }; | |
| #endif |
There was a problem hiding this comment.
Public interface changes need to be surrounded with these ifdefs. This includes anything in the include folder but not in the src folder.
Image loading with a shift causes the test data to contain negative values. This test passes when the data is limited to 0-127 instead.
This test data cannot be trivially shifted since rotate sets out-of-bounds values to 0, not -128. Limiting the range to 0-127 mostly works, but introduces rounding errors between output and gold.
This also slightly extends the interface macros.
verstatx
force-pushed
the
int8
branch
2 times, most recently
from
9ff74ea
to
94e7674
Compare
verstatx
changed the title
| @@ -1393,6 +1428,7 @@ namespace af | |||
| AFAPI array operator&(const array& lhs, const long long& rhs); | |||
| AFAPI array operator&(const array& lhs, const long& rhs); | |||
| AFAPI array operator&(const array& lhs, const short& rhs); | |||
| AFAPI array operator&(const array& lhs, const signed char& rhs); | |||
There was a problem hiding this comment.
Do the bitwise and logical AND operators need to be protected by the ifdefs? None of the other types have guards.
Description
Adds signed 8-bit integer support to all arrayfire backends. So far all backends build and most tests pass. Failing tests:
AnisotropicDiffusion/4.GradientGrayscaleAnisotropicDiffusion/4.GradientColorImageAnisotropicDiffusion/4.CurvatureGrayscaleAnisotropicDiffusion/4.CurvatureColorImageInverseDeconvolution/1.TikhonovOnGrayscaleIterativeDeconvolution/1.LandweberOnGrayscaleIterativeDeconvolution/1.RichardsonLucyOnGrayscale(skipped)RichardsonLucy doesn't work well with negative values. Test passes when data is 0-127.
RotateLinear/6.*(skipped)Test data for every non-cardinal rotation doesn't work well with s8.
TransformInt/6.PerspectiveBilinearTransformInt/6.PerspectiveBilinearInvertAll of the image loading tests appear to be failing due to unsigned->signed conversion overflow related to image loading.(fixed)Changes to Users
No additional build flags have been added for this feature.
Checklist