feat: add C ndarray implementation for blas/base/dscal
#2915
Conversation
stdlib-bot
added
the
BLAS
kgryte
added
Feature
| x[ i ] *= alpha; | ||
| } | ||
| return x; | ||
| return ndarray( N, alpha, x, stride, 0 ); |
There was a problem hiding this comment.
@aman-095 I believe we want to support negative strides here. That means using stride2offset to compute the offset before calling into the ndarray API.
There was a problem hiding this comment.
But, native dscal implementation from lapack doesn't allow for negative stride.
There was a problem hiding this comment.
That was my point about converting from negative to positive. But here, our ndarray.js file should support negative strides. In C, we should convert the stride before calling CBLAS. Meaning, we are free to support negative strides, but we need to ensure that if and when we call a library following LAPACK that we only pass along positive strides. In short, for all our Level 1 routines, we should be consistent in supporting negative strides. The current situation of some supporting and others not supporting is just confusing.
There was a problem hiding this comment.
It means that for our Js implementation, we should allow negative strides for all BLAS Level 1 and only for CBLAS, or other hardware-optimized libraries we use the same convention of stride conversion.
There was a problem hiding this comment.
Correct. We need to update our JS implementations to support negative strides.
| X[ i ] *= alpha; | ||
| } | ||
| return; | ||
| API_SUFFIX(c_dscal_ndarray)( N, alpha, X, stride, 0 ); |
There was a problem hiding this comment.
@aman-095 Similarly, we should go ahead and support negative strides here.
| * @param offset starting index | ||
| */ | ||
| void API_SUFFIX(c_dscal_ndarray)( const CBLAS_INT N, const double alpha, double *X, const CBLAS_INT stride, const CBLAS_INT offset ) { | ||
| X += stdlib_strided_min_view_buffer_index( N, stride, offset ); // adjust array pointer |
There was a problem hiding this comment.
@aman-095 Before L47, we should convert stride to a positive stride before calling into CBLAS, which may or may not support negative strides. By always passing in a positive stride, we avoid backward compat breaks.
| @@ -31,3 +32,17 @@ | |||
| void API_SUFFIX(c_dscal)( const CBLAS_INT N, const double alpha, double *X, const CBLAS_INT stride ) { | |||
| API_SUFFIX(cblas_dscal)( N, alpha, X, stride ); | |||
There was a problem hiding this comment.
@aman-095 We should also support negative strides here.
| @@ -31,3 +32,21 @@ | |||
| void API_SUFFIX(c_dscal)( const CBLAS_INT N, const double alpha, double *X, const CBLAS_INT stride ) { | |||
| dscal( &N, &alpha, X, &stride ); | |||
There was a problem hiding this comment.
@aman-095 We should also support negative strides here.
| */ | ||
| void API_SUFFIX(c_dscal_ndarray)( const CBLAS_INT N, const double alpha, double *X, const CBLAS_INT stride, const CBLAS_INT offset ) { | ||
| CBLAS_INT sx = stride; | ||
| if( sx < 0 ) { |
There was a problem hiding this comment.
| if( sx < 0 ) { | |
| if ( sx < 0 ) { |
kgryte
added
the
Needs Changes
Progresses #2039.
Description
This pull request:
ndarrayAPI forblas/base/dscal.Related Issues
This pull request:
Questions
No.
Other
No.
Checklist
@stdlib-js/reviewers