[libclc] Move logb/ilogb to CLC library; optimize #128028
Conversation
| #if __CLC_FPSIZE == 32 | ||
|
|
||
| _CLC_OVERLOAD _CLC_DEF __CLC_INTN __clc_ilogb(__CLC_GENTYPE x) { | ||
| __CLC_UINTN ux = __CLC_AS_UINTN(x); |
There was a problem hiding this comment.
This should be implemented in terms of frexp and avoid relying on the exponent bits.
Something like:
int ilogb(float x) {
int result;
(void)frexp(x, &result);
result = isnan(x) ? FP_ILOGBNAN : result;
result = isinf(x) ? INT_MAX : result;
return x == 0.0 ? FP_ILOGB0 : result;
}
Might be able to avoid the edge case handling too
There was a problem hiding this comment.
although I guess this better if you have a dedicated frexp_exp instruction. Could factor out the frexp_exp part into a utility function, which does the bithacking or the intrinsic
There was a problem hiding this comment.
Yeah, I'm not sure using frexp is currently better given that there's no native path for it. We talked about allowing that in the frexp PR.
But yeah even in that case, if we use the builtin/intrinsic for frexp it's not necessarily the case that it's faster than this implementation for ilogb. So I think the logb/ilogb-specific utility function would be the way to go.
I don't think we've got canonical examples of target-specific code paths for utility functions in libclc. We could do it with the SOURCES system, or preprocessor defines or some other system. I'm not yet sure.
This commit moves the logb and ilogb builtins to the CLC library. It simultaneously optimizes them both for vector types and for half types. Half types were previously promoting to float, whereas this commit provides them a native implementation. Everything passes the OpenCL-CTS. I had to intuit some magic numbers used by these implementations in order to generate the half variants. I gave them clearer definitions derived from what I believe are their actual component numbers, but named them 'magic' to convey that they weren't derived from first principles.
frasercrmck
force-pushed
the
libclc-clc-logb-ilogb
branch
from
5acc4a8 to
ebcca07
Compare
This commit moves the logb and ilogb builtins to the CLC library.
It simultaneously optimizes them both for vector types and for half types. Vector types were being scalarized in some cases. Half types were previously promoting to float, whereas this commit provides them a native implementation.
Everything passes the OpenCL-CTS.
I had to intuit some magic numbers used by these implementations in order to generate the half variants. I gave them clearer definitions derived from what I believe are their actual component numbers, but named them 'magic' to convey that they weren't derived from first principles.