topic Re: FP16 math functions not getting inlined in Intel® oneAPI DPC++/C++ Compiler

FP16 math functions not getting inlined

rabbits — Mon, 16 Sep 2024 07:25:13 GMT

icx 2024.2.0

-Ofast -xSAPPHIRERAPIDS

#include <immintrin.h> #include <mathimf.h> _Float16 test_fp16(_Float16 x, _Float16 y) { return sqrtf16(x) + sqrtf16(y); }

generates the following code:

test_fp16(_Float16, _Float16): push rax vmovsh word ptr [rsp + 4], xmm1 call sqrtf16 vmovsh word ptr [rsp], xmm0 vmovsh xmm0, word ptr [rsp + 4] call sqrtf16 vaddsh xmm0, xmm0, word ptr [rsp] pop rax ret

Is there a compiler flag I can use to make it generate the expected code instead?

test_fp16(_Float16, _Float16): vsqrtsh xmm0, xmm0, xmm0 vsqrtsh xmm1, xmm1, xmm1 vaddsh xmm0, xmm1, xmm0 ret

See https://godbolt.org/z/xWsqW8djE for more details.

Re: FP16 math functions not getting inlined

Alex_Y_Intel — Wed, 18 Sep 2024 00:27:04 GMT

I've escalated your issue to our internal team for further investigation, thanks.

Re:FP16 math functions not getting inlined

Alex_Y_Intel — Thu, 19 Sep 2024 07:09:47 GMT

In order for a function to be inlined, it has to have the header/source file when compiling. If you take a look at the "inline report" with: icx test.c -Ofast -xSAPPHIRERAPIDS -c -qopt-report=3 -qopt-report-file=stderr

Global optimization report for : test_fp32

=================================================================

Global optimization report for : test_fp16

=================================================================

Global optimization report for : test_fp16_fix

=================================================================

---- Begin Inlining Report ----

Option Values:

inline-threshold: 225

inlinehint-threshold: 325

inlinecold-threshold: 45

inlineoptsize-threshold: 15

DEAD STATIC FUNC: _mm_set_sh

DEAD STATIC FUNC: _mm_setzero_ph

DEAD STATIC FUNC: _mm_sqrt_sh

DEAD STATIC FUNC: sqrtf16_fast

COMPILE FUNC: test_fp32

-> llvm.sqrt.f32 test.c (6,12)

-> llvm.sqrt.f32 test.c (6,23)

COMPILE FUNC: test_fp16

-> EXTERN: sqrtf16 test.c (11,12)

-> EXTERN: sqrtf16 test.c (11,25)

COMPILE FUNC: test_fp16_fix

-> INLINE: sqrtf16_fast test.c (21,12) (-30<=487)

-> INLINE: _mm_sqrt_sh test.c (16,12)

-> INLINE: _mm_setzero_ph test.c

-> llvm.sqrt.f16 test.c

-> INLINE: _mm_setzero_ph test.c (16,24)

-> INLINE: _mm_set_sh test.c (16,42)

-> INLINE: sqrtf16_fast test.c (21,30) (-30<=487)

-> INLINE: _mm_sqrt_sh test.c (16,12)

-> INLINE: _mm_setzero_ph test.c

-> llvm.sqrt.f16 test.c

-> INLINE: _mm_setzero_ph test.c (16,24)

-> INLINE: _mm_set_sh test.c (16,42)

---- End Inlining Report ------

Please note that sqrtf16 uses an external function:

COMPILE FUNC: test_fp16

-> EXTERN: sqrtf16 test.c (11,12)

-> EXTERN: sqrtf16 test.c (11,25)

In fact, this sqrtf16 is actually extracted from libimf.a and linked by the linker, and libimf.a is a library object, not a header/source file.

In conclusion, icx does not provide the source code for sqrtf16 to be inlined.

Re: Re:FP16 math functions not getting inlined

rabbits — Thu, 19 Sep 2024 09:05:42 GMT

> In conclusion, icx does not provide the source code for sqrtf16 to be inlined.

swell!

Re: FP16 math functions not getting inlined

Alex_Y_Intel — Fri, 20 Sep 2024 00:00:03 GMT

I'm discussing this problem with another team and trying to find if we have other solutions.

Re: FP16 math functions not getting inlined

rabbits — Fri, 20 Sep 2024 10:05:37 GMT

Let me turn Gordon in his grave for you.