[v3] x86: Remove __mmask intrinsics in strstr-avx512.c
Checks
Context |
Check |
Description |
dj/TryBot-apply_patch |
success
|
Patch applied to master at the time it was sent
|
dj/TryBot-32bit |
success
|
Build for i686
|
Commit Message
The intrinsics are not available before GCC7 and using standard
operators generates code of equivilent or better quality.
Removed:
_cvtmask64_u64
_kshiftri_mask64
_kand_mask64
Geometric Mean of 5 Runs of Full Benchmark Suite New / Old: 0.958
---
sysdeps/x86_64/multiarch/strstr-avx512.c | 16 ++++++++++------
1 file changed, 10 insertions(+), 6 deletions(-)
Comments
On Tue, Jul 12, 2022 at 11:48 AM Noah Goldstein <goldstein.w.n@gmail.com> wrote:
>
> The intrinsics are not available before GCC7 and using standard
> operators generates code of equivilent or better quality.
generate equivalent
Otherwise LGTM.
Thanks.
>
> Removed:
> _cvtmask64_u64
> _kshiftri_mask64
> _kand_mask64
>
> Geometric Mean of 5 Runs of Full Benchmark Suite New / Old: 0.958
> ---
> sysdeps/x86_64/multiarch/strstr-avx512.c | 16 ++++++++++------
> 1 file changed, 10 insertions(+), 6 deletions(-)
>
> diff --git a/sysdeps/x86_64/multiarch/strstr-avx512.c b/sysdeps/x86_64/multiarch/strstr-avx512.c
> index 2ab9e96db8..e44c1a05dc 100644
> --- a/sysdeps/x86_64/multiarch/strstr-avx512.c
> +++ b/sysdeps/x86_64/multiarch/strstr-avx512.c
> @@ -26,6 +26,10 @@
> #define ZMM_SIZE_IN_BYTES 64
> #define PAGESIZE 4096
>
> +#define cvtmask64_u64(...) (uint64_t) (__VA_ARGS__)
> +#define kshiftri_mask64(x, y) ((x) >> (y))
> +#define kand_mask64(x, y) ((x) & (y))
> +
> /*
> Returns the index of the first edge within the needle, returns 0 if no edge
> is found. Example: 'ab' is the first edge in 'aaaaaaaaaabaarddg'
> @@ -133,15 +137,15 @@ __strstr_avx512 (const char *haystack, const char *ned)
> __m512i hay0 = _mm512_maskz_loadu_epi8 (loadmask, haystack + hay_index);
> /* Search for NULL and compare only till null char */
> uint64_t nullmask
> - = _cvtmask64_u64 (_mm512_mask_testn_epi8_mask (loadmask, hay0, hay0));
> + = cvtmask64_u64 (_mm512_mask_testn_epi8_mask (loadmask, hay0, hay0));
> uint64_t cmpmask = nullmask ^ (nullmask - ONE_64BIT);
> - cmpmask = cmpmask & _cvtmask64_u64 (loadmask);
> + cmpmask = cmpmask & cvtmask64_u64 (loadmask);
> /* Search for the 2 charaters of needle */
> __mmask64 k0 = _mm512_cmpeq_epi8_mask (hay0, ned0);
> __mmask64 k1 = _mm512_cmpeq_epi8_mask (hay0, ned1);
> - k1 = _kshiftri_mask64 (k1, 1);
> + k1 = kshiftri_mask64 (k1, 1);
> /* k2 masks tell us if both chars from needle match */
> - uint64_t k2 = _cvtmask64_u64 (_kand_mask64 (k0, k1)) & cmpmask;
> + uint64_t k2 = cvtmask64_u64 (kand_mask64 (k0, k1)) & cmpmask;
> /* For every match, search for the entire needle for a full match */
> while (k2)
> {
> @@ -178,13 +182,13 @@ __strstr_avx512 (const char *haystack, const char *ned)
> hay0 = _mm512_loadu_si512 (haystack + hay_index);
> hay1 = _mm512_load_si512 (haystack + hay_index
> + 1); // Always 64 byte aligned
> - nullmask = _cvtmask64_u64 (_mm512_testn_epi8_mask (hay1, hay1));
> + nullmask = cvtmask64_u64 (_mm512_testn_epi8_mask (hay1, hay1));
> /* Compare only till null char */
> cmpmask = nullmask ^ (nullmask - ONE_64BIT);
> k0 = _mm512_cmpeq_epi8_mask (hay0, ned0);
> k1 = _mm512_cmpeq_epi8_mask (hay1, ned1);
> /* k2 masks tell us if both chars from needle match */
> - k2 = _cvtmask64_u64 (_kand_mask64 (k0, k1)) & cmpmask;
> + k2 = cvtmask64_u64 (kand_mask64 (k0, k1)) & cmpmask;
> /* For every match, compare full strings for potential match */
> while (k2)
> {
> --
> 2.34.1
>
On Jul 12 2022, H.J. Lu via Libc-alpha wrote:
> On Tue, Jul 12, 2022 at 11:48 AM Noah Goldstein <goldstein.w.n@gmail.com> wrote:
>>
>> The intrinsics are not available before GCC7 and using standard
>> operators generates code of equivilent or better quality.
> generate equivalent
I think "using ... generates" is correct.
On Tue, Jul 12, 2022 at 12:48 PM Andreas Schwab <schwab@linux-m68k.org> wrote:
>
> On Jul 12 2022, H.J. Lu via Libc-alpha wrote:
>
> > On Tue, Jul 12, 2022 at 11:48 AM Noah Goldstein <goldstein.w.n@gmail.com> wrote:
> >>
> >> The intrinsics are not available before GCC7 and using standard
> >> operators generates code of equivilent or better quality.
> > generate equivalent
>
> I think "using ... generates" is correct.
Going to check in with just 'equivilent' -> 'equivalent'
>
> --
> Andreas Schwab, schwab@linux-m68k.org
> GPG Key fingerprint = 7578 EB47 D4E5 4D69 2510 2552 DF73 E780 A9DA AEC1
> "And now for something completely different."
@@ -26,6 +26,10 @@
#define ZMM_SIZE_IN_BYTES 64
#define PAGESIZE 4096
+#define cvtmask64_u64(...) (uint64_t) (__VA_ARGS__)
+#define kshiftri_mask64(x, y) ((x) >> (y))
+#define kand_mask64(x, y) ((x) & (y))
+
/*
Returns the index of the first edge within the needle, returns 0 if no edge
is found. Example: 'ab' is the first edge in 'aaaaaaaaaabaarddg'
@@ -133,15 +137,15 @@ __strstr_avx512 (const char *haystack, const char *ned)
__m512i hay0 = _mm512_maskz_loadu_epi8 (loadmask, haystack + hay_index);
/* Search for NULL and compare only till null char */
uint64_t nullmask
- = _cvtmask64_u64 (_mm512_mask_testn_epi8_mask (loadmask, hay0, hay0));
+ = cvtmask64_u64 (_mm512_mask_testn_epi8_mask (loadmask, hay0, hay0));
uint64_t cmpmask = nullmask ^ (nullmask - ONE_64BIT);
- cmpmask = cmpmask & _cvtmask64_u64 (loadmask);
+ cmpmask = cmpmask & cvtmask64_u64 (loadmask);
/* Search for the 2 charaters of needle */
__mmask64 k0 = _mm512_cmpeq_epi8_mask (hay0, ned0);
__mmask64 k1 = _mm512_cmpeq_epi8_mask (hay0, ned1);
- k1 = _kshiftri_mask64 (k1, 1);
+ k1 = kshiftri_mask64 (k1, 1);
/* k2 masks tell us if both chars from needle match */
- uint64_t k2 = _cvtmask64_u64 (_kand_mask64 (k0, k1)) & cmpmask;
+ uint64_t k2 = cvtmask64_u64 (kand_mask64 (k0, k1)) & cmpmask;
/* For every match, search for the entire needle for a full match */
while (k2)
{
@@ -178,13 +182,13 @@ __strstr_avx512 (const char *haystack, const char *ned)
hay0 = _mm512_loadu_si512 (haystack + hay_index);
hay1 = _mm512_load_si512 (haystack + hay_index
+ 1); // Always 64 byte aligned
- nullmask = _cvtmask64_u64 (_mm512_testn_epi8_mask (hay1, hay1));
+ nullmask = cvtmask64_u64 (_mm512_testn_epi8_mask (hay1, hay1));
/* Compare only till null char */
cmpmask = nullmask ^ (nullmask - ONE_64BIT);
k0 = _mm512_cmpeq_epi8_mask (hay0, ned0);
k1 = _mm512_cmpeq_epi8_mask (hay1, ned1);
/* k2 masks tell us if both chars from needle match */
- k2 = _cvtmask64_u64 (_kand_mask64 (k0, k1)) & cmpmask;
+ k2 = cvtmask64_u64 (kand_mask64 (k0, k1)) & cmpmask;
/* For every match, compare full strings for potential match */
while (k2)
{