From: Noah Goldstein <goldstein.w.n@gmail.com>
To: libc-alpha@sourceware.org
Cc: goldstein.w.n@gmail.com, hjl.tools@gmail.com,
andrey.kolesov@intel.com, carlos@systemhalted.org
Subject: [PATCH v1 13/27] x86/fpu: Optimize svml_s_atanf8_core_avx2.S
Date: Wed, 7 Dec 2022 00:52:22 -0800 [thread overview]
Message-ID: <20221207085236.1424424-13-goldstein.w.n@gmail.com> (raw)
In-Reply-To: <20221207085236.1424424-1-goldstein.w.n@gmail.com>
1. Cleanup some missed optimizations in instruction selection /
unnecissary repeated rodata references.
2. Remove unused rodata.
3. Use common data definitions where possible.
Code Size Change: -12 Bytes (163 - 175)
Input New Time / Old Time
0F (0x00000000) -> 0.8484
0F (0x0000ffff, Denorm) -> 0.9993
.1F (0x3dcccccd) -> 0.9368
5F (0x40a00000) -> 0.9476
2315255808F (0x4f0a0000) -> 0.9454
-NaN (0xffffffff) -> 0.9193
---
.../fpu/multiarch/svml_s_atanf8_core_avx2.S | 162 +++++++-----------
1 file changed, 58 insertions(+), 104 deletions(-)
diff --git a/sysdeps/x86_64/fpu/multiarch/svml_s_atanf8_core_avx2.S b/sysdeps/x86_64/fpu/multiarch/svml_s_atanf8_core_avx2.S
index ee49a3e10e..649277c682 100644
--- a/sysdeps/x86_64/fpu/multiarch/svml_s_atanf8_core_avx2.S
+++ b/sysdeps/x86_64/fpu/multiarch/svml_s_atanf8_core_avx2.S
@@ -28,120 +28,74 @@
*
*/
-/* Offsets for data table __svml_satan_data_internal
- */
-#define _sSIGN_MASK 0
-#define _sABS_MASK 32
-#define _sONE 64
-#define _sPIO2 96
-#define _sPC8 128
-#define _sPC7 160
-#define _sPC6 192
-#define _sPC5 224
-#define _sPC4 256
-#define _sPC3 288
-#define _sPC2 320
-#define _sPC1 352
-#define _sPC0 384
+
+#define LOCAL_DATA_NAME __svml_satan_data_internal
+#include "svml_s_common_avx2_rodata_offsets.h"
+/* Offsets for data table __svml_satan_data_internal. */
+#define _sPC8 0
+#define _sPC7 32
+#define _sPC6 64
+#define _sPC5 96
+#define _sPC4 128
+#define _sPC3 160
+#define _sPC2 192
+#define _sPC1 224
+#define _sPC0 256
#include <sysdep.h>
.section .text.avx2, "ax", @progbits
ENTRY(_ZGVdN8v_atanf_avx2)
- /*
- * 1) If x>1, then r=-1/x, PIO2=Pi/2
- * 2) If -1<=x<=1, then r=x, PIO2=0
- * 3) If x<-1, then r=-1/x, PIO2=-Pi/2
- */
- vmovups _sONE+__svml_satan_data_internal(%rip), %ymm2
- vmovups __svml_satan_data_internal(%rip), %ymm7
- vmovups _sPC7+__svml_satan_data_internal(%rip), %ymm13
+ /* 1) If x>1, then r=-1/x, PIO2=Pi/2
+ 2) If -1<=x<=1, then r=x, PIO2=0
+ 3) If x<-1, then r=-1/x, PIO2=-Pi/2. */
+ vmovups COMMON_DATA(_AbsMask)(%rip), %ymm7
+ vmovups COMMON_DATA(_OneF)(%rip), %ymm2
+
+ vandps %ymm0, %ymm7, %ymm3
+ /* Use minud\maxud operations for argument reduction. */
+ vpmaxud %ymm3, %ymm2, %ymm5
+ vpminud %ymm3, %ymm2, %ymm4
- /*
- * To use minps\maxps operations for argument reduction
- * uncomment _AT_USEMINMAX_ definition
- * Declarations
- * Variables
- * Constants
- */
- vandps _sABS_MASK+__svml_satan_data_internal(%rip), %ymm0, %ymm3
- vmaxps %ymm3, %ymm2, %ymm5
- vminps %ymm3, %ymm2, %ymm4
- vcmple_oqps %ymm2, %ymm3, %ymm6
- vdivps %ymm5, %ymm4, %ymm11
- vandps %ymm7, %ymm0, %ymm9
- vandnps %ymm7, %ymm6, %ymm8
- vxorps %ymm9, %ymm8, %ymm10
- vxorps %ymm11, %ymm10, %ymm15
+ vdivps %ymm5, %ymm4, %ymm4
- /* Polynomial. */
- vmulps %ymm15, %ymm15, %ymm14
- vmovups _sPC8+__svml_satan_data_internal(%rip), %ymm0
- vmulps %ymm14, %ymm14, %ymm12
- vfmadd213ps _sPC6+__svml_satan_data_internal(%rip), %ymm12, %ymm0
- vfmadd213ps _sPC5+__svml_satan_data_internal(%rip), %ymm12, %ymm13
- vfmadd213ps _sPC4+__svml_satan_data_internal(%rip), %ymm12, %ymm0
- vfmadd213ps _sPC3+__svml_satan_data_internal(%rip), %ymm12, %ymm13
- vfmadd213ps _sPC2+__svml_satan_data_internal(%rip), %ymm12, %ymm0
- vfmadd213ps _sPC1+__svml_satan_data_internal(%rip), %ymm12, %ymm13
- vfmadd213ps %ymm13, %ymm14, %ymm0
- vfmadd213ps _sPC0+__svml_satan_data_internal(%rip), %ymm14, %ymm0
- vandnps _sPIO2+__svml_satan_data_internal(%rip), %ymm6, %ymm1
- vxorps %ymm9, %ymm1, %ymm1
+ vpcmpgtd %ymm2, %ymm3, %ymm6
+ vandnps %ymm0, %ymm7, %ymm3
+ vandnps %ymm6, %ymm7, %ymm7
+ vxorps %ymm3, %ymm7, %ymm5
+ vxorps %ymm4, %ymm5, %ymm7
+ /* Polynomial. */
+ vmulps %ymm4, %ymm4, %ymm1
+ vmovups LOCAL_DATA(_sPC8)(%rip), %ymm0
+ vmovups LOCAL_DATA(_sPC7)(%rip), %ymm4
+ vmulps %ymm1, %ymm1, %ymm5
+ vfmadd213ps LOCAL_DATA(_sPC6)(%rip), %ymm5, %ymm0
+ vfmadd213ps LOCAL_DATA(_sPC5)(%rip), %ymm5, %ymm4
+ vfmadd213ps LOCAL_DATA(_sPC4)(%rip), %ymm5, %ymm0
+ vfmadd213ps LOCAL_DATA(_sPC3)(%rip), %ymm5, %ymm4
+ vfmadd213ps LOCAL_DATA(_sPC2)(%rip), %ymm5, %ymm0
+ vfmadd213ps LOCAL_DATA(_sPC1)(%rip), %ymm5, %ymm4
+ vfmadd213ps %ymm4, %ymm1, %ymm0
+ vfmadd213ps %ymm2, %ymm1, %ymm0
+ vandps COMMON_DATA(_TanSPI1_FMA)(%rip), %ymm6, %ymm1
+ vxorps %ymm3, %ymm1, %ymm1
- /* Reconstruction. */
- vfmadd213ps %ymm1, %ymm15, %ymm0
+ /* Reconstruction. */
+ vfmadd213ps %ymm1, %ymm7, %ymm0
ret
END(_ZGVdN8v_atanf_avx2)
- .section .rodata, "a"
- .align 32
-
-#ifdef __svml_satan_data_internal_typedef
-typedef unsigned int VUINT32;
-typedef struct {
- __declspec(align(32)) VUINT32 _sSIGN_MASK[8][1];
- __declspec(align(32)) VUINT32 _sABS_MASK[8][1];
- __declspec(align(32)) VUINT32 _sONE[8][1];
- __declspec(align(32)) VUINT32 _sPIO2[8][1];
- __declspec(align(32)) VUINT32 _sPC8[8][1];
- __declspec(align(32)) VUINT32 _sPC7[8][1];
- __declspec(align(32)) VUINT32 _sPC6[8][1];
- __declspec(align(32)) VUINT32 _sPC5[8][1];
- __declspec(align(32)) VUINT32 _sPC4[8][1];
- __declspec(align(32)) VUINT32 _sPC3[8][1];
- __declspec(align(32)) VUINT32 _sPC2[8][1];
- __declspec(align(32)) VUINT32 _sPC1[8][1];
- __declspec(align(32)) VUINT32 _sPC0[8][1];
-} __svml_satan_data_internal;
-#endif
-__svml_satan_data_internal:
- .long 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000 // _sSIGN_MASK
- .align 32
- .long 0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF // _sABS_MASK
- .align 32
- .long 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000 // _sONE
- .align 32
- .long 0x3FC90FDB, 0x3FC90FDB, 0x3FC90FDB, 0x3FC90FDB, 0x3FC90FDB, 0x3FC90FDB, 0x3FC90FDB, 0x3FC90FDB // _sPIO2
- .align 32
- .long 0x3B322CC0, 0x3B322CC0, 0x3B322CC0, 0x3B322CC0, 0x3B322CC0, 0x3B322CC0, 0x3B322CC0, 0x3B322CC0 // _sPC8
- .align 32
- .long 0xBC7F2631, 0xBC7F2631, 0xBC7F2631, 0xBC7F2631, 0xBC7F2631, 0xBC7F2631, 0xBC7F2631, 0xBC7F2631 // _sPC7
- .align 32
- .long 0x3D2BC384, 0x3D2BC384, 0x3D2BC384, 0x3D2BC384, 0x3D2BC384, 0x3D2BC384, 0x3D2BC384, 0x3D2BC384 // _sPC6
- .align 32
- .long 0xBD987629, 0xBD987629, 0xBD987629, 0xBD987629, 0xBD987629, 0xBD987629, 0xBD987629, 0xBD987629 // _sPC5
- .align 32
- .long 0x3DD96474, 0x3DD96474, 0x3DD96474, 0x3DD96474, 0x3DD96474, 0x3DD96474, 0x3DD96474, 0x3DD96474 // _sPC4
- .align 32
- .long 0xBE1161F8, 0xBE1161F8, 0xBE1161F8, 0xBE1161F8, 0xBE1161F8, 0xBE1161F8, 0xBE1161F8, 0xBE1161F8 // _sPC3
- .align 32
- .long 0x3E4CB79F, 0x3E4CB79F, 0x3E4CB79F, 0x3E4CB79F, 0x3E4CB79F, 0x3E4CB79F, 0x3E4CB79F, 0x3E4CB79F // _sPC2
- .align 32
- .long 0xBEAAAA49, 0xBEAAAA49, 0xBEAAAA49, 0xBEAAAA49, 0xBEAAAA49, 0xBEAAAA49, 0xBEAAAA49, 0xBEAAAA49 // _sPC1
- .align 32
- .long 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000 // _sPC0
+ .section .rodata.avx2, "a"
.align 32
- .type __svml_satan_data_internal, @object
- .size __svml_satan_data_internal, .-__svml_satan_data_internal
+LOCAL_DATA_NAME:
+ DATA_VEC (LOCAL_DATA_NAME, _sPC8, 0x3B322CC0)
+ DATA_VEC (LOCAL_DATA_NAME, _sPC7, 0xBC7F2631)
+ DATA_VEC (LOCAL_DATA_NAME, _sPC6, 0x3D2BC384)
+ DATA_VEC (LOCAL_DATA_NAME, _sPC5, 0xBD987629)
+ DATA_VEC (LOCAL_DATA_NAME, _sPC4, 0x3DD96474)
+ DATA_VEC (LOCAL_DATA_NAME, _sPC3, 0xBE1161F8)
+ DATA_VEC (LOCAL_DATA_NAME, _sPC2, 0x3E4CB79F)
+ DATA_VEC (LOCAL_DATA_NAME, _sPC1, 0xBEAAAA49)
+ .type LOCAL_DATA_NAME, @object
+ .size LOCAL_DATA_NAME, .-LOCAL_DATA_NAME
--
2.34.1
next prev parent reply other threads:[~2022-12-07 8:53 UTC|newest]
Thread overview: 38+ messages / expand[flat|nested] mbox.gz Atom feed top
2022-12-07 8:52 [PATCH v1 01/27] x86/fpu: Create helper file for common data macros Noah Goldstein
2022-12-07 8:52 ` [PATCH v1 02/27] x86/fpu: Add file for common data used across svml_s_*_avx2.S files Noah Goldstein
2022-12-07 8:52 ` [PATCH v1 03/27] x86/fpu: Add file for common data used across svml_s_*_avx512.S files Noah Goldstein
2022-12-07 8:52 ` [PATCH v1 04/27] x86/fpu: Add file for common data used across svml_s_*_sse4.S files Noah Goldstein
2022-12-07 8:52 ` [PATCH v1 05/27] x86/fpu: Build common data files for svml_s_*_{avx512,avx2,sse4}.S Noah Goldstein
2022-12-07 8:52 ` [PATCH v1 06/27] x86/fpu: Update rodata usage in svml_s_tanhf_*_{avx2,sse4} Noah Goldstein
2022-12-07 8:52 ` [PATCH v1 07/27] x86/fpu: Update rodata usage in svml_s_tanhf16_core_avx512.S Noah Goldstein
2022-12-16 17:05 ` H.J. Lu
2022-12-16 18:17 ` Noah Goldstein
2022-12-16 21:37 ` H.J. Lu
2022-12-16 21:51 ` Noah Goldstein
2022-12-16 22:01 ` H.J. Lu
2022-12-16 22:54 ` Sunil Pandey
2023-06-27 18:23 ` Noah Goldstein
2022-12-07 8:52 ` [PATCH v1 08/27] x86/fpu: Update rodata usage in svml_s_atanhf16_core_avx512.S Noah Goldstein
2022-12-07 8:52 ` [PATCH v1 09/27] x86/fpu: Update rodata usage in svml_s_atanhf4_core_sse4.S Noah Goldstein
2022-12-07 8:52 ` [PATCH v1 10/27] x86/fpu: Update rodata usage in svml_s_atanhf8_core_avx2.S Noah Goldstein
2022-12-07 8:52 ` [PATCH v1 11/27] x86/fpu: Optimize svml_s_atanf16_core_avx512.S Noah Goldstein
2022-12-07 8:52 ` [PATCH v1 12/27] x86/fpu: Optimize svml_s_atanf4_core_sse4.S Noah Goldstein
2022-12-07 8:52 ` Noah Goldstein [this message]
2022-12-07 8:52 ` [PATCH v1 14/27] x86/fpu: Add common rodata file for svml_s_tanf_*_{avx512,avx2,sse4}.S Noah Goldstein
2022-12-07 8:52 ` [PATCH v1 15/27] x86/fpu: Optimize svml_s_tanf16_core_avx512.S Noah Goldstein
2022-12-07 8:52 ` [PATCH v1 16/27] x86/fpu: Optimize svml_s_tanf4_core_sse4.S Noah Goldstein
2022-12-07 8:52 ` [PATCH v1 17/27] x86/fpu: Optimize svml_s_tanf8_core_avx2.S Noah Goldstein
2022-12-07 8:52 ` [PATCH v1 18/27] x86/fpu: Optimize svml_s_log10f16_core_avx512.S Noah Goldstein
2022-12-07 8:52 ` [PATCH v1 19/27] x86/fpu: Optimize svml_s_log10f4_core_sse4.S Noah Goldstein
2022-12-07 8:52 ` [PATCH v1 20/27] x86/fpu: Optimize svml_s_log10f8_core_avx2.S Noah Goldstein
2022-12-07 8:52 ` [PATCH v1 21/27] x86/fpu: Optimize svml_s_log2f16_core_avx512.S Noah Goldstein
2022-12-07 8:52 ` [PATCH v1 22/27] x86/fpu: Optimize svml_s_log2f4_core_sse4.S Noah Goldstein
2022-12-07 8:52 ` [PATCH v1 23/27] x86/fpu: Optimize svml_s_log2f8_core_avx2.S Noah Goldstein
2022-12-07 8:52 ` [PATCH v1 24/27] x86/fpu: Optimize svml_s_logf16_core_avx512.S Noah Goldstein
2022-12-07 8:52 ` [PATCH v1 25/27] x86/fpu: Optimize svml_s_logf4_core_sse4.S Noah Goldstein
2022-12-07 8:52 ` [PATCH v1 26/27] x86/fpu: Optimize svml_s_logf8_core_avx2.S Noah Goldstein
2022-12-07 8:52 ` [PATCH v1 27/27] x86/fpu: Remove unused svml_s_logf_data.S file Noah Goldstein
2022-12-07 23:53 ` [PATCH v1 01/27] x86/fpu: Create helper file for common data macros H.J. Lu
2022-12-08 0:13 ` Noah Goldstein
2022-12-08 0:22 ` H.J. Lu
2022-12-08 0:46 ` Noah Goldstein
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20221207085236.1424424-13-goldstein.w.n@gmail.com \
--to=goldstein.w.n@gmail.com \
--cc=andrey.kolesov@intel.com \
--cc=carlos@systemhalted.org \
--cc=hjl.tools@gmail.com \
--cc=libc-alpha@sourceware.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).