diff mbox series

[FFmpeg-devel,11/14] swscale/x86/range_convert: update sse2 and avx2 range_convert functions to new API

Message ID 20240923124017.33659-12-ramiro.polla@gmail.com
State New
Headers show
Series swscale/range_convert: fix mpeg ranges in yuv range conversion for non-8-bit pixel formats | expand

Checks

Context Check Description
yinshiyou/make_loongarch64 success Make finished
yinshiyou/make_fate_loongarch64 fail Make fate failed
andriy/make_x86 success Make finished
andriy/make_fate_x86 fail Make fate failed

Commit Message

Ramiro Polla Sept. 23, 2024, 12:40 p.m. UTC
chrRangeFromJpeg8_1920_c:                             3874.8 ( 1.00x)
chrRangeFromJpeg8_1920_sse2:                          1493.8 ( 2.59x)
chrRangeFromJpeg8_1920_avx2:                           741.8 ( 5.22x)
chrRangeToJpeg8_1920_c:                               5232.8 ( 1.00x)
chrRangeToJpeg8_1920_sse2:                            1673.3 ( 3.13x)
chrRangeToJpeg8_1920_avx2:                             850.6 ( 6.15x)
lumRangeFromJpeg8_1920_c:                             2416.3 ( 1.00x)
lumRangeFromJpeg8_1920_sse2:                           760.1 ( 3.18x)
lumRangeFromJpeg8_1920_avx2:                           379.6 ( 6.37x)
lumRangeToJpeg8_1920_c:                               3121.1 ( 1.00x)
lumRangeToJpeg8_1920_sse2:                             870.1 ( 3.59x)
lumRangeToJpeg8_1920_avx2:                             434.8 ( 7.18x)
---
 libswscale/x86/range_convert.asm | 112 ++++++++++++++++++-------------
 libswscale/x86/swscale.c         |  14 ++--
 2 files changed, 73 insertions(+), 53 deletions(-)

Comments

Rémi Denis-Courmont Sept. 24, 2024, 4:32 p.m. UTC | #1
Le maanantaina 23. syyskuuta 2024, 15.40.14 EEST Ramiro Polla a écrit :
> chrRangeFromJpeg8_1920_c:                             3874.8 ( 1.00x)
> chrRangeFromJpeg8_1920_sse2:                          1493.8 ( 2.59x)
> chrRangeFromJpeg8_1920_avx2:                           741.8 ( 5.22x)
> chrRangeToJpeg8_1920_c:                               5232.8 ( 1.00x)
> chrRangeToJpeg8_1920_sse2:                            1673.3 ( 3.13x)
> chrRangeToJpeg8_1920_avx2:                             850.6 ( 6.15x)
> lumRangeFromJpeg8_1920_c:                             2416.3 ( 1.00x)
> lumRangeFromJpeg8_1920_sse2:                           760.1 ( 3.18x)
> lumRangeFromJpeg8_1920_avx2:                           379.6 ( 6.37x)
> lumRangeToJpeg8_1920_c:                               3121.1 ( 1.00x)
> lumRangeToJpeg8_1920_sse2:                             870.1 ( 3.59x)
> lumRangeToJpeg8_1920_avx2:                             434.8 ( 7.18x)

How do these compare to previous benchmarks?
Ramiro Polla Sept. 25, 2024, 12:58 p.m. UTC | #2
On Tue, Sep 24, 2024 at 6:39 PM Rémi Denis-Courmont <remi@remlab.net> wrote:
> Le maanantaina 23. syyskuuta 2024, 15.40.14 EEST Ramiro Polla a écrit :
> > chrRangeFromJpeg8_1920_c:                             3874.8 ( 1.00x)
> > chrRangeFromJpeg8_1920_sse2:                          1493.8 ( 2.59x)
> > chrRangeFromJpeg8_1920_avx2:                           741.8 ( 5.22x)
> > chrRangeToJpeg8_1920_c:                               5232.8 ( 1.00x)
> > chrRangeToJpeg8_1920_sse2:                            1673.3 ( 3.13x)
> > chrRangeToJpeg8_1920_avx2:                             850.6 ( 6.15x)
> > lumRangeFromJpeg8_1920_c:                             2416.3 ( 1.00x)
> > lumRangeFromJpeg8_1920_sse2:                           760.1 ( 3.18x)
> > lumRangeFromJpeg8_1920_avx2:                           379.6 ( 6.37x)
> > lumRangeToJpeg8_1920_c:                               3121.1 ( 1.00x)
> > lumRangeToJpeg8_1920_sse2:                             870.1 ( 3.59x)
> > lumRangeToJpeg8_1920_avx2:                             434.8 ( 7.18x)
>
> How do these compare to previous benchmarks?

I'll add more comprehensive benchmarks in v2.
diff mbox series

Patch

diff --git a/libswscale/x86/range_convert.asm b/libswscale/x86/range_convert.asm
index 97c7525448..d1aff63d7c 100644
--- a/libswscale/x86/range_convert.asm
+++ b/libswscale/x86/range_convert.asm
@@ -20,55 +20,53 @@ 
 
 %include "libavutil/x86/x86util.asm"
 
-SECTION_RODATA
-
-chr_to_mult:        times 4 dw 4663, 0
-chr_to_offset:      times 4 dd -9289992
-%define chr_to_shift 12
-
-chr_from_mult:      times 4 dw 1799, 0
-chr_from_offset:    times 4 dd 4081085
-%define chr_from_shift 11
-
-lum_to_mult:        times 4 dw 19077, 0
-lum_to_offset:      times 4 dd -39057361
-%define lum_to_shift 14
-
-lum_from_mult:      times 4 dw 14071, 0
-lum_from_offset:    times 4 dd 33561947
-%define lum_from_shift 14
-
 SECTION .text
 
-; NOTE: there is no need to clamp the input when converting to jpeg range
-;       (like we do in the C code) because packssdw will saturate the output.
-
 ;-----------------------------------------------------------------------------
 ; lumConvertRange
 ;
-; void ff_lumRangeToJpeg_<opt>(int16_t *dst, int width);
-; void ff_lumRangeFromJpeg_<opt>(int16_t *dst, int width);
+; void ff_lumRangeToJpeg_<opt>(int16_t *dst, int width,
+;                              int amax, int coeff, int64_t offset);
+; void ff_lumRangeFromJpeg_<opt>(int16_t *dst, int width,
+;                                int amax, int coeff, int64_t offset);
 ;
 ;-----------------------------------------------------------------------------
 
-%macro LUMCONVERTRANGE 4
-cglobal %1, 2, 2, 5, dst, width
+%macro LUMCONVERTRANGE 1
+%ifidni %1,To
+cglobal lumRange%1Jpeg, 5, 5, 6, dst, width, amax, coeff, offset
+%else
+cglobal lumRange%1Jpeg, 5, 5, 5, dst, width, amax, coeff, offset
+%endif
     shl          widthd, 1
-    VBROADCASTI128   m2, [%2]
-    VBROADCASTI128   m3, [%3]
+    movd            xm2, coeffd
+    VBROADCASTSS     m2, xm2
+%if ARCH_X86_64
+    movq            xm3, offsetq
+%else
+    movq            xm3, offsetm
+%endif
+    VBROADCASTSS     m3, xm3
     pxor             m4, m4
+%ifidni %1,To
+    movd            xm5, amaxd
+    SPLATW           m5, xm5
+%endif
     add            dstq, widthq
     neg          widthq
 .loop:
     movu             m0, [dstq+widthq]
+%ifidni %1,To
+    pminsw           m0, m5
+%endif
     punpckhwd        m1, m0, m4
     punpcklwd        m0, m4
     pmaddwd          m0, m2
     pmaddwd          m1, m2
     paddd            m0, m3
     paddd            m1, m3
-    psrad            m0, %4
-    psrad            m1, %4
+    psrad            m0, 14
+    psrad            m1, 14
     packssdw         m0, m1
     movu  [dstq+widthq], m0
     add          widthq, mmsize
@@ -79,23 +77,43 @@  cglobal %1, 2, 2, 5, dst, width
 ;-----------------------------------------------------------------------------
 ; chrConvertRange
 ;
-; void ff_chrRangeToJpeg_<opt>(int16_t *dstU, int16_t *dstV, int width);
-; void ff_chrRangeFromJpeg_<opt>(int16_t *dstU, int16_t *dstV, int width);
+; void ff_chrRangeToJpeg_<opt>(int16_t *dstU, int16_t *dstV, int width,
+;                              int amax, int coeff, int64_t offset);
+; void ff_chrRangeFromJpeg_<opt>(int16_t *dstU, int16_t *dstV, int width,
+;                                int amax, int coeff, int64_t offset);
 ;
 ;-----------------------------------------------------------------------------
 
-%macro CHRCONVERTRANGE 4
-cglobal %1, 3, 3, 7, dstU, dstV, width
+%macro CHRCONVERTRANGE 1
+%ifidni %1,To
+cglobal chrRange%1Jpeg, 6, 6, 8, dstU, dstV, width, amax, coeff, offset
+%else
+cglobal chrRange%1Jpeg, 6, 6, 7, dstU, dstV, width, amax, coeff, offset
+%endif
     shl          widthd, 1
-    VBROADCASTI128   m4, [%2]
-    VBROADCASTI128   m5, [%3]
+    movd            xm4, coeffd
+    VBROADCASTSS     m4, xm4
+%if ARCH_X86_64
+    movq            xm5, offsetq
+%else
+    movq            xm5, offsetm
+%endif
+    VBROADCASTSS     m5, xm5
     pxor             m6, m6
+%ifidni %1,To
+    movd            xm7, amaxd
+    SPLATW           m7, xm7
+%endif
     add           dstUq, widthq
     add           dstVq, widthq
     neg          widthq
 .loop:
     movu             m0, [dstUq+widthq]
     movu             m2, [dstVq+widthq]
+%ifidni %1,To
+    pminsw           m0, m7
+    pminsw           m2, m7
+%endif
     punpckhwd        m1, m0, m6
     punpckhwd        m3, m2, m6
     punpcklwd        m0, m6
@@ -108,10 +126,10 @@  cglobal %1, 3, 3, 7, dstU, dstV, width
     paddd            m1, m5
     paddd            m2, m5
     paddd            m3, m5
-    psrad            m0, %4
-    psrad            m1, %4
-    psrad            m2, %4
-    psrad            m3, %4
+    psrad            m0, 14
+    psrad            m1, 14
+    psrad            m2, 14
+    psrad            m3, 14
     packssdw         m0, m1
     packssdw         m2, m3
     movu [dstUq+widthq], m0
@@ -122,15 +140,15 @@  cglobal %1, 3, 3, 7, dstU, dstV, width
 %endmacro
 
 INIT_XMM sse2
-LUMCONVERTRANGE lumRangeToJpeg,   lum_to_mult,   lum_to_offset,   lum_to_shift
-CHRCONVERTRANGE chrRangeToJpeg,   chr_to_mult,   chr_to_offset,   chr_to_shift
-LUMCONVERTRANGE lumRangeFromJpeg, lum_from_mult, lum_from_offset, lum_from_shift
-CHRCONVERTRANGE chrRangeFromJpeg, chr_from_mult, chr_from_offset, chr_from_shift
+LUMCONVERTRANGE To
+CHRCONVERTRANGE To
+LUMCONVERTRANGE From
+CHRCONVERTRANGE From
 
 %if HAVE_AVX2_EXTERNAL
 INIT_YMM avx2
-LUMCONVERTRANGE lumRangeToJpeg,   lum_to_mult,   lum_to_offset,   lum_to_shift
-CHRCONVERTRANGE chrRangeToJpeg,   chr_to_mult,   chr_to_offset,   chr_to_shift
-LUMCONVERTRANGE lumRangeFromJpeg, lum_from_mult, lum_from_offset, lum_from_shift
-CHRCONVERTRANGE chrRangeFromJpeg, chr_from_mult, chr_from_offset, chr_from_shift
+LUMCONVERTRANGE To
+CHRCONVERTRANGE To
+LUMCONVERTRANGE From
+CHRCONVERTRANGE From
 %endif
diff --git a/libswscale/x86/swscale.c b/libswscale/x86/swscale.c
index d55e45471f..2377365e91 100644
--- a/libswscale/x86/swscale.c
+++ b/libswscale/x86/swscale.c
@@ -464,24 +464,26 @@  INPUT_PLANAR_RGB_A_ALL_DECL(avx2);
 } while (0)
 
 #define RANGE_CONVERT_FUNCS_DECL(opt)                                       \
-void ff_lumRangeFromJpeg_ ##opt(int16_t *dst, int width);                   \
-void ff_chrRangeFromJpeg_ ##opt(int16_t *dstU, int16_t *dstV, int width);   \
-void ff_lumRangeToJpeg_ ##opt(int16_t *dst, int width);                     \
-void ff_chrRangeToJpeg_ ##opt(int16_t *dstU, int16_t *dstV, int width);     \
+void ff_lumRangeFromJpeg_ ##opt(int16_t *dst, int width,                    \
+                                int amax, int coeff, int64_t offset);       \
+void ff_chrRangeFromJpeg_ ##opt(int16_t *dstU, int16_t *dstV, int width,    \
+                                int amax, int coeff, int64_t offset);       \
+void ff_lumRangeToJpeg_ ##opt(int16_t *dst, int width,                      \
+                              int amax, int coeff, int64_t offset);         \
+void ff_chrRangeToJpeg_ ##opt(int16_t *dstU, int16_t *dstV, int width,      \
+                              int amax, int coeff, int64_t offset);         \
 
 RANGE_CONVERT_FUNCS_DECL(sse2);
 RANGE_CONVERT_FUNCS_DECL(avx2);
 
 av_cold void ff_sws_init_range_convert_x86(SwsContext *c)
 {
-#if 0
     int cpu_flags = av_get_cpu_flags();
     if (EXTERNAL_AVX2_FAST(cpu_flags)) {
         RANGE_CONVERT_FUNCS(avx2);
     } else if (EXTERNAL_SSE2(cpu_flags)) {
         RANGE_CONVERT_FUNCS(sse2);
     }
-#endif
 }
 
 av_cold void ff_sws_init_swscale_x86(SwsContext *c)