chrRangeFromJpeg16_1920_c: 3893.0 ( 1.00x)
chrRangeFromJpeg16_1920_sse2: 3249.0 ( 1.20x)
chrRangeFromJpeg16_1920_avx2: 1636.0 ( 2.38x)
chrRangeToJpeg16_1920_c: 7731.0 ( 1.00x)
chrRangeToJpeg16_1920_sse2: 4722.8 ( 1.64x)
chrRangeToJpeg16_1920_sse4: 3715.0 ( 2.08x)
chrRangeToJpeg16_1920_avx2: 1859.0 ( 4.16x)
lumRangeFromJpeg16_1920_c: 2385.5 ( 1.00x)
lumRangeFromJpeg16_1920_sse2: 1636.0 ( 1.46x)
lumRangeFromJpeg16_1920_avx2: 839.8 ( 2.84x)
lumRangeToJpeg16_1920_c: 3533.0 ( 1.00x)
lumRangeToJpeg16_1920_sse2: 2362.2 ( 1.50x)
lumRangeToJpeg16_1920_sse4: 1861.0 ( 1.90x)
lumRangeToJpeg16_1920_avx2: 946.0 ( 3.73x)
---
libswscale/x86/range_convert.asm | 133 ++++++++++++++++++++++++++-----
libswscale/x86/swscale.c | 50 +++++++++---
2 files changed, 149 insertions(+), 34 deletions(-)
diff --git a/libswscale/x86/range_convert.asm b/libswscale/x86/range_convert.asm
index d1aff63d7c..83dbe6c4b8 100644
--- a/libswscale/x86/range_convert.asm
+++ b/libswscale/x86/range_convert.asm
@@ -25,20 +25,28 @@ SECTION .text
;-----------------------------------------------------------------------------
; lumConvertRange
;
-; 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);
+; void ff_lumRangeToJpeg{8,16}_<opt>(int16_t *dst, int width,
+; int amax, int coeff, int64_t offset);
+; void ff_lumRangeFromJpeg{8,16}_<opt>(int16_t *dst, int width,
+; int amax, int coeff, int64_t offset);
;
;-----------------------------------------------------------------------------
-%macro LUMCONVERTRANGE 1
+%macro LUMCONVERTRANGE 2
%ifidni %1,To
-cglobal lumRange%1Jpeg, 5, 5, 6, dst, width, amax, coeff, offset
+%if %2 == 16
+cglobal lumRange%1Jpeg%2, 5, 5, 5, dst, width, amax, coeff, offset
+%elif %2 == 8
+cglobal lumRange%1Jpeg%2, 5, 5, 6, dst, width, amax, coeff, offset
+%endif ; %2 == 8/16
%else
-cglobal lumRange%1Jpeg, 5, 5, 5, dst, width, amax, coeff, offset
+cglobal lumRange%1Jpeg%2, 5, 5, 5, dst, width, amax, coeff, offset
%endif
+%if %2 == 16
+ shl widthd, 2
+%elif %2 == 8
shl widthd, 1
+%endif ; %2 == 8/16
movd xm2, coeffd
VBROADCASTSS m2, xm2
%if ARCH_X86_64
@@ -46,16 +54,40 @@ cglobal lumRange%1Jpeg, 5, 5, 5, dst, width, amax, coeff,
offset
%else
movq xm3, offsetm
%endif
+%if %2 == 16
+ VBROADCASTSD m3, xm3
+%elif %2 == 8
VBROADCASTSS m3, xm3
pxor m4, m4
+%endif ; %2 == 8/16
%ifidni %1,To
+%if %2 == 16
+ movd xm4, amaxd
+ VBROADCASTSS m4, xm4
+%elif %2 == 8
movd xm5, amaxd
SPLATW m5, xm5
+%endif ; %2 == 8/16
%endif
add dstq, widthq
neg widthq
.loop:
movu m0, [dstq+widthq]
+%if %2 == 16
+%ifidni %1,To
+ PMINSD m0, m4, m1
+%endif
+ pshufd m1, m0, 0xb1
+ pmuludq m0, m2
+ pmuludq m1, m2
+ paddq m0, m3
+ paddq m1, m3
+ psrlq m0, 18
+ psrlq m1, 18
+ pshufd m0, m0, 0xd8
+ pshufd m1, m1, 0xd8
+ punpckldq m0, m1
+%elif %2 == 8
%ifidni %1,To
pminsw m0, m5
%endif
@@ -68,6 +100,7 @@ cglobal lumRange%1Jpeg, 5, 5, 5, dst, width, amax, coeff,
offset
psrad m0, 14
psrad m1, 14
packssdw m0, m1
+%endif ; %2 == 8/16
movu [dstq+widthq], m0
add widthq, mmsize
jl .loop
@@ -77,20 +110,28 @@ cglobal lumRange%1Jpeg, 5, 5, 5, dst, width, amax, coeff,
offset
;-----------------------------------------------------------------------------
; chrConvertRange
;
-; 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);
+; void ff_chrRangeToJpeg{8,16}_<opt>(int16_t *dstU, int16_t *dstV, int width,
+; int amax, int coeff, int64_t offset);
+; void ff_chrRangeFromJpeg{8,16}_<opt>(int16_t *dstU, int16_t *dstV, int width,
+; int amax, int coeff, int64_t offset);
;
;-----------------------------------------------------------------------------
-%macro CHRCONVERTRANGE 1
+%macro CHRCONVERTRANGE 2
%ifidni %1,To
-cglobal chrRange%1Jpeg, 6, 6, 8, dstU, dstV, width, amax, coeff, offset
+%if %2 == 16
+cglobal chrRange%1Jpeg%2, 6, 6, 7, dstU, dstV, width, amax, coeff, offset
+%elif %2 == 8
+cglobal chrRange%1Jpeg%2, 6, 6, 8, dstU, dstV, width, amax, coeff, offset
+%endif ; %2 == 8/16
%else
-cglobal chrRange%1Jpeg, 6, 6, 7, dstU, dstV, width, amax, coeff, offset
+cglobal chrRange%1Jpeg%2, 6, 6, 7, dstU, dstV, width, amax, coeff, offset
%endif
+%if %2 == 16
+ shl widthd, 2
+%elif %2 == 8
shl widthd, 1
+%endif ; %2 == 8/16
movd xm4, coeffd
VBROADCASTSS m4, xm4
%if ARCH_X86_64
@@ -98,11 +139,20 @@ cglobal chrRange%1Jpeg, 6, 6, 7, dstU, dstV, width, amax,
coeff, offset
%else
movq xm5, offsetm
%endif
+%if %2 == 16
+ VBROADCASTSD m5, xm5
+%elif %2 == 8
VBROADCASTSS m5, xm5
pxor m6, m6
+%endif ; %2 == 8/16
%ifidni %1,To
+%if %2 == 16
+ movd xm6, amaxd
+ VBROADCASTSS m6, xm6
+%elif %2 == 8
movd xm7, amaxd
SPLATW m7, xm7
+%endif ; %2 == 8/16
%endif
add dstUq, widthq
add dstVq, widthq
@@ -110,6 +160,32 @@ cglobal chrRange%1Jpeg, 6, 6, 7, dstU, dstV, width, amax,
coeff, offset
.loop:
movu m0, [dstUq+widthq]
movu m2, [dstVq+widthq]
+%if %2 == 16
+%ifidni %1,To
+ PMINSD m0, m6, m1
+ PMINSD m2, m6, m3
+%endif
+ pshufd m1, m0, 0xb1
+ pshufd m3, m2, 0xb1
+ pmuludq m0, m4
+ pmuludq m1, m4
+ pmuludq m2, m4
+ pmuludq m3, m4
+ paddq m0, m5
+ paddq m1, m5
+ paddq m2, m5
+ paddq m3, m5
+ psrlq m0, 18
+ psrlq m1, 18
+ psrlq m2, 18
+ psrlq m3, 18
+ pshufd m0, m0, 0xd8
+ pshufd m1, m1, 0xd8
+ pshufd m2, m2, 0xd8
+ pshufd m3, m3, 0xd8
+ punpckldq m0, m1
+ punpckldq m2, m3
+%elif %2 == 8
%ifidni %1,To
pminsw m0, m7
pminsw m2, m7
@@ -132,6 +208,7 @@ cglobal chrRange%1Jpeg, 6, 6, 7, dstU, dstV, width, amax,
coeff, offset
psrad m3, 14
packssdw m0, m1
packssdw m2, m3
+%endif ; %2 == 8/16
movu [dstUq+widthq], m0
movu [dstVq+widthq], m2
add widthq, mmsize
@@ -140,15 +217,27 @@ cglobal chrRange%1Jpeg, 6, 6, 7, dstU, dstV, width, amax,
coeff, offset
%endmacro
INIT_XMM sse2
-LUMCONVERTRANGE To
-CHRCONVERTRANGE To
-LUMCONVERTRANGE From
-CHRCONVERTRANGE From
+LUMCONVERTRANGE To, 8
+LUMCONVERTRANGE To, 16
+CHRCONVERTRANGE To, 8
+CHRCONVERTRANGE To, 16
+LUMCONVERTRANGE From, 8
+LUMCONVERTRANGE From, 16
+CHRCONVERTRANGE From, 8
+CHRCONVERTRANGE From, 16
+
+INIT_XMM sse4
+LUMCONVERTRANGE To, 16
+CHRCONVERTRANGE To, 16
%if HAVE_AVX2_EXTERNAL
INIT_YMM avx2
-LUMCONVERTRANGE To
-CHRCONVERTRANGE To
-LUMCONVERTRANGE From
-CHRCONVERTRANGE From
+LUMCONVERTRANGE To, 8
+LUMCONVERTRANGE To, 16
+CHRCONVERTRANGE To, 8
+CHRCONVERTRANGE To, 16
+LUMCONVERTRANGE From, 8
+LUMCONVERTRANGE From, 16
+CHRCONVERTRANGE From, 8
+CHRCONVERTRANGE From, 16
%endif
diff --git a/libswscale/x86/swscale.c b/libswscale/x86/swscale.c
index 2377365e91..c49f32c5ef 100644
--- a/libswscale/x86/swscale.c
+++ b/libswscale/x86/swscale.c
@@ -454,26 +454,46 @@ INPUT_PLANAR_RGB_A_ALL_DECL(avx2);
#define RANGE_CONVERT_FUNCS(opt) do { \
if (c->dstBpc <= 14) { \
if (c->srcRange) { \
- c->lumConvertRange = ff_lumRangeFromJpeg_ ##opt; \
- c->chrConvertRange = ff_chrRangeFromJpeg_ ##opt; \
+ c->lumConvertRange = ff_lumRangeFromJpeg8_ ##opt; \
+ c->chrConvertRange = ff_chrRangeFromJpeg8_ ##opt; \
} else { \
- c->lumConvertRange = ff_lumRangeToJpeg_ ##opt; \
- c->chrConvertRange = ff_chrRangeToJpeg_ ##opt; \
+ c->lumConvertRange = ff_lumRangeToJpeg8_ ##opt; \
+ c->chrConvertRange = ff_chrRangeToJpeg8_ ##opt; \
+ } \
+ } else { \
+ if (c->srcRange) { \
+ c->lumConvertRange = ff_lumRangeFromJpeg16_ ##opt; \
+ c->chrConvertRange = ff_chrRangeFromJpeg16_ ##opt; \
+ } else { \
+ c->lumConvertRange = ff_lumRangeToJpeg16_ ##opt; \
+ c->chrConvertRange = ff_chrRangeToJpeg16_ ##opt; \
} \
} \
} while (0)
#define RANGE_CONVERT_FUNCS_DECL(opt) \
-void ff_lumRangeFromJpeg_ ##opt(int16_t *dst, int width, \
+void ff_lumRangeFromJpeg8_ ##opt(int16_t *dst, int width, \
+ int amax, int coeff, int64_t offset); \
+void ff_chrRangeFromJpeg8_ ##opt(int16_t *dstU, int16_t *dstV, int width, \
+ int amax, int coeff, int64_t offset); \
+void ff_lumRangeToJpeg8_ ##opt(int16_t *dst, int width, \
+ int amax, int coeff, int64_t offset); \
+void ff_chrRangeToJpeg8_ ##opt(int16_t *dstU, int16_t *dstV, int width, \
+ int amax, int coeff, int64_t offset); \
+void ff_lumRangeFromJpeg16_ ##opt(int16_t *dst, int width, \
+ int amax, int coeff, int64_t offset); \
+void ff_chrRangeFromJpeg16_ ##opt(int16_t *dstU, int16_t *dstV, int width, \
+ int amax, int coeff, int64_t offset); \
+void ff_lumRangeToJpeg16_ ##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, \
+void ff_chrRangeToJpeg16_ ##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);
+void ff_lumRangeToJpeg16_sse4(int16_t *dst, int width,
+ int amax, int coeff, int64_t offset);
+void ff_chrRangeToJpeg16_sse4(int16_t *dstU, int16_t *dstV, int width,
+ int amax, int coeff, int64_t offset);
RANGE_CONVERT_FUNCS_DECL(avx2);
av_cold void ff_sws_init_range_convert_x86(SwsContext *c)
@@ -481,8 +501,14 @@ av_cold void ff_sws_init_range_convert_x86(SwsContext *c)
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);
+ } else {
+ if (EXTERNAL_SSE2(cpu_flags)) {
+ RANGE_CONVERT_FUNCS(sse2);
+ }
+ if (EXTERNAL_SSE4(cpu_flags) && c->dstBpc > 14 && !c->srcRange) {
+ c->lumConvertRange = ff_lumRangeToJpeg16_sse4;
+ c->chrConvertRange = ff_chrRangeToJpeg16_sse4;
+ }
}
}
--
2.30.2
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
To unsubscribe, visit link above, or email
ffmpeg-devel-requ...@ffmpeg.org with subject "unsubscribe".
