@@ -25,6 +25,7 @@ SECTION_RODATA 32
cextern pd_1
pd_2: times 8 dd 2
+pd_8: times 8 dd 8
SECTION .text
@@ -246,7 +247,44 @@ RET
%endmacro
+%macro DD97_VERTICAL_HI 0
+
+cglobal dd97_vertical_hi, 6, 6, 8, b0, b1, b2, b3, b4, w
+ mova m7, [pd_8]
+ shl wd, 2
+ add b0q, wq
+ add b1q, wq
+ add b2q, wq
+ add b3q, wq
+ add b4q, wq
+ neg wq
+
+ ALIGN 16
+ .loop:
+ mova m0, [b0q + wq]
+ mova m1, [b1q + wq]
+ mova m2, [b2q + wq]
+ mova m3, [b3q + wq]
+ mova m4, [b4q + wq]
+ pslld m5, m1, 3
+ pslld m6, m3, 3
+ paddd m5, m1
+ paddd m6, m3
+ psubd m5, m0
+ psubd m6, m4
+ paddd m5, m7
+ paddd m5, m6
+ psrad m5, 4
+ paddd m2, m5
+ mova [b2q + wq], m2
+ add wq, mmsize
+ jl .loop
+RET
+
+%endmacro
+
INIT_XMM sse2
+DD97_VERTICAL_HI
HAAR_HORIZONTAL
HAAR_VERTICAL
LEGALL53_VERTICAL_HI
@@ -257,6 +295,7 @@ HAAR_HORIZONTAL
HAAR_VERTICAL
INIT_YMM avx2
+DD97_VERTICAL_HI
HAAR_HORIZONTAL
HAAR_VERTICAL
LEGALL53_VERTICAL_HI
@@ -23,6 +23,9 @@
#include "libavutil/x86/cpu.h"
#include "libavcodec/dirac_dwt.h"
+void ff_dd97_vertical_hi_sse2(int32_t *b0, int32_t *b1, int32_t *b2, int32_t *b3, int32_t *b4, int width);
+void ff_dd97_vertical_hi_avx2(int32_t *b0, int32_t *b1, int32_t *b2, int32_t *b3, int32_t *b4, int width);
+
void ff_legall53_vertical_hi_sse2(int32_t *b0, int32_t *b1, int32_t *b2, int width);
void ff_legall53_vertical_lo_sse2(int32_t *b0, int32_t *b1, int32_t *b2, int width);
void ff_legall53_vertical_hi_avx2(int32_t *b0, int32_t *b1, int32_t *b2, int width);
@@ -36,6 +39,24 @@ void ff_vertical_compose_haar_10bit_sse2(int32_t *b0, int32_t *b1, int width_ali
void ff_vertical_compose_haar_10bit_avx(int32_t *b0, int32_t *b1, int width_align);
void ff_vertical_compose_haar_10bit_avx2(int32_t *b0, int32_t *b1, int width_align);
+static void dd97_vertical_hi_sse2(int32_t *b0, int32_t *b1, int32_t *b2,
+ int32_t *b3, int32_t *b4, int width)
+{
+ int i = width & ~3;
+ ff_dd97_vertical_hi_sse2(b0, b1, b2, b3, b4, i);
+ for(; i<width; i++)
+ b2[i] = COMPOSE_DD97iH0(b0[i], b1[i], b2[i], b3[i], b4[i]);
+}
+
+static void dd97_vertical_hi_avx2(int32_t *b0, int32_t *b1, int32_t *b2,
+ int32_t *b3, int32_t *b4, int width)
+{
+ int i = width & ~7;
+ ff_dd97_vertical_hi_avx2(b0, b1, b2, b3, b4, i);
+ for(; i<width; i++)
+ b2[i] = COMPOSE_DD97iH0(b0[i], b1[i], b2[i], b3[i], b4[i]);
+}
+
av_cold void ff_spatial_idwt_init_10bit_x86(DWTContext *d, enum dwt_type type)
{
#if HAVE_X86ASM
@@ -43,6 +64,10 @@ av_cold void ff_spatial_idwt_init_10bit_x86(DWTContext *d, enum dwt_type type)
if (EXTERNAL_SSE2(cpu_flags)) {
switch (type) {
+ case DWT_DIRAC_DD9_7:
+ d->vertical_compose_h0 = (void*)dd97_vertical_hi_sse2;
+ d->vertical_compose_l0 = (void*)ff_legall53_vertical_lo_sse2;
+ break;
case DWT_DIRAC_LEGALL5_3:
d->vertical_compose_h0 = (void*)ff_legall53_vertical_hi_sse2;
d->vertical_compose_l0 = (void*)ff_legall53_vertical_lo_sse2;
@@ -71,6 +96,10 @@ av_cold void ff_spatial_idwt_init_10bit_x86(DWTContext *d, enum dwt_type type)
if (EXTERNAL_AVX2(cpu_flags)) {
switch (type) {
+ case DWT_DIRAC_DD9_7:
+ d->vertical_compose_h0 = (void*)dd97_vertical_hi_avx2;
+ d->vertical_compose_l0 = (void*)ff_legall53_vertical_lo_avx2;
+ break;
case DWT_DIRAC_LEGALL5_3:
d->vertical_compose_h0 = (void*)ff_legall53_vertical_hi_avx2;
d->vertical_compose_l0 = (void*)ff_legall53_vertical_lo_avx2;
Speed of ffmpeg when decoding a 720p yuv422p10 file encoded with the relevant transform. C: 84fps SSE2: 111fps AVX2: 115fps dd97 vertical hi sse2: 2.77x faster (31773 vs. 11457 decicycles) compared with C avx2: 3.83x faster (31773 vs. 8297 decicycles) compared with C --- libavcodec/x86/dirac_dwt_10bit.asm | 39 +++++++++++++++++++++++++++ libavcodec/x86/dirac_dwt_init_10bit.c | 29 ++++++++++++++++++++ 2 files changed, 68 insertions(+)