You've already forked FFmpeg
mirror of
https://github.com/FFmpeg/FFmpeg.git
synced 2025-10-06 05:47:18 +02:00
avcodec/x86/h264_qpel_8bit: Replace qpel8_h_lowpass_l2 MMXEXT by SSE2
Using xmm registers here is very natural, as it allows to operate on eight words at a time. It also saves 48B here and does not clobber the MMX state. Old benchmarks (only tests affected by the modified function are shown): avg_h264_qpel_8_mc11_8_c: 352.2 ( 1.00x) avg_h264_qpel_8_mc11_8_sse2: 70.4 ( 5.00x) avg_h264_qpel_8_mc11_8_ssse3: 53.9 ( 6.53x) avg_h264_qpel_8_mc13_8_c: 353.3 ( 1.00x) avg_h264_qpel_8_mc13_8_sse2: 72.8 ( 4.86x) avg_h264_qpel_8_mc13_8_ssse3: 53.8 ( 6.57x) avg_h264_qpel_8_mc21_8_c: 404.0 ( 1.00x) avg_h264_qpel_8_mc21_8_sse2: 116.1 ( 3.48x) avg_h264_qpel_8_mc21_8_ssse3: 94.3 ( 4.28x) avg_h264_qpel_8_mc23_8_c: 398.9 ( 1.00x) avg_h264_qpel_8_mc23_8_sse2: 118.6 ( 3.36x) avg_h264_qpel_8_mc23_8_ssse3: 94.8 ( 4.21x) avg_h264_qpel_8_mc31_8_c: 352.7 ( 1.00x) avg_h264_qpel_8_mc31_8_sse2: 71.4 ( 4.94x) avg_h264_qpel_8_mc31_8_ssse3: 53.8 ( 6.56x) avg_h264_qpel_8_mc33_8_c: 354.0 ( 1.00x) avg_h264_qpel_8_mc33_8_sse2: 70.6 ( 5.01x) avg_h264_qpel_8_mc33_8_ssse3: 53.7 ( 6.59x) avg_h264_qpel_16_mc11_8_c: 1417.0 ( 1.00x) avg_h264_qpel_16_mc11_8_sse2: 276.9 ( 5.12x) avg_h264_qpel_16_mc11_8_ssse3: 178.8 ( 7.92x) avg_h264_qpel_16_mc13_8_c: 1427.3 ( 1.00x) avg_h264_qpel_16_mc13_8_sse2: 277.4 ( 5.14x) avg_h264_qpel_16_mc13_8_ssse3: 179.7 ( 7.94x) avg_h264_qpel_16_mc21_8_c: 1634.1 ( 1.00x) avg_h264_qpel_16_mc21_8_sse2: 421.3 ( 3.88x) avg_h264_qpel_16_mc21_8_ssse3: 291.2 ( 5.61x) avg_h264_qpel_16_mc23_8_c: 1627.0 ( 1.00x) avg_h264_qpel_16_mc23_8_sse2: 420.8 ( 3.87x) avg_h264_qpel_16_mc23_8_ssse3: 291.0 ( 5.59x) avg_h264_qpel_16_mc31_8_c: 1418.4 ( 1.00x) avg_h264_qpel_16_mc31_8_sse2: 278.5 ( 5.09x) avg_h264_qpel_16_mc31_8_ssse3: 178.6 ( 7.94x) avg_h264_qpel_16_mc33_8_c: 1407.3 ( 1.00x) avg_h264_qpel_16_mc33_8_sse2: 277.6 ( 5.07x) avg_h264_qpel_16_mc33_8_ssse3: 179.9 ( 7.82x) put_h264_qpel_8_mc11_8_c: 348.1 ( 1.00x) put_h264_qpel_8_mc11_8_sse2: 69.1 ( 5.04x) put_h264_qpel_8_mc11_8_ssse3: 53.8 ( 6.47x) put_h264_qpel_8_mc13_8_c: 349.3 ( 1.00x) put_h264_qpel_8_mc13_8_sse2: 69.7 ( 5.01x) put_h264_qpel_8_mc13_8_ssse3: 53.7 ( 6.51x) put_h264_qpel_8_mc21_8_c: 398.5 ( 1.00x) put_h264_qpel_8_mc21_8_sse2: 115.0 ( 3.46x) put_h264_qpel_8_mc21_8_ssse3: 95.3 ( 4.18x) put_h264_qpel_8_mc23_8_c: 399.9 ( 1.00x) put_h264_qpel_8_mc23_8_sse2: 120.8 ( 3.31x) put_h264_qpel_8_mc23_8_ssse3: 95.4 ( 4.19x) put_h264_qpel_8_mc31_8_c: 350.4 ( 1.00x) put_h264_qpel_8_mc31_8_sse2: 69.6 ( 5.03x) put_h264_qpel_8_mc31_8_ssse3: 54.2 ( 6.47x) put_h264_qpel_8_mc33_8_c: 353.1 ( 1.00x) put_h264_qpel_8_mc33_8_sse2: 71.0 ( 4.97x) put_h264_qpel_8_mc33_8_ssse3: 54.2 ( 6.51x) put_h264_qpel_16_mc11_8_c: 1384.2 ( 1.00x) put_h264_qpel_16_mc11_8_sse2: 272.9 ( 5.07x) put_h264_qpel_16_mc11_8_ssse3: 178.3 ( 7.76x) put_h264_qpel_16_mc13_8_c: 1393.6 ( 1.00x) put_h264_qpel_16_mc13_8_sse2: 271.1 ( 5.14x) put_h264_qpel_16_mc13_8_ssse3: 178.3 ( 7.82x) put_h264_qpel_16_mc21_8_c: 1612.6 ( 1.00x) put_h264_qpel_16_mc21_8_sse2: 416.5 ( 3.87x) put_h264_qpel_16_mc21_8_ssse3: 289.1 ( 5.58x) put_h264_qpel_16_mc23_8_c: 1621.3 ( 1.00x) put_h264_qpel_16_mc23_8_sse2: 416.9 ( 3.89x) put_h264_qpel_16_mc23_8_ssse3: 289.4 ( 5.60x) put_h264_qpel_16_mc31_8_c: 1408.4 ( 1.00x) put_h264_qpel_16_mc31_8_sse2: 273.5 ( 5.15x) put_h264_qpel_16_mc31_8_ssse3: 176.9 ( 7.96x) put_h264_qpel_16_mc33_8_c: 1396.4 ( 1.00x) put_h264_qpel_16_mc33_8_sse2: 276.3 ( 5.05x) put_h264_qpel_16_mc33_8_ssse3: 176.4 ( 7.92x) New benchmarks: avg_h264_qpel_8_mc11_8_c: 352.1 ( 1.00x) avg_h264_qpel_8_mc11_8_sse2: 52.5 ( 6.71x) avg_h264_qpel_8_mc11_8_ssse3: 53.9 ( 6.54x) avg_h264_qpel_8_mc13_8_c: 350.8 ( 1.00x) avg_h264_qpel_8_mc13_8_sse2: 54.7 ( 6.42x) avg_h264_qpel_8_mc13_8_ssse3: 54.3 ( 6.46x) avg_h264_qpel_8_mc21_8_c: 400.1 ( 1.00x) avg_h264_qpel_8_mc21_8_sse2: 98.6 ( 4.06x) avg_h264_qpel_8_mc21_8_ssse3: 95.5 ( 4.19x) avg_h264_qpel_8_mc23_8_c: 400.4 ( 1.00x) avg_h264_qpel_8_mc23_8_sse2: 101.4 ( 3.95x) avg_h264_qpel_8_mc23_8_ssse3: 95.9 ( 4.18x) avg_h264_qpel_8_mc31_8_c: 352.4 ( 1.00x) avg_h264_qpel_8_mc31_8_sse2: 52.9 ( 6.67x) avg_h264_qpel_8_mc31_8_ssse3: 54.4 ( 6.48x) avg_h264_qpel_8_mc33_8_c: 354.5 ( 1.00x) avg_h264_qpel_8_mc33_8_sse2: 52.9 ( 6.70x) avg_h264_qpel_8_mc33_8_ssse3: 54.4 ( 6.52x) avg_h264_qpel_16_mc11_8_c: 1420.4 ( 1.00x) avg_h264_qpel_16_mc11_8_sse2: 204.8 ( 6.93x) avg_h264_qpel_16_mc11_8_ssse3: 177.9 ( 7.98x) avg_h264_qpel_16_mc13_8_c: 1409.8 ( 1.00x) avg_h264_qpel_16_mc13_8_sse2: 206.4 ( 6.83x) avg_h264_qpel_16_mc13_8_ssse3: 178.0 ( 7.92x) avg_h264_qpel_16_mc21_8_c: 1634.1 ( 1.00x) avg_h264_qpel_16_mc21_8_sse2: 349.6 ( 4.67x) avg_h264_qpel_16_mc21_8_ssse3: 290.0 ( 5.63x) avg_h264_qpel_16_mc23_8_c: 1624.1 ( 1.00x) avg_h264_qpel_16_mc23_8_sse2: 350.0 ( 4.64x) avg_h264_qpel_16_mc23_8_ssse3: 291.9 ( 5.56x) avg_h264_qpel_16_mc31_8_c: 1407.2 ( 1.00x) avg_h264_qpel_16_mc31_8_sse2: 205.8 ( 6.84x) avg_h264_qpel_16_mc31_8_ssse3: 178.2 ( 7.90x) avg_h264_qpel_16_mc33_8_c: 1400.5 ( 1.00x) avg_h264_qpel_16_mc33_8_sse2: 206.3 ( 6.79x) avg_h264_qpel_16_mc33_8_ssse3: 179.4 ( 7.81x) put_h264_qpel_8_mc11_8_c: 349.7 ( 1.00x) put_h264_qpel_8_mc11_8_sse2: 50.2 ( 6.96x) put_h264_qpel_8_mc11_8_ssse3: 51.3 ( 6.82x) put_h264_qpel_8_mc13_8_c: 349.8 ( 1.00x) put_h264_qpel_8_mc13_8_sse2: 50.7 ( 6.90x) put_h264_qpel_8_mc13_8_ssse3: 51.7 ( 6.76x) put_h264_qpel_8_mc21_8_c: 398.0 ( 1.00x) put_h264_qpel_8_mc21_8_sse2: 96.5 ( 4.13x) put_h264_qpel_8_mc21_8_ssse3: 92.3 ( 4.31x) put_h264_qpel_8_mc23_8_c: 401.4 ( 1.00x) put_h264_qpel_8_mc23_8_sse2: 102.3 ( 3.92x) put_h264_qpel_8_mc23_8_ssse3: 92.8 ( 4.32x) put_h264_qpel_8_mc31_8_c: 349.4 ( 1.00x) put_h264_qpel_8_mc31_8_sse2: 50.8 ( 6.88x) put_h264_qpel_8_mc31_8_ssse3: 51.8 ( 6.75x) put_h264_qpel_8_mc33_8_c: 351.1 ( 1.00x) put_h264_qpel_8_mc33_8_sse2: 52.2 ( 6.73x) put_h264_qpel_8_mc33_8_ssse3: 51.7 ( 6.79x) put_h264_qpel_16_mc11_8_c: 1391.1 ( 1.00x) put_h264_qpel_16_mc11_8_sse2: 196.6 ( 7.07x) put_h264_qpel_16_mc11_8_ssse3: 178.2 ( 7.81x) put_h264_qpel_16_mc13_8_c: 1385.2 ( 1.00x) put_h264_qpel_16_mc13_8_sse2: 195.6 ( 7.08x) put_h264_qpel_16_mc13_8_ssse3: 176.6 ( 7.84x) put_h264_qpel_16_mc21_8_c: 1607.5 ( 1.00x) put_h264_qpel_16_mc21_8_sse2: 341.0 ( 4.71x) put_h264_qpel_16_mc21_8_ssse3: 289.1 ( 5.56x) put_h264_qpel_16_mc23_8_c: 1616.7 ( 1.00x) put_h264_qpel_16_mc23_8_sse2: 340.8 ( 4.74x) put_h264_qpel_16_mc23_8_ssse3: 288.6 ( 5.60x) put_h264_qpel_16_mc31_8_c: 1397.6 ( 1.00x) put_h264_qpel_16_mc31_8_sse2: 197.3 ( 7.08x) put_h264_qpel_16_mc31_8_ssse3: 175.4 ( 7.97x) put_h264_qpel_16_mc33_8_c: 1394.3 ( 1.00x) put_h264_qpel_16_mc33_8_sse2: 197.7 ( 7.05x) put_h264_qpel_16_mc33_8_ssse3: 175.2 ( 7.96x) As can be seen, the SSE2 version is often neck-to-neck with the SSSE3 version (which also benefits from a better hv2_lowpass SSSE3 implementation for mc21 and mc23) for eight byte block sizes. Unsurprisingly, SSSE3 beats SSE2 for 16x16 blocks: For SSE2, these blocks are processed by calling the 8x8 function four times whereas SSSE3 has a dedicated function (on x64). This implementation should also be extendable to an AVX version for 16x16 blocks. Reviewed-by: James Almer <jamrial@gmail.com> Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
This commit is contained in:
Andreas Rheinhardt
@@ -58,7 +58,7 @@ void ff_avg_pixels16_l2_mmxext(uint8_t *dst, const uint8_t *src1, const uint8_t
|
||||
void ff_ ## OPNAME ## _h264_qpel4_h_lowpass_mmxext(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride);\
|
||||
void ff_ ## OPNAME ## _h264_qpel8_h_lowpass_ssse3(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride);\
|
||||
void ff_ ## OPNAME ## _h264_qpel4_h_lowpass_l2_mmxext(uint8_t *dst, const uint8_t *src, const uint8_t *src2, int dstStride, int src2Stride);\
|
||||
void ff_ ## OPNAME ## _h264_qpel8_h_lowpass_l2_mmxext(uint8_t *dst, const uint8_t *src, const uint8_t *src2, int dstStride, int src2Stride);\
|
||||
void ff_ ## OPNAME ## _h264_qpel8_h_lowpass_l2_sse2(uint8_t *dst, const uint8_t *src, const uint8_t *src2, int dstStride, int src2Stride);\
|
||||
void ff_ ## OPNAME ## _h264_qpel8_h_lowpass_l2_ssse3(uint8_t *dst, const uint8_t *src, const uint8_t *src2, int dstStride, int src2Stride);\
|
||||
void ff_ ## OPNAME ## _h264_qpel4_v_lowpass_mmxext(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride);\
|
||||
void ff_ ## OPNAME ## _h264_qpel8or16_v_lowpass_sse2(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h);\
|
||||
@@ -96,15 +96,16 @@ static av_always_inline void ff_ ## OPNAME ## h264_qpel8or16_hv2_lowpass_ ## MMX
|
||||
dst += 8;\
|
||||
}while(w--);\
|
||||
}\
|
||||
\
|
||||
static av_always_inline void ff_ ## OPNAME ## h264_qpel16_h_lowpass_l2_ ## MMX(uint8_t *dst, const uint8_t *src, const uint8_t *src2, int dstStride, int src2Stride){\
|
||||
ff_ ## OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst , src , src2 , dstStride, src2Stride);\
|
||||
ff_ ## OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst+8, src+8, src2+8, dstStride, src2Stride);\
|
||||
|
||||
#define QPEL_H264_H16(OPNAME, EXT) \
|
||||
static av_always_inline void ff_ ## OPNAME ## h264_qpel16_h_lowpass_l2_ ## EXT(uint8_t *dst, const uint8_t *src, const uint8_t *src2, int dstStride, int src2Stride){\
|
||||
ff_ ## OPNAME ## h264_qpel8_h_lowpass_l2_ ## EXT(dst , src , src2 , dstStride, src2Stride);\
|
||||
ff_ ## OPNAME ## h264_qpel8_h_lowpass_l2_ ## EXT(dst+8, src+8, src2+8, dstStride, src2Stride);\
|
||||
src += 8*dstStride;\
|
||||
dst += 8*dstStride;\
|
||||
src2 += 8*src2Stride;\
|
||||
ff_ ## OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst , src , src2 , dstStride, src2Stride);\
|
||||
ff_ ## OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst+8, src+8, src2+8, dstStride, src2Stride);\
|
||||
ff_ ## OPNAME ## h264_qpel8_h_lowpass_l2_ ## EXT(dst , src , src2 , dstStride, src2Stride);\
|
||||
ff_ ## OPNAME ## h264_qpel8_h_lowpass_l2_ ## EXT(dst+8, src+8, src2+8, dstStride, src2Stride);\
|
||||
}\
|
||||
|
||||
|
||||
@@ -115,16 +116,7 @@ void ff_avg_h264_qpel16_h_lowpass_l2_ssse3(uint8_t *dst, const uint8_t *src, con
|
||||
void ff_put_h264_qpel16_h_lowpass_l2_ssse3(uint8_t *dst, const uint8_t *src, const uint8_t *src2, int dstStride, int src2Stride);
|
||||
|
||||
#else // ARCH_X86_64
|
||||
#define QPEL_H264_H16_XMM(OPNAME, OP, MMX)\
|
||||
static av_always_inline void ff_ ## OPNAME ## h264_qpel16_h_lowpass_l2_ ## MMX(uint8_t *dst, const uint8_t *src, const uint8_t *src2, int dstStride, int src2Stride){\
|
||||
ff_ ## OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst , src , src2 , dstStride, src2Stride);\
|
||||
ff_ ## OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst+8, src+8, src2+8, dstStride, src2Stride);\
|
||||
src += 8*dstStride;\
|
||||
dst += 8*dstStride;\
|
||||
src2 += 8*src2Stride;\
|
||||
ff_ ## OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst , src , src2 , dstStride, src2Stride);\
|
||||
ff_ ## OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst+8, src+8, src2+8, dstStride, src2Stride);\
|
||||
}
|
||||
#define QPEL_H264_H16_XMM(OPNAME, OP, EXT) QPEL_H264_H16(OPNAME, EXT)
|
||||
#endif // ARCH_X86_64
|
||||
|
||||
#define QPEL_H264_H_XMM(OPNAME, OP, MMX)\
|
||||
@@ -185,10 +177,8 @@ ff_ ## OPNAME ## _h264_qpel8or16_hv2_lowpass_ssse3(uint8_t *dst, int16_t *tmp, i
|
||||
SSSE3_HV2_LOWPASS_WRAPPER(avg)
|
||||
SSSE3_HV2_LOWPASS_WRAPPER(put)
|
||||
|
||||
#define ff_put_h264_qpel8_h_lowpass_l2_sse2 ff_put_h264_qpel8_h_lowpass_l2_mmxext
|
||||
#define ff_avg_h264_qpel8_h_lowpass_l2_sse2 ff_avg_h264_qpel8_h_lowpass_l2_mmxext
|
||||
#define ff_put_h264_qpel16_h_lowpass_l2_sse2 ff_put_h264_qpel16_h_lowpass_l2_mmxext
|
||||
#define ff_avg_h264_qpel16_h_lowpass_l2_sse2 ff_avg_h264_qpel16_h_lowpass_l2_mmxext
|
||||
QPEL_H264_H16(avg_, sse2)
|
||||
QPEL_H264_H16(put_, sse2)
|
||||
|
||||
#define ff_put_h264_qpel8_v_lowpass_ssse3 ff_put_h264_qpel8_v_lowpass_sse2
|
||||
#define ff_avg_h264_qpel8_v_lowpass_ssse3 ff_avg_h264_qpel8_v_lowpass_sse2
|
||||
|
||||
@@ -230,56 +230,39 @@ QPEL4_H_LOWPASS_L2_OP avg
|
||||
|
||||
|
||||
%macro QPEL8_H_LOWPASS_L2_OP 1
|
||||
cglobal %1_h264_qpel8_h_lowpass_l2, 5,6 ; dst, src, src2, dstStride, srcStride
|
||||
cglobal %1_h264_qpel8_h_lowpass_l2, 5,6,6 ; dst, src, src2, dstStride, srcStride
|
||||
movsxdifnidn r3, r3d
|
||||
movsxdifnidn r4, r4d
|
||||
mova m3, [pw_16]
|
||||
mov r5d, 8
|
||||
pxor m7, m7
|
||||
mova m6, [pw_5]
|
||||
pxor m5, m5
|
||||
mova m4, [pw_5]
|
||||
.loop:
|
||||
mova m0, [r1]
|
||||
mova m2, [r1+1]
|
||||
mova m1, m0
|
||||
mova m3, m2
|
||||
punpcklbw m0, m7
|
||||
punpckhbw m1, m7
|
||||
punpcklbw m2, m7
|
||||
punpckhbw m3, m7
|
||||
paddw m0, m2
|
||||
paddw m1, m3
|
||||
movh m0, [r1]
|
||||
movh m1, [r1+1]
|
||||
punpcklbw m0, m5
|
||||
punpcklbw m1, m5
|
||||
paddw m0, m1
|
||||
psllw m0, 2
|
||||
psllw m1, 2
|
||||
mova m2, [r1-1]
|
||||
mova m4, [r1+2]
|
||||
mova m3, m2
|
||||
mova m5, m4
|
||||
punpcklbw m2, m7
|
||||
punpckhbw m3, m7
|
||||
punpcklbw m4, m7
|
||||
punpckhbw m5, m7
|
||||
paddw m2, m4
|
||||
paddw m5, m3
|
||||
psubw m0, m2
|
||||
psubw m1, m5
|
||||
pmullw m0, m6
|
||||
pmullw m1, m6
|
||||
movd m2, [r1-2]
|
||||
movd m5, [r1+7]
|
||||
punpcklbw m2, m7
|
||||
punpcklbw m5, m7
|
||||
paddw m2, m3
|
||||
paddw m4, m5
|
||||
mova m5, [pw_16]
|
||||
paddw m2, m5
|
||||
paddw m4, m5
|
||||
movh m1, [r1-1]
|
||||
movh m2, [r1+2]
|
||||
punpcklbw m1, m5
|
||||
punpcklbw m2, m5
|
||||
paddw m1, m2
|
||||
psubw m0, m1
|
||||
pmullw m0, m4
|
||||
movh m1, [r1-2]
|
||||
movh m2, [r1+3]
|
||||
punpcklbw m1, m5
|
||||
punpcklbw m2, m5
|
||||
paddw m0, m1
|
||||
paddw m0, m2
|
||||
paddw m1, m4
|
||||
paddw m0, m3
|
||||
psraw m0, 5
|
||||
psraw m1, 5
|
||||
mova m4, [r2]
|
||||
packuswb m0, m1
|
||||
pavgb m0, m4
|
||||
op_%1 m0, [r0], m4
|
||||
packuswb m0, m5
|
||||
movh m2, [r2]
|
||||
pavgb m0, m2
|
||||
op_%1h m0, [r0], m2
|
||||
add r0, r3
|
||||
add r1, r3
|
||||
add r2, r4
|
||||
@@ -288,7 +271,7 @@ cglobal %1_h264_qpel8_h_lowpass_l2, 5,6 ; dst, src, src2, dstStride, srcStride
|
||||
RET
|
||||
%endmacro
|
||||
|
||||
INIT_MMX mmxext
|
||||
INIT_XMM sse2
|
||||
QPEL8_H_LOWPASS_L2_OP put
|
||||
QPEL8_H_LOWPASS_L2_OP avg
|
||||
|
||||
|
||||
Reference in New Issue
Block a user