X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=modules%2Fcodec%2Fxvmc%2Fmotion_comp_mmx.c;h=5b68a290e452917eb2f90a7bc811b3e67a8bf8a3;hb=6ee1e193fd896ab9a4729fde14f009d9ce629815;hp=bb614931571a37c4bdcc979362f67931a1a8e496;hpb=3305b049e7f587b23359a1c9047fb5763d19c1dc;p=vlc diff --git a/modules/codec/xvmc/motion_comp_mmx.c b/modules/codec/xvmc/motion_comp_mmx.c index bb61493157..5b68a290e4 100644 --- a/modules/codec/xvmc/motion_comp_mmx.c +++ b/modules/codec/xvmc/motion_comp_mmx.c @@ -67,175 +67,175 @@ static inline void mmx_zero_reg () } static inline void mmx_average_2_U8 (uint8_t * dest, const uint8_t * src1, - const uint8_t * src2) + const uint8_t * src2) { /* *dest = (*src1 + *src2 + 1)/ 2; */ - movq_m2r (*src1, mm1); /* load 8 src1 bytes */ - movq_r2r (mm1, mm2); /* copy 8 src1 bytes */ + movq_m2r (*src1, mm1); /* load 8 src1 bytes */ + movq_r2r (mm1, mm2); /* copy 8 src1 bytes */ - movq_m2r (*src2, mm3); /* load 8 src2 bytes */ - movq_r2r (mm3, mm4); /* copy 8 src2 bytes */ + movq_m2r (*src2, mm3); /* load 8 src2 bytes */ + movq_r2r (mm3, mm4); /* copy 8 src2 bytes */ - pxor_r2r (mm1, mm3); /* xor src1 and src2 */ - pand_m2r (mask1, mm3); /* mask lower bits */ - psrlq_i2r (1, mm3); /* /2 */ - por_r2r (mm2, mm4); /* or src1 and src2 */ - psubb_r2r (mm3, mm4); /* subtract subresults */ - movq_r2m (mm4, *dest); /* store result in dest */ + pxor_r2r (mm1, mm3); /* xor src1 and src2 */ + pand_m2r (mask1, mm3); /* mask lower bits */ + psrlq_i2r (1, mm3); /* /2 */ + por_r2r (mm2, mm4); /* or src1 and src2 */ + psubb_r2r (mm3, mm4); /* subtract subresults */ + movq_r2m (mm4, *dest); /* store result in dest */ } static inline void mmx_interp_average_2_U8 (uint8_t * dest, - const uint8_t * src1, - const uint8_t * src2) + const uint8_t * src1, + const uint8_t * src2) { /* *dest = (*dest + (*src1 + *src2 + 1)/ 2 + 1)/ 2; */ - movq_m2r (*dest, mm1); /* load 8 dest bytes */ - movq_r2r (mm1, mm2); /* copy 8 dest bytes */ + movq_m2r (*dest, mm1); /* load 8 dest bytes */ + movq_r2r (mm1, mm2); /* copy 8 dest bytes */ - movq_m2r (*src1, mm3); /* load 8 src1 bytes */ - movq_r2r (mm3, mm4); /* copy 8 src1 bytes */ + movq_m2r (*src1, mm3); /* load 8 src1 bytes */ + movq_r2r (mm3, mm4); /* copy 8 src1 bytes */ - movq_m2r (*src2, mm5); /* load 8 src2 bytes */ - movq_r2r (mm5, mm6); /* copy 8 src2 bytes */ + movq_m2r (*src2, mm5); /* load 8 src2 bytes */ + movq_r2r (mm5, mm6); /* copy 8 src2 bytes */ - pxor_r2r (mm3, mm5); /* xor src1 and src2 */ - pand_m2r (mask1, mm5); /* mask lower bits */ - psrlq_i2r (1, mm5); /* /2 */ - por_r2r (mm4, mm6); /* or src1 and src2 */ - psubb_r2r (mm5, mm6); /* subtract subresults */ - movq_r2r (mm6, mm5); /* copy subresult */ + pxor_r2r (mm3, mm5); /* xor src1 and src2 */ + pand_m2r (mask1, mm5); /* mask lower bits */ + psrlq_i2r (1, mm5); /* /2 */ + por_r2r (mm4, mm6); /* or src1 and src2 */ + psubb_r2r (mm5, mm6); /* subtract subresults */ + movq_r2r (mm6, mm5); /* copy subresult */ - pxor_r2r (mm1, mm5); /* xor srcavg and dest */ - pand_m2r (mask1, mm5); /* mask lower bits */ - psrlq_i2r (1, mm5); /* /2 */ - por_r2r (mm2, mm6); /* or srcavg and dest */ - psubb_r2r (mm5, mm6); /* subtract subresults */ - movq_r2m (mm6, *dest); /* store result in dest */ + pxor_r2r (mm1, mm5); /* xor srcavg and dest */ + pand_m2r (mask1, mm5); /* mask lower bits */ + psrlq_i2r (1, mm5); /* /2 */ + por_r2r (mm2, mm6); /* or srcavg and dest */ + psubb_r2r (mm5, mm6); /* subtract subresults */ + movq_r2m (mm6, *dest); /* store result in dest */ } static inline void mmx_average_4_U8 (uint8_t * dest, const uint8_t * src1, - const uint8_t * src2, - const uint8_t * src3, - const uint8_t * src4) + const uint8_t * src2, + const uint8_t * src3, + const uint8_t * src4) { /* *dest = (*src1 + *src2 + *src3 + *src4 + 2)/ 4; */ - movq_m2r (*src1, mm1); /* load 8 src1 bytes */ - movq_r2r (mm1, mm2); /* copy 8 src1 bytes */ + movq_m2r (*src1, mm1); /* load 8 src1 bytes */ + movq_r2r (mm1, mm2); /* copy 8 src1 bytes */ - punpcklbw_r2r (mm0, mm1); /* unpack low src1 bytes */ - punpckhbw_r2r (mm0, mm2); /* unpack high src1 bytes */ + punpcklbw_r2r (mm0, mm1); /* unpack low src1 bytes */ + punpckhbw_r2r (mm0, mm2); /* unpack high src1 bytes */ - movq_m2r (*src2, mm3); /* load 8 src2 bytes */ - movq_r2r (mm3, mm4); /* copy 8 src2 bytes */ + movq_m2r (*src2, mm3); /* load 8 src2 bytes */ + movq_r2r (mm3, mm4); /* copy 8 src2 bytes */ - punpcklbw_r2r (mm0, mm3); /* unpack low src2 bytes */ - punpckhbw_r2r (mm0, mm4); /* unpack high src2 bytes */ + punpcklbw_r2r (mm0, mm3); /* unpack low src2 bytes */ + punpckhbw_r2r (mm0, mm4); /* unpack high src2 bytes */ - paddw_r2r (mm3, mm1); /* add lows */ - paddw_r2r (mm4, mm2); /* add highs */ + paddw_r2r (mm3, mm1); /* add lows */ + paddw_r2r (mm4, mm2); /* add highs */ /* now have partials in mm1 and mm2 */ - movq_m2r (*src3, mm3); /* load 8 src3 bytes */ - movq_r2r (mm3, mm4); /* copy 8 src3 bytes */ + movq_m2r (*src3, mm3); /* load 8 src3 bytes */ + movq_r2r (mm3, mm4); /* copy 8 src3 bytes */ - punpcklbw_r2r (mm0, mm3); /* unpack low src3 bytes */ - punpckhbw_r2r (mm0, mm4); /* unpack high src3 bytes */ + punpcklbw_r2r (mm0, mm3); /* unpack low src3 bytes */ + punpckhbw_r2r (mm0, mm4); /* unpack high src3 bytes */ - paddw_r2r (mm3, mm1); /* add lows */ - paddw_r2r (mm4, mm2); /* add highs */ + paddw_r2r (mm3, mm1); /* add lows */ + paddw_r2r (mm4, mm2); /* add highs */ - movq_m2r (*src4, mm5); /* load 8 src4 bytes */ - movq_r2r (mm5, mm6); /* copy 8 src4 bytes */ + movq_m2r (*src4, mm5); /* load 8 src4 bytes */ + movq_r2r (mm5, mm6); /* copy 8 src4 bytes */ - punpcklbw_r2r (mm0, mm5); /* unpack low src4 bytes */ - punpckhbw_r2r (mm0, mm6); /* unpack high src4 bytes */ + punpcklbw_r2r (mm0, mm5); /* unpack low src4 bytes */ + punpckhbw_r2r (mm0, mm6); /* unpack high src4 bytes */ - paddw_r2r (mm5, mm1); /* add lows */ - paddw_r2r (mm6, mm2); /* add highs */ + paddw_r2r (mm5, mm1); /* add lows */ + paddw_r2r (mm6, mm2); /* add highs */ /* now have subtotal in mm1 and mm2 */ paddw_m2r (round4, mm1); - psraw_i2r (2, mm1); /* /4 */ + psraw_i2r (2, mm1); /* /4 */ paddw_m2r (round4, mm2); - psraw_i2r (2, mm2); /* /4 */ + psraw_i2r (2, mm2); /* /4 */ - packuswb_r2r (mm2, mm1); /* pack (w/ saturation) */ - movq_r2m (mm1, *dest); /* store result in dest */ + packuswb_r2r (mm2, mm1); /* pack (w/ saturation) */ + movq_r2m (mm1, *dest); /* store result in dest */ } static inline void mmx_interp_average_4_U8 (uint8_t * dest, - const uint8_t * src1, - const uint8_t * src2, - const uint8_t * src3, - const uint8_t * src4) + const uint8_t * src1, + const uint8_t * src2, + const uint8_t * src3, + const uint8_t * src4) { /* *dest = (*dest + (*src1 + *src2 + *src3 + *src4 + 2)/ 4 + 1)/ 2; */ - movq_m2r (*src1, mm1); /* load 8 src1 bytes */ - movq_r2r (mm1, mm2); /* copy 8 src1 bytes */ + movq_m2r (*src1, mm1); /* load 8 src1 bytes */ + movq_r2r (mm1, mm2); /* copy 8 src1 bytes */ - punpcklbw_r2r (mm0, mm1); /* unpack low src1 bytes */ - punpckhbw_r2r (mm0, mm2); /* unpack high src1 bytes */ + punpcklbw_r2r (mm0, mm1); /* unpack low src1 bytes */ + punpckhbw_r2r (mm0, mm2); /* unpack high src1 bytes */ - movq_m2r (*src2, mm3); /* load 8 src2 bytes */ - movq_r2r (mm3, mm4); /* copy 8 src2 bytes */ + movq_m2r (*src2, mm3); /* load 8 src2 bytes */ + movq_r2r (mm3, mm4); /* copy 8 src2 bytes */ - punpcklbw_r2r (mm0, mm3); /* unpack low src2 bytes */ - punpckhbw_r2r (mm0, mm4); /* unpack high src2 bytes */ + punpcklbw_r2r (mm0, mm3); /* unpack low src2 bytes */ + punpckhbw_r2r (mm0, mm4); /* unpack high src2 bytes */ - paddw_r2r (mm3, mm1); /* add lows */ - paddw_r2r (mm4, mm2); /* add highs */ + paddw_r2r (mm3, mm1); /* add lows */ + paddw_r2r (mm4, mm2); /* add highs */ /* now have partials in mm1 and mm2 */ - movq_m2r (*src3, mm3); /* load 8 src3 bytes */ - movq_r2r (mm3, mm4); /* copy 8 src3 bytes */ + movq_m2r (*src3, mm3); /* load 8 src3 bytes */ + movq_r2r (mm3, mm4); /* copy 8 src3 bytes */ - punpcklbw_r2r (mm0, mm3); /* unpack low src3 bytes */ - punpckhbw_r2r (mm0, mm4); /* unpack high src3 bytes */ + punpcklbw_r2r (mm0, mm3); /* unpack low src3 bytes */ + punpckhbw_r2r (mm0, mm4); /* unpack high src3 bytes */ - paddw_r2r (mm3, mm1); /* add lows */ - paddw_r2r (mm4, mm2); /* add highs */ + paddw_r2r (mm3, mm1); /* add lows */ + paddw_r2r (mm4, mm2); /* add highs */ - movq_m2r (*src4, mm5); /* load 8 src4 bytes */ - movq_r2r (mm5, mm6); /* copy 8 src4 bytes */ + movq_m2r (*src4, mm5); /* load 8 src4 bytes */ + movq_r2r (mm5, mm6); /* copy 8 src4 bytes */ - punpcklbw_r2r (mm0, mm5); /* unpack low src4 bytes */ - punpckhbw_r2r (mm0, mm6); /* unpack high src4 bytes */ + punpcklbw_r2r (mm0, mm5); /* unpack low src4 bytes */ + punpckhbw_r2r (mm0, mm6); /* unpack high src4 bytes */ - paddw_r2r (mm5, mm1); /* add lows */ - paddw_r2r (mm6, mm2); /* add highs */ + paddw_r2r (mm5, mm1); /* add lows */ + paddw_r2r (mm6, mm2); /* add highs */ paddw_m2r (round4, mm1); - psraw_i2r (2, mm1); /* /4 */ + psraw_i2r (2, mm1); /* /4 */ paddw_m2r (round4, mm2); - psraw_i2r (2, mm2); /* /4 */ + psraw_i2r (2, mm2); /* /4 */ /* now have subtotal/4 in mm1 and mm2 */ - movq_m2r (*dest, mm3); /* load 8 dest bytes */ - movq_r2r (mm3, mm4); /* copy 8 dest bytes */ + movq_m2r (*dest, mm3); /* load 8 dest bytes */ + movq_r2r (mm3, mm4); /* copy 8 dest bytes */ - packuswb_r2r (mm2, mm1); /* pack (w/ saturation) */ - movq_r2r (mm1,mm2); /* copy subresult */ + packuswb_r2r (mm2, mm1); /* pack (w/ saturation) */ + movq_r2r (mm1,mm2); /* copy subresult */ - pxor_r2r (mm1, mm3); /* xor srcavg and dest */ - pand_m2r (mask1, mm3); /* mask lower bits */ - psrlq_i2r (1, mm3); /* /2 */ - por_r2r (mm2, mm4); /* or srcavg and dest */ - psubb_r2r (mm3, mm4); /* subtract subresults */ - movq_r2m (mm4, *dest); /* store result in dest */ + pxor_r2r (mm1, mm3); /* xor srcavg and dest */ + pand_m2r (mask1, mm3); /* mask lower bits */ + psrlq_i2r (1, mm3); /* /2 */ + por_r2r (mm2, mm4); /* or srcavg and dest */ + psubb_r2r (mm3, mm4); /* subtract subresults */ + movq_r2m (mm4, *dest); /* store result in dest */ } /*-----------------------------------------------------------------------*/ static inline void MC_avg_mmx (const int width, int height, uint8_t * dest, - const uint8_t * ref, const int stride) + const uint8_t * ref, const int stride) { mmx_zero_reg (); @@ -252,13 +252,13 @@ static inline void MC_avg_mmx (const int width, int height, uint8_t * dest, } static void MC_avg_o_16_mmx (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_avg_mmx (16, height, dest, ref, stride); } static void MC_avg_o_8_mmx (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_avg_mmx (8, height, dest, ref, stride); } @@ -266,19 +266,19 @@ static void MC_avg_o_8_mmx (uint8_t * dest, const uint8_t * ref, /*-----------------------------------------------------------------------*/ static inline void MC_put_mmx (const int width, int height, uint8_t * dest, - const uint8_t * ref, const int stride) + const uint8_t * ref, const int stride) { mmx_zero_reg (); do { - movq_m2r (* ref, mm1); /* load 8 ref bytes */ - movq_r2m (mm1,* dest); /* store 8 bytes at curr */ + movq_m2r (* ref, mm1); /* load 8 ref bytes */ + movq_r2m (mm1,* dest); /* store 8 bytes at curr */ if( width == 16 ) { - movq_m2r (* (ref+8), mm1); /* load 8 ref bytes */ - movq_r2m (mm1,* (dest+8)); /* store 8 bytes at curr */ + movq_m2r (* (ref+8), mm1); /* load 8 ref bytes */ + movq_r2m (mm1,* (dest+8)); /* store 8 bytes at curr */ } dest += stride; ref += stride; @@ -286,13 +286,13 @@ static inline void MC_put_mmx (const int width, int height, uint8_t * dest, } static void MC_put_o_16_mmx (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_put_mmx (16, height, dest, ref, stride); } static void MC_put_o_8_mmx (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_put_mmx (8, height, dest, ref, stride); } @@ -301,7 +301,7 @@ static void MC_put_o_8_mmx (uint8_t * dest, const uint8_t * ref, /* Half pixel interpolation in the x direction */ static inline void MC_avg_x_mmx (const int width, int height, uint8_t * dest, - const uint8_t * ref, const int stride) + const uint8_t * ref, const int stride) { mmx_zero_reg (); @@ -318,13 +318,13 @@ static inline void MC_avg_x_mmx (const int width, int height, uint8_t * dest, } static void MC_avg_x_16_mmx (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_avg_x_mmx (16, height, dest, ref, stride); } static void MC_avg_x_8_mmx (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_avg_x_mmx (8, height, dest, ref, stride); } @@ -332,7 +332,7 @@ static void MC_avg_x_8_mmx (uint8_t * dest, const uint8_t * ref, /*-----------------------------------------------------------------------*/ static inline void MC_put_x_mmx (const int width, int height, uint8_t * dest, - const uint8_t * ref, const int stride) + const uint8_t * ref, const int stride) { mmx_zero_reg (); @@ -349,13 +349,13 @@ static inline void MC_put_x_mmx (const int width, int height, uint8_t * dest, } static void MC_put_x_16_mmx (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_put_x_mmx (16, height, dest, ref, stride); } static void MC_put_x_8_mmx (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_put_x_mmx (8, height, dest, ref, stride); } @@ -363,7 +363,7 @@ static void MC_put_x_8_mmx (uint8_t * dest, const uint8_t * ref, /*-----------------------------------------------------------------------*/ static inline void MC_avg_xy_mmx (const int width, int height, uint8_t * dest, - const uint8_t * ref, const int stride) + const uint8_t * ref, const int stride) { const uint8_t * ref_next = ref + stride; @@ -384,13 +384,13 @@ static inline void MC_avg_xy_mmx (const int width, int height, uint8_t * dest, } static void MC_avg_xy_16_mmx (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_avg_xy_mmx (16, height, dest, ref, stride); } static void MC_avg_xy_8_mmx (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_avg_xy_mmx (8, height, dest, ref, stride); } @@ -398,7 +398,7 @@ static void MC_avg_xy_8_mmx (uint8_t * dest, const uint8_t * ref, /*-----------------------------------------------------------------------*/ static inline void MC_put_xy_mmx (const int width, int height, uint8_t * dest, - const uint8_t * ref, const int stride) + const uint8_t * ref, const int stride) { const uint8_t * ref_next = ref + stride; @@ -418,13 +418,13 @@ static inline void MC_put_xy_mmx (const int width, int height, uint8_t * dest, } static void MC_put_xy_16_mmx (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_put_xy_mmx (16, height, dest, ref, stride); } static void MC_put_xy_8_mmx (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_put_xy_mmx (8, height, dest, ref, stride); } @@ -432,7 +432,7 @@ static void MC_put_xy_8_mmx (uint8_t * dest, const uint8_t * ref, /*-----------------------------------------------------------------------*/ static inline void MC_avg_y_mmx (const int width, int height, uint8_t * dest, - const uint8_t * ref, const int stride) + const uint8_t * ref, const int stride) { const uint8_t * ref_next = ref + stride; @@ -452,13 +452,13 @@ static inline void MC_avg_y_mmx (const int width, int height, uint8_t * dest, } static void MC_avg_y_16_mmx (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_avg_y_mmx (16, height, dest, ref, stride); } static void MC_avg_y_8_mmx (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_avg_y_mmx (8, height, dest, ref, stride); } @@ -466,7 +466,7 @@ static void MC_avg_y_8_mmx (uint8_t * dest, const uint8_t * ref, /*-----------------------------------------------------------------------*/ static inline void MC_put_y_mmx (const int width, int height, uint8_t * dest, - const uint8_t * ref, const int stride) + const uint8_t * ref, const int stride) { const uint8_t * ref_next = ref + stride; @@ -486,13 +486,13 @@ static inline void MC_put_y_mmx (const int width, int height, uint8_t * dest, } static void MC_put_y_16_mmx (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_put_y_mmx (16, height, dest, ref, stride); } static void MC_put_y_8_mmx (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_put_y_mmx (8, height, dest, ref, stride); } @@ -502,20 +502,20 @@ MPEG2_MC_EXTERN (mmx) /* CPU_MMXEXT/CPU_3DNOW adaptation layer */ -#define pavg_r2r(src,dest) \ -do { \ - if (cpu == CPU_MMXEXT) \ - pavgb_r2r (src, dest); \ - else \ - pavgusb_r2r (src, dest); \ +#define pavg_r2r(src,dest) \ +do { \ + if (cpu == CPU_MMXEXT) \ + pavgb_r2r (src, dest); \ + else \ + pavgusb_r2r (src, dest); \ } while (0) -#define pavg_m2r(src,dest) \ -do { \ - if (cpu == CPU_MMXEXT) \ - pavgb_m2r (src, dest); \ - else \ - pavgusb_m2r (src, dest); \ +#define pavg_m2r(src,dest) \ +do { \ + if (cpu == CPU_MMXEXT) \ + pavgb_m2r (src, dest); \ + else \ + pavgusb_m2r (src, dest); \ } while (0) @@ -523,7 +523,7 @@ do { \ static inline void MC_put1_8 (int height, uint8_t * dest, const uint8_t * ref, - const int stride) + const int stride) { do { movq_m2r (*ref, mm0); @@ -534,7 +534,7 @@ static inline void MC_put1_8 (int height, uint8_t * dest, const uint8_t * ref, } static inline void MC_put1_16 (int height, uint8_t * dest, const uint8_t * ref, - const int stride) + const int stride) { do { movq_m2r (*ref, mm0); @@ -547,7 +547,7 @@ static inline void MC_put1_16 (int height, uint8_t * dest, const uint8_t * ref, } static inline void MC_avg1_8 (int height, uint8_t * dest, const uint8_t * ref, - const int stride, const int cpu) + const int stride, const int cpu) { do { movq_m2r (*ref, mm0); @@ -559,7 +559,7 @@ static inline void MC_avg1_8 (int height, uint8_t * dest, const uint8_t * ref, } static inline void MC_avg1_16 (int height, uint8_t * dest, const uint8_t * ref, - const int stride, const int cpu) + const int stride, const int cpu) { do { movq_m2r (*ref, mm0); @@ -574,8 +574,8 @@ static inline void MC_avg1_16 (int height, uint8_t * dest, const uint8_t * ref, } static inline void MC_put2_8 (int height, uint8_t * dest, const uint8_t * ref, - const int stride, const int offset, - const int cpu) + const int stride, const int offset, + const int cpu) { do { movq_m2r (*ref, mm0); @@ -587,8 +587,8 @@ static inline void MC_put2_8 (int height, uint8_t * dest, const uint8_t * ref, } static inline void MC_put2_16 (int height, uint8_t * dest, const uint8_t * ref, - const int stride, const int offset, - const int cpu) + const int stride, const int offset, + const int cpu) { do { movq_m2r (*ref, mm0); @@ -603,8 +603,8 @@ static inline void MC_put2_16 (int height, uint8_t * dest, const uint8_t * ref, } static inline void MC_avg2_8 (int height, uint8_t * dest, const uint8_t * ref, - const int stride, const int offset, - const int cpu) + const int stride, const int offset, + const int cpu) { do { movq_m2r (*ref, mm0); @@ -617,8 +617,8 @@ static inline void MC_avg2_8 (int height, uint8_t * dest, const uint8_t * ref, } static inline void MC_avg2_16 (int height, uint8_t * dest, const uint8_t * ref, - const int stride, const int offset, - const int cpu) + const int stride, const int offset, + const int cpu) { do { movq_m2r (*ref, mm0); @@ -637,7 +637,7 @@ static inline void MC_avg2_16 (int height, uint8_t * dest, const uint8_t * ref, static mmx_t mask_one = {0x0101010101010101LL}; static inline void MC_put4_8 (int height, uint8_t * dest, const uint8_t * ref, - const int stride, const int cpu) + const int stride, const int cpu) { movq_m2r (*ref, mm0); movq_m2r (*(ref+1), mm1); @@ -670,13 +670,13 @@ static inline void MC_put4_8 (int height, uint8_t * dest, const uint8_t * ref, movq_r2m (mm0, *dest); dest += stride; - movq_r2r (mm6, mm7); /* unroll ! */ - movq_r2r (mm2, mm0); /* unroll ! */ + movq_r2r (mm6, mm7); /* unroll ! */ + movq_r2r (mm2, mm0); /* unroll ! */ } while (--height); } static inline void MC_put4_16 (int height, uint8_t * dest, const uint8_t * ref, - const int stride, const int cpu) + const int stride, const int cpu) { do { movq_m2r (*ref, mm0); @@ -722,7 +722,7 @@ static inline void MC_put4_16 (int height, uint8_t * dest, const uint8_t * ref, } static inline void MC_avg4_8 (int height, uint8_t * dest, const uint8_t * ref, - const int stride, const int cpu) + const int stride, const int cpu) { do { movq_m2r (*ref, mm0); @@ -751,7 +751,7 @@ static inline void MC_avg4_8 (int height, uint8_t * dest, const uint8_t * ref, } static inline void MC_avg4_16 (int height, uint8_t * dest, const uint8_t * ref, - const int stride, const int cpu) + const int stride, const int cpu) { do { movq_m2r (*ref, mm0); @@ -801,97 +801,97 @@ static inline void MC_avg4_16 (int height, uint8_t * dest, const uint8_t * ref, } static void MC_avg_o_16_mmxext (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_avg1_16 (height, dest, ref, stride, CPU_MMXEXT); } static void MC_avg_o_8_mmxext (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_avg1_8 (height, dest, ref, stride, CPU_MMXEXT); } static void MC_put_o_16_mmxext (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_put1_16 (height, dest, ref, stride); } static void MC_put_o_8_mmxext (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_put1_8 (height, dest, ref, stride); } static void MC_avg_x_16_mmxext (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_avg2_16 (height, dest, ref, stride, 1, CPU_MMXEXT); } static void MC_avg_x_8_mmxext (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_avg2_8 (height, dest, ref, stride, 1, CPU_MMXEXT); } static void MC_put_x_16_mmxext (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_put2_16 (height, dest, ref, stride, 1, CPU_MMXEXT); } static void MC_put_x_8_mmxext (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_put2_8 (height, dest, ref, stride, 1, CPU_MMXEXT); } static void MC_avg_y_16_mmxext (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_avg2_16 (height, dest, ref, stride, stride, CPU_MMXEXT); } static void MC_avg_y_8_mmxext (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_avg2_8 (height, dest, ref, stride, stride, CPU_MMXEXT); } static void MC_put_y_16_mmxext (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_put2_16 (height, dest, ref, stride, stride, CPU_MMXEXT); } static void MC_put_y_8_mmxext (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_put2_8 (height, dest, ref, stride, stride, CPU_MMXEXT); } static void MC_avg_xy_16_mmxext (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_avg4_16 (height, dest, ref, stride, CPU_MMXEXT); } static void MC_avg_xy_8_mmxext (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_avg4_8 (height, dest, ref, stride, CPU_MMXEXT); } static void MC_put_xy_16_mmxext (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_put4_16 (height, dest, ref, stride, CPU_MMXEXT); } static void MC_put_xy_8_mmxext (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_put4_8 (height, dest, ref, stride, CPU_MMXEXT); } @@ -899,97 +899,97 @@ static void MC_put_xy_8_mmxext (uint8_t * dest, const uint8_t * ref, MPEG2_MC_EXTERN (mmxext) static void MC_avg_o_16_3dnow (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_avg1_16 (height, dest, ref, stride, CPU_3DNOW); } static void MC_avg_o_8_3dnow (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_avg1_8 (height, dest, ref, stride, CPU_3DNOW); } static void MC_put_o_16_3dnow (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_put1_16 (height, dest, ref, stride); } static void MC_put_o_8_3dnow (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_put1_8 (height, dest, ref, stride); } static void MC_avg_x_16_3dnow (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_avg2_16 (height, dest, ref, stride, 1, CPU_3DNOW); } static void MC_avg_x_8_3dnow (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_avg2_8 (height, dest, ref, stride, 1, CPU_3DNOW); } static void MC_put_x_16_3dnow (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_put2_16 (height, dest, ref, stride, 1, CPU_3DNOW); } static void MC_put_x_8_3dnow (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_put2_8 (height, dest, ref, stride, 1, CPU_3DNOW); } static void MC_avg_y_16_3dnow (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_avg2_16 (height, dest, ref, stride, stride, CPU_3DNOW); } static void MC_avg_y_8_3dnow (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_avg2_8 (height, dest, ref, stride, stride, CPU_3DNOW); } static void MC_put_y_16_3dnow (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_put2_16 (height, dest, ref, stride, stride, CPU_3DNOW); } static void MC_put_y_8_3dnow (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_put2_8 (height, dest, ref, stride, stride, CPU_3DNOW); } static void MC_avg_xy_16_3dnow (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_avg4_16 (height, dest, ref, stride, CPU_3DNOW); } static void MC_avg_xy_8_3dnow (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_avg4_8 (height, dest, ref, stride, CPU_3DNOW); } static void MC_put_xy_16_3dnow (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_put4_16 (height, dest, ref, stride, CPU_3DNOW); } static void MC_put_xy_8_3dnow (uint8_t * dest, const uint8_t * ref, - int stride, int height) + int stride, int height) { MC_put4_8 (height, dest, ref, stride, CPU_3DNOW); }