static inline void mmx_interp_average_2_U8 (uint8_t * dest,
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 (*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 */
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 */
}
static inline void MC_avg4_8 (int height, uint8_t * dest, const uint8_t * ref,
const int stride, const int cpu)
{
do {
movq_m2r (*ref, mm0);
movq_m2r (*(ref+stride+1), mm1);
movq_r2r (mm0, mm7);
movq_m2r (*(ref+1), mm2);
pxor_r2r (mm1, mm7);
movq_m2r (*(ref+stride), mm3);
movq_r2r (mm2, mm6);
pxor_r2r (mm3, mm6);
pavg_r2r (mm1, mm0);
pavg_r2r (mm3, mm2);
por_r2r (mm6, mm7);
movq_r2r (mm0, mm6);
pxor_r2r (mm2, mm6);
pand_r2r (mm6, mm7);
pand_m2r (mask_one, mm7);
pavg_r2r (mm2, mm0);
psubusb_r2r (mm7, mm0);
movq_m2r (*dest, mm1);
pavg_r2r (mm1, mm0);
ref += stride;
movq_r2m (mm0, *dest);
dest += stride;
} while (--height);
}
int main(int ac, char **av)
{
int i, j, k, n;
unsigned char dat0[8] = { 0x01, 0xf2, 0x03, 0x04, 0x05, 0x06, 0xf7, 0x08 };
long long *datp = (long long *)&dat0;
int16_t dat1[8] = { 0x10, 0x20, -0x130, -0x140, 0x50, -0x160, -0x170, 0x80 };
volatile uint8_t *rfp = dat0;
volatile int16_t *bp = dat1;
unsigned char ans1[8], ans2[8];
n = 0;
for( i=-32768; i<32768; ++i ) {
j = 0;
while( j < 256 ) {
for( k=0; k<8; ++k ) {
dat0[k] = i;
dat1[k] = j++;
}
movq_m2r(m_(&rfp[0]),mm1); /* rfp[0..7] */
pxor_r2r(mm3,mm3);
pxor_r2r(mm4,mm4);
movq_m2r(m_(&bp[0]),mm5); /* bp[0..3] */
movq_r2r(mm1,mm2);
movq_m2r(m_(&bp[4]),mm6); /* bp[4..7] */
punpcklbw_r2r(mm3,mm1); /* rfp[0,2,4,6] */
punpckhbw_r2r(mm3,mm2); /* rfp[1,3,5,7] */
paddsw_r2r(mm5,mm1); /* bp[0..3] */
paddsw_r2r(mm6,mm2); /* bp[4..7] */
pcmpgtw_r2r(mm1,mm3);
pcmpgtw_r2r(mm2,mm4);
pandn_r2r(mm1,mm3);
pandn_r2r(mm2,mm4);
packuswb_r2r(mm4,mm3);
movq_r2m(mm3,m_(&ans1[0]));
emms();
ans2[0] = clip(bp[0] + rfp[0]);
ans2[1] = clip(bp[1] + rfp[1]);
ans2[2] = clip(bp[2] + rfp[2]);
ans2[3] = clip(bp[3] + rfp[3]);
ans2[4] = clip(bp[4] + rfp[4]);
ans2[5] = clip(bp[5] + rfp[5]);
ans2[6] = clip(bp[6] + rfp[6]);
ans2[7] = clip(bp[7] + rfp[7]);
if( *(uint64_t *)&ans1[0] != *(uint64_t *)&ans2[0] )
{
printf(" i=%5d %02x %02x %02x %02x %02x %02x %02x %02x\n", i,
ans1[0], ans1[1], ans1[2], ans1[3], ans1[4], ans1[5], ans1[6], ans1[7]);
printf(" j=%5d %02x %02x %02x %02x %02x %02x %02x %02x\n", j,
ans2[0], ans2[1], ans2[2], ans2[3], ans2[4], ans2[5], ans2[6], ans2[7]);
// exit(0);
}
n += 8;
}
}
printf("n=%d\n",n);
return 0;
}
static void scale_uint8_x_4_x_generic_mmx(gavl_video_scale_context_t * ctx, int scanline, uint8_t * dest_start)
{
int i, j;
uint8_t * src, * dst, *src_start;
int32_t * factors;
// mmx_t tmp_mm;
/*
* mm0: Input
* mm1: factor_mask
* mm2: Factor
* mm3: Output
* mm4:
* mm5:
* mm6: 0
* mm7: scratch
*
*/
src_start = ctx->src + scanline * ctx->src_stride;
pxor_r2r(mm6, mm6);
movq_m2r(factor_mask, mm1);
dst = dest_start;
for(i = 0; i < ctx->dst_size; i++)
{
src = src_start + 4*ctx->table_h.pixels[i].index;
factors = ctx->table_h.pixels[i].factor_i;
pxor_r2r(mm3, mm3);
for(j = 0; j < ctx->table_h.factors_per_pixel; j++)
{
/* Load pixels */
movd_m2r(*(src), mm0);
punpcklbw_r2r(mm6, mm0);
psllw_i2r(7, mm0);
/* Load factors */
LOAD_FACTOR_1_4;
/* Multiply */
pmulhw_r2r(mm7, mm0);
paddw_r2r(mm0, mm3);
// DUMP_MM("mm3_2", mm3);
src += 4;
factors++;
}
psraw_i2r(5, mm3);
packuswb_r2r(mm6, mm3);
movd_r2m(mm3, *dst);
dst+=4;
}
ctx->need_emms = 1;
}
static __inline__ int qblock_sad_mmxe(uint8_t *refblk,
uint32_t h,
uint32_t rowstride)
{
int res;
pxor_r2r (mm4,mm4);
movq_r2r (mm0,mm5); /* First row */
movd_m2r (*refblk, mm6);
pxor_r2r ( mm7, mm7);
refblk += rowstride;
punpcklbw_r2r ( mm7, mm5);
punpcklbw_r2r ( mm7, mm6);
psadbw_r2r ( mm5, mm6);
paddw_r2r ( mm6, mm4 );
movq_r2r (mm1,mm5); /* Second row */
movd_m2r (*refblk, mm6);
refblk += rowstride;
punpcklbw_r2r ( mm7, mm5);
punpcklbw_r2r ( mm7, mm6);
psadbw_r2r ( mm5, mm6);
paddw_r2r ( mm6, mm4 );
if( h == 4 )
{
movq_r2r (mm2,mm5); /* Third row */
movd_m2r (*refblk, mm6);
refblk += rowstride;
punpcklbw_r2r ( mm7, mm5);
punpcklbw_r2r ( mm7, mm6);
psadbw_r2r ( mm5, mm6);
paddw_r2r ( mm6, mm4 );
movq_r2r (mm3,mm5); /* Fourth row */
movd_m2r (*refblk, mm6);
punpcklbw_r2r ( mm7, mm5);
punpcklbw_r2r ( mm7, mm6);
psadbw_r2r ( mm5, mm6);
paddw_r2r ( mm6, mm4 );
}
movd_r2m ( mm4, res );
return res;
}
static inline void mmx_unpack_16rgb (uint8_t * image, const int cpu)
{
static mmx_t mmx_bluemask = {0xf8f8f8f8f8f8f8f8LL};
static mmx_t mmx_greenmask = {0xfcfcfcfcfcfcfcfcLL};
static mmx_t mmx_redmask = {0xf8f8f8f8f8f8f8f8LL};
/*
* convert RGB plane to RGB 16 bits
* mm0 -> B, mm1 -> R, mm2 -> G
* mm4 -> GB, mm5 -> AR pixel 4-7
* mm6 -> GB, mm7 -> AR pixel 0-3
*/
pand_m2r (mmx_bluemask, mm0); /* mm0 = b7b6b5b4b3______ */
pand_m2r (mmx_greenmask, mm2); /* mm2 = g7g6g5g4g3g2____ */
pand_m2r (mmx_redmask, mm1); /* mm1 = r7r6r5r4r3______ */
psrlq_i2r (3, mm0); /* mm0 = ______b7b6b5b4b3 */
pxor_r2r (mm4, mm4); /* mm4 = 0 */
movq_r2r (mm0, mm5); /* mm5 = ______b7b6b5b4b3 */
movq_r2r (mm2, mm7); /* mm7 = g7g6g5g4g3g2____ */
punpcklbw_r2r (mm4, mm2);
punpcklbw_r2r (mm1, mm0);
psllq_i2r (3, mm2);
por_r2r (mm2, mm0);
movntq (mm0, *image);
punpckhbw_r2r (mm4, mm7);
punpckhbw_r2r (mm1, mm5);
psllq_i2r (3, mm7);
por_r2r (mm7, mm5);
movntq (mm5, *(image+8));
}
static inline void mmx_end(uint8_t *src3, uint8_t *src5,
uint8_t *dst, int X)
{
punpcklbw_m2r (mm_cpool[0], mm4);
punpckhbw_m2r (mm_cpool[0], mm5);
psubusw_r2r (mm2, mm0);
psubusw_r2r (mm3, mm1);
movq_m2r (src5[X], mm2);
movq_m2r (src5[X], mm3);
punpcklbw_m2r (mm_cpool[0], mm2);
punpckhbw_m2r (mm_cpool[0], mm3);
psubusw_r2r (mm2, mm0);
psubusw_r2r (mm3, mm1);
psrlw_i2r (3, mm0);
psrlw_i2r (3, mm1);
psubw_r2r (mm6, mm4);
psubw_r2r (mm7, mm5);
packuswb_r2r (mm1,mm0);
movq_r2r (mm4, mm6);
movq_r2r (mm5, mm7);
pcmpgtw_m2r (mm_lthr, mm4);
pcmpgtw_m2r (mm_lthr, mm5);
pcmpgtw_m2r (mm_hthr, mm6);
pcmpgtw_m2r (mm_hthr, mm7);
packsswb_r2r (mm5, mm4);
packsswb_r2r (mm7, mm6);
pxor_r2r (mm6, mm4);
movq_r2r (mm4, mm5);
pandn_r2r (mm0, mm4);
pand_m2r (src3[X], mm5);
por_r2r (mm4, mm5);
movq_r2m (mm5, dst[X]);
}
static void frame_i2f_sse(u_char *src,float *dst,int l)
{
int i;
pxor_r2r(mm7,mm7);
for( i=0; i<l; i+=8 ) {
movq_m2r(*src,mm0);
movq_r2r(mm0, mm2);
punpcklbw_r2r(mm7, mm0);
punpckhbw_r2r(mm7, mm2);
movq_r2r(mm0, mm1);
movq_r2r(mm2, mm3);
punpcklwd_r2r(mm7, mm0);
punpckhwd_r2r(mm7, mm1);
punpcklwd_r2r(mm7, mm2);
punpckhwd_r2r(mm7, mm3);
cvtpi2ps_r2r(mm0,xmm0);
cvtpi2ps_r2r(mm1,xmm1);
cvtpi2ps_r2r(mm2,xmm2);
cvtpi2ps_r2r(mm3,xmm3);
movlps_r2m(xmm0,dst[0]);
movlps_r2m(xmm1,dst[2]);
movlps_r2m(xmm2,dst[4]);
movlps_r2m(xmm3,dst[6]);
src+=8;
dst+=8;
}
emms();
}
static void
scale_uint8_x_1_x_bicubic_noclip_mmx(gavl_video_scale_context_t * ctx, int scanline, uint8_t * dest_start)
{
int i;
uint8_t * src, * dst, *src_start;
int32_t * factors;
mmx_t tmp_mm;
// fprintf(stderr, "scale_uint8_x_1_x_bicubic_noclip_mmx\n");
src_start = ctx->src + scanline * ctx->src_stride;
pxor_r2r(mm6, mm6);
dst = dest_start;
for(i = 0; i < ctx->dst_size; i++)
{
src = src_start + ctx->table_h.pixels[i].index;
factors = ctx->table_h.pixels[i].factor_i;
/* Load pixels */
movd_m2r(*(src), mm0);
punpcklbw_r2r(mm6, mm0);
/* Load factors */
movq_m2r(*factors, mm2);
movq_m2r(*(factors+2), mm3);
packssdw_r2r(mm3, mm2);
/* Multiply */
pmaddwd_r2r(mm2, mm0);
psrld_i2r(14, mm0);
MOVQ_R2M(mm0, tmp_mm);
*(dst++) = tmp_mm.d[0] + tmp_mm.d[1];
}
ctx->need_emms = 1;
}
static inline void mmx_unpack_32rgb (uint8_t * image, const int cpu)
{
/*
* convert RGB plane to RGB packed format,
* mm0 -> B, mm1 -> R, mm2 -> G, mm3 -> 0,
* mm4 -> GB, mm5 -> AR pixel 4-7,
* mm6 -> GB, mm7 -> AR pixel 0-3
*/
pxor_r2r (mm3, mm3);
movq_r2r (mm0, mm6);
movq_r2r (mm1, mm7);
movq_r2r (mm0, mm4);
movq_r2r (mm1, mm5);
punpcklbw_r2r (mm2, mm6);
punpcklbw_r2r (mm3, mm7);
punpcklwd_r2r (mm7, mm6);
movntq (mm6, *image);
movq_r2r (mm0, mm6);
punpcklbw_r2r (mm2, mm6);
punpckhwd_r2r (mm7, mm6);
movntq (mm6, *(image+8));
punpckhbw_r2r (mm2, mm4);
punpckhbw_r2r (mm3, mm5);
punpcklwd_r2r (mm5, mm4);
movntq (mm4, *(image+16));
movq_r2r (mm0, mm4);
punpckhbw_r2r (mm2, mm4);
punpckhwd_r2r (mm5, mm4);
movntq (mm4, *(image+24));
}
/* For a 16*h block, this computes
(((((*pf + *pf2 + 1)>>1) + ((*pb + *pb2 + 1)>>1) + 1)>>1) + *p2 + 1)>>1
*/
static int bsad_0quad_mmxe(uint8_t *pf,uint8_t *pf2,uint8_t *pb,uint8_t *pb2,uint8_t *p2,int lx,int h)
{
int32_t s=0;
pxor_r2r(mm7, mm7);
do {
movq_m2r(pf2[0],mm0);
movq_m2r(pf2[8],mm2);
movq_m2r(pb2[0],mm1);
movq_m2r(pb2[8],mm3);
pavgb_m2r(pf[0],mm0);
pavgb_m2r(pf[8],mm2);
pavgb_m2r(pb[0],mm1);
pavgb_m2r(pb[8],mm3);
pavgb_r2r(mm1,mm0);
pavgb_r2r(mm3,mm2);
psadbw_m2r(p2[0],mm0);
psadbw_m2r(p2[8],mm2);
paddd_r2r(mm0,mm7);
paddd_r2r(mm2,mm7);
pf+=lx;
pf2+=lx;
pb+=lx;
pb2+=lx;
p2+=lx;
h--;
} while (h);
movd_r2g(mm7,s);
emms();
return s;
}
void
evas_common_cpu_mmx_test(void)
{
#ifdef BUILD_MMX
pxor_r2r(mm4, mm4);
#endif
}
static void
_op_mul_p_mas_dp_mmx(DATA32 *s, DATA8 *m, DATA32 c, DATA32 *d, int l) {
DATA32 *e = d + l;
MOV_A2R(ALPHA_255, mm5)
pxor_r2r(mm0, mm0);
while (d < e) {
c = *m;
switch(c)
{
case 0:
break;
case 255:
MOV_P2R(*d, mm1, mm0)
MOV_P2R(*s, mm2, mm0)
MUL4_SYM_R2R(mm2, mm1, mm5)
MOV_R2P(mm1, *d, mm0)
break;
default:
c++;
MOV_A2R(c, mm1)
c = ~(*s);
MOV_P2R(c, mm3, mm0)
MUL4_256_R2R(mm3, mm1)
movq_r2r(mm5, mm4);
psubw_r2r(mm1, mm4);
MOV_P2R(*d, mm1, mm0)
MUL4_SYM_R2R(mm4, mm1, mm5)
MOV_R2P(mm1, *d, mm0)
break;
}
s++; m++; d++;
}
}
static void scale_uint8_x_4_x_bilinear_mmx(gavl_video_scale_context_t * ctx, int scanline, uint8_t * dest_start)
{
int i;
uint8_t * src, * dst, *src_start;
int32_t * factors;
// mmx_t tmp_mm;
/*
* mm0: Input1
* mm1: Factor mask
* mm2:
* mm3: Output
* mm4:
* mm5: Input2
* mm6: 0
* mm7: Factor
*
*/
// fprintf(stderr, "scale_uint8_x_4_x_bilinear_mmx\n");
src_start = ctx->src + scanline * ctx->src_stride;
pxor_r2r(mm6, mm6);
movq_m2r(factor_mask, mm1);
dst = dest_start;
for(i = 0; i < ctx->dst_size; i++)
{
src = src_start + 4*ctx->table_h.pixels[i].index;
factors = ctx->table_h.pixels[i].factor_i;
/* Load pixels */
movd_m2r(*(src), mm0);
punpcklbw_r2r(mm6, mm0);
psllw_i2r(6, mm0); /* 14 bit */
/* Load pixels */
movd_m2r(*(src+4), mm5);
punpcklbw_r2r(mm6, mm5);
psllw_i2r(6, mm5); /* 14 bit */
/* Load factors */
LOAD_FACTOR_1_4_NOCLIP; /* 14 bit */
/* Subtract */
psubsw_r2r(mm5, mm0); /* s1(mm0) - s2(mm5) -> mm0 (14 bit) */
pmulhw_r2r(mm7, mm0); /* factor * (s2 - s1) -> mm0 (12 bit) */
psllw_i2r(2, mm0); /* (14 bit) */
paddsw_r2r(mm5, mm0);/* (15 bit) */
psraw_i2r(6, mm0);/* (8 bit) */
packuswb_r2r(mm6, mm0);
movd_r2m(mm0, *dst);
dst+=4;
}
ctx->need_emms = 1;
}
static inline void mean8(unsigned char *refpix,unsigned char *pixel,int radius_count,int row_stride,int threshold,int8_t *diff,unsigned char *count)
{
int a,b;
pxor_r2r(mm6,mm6); // mm6 (aka count) = 0
pxor_r2r(mm7,mm7); // mm7 (aka diff) = 0
movq_m2r(*refpix,mm3); // mm3 = refpix[0]
movd_g2r(0x80808080,mm4); // mm4 = 128
punpcklbw_r2r(mm4,mm4);
pxor_r2r(mm4,mm3); // mm3 = refpix[0]-128
movd_g2r(threshold,mm5); // mm5 = threshold
punpcklbw_r2r(mm5,mm5);
punpcklbw_r2r(mm5,mm5);
punpcklbw_r2r(mm5,mm5);
for( b=0; b<radius_count; b++ ) {
for( a=0; a<radius_count; a++ ) {
movq_m2r(*pixel,mm0); // mm0 = pixel[0]
pxor_r2r(mm4,mm0); // mm0 = pixel[0]-128
movq_r2r(mm3,mm2); // mm2 = refpix[0]-128
psubsb_r2r(mm0,mm2); // mm2 = refpix[0]-pixel[0]
psubsb_r2r(mm3,mm0); // mm0 = pixel[0]-refpix[0]
pminub_r2r(mm0,mm2); // mm2 = abs(pixel[0]-refpix[0])
movq_r2r(mm5,mm1); // mm1 = threshold
pcmpgtb_r2r(mm2,mm1); // mm1 = (threshold > abs(pixel[0]-refpix[0])) ? -1 : 0
psubb_r2r(mm1,mm6); // mm6 += (threshold > abs(pixel[0]-refpix[0]))
pand_r2r(mm1,mm0); // mm0 = (threshold > abs(pixel[0]-refpix[0])) ? pixel[0]-refpix[0] : 0
paddb_r2r(mm0,mm7); // mm7 += (threshold > abs(pixel[0]-refpix[0])) ? pixel[0]-refpix[0] : 0
++pixel;
}
pixel += row_stride - radius_count;
}
movq_r2m(mm6,*count);
movq_r2m(mm7,*diff);
emms();
}
static inline void MC_put4_8 (int height, uint8_t * dest, uint8_t * ref,
int stride, int cpu)
{
movq_m2r (*ref, mm0);
movq_m2r (*(ref+1), mm1);
movq_r2r (mm0, mm7);
pxor_r2r (mm1, mm7);
pavg_r2r (mm1, mm0);
ref += stride;
do {
movq_m2r (*ref, mm2);
movq_r2r (mm0, mm5);
movq_m2r (*(ref+1), mm3);
movq_r2r (mm2, mm6);
pxor_r2r (mm3, mm6);
pavg_r2r (mm3, mm2);
por_r2r (mm6, mm7);
pxor_r2r (mm2, mm5);
pand_r2r (mm5, mm7);
pavg_r2r (mm2, mm0);
pand_m2r (mask_one, mm7);
psubusb_r2r (mm7, mm0);
ref += stride;
movq_r2m (mm0, *dest);
dest += stride;
movq_r2r (mm6, mm7); // unroll !
movq_r2r (mm2, mm0); // unroll !
} while (--height);
}
static void
deinterlace_line_mmx (uint8_t * dst, uint8_t * lum_m4,
uint8_t * lum_m3, uint8_t * lum_m2,
uint8_t * lum_m1, uint8_t * lum, int size)
{
mmx_t rounder;
rounder.uw[0] = 4;
rounder.uw[1] = 4;
rounder.uw[2] = 4;
rounder.uw[3] = 4;
pxor_r2r (mm7, mm7);
movq_m2r (rounder, mm6);
for (; size > 3; size -= 4) {
movd_m2r (*lum_m4, mm0);
movd_m2r (*lum_m3, mm1);
movd_m2r (*lum_m2, mm2);
movd_m2r (*lum_m1, mm3);
movd_m2r (*lum, mm4);
punpcklbw_r2r (mm7, mm0);
punpcklbw_r2r (mm7, mm1);
punpcklbw_r2r (mm7, mm2);
punpcklbw_r2r (mm7, mm3);
punpcklbw_r2r (mm7, mm4);
paddw_r2r (mm3, mm1);
psllw_i2r (1, mm2);
paddw_r2r (mm4, mm0);
psllw_i2r (2, mm1); // 2
paddw_r2r (mm6, mm2);
paddw_r2r (mm2, mm1);
psubusw_r2r (mm0, mm1);
psrlw_i2r (3, mm1); // 3
packuswb_r2r (mm7, mm1);
movd_r2m (mm1, *dst);
lum_m4 += 4;
lum_m3 += 4;
lum_m2 += 4;
lum_m1 += 4;
lum += 4;
dst += 4;
}
emms ();
/* Handle odd widths */
if (size > 0)
deinterlace_line_c (dst, lum_m4, lum_m3, lum_m2, lum_m1, lum, size);
}
/* filter parameters: [-1 4 2 4 -1] // 8 */
static inline void deinterlace_line(uint8_t *dst,
const uint8_t *lum_m4, const uint8_t *lum_m3,
const uint8_t *lum_m2, const uint8_t *lum_m1,
const uint8_t *lum,
int size)
{
#ifndef USE_MMX
uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;
int sum;
for(;size > 0;size--) {
sum = -lum_m4[0];
sum += lum_m3[0] << 2;
sum += lum_m2[0] << 1;
sum += lum_m1[0] << 2;
sum += -lum[0];
dst[0] = cm[(sum + 4) >> 3];
lum_m4++;
lum_m3++;
lum_m2++;
lum_m1++;
lum++;
dst++;
}
#else
{
mmx_t rounder;
rounder.uw[0]=4;
rounder.uw[1]=4;
rounder.uw[2]=4;
rounder.uw[3]=4;
pxor_r2r(mm7,mm7);
movq_m2r(rounder,mm6);
}
for (;size > 3; size-=4) {
DEINT_LINE_LUM
lum_m4+=4;
lum_m3+=4;
lum_m2+=4;
lum_m1+=4;
lum+=4;
dst+=4;
}
#endif
}
static inline void XDeint8x8MergeMMXEXT( uint8_t *dst, int i_dst,
uint8_t *src1, int i_src1,
uint8_t *src2, int i_src2 )
{
static const uint64_t m_4 = INT64_C(0x0004000400040004);
int y, x;
/* Progressive */
pxor_r2r( mm7, mm7 );
for( y = 0; y < 8; y += 2 )
{
for( x = 0; x < 8; x +=4 )
{
movd_m2r( src1[x], mm0 );
movd_r2m( mm0, dst[x] );
movd_m2r( src2[x], mm1 );
movd_m2r( src1[i_src1+x], mm2 );
punpcklbw_r2r( mm7, mm0 );
punpcklbw_r2r( mm7, mm1 );
punpcklbw_r2r( mm7, mm2 );
paddw_r2r( mm1, mm1 );
movq_r2r( mm1, mm3 );
paddw_r2r( mm3, mm3 );
paddw_r2r( mm2, mm0 );
paddw_r2r( mm3, mm1 );
paddw_m2r( m_4, mm1 );
paddw_r2r( mm1, mm0 );
psraw_i2r( 3, mm0 );
packuswb_r2r( mm7, mm0 );
movd_r2m( mm0, dst[i_dst+x] );
}
dst += 2*i_dst;
src1 += i_src1;
src2 += i_src2;
}
}
static void scale_uint16_x_1_x_bicubic_mmx(gavl_video_scale_context_t * ctx, int scanline, uint8_t * dest_start)
{
int i;
uint16_t * dst;
uint8_t * src, *src_start;
int32_t * factors;
mmx_t tmp_mm;
int32_t tmp;
// fprintf(stderr, "scale_uint8_x_1_x_bicubic_mmx\n");
src_start = ctx->src + scanline * ctx->src_stride;
pxor_r2r(mm6, mm6);
dst = (uint16_t*)dest_start;
for(i = 0; i < ctx->dst_size; i++)
{
src = src_start + 2*ctx->table_h.pixels[i].index;
factors = ctx->table_h.pixels[i].factor_i;
/* Load pixels */
movq_m2r(*(src), mm0);
psrlw_i2r(1, mm0);
// DUMP_MM("mm0", mm0);
/* Load factors */
movq_m2r(*factors, mm2);
movq_m2r(*(factors+2), mm3);
packssdw_r2r(mm3, mm2);
/* Multiply */
pmaddwd_r2r(mm2, mm0);
MOVQ_R2M(mm0, tmp_mm);
tmp = tmp_mm.d[0] + tmp_mm.d[1];
tmp >>= 13;
RECLIP(tmp, ctx->plane);
*(dst++) = tmp;
}
ctx->need_emms = 1;
}
int field_dct_best_mmx( uint8_t *cur_lum_mb, uint8_t *pred_lum_mb)
{
/*
* calculate prediction error (cur-pred) for top (blk0)
* and bottom field (blk1)
*/
double r,d;
int rowoffs = 0;
int sumtop, sumbot, sumsqtop, sumsqbot, sumbottop;
int j;
int dct_type;
int topvar, botvar;
mmx_t sumtop_accs, sumbot_accs;
mmx_t sumsqtop_accs, sumsqbot_accs, sumxprod_accs;
int32_t sumtop_acc, sumbot_acc;
int32_t sumsqtop_acc, sumsqbot_acc, sumxprod_acc;
pxor_r2r(mm0,mm0);
movq_r2m( mm0, *(&sumtop_accs) );
movq_r2m( mm0, *(&sumbot_accs) );
movq_r2m( mm0, *(&sumsqtop_accs) );
movq_r2m( mm0, *(&sumsqbot_accs) );
movq_r2m( mm0, *(&sumxprod_accs) );
sumtop = sumsqtop = sumbot = sumsqbot = sumbottop = 0;
sumtop_acc = sumbot_acc = sumsqtop_acc = sumsqbot_acc = sumxprod_acc = 0;
for (j=0; j<8; j++)
{
#ifdef ORIGINAL_CODE
for (i=0; i<16; i++)
{
register int toppix =
cur_lum_mb[rowoffs+i] - pred_lum_mb[rowoffs+i];
register int botpix =
cur_lum_mb[rowoffs+width+i] - pred_lum_mb[rowoffs+width+i];
sumtop += toppix;
sumsqtop += toppix*toppix;
sumbot += botpix;
sumsqbot += botpix*botpix;
sumbottop += toppix*botpix;
}
#endif
sum_sumsq_8bytes( &cur_lum_mb[rowoffs], &pred_lum_mb[rowoffs],
&sumtop_accs, &sumbot_accs,
&sumsqtop_accs, &sumsqbot_accs, &sumxprod_accs
);
sum_sumsq_8bytes( &cur_lum_mb[rowoffs+8], &pred_lum_mb[rowoffs+8],
&sumtop_accs, &sumbot_accs,
&sumsqtop_accs, &sumsqbot_accs, &sumxprod_accs );
rowoffs += (opt->phy_width<<1);
}
mmx_sum_4_word_accs( &sumtop_accs, &sumtop );
mmx_sum_4_word_accs( &sumbot_accs, &sumbot );
emms();
sumsqtop = sumsqtop_accs.d[0] + sumsqtop_accs.d[1];
sumsqbot = sumsqbot_accs.d[0] + sumsqbot_accs.d[1];
sumbottop = sumxprod_accs.d[0] + sumxprod_accs.d[1];
/* Calculate Variances top and bottom. If they're of similar
sign estimate correlation if its good use frame DCT otherwise
use field.
*/
r = 0.0;
topvar = sumsqtop-sumtop*sumtop/128;
botvar = sumsqbot-sumbot*sumbot/128;
if ( !((topvar <= 0) ^ (botvar <= 0)) )
{
d = ((double) topvar) * ((double)botvar);
r = (sumbottop-(sumtop*sumbot)/128);
if (r>0.5*sqrt(d))
return 0; /* frame DCT */
else
return 1; /* field DCT */
}
else
return 1; /* field DCT */
return dct_type;
}
static inline void mmx_yuv2rgb (uint8_t * py, uint8_t * pu, uint8_t * pv)
{
static mmx_t mmx_80w = {0x0080008000800080LL};
static mmx_t mmx_U_green = {0xf37df37df37df37dLL};
static mmx_t mmx_U_blue = {0x4093409340934093LL};
static mmx_t mmx_V_red = {0x3312331233123312LL};
static mmx_t mmx_V_green = {0xe5fce5fce5fce5fcLL};
static mmx_t mmx_10w = {0x1010101010101010LL};
static mmx_t mmx_00ffw = {0x00ff00ff00ff00ffLL};
static mmx_t mmx_Y_coeff = {0x253f253f253f253fLL};
movd_m2r (*pu, mm0); /* mm0 = 00 00 00 00 u3 u2 u1 u0 */
movd_m2r (*pv, mm1); /* mm1 = 00 00 00 00 v3 v2 v1 v0 */
movq_m2r (*py, mm6); /* mm6 = Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0 */
pxor_r2r (mm4, mm4); /* mm4 = 0 */
/* XXX might do cache preload for image here */
/*
* Do the multiply part of the conversion for even and odd pixels
* register usage:
* mm0 -> Cblue, mm1 -> Cred, mm2 -> Cgreen even pixels
* mm3 -> Cblue, mm4 -> Cred, mm5 -> Cgreen odd pixels
* mm6 -> Y even, mm7 -> Y odd
*/
punpcklbw_r2r (mm4, mm0); /* mm0 = u3 u2 u1 u0 */
punpcklbw_r2r (mm4, mm1); /* mm1 = v3 v2 v1 v0 */
psubsw_m2r (mmx_80w, mm0); /* u -= 128 */
psubsw_m2r (mmx_80w, mm1); /* v -= 128 */
psllw_i2r (3, mm0); /* promote precision */
psllw_i2r (3, mm1); /* promote precision */
movq_r2r (mm0, mm2); /* mm2 = u3 u2 u1 u0 */
movq_r2r (mm1, mm3); /* mm3 = v3 v2 v1 v0 */
pmulhw_m2r (mmx_U_green, mm2); /* mm2 = u * u_green */
pmulhw_m2r (mmx_V_green, mm3); /* mm3 = v * v_green */
pmulhw_m2r (mmx_U_blue, mm0); /* mm0 = chroma_b */
pmulhw_m2r (mmx_V_red, mm1); /* mm1 = chroma_r */
paddsw_r2r (mm3, mm2); /* mm2 = chroma_g */
psubusb_m2r (mmx_10w, mm6); /* Y -= 16 */
movq_r2r (mm6, mm7); /* mm7 = Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0 */
pand_m2r (mmx_00ffw, mm6); /* mm6 = Y6 Y4 Y2 Y0 */
psrlw_i2r (8, mm7); /* mm7 = Y7 Y5 Y3 Y1 */
psllw_i2r (3, mm6); /* promote precision */
psllw_i2r (3, mm7); /* promote precision */
pmulhw_m2r (mmx_Y_coeff, mm6); /* mm6 = luma_rgb even */
pmulhw_m2r (mmx_Y_coeff, mm7); /* mm7 = luma_rgb odd */
/*
* Do the addition part of the conversion for even and odd pixels
* register usage:
* mm0 -> Cblue, mm1 -> Cred, mm2 -> Cgreen even pixels
* mm3 -> Cblue, mm4 -> Cred, mm5 -> Cgreen odd pixels
* mm6 -> Y even, mm7 -> Y odd
*/
movq_r2r (mm0, mm3); /* mm3 = chroma_b */
movq_r2r (mm1, mm4); /* mm4 = chroma_r */
movq_r2r (mm2, mm5); /* mm5 = chroma_g */
paddsw_r2r (mm6, mm0); /* mm0 = B6 B4 B2 B0 */
paddsw_r2r (mm7, mm3); /* mm3 = B7 B5 B3 B1 */
paddsw_r2r (mm6, mm1); /* mm1 = R6 R4 R2 R0 */
paddsw_r2r (mm7, mm4); /* mm4 = R7 R5 R3 R1 */
paddsw_r2r (mm6, mm2); /* mm2 = G6 G4 G2 G0 */
paddsw_r2r (mm7, mm5); /* mm5 = G7 G5 G3 G1 */
packuswb_r2r (mm0, mm0); /* saturate to 0-255 */
packuswb_r2r (mm1, mm1); /* saturate to 0-255 */
packuswb_r2r (mm2, mm2); /* saturate to 0-255 */
packuswb_r2r (mm3, mm3); /* saturate to 0-255 */
packuswb_r2r (mm4, mm4); /* saturate to 0-255 */
packuswb_r2r (mm5, mm5); /* saturate to 0-255 */
punpcklbw_r2r (mm3, mm0); /* mm0 = B7 B6 B5 B4 B3 B2 B1 B0 */
punpcklbw_r2r (mm4, mm1); /* mm1 = R7 R6 R5 R4 R3 R2 R1 R0 */
punpcklbw_r2r (mm5, mm2); /* mm2 = G7 G6 G5 G4 G3 G2 G1 G0 */
}
static int bsad_1quad_mmxe(uint8_t *pf, uint8_t *pb, uint8_t *pb2, uint8_t *p2, int lx, int h)
{
int s;
s = 0; /* the accumulator */
if (h > 0)
{
pcmpeqw_r2r(mm6, mm6);
psrlw_i2r(15, mm6);
paddw_r2r(mm6, mm6);
pxor_r2r(mm7, mm7);
pxor_r2r(mm5, mm5);
do {
BSAD_LOAD(pf[0],mm0,mm1);
BSAD_LOAD_ACC(pf[1],mm2,mm3,mm0,mm1);
BSAD_LOAD_ACC(pf[lx],mm2,mm3,mm0,mm1);
BSAD_LOAD_ACC(pf[lx+1],mm2,mm3,mm0,mm1);
paddw_r2r(mm6, mm0);
paddw_r2r(mm6, mm1);
psrlw_i2r(2, mm0);
psrlw_i2r(2, mm1);
packuswb_r2r(mm1, mm0);
movq_m2r(pb2[0],mm1);
pavgb_m2r(pb[0],mm1);
pavgb_r2r(mm1, mm0);
psadbw_m2r(p2[0],mm0);
paddd_r2r(mm0,mm5);
BSAD_LOAD(pf[8],mm0,mm1);
BSAD_LOAD_ACC(pf[9],mm2,mm3,mm0,mm1);
BSAD_LOAD_ACC(pf[lx+8],mm2,mm3,mm0,mm1);
BSAD_LOAD_ACC(pf[lx+9],mm2,mm3,mm0,mm1);
paddw_r2r(mm6, mm0);
paddw_r2r(mm6, mm1);
psrlw_i2r(2, mm0);
psrlw_i2r(2, mm1);
packuswb_r2r(mm1, mm0);
movq_m2r(pb2[8],mm1);
pavgb_m2r(pb[8],mm1);
pavgb_r2r(mm1, mm0);
psadbw_m2r(p2[8],mm0);
paddd_r2r(mm0,mm5);
p2 += lx;
pf += lx;
pb += lx;
pb2 += lx;
h--;
} while (h > 0);
}
movd_r2g(mm5,s);
emms();
return s;
}
VLC_MMX
static int TestForMotionInBlockMMX( uint8_t *p_pix_p, uint8_t *p_pix_c,
int i_pitch_prev, int i_pitch_curr,
int* pi_top, int* pi_bot )
{
int32_t i_motion = 0;
int32_t i_top_motion = 0;
int32_t i_bot_motion = 0;
static alignas (8) const mmx_t bT = { .ub = { T, T, T, T, T, T, T, T } };
pxor_r2r( mm6, mm6 ); /* zero, used in psadbw */
movq_m2r( bT, mm5 );
pxor_r2r( mm3, mm3 ); /* score (top field) */
pxor_r2r( mm4, mm4 ); /* score (bottom field) */
for( int y = 0; y < 8; y+=2 )
{
/* top field */
movq_m2r( *((uint64_t*)p_pix_c), mm0 );
movq_m2r( *((uint64_t*)p_pix_p), mm1 );
movq_r2r( mm0, mm2 );
psubusb_r2r( mm1, mm2 );
psubusb_r2r( mm0, mm1 );
pcmpgtb_r2r( mm5, mm2 );
pcmpgtb_r2r( mm5, mm1 );
psadbw_r2r( mm6, mm2 );
psadbw_r2r( mm6, mm1 );
paddd_r2r( mm2, mm1 );
paddd_r2r( mm1, mm3 ); /* add to top field score */
p_pix_c += i_pitch_curr;
p_pix_p += i_pitch_prev;
/* bottom field - handling identical to top field, except... */
movq_m2r( *((uint64_t*)p_pix_c), mm0 );
movq_m2r( *((uint64_t*)p_pix_p), mm1 );
movq_r2r( mm0, mm2 );
psubusb_r2r( mm1, mm2 );
psubusb_r2r( mm0, mm1 );
pcmpgtb_r2r( mm5, mm2 );
pcmpgtb_r2r( mm5, mm1 );
psadbw_r2r( mm6, mm2 );
psadbw_r2r( mm6, mm1 );
paddd_r2r( mm2, mm1 );
paddd_r2r( mm1, mm4 ); /* ...here we add to bottom field score */
p_pix_c += i_pitch_curr;
p_pix_p += i_pitch_prev;
}
movq_r2r( mm3, mm7 ); /* score (total) */
paddd_r2r( mm4, mm7 );
movd_r2m( mm3, i_top_motion );
movd_r2m( mm4, i_bot_motion );
movd_r2m( mm7, i_motion );
/* The loop counts actual score * 255. */
i_top_motion /= 255;
i_bot_motion /= 255;
i_motion /= 255;
emms();
(*pi_top) = ( i_top_motion >= 8 );
(*pi_bot) = ( i_bot_motion >= 8 );
return (i_motion >= 8);
}
VLC_MMX
static int CalculateInterlaceScoreMMX( const picture_t* p_pic_top,
const picture_t* p_pic_bot )
{
assert( p_pic_top->i_planes == p_pic_bot->i_planes );
/* Amount of bits must be known for MMX, thus int32_t.
Doesn't hurt the C implementation. */
int32_t i_score_mmx = 0; /* this must be divided by 255 when finished */
int32_t i_score_c = 0; /* this counts as-is (used for non-MMX parts) */
pxor_r2r( mm7, mm7 ); /* we will keep score in mm7 */
for( int i_plane = 0 ; i_plane < p_pic_top->i_planes ; ++i_plane )
{
/* Sanity check */
if( p_pic_top->p[i_plane].i_visible_lines !=
p_pic_bot->p[i_plane].i_visible_lines )
return -1;
const int i_lasty = p_pic_top->p[i_plane].i_visible_lines-1;
const int w = FFMIN( p_pic_top->p[i_plane].i_visible_pitch,
p_pic_bot->p[i_plane].i_visible_pitch );
const int wm8 = w % 8; /* remainder */
const int w8 = w - wm8; /* part of width that is divisible by 8 */
/* Current line / neighbouring lines picture pointers */
const picture_t *cur = p_pic_bot;
const picture_t *ngh = p_pic_top;
int wc = cur->p[i_plane].i_pitch;
int wn = ngh->p[i_plane].i_pitch;
/* Transcode 1.1.5 only checks every other line. Checking every line
works better for anime, which may contain horizontal,
one pixel thick cartoon outlines.
*/
for( int y = 1; y < i_lasty; ++y )
{
uint8_t *p_c = &cur->p[i_plane].p_pixels[y*wc]; /* this line */
uint8_t *p_p = &ngh->p[i_plane].p_pixels[(y-1)*wn]; /* prev line */
uint8_t *p_n = &ngh->p[i_plane].p_pixels[(y+1)*wn]; /* next line */
int x = 0;
/* Easy-to-read C version further below.
Assumptions: 0 < T < 127
# of pixels < (2^32)/255
Note: calculates score * 255
*/
static alignas (8) const mmx_t b0 = {
.uq = 0x0000000000000000ULL };
static alignas (8) const mmx_t b128 = {
.uq = 0x8080808080808080ULL };
static alignas (8) const mmx_t bT = {
.ub = { T, T, T, T, T, T, T, T } };
for( ; x < w8; x += 8 )
{
movq_m2r( *((int64_t*)p_c), mm0 );
movq_m2r( *((int64_t*)p_p), mm1 );
movq_m2r( *((int64_t*)p_n), mm2 );
psubb_m2r( b128, mm0 );
psubb_m2r( b128, mm1 );
psubb_m2r( b128, mm2 );
psubsb_r2r( mm0, mm1 );
psubsb_r2r( mm0, mm2 );
pxor_r2r( mm3, mm3 );
pxor_r2r( mm4, mm4 );
pxor_r2r( mm5, mm5 );
pxor_r2r( mm6, mm6 );
punpcklbw_r2r( mm1, mm3 );
punpcklbw_r2r( mm2, mm4 );
punpckhbw_r2r( mm1, mm5 );
punpckhbw_r2r( mm2, mm6 );
pmulhw_r2r( mm3, mm4 );
pmulhw_r2r( mm5, mm6 );
packsswb_r2r(mm4, mm6);
pcmpgtb_m2r( bT, mm6 );
psadbw_m2r( b0, mm6 );
paddd_r2r( mm6, mm7 );
p_c += 8;
p_p += 8;
p_n += 8;
}
for( ; x < w; ++x )
{
/* Worst case: need 17 bits for "comb". */
int_fast32_t C = *p_c;
int_fast32_t P = *p_p;
int_fast32_t N = *p_n;
/* Comments in Transcode's filter_ivtc.c attribute this
combing metric to Gunnar Thalin.
The idea is that if the picture is interlaced, both
expressions will have the same sign, and this comes
up positive. The value T = 100 has been chosen such
that a pixel difference of 10 (on average) will
trigger the detector.
*/
int_fast32_t comb = (P - C) * (N - C);
if( comb > T )
++i_score_c;
++p_c;
++p_p;
++p_n;
}
/* Now the other field - swap current and neighbour pictures */
const picture_t *tmp = cur;
cur = ngh;
ngh = tmp;
int tmp_pitch = wc;
wc = wn;
wn = tmp_pitch;
}
}
movd_r2m( mm7, i_score_mmx );
emms();
return i_score_mmx/255 + i_score_c;
}
#endif
/* See header for function doc. */
int CalculateInterlaceScore( const picture_t* p_pic_top,
const picture_t* p_pic_bot )
{
/*
We use the comb metric from the IVTC filter of Transcode 1.1.5.
This was found to work better for the particular purpose of IVTC
than RenderX()'s comb metric.
Note that we *must not* subsample at all in order to catch interlacing
in telecined frames with localized motion (e.g. anime with characters
talking, where only mouths move and everything else stays still.)
*/
assert( p_pic_top != NULL );
assert( p_pic_bot != NULL );
if( p_pic_top->i_planes != p_pic_bot->i_planes )
return -1;
#ifdef CAN_COMPILE_MMXEXT
if (vlc_CPU_MMXEXT())
return CalculateInterlaceScoreMMX( p_pic_top, p_pic_bot );
#endif
int32_t i_score = 0;
for( int i_plane = 0 ; i_plane < p_pic_top->i_planes ; ++i_plane )
{
/* Sanity check */
if( p_pic_top->p[i_plane].i_visible_lines !=
p_pic_bot->p[i_plane].i_visible_lines )
return -1;
const int i_lasty = p_pic_top->p[i_plane].i_visible_lines-1;
const int w = FFMIN( p_pic_top->p[i_plane].i_visible_pitch,
p_pic_bot->p[i_plane].i_visible_pitch );
/* Current line / neighbouring lines picture pointers */
const picture_t *cur = p_pic_bot;
const picture_t *ngh = p_pic_top;
int wc = cur->p[i_plane].i_pitch;
int wn = ngh->p[i_plane].i_pitch;
/* Transcode 1.1.5 only checks every other line. Checking every line
works better for anime, which may contain horizontal,
one pixel thick cartoon outlines.
*/
for( int y = 1; y < i_lasty; ++y )
{
uint8_t *p_c = &cur->p[i_plane].p_pixels[y*wc]; /* this line */
uint8_t *p_p = &ngh->p[i_plane].p_pixels[(y-1)*wn]; /* prev line */
uint8_t *p_n = &ngh->p[i_plane].p_pixels[(y+1)*wn]; /* next line */
for( int x = 0; x < w; ++x )
{
/* Worst case: need 17 bits for "comb". */
int_fast32_t C = *p_c;
int_fast32_t P = *p_p;
int_fast32_t N = *p_n;
/* Comments in Transcode's filter_ivtc.c attribute this
combing metric to Gunnar Thalin.
The idea is that if the picture is interlaced, both
expressions will have the same sign, and this comes
up positive. The value T = 100 has been chosen such
that a pixel difference of 10 (on average) will
trigger the detector.
*/
int_fast32_t comb = (P - C) * (N - C);
if( comb > T )
++i_score;
++p_c;
++p_p;
++p_n;
}
/* Now the other field - swap current and neighbour pictures */
const picture_t *tmp = cur;
cur = ngh;
ngh = tmp;
int tmp_pitch = wc;
wc = wn;
wn = tmp_pitch;
}
}
return i_score;
}
int bsad_mmx(uint8_t *pf, uint8_t *pb, uint8_t *p2, int lx, int hxf, int hyf, int hxb, int hyb, int h)
{
uint8_t *pfa,*pfb,*pfc,*pba,*pbb,*pbc;
int s, s1, s2;
pfa = pf + hxf;
pfb = pf + lx * hyf;
pfc = pfb + hxf;
pba = pb + hxb;
pbb = pb + lx * hyb;
pbc = pbb + hxb;
s = 0; /* the accumulator */
if (h > 0)
{
pxor_r2r(mm7, mm7);
pxor_r2r(mm6, mm6);
pcmpeqw_r2r(mm5, mm5);
psubw_r2r(mm5, mm6);
psllw_i2r(1, mm6);
do {
BSAD_LOAD(pf[0],mm0,mm1);
BSAD_LOAD_ACC(pfa[0],mm2,mm3,mm0,mm1);
BSAD_LOAD_ACC(pfb[0],mm2,mm3,mm0,mm1);
BSAD_LOAD_ACC(pfc[0],mm2,mm3,mm0,mm1);
paddw_r2r(mm6, mm0);
paddw_r2r(mm6, mm1);
psrlw_i2r(2, mm0);
psrlw_i2r(2, mm1);
BSAD_LOAD(pb[0],mm2,mm3);
BSAD_LOAD_ACC(pba[0],mm4,mm5,mm2,mm3);
BSAD_LOAD_ACC(pbb[0],mm4,mm5,mm2,mm3);
BSAD_LOAD_ACC(pbc[0],mm4,mm5,mm2,mm3);
paddw_r2r(mm6, mm2);
paddw_r2r(mm6, mm3);
psrlw_i2r(2, mm2);
psrlw_i2r(2, mm3);
paddw_r2r(mm2, mm0);
paddw_r2r(mm3, mm1);
psrlw_i2r(1, mm6);
paddw_r2r(mm6, mm0);
paddw_r2r(mm6, mm1);
psllw_i2r(1, mm6);
psrlw_i2r(1, mm0);
psrlw_i2r(1, mm1);
packuswb_r2r(mm1, mm0);
movq_m2r(p2[0], mm1);
movq_r2r(mm0, mm2);
psubusb_r2r(mm1, mm0);
psubusb_r2r(mm2, mm1);
por_r2r(mm1, mm0);
movq_r2r(mm0, mm1);
punpcklbw_r2r(mm7, mm0);
punpckhbw_r2r(mm7, mm1);
paddw_r2r(mm1, mm0);
movq_r2r(mm0, mm1);
punpcklwd_r2r(mm7, mm0);
punpckhwd_r2r(mm7, mm1);
paddd_r2r(mm1, mm0);
movd_r2g(mm0, s1);
psrlq_i2r(32, mm0);
movd_r2g(mm0, s2);
s += s1 + s2;
BSAD_LOAD(pf[8],mm0,mm1);
BSAD_LOAD_ACC(pfa[8],mm2,mm3,mm0,mm1);
BSAD_LOAD_ACC(pfb[8],mm2,mm3,mm0,mm1);
BSAD_LOAD_ACC(pfc[8],mm2,mm3,mm0,mm1);
paddw_r2r(mm6, mm0);
paddw_r2r(mm6, mm1);
psrlw_i2r(2, mm0);
psrlw_i2r(2, mm1);
BSAD_LOAD(pb[8],mm2,mm3);
BSAD_LOAD_ACC(pba[8],mm4,mm5,mm2,mm3);
BSAD_LOAD_ACC(pbb[8],mm4,mm5,mm2,mm3);
BSAD_LOAD_ACC(pbc[8],mm4,mm5,mm2,mm3);
paddw_r2r(mm6, mm2);
paddw_r2r(mm6, mm3);
psrlw_i2r(2, mm2);
psrlw_i2r(2, mm3);
paddw_r2r(mm2, mm0);
paddw_r2r(mm3, mm1);
psrlw_i2r(1, mm6);
paddw_r2r(mm6, mm0);
paddw_r2r(mm6, mm1);
psllw_i2r(1, mm6);
psrlw_i2r(1, mm0);
psrlw_i2r(1, mm1);
packuswb_r2r(mm1, mm0);
movq_m2r(p2[8], mm1);
movq_r2r(mm0, mm2);
psubusb_r2r(mm1, mm0);
psubusb_r2r(mm2, mm1);
por_r2r(mm1, mm0);
movq_r2r(mm0, mm1);
punpcklbw_r2r(mm7, mm0);
punpckhbw_r2r(mm7, mm1);
paddw_r2r(mm1, mm0);
movq_r2r(mm0, mm1);
punpcklwd_r2r(mm7, mm0);
punpckhwd_r2r(mm7, mm1);
paddd_r2r(mm1, mm0);
movd_r2g(mm0, s1);
psrlq_i2r(32, mm0);
movd_r2g(mm0, s2);
s += s1 + s2;
p2 += lx;
pf += lx;
pfa += lx;
pfb += lx;
pfc += lx;
pb += lx;
pba += lx;
pbb += lx;
pbc += lx;
h--;
} while (h > 0);
}
emms();
return s;
}
static void
_evas_yv12torgb_sse(unsigned char **yuv, unsigned char *rgb, int w, int h)
{
#ifdef BUILD_MMX
int xx, yy;
register unsigned char *yp1, *up, *vp;
unsigned char *dp1;
/* destination pointers */
dp1 = rgb;
for (yy = 0; yy < h; yy++)
{
/* plane pointers */
yp1 = yuv[yy];
up = yuv[h + (yy / 2)];
vp = yuv[h + (h / 2) + (yy / 2)];
for (xx = 0; xx < (w - 7); xx += 8)
{
movd_m2r(*up, mm3);
movd_m2r(*vp, mm2);
movq_m2r(*yp1, mm0);
pxor_r2r(mm7, mm7);
punpcklbw_r2r(mm7, mm2);
punpcklbw_r2r(mm7, mm3);
movq_r2r(mm0, mm1);
psrlw_i2r(8, mm0);
psllw_i2r(8, mm1);
psrlw_i2r(8, mm1);
movq_m2r(CONST_16, mm4);
psubsw_r2r(mm4, mm0);
psubsw_r2r(mm4, mm1);
movq_m2r(CONST_128, mm5);
psubsw_r2r(mm5, mm2);
psubsw_r2r(mm5, mm3);
movq_m2r(CONST_YMUL, mm4);
pmullw_r2r(mm4, mm0);
pmullw_r2r(mm4, mm1);
movq_m2r(CONST_CRVCRV, mm7);
pmullw_r2r(mm3, mm7);
movq_m2r(CONST_CBUCBU, mm6);
pmullw_r2r(mm2, mm6);
movq_m2r(CONST_CGUCGU, mm5);
pmullw_r2r(mm2, mm5);
movq_m2r(CONST_CGVCGV, mm4);
pmullw_r2r(mm3, mm4);
movq_r2r(mm0, mm2);
paddsw_r2r(mm7, mm2);
paddsw_r2r(mm1, mm7);
psraw_i2r(RES, mm2);
psraw_i2r(RES, mm7);
packuswb_r2r(mm7, mm2);
pxor_r2r(mm7, mm7);
movq_r2r(mm2, mm3);
punpckhbw_r2r(mm7, mm2);
punpcklbw_r2r(mm3, mm7);
por_r2r(mm7, mm2);
movq_r2r(mm0, mm3);
psubsw_r2r(mm5, mm3);
psubsw_r2r(mm4, mm3);
paddsw_m2r(CONST_32, mm3);
movq_r2r(mm1, mm7);
psubsw_r2r(mm5, mm7);
psubsw_r2r(mm4, mm7);
paddsw_m2r(CONST_32, mm7);
psraw_i2r(RES, mm3);
psraw_i2r(RES, mm7);
packuswb_r2r(mm7, mm3);
pxor_r2r(mm7, mm7);
movq_r2r(mm3, mm4);
punpckhbw_r2r(mm7, mm3);
punpcklbw_r2r(mm4, mm7);
por_r2r(mm7, mm3);
movq_m2r(CONST_32, mm4);
paddsw_r2r(mm6, mm0);
paddsw_r2r(mm6, mm1);
paddsw_r2r(mm4, mm0);
paddsw_r2r(mm4, mm1);
psraw_i2r(RES, mm0);
psraw_i2r(RES, mm1);
packuswb_r2r(mm1, mm0);
pxor_r2r(mm7, mm7);
movq_r2r(mm0, mm5);
punpckhbw_r2r(mm7, mm0);
punpcklbw_r2r(mm5, mm7);
por_r2r(mm7, mm0);
movq_m2r(CONST_FF, mm1);
movq_r2r(mm0, mm5);
movq_r2r(mm3, mm6);
movq_r2r(mm2, mm7);
punpckhbw_r2r(mm3, mm2);
punpcklbw_r2r(mm6, mm7);
punpckhbw_r2r(mm1, mm0);
punpcklbw_r2r(mm1, mm5);
movq_r2r(mm7, mm1);
punpckhwd_r2r(mm5, mm7);
punpcklwd_r2r(mm5, mm1);
movq_r2r(mm2, mm4);
punpckhwd_r2r(mm0, mm2);
punpcklwd_r2r(mm0, mm4);
movntq_r2m(mm1, *(dp1));
movntq_r2m(mm7, *(dp1 + 8));
movntq_r2m(mm4, *(dp1 + 16));
movntq_r2m(mm2, *(dp1 + 24));
yp1 += 8;
up += 4;
vp += 4;
dp1 += 8 * 4;
}
/* cleanup pixles that arent a multiple of 8 pixels wide */
if (xx < w)
{
int y, u, v, r, g, b;
for (; xx < w; xx += 2)
{
u = (*up++) - 128;
v = (*vp++) - 128;
y = RZ(YMUL) * ((*yp1++) - 16);
r = LUT_CLIP((y + (_crv * v)) >> RES);
g = LUT_CLIP((y - (_cgu * u) - (_cgv * v) + RZ(OFF)) >> RES);
b = LUT_CLIP((y + (_cbu * u) + RZ(OFF)) >> RES);
*((DATA32 *) dp1) = 0xff000000 + RGB_JOIN(r,g,b);
dp1 += 4;
y = RZ(YMUL) * ((*yp1++) - 16);
r = LUT_CLIP((y + (_crv * v)) >> RES);
g = LUT_CLIP((y - (_cgu * u) - (_cgv * v) + RZ(OFF)) >> RES);
b = LUT_CLIP((y + (_cbu * u) + RZ(OFF)) >> RES);
*((DATA32 *) dp1) = 0xff000000 + RGB_JOIN(r,g,b);
dp1 += 4;
}
}
}
static inline void mmx_zero_reg (void)
{
/* load 0 into mm0 */
pxor_r2r (mm0, mm0);
}
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)
{
/* *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 */
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 */
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 */
/* now have partials in mm1 and mm2 */
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 */
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 */
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_m2r (round4, mm1);
psraw_i2r (2, mm1); /* /4 */
paddw_m2r (round4, mm2);
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 */
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 */
}
static __inline__ void
sum_sumsq_8bytes( uint8_t *cur_lum_mb,
uint8_t *pred_lum_mb,
mmx_t *sumtop_accs,
mmx_t *sumbot_accs,
mmx_t *sumsqtop_accs,
mmx_t *sumsqbot_accs,
mmx_t *sumxprod_accs
)
{
pxor_r2r(mm0,mm0);
/* Load pixels from top field into mm1.w,mm2.w
*/
movq_m2r( *((mmx_t*)cur_lum_mb), mm1 );
movq_m2r( *((mmx_t*)pred_lum_mb), mm2 );
/* mm3 := mm1 mm4 := mm2
mm1.w[0..3] := mm1.b[0..3]-mm2.b[0..3]
*/
movq_r2r( mm1, mm3 );
punpcklbw_r2r( mm0, mm1 );
movq_r2r( mm2, mm4 );
punpcklbw_r2r( mm0, mm2 );
psubw_r2r( mm2, mm1 );
/* mm3.w[0..3] := mm3.b[4..7]-mm4.b[4..7]
*/
punpckhbw_r2r( mm0, mm3 );
punpckhbw_r2r( mm0, mm4 );
psubw_r2r( mm4, mm3 );
/* sumtop_accs->w[0..3] += mm1.w[0..3];
sumtop_accs->w[0..3] += mm3.w[0..3];
mm6 = mm1; mm7 = mm3;
*/
movq_m2r( *sumtop_accs, mm5 );
paddw_r2r( mm1, mm5 );
paddw_r2r( mm3, mm5 );
movq_r2r( mm1, mm6 );
movq_r2r( mm3, mm7 );
movq_r2m( mm5, *sumtop_accs );
/*
*sumsq_top_acc += mm1.w[0..3] * mm1.w[0..3];
*sumsq_top_acc += mm3.w[0..3] * mm3.w[0..3];
*/
pmaddwd_r2r( mm1, mm1 );
movq_m2r( *sumsqtop_accs, mm5 );
pmaddwd_r2r( mm3, mm3 );
paddd_r2r( mm1, mm5 );
paddd_r2r( mm3, mm5 );
movq_r2m( mm5, *sumsqtop_accs );
/* Load pixels from bot field into mm1.w,mm2.w
*/
movq_m2r( *((mmx_t*)(cur_lum_mb+opt->phy_width)), mm1 );
movq_m2r( *((mmx_t*)(pred_lum_mb+opt->phy_width)), mm2 );
/* mm2 := mm1 mm4 := mm2
mm1.w[0..3] := mm1.b[0..3]-mm2.b[0..3]
*/
movq_r2r( mm1, mm3 );
punpcklbw_r2r( mm0, mm1 );
movq_r2r( mm2, mm4 );
punpcklbw_r2r( mm0, mm2 );
psubw_r2r( mm2, mm1 );
/* mm3.w[0..3] := mm3.b[4..7]-mm4.b[4..7]
*/
punpckhbw_r2r( mm0, mm3 );
punpckhbw_r2r( mm0, mm4 );
psubw_r2r( mm4, mm3 );
/*
sumbot_accs->w[0..3] += mm1.w[0..3];
sumbot_accs->w[0..3] += mm3.w[0..3];
mm2 := mm1; mm4 := mm3;
*/
movq_m2r( *sumbot_accs, mm5 );
paddw_r2r( mm1, mm5 );
movq_r2r( mm1, mm2 );
paddw_r2r( mm3, mm5 );
movq_r2r( mm3, mm4 );
movq_r2m( mm5, *sumbot_accs );
/*
*sumsqbot_acc += mm1.w[0..3] * mm1.w[0..3];
*sumsqbot_acc += mm3.w[0..3] * mm3.w[0..3];
*/
pmaddwd_r2r( mm1, mm1 );
movq_m2r( *sumsqbot_accs, mm5 );
pmaddwd_r2r( mm3, mm3 );
paddd_r2r( mm1, mm5 );
paddd_r2r( mm3, mm5 );
movq_r2m( mm5, *sumsqbot_accs );
/* Accumulate cross-product
*sum_xprod_acc += mm1.w[0..3] * mm6[0..3];
*sum_xprod_acc += mm3.w[0..3] * mm7[0..3];
*/
movq_m2r( *sumxprod_accs, mm5 );
pmaddwd_r2r( mm6, mm2);
pmaddwd_r2r( mm7, mm4);
paddd_r2r( mm2, mm5 );
paddd_r2r( mm4, mm5 );
movq_r2m( mm5, *sumxprod_accs );
emms();
}
