yading@10: /* yading@10: * Optimized for ia32 CPUs by Nick Kurshev yading@10: * h263, mpeg1, mpeg2 dequantizer & draw_edges by Michael Niedermayer yading@10: * yading@10: * This file is part of FFmpeg. yading@10: * yading@10: * FFmpeg is free software; you can redistribute it and/or yading@10: * modify it under the terms of the GNU Lesser General Public yading@10: * License as published by the Free Software Foundation; either yading@10: * version 2.1 of the License, or (at your option) any later version. yading@10: * yading@10: * FFmpeg is distributed in the hope that it will be useful, yading@10: * but WITHOUT ANY WARRANTY; without even the implied warranty of yading@10: * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU yading@10: * Lesser General Public License for more details. yading@10: * yading@10: * You should have received a copy of the GNU Lesser General Public yading@10: * License along with FFmpeg; if not, write to the Free Software yading@10: * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA yading@10: */ yading@10: yading@10: #include "libavutil/attributes.h" yading@10: #include "libavutil/cpu.h" yading@10: #include "libavutil/x86/asm.h" yading@10: #include "libavcodec/avcodec.h" yading@10: #include "libavcodec/mpegvideo.h" yading@10: #include "dsputil_mmx.h" yading@10: yading@10: #if HAVE_INLINE_ASM yading@10: yading@10: static void dct_unquantize_h263_intra_mmx(MpegEncContext *s, yading@10: int16_t *block, int n, int qscale) yading@10: { yading@10: x86_reg level, qmul, qadd, nCoeffs; yading@10: yading@10: qmul = qscale << 1; yading@10: yading@10: av_assert2(s->block_last_index[n]>=0 || s->h263_aic); yading@10: yading@10: if (!s->h263_aic) { yading@10: if (n < 4) yading@10: level = block[0] * s->y_dc_scale; yading@10: else yading@10: level = block[0] * s->c_dc_scale; yading@10: qadd = (qscale - 1) | 1; yading@10: }else{ yading@10: qadd = 0; yading@10: level= block[0]; yading@10: } yading@10: if(s->ac_pred) yading@10: nCoeffs=63; yading@10: else yading@10: nCoeffs= s->inter_scantable.raster_end[ s->block_last_index[n] ]; yading@10: yading@10: __asm__ volatile( yading@10: "movd %1, %%mm6 \n\t" //qmul yading@10: "packssdw %%mm6, %%mm6 \n\t" yading@10: "packssdw %%mm6, %%mm6 \n\t" yading@10: "movd %2, %%mm5 \n\t" //qadd yading@10: "pxor %%mm7, %%mm7 \n\t" yading@10: "packssdw %%mm5, %%mm5 \n\t" yading@10: "packssdw %%mm5, %%mm5 \n\t" yading@10: "psubw %%mm5, %%mm7 \n\t" yading@10: "pxor %%mm4, %%mm4 \n\t" yading@10: ".p2align 4 \n\t" yading@10: "1: \n\t" yading@10: "movq (%0, %3), %%mm0 \n\t" yading@10: "movq 8(%0, %3), %%mm1 \n\t" yading@10: yading@10: "pmullw %%mm6, %%mm0 \n\t" yading@10: "pmullw %%mm6, %%mm1 \n\t" yading@10: yading@10: "movq (%0, %3), %%mm2 \n\t" yading@10: "movq 8(%0, %3), %%mm3 \n\t" yading@10: yading@10: "pcmpgtw %%mm4, %%mm2 \n\t" // block[i] < 0 ? -1 : 0 yading@10: "pcmpgtw %%mm4, %%mm3 \n\t" // block[i] < 0 ? -1 : 0 yading@10: yading@10: "pxor %%mm2, %%mm0 \n\t" yading@10: "pxor %%mm3, %%mm1 \n\t" yading@10: yading@10: "paddw %%mm7, %%mm0 \n\t" yading@10: "paddw %%mm7, %%mm1 \n\t" yading@10: yading@10: "pxor %%mm0, %%mm2 \n\t" yading@10: "pxor %%mm1, %%mm3 \n\t" yading@10: yading@10: "pcmpeqw %%mm7, %%mm0 \n\t" // block[i] == 0 ? -1 : 0 yading@10: "pcmpeqw %%mm7, %%mm1 \n\t" // block[i] == 0 ? -1 : 0 yading@10: yading@10: "pandn %%mm2, %%mm0 \n\t" yading@10: "pandn %%mm3, %%mm1 \n\t" yading@10: yading@10: "movq %%mm0, (%0, %3) \n\t" yading@10: "movq %%mm1, 8(%0, %3) \n\t" yading@10: yading@10: "add $16, %3 \n\t" yading@10: "jng 1b \n\t" yading@10: ::"r" (block+nCoeffs), "rm"(qmul), "rm" (qadd), "r" (2*(-nCoeffs)) yading@10: : "memory" yading@10: ); yading@10: block[0]= level; yading@10: } yading@10: yading@10: yading@10: static void dct_unquantize_h263_inter_mmx(MpegEncContext *s, yading@10: int16_t *block, int n, int qscale) yading@10: { yading@10: x86_reg qmul, qadd, nCoeffs; yading@10: yading@10: qmul = qscale << 1; yading@10: qadd = (qscale - 1) | 1; yading@10: yading@10: av_assert2(s->block_last_index[n]>=0 || s->h263_aic); yading@10: yading@10: nCoeffs= s->inter_scantable.raster_end[ s->block_last_index[n] ]; yading@10: yading@10: __asm__ volatile( yading@10: "movd %1, %%mm6 \n\t" //qmul yading@10: "packssdw %%mm6, %%mm6 \n\t" yading@10: "packssdw %%mm6, %%mm6 \n\t" yading@10: "movd %2, %%mm5 \n\t" //qadd yading@10: "pxor %%mm7, %%mm7 \n\t" yading@10: "packssdw %%mm5, %%mm5 \n\t" yading@10: "packssdw %%mm5, %%mm5 \n\t" yading@10: "psubw %%mm5, %%mm7 \n\t" yading@10: "pxor %%mm4, %%mm4 \n\t" yading@10: ".p2align 4 \n\t" yading@10: "1: \n\t" yading@10: "movq (%0, %3), %%mm0 \n\t" yading@10: "movq 8(%0, %3), %%mm1 \n\t" yading@10: yading@10: "pmullw %%mm6, %%mm0 \n\t" yading@10: "pmullw %%mm6, %%mm1 \n\t" yading@10: yading@10: "movq (%0, %3), %%mm2 \n\t" yading@10: "movq 8(%0, %3), %%mm3 \n\t" yading@10: yading@10: "pcmpgtw %%mm4, %%mm2 \n\t" // block[i] < 0 ? -1 : 0 yading@10: "pcmpgtw %%mm4, %%mm3 \n\t" // block[i] < 0 ? -1 : 0 yading@10: yading@10: "pxor %%mm2, %%mm0 \n\t" yading@10: "pxor %%mm3, %%mm1 \n\t" yading@10: yading@10: "paddw %%mm7, %%mm0 \n\t" yading@10: "paddw %%mm7, %%mm1 \n\t" yading@10: yading@10: "pxor %%mm0, %%mm2 \n\t" yading@10: "pxor %%mm1, %%mm3 \n\t" yading@10: yading@10: "pcmpeqw %%mm7, %%mm0 \n\t" // block[i] == 0 ? -1 : 0 yading@10: "pcmpeqw %%mm7, %%mm1 \n\t" // block[i] == 0 ? -1 : 0 yading@10: yading@10: "pandn %%mm2, %%mm0 \n\t" yading@10: "pandn %%mm3, %%mm1 \n\t" yading@10: yading@10: "movq %%mm0, (%0, %3) \n\t" yading@10: "movq %%mm1, 8(%0, %3) \n\t" yading@10: yading@10: "add $16, %3 \n\t" yading@10: "jng 1b \n\t" yading@10: ::"r" (block+nCoeffs), "rm"(qmul), "rm" (qadd), "r" (2*(-nCoeffs)) yading@10: : "memory" yading@10: ); yading@10: } yading@10: yading@10: yading@10: /* yading@10: We can suppose that result of two multiplications can't be greater than 0xFFFF yading@10: i.e. is 16-bit, so we use here only PMULLW instruction and can avoid yading@10: a complex multiplication. yading@10: ===================================================== yading@10: Full formula for multiplication of 2 integer numbers yading@10: which are represent as high:low words: yading@10: input: value1 = high1:low1 yading@10: value2 = high2:low2 yading@10: output: value3 = value1*value2 yading@10: value3=high3:low3 (on overflow: modulus 2^32 wrap-around) yading@10: this mean that for 0x123456 * 0x123456 correct result is 0x766cb0ce4 yading@10: but this algorithm will compute only 0x66cb0ce4 yading@10: this limited by 16-bit size of operands yading@10: --------------------------------- yading@10: tlow1 = high1*low2 yading@10: tlow2 = high2*low1 yading@10: tlow1 = tlow1 + tlow2 yading@10: high3:low3 = low1*low2 yading@10: high3 += tlow1 yading@10: */ yading@10: static void dct_unquantize_mpeg1_intra_mmx(MpegEncContext *s, yading@10: int16_t *block, int n, int qscale) yading@10: { yading@10: x86_reg nCoeffs; yading@10: const uint16_t *quant_matrix; yading@10: int block0; yading@10: yading@10: av_assert2(s->block_last_index[n]>=0); yading@10: yading@10: nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]+1; yading@10: yading@10: if (n < 4) yading@10: block0 = block[0] * s->y_dc_scale; yading@10: else yading@10: block0 = block[0] * s->c_dc_scale; yading@10: /* XXX: only mpeg1 */ yading@10: quant_matrix = s->intra_matrix; yading@10: __asm__ volatile( yading@10: "pcmpeqw %%mm7, %%mm7 \n\t" yading@10: "psrlw $15, %%mm7 \n\t" yading@10: "movd %2, %%mm6 \n\t" yading@10: "packssdw %%mm6, %%mm6 \n\t" yading@10: "packssdw %%mm6, %%mm6 \n\t" yading@10: "mov %3, %%"REG_a" \n\t" yading@10: ".p2align 4 \n\t" yading@10: "1: \n\t" yading@10: "movq (%0, %%"REG_a"), %%mm0 \n\t" yading@10: "movq 8(%0, %%"REG_a"), %%mm1 \n\t" yading@10: "movq (%1, %%"REG_a"), %%mm4 \n\t" yading@10: "movq 8(%1, %%"REG_a"), %%mm5 \n\t" yading@10: "pmullw %%mm6, %%mm4 \n\t" // q=qscale*quant_matrix[i] yading@10: "pmullw %%mm6, %%mm5 \n\t" // q=qscale*quant_matrix[i] yading@10: "pxor %%mm2, %%mm2 \n\t" yading@10: "pxor %%mm3, %%mm3 \n\t" yading@10: "pcmpgtw %%mm0, %%mm2 \n\t" // block[i] < 0 ? -1 : 0 yading@10: "pcmpgtw %%mm1, %%mm3 \n\t" // block[i] < 0 ? -1 : 0 yading@10: "pxor %%mm2, %%mm0 \n\t" yading@10: "pxor %%mm3, %%mm1 \n\t" yading@10: "psubw %%mm2, %%mm0 \n\t" // abs(block[i]) yading@10: "psubw %%mm3, %%mm1 \n\t" // abs(block[i]) yading@10: "pmullw %%mm4, %%mm0 \n\t" // abs(block[i])*q yading@10: "pmullw %%mm5, %%mm1 \n\t" // abs(block[i])*q yading@10: "pxor %%mm4, %%mm4 \n\t" yading@10: "pxor %%mm5, %%mm5 \n\t" // FIXME slow yading@10: "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0 yading@10: "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0 yading@10: "psraw $3, %%mm0 \n\t" yading@10: "psraw $3, %%mm1 \n\t" yading@10: "psubw %%mm7, %%mm0 \n\t" yading@10: "psubw %%mm7, %%mm1 \n\t" yading@10: "por %%mm7, %%mm0 \n\t" yading@10: "por %%mm7, %%mm1 \n\t" yading@10: "pxor %%mm2, %%mm0 \n\t" yading@10: "pxor %%mm3, %%mm1 \n\t" yading@10: "psubw %%mm2, %%mm0 \n\t" yading@10: "psubw %%mm3, %%mm1 \n\t" yading@10: "pandn %%mm0, %%mm4 \n\t" yading@10: "pandn %%mm1, %%mm5 \n\t" yading@10: "movq %%mm4, (%0, %%"REG_a") \n\t" yading@10: "movq %%mm5, 8(%0, %%"REG_a") \n\t" yading@10: yading@10: "add $16, %%"REG_a" \n\t" yading@10: "js 1b \n\t" yading@10: ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "rm" (qscale), "g" (-2*nCoeffs) yading@10: : "%"REG_a, "memory" yading@10: ); yading@10: block[0]= block0; yading@10: } yading@10: yading@10: static void dct_unquantize_mpeg1_inter_mmx(MpegEncContext *s, yading@10: int16_t *block, int n, int qscale) yading@10: { yading@10: x86_reg nCoeffs; yading@10: const uint16_t *quant_matrix; yading@10: yading@10: av_assert2(s->block_last_index[n]>=0); yading@10: yading@10: nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]+1; yading@10: yading@10: quant_matrix = s->inter_matrix; yading@10: __asm__ volatile( yading@10: "pcmpeqw %%mm7, %%mm7 \n\t" yading@10: "psrlw $15, %%mm7 \n\t" yading@10: "movd %2, %%mm6 \n\t" yading@10: "packssdw %%mm6, %%mm6 \n\t" yading@10: "packssdw %%mm6, %%mm6 \n\t" yading@10: "mov %3, %%"REG_a" \n\t" yading@10: ".p2align 4 \n\t" yading@10: "1: \n\t" yading@10: "movq (%0, %%"REG_a"), %%mm0 \n\t" yading@10: "movq 8(%0, %%"REG_a"), %%mm1 \n\t" yading@10: "movq (%1, %%"REG_a"), %%mm4 \n\t" yading@10: "movq 8(%1, %%"REG_a"), %%mm5 \n\t" yading@10: "pmullw %%mm6, %%mm4 \n\t" // q=qscale*quant_matrix[i] yading@10: "pmullw %%mm6, %%mm5 \n\t" // q=qscale*quant_matrix[i] yading@10: "pxor %%mm2, %%mm2 \n\t" yading@10: "pxor %%mm3, %%mm3 \n\t" yading@10: "pcmpgtw %%mm0, %%mm2 \n\t" // block[i] < 0 ? -1 : 0 yading@10: "pcmpgtw %%mm1, %%mm3 \n\t" // block[i] < 0 ? -1 : 0 yading@10: "pxor %%mm2, %%mm0 \n\t" yading@10: "pxor %%mm3, %%mm1 \n\t" yading@10: "psubw %%mm2, %%mm0 \n\t" // abs(block[i]) yading@10: "psubw %%mm3, %%mm1 \n\t" // abs(block[i]) yading@10: "paddw %%mm0, %%mm0 \n\t" // abs(block[i])*2 yading@10: "paddw %%mm1, %%mm1 \n\t" // abs(block[i])*2 yading@10: "paddw %%mm7, %%mm0 \n\t" // abs(block[i])*2 + 1 yading@10: "paddw %%mm7, %%mm1 \n\t" // abs(block[i])*2 + 1 yading@10: "pmullw %%mm4, %%mm0 \n\t" // (abs(block[i])*2 + 1)*q yading@10: "pmullw %%mm5, %%mm1 \n\t" // (abs(block[i])*2 + 1)*q yading@10: "pxor %%mm4, %%mm4 \n\t" yading@10: "pxor %%mm5, %%mm5 \n\t" // FIXME slow yading@10: "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0 yading@10: "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0 yading@10: "psraw $4, %%mm0 \n\t" yading@10: "psraw $4, %%mm1 \n\t" yading@10: "psubw %%mm7, %%mm0 \n\t" yading@10: "psubw %%mm7, %%mm1 \n\t" yading@10: "por %%mm7, %%mm0 \n\t" yading@10: "por %%mm7, %%mm1 \n\t" yading@10: "pxor %%mm2, %%mm0 \n\t" yading@10: "pxor %%mm3, %%mm1 \n\t" yading@10: "psubw %%mm2, %%mm0 \n\t" yading@10: "psubw %%mm3, %%mm1 \n\t" yading@10: "pandn %%mm0, %%mm4 \n\t" yading@10: "pandn %%mm1, %%mm5 \n\t" yading@10: "movq %%mm4, (%0, %%"REG_a") \n\t" yading@10: "movq %%mm5, 8(%0, %%"REG_a") \n\t" yading@10: yading@10: "add $16, %%"REG_a" \n\t" yading@10: "js 1b \n\t" yading@10: ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "rm" (qscale), "g" (-2*nCoeffs) yading@10: : "%"REG_a, "memory" yading@10: ); yading@10: } yading@10: yading@10: static void dct_unquantize_mpeg2_intra_mmx(MpegEncContext *s, yading@10: int16_t *block, int n, int qscale) yading@10: { yading@10: x86_reg nCoeffs; yading@10: const uint16_t *quant_matrix; yading@10: int block0; yading@10: yading@10: av_assert2(s->block_last_index[n]>=0); yading@10: yading@10: if(s->alternate_scan) nCoeffs= 63; //FIXME yading@10: else nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]; yading@10: yading@10: if (n < 4) yading@10: block0 = block[0] * s->y_dc_scale; yading@10: else yading@10: block0 = block[0] * s->c_dc_scale; yading@10: quant_matrix = s->intra_matrix; yading@10: __asm__ volatile( yading@10: "pcmpeqw %%mm7, %%mm7 \n\t" yading@10: "psrlw $15, %%mm7 \n\t" yading@10: "movd %2, %%mm6 \n\t" yading@10: "packssdw %%mm6, %%mm6 \n\t" yading@10: "packssdw %%mm6, %%mm6 \n\t" yading@10: "mov %3, %%"REG_a" \n\t" yading@10: ".p2align 4 \n\t" yading@10: "1: \n\t" yading@10: "movq (%0, %%"REG_a"), %%mm0 \n\t" yading@10: "movq 8(%0, %%"REG_a"), %%mm1 \n\t" yading@10: "movq (%1, %%"REG_a"), %%mm4 \n\t" yading@10: "movq 8(%1, %%"REG_a"), %%mm5 \n\t" yading@10: "pmullw %%mm6, %%mm4 \n\t" // q=qscale*quant_matrix[i] yading@10: "pmullw %%mm6, %%mm5 \n\t" // q=qscale*quant_matrix[i] yading@10: "pxor %%mm2, %%mm2 \n\t" yading@10: "pxor %%mm3, %%mm3 \n\t" yading@10: "pcmpgtw %%mm0, %%mm2 \n\t" // block[i] < 0 ? -1 : 0 yading@10: "pcmpgtw %%mm1, %%mm3 \n\t" // block[i] < 0 ? -1 : 0 yading@10: "pxor %%mm2, %%mm0 \n\t" yading@10: "pxor %%mm3, %%mm1 \n\t" yading@10: "psubw %%mm2, %%mm0 \n\t" // abs(block[i]) yading@10: "psubw %%mm3, %%mm1 \n\t" // abs(block[i]) yading@10: "pmullw %%mm4, %%mm0 \n\t" // abs(block[i])*q yading@10: "pmullw %%mm5, %%mm1 \n\t" // abs(block[i])*q yading@10: "pxor %%mm4, %%mm4 \n\t" yading@10: "pxor %%mm5, %%mm5 \n\t" // FIXME slow yading@10: "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0 yading@10: "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0 yading@10: "psraw $3, %%mm0 \n\t" yading@10: "psraw $3, %%mm1 \n\t" yading@10: "pxor %%mm2, %%mm0 \n\t" yading@10: "pxor %%mm3, %%mm1 \n\t" yading@10: "psubw %%mm2, %%mm0 \n\t" yading@10: "psubw %%mm3, %%mm1 \n\t" yading@10: "pandn %%mm0, %%mm4 \n\t" yading@10: "pandn %%mm1, %%mm5 \n\t" yading@10: "movq %%mm4, (%0, %%"REG_a") \n\t" yading@10: "movq %%mm5, 8(%0, %%"REG_a") \n\t" yading@10: yading@10: "add $16, %%"REG_a" \n\t" yading@10: "jng 1b \n\t" yading@10: ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "rm" (qscale), "g" (-2*nCoeffs) yading@10: : "%"REG_a, "memory" yading@10: ); yading@10: block[0]= block0; yading@10: //Note, we do not do mismatch control for intra as errors cannot accumulate yading@10: } yading@10: yading@10: static void dct_unquantize_mpeg2_inter_mmx(MpegEncContext *s, yading@10: int16_t *block, int n, int qscale) yading@10: { yading@10: x86_reg nCoeffs; yading@10: const uint16_t *quant_matrix; yading@10: yading@10: av_assert2(s->block_last_index[n]>=0); yading@10: yading@10: if(s->alternate_scan) nCoeffs= 63; //FIXME yading@10: else nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]; yading@10: yading@10: quant_matrix = s->inter_matrix; yading@10: __asm__ volatile( yading@10: "pcmpeqw %%mm7, %%mm7 \n\t" yading@10: "psrlq $48, %%mm7 \n\t" yading@10: "movd %2, %%mm6 \n\t" yading@10: "packssdw %%mm6, %%mm6 \n\t" yading@10: "packssdw %%mm6, %%mm6 \n\t" yading@10: "mov %3, %%"REG_a" \n\t" yading@10: ".p2align 4 \n\t" yading@10: "1: \n\t" yading@10: "movq (%0, %%"REG_a"), %%mm0 \n\t" yading@10: "movq 8(%0, %%"REG_a"), %%mm1 \n\t" yading@10: "movq (%1, %%"REG_a"), %%mm4 \n\t" yading@10: "movq 8(%1, %%"REG_a"), %%mm5 \n\t" yading@10: "pmullw %%mm6, %%mm4 \n\t" // q=qscale*quant_matrix[i] yading@10: "pmullw %%mm6, %%mm5 \n\t" // q=qscale*quant_matrix[i] yading@10: "pxor %%mm2, %%mm2 \n\t" yading@10: "pxor %%mm3, %%mm3 \n\t" yading@10: "pcmpgtw %%mm0, %%mm2 \n\t" // block[i] < 0 ? -1 : 0 yading@10: "pcmpgtw %%mm1, %%mm3 \n\t" // block[i] < 0 ? -1 : 0 yading@10: "pxor %%mm2, %%mm0 \n\t" yading@10: "pxor %%mm3, %%mm1 \n\t" yading@10: "psubw %%mm2, %%mm0 \n\t" // abs(block[i]) yading@10: "psubw %%mm3, %%mm1 \n\t" // abs(block[i]) yading@10: "paddw %%mm0, %%mm0 \n\t" // abs(block[i])*2 yading@10: "paddw %%mm1, %%mm1 \n\t" // abs(block[i])*2 yading@10: "pmullw %%mm4, %%mm0 \n\t" // abs(block[i])*2*q yading@10: "pmullw %%mm5, %%mm1 \n\t" // abs(block[i])*2*q yading@10: "paddw %%mm4, %%mm0 \n\t" // (abs(block[i])*2 + 1)*q yading@10: "paddw %%mm5, %%mm1 \n\t" // (abs(block[i])*2 + 1)*q yading@10: "pxor %%mm4, %%mm4 \n\t" yading@10: "pxor %%mm5, %%mm5 \n\t" // FIXME slow yading@10: "pcmpeqw (%0, %%"REG_a"), %%mm4 \n\t" // block[i] == 0 ? -1 : 0 yading@10: "pcmpeqw 8(%0, %%"REG_a"), %%mm5\n\t" // block[i] == 0 ? -1 : 0 yading@10: "psrlw $4, %%mm0 \n\t" yading@10: "psrlw $4, %%mm1 \n\t" yading@10: "pxor %%mm2, %%mm0 \n\t" yading@10: "pxor %%mm3, %%mm1 \n\t" yading@10: "psubw %%mm2, %%mm0 \n\t" yading@10: "psubw %%mm3, %%mm1 \n\t" yading@10: "pandn %%mm0, %%mm4 \n\t" yading@10: "pandn %%mm1, %%mm5 \n\t" yading@10: "pxor %%mm4, %%mm7 \n\t" yading@10: "pxor %%mm5, %%mm7 \n\t" yading@10: "movq %%mm4, (%0, %%"REG_a") \n\t" yading@10: "movq %%mm5, 8(%0, %%"REG_a") \n\t" yading@10: yading@10: "add $16, %%"REG_a" \n\t" yading@10: "jng 1b \n\t" yading@10: "movd 124(%0, %3), %%mm0 \n\t" yading@10: "movq %%mm7, %%mm6 \n\t" yading@10: "psrlq $32, %%mm7 \n\t" yading@10: "pxor %%mm6, %%mm7 \n\t" yading@10: "movq %%mm7, %%mm6 \n\t" yading@10: "psrlq $16, %%mm7 \n\t" yading@10: "pxor %%mm6, %%mm7 \n\t" yading@10: "pslld $31, %%mm7 \n\t" yading@10: "psrlq $15, %%mm7 \n\t" yading@10: "pxor %%mm7, %%mm0 \n\t" yading@10: "movd %%mm0, 124(%0, %3) \n\t" yading@10: yading@10: ::"r" (block+nCoeffs), "r"(quant_matrix+nCoeffs), "rm" (qscale), "r" (-2*nCoeffs) yading@10: : "%"REG_a, "memory" yading@10: ); yading@10: } yading@10: yading@10: static void denoise_dct_mmx(MpegEncContext *s, int16_t *block){ yading@10: const int intra= s->mb_intra; yading@10: int *sum= s->dct_error_sum[intra]; yading@10: uint16_t *offset= s->dct_offset[intra]; yading@10: yading@10: s->dct_count[intra]++; yading@10: yading@10: __asm__ volatile( yading@10: "pxor %%mm7, %%mm7 \n\t" yading@10: "1: \n\t" yading@10: "pxor %%mm0, %%mm0 \n\t" yading@10: "pxor %%mm1, %%mm1 \n\t" yading@10: "movq (%0), %%mm2 \n\t" yading@10: "movq 8(%0), %%mm3 \n\t" yading@10: "pcmpgtw %%mm2, %%mm0 \n\t" yading@10: "pcmpgtw %%mm3, %%mm1 \n\t" yading@10: "pxor %%mm0, %%mm2 \n\t" yading@10: "pxor %%mm1, %%mm3 \n\t" yading@10: "psubw %%mm0, %%mm2 \n\t" yading@10: "psubw %%mm1, %%mm3 \n\t" yading@10: "movq %%mm2, %%mm4 \n\t" yading@10: "movq %%mm3, %%mm5 \n\t" yading@10: "psubusw (%2), %%mm2 \n\t" yading@10: "psubusw 8(%2), %%mm3 \n\t" yading@10: "pxor %%mm0, %%mm2 \n\t" yading@10: "pxor %%mm1, %%mm3 \n\t" yading@10: "psubw %%mm0, %%mm2 \n\t" yading@10: "psubw %%mm1, %%mm3 \n\t" yading@10: "movq %%mm2, (%0) \n\t" yading@10: "movq %%mm3, 8(%0) \n\t" yading@10: "movq %%mm4, %%mm2 \n\t" yading@10: "movq %%mm5, %%mm3 \n\t" yading@10: "punpcklwd %%mm7, %%mm4 \n\t" yading@10: "punpckhwd %%mm7, %%mm2 \n\t" yading@10: "punpcklwd %%mm7, %%mm5 \n\t" yading@10: "punpckhwd %%mm7, %%mm3 \n\t" yading@10: "paddd (%1), %%mm4 \n\t" yading@10: "paddd 8(%1), %%mm2 \n\t" yading@10: "paddd 16(%1), %%mm5 \n\t" yading@10: "paddd 24(%1), %%mm3 \n\t" yading@10: "movq %%mm4, (%1) \n\t" yading@10: "movq %%mm2, 8(%1) \n\t" yading@10: "movq %%mm5, 16(%1) \n\t" yading@10: "movq %%mm3, 24(%1) \n\t" yading@10: "add $16, %0 \n\t" yading@10: "add $32, %1 \n\t" yading@10: "add $16, %2 \n\t" yading@10: "cmp %3, %0 \n\t" yading@10: " jb 1b \n\t" yading@10: : "+r" (block), "+r" (sum), "+r" (offset) yading@10: : "r"(block+64) yading@10: ); yading@10: } yading@10: yading@10: static void denoise_dct_sse2(MpegEncContext *s, int16_t *block){ yading@10: const int intra= s->mb_intra; yading@10: int *sum= s->dct_error_sum[intra]; yading@10: uint16_t *offset= s->dct_offset[intra]; yading@10: yading@10: s->dct_count[intra]++; yading@10: yading@10: __asm__ volatile( yading@10: "pxor %%xmm7, %%xmm7 \n\t" yading@10: "1: \n\t" yading@10: "pxor %%xmm0, %%xmm0 \n\t" yading@10: "pxor %%xmm1, %%xmm1 \n\t" yading@10: "movdqa (%0), %%xmm2 \n\t" yading@10: "movdqa 16(%0), %%xmm3 \n\t" yading@10: "pcmpgtw %%xmm2, %%xmm0 \n\t" yading@10: "pcmpgtw %%xmm3, %%xmm1 \n\t" yading@10: "pxor %%xmm0, %%xmm2 \n\t" yading@10: "pxor %%xmm1, %%xmm3 \n\t" yading@10: "psubw %%xmm0, %%xmm2 \n\t" yading@10: "psubw %%xmm1, %%xmm3 \n\t" yading@10: "movdqa %%xmm2, %%xmm4 \n\t" yading@10: "movdqa %%xmm3, %%xmm5 \n\t" yading@10: "psubusw (%2), %%xmm2 \n\t" yading@10: "psubusw 16(%2), %%xmm3 \n\t" yading@10: "pxor %%xmm0, %%xmm2 \n\t" yading@10: "pxor %%xmm1, %%xmm3 \n\t" yading@10: "psubw %%xmm0, %%xmm2 \n\t" yading@10: "psubw %%xmm1, %%xmm3 \n\t" yading@10: "movdqa %%xmm2, (%0) \n\t" yading@10: "movdqa %%xmm3, 16(%0) \n\t" yading@10: "movdqa %%xmm4, %%xmm6 \n\t" yading@10: "movdqa %%xmm5, %%xmm0 \n\t" yading@10: "punpcklwd %%xmm7, %%xmm4 \n\t" yading@10: "punpckhwd %%xmm7, %%xmm6 \n\t" yading@10: "punpcklwd %%xmm7, %%xmm5 \n\t" yading@10: "punpckhwd %%xmm7, %%xmm0 \n\t" yading@10: "paddd (%1), %%xmm4 \n\t" yading@10: "paddd 16(%1), %%xmm6 \n\t" yading@10: "paddd 32(%1), %%xmm5 \n\t" yading@10: "paddd 48(%1), %%xmm0 \n\t" yading@10: "movdqa %%xmm4, (%1) \n\t" yading@10: "movdqa %%xmm6, 16(%1) \n\t" yading@10: "movdqa %%xmm5, 32(%1) \n\t" yading@10: "movdqa %%xmm0, 48(%1) \n\t" yading@10: "add $32, %0 \n\t" yading@10: "add $64, %1 \n\t" yading@10: "add $32, %2 \n\t" yading@10: "cmp %3, %0 \n\t" yading@10: " jb 1b \n\t" yading@10: : "+r" (block), "+r" (sum), "+r" (offset) yading@10: : "r"(block+64) yading@10: XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3", yading@10: "%xmm4", "%xmm5", "%xmm6", "%xmm7") yading@10: ); yading@10: } yading@10: yading@10: #endif /* HAVE_INLINE_ASM */ yading@10: yading@10: av_cold void ff_MPV_common_init_x86(MpegEncContext *s) yading@10: { yading@10: #if HAVE_INLINE_ASM yading@10: int mm_flags = av_get_cpu_flags(); yading@10: yading@10: if (mm_flags & AV_CPU_FLAG_MMX) { yading@10: s->dct_unquantize_h263_intra = dct_unquantize_h263_intra_mmx; yading@10: s->dct_unquantize_h263_inter = dct_unquantize_h263_inter_mmx; yading@10: s->dct_unquantize_mpeg1_intra = dct_unquantize_mpeg1_intra_mmx; yading@10: s->dct_unquantize_mpeg1_inter = dct_unquantize_mpeg1_inter_mmx; yading@10: if(!(s->flags & CODEC_FLAG_BITEXACT)) yading@10: s->dct_unquantize_mpeg2_intra = dct_unquantize_mpeg2_intra_mmx; yading@10: s->dct_unquantize_mpeg2_inter = dct_unquantize_mpeg2_inter_mmx; yading@10: yading@10: if (mm_flags & AV_CPU_FLAG_SSE2) { yading@10: s->denoise_dct= denoise_dct_sse2; yading@10: } else { yading@10: s->denoise_dct= denoise_dct_mmx; yading@10: } yading@10: } yading@10: #endif /* HAVE_INLINE_ASM */ yading@10: }