michael@0: /*
michael@0:  *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
michael@0:  *
michael@0:  *  Use of this source code is governed by a BSD-style license
michael@0:  *  that can be found in the LICENSE file in the root of the source
michael@0:  *  tree. An additional intellectual property rights grant can be found
michael@0:  *  in the file PATENTS.  All contributing project authors may
michael@0:  *  be found in the AUTHORS file in the root of the source tree.
michael@0:  */
michael@0: 
michael@0: #include "vpx_config.h"
michael@0: #include "vp8_rtcd.h"
michael@0: #include "vp8/common/blockd.h"
michael@0: #include "vpx_mem/vpx_mem.h"
michael@0: 
michael@0: extern void vp8_dequantize_b_impl_mmx(short *sq, short *dq, short *q);
michael@0: 
michael@0: void vp8_dequantize_b_mmx(BLOCKD *d, short *DQC)
michael@0: {
michael@0:     short *sq = (short *) d->qcoeff;
michael@0:     short *dq = (short *) d->dqcoeff;
michael@0: 
michael@0:     vp8_dequantize_b_impl_mmx(sq, dq, DQC);
michael@0: }
michael@0: 
michael@0: void vp8_dequant_idct_add_y_block_mmx
michael@0:             (short *q, short *dq,
michael@0:              unsigned char *dst, int stride, char *eobs)
michael@0: {
michael@0:     int i;
michael@0: 
michael@0:     for (i = 0; i < 4; i++)
michael@0:     {
michael@0:         if (eobs[0] > 1)
michael@0:             vp8_dequant_idct_add_mmx (q, dq, dst, stride);
michael@0:         else if (eobs[0] == 1)
michael@0:         {
michael@0:             vp8_dc_only_idct_add_mmx (q[0]*dq[0], dst, stride, dst, stride);
michael@0:             vpx_memset(q, 0, 2 * sizeof(q[0]));
michael@0:         }
michael@0: 
michael@0:         if (eobs[1] > 1)
michael@0:             vp8_dequant_idct_add_mmx (q+16, dq, dst+4, stride);
michael@0:         else if (eobs[1] == 1)
michael@0:         {
michael@0:             vp8_dc_only_idct_add_mmx (q[16]*dq[0], dst+4, stride,
michael@0:                                       dst+4, stride);
michael@0:             vpx_memset(q + 16, 0, 2 * sizeof(q[0]));
michael@0:         }
michael@0: 
michael@0:         if (eobs[2] > 1)
michael@0:             vp8_dequant_idct_add_mmx (q+32, dq, dst+8, stride);
michael@0:         else if (eobs[2] == 1)
michael@0:         {
michael@0:             vp8_dc_only_idct_add_mmx (q[32]*dq[0], dst+8, stride,
michael@0:                                       dst+8, stride);
michael@0:             vpx_memset(q + 32, 0, 2 * sizeof(q[0]));
michael@0:         }
michael@0: 
michael@0:         if (eobs[3] > 1)
michael@0:             vp8_dequant_idct_add_mmx (q+48, dq, dst+12, stride);
michael@0:         else if (eobs[3] == 1)
michael@0:         {
michael@0:             vp8_dc_only_idct_add_mmx (q[48]*dq[0], dst+12, stride,
michael@0:                                       dst+12, stride);
michael@0:             vpx_memset(q + 48, 0, 2 * sizeof(q[0]));
michael@0:         }
michael@0: 
michael@0:         q    += 64;
michael@0:         dst  += 4*stride;
michael@0:         eobs += 4;
michael@0:     }
michael@0: }
michael@0: 
michael@0: void vp8_dequant_idct_add_uv_block_mmx
michael@0:             (short *q, short *dq,
michael@0:              unsigned char *dstu, unsigned char *dstv, int stride, char *eobs)
michael@0: {
michael@0:     int i;
michael@0: 
michael@0:     for (i = 0; i < 2; i++)
michael@0:     {
michael@0:         if (eobs[0] > 1)
michael@0:             vp8_dequant_idct_add_mmx (q, dq, dstu, stride);
michael@0:         else if (eobs[0] == 1)
michael@0:         {
michael@0:             vp8_dc_only_idct_add_mmx (q[0]*dq[0], dstu, stride, dstu, stride);
michael@0:             vpx_memset(q, 0, 2 * sizeof(q[0]));
michael@0:         }
michael@0: 
michael@0:         if (eobs[1] > 1)
michael@0:             vp8_dequant_idct_add_mmx (q+16, dq, dstu+4, stride);
michael@0:         else if (eobs[1] == 1)
michael@0:         {
michael@0:             vp8_dc_only_idct_add_mmx (q[16]*dq[0], dstu+4, stride,
michael@0:                                       dstu+4, stride);
michael@0:             vpx_memset(q + 16, 0, 2 * sizeof(q[0]));
michael@0:         }
michael@0: 
michael@0:         q    += 32;
michael@0:         dstu += 4*stride;
michael@0:         eobs += 2;
michael@0:     }
michael@0: 
michael@0:     for (i = 0; i < 2; i++)
michael@0:     {
michael@0:         if (eobs[0] > 1)
michael@0:             vp8_dequant_idct_add_mmx (q, dq, dstv, stride);
michael@0:         else if (eobs[0] == 1)
michael@0:         {
michael@0:             vp8_dc_only_idct_add_mmx (q[0]*dq[0], dstv, stride, dstv, stride);
michael@0:             vpx_memset(q, 0, 2 * sizeof(q[0]));
michael@0:         }
michael@0: 
michael@0:         if (eobs[1] > 1)
michael@0:             vp8_dequant_idct_add_mmx (q+16, dq, dstv+4, stride);
michael@0:         else if (eobs[1] == 1)
michael@0:         {
michael@0:             vp8_dc_only_idct_add_mmx (q[16]*dq[0], dstv+4, stride,
michael@0:                                       dstv+4, stride);
michael@0:             vpx_memset(q + 16, 0, 2 * sizeof(q[0]));
michael@0:         }
michael@0: 
michael@0:         q    += 32;
michael@0:         dstv += 4*stride;
michael@0:         eobs += 2;
michael@0:     }
michael@0: }