Quelle  convolve8_vert_dspr2.c   Sprache: C

/*
 *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
 *
 *  Use of this source code is governed by a BSD-style license
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
 *  in the file PATENTS.  All contributing project authors may
 *  be found in the AUTHORS file in the root of the source tree.
 */

#include <assert.h>
#include <stdio.h>

#include "./vpx_dsp_rtcd.h"
#include "vpx_dsp/mips/convolve_common_dspr2.h"
#include "vpx_dsp/vpx_dsp_common.h"
#include "vpx_dsp/vpx_filter.h"
#include "vpx_ports/mem.h"

#if HAVE_DSPR2
static void convolve_vert_4_dspr2(const uint8_t *src, int32_t src_stride,
                                  uint8_t *dst, int32_t dst_stride,
                                  const int16_t *filter_y, int32_t w,
                                  int32_t h) {
  int32_t x, y;
  const uint8_t *src_ptr;
  uint8_t *dst_ptr;
  uint8_t *cm = vpx_ff_cropTbl;
  uint32_t vector4a = 64;
  uint32_t load1, load2, load3, load4;
  uint32_t p1, p2;
  uint32_t n1, n2;
  uint32_t scratch1, scratch2;
  uint32_t store1, store2;
  int32_t vector1b, vector2b, vector3b, vector4b;
  int32_t Temp1, Temp2;

  vector1b = ((const int32_t *)filter_y)[0];
  vector2b = ((const int32_t *)filter_y)[1];
  vector3b = ((const int32_t *)filter_y)[2];
  vector4b = ((const int32_t *)filter_y)[3];

  src -= 3 * src_stride;

  for (y = h; y--;) {
    /* prefetch data to cache memory */
    prefetch_store(dst + dst_stride);

    for (x = 0; x < w; x += 4) {
      src_ptr = src + x;
      dst_ptr = dst + x;

      __asm__ __volatile__(
          "ulw %[load1], 0(%[src_ptr]) \n\t"
          "add %[src_ptr], %[src_ptr], %[src_stride] \n\t"
          "ulw %[load2], 0(%[src_ptr]) \n\t"
          "add %[src_ptr], %[src_ptr], %[src_stride] \n\t"
          "ulw %[load3], 0(%[src_ptr]) \n\t"
          "add %[src_ptr], %[src_ptr], %[src_stride] \n\t"
          "ulw %[load4], 0(%[src_ptr]) \n\t"

          "mtlo %[vector4a], $ac0 \n\t"
          "mtlo %[vector4a], $ac1 \n\t"
          "mtlo %[vector4a], $ac2 \n\t"
          "mtlo %[vector4a], $ac3 \n\t"
          "mthi $zero, $ac0 \n\t"
          "mthi $zero, $ac1 \n\t"
          "mthi $zero, $ac2 \n\t"
          "mthi $zero, $ac3 \n\t"

          "preceu.ph.qbr %[scratch1], %[load1] \n\t"
          "preceu.ph.qbr %[p1], %[load2] \n\t"
          "precrq.ph.w %[n1], %[p1], %[scratch1] \n\t" /* pixel 2 */
          "append %[p1], %[scratch1], 16 \n\t" /* pixel 1 */
          "preceu.ph.qbr %[scratch2], %[load3] \n\t"
          "preceu.ph.qbr %[p2], %[load4] \n\t"
          "precrq.ph.w %[n2], %[p2], %[scratch2] \n\t" /* pixel 2 */
          "append %[p2], %[scratch2], 16 \n\t" /* pixel 1 */

          "dpa.w.ph $ac0, %[p1], %[vector1b] \n\t"
          "dpa.w.ph $ac0, %[p2], %[vector2b] \n\t"
          "dpa.w.ph $ac1, %[n1], %[vector1b] \n\t"
          "dpa.w.ph $ac1, %[n2], %[vector2b] \n\t"

          "preceu.ph.qbl %[scratch1], %[load1] \n\t"
          "preceu.ph.qbl %[p1], %[load2] \n\t"
          "precrq.ph.w %[n1], %[p1], %[scratch1] \n\t" /* pixel 2 */
          "append %[p1], %[scratch1], 16 \n\t" /* pixel 1 */
          "preceu.ph.qbl %[scratch2], %[load3] \n\t"
          "preceu.ph.qbl %[p2], %[load4] \n\t"
          "precrq.ph.w %[n2], %[p2], %[scratch2] \n\t" /* pixel 2 */
          "append %[p2], %[scratch2], 16 \n\t" /* pixel 1 */

          "dpa.w.ph $ac2, %[p1], %[vector1b] \n\t"
          "dpa.w.ph $ac2, %[p2], %[vector2b] \n\t"
          "dpa.w.ph $ac3, %[n1], %[vector1b] \n\t"
          "dpa.w.ph $ac3, %[n2], %[vector2b] \n\t"

          "add %[src_ptr], %[src_ptr], %[src_stride] \n\t"
          "ulw %[load1], 0(%[src_ptr]) \n\t"
          "add %[src_ptr], %[src_ptr], %[src_stride] \n\t"
          "ulw %[load2], 0(%[src_ptr]) \n\t"
          "add %[src_ptr], %[src_ptr], %[src_stride] \n\t"
          "ulw %[load3], 0(%[src_ptr]) \n\t"
          "add %[src_ptr], %[src_ptr], %[src_stride] \n\t"
          "ulw %[load4], 0(%[src_ptr]) \n\t"

          "preceu.ph.qbr %[scratch1], %[load1] \n\t"
          "preceu.ph.qbr %[p1], %[load2] \n\t"
          "precrq.ph.w %[n1], %[p1], %[scratch1] \n\t" /* pixel 2 */
          "append %[p1], %[scratch1], 16 \n\t" /* pixel 1 */
          "preceu.ph.qbr %[scratch2], %[load3] \n\t"
          "preceu.ph.qbr %[p2], %[load4] \n\t"
          "precrq.ph.w %[n2], %[p2], %[scratch2] \n\t" /* pixel 2 */
          "append %[p2], %[scratch2], 16 \n\t" /* pixel 1 */

          "dpa.w.ph $ac0, %[p1], %[vector3b] \n\t"
          "dpa.w.ph $ac0, %[p2], %[vector4b] \n\t"
          "extp %[Temp1], $ac0, 31 \n\t"
          "dpa.w.ph $ac1, %[n1], %[vector3b] \n\t"
          "dpa.w.ph $ac1, %[n2], %[vector4b] \n\t"
          "extp %[Temp2], $ac1, 31 \n\t"

          "preceu.ph.qbl %[scratch1], %[load1] \n\t"
          "preceu.ph.qbl %[p1], %[load2] \n\t"
          "precrq.ph.w %[n1], %[p1], %[scratch1] \n\t" /* pixel 2 */
          "append %[p1], %[scratch1], 16 \n\t" /* pixel 1 */
          "preceu.ph.qbl %[scratch2], %[load3] \n\t"
          "preceu.ph.qbl %[p2], %[load4] \n\t"
          "precrq.ph.w %[n2], %[p2], %[scratch2] \n\t" /* pixel 2 */
          "append %[p2], %[scratch2], 16 \n\t" /* pixel 1 */

          "lbux %[store1], %[Temp1](%[cm]) \n\t"
          "dpa.w.ph $ac2, %[p1], %[vector3b] \n\t"
          "dpa.w.ph $ac2, %[p2], %[vector4b] \n\t"
          "extp %[Temp1], $ac2, 31 \n\t"

          "lbux %[store2], %[Temp2](%[cm]) \n\t"
          "dpa.w.ph $ac3, %[n1], %[vector3b] \n\t"
          "dpa.w.ph $ac3, %[n2], %[vector4b] \n\t"
          "extp %[Temp2], $ac3, 31 \n\t"

          "sb %[store1], 0(%[dst_ptr]) \n\t"
          "sb %[store2], 1(%[dst_ptr]) \n\t"

          "lbux %[store1], %[Temp1](%[cm]) \n\t"
          "lbux %[store2], %[Temp2](%[cm]) \n\t"

          "sb %[store1], 2(%[dst_ptr]) \n\t"
          "sb %[store2], 3(%[dst_ptr]) \n\t"

          : [load1] "=&r"(load1), [load2] "=&r"(load2), [load3] "=&r"(load3),
            [load4] "=&r"(load4), [p1] "=&r"(p1), [p2] "=&r"(p2),
            [n1] "=&r"(n1), [n2] "=&r"(n2), [scratch1] "=&r"(scratch1),
            [scratch2] "=&r"(scratch2), [Temp1] "=&r"(Temp1),
            [Temp2] "=&r"(Temp2), [store1] "=&r"(store1),
            [store2] "=&r"(store2), [src_ptr] "+r"(src_ptr)
          : [vector1b] "r"(vector1b), [vector2b] "r"(vector2b),
            [vector3b] "r"(vector3b), [vector4b] "r"(vector4b),
            [vector4a] "r"(vector4a), [src_stride] "r"(src_stride),
            [cm] "r"(cm), [dst_ptr] "r"(dst_ptr));
    }

    /* Next row... */
    src += src_stride;
    dst += dst_stride;
  }
}

static void convolve_vert_64_dspr2(const uint8_t *src, int32_t src_stride,
                                   uint8_t *dst, int32_t dst_stride,
                                   const int16_t *filter_y, int32_t h) {
  int32_t x, y;
  const uint8_t *src_ptr;
  uint8_t *dst_ptr;
  uint8_t *cm = vpx_ff_cropTbl;
  uint32_t vector4a = 64;
  uint32_t load1, load2, load3, load4;
  uint32_t p1, p2;
  uint32_t n1, n2;
  uint32_t scratch1, scratch2;
  uint32_t store1, store2;
  int32_t vector1b, vector2b, vector3b, vector4b;
  int32_t Temp1, Temp2;

  vector1b = ((const int32_t *)filter_y)[0];
  vector2b = ((const int32_t *)filter_y)[1];
  vector3b = ((const int32_t *)filter_y)[2];
  vector4b = ((const int32_t *)filter_y)[3];

  src -= 3 * src_stride;

  for (y = h; y--;) {
    /* prefetch data to cache memory */
    prefetch_store(dst + dst_stride);
    prefetch_store(dst + dst_stride + 32);

    for (x = 0; x < 64; x += 4) {
      src_ptr = src + x;
      dst_ptr = dst + x;

      __asm__ __volatile__(
          "ulw %[load1], 0(%[src_ptr]) \n\t"
          "add %[src_ptr], %[src_ptr], %[src_stride] \n\t"
          "ulw %[load2], 0(%[src_ptr]) \n\t"
          "add %[src_ptr], %[src_ptr], %[src_stride] \n\t"
          "ulw %[load3], 0(%[src_ptr]) \n\t"
          "add %[src_ptr], %[src_ptr], %[src_stride] \n\t"
          "ulw %[load4], 0(%[src_ptr]) \n\t"

          "mtlo %[vector4a], $ac0 \n\t"
          "mtlo %[vector4a], $ac1 \n\t"
          "mtlo %[vector4a], $ac2 \n\t"
          "mtlo %[vector4a], $ac3 \n\t"
          "mthi $zero, $ac0 \n\t"
          "mthi $zero, $ac1 \n\t"
          "mthi $zero, $ac2 \n\t"
          "mthi $zero, $ac3 \n\t"

          "preceu.ph.qbr %[scratch1], %[load1] \n\t"
          "preceu.ph.qbr %[p1], %[load2] \n\t"
          "precrq.ph.w %[n1], %[p1], %[scratch1] \n\t" /* pixel 2 */
          "append %[p1], %[scratch1], 16 \n\t" /* pixel 1 */
          "preceu.ph.qbr %[scratch2], %[load3] \n\t"
          "preceu.ph.qbr %[p2], %[load4] \n\t"
          "precrq.ph.w %[n2], %[p2], %[scratch2] \n\t" /* pixel 2 */
          "append %[p2], %[scratch2], 16 \n\t" /* pixel 1 */

          "dpa.w.ph $ac0, %[p1], %[vector1b] \n\t"
          "dpa.w.ph $ac0, %[p2], %[vector2b] \n\t"
          "dpa.w.ph $ac1, %[n1], %[vector1b] \n\t"
          "dpa.w.ph $ac1, %[n2], %[vector2b] \n\t"

          "preceu.ph.qbl %[scratch1], %[load1] \n\t"
          "preceu.ph.qbl %[p1], %[load2] \n\t"
          "precrq.ph.w %[n1], %[p1], %[scratch1] \n\t" /* pixel 2 */
          "append %[p1], %[scratch1], 16 \n\t" /* pixel 1 */
          "preceu.ph.qbl %[scratch2], %[load3] \n\t"
          "preceu.ph.qbl %[p2], %[load4] \n\t"
          "precrq.ph.w %[n2], %[p2], %[scratch2] \n\t" /* pixel 2 */
          "append %[p2], %[scratch2], 16 \n\t" /* pixel 1 */

          "dpa.w.ph $ac2, %[p1], %[vector1b] \n\t"
          "dpa.w.ph $ac2, %[p2], %[vector2b] \n\t"
          "dpa.w.ph $ac3, %[n1], %[vector1b] \n\t"
          "dpa.w.ph $ac3, %[n2], %[vector2b] \n\t"

          "add %[src_ptr], %[src_ptr], %[src_stride] \n\t"
          "ulw %[load1], 0(%[src_ptr]) \n\t"
          "add %[src_ptr], %[src_ptr], %[src_stride] \n\t"
          "ulw %[load2], 0(%[src_ptr]) \n\t"
          "add %[src_ptr], %[src_ptr], %[src_stride] \n\t"
          "ulw %[load3], 0(%[src_ptr]) \n\t"
          "add %[src_ptr], %[src_ptr], %[src_stride] \n\t"
          "ulw %[load4], 0(%[src_ptr]) \n\t"

          "preceu.ph.qbr %[scratch1], %[load1] \n\t"
          "preceu.ph.qbr %[p1], %[load2] \n\t"
          "precrq.ph.w %[n1], %[p1], %[scratch1] \n\t" /* pixel 2 */
          "append %[p1], %[scratch1], 16 \n\t" /* pixel 1 */
          "preceu.ph.qbr %[scratch2], %[load3] \n\t"
          "preceu.ph.qbr %[p2], %[load4] \n\t"
          "precrq.ph.w %[n2], %[p2], %[scratch2] \n\t" /* pixel 2 */
          "append %[p2], %[scratch2], 16 \n\t" /* pixel 1 */

          "dpa.w.ph $ac0, %[p1], %[vector3b] \n\t"
          "dpa.w.ph $ac0, %[p2], %[vector4b] \n\t"
          "extp %[Temp1], $ac0, 31 \n\t"
          "dpa.w.ph $ac1, %[n1], %[vector3b] \n\t"
          "dpa.w.ph $ac1, %[n2], %[vector4b] \n\t"
          "extp %[Temp2], $ac1, 31 \n\t"

          "preceu.ph.qbl %[scratch1], %[load1] \n\t"
          "preceu.ph.qbl %[p1], %[load2] \n\t"
          "precrq.ph.w %[n1], %[p1], %[scratch1] \n\t" /* pixel 2 */
          "append %[p1], %[scratch1], 16 \n\t" /* pixel 1 */
          "preceu.ph.qbl %[scratch2], %[load3] \n\t"
          "preceu.ph.qbl %[p2], %[load4] \n\t"
          "precrq.ph.w %[n2], %[p2], %[scratch2] \n\t" /* pixel 2 */
          "append %[p2], %[scratch2], 16 \n\t" /* pixel 1 */

          "lbux %[store1], %[Temp1](%[cm]) \n\t"
          "dpa.w.ph $ac2, %[p1], %[vector3b] \n\t"
          "dpa.w.ph $ac2, %[p2], %[vector4b] \n\t"
          "extp %[Temp1], $ac2, 31 \n\t"

          "lbux %[store2], %[Temp2](%[cm]) \n\t"
          "dpa.w.ph $ac3, %[n1], %[vector3b] \n\t"
          "dpa.w.ph $ac3, %[n2], %[vector4b] \n\t"
          "extp %[Temp2], $ac3, 31 \n\t"

          "sb %[store1], 0(%[dst_ptr]) \n\t"
          "sb %[store2], 1(%[dst_ptr]) \n\t"

          "lbux %[store1], %[Temp1](%[cm]) \n\t"
          "lbux %[store2], %[Temp2](%[cm]) \n\t"

          "sb %[store1], 2(%[dst_ptr]) \n\t"
          "sb %[store2], 3(%[dst_ptr]) \n\t"

          : [load1] "=&r"(load1), [load2] "=&r"(load2), [load3] "=&r"(load3),
            [load4] "=&r"(load4), [p1] "=&r"(p1), [p2] "=&r"(p2),
            [n1] "=&r"(n1), [n2] "=&r"(n2), [scratch1] "=&r"(scratch1),
            [scratch2] "=&r"(scratch2), [Temp1] "=&r"(Temp1),
            [Temp2] "=&r"(Temp2), [store1] "=&r"(store1),
            [store2] "=&r"(store2), [src_ptr] "+r"(src_ptr)
          : [vector1b] "r"(vector1b), [vector2b] "r"(vector2b),
            [vector3b] "r"(vector3b), [vector4b] "r"(vector4b),
            [vector4a] "r"(vector4a), [src_stride] "r"(src_stride),
            [cm] "r"(cm), [dst_ptr] "r"(dst_ptr));
    }

    /* Next row... */
    src += src_stride;
    dst += dst_stride;
  }
}

void vpx_convolve8_vert_dspr2(const uint8_t *src, ptrdiff_t src_stride,
                              uint8_t *dst, ptrdiff_t dst_stride,
                              const InterpKernel *filter, int x0_q4,
                              int x_step_q4, int y0_q4, int y_step_q4, int w,
                              int h) {
  const int16_t *const filter_y = filter[y0_q4];
  assert(y_step_q4 == 16);
  assert(((const int32_t *)filter_y)[1] != 0x800000);

  if (vpx_get_filter_taps(filter_y) == 2) {
    vpx_convolve2_vert_dspr2(src, src_stride, dst, dst_stride, filter, x0_q4,
                             x_step_q4, y0_q4, y_step_q4, w, h);
  } else {
    uint32_t pos = 38;

    /* bit positon for extract from acc */
    __asm__ __volatile__("wrdsp %[pos], 1 \n\t"
                         :
                         : [pos] "r"(pos));

    prefetch_store(dst);

    switch (w) {
      case 4:
      case 8:
      case 16:
      case 32:
        convolve_vert_4_dspr2(src, src_stride, dst, dst_stride, filter_y, w, h);
        break;
      case 64:
        prefetch_store(dst + 32);
        convolve_vert_64_dspr2(src, src_stride, dst, dst_stride, filter_y, h);
        break;
      default:
        vpx_convolve8_vert_c(src, src_stride, dst, dst_stride, filter, x0_q4,
                             x_step_q4, y0_q4, y_step_q4, w, h);
        break;
    }
  }
}

#endif