michael@0: /*
michael@0:  *  Copyright (c) 2013 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 <assert.h>
michael@0: 
michael@0: #include "./vpx_config.h"
michael@0: #include "./vp9_rtcd.h"
michael@0: #include "vp9/common/vp9_common.h"
michael@0: #include "vp9/common/vp9_convolve.h"
michael@0: #include "vp9/common/vp9_filter.h"
michael@0: #include "vpx/vpx_integer.h"
michael@0: #include "vpx_ports/mem.h"
michael@0: 
michael@0: static void convolve_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
michael@0:                              uint8_t *dst, ptrdiff_t dst_stride,
michael@0:                              const int16_t *filter_x0, int x_step_q4,
michael@0:                              const int16_t *filter_y, int y_step_q4,
michael@0:                              int w, int h, int taps) {
michael@0:   int x, y, k;
michael@0: 
michael@0:   /* NOTE: This assumes that the filter table is 256-byte aligned. */
michael@0:   /* TODO(agrange) Modify to make independent of table alignment. */
michael@0:   const int16_t *const filter_x_base =
michael@0:       (const int16_t *)(((intptr_t)filter_x0) & ~(intptr_t)0xff);
michael@0: 
michael@0:   /* Adjust base pointer address for this source line */
michael@0:   src -= taps / 2 - 1;
michael@0: 
michael@0:   for (y = 0; y < h; ++y) {
michael@0:     /* Initial phase offset */
michael@0:     int x_q4 = (int)(filter_x0 - filter_x_base) / taps;
michael@0: 
michael@0:     for (x = 0; x < w; ++x) {
michael@0:       /* Per-pixel src offset */
michael@0:       const int src_x = x_q4 >> SUBPEL_BITS;
michael@0:       int sum = 0;
michael@0: 
michael@0:       /* Pointer to filter to use */
michael@0:       const int16_t *const filter_x = filter_x_base +
michael@0:           (x_q4 & SUBPEL_MASK) * taps;
michael@0: 
michael@0:       for (k = 0; k < taps; ++k)
michael@0:         sum += src[src_x + k] * filter_x[k];
michael@0: 
michael@0:       dst[x] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
michael@0: 
michael@0:       /* Move to the next source pixel */
michael@0:       x_q4 += x_step_q4;
michael@0:     }
michael@0:     src += src_stride;
michael@0:     dst += dst_stride;
michael@0:   }
michael@0: }
michael@0: 
michael@0: static void convolve_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
michael@0:                                  uint8_t *dst, ptrdiff_t dst_stride,
michael@0:                                  const int16_t *filter_x0, int x_step_q4,
michael@0:                                  const int16_t *filter_y, int y_step_q4,
michael@0:                                  int w, int h, int taps) {
michael@0:   int x, y, k;
michael@0: 
michael@0:   /* NOTE: This assumes that the filter table is 256-byte aligned. */
michael@0:   /* TODO(agrange) Modify to make independent of table alignment. */
michael@0:   const int16_t *const filter_x_base =
michael@0:       (const int16_t *)(((intptr_t)filter_x0) & ~(intptr_t)0xff);
michael@0: 
michael@0:   /* Adjust base pointer address for this source line */
michael@0:   src -= taps / 2 - 1;
michael@0: 
michael@0:   for (y = 0; y < h; ++y) {
michael@0:     /* Initial phase offset */
michael@0:     int x_q4 = (int)(filter_x0 - filter_x_base) / taps;
michael@0: 
michael@0:     for (x = 0; x < w; ++x) {
michael@0:       /* Per-pixel src offset */
michael@0:       const int src_x = x_q4 >> SUBPEL_BITS;
michael@0:       int sum = 0;
michael@0: 
michael@0:       /* Pointer to filter to use */
michael@0:       const int16_t *const filter_x = filter_x_base +
michael@0:           (x_q4 & SUBPEL_MASK) * taps;
michael@0: 
michael@0:       for (k = 0; k < taps; ++k)
michael@0:         sum += src[src_x + k] * filter_x[k];
michael@0: 
michael@0:       dst[x] = ROUND_POWER_OF_TWO(dst[x] +
michael@0:                    clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)), 1);
michael@0: 
michael@0:       /* Move to the next source pixel */
michael@0:       x_q4 += x_step_q4;
michael@0:     }
michael@0:     src += src_stride;
michael@0:     dst += dst_stride;
michael@0:   }
michael@0: }
michael@0: 
michael@0: static void convolve_vert_c(const uint8_t *src, ptrdiff_t src_stride,
michael@0:                             uint8_t *dst, ptrdiff_t dst_stride,
michael@0:                             const int16_t *filter_x, int x_step_q4,
michael@0:                             const int16_t *filter_y0, int y_step_q4,
michael@0:                             int w, int h, int taps) {
michael@0:   int x, y, k;
michael@0: 
michael@0:   /* NOTE: This assumes that the filter table is 256-byte aligned. */
michael@0:   /* TODO(agrange) Modify to make independent of table alignment. */
michael@0:   const int16_t *const filter_y_base =
michael@0:       (const int16_t *)(((intptr_t)filter_y0) & ~(intptr_t)0xff);
michael@0: 
michael@0:   /* Adjust base pointer address for this source column */
michael@0:   src -= src_stride * (taps / 2 - 1);
michael@0: 
michael@0:   for (x = 0; x < w; ++x) {
michael@0:     /* Initial phase offset */
michael@0:     int y_q4 = (int)(filter_y0 - filter_y_base) / taps;
michael@0: 
michael@0:     for (y = 0; y < h; ++y) {
michael@0:       /* Per-pixel src offset */
michael@0:       const int src_y = y_q4 >> SUBPEL_BITS;
michael@0:       int sum = 0;
michael@0: 
michael@0:       /* Pointer to filter to use */
michael@0:       const int16_t *const filter_y = filter_y_base +
michael@0:           (y_q4 & SUBPEL_MASK) * taps;
michael@0: 
michael@0:       for (k = 0; k < taps; ++k)
michael@0:         sum += src[(src_y + k) * src_stride] * filter_y[k];
michael@0: 
michael@0:       dst[y * dst_stride] =
michael@0:           clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
michael@0: 
michael@0:       /* Move to the next source pixel */
michael@0:       y_q4 += y_step_q4;
michael@0:     }
michael@0:     ++src;
michael@0:     ++dst;
michael@0:   }
michael@0: }
michael@0: 
michael@0: static void convolve_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
michael@0:                                 uint8_t *dst, ptrdiff_t dst_stride,
michael@0:                                 const int16_t *filter_x, int x_step_q4,
michael@0:                                 const int16_t *filter_y0, int y_step_q4,
michael@0:                                 int w, int h, int taps) {
michael@0:   int x, y, k;
michael@0: 
michael@0:   /* NOTE: This assumes that the filter table is 256-byte aligned. */
michael@0:   /* TODO(agrange) Modify to make independent of table alignment. */
michael@0:   const int16_t *const filter_y_base =
michael@0:       (const int16_t *)(((intptr_t)filter_y0) & ~(intptr_t)0xff);
michael@0: 
michael@0:   /* Adjust base pointer address for this source column */
michael@0:   src -= src_stride * (taps / 2 - 1);
michael@0: 
michael@0:   for (x = 0; x < w; ++x) {
michael@0:     /* Initial phase offset */
michael@0:     int y_q4 = (int)(filter_y0 - filter_y_base) / taps;
michael@0: 
michael@0:     for (y = 0; y < h; ++y) {
michael@0:       /* Per-pixel src offset */
michael@0:       const int src_y = y_q4 >> SUBPEL_BITS;
michael@0:       int sum = 0;
michael@0: 
michael@0:       /* Pointer to filter to use */
michael@0:       const int16_t *const filter_y = filter_y_base +
michael@0:           (y_q4 & SUBPEL_MASK) * taps;
michael@0: 
michael@0:       for (k = 0; k < taps; ++k)
michael@0:         sum += src[(src_y + k) * src_stride] * filter_y[k];
michael@0: 
michael@0:       dst[y * dst_stride] = ROUND_POWER_OF_TWO(dst[y * dst_stride] +
michael@0:            clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)), 1);
michael@0: 
michael@0:       /* Move to the next source pixel */
michael@0:       y_q4 += y_step_q4;
michael@0:     }
michael@0:     ++src;
michael@0:     ++dst;
michael@0:   }
michael@0: }
michael@0: 
michael@0: static void convolve_c(const uint8_t *src, ptrdiff_t src_stride,
michael@0:                        uint8_t *dst, ptrdiff_t dst_stride,
michael@0:                        const int16_t *filter_x, int x_step_q4,
michael@0:                        const int16_t *filter_y, int y_step_q4,
michael@0:                        int w, int h, int taps) {
michael@0:   /* Fixed size intermediate buffer places limits on parameters.
michael@0:    * Maximum intermediate_height is 324, for y_step_q4 == 80,
michael@0:    * h == 64, taps == 8.
michael@0:    * y_step_q4 of 80 allows for 1/10 scale for 5 layer svc
michael@0:    */
michael@0:   uint8_t temp[64 * 324];
michael@0:   int intermediate_height = (((h - 1) * y_step_q4 + 15) >> 4) + taps;
michael@0: 
michael@0:   assert(w <= 64);
michael@0:   assert(h <= 64);
michael@0:   assert(taps <= 8);
michael@0:   assert(y_step_q4 <= 80);
michael@0:   assert(x_step_q4 <= 80);
michael@0: 
michael@0:   if (intermediate_height < h)
michael@0:     intermediate_height = h;
michael@0: 
michael@0:   convolve_horiz_c(src - src_stride * (taps / 2 - 1), src_stride, temp, 64,
michael@0:                    filter_x, x_step_q4, filter_y, y_step_q4, w,
michael@0:                    intermediate_height, taps);
michael@0:   convolve_vert_c(temp + 64 * (taps / 2 - 1), 64, dst, dst_stride, filter_x,
michael@0:                   x_step_q4, filter_y, y_step_q4, w, h, taps);
michael@0: }
michael@0: 
michael@0: void vp9_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
michael@0:                            uint8_t *dst, ptrdiff_t dst_stride,
michael@0:                            const int16_t *filter_x, int x_step_q4,
michael@0:                            const int16_t *filter_y, int y_step_q4,
michael@0:                            int w, int h) {
michael@0:   convolve_horiz_c(src, src_stride, dst, dst_stride,
michael@0:                    filter_x, x_step_q4, filter_y, y_step_q4, w, h, 8);
michael@0: }
michael@0: 
michael@0: void vp9_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
michael@0:                                uint8_t *dst, ptrdiff_t dst_stride,
michael@0:                                const int16_t *filter_x, int x_step_q4,
michael@0:                                const int16_t *filter_y, int y_step_q4,
michael@0:                                int w, int h) {
michael@0:   convolve_avg_horiz_c(src, src_stride, dst, dst_stride,
michael@0:                        filter_x, x_step_q4, filter_y, y_step_q4, w, h, 8);
michael@0: }
michael@0: 
michael@0: void vp9_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride,
michael@0:                           uint8_t *dst, ptrdiff_t dst_stride,
michael@0:                           const int16_t *filter_x, int x_step_q4,
michael@0:                           const int16_t *filter_y, int y_step_q4,
michael@0:                           int w, int h) {
michael@0:   convolve_vert_c(src, src_stride, dst, dst_stride,
michael@0:                   filter_x, x_step_q4, filter_y, y_step_q4, w, h, 8);
michael@0: }
michael@0: 
michael@0: void vp9_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
michael@0:                               uint8_t *dst, ptrdiff_t dst_stride,
michael@0:                               const int16_t *filter_x, int x_step_q4,
michael@0:                               const int16_t *filter_y, int y_step_q4,
michael@0:                               int w, int h) {
michael@0:   convolve_avg_vert_c(src, src_stride, dst, dst_stride,
michael@0:                       filter_x, x_step_q4, filter_y, y_step_q4, w, h, 8);
michael@0: }
michael@0: 
michael@0: void vp9_convolve8_c(const uint8_t *src, ptrdiff_t src_stride,
michael@0:                      uint8_t *dst, ptrdiff_t dst_stride,
michael@0:                      const int16_t *filter_x, int x_step_q4,
michael@0:                      const int16_t *filter_y, int y_step_q4,
michael@0:                      int w, int h) {
michael@0:   convolve_c(src, src_stride, dst, dst_stride,
michael@0:              filter_x, x_step_q4, filter_y, y_step_q4, w, h, 8);
michael@0: }
michael@0: 
michael@0: void vp9_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride,
michael@0:                          uint8_t *dst, ptrdiff_t dst_stride,
michael@0:                          const int16_t *filter_x, int x_step_q4,
michael@0:                          const int16_t *filter_y, int y_step_q4,
michael@0:                          int w, int h) {
michael@0:   /* Fixed size intermediate buffer places limits on parameters. */
michael@0:   DECLARE_ALIGNED_ARRAY(16, uint8_t, temp, 64 * 64);
michael@0:   assert(w <= 64);
michael@0:   assert(h <= 64);
michael@0: 
michael@0:   vp9_convolve8(src, src_stride, temp, 64,
michael@0:                filter_x, x_step_q4, filter_y, y_step_q4, w, h);
michael@0:   vp9_convolve_avg(temp, 64, dst, dst_stride, NULL, 0, NULL, 0, w, h);
michael@0: }
michael@0: 
michael@0: void vp9_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride,
michael@0:                          uint8_t *dst, ptrdiff_t dst_stride,
michael@0:                          const int16_t *filter_x, int filter_x_stride,
michael@0:                          const int16_t *filter_y, int filter_y_stride,
michael@0:                          int w, int h) {
michael@0:   int r;
michael@0: 
michael@0:   for (r = h; r > 0; --r) {
michael@0:     vpx_memcpy(dst, src, w);
michael@0:     src += src_stride;
michael@0:     dst += dst_stride;
michael@0:   }
michael@0: }
michael@0: 
michael@0: void vp9_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride,
michael@0:                         uint8_t *dst, ptrdiff_t dst_stride,
michael@0:                         const int16_t *filter_x, int filter_x_stride,
michael@0:                         const int16_t *filter_y, int filter_y_stride,
michael@0:                         int w, int h) {
michael@0:   int x, y;
michael@0: 
michael@0:   for (y = 0; y < h; ++y) {
michael@0:     for (x = 0; x < w; ++x)
michael@0:       dst[x] = ROUND_POWER_OF_TWO(dst[x] + src[x], 1);
michael@0: 
michael@0:     src += src_stride;
michael@0:     dst += dst_stride;
michael@0:   }
michael@0: }