ref: 2c560c3c22d097981e842499551f0146145567b1
parent: 3c700052bef486efd155269a6b86eb755b76fcc2
author: Linfeng Zhang <linfengz@google.com>
date: Wed Sep 20 06:58:39 EDT 2017
Specialize 4 to 3 frame scaling in C Scale 3x3 block instead of 16x16 block in each loop. Disabled by default. Benefits: 1. Reduced number of different phase_scaler from 16 to 3. Optimization code will be smaller and faster. 2. Maximum phase_scaler drifting will be reduced from 5/16 to 1/24. (The drifting is 1/(3*16) in each step.) BUG=webm:1419 Change-Id: I59a1f7496d89a1b090498c935d30cfcf1d0c282b
--- a/vp9/encoder/vp9_frame_scale.c
+++ b/vp9/encoder/vp9_frame_scale.c
@@ -20,8 +20,6 @@
INTERP_FILTER filter_type, int phase_scaler) {const int src_w = src->y_crop_width;
const int src_h = src->y_crop_height;
- const int dst_w = dst->y_crop_width;
- const int dst_h = dst->y_crop_height;
const uint8_t *const srcs[3] = { src->y_buffer, src->u_buffer,src->v_buffer };
const int src_strides[3] = { src->y_stride, src->uv_stride, src->uv_stride };@@ -30,22 +28,86 @@
const InterpKernel *const kernel = vp9_filter_kernels[filter_type];
int x, y, i;
- for (i = 0; i < MAX_MB_PLANE; ++i) {- const int factor = (i == 0 || i == 3 ? 1 : 2);
- const int src_stride = src_strides[i];
- const int dst_stride = dst_strides[i];
- for (y = 0; y < dst_h; y += 16) {- const int y_q4 = y * (16 / factor) * src_h / dst_h + phase_scaler;
- for (x = 0; x < dst_w; x += 16) {- const int x_q4 = x * (16 / factor) * src_w / dst_w + phase_scaler;
- const uint8_t *src_ptr = srcs[i] +
- (y / factor) * src_h / dst_h * src_stride +
- (x / factor) * src_w / dst_w;
- uint8_t *dst_ptr = dsts[i] + (y / factor) * dst_stride + (x / factor);
-
- vpx_scaled_2d(src_ptr, src_stride, dst_ptr, dst_stride, kernel,
- x_q4 & 0xf, 16 * src_w / dst_w, y_q4 & 0xf,
- 16 * src_h / dst_h, 16 / factor, 16 / factor);
+#if 0
+ // TODO(linfengz): The 4:3 specialized C code is disabled by default since
+ // it's much slower than the general version which calls vpx_scaled_2d() even
+ // if vpx_scaled_2d() is not optimized. It will only be enabled as a reference
+ // for the platforms which have faster optimization.
+ if (4 * dst->y_crop_width == 3 * src_w &&
+ 4 * dst->y_crop_height == 3 * src_h) {+ // Specialize 4 to 3 scaling.
+ // Example pixel locations.
+ // (O: Original pixel. S: Scaled pixel. X: Overlapped pixel.)
+ // phase_scaler = 0 | phase_scaler = 8
+ // |
+ // X O S O S O X | O O O O O
+ // |
+ // |
+ // | S S S
+ // |
+ // |
+ // O O O O O | O O O O O
+ // |
+ // S S S S |
+ // |
+ // |
+ // | S S S
+ // O O O O O | O O O O O
+ // |
+ // |
+ // |
+ // S S S S |
+ // |
+ // O O O O O | O O O O O
+ // | S S S
+ // |
+ // |
+ // |
+ // |
+ // X O S O S O X | O O O O O
+
+ const int dst_ws[3] = { dst->y_crop_width, dst->uv_crop_width,+ dst->uv_crop_width };
+ const int dst_hs[3] = { dst->y_crop_height, dst->uv_crop_height,+ dst->uv_crop_height };
+ for (i = 0; i < MAX_MB_PLANE; ++i) {+ const int dst_w = dst_ws[i];
+ const int dst_h = dst_hs[i];
+ const int src_stride = src_strides[i];
+ const int dst_stride = dst_strides[i];
+ for (y = 0; y < dst_h; y += 3) {+ for (x = 0; x < dst_w; x += 3) {+ const uint8_t *src_ptr = srcs[i] + 4 * y / 3 * src_stride + 4 * x / 3;
+ uint8_t *dst_ptr = dsts[i] + y * dst_stride + x;
+
+ // Must call c function because its optimization doesn't support 3x3.
+ vpx_scaled_2d_c(src_ptr, src_stride, dst_ptr, dst_stride, kernel,
+ phase_scaler, 64 / 3, phase_scaler, 64 / 3, 3, 3);
+ }
+ }
+ }
+ } else
+#endif
+ {+ const int dst_w = dst->y_crop_width;
+ const int dst_h = dst->y_crop_height;
+ for (i = 0; i < MAX_MB_PLANE; ++i) {+ const int factor = (i == 0 || i == 3 ? 1 : 2);
+ const int src_stride = src_strides[i];
+ const int dst_stride = dst_strides[i];
+ for (y = 0; y < dst_h; y += 16) {+ const int y_q4 = y * (16 / factor) * src_h / dst_h + phase_scaler;
+ for (x = 0; x < dst_w; x += 16) {+ const int x_q4 = x * (16 / factor) * src_w / dst_w + phase_scaler;
+ const uint8_t *src_ptr = srcs[i] +
+ (y / factor) * src_h / dst_h * src_stride +
+ (x / factor) * src_w / dst_w;
+ uint8_t *dst_ptr = dsts[i] + (y / factor) * dst_stride + (x / factor);
+
+ vpx_scaled_2d(src_ptr, src_stride, dst_ptr, dst_stride, kernel,
+ x_q4 & 0xf, 16 * src_w / dst_w, y_q4 & 0xf,
+ 16 * src_h / dst_h, 16 / factor, 16 / factor);
+ }
}
}
}