[M102-LTS] Use non normative scaler for non optimized ratio
M102 merge issues:
Had to change the author (jianj@google.com) to be able to upload the
CL because of the following error:
remote: ERROR: commit 83544bf: email address jianj@google.com is not registered
in your account, and you lack 'forge author' permission.
There are only optimized scalers for 1/4, 1/2 and 3/4 scaling ratio.
SSSE3 also has 2x upscaling optimization.
Use non normative scalers for all other scaling ratios.
Bug: chromium:1346938
Bug: chromium:1338114
Change-Id: I2a01717b56c53c42906440d5a3f95ca2c00dc571
(cherry picked from commit ff7b753a63a536423a91b64a066bd385c52ceacc)
diff --git a/av1/common/resize.c b/av1/common/resize.c
index a3c3c0e..322363fa 100644
--- a/av1/common/resize.c
+++ b/av1/common/resize.c
@@ -1366,15 +1366,20 @@
aom_internal_error(cm->error, AOM_CODEC_MEM_ERROR,
"Failed to allocate scaled buffer");
+ const bool has_optimized_scaler = av1_has_optimized_scaler(
+ unscaled->y_crop_width, unscaled->y_crop_height, scaled_width,
+ scaled_height);
+
#if CONFIG_AV1_HIGHBITDEPTH
- if (use_optimized_scaler && cm->seq_params->bit_depth == AOM_BITS_8) {
+ if (use_optimized_scaler && has_optimized_scaler &&
+ cm->seq_params->bit_depth == AOM_BITS_8) {
av1_resize_and_extend_frame(unscaled, scaled, filter, phase, num_planes);
} else {
av1_resize_and_extend_frame_nonnormative(
unscaled, scaled, (int)cm->seq_params->bit_depth, num_planes);
}
#else
- if (use_optimized_scaler) {
+ if (use_optimized_scaler && has_optimized_scaler) {
av1_resize_and_extend_frame(unscaled, scaled, filter, phase, num_planes);
} else {
av1_resize_and_extend_frame_nonnormative(
diff --git a/av1/common/resize.h b/av1/common/resize.h
index 75abe62..9bc23b3 100644
--- a/av1/common/resize.h
+++ b/av1/common/resize.h
@@ -105,6 +105,24 @@
return !(cm->width == cm->superres_upscaled_width);
}
+// There's SIMD optimizations for 1/4, 1/2 and 3/4 downscaling.
+// SSSE3 also has optimizations for 2x upscaling.
+// Use non normative scalers for other scaling ratios.
+static INLINE bool av1_has_optimized_scaler(const int src_width,
+ const int src_height,
+ const int dst_width,
+ const int dst_height) {
+ const bool has_optimized_scaler =
+ (dst_width * 4 == src_width && dst_height * 4 == src_height) ||
+ (dst_width * 2 == src_width && dst_height * 2 == src_height) ||
+ (dst_width * 4 == src_width * 3 && dst_height * 4 == src_height * 3);
+#if HAVE_SSSE3
+ return has_optimized_scaler ||
+ (dst_width == src_width * 2 && dst_height == src_height * 2);
+#endif
+ return has_optimized_scaler;
+}
+
#define UPSCALE_NORMATIVE_TAPS 8
extern const int16_t av1_resize_filter_normative[1 << RS_SUBPEL_BITS]
[UPSCALE_NORMATIVE_TAPS];
diff --git a/av1/encoder/encoder_utils.c b/av1/encoder/encoder_utils.c
index fd8be7b..cebea60 100644
--- a/av1/encoder/encoder_utils.c
+++ b/av1/encoder/encoder_utils.c
@@ -733,15 +733,19 @@
aom_internal_error(cm->error, AOM_CODEC_MEM_ERROR,
"Failed to allocate frame buffer");
}
+ const bool has_optimized_scaler = av1_has_optimized_scaler(
+ cm->width, cm->height, new_fb->buf.y_crop_width,
+ new_fb->buf.y_crop_height);
#if CONFIG_AV1_HIGHBITDEPTH
- if (use_optimized_scaler && cm->seq_params->bit_depth == AOM_BITS_8)
+ if (use_optimized_scaler && has_optimized_scaler &&
+ cm->seq_params->bit_depth == AOM_BITS_8)
av1_resize_and_extend_frame(ref, &new_fb->buf, filter, phase,
num_planes);
else
av1_resize_and_extend_frame_nonnormative(
ref, &new_fb->buf, (int)cm->seq_params->bit_depth, num_planes);
#else
- if (use_optimized_scaler)
+ if (use_optimized_scaler && has_optimized_scaler)
av1_resize_and_extend_frame(ref, &new_fb->buf, filter, phase,
num_planes);
else
