• file
• revisions
• annotate
• diff
• comparison
• raw

media/libvpx/vp9/common/arm/neon/vp9_convolve_neon.c@b8a032363ba2 (annotated)

media/libvpx/vp9/common/arm/neon/vp9_convolve_neon.c

Thu, 22 Jan 2015 13:21:57 +0100

author

Michael Schloh von Bennewitz <michael@schloh.com>

date

Thu, 22 Jan 2015 13:21:57 +0100

branch

TOR_BUG_9701

changeset 15

b8a032363ba2

permissions

-rw-r--r--

Incorporate requested changes from Mozilla in review:
https://bugzilla.mozilla.org/show_bug.cgi?id=1123480#c6

michael@0	1	/*
michael@0	2	* Copyright (c) 2013 The WebM project authors. All Rights Reserved.
michael@0	3	*
michael@0	4	* Use of this source code is governed by a BSD-style license
michael@0	5	* that can be found in the LICENSE file in the root of the source
michael@0	6	* tree. An additional intellectual property rights grant can be found
michael@0	7	* in the file PATENTS. All contributing project authors may
michael@0	8	* be found in the AUTHORS file in the root of the source tree.
michael@0	9	*/
michael@0	10
michael@0	11	#include "./vp9_rtcd.h"
michael@0	12	#include "vp9/common/vp9_common.h"
michael@0	13	#include "vpx_ports/mem.h"
michael@0	14
michael@0	15	void vp9_convolve8_neon(const uint8_t *src, ptrdiff_t src_stride,
michael@0	16	uint8_t *dst, ptrdiff_t dst_stride,
michael@0	17	const int16_t *filter_x, int x_step_q4,
michael@0	18	const int16_t *filter_y, int y_step_q4,
michael@0	19	int w, int h) {
michael@0	20	/* Given our constraints: w <= 64, h <= 64, taps == 8 we can reduce the
michael@0	21	* maximum buffer size to 64 * 64 + 7 (+ 1 to make it divisible by 4).
michael@0	22	*/
michael@0	23	DECLARE_ALIGNED_ARRAY(8, uint8_t, temp, 64 * 72);
michael@0	24
michael@0	25	// Account for the vertical phase needing 3 lines prior and 4 lines post
michael@0	26	int intermediate_height = h + 7;
michael@0	27
michael@0	28	if (x_step_q4 != 16 || y_step_q4 != 16)
michael@0	29	return vp9_convolve8_c(src, src_stride,
michael@0	30	dst, dst_stride,
michael@0	31	filter_x, x_step_q4,
michael@0	32	filter_y, y_step_q4,
michael@0	33	w, h);
michael@0	34
michael@0	35	/* Filter starting 3 lines back. The neon implementation will ignore the
michael@0	36	* given height and filter a multiple of 4 lines. Since this goes in to
michael@0	37	* the temp buffer which has lots of extra room and is subsequently discarded
michael@0	38	* this is safe if somewhat less than ideal.
michael@0	39	*/
michael@0	40	vp9_convolve8_horiz_neon(src - src_stride * 3, src_stride,
michael@0	41	temp, 64,
michael@0	42	filter_x, x_step_q4, filter_y, y_step_q4,
michael@0	43	w, intermediate_height);
michael@0	44
michael@0	45	/* Step into the temp buffer 3 lines to get the actual frame data */
michael@0	46	vp9_convolve8_vert_neon(temp + 64 * 3, 64,
michael@0	47	dst, dst_stride,
michael@0	48	filter_x, x_step_q4, filter_y, y_step_q4,
michael@0	49	w, h);
michael@0	50	}
michael@0	51
michael@0	52	void vp9_convolve8_avg_neon(const uint8_t *src, ptrdiff_t src_stride,
michael@0	53	uint8_t *dst, ptrdiff_t dst_stride,
michael@0	54	const int16_t *filter_x, int x_step_q4,
michael@0	55	const int16_t *filter_y, int y_step_q4,
michael@0	56	int w, int h) {
michael@0	57	DECLARE_ALIGNED_ARRAY(8, uint8_t, temp, 64 * 72);
michael@0	58	int intermediate_height = h + 7;
michael@0	59
michael@0	60	if (x_step_q4 != 16 || y_step_q4 != 16)
michael@0	61	return vp9_convolve8_avg_c(src, src_stride,
michael@0	62	dst, dst_stride,
michael@0	63	filter_x, x_step_q4,
michael@0	64	filter_y, y_step_q4,
michael@0	65	w, h);
michael@0	66
michael@0	67	/* This implementation has the same issues as above. In addition, we only want
michael@0	68	* to average the values after both passes.
michael@0	69	*/
michael@0	70	vp9_convolve8_horiz_neon(src - src_stride * 3, src_stride,
michael@0	71	temp, 64,
michael@0	72	filter_x, x_step_q4, filter_y, y_step_q4,
michael@0	73	w, intermediate_height);
michael@0	74	vp9_convolve8_avg_vert_neon(temp + 64 * 3,
michael@0	75	64, dst, dst_stride,
michael@0	76	filter_x, x_step_q4, filter_y, y_step_q4,
michael@0	77	w, h);
michael@0	78	}