/* * Copyright (c) 2016, Alliance for Open Media. All rights reserved. * * This source code is subject to the terms of the BSD 2 Clause License and * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License * was not distributed with this source code in the LICENSE file, you can * obtain it at www.aomedia.org/license/software. If the Alliance for Open * Media Patent License 1.0 was not distributed with this source code in the * PATENTS file, you can obtain it at www.aomedia.org/license/patent. #include <immintrin.h> // AVX2 #include "config/aom_dsp_rtcd.h" #include "aom_ports/mem.h" /* clang-format off */ static const uint8_t, bilinear_filters_avx2[512]) = {/* clang-format on */ #define FILTER_SRC(filter) \/* filter the source */ \ /* add 8 to source */ \ /* divide source by 16 */ \ #define MERGE_WITH_SRC(src_reg, reg) \#define LOAD_SRC_DST \/* load source and destination */ \ const *)(src)); \const *)(dst));#define AVG_NEXT_SRC(src_reg, size_stride) \const *)(src + size_stride)); \/* average between current and next stride source */ \ #define MERGE_NEXT_SRC(src_reg, size_stride) \const *)(src + size_stride)); \#define CALC_SUM_SSE_INSIDE_LOOP \/* expand each byte to 2 bytes */ \ /* source - dest */ \ /* caculate sum */ \ /* calculate sse */ \ // final calculation to sum and sse #define CALC_SUM_AND_SSE \int *)sse) = _mm_cvtsi128_si32(_mm256_castsi256_si128(sse_reg)) + \// Functions related to sub pixel variance width 16 #define LOAD_SRC_DST_INSERT(src_stride, dst_stride) \/* load source and destination of 2 rows and insert*/ \ #define AVG_NEXT_SRC_INSERT(src_reg, size_stride) \/* average between current and next stride source */ \ #define MERGE_NEXT_SRC_INSERT(src_reg, size_stride) \#define LOAD_SRC_NEXT_BYTE_INSERT \/* load source and another source from next row */ \ /* load source and next row source from 1 byte onwards */ \ #define LOAD_DST_INSERT \#define LOAD_SRC_MERGE_128BIT(filter) \#define FILTER_SRC_128BIT(filter) \/* filter the source */ \ /* add 8 to source */ \ /* divide source by 16 */ \ // TODO(chiyotsai@google.com): These variance functions are macro-fied so we // don't have to manually optimize the individual for-loops. We could save some // binary size by optimizing the loops more carefully without duplicating the // codes with a macro. #define MAKE_SUB_PIXEL_VAR_32XH(height, log2height) \static inline int aom_sub_pixel_variance32x## height## _imp_avx2( \const uint8_t *src, int src_stride, int x_offset, int y_offset, \const uint8_t *dst, int dst_stride, unsigned int *sse) { \int i, sum; \/* x_offset = 0 and y_offset = 0 */ \ if (x_offset == 0) { \if (y_offset == 0) { \for (i = 0; i < height; i++) { \/* expend each byte to 2 bytes */ \ /* x_offset = 0 and y_offset = 4 */ \ else if (y_offset == 4) { \for (i = 0; i < height; i++) { \/* expend each byte to 2 bytes */ \ /* x_offset = 0 and y_offset = bilin interpolation */ \ else { \const *)(bilinear_filters_avx2 + y_offset)); \for (i = 0; i < height; i++) { \/* x_offset = 4 and y_offset = 0 */ \ else if (x_offset == 4) { \if (y_offset == 0) { \for (i = 0; i < height; i++) { \/* expand each byte to 2 bytes */ \ /* x_offset = 4 and y_offset = 4 */ \ else if (y_offset == 4) { \/* load source and another source starting from the next */ \ /* following byte */ \ const *)(src)); \for (i = 0; i < height; i++) { \/* average between previous average to current average */ \ /* expand each byte to 2 bytes */ \ /* save current source average */ \ /* x_offset = 4 and y_offset = bilin interpolation */ \ else { \const *)(bilinear_filters_avx2 + y_offset)); \/* load source and another source starting from the next */ \ /* following byte */ \ const *)(src)); \for (i = 0; i < height; i++) { \/* save current source average */ \ /* x_offset = bilin interpolation and y_offset = 0 */ \ else { \if (y_offset == 0) { \const *)(bilinear_filters_avx2 + x_offset)); \for (i = 0; i < height; i++) { \/* x_offset = bilin interpolation and y_offset = 4 */ \ else if (y_offset == 4) { \const *)(bilinear_filters_avx2 + x_offset)); \const *)(src)); \/* convert each 16 bit to 8 bit to each low and high lane source */ \ for (i = 0; i < height; i++) { \/* average between previous pack to the current */ \ /* x_offset = bilin interpolation and y_offset = bilin interpolation \ else { \const *)(bilinear_filters_avx2 + x_offset)); \const *)(bilinear_filters_avx2 + y_offset)); \/* load source and another source starting from the next */ \ /* following byte */ \ const *)(src)); \/* convert each 16 bit to 8 bit to each low and high lane source */ \ for (i = 0; i < height; i++) { \/* merge previous pack to current pack source */ \ /* filter the source */ \ return sum; \unsigned int aom_sub_pixel_variance32x## height## _avx2( \const uint8_t *src, int src_stride, int x_offset, int y_offset, \const uint8_t *dst, int dst_stride, unsigned int *sse) { \const int sum = aom_sub_pixel_variance32x## height## _imp_avx2( \return *sse - (unsigned int )(((int64_t)sum * sum) >> (5 + log2height)); \#define AOM_SUB_PIXEL_VAR_AVX2(w, h, wf, hf, wlog2, hlog2) \unsigned int aom_sub_pixel_variance## w## x## h## _avx2( \const uint8_t *src, int src_stride, int x_offset, int y_offset, \const uint8_t *dst, int dst_stride, unsigned int *sse_ptr) { \unsigned int sse = 0; \int se = 0; \for (int i = 0; i < (w / wf); ++i) { \const uint8_t *src_ptr = src; \const uint8_t *dst_ptr = dst; \for (int j = 0; j < (h / hf); ++j) { \unsigned int sse2; \const int se2 = aom_sub_pixel_variance## wf## x## hf## _imp_avx2( \return sse - (unsigned int )(((int64_t)se * se) >> (wlog2 + hlog2)); \// Note: hf = AOMMIN(h, 64) to avoid overflow in helper by capping height. #define MAKE_SUB_PIXEL_VAR_16XH(height, log2height) \unsigned int aom_sub_pixel_variance16x## height## _avx2( \const uint8_t *src, int src_stride, int x_offset, int y_offset, \const uint8_t *dst, int dst_stride, unsigned int *sse) { \int i, sum; \/* x_offset = 0 and y_offset = 0 */ \ if (x_offset == 0) { \if (y_offset == 0) { \for (i = 0; i < height; i += 2) { \/* expend each byte to 2 bytes */ \ /* x_offset = 0 and y_offset = 4 */ \ else if (y_offset == 4) { \for (i = 0; i < height; i += 2) { \/* expend each byte to 2 bytes */ \ /* x_offset = 0 and y_offset = bilin interpolation */ \ else { \const *)(bilinear_filters_avx2 + y_offset)); \for (i = 0; i < height; i += 2) { \/* x_offset = 4 and y_offset = 0 */ \ else if (x_offset == 4) { \if (y_offset == 0) { \for (i = 0; i < height; i += 2) { \/* average between current and next stride source */ \ /* expand each byte to 2 bytes */ \ /* x_offset = 4 and y_offset = 4 */ \ else if (y_offset == 4) { \/* load and insert source and next row source */ \ for (i = 0; i < height - 2; i += 2) { \/* expand each byte to 2 bytes */ \ /* save current source average */ \ /* last 2 rows processing happens here */ \ else { \/* x_offset = 4 and y_offset = bilin interpolation */ \ const *)(bilinear_filters_avx2 + y_offset)); \/* load and insert source and next row source */ \ for (i = 0; i < height - 2; i += 2) { \/* save current source average */ \ /* last 2 rows processing happens here */ \ /* x_offset = bilin interpolation and y_offset = 0 */ \ else { \if (y_offset == 0) { \const *)(bilinear_filters_avx2 + x_offset)); \for (i = 0; i < height; i += 2) { \/* x_offset = bilin interpolation and y_offset = 4 */ \ else if (y_offset == 4) { \const *)(bilinear_filters_avx2 + x_offset)); \/* load and insert source and next row source */ \ /* convert each 16 bit to 8 bit to each low and high lane source */ \ for (i = 0; i < height - 2; i += 2) { \/* average between previous pack to the current */ \ /* last 2 rows processing happens here */ \ /* average between previous pack to the current */ \ else { \/* x_offset = bilin interpolation and y_offset = bilin interpolation \ const *)(bilinear_filters_avx2 + x_offset)); \const *)(bilinear_filters_avx2 + y_offset)); \/* load and insert source and next row source */ \ /* convert each 16 bit to 8 bit to each low and high lane source */ \ for (i = 0; i < height - 2; i += 2) { \/* average between previous pack to the current */ \ /* filter the source */ \ /* last 2 rows processing happens here */ \ return *sse - (unsigned int )(((int64_t)sum * sum) >> (4 + log2height)); \#if !CONFIG_REALTIME_ONLY#endif #define MAKE_SUB_PIXEL_AVG_VAR_32XH(height, log2height) \static int sub_pixel_avg_variance32x## height## _imp_avx2( \const uint8_t *src, int src_stride, int x_offset, int y_offset, \const uint8_t *dst, int dst_stride, const uint8_t *sec, int sec_stride, \unsigned int *sse) { \int i, sum; \/* x_offset = 0 and y_offset = 0 */ \ if (x_offset == 0) { \if (y_offset == 0) { \for (i = 0; i < height; i++) { \const *)(sec)); \/* expend each byte to 2 bytes */ \ else if (y_offset == 4) { \for (i = 0; i < height; i++) { \const *)(sec)); \/* expend each byte to 2 bytes */ \ /* x_offset = 0 and y_offset = bilin interpolation */ \ else { \const *)(bilinear_filters_avx2 + y_offset)); \for (i = 0; i < height; i++) { \const *)(sec)); \/* x_offset = 4 and y_offset = 0 */ \ else if (x_offset == 4) { \if (y_offset == 0) { \for (i = 0; i < height; i++) { \const *)(sec)); \/* expand each byte to 2 bytes */ \ /* x_offset = 4 and y_offset = 4 */ \ else if (y_offset == 4) { \/* load source and another source starting from the next */ \ /* following byte */ \ const *)(src)); \for (i = 0; i < height; i++) { \/* save current source average */ \ /* average between previous average to current average */ \ const *)(sec)); \/* expand each byte to 2 bytes */ \ /* x_offset = 4 and y_offset = bilin interpolation */ \ else { \const *)(bilinear_filters_avx2 + y_offset)); \/* load source and another source starting from the next */ \ /* following byte */ \ const *)(src)); \for (i = 0; i < height; i++) { \/* save current source average */ \ const *)(sec)); \/* expand each byte to 2 bytes */ \ /* x_offset = bilin interpolation and y_offset = 0 */ \ else { \if (y_offset == 0) { \const *)(bilinear_filters_avx2 + x_offset)); \for (i = 0; i < height; i++) { \const *)(sec)); \/* x_offset = bilin interpolation and y_offset = 4 */ \ else if (y_offset == 4) { \const *)(bilinear_filters_avx2 + x_offset)); \const *)(src)); \/* convert each 16 bit to 8 bit to each low and high lane source */ \ for (i = 0; i < height; i++) { \/* average between previous pack to the current */ \ const *)(sec)); \/* x_offset = bilin interpolation and y_offset = bilin interpolation \ else { \const *)(bilinear_filters_avx2 + x_offset)); \const *)(bilinear_filters_avx2 + y_offset)); \/* load source and another source starting from the next */ \ /* following byte */ \ const *)(src)); \/* convert each 16 bit to 8 bit to each low and high lane source */ \ for (i = 0; i < height; i++) { \/* merge previous pack to current pack source */ \ /* filter the source */ \ const *)(sec)); \return sum; \unsigned int aom_sub_pixel_avg_variance32x## height## _avx2( \const uint8_t *src, int src_stride, int x_offset, int y_offset, \const uint8_t *dst, int dst_stride, unsigned int *sse, \const uint8_t *sec_ptr) { \const int sum = sub_pixel_avg_variance32x## height## _imp_avx2( \return *sse - (unsigned int )(((int64_t)sum * sum) >> (5 + log2height)); \#define AOM_SUB_PIXEL_AVG_VAR_AVX2(w, h, wf, hf, wlog2, hlog2) \unsigned int aom_sub_pixel_avg_variance## w## x## h## _avx2( \const uint8_t *src, int src_stride, int x_offset, int y_offset, \const uint8_t *dst, int dst_stride, unsigned int *sse_ptr, \const uint8_t *sec) { \unsigned int sse = 0; \int se = 0; \
Messung V0.5 C=95 H=91 G=92
¤ Dauer der Verarbeitung: 0.23 Sekunden
(vorverarbeitet)
¤ *© Formatika GbR, Deutschland
