/* * 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>#include "config/aom_dsp_rtcd.h" #include "aom/aom_integer.h" #include "aom_dsp/x86/bitdepth_conversion_sse2.h" #include "aom_dsp/x86/mem_sse2.h" #include "aom_dsp/x86/synonyms.h" #include "aom_ports/mem.h" static inline void sign_extend_16bit_to_32bit_sse2(__m128i in, __m128i zero,const __m128i sign_bits = _mm_cmplt_epi16(in, zero);static inline __m128i invert_sign_32_sse2(__m128i a, __m128i sign) {return _mm_sub_epi32(a, sign);void aom_minmax_8x8_sse2(const uint8_t *s, int p, const uint8_t *d, int dp,int *min, int *max) {// Row 0 const __m128i *)(s)), u0);const __m128i *)(d)), u0);// Row 1 const __m128i *)(s + p)), u0);const __m128i *)(d + dp)), u0);// Row 2 const __m128i *)(s + 2 * p)), u0);const __m128i *)(d + 2 * dp)), u0);// Row 3 const __m128i *)(s + 3 * p)), u0);const __m128i *)(d + 3 * dp)), u0);// Row 4 const __m128i *)(s + 4 * p)), u0);const __m128i *)(d + 4 * dp)), u0);// Row 5 const __m128i *)(s + 5 * p)), u0);const __m128i *)(d + 5 * dp)), u0);// Row 6 const __m128i *)(s + 6 * p)), u0);const __m128i *)(d + 6 * dp)), u0);// Row 7 const __m128i *)(s + 7 * p)), u0);const __m128i *)(d + 7 * dp)), u0);unsigned int aom_avg_8x8_sse2(const uint8_t *s, int p) {unsigned int avg = 0;const __m128i *)(s + p),const __m128i *)(s)));const __m128i *)(s + 3 * p),const __m128i *)(s + 2 * p)));const __m128i *)(s + 5 * p),const __m128i *)(s + 4 * p)));const __m128i *)(s + 7 * p),const __m128i *)(s + 6 * p)));return (avg + 32) >> 6;static void calc_avg_8x8_dual_sse2(const uint8_t *s, int p, int *avg) {const __m128i *)(s)), u0);const __m128i *)(s + p)), u0);const __m128i *)(s + 2 * p)), u0);const __m128i *)(s + 3 * p)), u0);const __m128i *)(s + 4 * p)), u0);const __m128i *)(s + 5 * p)), u0);const __m128i *)(s + 6 * p)), u0);const __m128i *)(s + 7 * p)), u0);// (avg + 32) >> 6 void aom_avg_8x8_quad_sse2(const uint8_t *s, int p, int x16_idx, int y16_idx,int *avg) {const uint8_t *s_ptr = s + y16_idx * p + x16_idx;for (int k = 0; k < 2; k++) {unsigned int aom_avg_4x4_sse2(const uint8_t *s, int p) {unsigned int avg = 0;return (avg + 8) >> 4;static inline void hadamard_col4_sse2(__m128i *in, int iter) {const __m128i a0 = in[0];const __m128i a1 = in[1];const __m128i a2 = in[2];const __m128i a3 = in[3];const __m128i b0 = _mm_srai_epi16(_mm_add_epi16(a0, a1), 1);const __m128i b1 = _mm_srai_epi16(_mm_sub_epi16(a0, a1), 1);const __m128i b2 = _mm_srai_epi16(_mm_add_epi16(a2, a3), 1);const __m128i b3 = _mm_srai_epi16(_mm_sub_epi16(a2, a3), 1);if (iter == 0) {const __m128i ba = _mm_unpacklo_epi16(in[0], in[1]);const __m128i dc = _mm_unpacklo_epi16(in[2], in[3]);const __m128i dcba_lo = _mm_unpacklo_epi32(ba, dc);const __m128i dcba_hi = _mm_unpackhi_epi32(ba, dc);void aom_hadamard_4x4_sse2(const int16_t *src_diff, ptrdiff_t src_stride,const __m128i *)src_diff);const __m128i *)(src_diff += src_stride));const __m128i *)(src_diff += src_stride));const __m128i *)(src_diff + src_stride));static inline void hadamard_col8_sse2(__m128i *in, int iter) {if (iter == 0) {else {static inline void hadamard_8x8_sse2(const int16_t *src_diff,int is_final) {const __m128i *)src_diff);const __m128i *)(src_diff += src_stride));const __m128i *)(src_diff += src_stride));const __m128i *)(src_diff += src_stride));const __m128i *)(src_diff += src_stride));const __m128i *)(src_diff += src_stride));const __m128i *)(src_diff += src_stride));const __m128i *)(src_diff + src_stride));if (is_final) {else {void aom_hadamard_8x8_sse2(const int16_t *src_diff, ptrdiff_t src_stride,static inline void hadamard_lp_8x8_sse2(const int16_t *src_diff,const __m128i *)src_diff);const __m128i *)(src_diff += src_stride));const __m128i *)(src_diff += src_stride));const __m128i *)(src_diff += src_stride));const __m128i *)(src_diff += src_stride));const __m128i *)(src_diff += src_stride));const __m128i *)(src_diff += src_stride));const __m128i *)(src_diff + src_stride));void aom_hadamard_lp_8x8_sse2(const int16_t *src_diff, ptrdiff_t src_stride,void aom_hadamard_lp_8x8_dual_sse2(const int16_t *src_diff,for (int i = 0; i < 2; i++) {void aom_hadamard_lp_16x16_sse2(const int16_t *src_diff, ptrdiff_t src_stride,for (int idx = 0; idx < 4; ++idx) {const int16_t *src_ptr =for (int idx = 0; idx < 64; idx += 8) {const __m128i *)t_coeff);const __m128i *)(t_coeff + 64));const __m128i *)(t_coeff + 128));const __m128i *)(t_coeff + 192));static inline void hadamard_16x16_sse2(const int16_t *src_diff,int is_final) {// For high bitdepths, it is unnecessary to store_tran_low // (mult/unpack/store), then load_tran_low (load/pack) the same memory in the // next stage. Output to an intermediate buffer first, then store_tran_low() // in the final stage. int idx;for (idx = 0; idx < 4; ++idx) {const int16_t *src_ptr =for (idx = 0; idx < 64; idx += 8) {const __m128i *)t_coeff);const __m128i *)(t_coeff + 64));const __m128i *)(t_coeff + 128));const __m128i *)(t_coeff + 192));if (is_final) {else {// Increment the pointer additionally by 0 and 8 in alternate // iterations(instead of 8) to ensure the coherency with the implementation // of store_tran_low_offset_4() void aom_hadamard_16x16_sse2(const int16_t *src_diff, ptrdiff_t src_stride,void aom_hadamard_32x32_sse2(const int16_t *src_diff, ptrdiff_t src_stride,// For high bitdepths, it is unnecessary to store_tran_low // (mult/unpack/store), then load_tran_low (load/pack) the same memory in the // next stage. Output to an intermediate buffer first, then store_tran_low() // in the final stage. int idx;const __m128i zero = _mm_setzero_si128();for (idx = 0; idx < 4; ++idx) {const int16_t *src_ptr =for (idx = 0; idx < 256; idx += 8) {const __m128i *)t_coeff);const __m128i *)(t_coeff + 256));const __m128i *)(t_coeff + 512));const __m128i *)(t_coeff + 768));// Sign extend 16 bit to 32 bit. // Increment the pointer by 4 and 12 in alternate iterations(instead of 8) // to ensure the coherency with the implementation of // store_tran_low_offset_4() int aom_satd_sse2(const tran_low_t *coeff, int length) {int i;const __m128i zero = _mm_setzero_si128();for (i = 0; i < length; i += 4) {const __m128i src_line = _mm_load_si128((const __m128i *)coeff);const __m128i coeff_sign = _mm_srai_epi32(src_line, 31);const __m128i abs_coeff = invert_sign_32_sse2(src_line, coeff_sign);// cascading summation of accum return _mm_cvtsi128_si32(accum);int aom_satd_lp_sse2(const int16_t *coeff, int length) {const __m128i zero = _mm_setzero_si128();const __m128i one = _mm_set1_epi16(1);for (int i = 0; i < length; i += 16) {const __m128i src_line0 = _mm_loadu_si128((const __m128i *)coeff);const __m128i src_line1 = _mm_loadu_si128((const __m128i *)(coeff + 8));const __m128i inv0 = _mm_sub_epi16(zero, src_line0);const __m128i inv1 = _mm_sub_epi16(zero, src_line1);const __m128i abs0 = _mm_max_epi16(src_line0, inv0); // abs(src_line) const __m128i abs1 = _mm_max_epi16(src_line1, inv1); // abs(src_line) const __m128i sum0 = _mm_madd_epi16(abs0, one);const __m128i sum1 = _mm_madd_epi16(abs1, one);// cascading summation of accum return _mm_cvtsi128_si32(accum);void aom_int_pro_row_sse2(int16_t *hbuf, const uint8_t *ref,const int ref_stride, const int width,const int height, int norm_factor) {// SIMD implementation assumes width and height to be multiple of 16 and 2 // respectively. For any odd width or height, SIMD support needs to be added. for (int wd = 0; wd < width; wd += 16) {const uint8_t *ref_tmp = ref + wd;int idx = 0;do {const __m128i *)ref_tmp);const __m128i *)ref_tmp);while (idx < height);void aom_int_pro_col_sse2(int16_t *vbuf, const uint8_t *ref,const int ref_stride, const int width,const int height, int norm_factor) {// SIMD implementation assumes width to be multiple of 16. for (int ht = 0; ht < height; ht++) {const uint8_t *ref_tmp = ref + (ht * ref_stride);for (int i = 0; i < width; i += 16) {const __m128i *)ref_tmp);
Messung V0.5 C=90 H=95 G=92
¤ Dauer der Verarbeitung: 0.15 Sekunden
(vorverarbeitet)
¤ *© Formatika GbR, Deutschland
