From 42cb3dbf94706dd8477f9313ceba1fc3a9a14e92 Mon Sep 17 00:00:00 2001 From: Salome Thirot Date: Tue, 7 Feb 2023 14:08:33 +0000 Subject: [PATCH] Optimize Neon high bitdepth subpel variance functions Use the same general code style as in the standard bitdepth Neon implementation. Additionally, do not unnecessarily widen to 32-bit data types when doing bilinear filtering - allowing us to process twice as many elements per instruction. Change-Id: I1e178991d2aa71f5f77a376e145d19257481e90f --- vpx_dsp/arm/highbd_subpel_variance_neon.c | 307 ++++++++++++++++++++++++++++++ vpx_dsp/arm/highbd_variance_neon.c | 288 ++-------------------------- vpx_dsp/arm/mem_neon.h | 14 ++ vpx_dsp/vpx_dsp.mk | 1 + 4 files changed, 336 insertions(+), 274 deletions(-) create mode 100644 vpx_dsp/arm/highbd_subpel_variance_neon.c diff --git a/vpx_dsp/arm/highbd_subpel_variance_neon.c b/vpx_dsp/arm/highbd_subpel_variance_neon.c new file mode 100644 index 0000000..81943ee --- /dev/null +++ b/vpx_dsp/arm/highbd_subpel_variance_neon.c @@ -0,0 +1,307 @@ +/* + * Copyright (c) 2023 The WebM project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include +#include + +#include "./vpx_dsp_rtcd.h" +#include "./vpx_config.h" + +#include "vpx/vpx_integer.h" +#include "vpx_dsp/arm/mem_neon.h" + +// The bilinear filters look like this: +// +// {{ 128, 0 }, { 112, 16 }, { 96, 32 }, { 80, 48 }, +// { 64, 64 }, { 48, 80 }, { 32, 96 }, { 16, 112 }} +// +// We can factor out the highest common multiple, such that the sum of both +// weights will be 8 instead of 128. The benefits of this are two-fold: +// +// 1) We can infer the filter values from the filter_offset parameter in the +// bilinear filter functions below - we don't have to actually load the values +// from memory: +// f0 = 8 - filter_offset +// f1 = filter_offset +// +// 2) Scaling the pixel values by 8, instead of 128 enables us to operate on +// 16-bit data types at all times, rather than widening out to 32-bit and +// requiring double the number of data processing instructions. (12-bit * 8 = +// 15-bit.) + +// Process a block exactly 4 wide and a multiple of 2 high. +static void highbd_var_filter_block2d_bil_w4(const uint16_t *src_ptr, + uint16_t *dst_ptr, int src_stride, + int pixel_step, int dst_height, + int filter_offset) { + const uint16x8_t f0 = vdupq_n_u16(8 - filter_offset); + const uint16x8_t f1 = vdupq_n_u16(filter_offset); + + int i = dst_height; + do { + uint16x8_t s0 = load_unaligned_u16q(src_ptr, src_stride); + uint16x8_t s1 = load_unaligned_u16q(src_ptr + pixel_step, src_stride); + + uint16x8_t blend = vmulq_u16(s0, f0); + blend = vmlaq_u16(blend, s1, f1); + blend = vrshrq_n_u16(blend, 3); + + vst1q_u16(dst_ptr, blend); + + src_ptr += 2 * src_stride; + dst_ptr += 8; + i -= 2; + } while (i != 0); +} + +// Process a block which is a multiple of 8 and any height. +static void highbd_var_filter_block2d_bil_large(const uint16_t *src_ptr, + uint16_t *dst_ptr, + int src_stride, int pixel_step, + int dst_width, int dst_height, + int filter_offset) { + const uint16x8_t f0 = vdupq_n_u16(8 - filter_offset); + const uint16x8_t f1 = vdupq_n_u16(filter_offset); + + int i = dst_height; + do { + int j = 0; + do { + uint16x8_t s0 = vld1q_u16(src_ptr + j); + uint16x8_t s1 = vld1q_u16(src_ptr + j + pixel_step); + + uint16x8_t blend = vmulq_u16(s0, f0); + blend = vmlaq_u16(blend, s1, f1); + blend = vrshrq_n_u16(blend, 3); + + vst1q_u16(dst_ptr + j, blend); + + j += 8; + } while (j < dst_width); + + src_ptr += src_stride; + dst_ptr += dst_width; + } while (--i != 0); +} + +static void highbd_var_filter_block2d_bil_w8(const uint16_t *src_ptr, + uint16_t *dst_ptr, int src_stride, + int pixel_step, int dst_height, + int filter_offset) { + highbd_var_filter_block2d_bil_large(src_ptr, dst_ptr, src_stride, pixel_step, + 8, dst_height, filter_offset); +} +static void highbd_var_filter_block2d_bil_w16(const uint16_t *src_ptr, + uint16_t *dst_ptr, int src_stride, + int pixel_step, int dst_height, + int filter_offset) { + highbd_var_filter_block2d_bil_large(src_ptr, dst_ptr, src_stride, pixel_step, + 16, dst_height, filter_offset); +} +static void highbd_var_filter_block2d_bil_w32(const uint16_t *src_ptr, + uint16_t *dst_ptr, int src_stride, + int pixel_step, int dst_height, + int filter_offset) { + highbd_var_filter_block2d_bil_large(src_ptr, dst_ptr, src_stride, pixel_step, + 32, dst_height, filter_offset); +} +static void highbd_var_filter_block2d_bil_w64(const uint16_t *src_ptr, + uint16_t *dst_ptr, int src_stride, + int pixel_step, int dst_height, + int filter_offset) { + highbd_var_filter_block2d_bil_large(src_ptr, dst_ptr, src_stride, pixel_step, + 64, dst_height, filter_offset); +} + +#define HBD_SUBPEL_VARIANCE_WXH_NEON(w, h, padding) \ + unsigned int vpx_highbd_8_sub_pixel_variance##w##x##h##_neon( \ + const uint8_t *src, int src_stride, int xoffset, int yoffset, \ + const uint8_t *ref, int ref_stride, uint32_t *sse) { \ + uint16_t tmp0[w * (h + padding)]; \ + uint16_t tmp1[w * h]; \ + uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src); \ + \ + highbd_var_filter_block2d_bil_w##w(src_ptr, tmp0, src_stride, 1, \ + (h + padding), xoffset); \ + highbd_var_filter_block2d_bil_w##w(tmp0, tmp1, w, w, h, yoffset); \ + \ + return vpx_highbd_8_variance##w##x##h(CONVERT_TO_BYTEPTR(tmp1), w, ref, \ + ref_stride, sse); \ + } \ + \ + unsigned int vpx_highbd_10_sub_pixel_variance##w##x##h##_neon( \ + const uint8_t *src, int src_stride, int xoffset, int yoffset, \ + const uint8_t *ref, int ref_stride, uint32_t *sse) { \ + uint16_t tmp0[w * (h + padding)]; \ + uint16_t tmp1[w * h]; \ + uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src); \ + \ + highbd_var_filter_block2d_bil_w##w(src_ptr, tmp0, src_stride, 1, \ + (h + padding), xoffset); \ + highbd_var_filter_block2d_bil_w##w(tmp0, tmp1, w, w, h, yoffset); \ + \ + return vpx_highbd_10_variance##w##x##h(CONVERT_TO_BYTEPTR(tmp1), w, ref, \ + ref_stride, sse); \ + } \ + unsigned int vpx_highbd_12_sub_pixel_variance##w##x##h##_neon( \ + const uint8_t *src, int src_stride, int xoffset, int yoffset, \ + const uint8_t *ref, int ref_stride, uint32_t *sse) { \ + uint16_t tmp0[w * (h + padding)]; \ + uint16_t tmp1[w * h]; \ + uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src); \ + \ + highbd_var_filter_block2d_bil_w##w(src_ptr, tmp0, src_stride, 1, \ + (h + padding), xoffset); \ + highbd_var_filter_block2d_bil_w##w(tmp0, tmp1, w, w, h, yoffset); \ + \ + return vpx_highbd_12_variance##w##x##h(CONVERT_TO_BYTEPTR(tmp1), w, ref, \ + ref_stride, sse); \ + } + +// 4x blocks are processed two rows at a time, so require an extra row of +// padding. +HBD_SUBPEL_VARIANCE_WXH_NEON(4, 4, 2) +HBD_SUBPEL_VARIANCE_WXH_NEON(4, 8, 2) + +HBD_SUBPEL_VARIANCE_WXH_NEON(8, 4, 1) +HBD_SUBPEL_VARIANCE_WXH_NEON(8, 8, 1) +HBD_SUBPEL_VARIANCE_WXH_NEON(8, 16, 1) + +HBD_SUBPEL_VARIANCE_WXH_NEON(16, 8, 1) +HBD_SUBPEL_VARIANCE_WXH_NEON(16, 16, 1) +HBD_SUBPEL_VARIANCE_WXH_NEON(16, 32, 1) + +HBD_SUBPEL_VARIANCE_WXH_NEON(32, 16, 1) +HBD_SUBPEL_VARIANCE_WXH_NEON(32, 32, 1) +HBD_SUBPEL_VARIANCE_WXH_NEON(32, 64, 1) + +HBD_SUBPEL_VARIANCE_WXH_NEON(64, 32, 1) +HBD_SUBPEL_VARIANCE_WXH_NEON(64, 64, 1) + +void vpx_highbd_comp_avg_pred_neon(uint16_t *comp_pred, const uint16_t *pred, + int width, int height, const uint16_t *ref, + int ref_stride) { + int i, j; + uint32x4_t one_u32 = vdupq_n_u32(1); + if (width >= 8) { + for (i = 0; i < height; ++i) { + for (j = 0; j < width; j += 8) { + const uint16x8_t pred_u16 = vld1q_u16(&pred[j]); + const uint16x8_t ref_u16 = vld1q_u16(&ref[j]); + const uint32x4_t sum1_u32 = + vaddl_u16(vget_low_u16(pred_u16), vget_low_u16(ref_u16)); + const uint32x4_t sum2_u32 = + vaddl_u16(vget_high_u16(pred_u16), vget_high_u16(ref_u16)); + const uint16x4_t sum1_u16 = + vshrn_n_u32(vaddq_u32(sum1_u32, one_u32), 1); + const uint16x4_t sum2_u16 = + vshrn_n_u32(vaddq_u32(sum2_u32, one_u32), 1); + const uint16x8_t vcomp_pred = vcombine_u16(sum1_u16, sum2_u16); + vst1q_u16(&comp_pred[j], vcomp_pred); + } + comp_pred += width; + pred += width; + ref += ref_stride; + } + } else { + assert(width >= 4); + for (i = 0; i < height; ++i) { + for (j = 0; j < width; j += 4) { + const uint16x4_t pred_u16 = vld1_u16(&pred[j]); + const uint16x4_t ref_u16 = vld1_u16(&ref[j]); + const uint32x4_t sum_u32 = vaddl_u16(pred_u16, ref_u16); + const uint16x4_t vcomp_pred = + vshrn_n_u32(vaddq_u32(sum_u32, one_u32), 1); + vst1_u16(&comp_pred[j], vcomp_pred); + } + comp_pred += width; + pred += width; + ref += ref_stride; + } + } +} + +#define SUBPEL_AVG_VARIANCE_WXH_NEON(w, h, padding) \ + uint32_t vpx_highbd_8_sub_pixel_avg_variance##w##x##h##_neon( \ + const uint8_t *src, int src_stride, int xoffset, int yoffset, \ + const uint8_t *ref, int ref_stride, uint32_t *sse, \ + const uint8_t *second_pred) { \ + uint16_t tmp0[w * (h + padding)]; \ + uint16_t tmp1[w * h]; \ + uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src); \ + \ + highbd_var_filter_block2d_bil_w##w(src_ptr, tmp0, src_stride, 1, \ + (h + padding), xoffset); \ + highbd_var_filter_block2d_bil_w##w(tmp0, tmp1, w, w, h, yoffset); \ + \ + vpx_highbd_comp_avg_pred_neon(tmp0, CONVERT_TO_SHORTPTR(second_pred), w, \ + h, tmp1, w); \ + \ + return vpx_highbd_8_variance##w##x##h##_neon(CONVERT_TO_BYTEPTR(tmp0), w, \ + ref, ref_stride, sse); \ + } \ + \ + uint32_t vpx_highbd_10_sub_pixel_avg_variance##w##x##h##_neon( \ + const uint8_t *src, int src_stride, int xoffset, int yoffset, \ + const uint8_t *ref, int ref_stride, uint32_t *sse, \ + const uint8_t *second_pred) { \ + uint16_t tmp0[w * (h + padding)]; \ + uint16_t tmp1[w * h]; \ + uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src); \ + \ + highbd_var_filter_block2d_bil_w##w(src_ptr, tmp0, src_stride, 1, \ + (h + padding), xoffset); \ + highbd_var_filter_block2d_bil_w##w(tmp0, tmp1, w, w, h, yoffset); \ + \ + vpx_highbd_comp_avg_pred_neon(tmp0, CONVERT_TO_SHORTPTR(second_pred), w, \ + h, tmp1, w); \ + \ + return vpx_highbd_10_variance##w##x##h##_neon(CONVERT_TO_BYTEPTR(tmp0), w, \ + ref, ref_stride, sse); \ + } \ + \ + uint32_t vpx_highbd_12_sub_pixel_avg_variance##w##x##h##_neon( \ + const uint8_t *src, int src_stride, int xoffset, int yoffset, \ + const uint8_t *ref, int ref_stride, uint32_t *sse, \ + const uint8_t *second_pred) { \ + uint16_t tmp0[w * (h + padding)]; \ + uint16_t tmp1[w * h]; \ + uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src); \ + \ + highbd_var_filter_block2d_bil_w##w(src_ptr, tmp0, src_stride, 1, \ + (h + padding), xoffset); \ + highbd_var_filter_block2d_bil_w##w(tmp0, tmp1, w, w, h, yoffset); \ + \ + vpx_highbd_comp_avg_pred_neon(tmp0, CONVERT_TO_SHORTPTR(second_pred), w, \ + h, tmp1, w); \ + \ + return vpx_highbd_12_variance##w##x##h##_neon(CONVERT_TO_BYTEPTR(tmp0), w, \ + ref, ref_stride, sse); \ + } + +// 4x blocks are processed two rows at a time, so require an extra row of +// padding. +SUBPEL_AVG_VARIANCE_WXH_NEON(4, 4, 2) +SUBPEL_AVG_VARIANCE_WXH_NEON(4, 8, 2) + +SUBPEL_AVG_VARIANCE_WXH_NEON(8, 4, 1) +SUBPEL_AVG_VARIANCE_WXH_NEON(8, 8, 1) +SUBPEL_AVG_VARIANCE_WXH_NEON(8, 16, 1) + +SUBPEL_AVG_VARIANCE_WXH_NEON(16, 8, 1) +SUBPEL_AVG_VARIANCE_WXH_NEON(16, 16, 1) +SUBPEL_AVG_VARIANCE_WXH_NEON(16, 32, 1) + +SUBPEL_AVG_VARIANCE_WXH_NEON(32, 16, 1) +SUBPEL_AVG_VARIANCE_WXH_NEON(32, 32, 1) +SUBPEL_AVG_VARIANCE_WXH_NEON(32, 64, 1) + +SUBPEL_AVG_VARIANCE_WXH_NEON(64, 32, 1) +SUBPEL_AVG_VARIANCE_WXH_NEON(64, 64, 1) diff --git a/vpx_dsp/arm/highbd_variance_neon.c b/vpx_dsp/arm/highbd_variance_neon.c index 96a35af..985cc35 100644 --- a/vpx_dsp/arm/highbd_variance_neon.c +++ b/vpx_dsp/arm/highbd_variance_neon.c @@ -18,11 +18,6 @@ #include "vpx_dsp/arm/sum_neon.h" #include "vpx_ports/mem.h" -static const uint8_t bilinear_filters[8][2] = { - { 128, 0 }, { 112, 16 }, { 96, 32 }, { 80, 48 }, - { 64, 64 }, { 48, 80 }, { 32, 96 }, { 16, 112 }, -}; - static INLINE void highbd_variance16(const uint16_t *src_ptr, int src_stride, const uint16_t *ref_ptr, int ref_stride, int w, int h, uint64_t *sse, @@ -136,7 +131,7 @@ static INLINE void highbd_12_variance(const uint8_t *src8_ptr, int src_stride, *sum = (int)ROUND_POWER_OF_TWO(sum_long, 4); } -#define HIGHBD_VAR(W, H) \ +#define HBD_VARIANCE_WXH_NEON(W, H) \ uint32_t vpx_highbd_8_variance##W##x##H##_neon( \ const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, \ int ref_stride, uint32_t *sse) { \ @@ -218,274 +213,19 @@ static INLINE void highbd_12_variance(const uint8_t *src8_ptr, int src_stride, return *sse; \ } -static INLINE void highbd_var_filter_block2d_bil_first_pass( - const uint8_t *src_ptr8, uint16_t *output_ptr, - unsigned int src_pixels_per_line, int pixel_step, - unsigned int output_height, unsigned int output_width, - const uint8_t *filter) { - uint32_t i, j; - uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src_ptr8); - - uint32x4_t round_u32 = vshlq_n_u32(vdupq_n_u32(1), FILTER_BITS - 1); - uint16x4_t filter1_u16 = vdup_n_u16(filter[0]); - uint16x4_t filter2_u16 = vdup_n_u16(filter[1]); - - if (output_width >= 8) { - for (i = 0; i < output_height; ++i) { - for (j = 0; j < output_width; j += 8) { - const uint16x8_t src1_u16 = vld1q_u16(&src_ptr[j]); - const uint16x8_t src2_u16 = vld1q_u16(&src_ptr[j + pixel_step]); - uint32x4_t sum1_u32 = vmull_u16(filter1_u16, vget_low_u16(src1_u16)); - uint32x4_t sum2_u32 = vmull_u16(filter1_u16, vget_high_u16(src1_u16)); - uint16x4_t out1_u16; - uint16x4_t out2_u16; - sum1_u32 = vmlal_u16(sum1_u32, filter2_u16, vget_low_u16(src2_u16)); - sum2_u32 = vmlal_u16(sum2_u32, filter2_u16, vget_high_u16(src2_u16)); - out1_u16 = vshrn_n_u32(vaddq_u32(sum1_u32, round_u32), FILTER_BITS); - out2_u16 = vshrn_n_u32(vaddq_u32(sum2_u32, round_u32), FILTER_BITS); - vst1q_u16(&output_ptr[j], vcombine_u16(out1_u16, out2_u16)); - } - // Next row... - src_ptr += src_pixels_per_line; - output_ptr += output_width; - } - } else { - assert(output_width >= 4); - for (i = 0; i < output_height; ++i) { - for (j = 0; j < output_width; j += 4) { - const uint16x4_t src1_u16 = vld1_u16(&src_ptr[j]); - const uint16x4_t src2_u16 = vld1_u16(&src_ptr[j + pixel_step]); - uint32x4_t sum_u32 = vmull_u16(filter1_u16, src1_u16); - uint16x4_t out_u16; - sum_u32 = vmlal_u16(sum_u32, filter2_u16, src2_u16); - out_u16 = vshrn_n_u32(vaddq_u32(sum_u32, round_u32), FILTER_BITS); - vst1_u16(&output_ptr[j], out_u16); - } - // Next row... - src_ptr += src_pixels_per_line; - output_ptr += output_width; - } - } -} - -static INLINE void highbd_var_filter_block2d_bil_second_pass( - const uint16_t *src_ptr, uint16_t *output_ptr, - unsigned int src_pixels_per_line, unsigned int pixel_step, - unsigned int output_height, unsigned int output_width, - const uint8_t *filter) { - uint32_t i, j; - - uint32x4_t round_u32 = vshlq_n_u32(vdupq_n_u32(1), FILTER_BITS - 1); - uint16x4_t filter1_u16 = vdup_n_u16(filter[0]); - uint16x4_t filter2_u16 = vdup_n_u16(filter[1]); - - if (output_width >= 8) { - for (i = 0; i < output_height; ++i) { - for (j = 0; j < output_width; j += 8) { - const uint16x8_t src1_u16 = vld1q_u16(&src_ptr[j]); - const uint16x8_t src2_u16 = vld1q_u16(&src_ptr[j + pixel_step]); - uint32x4_t sum1_u32 = vmull_u16(filter1_u16, vget_low_u16(src1_u16)); - uint32x4_t sum2_u32 = vmull_u16(filter1_u16, vget_high_u16(src1_u16)); - uint16x4_t out1_u16; - uint16x4_t out2_u16; - sum1_u32 = vmlal_u16(sum1_u32, filter2_u16, vget_low_u16(src2_u16)); - sum2_u32 = vmlal_u16(sum2_u32, filter2_u16, vget_high_u16(src2_u16)); - out1_u16 = vshrn_n_u32(vaddq_u32(sum1_u32, round_u32), FILTER_BITS); - out2_u16 = vshrn_n_u32(vaddq_u32(sum2_u32, round_u32), FILTER_BITS); - vst1q_u16(&output_ptr[j], vcombine_u16(out1_u16, out2_u16)); - } - // Next row... - src_ptr += src_pixels_per_line; - output_ptr += output_width; - } - } else { - assert(output_width >= 4); - for (i = 0; i < output_height; ++i) { - for (j = 0; j < output_width; j += 4) { - const uint16x4_t src1_u16 = vld1_u16(&src_ptr[j]); - const uint16x4_t src2_u16 = vld1_u16(&src_ptr[j + pixel_step]); - uint32x4_t sum_u32 = vmull_u16(filter1_u16, src1_u16); - uint16x4_t out_u16; - sum_u32 = vmlal_u16(sum_u32, filter2_u16, src2_u16); - out_u16 = vshrn_n_u32(vaddq_u32(sum_u32, round_u32), FILTER_BITS); - vst1_u16(&output_ptr[j], out_u16); - } - // Next row... - src_ptr += src_pixels_per_line; - output_ptr += output_width; - } - } -} - -#define HIGHBD_SUBPIX_VAR(W, H) \ - uint32_t vpx_highbd_8_sub_pixel_variance##W##x##H##_neon( \ - const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, \ - const uint8_t *ref_ptr, int ref_stride, uint32_t *sse) { \ - uint16_t fdata3[(H + 1) * W]; \ - uint16_t temp2[H * W]; \ - \ - highbd_var_filter_block2d_bil_first_pass( \ - src_ptr, fdata3, src_stride, 1, H + 1, W, bilinear_filters[x_offset]); \ - highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, W, W, H, W, \ - bilinear_filters[y_offset]); \ - \ - return vpx_highbd_8_variance##W##x##H##_neon(CONVERT_TO_BYTEPTR(temp2), W, \ - ref_ptr, ref_stride, sse); \ - } \ - \ - uint32_t vpx_highbd_10_sub_pixel_variance##W##x##H##_neon( \ - const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, \ - const uint8_t *ref_ptr, int ref_stride, uint32_t *sse) { \ - uint16_t fdata3[(H + 1) * W]; \ - uint16_t temp2[H * W]; \ - \ - highbd_var_filter_block2d_bil_first_pass( \ - src_ptr, fdata3, src_stride, 1, H + 1, W, bilinear_filters[x_offset]); \ - highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, W, W, H, W, \ - bilinear_filters[y_offset]); \ - \ - return vpx_highbd_10_variance##W##x##H##_neon( \ - CONVERT_TO_BYTEPTR(temp2), W, ref_ptr, ref_stride, sse); \ - } \ - \ - uint32_t vpx_highbd_12_sub_pixel_variance##W##x##H##_neon( \ - const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, \ - const uint8_t *ref_ptr, int ref_stride, uint32_t *sse) { \ - uint16_t fdata3[(H + 1) * W]; \ - uint16_t temp2[H * W]; \ - \ - highbd_var_filter_block2d_bil_first_pass( \ - src_ptr, fdata3, src_stride, 1, H + 1, W, bilinear_filters[x_offset]); \ - highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, W, W, H, W, \ - bilinear_filters[y_offset]); \ - \ - return vpx_highbd_12_variance##W##x##H##_neon( \ - CONVERT_TO_BYTEPTR(temp2), W, ref_ptr, ref_stride, sse); \ - } - -#define HIGHBD_SUBPIX_AVG_VAR(W, H) \ - uint32_t vpx_highbd_8_sub_pixel_avg_variance##W##x##H##_neon( \ - const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, \ - const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, \ - const uint8_t *second_pred) { \ - uint16_t fdata3[(H + 1) * W]; \ - uint16_t temp2[H * W]; \ - DECLARE_ALIGNED(16, uint16_t, temp3[H * W]); \ - \ - highbd_var_filter_block2d_bil_first_pass( \ - src_ptr, fdata3, src_stride, 1, H + 1, W, bilinear_filters[x_offset]); \ - highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, W, W, H, W, \ - bilinear_filters[y_offset]); \ - \ - vpx_highbd_comp_avg_pred_neon(temp3, CONVERT_TO_SHORTPTR(second_pred), W, \ - H, temp2, W); \ - \ - return vpx_highbd_8_variance##W##x##H##_neon(CONVERT_TO_BYTEPTR(temp3), W, \ - ref_ptr, ref_stride, sse); \ - } \ - \ - uint32_t vpx_highbd_10_sub_pixel_avg_variance##W##x##H##_neon( \ - const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, \ - const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, \ - const uint8_t *second_pred) { \ - uint16_t fdata3[(H + 1) * W]; \ - uint16_t temp2[H * W]; \ - DECLARE_ALIGNED(16, uint16_t, temp3[H * W]); \ - \ - highbd_var_filter_block2d_bil_first_pass( \ - src_ptr, fdata3, src_stride, 1, H + 1, W, bilinear_filters[x_offset]); \ - highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, W, W, H, W, \ - bilinear_filters[y_offset]); \ - \ - vpx_highbd_comp_avg_pred_neon(temp3, CONVERT_TO_SHORTPTR(second_pred), W, \ - H, temp2, W); \ - \ - return vpx_highbd_10_variance##W##x##H##_neon( \ - CONVERT_TO_BYTEPTR(temp3), W, ref_ptr, ref_stride, sse); \ - } \ - \ - uint32_t vpx_highbd_12_sub_pixel_avg_variance##W##x##H##_neon( \ - const uint8_t *src_ptr, int src_stride, int x_offset, int y_offset, \ - const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, \ - const uint8_t *second_pred) { \ - uint16_t fdata3[(H + 1) * W]; \ - uint16_t temp2[H * W]; \ - DECLARE_ALIGNED(16, uint16_t, temp3[H * W]); \ - \ - highbd_var_filter_block2d_bil_first_pass( \ - src_ptr, fdata3, src_stride, 1, H + 1, W, bilinear_filters[x_offset]); \ - highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, W, W, H, W, \ - bilinear_filters[y_offset]); \ - \ - vpx_highbd_comp_avg_pred_neon(temp3, CONVERT_TO_SHORTPTR(second_pred), W, \ - H, temp2, W); \ - \ - return vpx_highbd_12_variance##W##x##H##_neon( \ - CONVERT_TO_BYTEPTR(temp3), W, ref_ptr, ref_stride, sse); \ - } - -void vpx_highbd_comp_avg_pred_neon(uint16_t *comp_pred, const uint16_t *pred, - int width, int height, const uint16_t *ref, - int ref_stride) { - int i, j; - uint32x4_t one_u32 = vdupq_n_u32(1); - if (width >= 8) { - for (i = 0; i < height; ++i) { - for (j = 0; j < width; j += 8) { - const uint16x8_t pred_u16 = vld1q_u16(&pred[j]); - const uint16x8_t ref_u16 = vld1q_u16(&ref[j]); - const uint32x4_t sum1_u32 = - vaddl_u16(vget_low_u16(pred_u16), vget_low_u16(ref_u16)); - const uint32x4_t sum2_u32 = - vaddl_u16(vget_high_u16(pred_u16), vget_high_u16(ref_u16)); - const uint16x4_t sum1_u16 = - vshrn_n_u32(vaddq_u32(sum1_u32, one_u32), 1); - const uint16x4_t sum2_u16 = - vshrn_n_u32(vaddq_u32(sum2_u32, one_u32), 1); - const uint16x8_t vcomp_pred = vcombine_u16(sum1_u16, sum2_u16); - vst1q_u16(&comp_pred[j], vcomp_pred); - } - comp_pred += width; - pred += width; - ref += ref_stride; - } - } else { - assert(width >= 4); - for (i = 0; i < height; ++i) { - for (j = 0; j < width; j += 4) { - const uint16x4_t pred_u16 = vld1_u16(&pred[j]); - const uint16x4_t ref_u16 = vld1_u16(&ref[j]); - const uint32x4_t sum_u32 = vaddl_u16(pred_u16, ref_u16); - const uint16x4_t vcomp_pred = - vshrn_n_u32(vaddq_u32(sum_u32, one_u32), 1); - vst1_u16(&comp_pred[j], vcomp_pred); - } - comp_pred += width; - pred += width; - ref += ref_stride; - } - } -} - -/* All three forms of the variance are available in the same sizes. */ -#define HIGHBD_VARIANCES(W, H) \ - HIGHBD_VAR(W, H) \ - HIGHBD_SUBPIX_VAR(W, H) \ - HIGHBD_SUBPIX_AVG_VAR(W, H) - -HIGHBD_VARIANCES(64, 64) -HIGHBD_VARIANCES(64, 32) -HIGHBD_VARIANCES(32, 64) -HIGHBD_VARIANCES(32, 32) -HIGHBD_VARIANCES(32, 16) -HIGHBD_VARIANCES(16, 32) -HIGHBD_VARIANCES(16, 16) -HIGHBD_VARIANCES(16, 8) -HIGHBD_VARIANCES(8, 16) -HIGHBD_VARIANCES(8, 8) -HIGHBD_VARIANCES(8, 4) -HIGHBD_VARIANCES(4, 8) -HIGHBD_VARIANCES(4, 4) +HBD_VARIANCE_WXH_NEON(64, 64) +HBD_VARIANCE_WXH_NEON(64, 32) +HBD_VARIANCE_WXH_NEON(32, 64) +HBD_VARIANCE_WXH_NEON(32, 32) +HBD_VARIANCE_WXH_NEON(32, 16) +HBD_VARIANCE_WXH_NEON(16, 32) +HBD_VARIANCE_WXH_NEON(16, 16) +HBD_VARIANCE_WXH_NEON(16, 8) +HBD_VARIANCE_WXH_NEON(8, 16) +HBD_VARIANCE_WXH_NEON(8, 8) +HBD_VARIANCE_WXH_NEON(8, 4) +HBD_VARIANCE_WXH_NEON(4, 8) +HBD_VARIANCE_WXH_NEON(4, 4) HIGHBD_GET_VAR(8) HIGHBD_GET_VAR(16) diff --git a/vpx_dsp/arm/mem_neon.h b/vpx_dsp/arm/mem_neon.h index 19cfc7c..866be74 100644 --- a/vpx_dsp/arm/mem_neon.h +++ b/vpx_dsp/arm/mem_neon.h @@ -126,6 +126,20 @@ static INLINE uint8x8_t load_unaligned_u8(const uint8_t *buf, return vreinterpret_u8_u32(a_u32); } +// Load 2 sets of 8 bytes when alignment is not guaranteed. +static INLINE uint16x8_t load_unaligned_u16q(const uint16_t *buf, + ptrdiff_t stride) { + uint64_t a; + uint64x2_t a_u64; + if (stride == 4) return vld1q_u16(buf); + memcpy(&a, buf, 8); + buf += stride; + a_u64 = vdupq_n_u64(a); + memcpy(&a, buf, 8); + a_u64 = vsetq_lane_u64(a, a_u64, 1); + return vreinterpretq_u16_u64(a_u64); +} + // Store 2 sets of 4 bytes when alignment is not guaranteed. static INLINE void store_unaligned_u8(uint8_t *buf, ptrdiff_t stride, const uint8x8_t a) { diff --git a/vpx_dsp/vpx_dsp.mk b/vpx_dsp/vpx_dsp.mk index 3b04e97..f10b7cc 100644 --- a/vpx_dsp/vpx_dsp.mk +++ b/vpx_dsp/vpx_dsp.mk @@ -434,6 +434,7 @@ DSP_SRCS-$(HAVE_SSE2) += x86/highbd_variance_sse2.c DSP_SRCS-$(HAVE_SSE2) += x86/highbd_variance_impl_sse2.asm DSP_SRCS-$(HAVE_SSE2) += x86/highbd_subpel_variance_impl_sse2.asm DSP_SRCS-$(HAVE_NEON) += arm/highbd_variance_neon.c +DSP_SRCS-$(HAVE_NEON) += arm/highbd_subpel_variance_neon.c endif # CONFIG_VP9_HIGHBITDEPTH endif # CONFIG_ENCODERS || CONFIG_POSTPROC || CONFIG_VP9_POSTPROC -- 2.7.4