From d35f992599f1f5460bd6a9320247f12b90937907 Mon Sep 17 00:00:00 2001 From: Parag Salasakar Date: Sat, 1 Aug 2015 08:05:25 +0530 Subject: [PATCH] mips msa vp8 denoising filter optimization average improvement ~2x-3x Change-Id: I6c17012c731fa4d56e0343f8de0df47b2dde289b --- vp8/common/rtcd_defs.pl | 4 +- vp8/encoder/mips/msa/denoising_msa.c | 624 +++++++++++++++++++++++++++ vp8/vp8cx.mk | 4 + 3 files changed, 630 insertions(+), 2 deletions(-) create mode 100644 vp8/encoder/mips/msa/denoising_msa.c diff --git a/vp8/common/rtcd_defs.pl b/vp8/common/rtcd_defs.pl index 6fe070b42..d65ca604d 100644 --- a/vp8/common/rtcd_defs.pl +++ b/vp8/common/rtcd_defs.pl @@ -334,9 +334,9 @@ if (vpx_config("CONFIG_REALTIME_ONLY") ne "yes") { # if (vpx_config("CONFIG_TEMPORAL_DENOISING") eq "yes") { add_proto qw/int vp8_denoiser_filter/, "unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising"; - specialize qw/vp8_denoiser_filter sse2 neon/; + specialize qw/vp8_denoiser_filter sse2 neon msa/; add_proto qw/int vp8_denoiser_filter_uv/, "unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising"; - specialize qw/vp8_denoiser_filter_uv sse2 neon/; + specialize qw/vp8_denoiser_filter_uv sse2 neon msa/; } # End of encoder only functions diff --git a/vp8/encoder/mips/msa/denoising_msa.c b/vp8/encoder/mips/msa/denoising_msa.c new file mode 100644 index 000000000..66965c668 --- /dev/null +++ b/vp8/encoder/mips/msa/denoising_msa.c @@ -0,0 +1,624 @@ +/* + * Copyright (c) 2015 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 "./vp8_rtcd.h" +#include "vp8/common/mips/msa/vp8_macros_msa.h" +#include "vp8/encoder/denoising.h" + +int32_t vp8_denoiser_filter_msa(uint8_t *mc_running_avg_y_ptr, + int32_t mc_avg_y_stride, + uint8_t *running_avg_y_ptr, + int32_t avg_y_stride, + uint8_t *sig_ptr, int32_t sig_stride, + uint32_t motion_magnitude, + int32_t increase_denoising) +{ + uint8_t *running_avg_y_start = running_avg_y_ptr; + uint8_t *sig_start = sig_ptr; + int32_t cnt = 0; + int32_t sum_diff = 0; + int32_t shift_inc1 = 3; + int32_t delta = 0; + int32_t sum_diff_thresh; + v16u8 src0, src1, src2, src3, src4, src5, src6, src7; + v16u8 src8, src9, src10, src11, src12, src13, src14, src15; + v16u8 mc_running_avg_y0, running_avg_y, sig0; + v16u8 mc_running_avg_y1, running_avg_y1, sig1; + v16u8 coeff0, coeff1; + v8i16 diff0, diff1, abs_diff0, abs_diff1, abs_diff_neg0, abs_diff_neg1; + v8i16 adjust0, adjust1, adjust2, adjust3; + v8i16 shift_inc1_vec = { 0 }; + v8i16 col_sum0 = { 0 }; + v8i16 col_sum1 = { 0 }; + v8i16 col_sum2 = { 0 }; + v8i16 col_sum3 = { 0 }; + v8i16 temp0_h, temp1_h, temp2_h, temp3_h, cmp, delta_vec; + v4i32 temp0_w; + v2i64 temp0_d, temp1_d; + v8i16 zero = { 0 }; + v8i16 one = __msa_ldi_h(1); + v8i16 four = __msa_ldi_h(4); + v8i16 val_127 = __msa_ldi_h(127); + v8i16 adj_val = { 6, 4, 3, 0, -6, -4, -3, 0 }; + + if (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) + { + adj_val = __msa_add_a_h(adj_val, one); + if (increase_denoising) + { + adj_val = __msa_add_a_h(adj_val, one); + shift_inc1 = 4; + } + + temp0_h = zero - adj_val; + adj_val = (v8i16)__msa_ilvev_d((v2i64)temp0_h, (v2i64)adj_val); + } + + adj_val = __msa_insert_h(adj_val, 3, cnt); + adj_val = __msa_insert_h(adj_val, 7, cnt); + shift_inc1_vec = __msa_fill_h(shift_inc1); + + for (cnt = 8; cnt--;) + { + v8i16 mask0 = { 0 }; + v8i16 mask1 = { 0 }; + + mc_running_avg_y0 = LD_UB(mc_running_avg_y_ptr); + sig0 = LD_UB(sig_ptr); + sig_ptr += sig_stride; + mc_running_avg_y_ptr += mc_avg_y_stride; + + mc_running_avg_y1 = LD_UB(mc_running_avg_y_ptr); + sig1 = LD_UB(sig_ptr); + + ILVRL_B2_UB(mc_running_avg_y0, sig0, coeff0, coeff1); + HSUB_UB2_SH(coeff0, coeff1, diff0, diff1); + abs_diff0 = __msa_add_a_h(diff0, zero); + abs_diff1 = __msa_add_a_h(diff1, zero); + cmp = __msa_clei_s_h(abs_diff0, 15); + cmp = cmp & one; + mask0 += cmp; + cmp = __msa_clei_s_h(abs_diff0, 7); + cmp = cmp & one; + mask0 += cmp; + cmp = abs_diff0 < shift_inc1_vec; + cmp = cmp & one; + mask0 += cmp; + cmp = __msa_clei_s_h(abs_diff1, 15); + cmp = cmp & one; + mask1 += cmp; + cmp = __msa_clei_s_h(abs_diff1, 7); + cmp = cmp & one; + mask1 += cmp; + cmp = abs_diff1 < shift_inc1_vec; + cmp = cmp & one; + mask1 += cmp; + temp0_h = __msa_clei_s_h(diff0, 0); + temp0_h = temp0_h & four; + mask0 += temp0_h; + temp1_h = __msa_clei_s_h(diff1, 0); + temp1_h = temp1_h & four; + mask1 += temp1_h; + VSHF_H2_SH(adj_val, adj_val, adj_val, adj_val, mask0, mask1, adjust0, + adjust1); + temp2_h = __msa_ceqi_h(adjust0, 0); + temp3_h = __msa_ceqi_h(adjust1, 0); + adjust0 = (v8i16)__msa_bmnz_v((v16u8)adjust0, (v16u8)diff0, + (v16u8)temp2_h); + adjust1 = (v8i16)__msa_bmnz_v((v16u8)adjust1, (v16u8)diff1, + (v16u8)temp3_h); + ADD2(col_sum0, adjust0, col_sum1, adjust1, col_sum0, col_sum1); + UNPCK_UB_SH(sig0, temp0_h, temp1_h); + ADD2(temp0_h, adjust0, temp1_h, adjust1, temp0_h, temp1_h); + MAXI_SH2_SH(temp0_h, temp1_h, 0); + SAT_UH2_SH(temp0_h, temp1_h, 7); + temp2_h = (v8i16)__msa_pckev_b((v16i8)temp3_h, (v16i8)temp2_h); + running_avg_y = (v16u8)__msa_pckev_b((v16i8)temp1_h, (v16i8)temp0_h); + running_avg_y = __msa_bmnz_v(running_avg_y, mc_running_avg_y0, + (v16u8)temp2_h); + ST_UB(running_avg_y, running_avg_y_ptr); + running_avg_y_ptr += avg_y_stride; + + mask0 = zero; + mask1 = zero; + ILVRL_B2_UB(mc_running_avg_y1, sig1, coeff0, coeff1); + HSUB_UB2_SH(coeff0, coeff1, diff0, diff1); + abs_diff0 = __msa_add_a_h(diff0, zero); + abs_diff1 = __msa_add_a_h(diff1, zero); + cmp = __msa_clei_s_h(abs_diff0, 15); + cmp = cmp & one; + mask0 += cmp; + cmp = __msa_clei_s_h(abs_diff0, 7); + cmp = cmp & one; + mask0 += cmp; + cmp = abs_diff0 < shift_inc1_vec; + cmp = cmp & one; + mask0 += cmp; + cmp = __msa_clei_s_h(abs_diff1, 15); + cmp = cmp & one; + mask1 += cmp; + cmp = __msa_clei_s_h(abs_diff1, 7); + cmp = cmp & one; + mask1 += cmp; + cmp = abs_diff1 < shift_inc1_vec; + cmp = cmp & one; + mask1 += cmp; + temp0_h = __msa_clei_s_h(diff0, 0); + temp0_h = temp0_h & four; + mask0 += temp0_h; + temp1_h = __msa_clei_s_h(diff1, 0); + temp1_h = temp1_h & four; + mask1 += temp1_h; + VSHF_H2_SH(adj_val, adj_val, adj_val, adj_val, mask0, mask1, adjust0, + adjust1); + temp2_h = __msa_ceqi_h(adjust0, 0); + temp3_h = __msa_ceqi_h(adjust1, 0); + adjust0 = (v8i16)__msa_bmnz_v((v16u8)adjust0, (v16u8)diff0, + (v16u8)temp2_h); + adjust1 = (v8i16)__msa_bmnz_v((v16u8)adjust1, (v16u8)diff1, + (v16u8)temp3_h); + ADD2(col_sum0, adjust0, col_sum1, adjust1, col_sum0, col_sum1); + UNPCK_UB_SH(sig1, temp0_h, temp1_h); + ADD2(temp0_h, adjust0, temp1_h, adjust1, temp0_h, temp1_h); + MAXI_SH2_SH(temp0_h, temp1_h, 0); + SAT_UH2_SH(temp0_h, temp1_h, 7); + temp2_h = (v8i16)__msa_pckev_b((v16i8)temp3_h, (v16i8)temp2_h); + running_avg_y = (v16u8)__msa_pckev_b((v16i8)temp1_h, (v16i8)temp0_h); + running_avg_y = __msa_bmnz_v(running_avg_y, mc_running_avg_y1, + (v16u8)temp2_h); + ST_UB(running_avg_y, running_avg_y_ptr); + sig_ptr += sig_stride; + mc_running_avg_y_ptr += mc_avg_y_stride; + running_avg_y_ptr += avg_y_stride; + } + + col_sum0 = __msa_min_s_h(col_sum0, val_127); + col_sum1 = __msa_min_s_h(col_sum1, val_127); + temp0_h = col_sum0 + col_sum1; + temp0_w = __msa_hadd_s_w(temp0_h, temp0_h); + temp0_d = __msa_hadd_s_d(temp0_w, temp0_w); + temp1_d = __msa_splati_d(temp0_d, 1); + temp0_d += temp1_d; + sum_diff = __msa_copy_s_w((v4i32)temp0_d, 0); + sig_ptr -= sig_stride * 16; + mc_running_avg_y_ptr -= mc_avg_y_stride * 16; + running_avg_y_ptr -= avg_y_stride * 16; + + if (increase_denoising) + { + sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH; + } + + if (abs(sum_diff) > sum_diff_thresh) + { + delta = ((abs(sum_diff) - sum_diff_thresh) >> 8) + 1; + delta_vec = __msa_fill_h(delta); + if (delta < 4) + { + for (cnt = 8; cnt--;) + { + running_avg_y = LD_UB(running_avg_y_ptr); + mc_running_avg_y0 = LD_UB(mc_running_avg_y_ptr); + sig0 = LD_UB(sig_ptr); + sig_ptr += sig_stride; + mc_running_avg_y_ptr += mc_avg_y_stride; + running_avg_y_ptr += avg_y_stride; + mc_running_avg_y1 = LD_UB(mc_running_avg_y_ptr); + sig1 = LD_UB(sig_ptr); + running_avg_y1 = LD_UB(running_avg_y_ptr); + ILVRL_B2_UB(mc_running_avg_y0, sig0, coeff0, coeff1); + HSUB_UB2_SH(coeff0, coeff1, diff0, diff1); + abs_diff0 = __msa_add_a_h(diff0, zero); + abs_diff1 = __msa_add_a_h(diff1, zero); + temp0_h = abs_diff0 < delta_vec; + temp1_h = abs_diff1 < delta_vec; + abs_diff0 = (v8i16)__msa_bmz_v((v16u8)abs_diff0, + (v16u8)delta_vec, + (v16u8)temp0_h); + abs_diff1 = (v8i16)__msa_bmz_v((v16u8)abs_diff1, + (v16u8)delta_vec, + (v16u8)temp1_h); + SUB2(zero, abs_diff0, zero, abs_diff1, abs_diff_neg0, + abs_diff_neg1); + abs_diff_neg0 = zero - abs_diff0; + abs_diff_neg1 = zero - abs_diff1; + temp0_h = __msa_clei_s_h(diff0, 0); + temp1_h = __msa_clei_s_h(diff1, 0); + adjust0 = (v8i16)__msa_bmnz_v((v16u8)abs_diff0, + (v16u8)abs_diff_neg0, + (v16u8)temp0_h); + adjust1 = (v8i16)__msa_bmnz_v((v16u8)abs_diff1, + (v16u8)abs_diff_neg1, + (v16u8)temp1_h); + ILVRL_B2_SH(zero, running_avg_y, temp2_h, temp3_h); + ADD2(temp2_h, adjust0, temp3_h, adjust1, adjust2, adjust3); + MAXI_SH2_SH(adjust2, adjust3, 0); + SAT_UH2_SH(adjust2, adjust3, 7); + temp0_h = __msa_ceqi_h(diff0, 0); + temp1_h = __msa_ceqi_h(diff1, 0); + adjust2 = (v8i16)__msa_bmz_v((v16u8)adjust2, (v16u8)temp2_h, + (v16u8)temp0_h); + adjust3 = (v8i16)__msa_bmz_v((v16u8)adjust3, (v16u8)temp3_h, + (v16u8)temp1_h); + adjust0 = (v8i16)__msa_bmnz_v((v16u8)adjust0, (v16u8)zero, + (v16u8)temp0_h); + adjust1 = (v8i16)__msa_bmnz_v((v16u8)adjust1, (v16u8)zero, + (v16u8)temp1_h); + ADD2(col_sum2, adjust0, col_sum3, adjust1, col_sum2, col_sum3); + running_avg_y = (v16u8)__msa_pckev_b((v16i8)adjust3, + (v16i8)adjust2); + ST_UB(running_avg_y, running_avg_y_ptr - avg_y_stride); + ILVRL_B2_UB(mc_running_avg_y1, sig1, coeff0, coeff1); + HSUB_UB2_SH(coeff0, coeff1, diff0, diff1); + abs_diff0 = __msa_add_a_h(diff0, zero); + abs_diff1 = __msa_add_a_h(diff1, zero); + temp0_h = abs_diff0 < delta_vec; + temp1_h = abs_diff1 < delta_vec; + abs_diff0 = (v8i16)__msa_bmz_v((v16u8)abs_diff0, + (v16u8)delta_vec, + (v16u8)temp0_h); + abs_diff1 = (v8i16)__msa_bmz_v((v16u8)abs_diff1, + (v16u8)delta_vec, + (v16u8)temp1_h); + SUB2(zero, abs_diff0, zero, abs_diff1, abs_diff_neg0, + abs_diff_neg1); + temp0_h = __msa_clei_s_h(diff0, 0); + temp1_h = __msa_clei_s_h(diff1, 0); + adjust0 = (v8i16)__msa_bmnz_v((v16u8)abs_diff0, + (v16u8)abs_diff_neg0, + (v16u8)temp0_h); + adjust1 = (v8i16)__msa_bmnz_v((v16u8)abs_diff1, + (v16u8)abs_diff_neg1, + (v16u8)temp1_h); + ILVRL_H2_SH(zero, running_avg_y1, temp2_h, temp3_h); + ADD2(temp2_h, adjust0, temp3_h, adjust1, adjust2, adjust3); + MAXI_SH2_SH(adjust2, adjust3, 0); + SAT_UH2_SH(adjust2, adjust3, 7); + temp0_h = __msa_ceqi_h(diff0, 0); + temp1_h = __msa_ceqi_h(diff1, 0); + adjust2 = (v8i16)__msa_bmz_v((v16u8)adjust2, (v16u8)temp2_h, + (v16u8)temp0_h); + adjust3 = (v8i16)__msa_bmz_v((v16u8)adjust3, (v16u8)temp3_h, + (v16u8)temp1_h); + adjust0 = (v8i16)__msa_bmz_v((v16u8)adjust0, (v16u8)zero, + (v16u8)temp0_h); + adjust1 = (v8i16)__msa_bmz_v((v16u8)adjust1, (v16u8)zero, + (v16u8)temp1_h); + ADD2(col_sum2, adjust0, col_sum3, adjust1, col_sum2, col_sum3); + running_avg_y = (v16u8)__msa_pckev_b((v16i8)adjust3, + (v16i8)adjust2); + ST_UB(running_avg_y, running_avg_y_ptr); + running_avg_y_ptr += avg_y_stride; + } + + col_sum2 = __msa_min_s_h(col_sum2, val_127); + col_sum3 = __msa_min_s_h(col_sum3, val_127); + temp0_h = col_sum2 + col_sum3; + temp0_w = __msa_hadd_s_w(temp0_h, temp0_h); + temp0_d = __msa_hadd_s_d(temp0_w, temp0_w); + temp1_d = __msa_splati_d(temp0_d, 1); + temp0_d += (v2i64)temp1_d; + sum_diff = __msa_copy_s_w((v4i32)temp0_d, 0); + if (abs(sum_diff) > SUM_DIFF_THRESHOLD) + { + return COPY_BLOCK; + } + } + else + { + return COPY_BLOCK; + } + } + + LD_UB8(sig_start, sig_stride, src0, src1, src2, src3, src4, src5, src6, + src7); + sig_start += (8 * sig_stride); + LD_UB8(sig_start, sig_stride, src8, src9, src10, src11, src12, src13, + src14, src15); + + ST_UB8(src0, src1, src2, src3, src4, src5, src6, src7, running_avg_y_start, + avg_y_stride); + running_avg_y_start += (8 * avg_y_stride); + ST_UB8(src8, src9, src10, src11, src12, src13, src14, src15, + running_avg_y_start, avg_y_stride); + + return FILTER_BLOCK; +} + +int32_t vp8_denoiser_filter_uv_msa(uint8_t *mc_running_avg_y_ptr, + int32_t mc_avg_y_stride, + uint8_t *running_avg_y_ptr, + int32_t avg_y_stride, + uint8_t *sig_ptr, + int32_t sig_stride, + uint32_t motion_magnitude, + int32_t increase_denoising) +{ + uint8_t *running_avg_y_start = running_avg_y_ptr; + uint8_t *sig_start = sig_ptr; + int32_t cnt = 0; + int32_t sum_diff = 0; + int32_t shift_inc1 = 3; + int32_t delta = 0; + int32_t sum_block = 0; + int32_t sum_diff_thresh; + int64_t dst0, dst1, src0, src1, src2, src3; + v16u8 mc_running_avg_y0, running_avg_y, sig0; + v16u8 mc_running_avg_y1, running_avg_y1, sig1; + v16u8 sig2, sig3, sig4, sig5, sig6, sig7; + v16u8 coeff0; + v8i16 diff0, abs_diff0, abs_diff_neg0; + v8i16 adjust0, adjust2; + v8i16 shift_inc1_vec = { 0 }; + v8i16 col_sum0 = { 0 }; + v8i16 temp0_h, temp2_h, cmp, delta_vec; + v4i32 temp0_w; + v2i64 temp0_d, temp1_d; + v16i8 zero = { 0 }; + v8i16 one = __msa_ldi_h(1); + v8i16 four = __msa_ldi_h(4); + v8i16 adj_val = { 6, 4, 3, 0, -6, -4, -3, 0 }; + + + sig0 = LD_UB(sig_ptr); + sig_ptr += sig_stride; + temp0_h = (v8i16)__msa_ilvr_b(zero, (v16i8)sig0); + sig1 = LD_UB(sig_ptr); + sig_ptr += sig_stride; + temp0_h += (v8i16)__msa_ilvr_b(zero, (v16i8)sig1); + sig2 = LD_UB(sig_ptr); + sig_ptr += sig_stride; + temp0_h += (v8i16)__msa_ilvr_b(zero, (v16i8)sig2); + sig3 = LD_UB(sig_ptr); + sig_ptr += sig_stride; + temp0_h += (v8i16)__msa_ilvr_b(zero, (v16i8)sig3); + sig4 = LD_UB(sig_ptr); + sig_ptr += sig_stride; + temp0_h += (v8i16)__msa_ilvr_b(zero, (v16i8)sig4); + sig5 = LD_UB(sig_ptr); + sig_ptr += sig_stride; + temp0_h += (v8i16)__msa_ilvr_b(zero, (v16i8)sig5); + sig6 = LD_UB(sig_ptr); + sig_ptr += sig_stride; + temp0_h += (v8i16)__msa_ilvr_b(zero, (v16i8)sig6); + sig7 = LD_UB(sig_ptr); + sig_ptr += sig_stride; + temp0_h += (v8i16)__msa_ilvr_b(zero, (v16i8)sig7); + temp0_w = __msa_hadd_s_w(temp0_h, temp0_h); + temp0_d = __msa_hadd_s_d(temp0_w, temp0_w); + temp1_d = __msa_splati_d(temp0_d, 1); + temp0_d += temp1_d; + sum_block = __msa_copy_s_w((v4i32)temp0_d, 0); + sig_ptr -= sig_stride * 8; + + if (abs(sum_block - (128 * 8 * 8)) < SUM_DIFF_FROM_AVG_THRESH_UV) + { + return COPY_BLOCK; + } + + if (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) + { + adj_val = __msa_add_a_h(adj_val, one); + + if (increase_denoising) + { + adj_val = __msa_add_a_h(adj_val, one); + shift_inc1 = 4; + } + + temp0_h = (v8i16)zero - adj_val; + adj_val = (v8i16)__msa_ilvev_d((v2i64)temp0_h, (v2i64)adj_val); + } + + adj_val = __msa_insert_h(adj_val, 3, cnt); + adj_val = __msa_insert_h(adj_val, 7, cnt); + shift_inc1_vec = __msa_fill_h(shift_inc1); + for (cnt = 4; cnt--;) + { + v8i16 mask0 = { 0 }; + mc_running_avg_y0 = LD_UB(mc_running_avg_y_ptr); + sig0 = LD_UB(sig_ptr); + sig_ptr += sig_stride; + mc_running_avg_y_ptr += mc_avg_y_stride; + mc_running_avg_y1 = LD_UB(mc_running_avg_y_ptr); + sig1 = LD_UB(sig_ptr); + coeff0 = (v16u8)__msa_ilvr_b((v16i8)mc_running_avg_y0, (v16i8)sig0); + diff0 = __msa_hsub_u_h(coeff0, coeff0); + abs_diff0 = __msa_add_a_h(diff0, (v8i16)zero); + cmp = __msa_clei_s_h(abs_diff0, 15); + cmp = cmp & one; + mask0 += cmp; + cmp = __msa_clei_s_h(abs_diff0, 7); + cmp = cmp & one; + mask0 += cmp; + cmp = abs_diff0 < shift_inc1_vec; + cmp = cmp & one; + mask0 += cmp; + temp0_h = __msa_clei_s_h(diff0, 0); + temp0_h = temp0_h & four; + mask0 += temp0_h; + adjust0 = __msa_vshf_h(mask0, adj_val, adj_val); + temp2_h = __msa_ceqi_h(adjust0, 0); + adjust0 = (v8i16)__msa_bmnz_v((v16u8)adjust0, (v16u8)diff0, + (v16u8)temp2_h); + col_sum0 += adjust0; + temp0_h = (v8i16)__msa_ilvr_b(zero, (v16i8)sig0); + temp0_h += adjust0; + temp0_h = __msa_maxi_s_h(temp0_h, 0); + temp0_h = (v8i16)__msa_sat_u_h((v8u16)temp0_h, 7); + temp2_h = (v8i16)__msa_pckev_b((v16i8)temp2_h, (v16i8)temp2_h); + running_avg_y = (v16u8)__msa_pckev_b((v16i8)temp0_h, (v16i8)temp0_h); + running_avg_y = __msa_bmnz_v(running_avg_y, mc_running_avg_y0, + (v16u8)temp2_h); + dst0 = __msa_copy_s_d((v2i64)running_avg_y, 0); + SD(dst0, running_avg_y_ptr); + running_avg_y_ptr += avg_y_stride; + + mask0 = __msa_ldi_h(0); + coeff0 = (v16u8)__msa_ilvr_b((v16i8)mc_running_avg_y1, (v16i8)sig1); + diff0 = __msa_hsub_u_h(coeff0, coeff0); + abs_diff0 = __msa_add_a_h(diff0, (v8i16)zero); + cmp = __msa_clei_s_h(abs_diff0, 15); + cmp = cmp & one; + mask0 += cmp; + cmp = __msa_clei_s_h(abs_diff0, 7); + cmp = cmp & one; + mask0 += cmp; + cmp = abs_diff0 < shift_inc1_vec; + cmp = cmp & one; + mask0 += cmp; + temp0_h = __msa_clei_s_h(diff0, 0); + temp0_h = temp0_h & four; + mask0 += temp0_h; + adjust0 = __msa_vshf_h(mask0, adj_val, adj_val); + temp2_h = __msa_ceqi_h(adjust0, 0); + adjust0 = (v8i16)__msa_bmnz_v((v16u8)adjust0, (v16u8)diff0, + (v16u8)temp2_h); + col_sum0 += adjust0; + temp0_h = (v8i16)__msa_ilvr_b(zero, (v16i8)sig1); + temp0_h += adjust0; + temp0_h = __msa_maxi_s_h(temp0_h, 0); + temp0_h = (v8i16)__msa_sat_u_h((v8u16)temp0_h, 7); + + temp2_h = (v8i16)__msa_pckev_b((v16i8)temp2_h, (v16i8)temp2_h); + running_avg_y = (v16u8)__msa_pckev_b((v16i8)temp0_h, (v16i8)temp0_h); + running_avg_y = __msa_bmnz_v(running_avg_y, mc_running_avg_y1, + (v16u8)temp2_h); + dst1 = __msa_copy_s_d((v2i64)running_avg_y, 0); + SD(dst1, running_avg_y_ptr); + + sig_ptr += sig_stride; + mc_running_avg_y_ptr += mc_avg_y_stride; + running_avg_y_ptr += avg_y_stride; + } + + temp0_h = col_sum0; + temp0_w = __msa_hadd_s_w(temp0_h, temp0_h); + temp0_d = __msa_hadd_s_d(temp0_w, temp0_w); + temp1_d = __msa_splati_d(temp0_d, 1); + temp0_d += temp1_d; + sum_diff = __msa_copy_s_w((v4i32)temp0_d, 0); + sig_ptr -= sig_stride * 8; + mc_running_avg_y_ptr -= mc_avg_y_stride * 8; + running_avg_y_ptr -= avg_y_stride * 8; + sum_diff_thresh = SUM_DIFF_THRESHOLD_UV; + + if (increase_denoising) + { + sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH_UV; + } + + if (abs(sum_diff) > sum_diff_thresh) + { + delta = ((abs(sum_diff) - sum_diff_thresh) >> 8) + 1; + delta_vec = __msa_fill_h(delta); + if (delta < 4) + { + for (cnt = 4; cnt--;) + { + running_avg_y = LD_UB(running_avg_y_ptr); + mc_running_avg_y0 = LD_UB(mc_running_avg_y_ptr); + sig0 = LD_UB(sig_ptr); + /* Update pointers for next iteration. */ + sig_ptr += sig_stride; + mc_running_avg_y_ptr += mc_avg_y_stride; + running_avg_y_ptr += avg_y_stride; + + mc_running_avg_y1 = LD_UB(mc_running_avg_y_ptr); + sig1 = LD_UB(sig_ptr); + running_avg_y1 = LD_UB(running_avg_y_ptr); + + coeff0 = (v16u8)__msa_ilvr_b((v16i8)mc_running_avg_y0, + (v16i8)sig0); + diff0 = __msa_hsub_u_h(coeff0, coeff0); + abs_diff0 = __msa_add_a_h(diff0, (v8i16)zero); + temp0_h = delta_vec < abs_diff0; + abs_diff0 = (v8i16)__msa_bmnz_v((v16u8)abs_diff0, + (v16u8)delta_vec, + (v16u8)temp0_h); + abs_diff_neg0 = (v8i16)zero - abs_diff0; + temp0_h = __msa_clei_s_h(diff0, 0); + adjust0 = (v8i16)__msa_bmz_v((v16u8)abs_diff0, + (v16u8)abs_diff_neg0, + (v16u8)temp0_h); + temp2_h = (v8i16)__msa_ilvr_b(zero, (v16i8)running_avg_y); + adjust2 = temp2_h + adjust0; + adjust2 = __msa_maxi_s_h(adjust2, 0); + adjust2 = (v8i16)__msa_sat_u_h((v8u16)adjust2, 7); + temp0_h = __msa_ceqi_h(diff0, 0); + adjust2 = (v8i16)__msa_bmnz_v((v16u8)adjust2, (v16u8)temp2_h, + (v16u8)temp0_h); + adjust0 = (v8i16)__msa_bmnz_v((v16u8)adjust0, (v16u8)zero, + (v16u8)temp0_h); + col_sum0 += adjust0; + running_avg_y = (v16u8)__msa_pckev_b((v16i8)adjust2, + (v16i8)adjust2); + dst0 = __msa_copy_s_d((v2i64)running_avg_y, 0); + SD(dst0, running_avg_y_ptr - avg_y_stride); + + coeff0 = (v16u8)__msa_ilvr_b((v16i8)mc_running_avg_y1, + (v16i8)sig1); + diff0 = __msa_hsub_u_h(coeff0, coeff0); + abs_diff0 = __msa_add_a_h(diff0, (v8i16)zero); + temp0_h = delta_vec < abs_diff0; + abs_diff0 = (v8i16)__msa_bmnz_v((v16u8)abs_diff0, + (v16u8)delta_vec, + (v16u8)temp0_h); + abs_diff_neg0 = (v8i16)zero - abs_diff0; + temp0_h = __msa_clei_s_h(diff0, 0); + adjust0 = (v8i16)__msa_bmz_v((v16u8)abs_diff0, + (v16u8)abs_diff_neg0, + (v16u8)temp0_h); + temp2_h = (v8i16)__msa_ilvr_b(zero, (v16i8)running_avg_y1); + adjust2 = temp2_h + adjust0; + adjust2 = __msa_maxi_s_h(adjust2, 0); + adjust2 = (v8i16)__msa_sat_u_h((v8u16)adjust2, 7); + temp0_h = __msa_ceqi_h(diff0, 0); + adjust2 = (v8i16)__msa_bmnz_v((v16u8)adjust2, (v16u8)temp2_h, + (v16u8)temp0_h); + adjust0 = (v8i16)__msa_bmnz_v((v16u8)adjust0, (v16u8)zero, + (v16u8)temp0_h); + col_sum0 += adjust0; + running_avg_y = (v16u8)__msa_pckev_b((v16i8)adjust2, + (v16i8)adjust2); + dst1 = __msa_copy_s_d((v2i64)running_avg_y, 0); + SD(dst1, running_avg_y_ptr); + running_avg_y_ptr += avg_y_stride; + } + + temp0_h = col_sum0; + temp0_w = __msa_hadd_s_w(temp0_h, temp0_h); + temp0_d = __msa_hadd_s_d(temp0_w, temp0_w); + temp1_d = __msa_splati_d(temp0_d, 1); + temp0_d += temp1_d; + sum_diff = __msa_copy_s_w((v4i32)temp0_d, 0); + + if (abs(sum_diff) > sum_diff_thresh) + { + return COPY_BLOCK; + } + } + else + { + return COPY_BLOCK; + } + } + + LD4(sig_start, sig_stride, src0, src1, src2, src3); + sig_start += (4 * sig_stride); + SD4(src0, src1, src2, src3, running_avg_y_start, avg_y_stride); + running_avg_y_start += (4 * avg_y_stride); + + LD4(sig_start, sig_stride, src0, src1, src2, src3); + SD4(src0, src1, src2, src3, running_avg_y_start, avg_y_stride); + + return FILTER_BLOCK; +} diff --git a/vp8/vp8cx.mk b/vp8/vp8cx.mk index 1bafbfea0..59f9fe839 100644 --- a/vp8/vp8cx.mk +++ b/vp8/vp8cx.mk @@ -108,6 +108,10 @@ VP8_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/encodeopt_msa.c VP8_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/quantize_msa.c VP8_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/temporal_filter_msa.c +ifeq ($(CONFIG_TEMPORAL_DENOISING),yes) +VP8_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/denoising_msa.c +endif + ifeq ($(CONFIG_REALTIME_ONLY),yes) VP8_CX_SRCS_REMOVE-$(HAVE_MSA) += encoder/mips/msa/temporal_filter_msa.c endif -- 2.40.0