From 624b1367004801639ed35759d5f1759a092c8410 Mon Sep 17 00:00:00 2001 From: yuanhecai Date: Tue, 1 Mar 2022 09:48:13 +0800 Subject: [PATCH] vp9[loongarch]: Optimize horizontal/vertical_8_c 1. vpx_lpf_vertical_8_lsx 2. vpx_lpf_horizontal_8_lsx Bug: webm:1755 Change-Id: I6b05d6b1b2ac4d2a75beb9c9ca9700976fc3af55 --- test/lpf_test.cc | 12 +++ vpx_dsp/loongarch/loopfilter_8_lsx.c | 199 +++++++++++++++++++++++++++++++++++ vpx_dsp/vpx_dsp.mk | 2 + vpx_dsp/vpx_dsp_rtcd_defs.pl | 4 +- 4 files changed, 215 insertions(+), 2 deletions(-) create mode 100644 vpx_dsp/loongarch/loopfilter_8_lsx.c diff --git a/test/lpf_test.cc b/test/lpf_test.cc index 62c6f30..833dfb9 100644 --- a/test/lpf_test.cc +++ b/test/lpf_test.cc @@ -692,4 +692,16 @@ INSTANTIATE_TEST_SUITE_P( &vpx_lpf_vertical_8_dual_c, 8))); #endif // HAVE_MSA && (!CONFIG_VP9_HIGHBITDEPTH) +#if HAVE_LSX && (!CONFIG_VP9_HIGHBITDEPTH) +INSTANTIATE_TEST_SUITE_P( + LSX, Loop8Test6Param, + ::testing::Values( + make_tuple(&vpx_lpf_horizontal_8_lsx, &vpx_lpf_horizontal_8_c, 8), + make_tuple(&vpx_lpf_horizontal_16_dual_lsx, + &vpx_lpf_horizontal_16_dual_c, 8), + make_tuple(&vpx_lpf_vertical_8_lsx, &vpx_lpf_vertical_8_c, 8), + make_tuple(&vpx_lpf_vertical_16_dual_lsx, &vpx_lpf_vertical_16_dual_c, + 8))); +#endif // HAVE_LSX && (!CONFIG_VP9_HIGHBITDEPTH) + } // namespace diff --git a/vpx_dsp/loongarch/loopfilter_8_lsx.c b/vpx_dsp/loongarch/loopfilter_8_lsx.c new file mode 100644 index 0000000..facf6f3 --- /dev/null +++ b/vpx_dsp/loongarch/loopfilter_8_lsx.c @@ -0,0 +1,199 @@ +/* + * Copyright (c) 2022 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 "./vpx_dsp_rtcd.h" +#include "vpx_dsp/loongarch/loopfilter_lsx.h" + +void vpx_lpf_horizontal_8_lsx(uint8_t *dst, int32_t stride, + const uint8_t *b_limit_ptr, + const uint8_t *limit_ptr, + const uint8_t *thresh_ptr) { + __m128i mask, hev, flat, thresh, b_limit, limit; + __m128i p3, p2, p1, p0, q3, q2, q1, q0; + __m128i p2_out, p1_out, p0_out, q0_out, q1_out, q2_out; + __m128i p2_filter8, p1_filter8, p0_filter8; + __m128i q0_filter8, q1_filter8, q2_filter8; + __m128i p3_l, p2_l, p1_l, p0_l, q3_l, q2_l, q1_l, q0_l; + __m128i zero = __lsx_vldi(0); + + int32_t stride2 = stride << 1; + int32_t stride3 = stride2 + stride; + int32_t stride4 = stride2 << 1; + + /* load vector elements */ + DUP4_ARG2(__lsx_vldx, dst, -stride4, dst, -stride3, dst, -stride2, dst, + -stride, p3, p2, p1, p0); + q0 = __lsx_vld(dst, 0); + DUP2_ARG2(__lsx_vldx, dst, stride, dst, stride2, q1, q2); + q3 = __lsx_vldx(dst, stride3); + + thresh = __lsx_vreplgr2vr_b(*thresh_ptr); + b_limit = __lsx_vreplgr2vr_b(*b_limit_ptr); + limit = __lsx_vreplgr2vr_b(*limit_ptr); + + LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, hev, + mask, flat); + VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat); + VP9_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out); + + flat = __lsx_vilvl_d(zero, flat); + + if (__lsx_bz_v(flat)) { + __lsx_vstelm_d(p1_out, dst - stride2, 0, 0); + __lsx_vstelm_d(p0_out, dst - stride, 0, 0); + __lsx_vstelm_d(q0_out, dst, 0, 0); + __lsx_vstelm_d(q1_out, dst + stride, 0, 0); + } else { + DUP4_ARG2(__lsx_vilvl_b, zero, p3, zero, p2, zero, p1, zero, p0, p3_l, p2_l, + p1_l, p0_l); + DUP4_ARG2(__lsx_vilvl_b, zero, q0, zero, q1, zero, q2, zero, q3, q0_l, q1_l, + q2_l, q3_l); + VP9_FILTER8(p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l, p2_filter8, + p1_filter8, p0_filter8, q0_filter8, q1_filter8, q2_filter8); + + /* convert 16 bit output data into 8 bit */ + DUP4_ARG2(__lsx_vpickev_b, zero, p2_filter8, zero, p1_filter8, zero, + p0_filter8, zero, q0_filter8, p2_filter8, p1_filter8, p0_filter8, + q0_filter8); + DUP2_ARG2(__lsx_vpickev_b, zero, q1_filter8, zero, q2_filter8, q1_filter8, + q2_filter8); + DUP4_ARG3(__lsx_vbitsel_v, p2, p2_filter8, flat, p1_out, p1_filter8, flat, + p0_out, p0_filter8, flat, q0_out, q0_filter8, flat, p2_out, + p1_out, p0_out, q0_out); + DUP2_ARG3(__lsx_vbitsel_v, q1_out, q1_filter8, flat, q2, q2_filter8, flat, + q1_out, q2_out); + dst -= stride3; + + __lsx_vstelm_d(p2_out, dst, 0, 0); + __lsx_vstelm_d(p1_out, dst + stride, 0, 0); + __lsx_vstelm_d(p0_out, dst + stride2, 0, 0); + __lsx_vstelm_d(q0_out, dst + stride3, 0, 0); + + dst += stride4; + __lsx_vstelm_d(q1_out, dst, 0, 0); + dst += stride; + __lsx_vstelm_d(q2_out, dst, 0, 0); + } +} + +void vpx_lpf_vertical_8_lsx(uint8_t *dst, int32_t stride, + const uint8_t *b_limit_ptr, + const uint8_t *limit_ptr, + const uint8_t *thresh_ptr) { + __m128i p3, p2, p1, p0, q3, q2, q1, q0; + __m128i p1_out, p0_out, q0_out, q1_out; + __m128i flat, mask, hev, thresh, b_limit, limit; + __m128i p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l; + __m128i p2_filt8_l, p1_filt8_l, p0_filt8_l; + __m128i q0_filt8_l, q1_filt8_l, q2_filt8_l; + __m128i vec0, vec1, vec2, vec3, vec4; + __m128i zero = __lsx_vldi(0); + + int32_t stride2 = stride << 1; + int32_t stride3 = stride2 + stride; + int32_t stride4 = stride2 << 1; + uint8_t *dst_tmp = dst - 4; + + /* load vector elements */ + p3 = __lsx_vld(dst_tmp, 0); + DUP2_ARG2(__lsx_vldx, dst_tmp, stride, dst_tmp, stride2, p2, p1); + p0 = __lsx_vldx(dst_tmp, stride3); + dst_tmp += stride4; + q0 = __lsx_vld(dst_tmp, 0); + DUP2_ARG2(__lsx_vldx, dst_tmp, stride, dst_tmp, stride2, q1, q2); + q3 = __lsx_vldx(dst_tmp, stride3); + + LSX_TRANSPOSE8x8_B(p3, p2, p1, p0, q0, q1, q2, q3, p3, p2, p1, p0, q0, q1, q2, + q3); + + thresh = __lsx_vreplgr2vr_b(*thresh_ptr); + b_limit = __lsx_vreplgr2vr_b(*b_limit_ptr); + limit = __lsx_vreplgr2vr_b(*limit_ptr); + + /* mask and hev */ + LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, hev, + mask, flat); + /* flat4 */ + VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat); + /* filter4 */ + VP9_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out); + + flat = __lsx_vilvl_d(zero, flat); + + /* if flat is zero for all pixels, then no need to calculate other filter */ + if (__lsx_bz_v(flat)) { + /* Store 4 pixels p1-_q1 */ + DUP2_ARG2(__lsx_vilvl_b, p0_out, p1_out, q1_out, q0_out, vec0, vec1); + vec2 = __lsx_vilvl_h(vec1, vec0); + vec3 = __lsx_vilvh_h(vec1, vec0); + + dst -= 2; + __lsx_vstelm_w(vec2, dst, 0, 0); + __lsx_vstelm_w(vec2, dst + stride, 0, 1); + __lsx_vstelm_w(vec2, dst + stride2, 0, 2); + __lsx_vstelm_w(vec2, dst + stride3, 0, 3); + dst += stride4; + __lsx_vstelm_w(vec3, dst, 0, 0); + __lsx_vstelm_w(vec3, dst + stride, 0, 1); + __lsx_vstelm_w(vec3, dst + stride2, 0, 2); + __lsx_vstelm_w(vec3, dst + stride3, 0, 3); + } else { + DUP4_ARG2(__lsx_vilvl_b, zero, p3, zero, p2, zero, p1, zero, p0, p3_l, p2_l, + p1_l, p0_l); + DUP4_ARG2(__lsx_vilvl_b, zero, q0, zero, q1, zero, q2, zero, q3, q0_l, q1_l, + q2_l, q3_l); + VP9_FILTER8(p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l, p2_filt8_l, + p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l); + /* convert 16 bit output data into 8 bit */ + DUP4_ARG2(__lsx_vpickev_b, p2_filt8_l, p2_filt8_l, p1_filt8_l, p1_filt8_l, + p0_filt8_l, p0_filt8_l, q0_filt8_l, q0_filt8_l, p2_filt8_l, + p1_filt8_l, p0_filt8_l, q0_filt8_l); + DUP2_ARG2(__lsx_vpickev_b, q1_filt8_l, q1_filt8_l, q2_filt8_l, q2_filt8_l, + q1_filt8_l, q2_filt8_l); + /* store pixel values */ + p2 = __lsx_vbitsel_v(p2, p2_filt8_l, flat); + p1 = __lsx_vbitsel_v(p1_out, p1_filt8_l, flat); + p0 = __lsx_vbitsel_v(p0_out, p0_filt8_l, flat); + q0 = __lsx_vbitsel_v(q0_out, q0_filt8_l, flat); + q1 = __lsx_vbitsel_v(q1_out, q1_filt8_l, flat); + q2 = __lsx_vbitsel_v(q2, q2_filt8_l, flat); + + /* Store 6 pixels p2-_q2 */ + DUP2_ARG2(__lsx_vilvl_b, p1, p2, q0, p0, vec0, vec1); + vec2 = __lsx_vilvl_h(vec1, vec0); + vec3 = __lsx_vilvh_h(vec1, vec0); + vec4 = __lsx_vilvl_b(q2, q1); + + dst -= 3; + __lsx_vstelm_w(vec2, dst, 0, 0); + __lsx_vstelm_h(vec4, dst, 4, 0); + dst += stride; + __lsx_vstelm_w(vec2, dst, 0, 1); + __lsx_vstelm_h(vec4, dst, 4, 1); + dst += stride; + __lsx_vstelm_w(vec2, dst, 0, 2); + __lsx_vstelm_h(vec4, dst, 4, 2); + dst += stride; + __lsx_vstelm_w(vec2, dst, 0, 3); + __lsx_vstelm_h(vec4, dst, 4, 3); + dst += stride; + __lsx_vstelm_w(vec3, dst, 0, 0); + __lsx_vstelm_h(vec4, dst, 4, 4); + dst += stride; + __lsx_vstelm_w(vec3, dst, 0, 1); + __lsx_vstelm_h(vec4, dst, 4, 5); + dst += stride; + __lsx_vstelm_w(vec3, dst, 0, 2); + __lsx_vstelm_h(vec4, dst, 4, 6); + dst += stride; + __lsx_vstelm_w(vec3, dst, 0, 3); + __lsx_vstelm_h(vec4, dst, 4, 7); + } +} diff --git a/vpx_dsp/vpx_dsp.mk b/vpx_dsp/vpx_dsp.mk index eb530db..976c652 100644 --- a/vpx_dsp/vpx_dsp.mk +++ b/vpx_dsp/vpx_dsp.mk @@ -189,7 +189,9 @@ DSP_SRCS-$(HAVE_DSPR2) += mips/loopfilter_mb_dspr2.c DSP_SRCS-$(HAVE_DSPR2) += mips/loopfilter_mb_horiz_dspr2.c DSP_SRCS-$(HAVE_DSPR2) += mips/loopfilter_mb_vert_dspr2.c +DSP_SRCS-$(HAVE_LSX) += loongarch/loopfilter_lsx.h DSP_SRCS-$(HAVE_LSX) += loongarch/loopfilter_16_lsx.c +DSP_SRCS-$(HAVE_LSX) += loongarch/loopfilter_8_lsx.c ifeq ($(CONFIG_VP9_HIGHBITDEPTH),yes) DSP_SRCS-$(HAVE_NEON) += arm/highbd_loopfilter_neon.c diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl index 9cd58d3..ce0780f 100644 --- a/vpx_dsp/vpx_dsp_rtcd_defs.pl +++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl @@ -445,7 +445,7 @@ add_proto qw/void vpx_lpf_vertical_16_dual/, "uint8_t *s, int pitch, const uint8 specialize qw/vpx_lpf_vertical_16_dual sse2 neon dspr2 msa lsx/; add_proto qw/void vpx_lpf_vertical_8/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh"; -specialize qw/vpx_lpf_vertical_8 sse2 neon dspr2 msa/; +specialize qw/vpx_lpf_vertical_8 sse2 neon dspr2 msa lsx/; add_proto qw/void vpx_lpf_vertical_8_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1"; specialize qw/vpx_lpf_vertical_8_dual sse2 neon dspr2 msa/; @@ -463,7 +463,7 @@ add_proto qw/void vpx_lpf_horizontal_16_dual/, "uint8_t *s, int pitch, const uin specialize qw/vpx_lpf_horizontal_16_dual sse2 avx2 neon dspr2 msa lsx/; add_proto qw/void vpx_lpf_horizontal_8/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh"; -specialize qw/vpx_lpf_horizontal_8 sse2 neon dspr2 msa/; +specialize qw/vpx_lpf_horizontal_8 sse2 neon dspr2 msa lsx/; add_proto qw/void vpx_lpf_horizontal_8_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1"; specialize qw/vpx_lpf_horizontal_8_dual sse2 neon dspr2 msa/; -- 2.7.4