From 220643c8627d158f75acbf7e1b7dcd9ae642261c Mon Sep 17 00:00:00 2001 From: yuanhecai Date: Tue, 1 Mar 2022 16:33:47 +0800 Subject: [PATCH] vp9[loongarch]: Optimize convolve8_horiz/vert/c 1. vpx_convolve8_lsx 2. vpx_convolve8_vert_lsx 3. vpx_convolve8_horiz_lsx Bug: webm:1755 Change-Id: I9897e1ed6a904ac74d1078bd22b275af44db142d --- test/convolve_test.cc | 13 + vpx_dsp/loongarch/vpx_convolve8_horiz_lsx.c | 852 +++++++++++++++++++++++++++ vpx_dsp/loongarch/vpx_convolve8_lsx.c | 718 +++++++++++++++++++++++ vpx_dsp/loongarch/vpx_convolve8_vert_lsx.c | 870 ++++++++++++++++++++++++++++ vpx_dsp/loongarch/vpx_convolve_lsx.h | 117 ++++ vpx_dsp/vpx_dsp.mk | 6 + vpx_dsp/vpx_dsp_rtcd_defs.pl | 6 +- 7 files changed, 2579 insertions(+), 3 deletions(-) create mode 100644 vpx_dsp/loongarch/vpx_convolve8_horiz_lsx.c create mode 100644 vpx_dsp/loongarch/vpx_convolve8_lsx.c create mode 100644 vpx_dsp/loongarch/vpx_convolve8_vert_lsx.c create mode 100644 vpx_dsp/loongarch/vpx_convolve_lsx.h diff --git a/test/convolve_test.cc b/test/convolve_test.cc index 4b2dade..94b2814 100644 --- a/test/convolve_test.cc +++ b/test/convolve_test.cc @@ -1449,6 +1449,19 @@ INSTANTIATE_TEST_SUITE_P(MSA, ConvolveTest, ::testing::ValuesIn(kArrayConvolve8_msa)); #endif // HAVE_MSA +#if HAVE_LSX +const ConvolveFunctions convolve8_lsx( + vpx_convolve_copy_c, vpx_convolve_avg_c, vpx_convolve8_horiz_lsx, + vpx_convolve8_avg_horiz_c, vpx_convolve8_vert_lsx, vpx_convolve8_avg_vert_c, + vpx_convolve8_lsx, vpx_convolve8_avg_c, vpx_scaled_horiz_c, + vpx_scaled_avg_horiz_c, vpx_scaled_vert_c, vpx_scaled_avg_vert_c, + vpx_scaled_2d_c, vpx_scaled_avg_2d_c, 0); + +const ConvolveParam kArrayConvolve8_lsx[] = { ALL_SIZES(convolve8_lsx) }; +INSTANTIATE_TEST_SUITE_P(LSX, ConvolveTest, + ::testing::ValuesIn(kArrayConvolve8_lsx)); +#endif // HAVE_LSX + #if HAVE_VSX const ConvolveFunctions convolve8_vsx( vpx_convolve_copy_vsx, vpx_convolve_avg_vsx, vpx_convolve8_horiz_vsx, diff --git a/vpx_dsp/loongarch/vpx_convolve8_horiz_lsx.c b/vpx_dsp/loongarch/vpx_convolve8_horiz_lsx.c new file mode 100644 index 0000000..3608fe3 --- /dev/null +++ b/vpx_dsp/loongarch/vpx_convolve8_horiz_lsx.c @@ -0,0 +1,852 @@ +/* + * 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 +#include "./vpx_dsp_rtcd.h" +#include "vpx_dsp/loongarch/vpx_convolve_lsx.h" + +#define LSX_LD_4(_src, _stride, _src0, _src1, _src2, _src3) \ + { \ + _src0 = __lsx_vld(_src, 0); \ + _src += _stride; \ + _src1 = __lsx_vld(_src, 0); \ + _src += _stride; \ + _src2 = __lsx_vld(_src, 0); \ + _src += _stride; \ + _src3 = __lsx_vld(_src, 0); \ + } + +static const uint8_t mc_filt_mask_arr[16 * 3] = { + /* 8 width cases */ + 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, + /* 4 width cases */ + 0, 1, 1, 2, 2, 3, 3, 4, 16, 17, 17, 18, 18, 19, 19, 20, + /* 4 width cases */ + 8, 9, 9, 10, 10, 11, 11, 12, 24, 25, 25, 26, 26, 27, 27, 28 +}; + +static void common_hz_8t_4x4_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + const int8_t *filter) { + __m128i src0, src1, src2, src3; + __m128i filter0, filter1, filter2, filter3; + __m128i mask0, mask1, mask2, mask3; + __m128i out, out0, out1; + + mask0 = __lsx_vld(mc_filt_mask_arr, 16); + src -= 3; + DUP4_ARG2(__lsx_vldrepl_h, filter, 0, filter, 2, filter, 4, filter, 6, + filter0, filter1, filter2, filter3); + DUP2_ARG2(__lsx_vaddi_bu, mask0, 2, mask0, 4, mask1, mask2); + mask3 = __lsx_vaddi_bu(mask0, 6); + + LSX_LD_4(src, src_stride, src0, src1, src2, src3); + DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128, src0, + src1, src2, src3); + HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, mask3, + filter0, filter1, filter2, filter3, out0, out1); + out = __lsx_vssrarni_b_h(out1, out0, 7); + out = __lsx_vxori_b(out, 128); + __lsx_vstelm_w(out, dst, 0, 0); + dst += dst_stride; + __lsx_vstelm_w(out, dst, 0, 1); + dst += dst_stride; + __lsx_vstelm_w(out, dst, 0, 2); + dst += dst_stride; + __lsx_vstelm_w(out, dst, 0, 3); +} + +static void common_hz_8t_4x8_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + const int8_t *filter) { + int32_t src_stride2 = src_stride << 1; + int32_t src_stride3 = src_stride + src_stride2; + int32_t src_stride4 = src_stride2 << 1; + __m128i src0, src1, src2, src3; + __m128i filter0, filter1, filter2, filter3; + __m128i mask0, mask1, mask2, mask3; + __m128i out0, out1, out2, out3; + uint8_t *_src = (uint8_t *)src - 3; + + mask0 = __lsx_vld(mc_filt_mask_arr, 16); + DUP2_ARG2(__lsx_vaddi_bu, mask0, 2, mask0, 4, mask1, mask2); + mask3 = __lsx_vaddi_bu(mask0, 6); + DUP4_ARG2(__lsx_vldrepl_h, filter, 0, filter, 2, filter, 4, filter, 6, + filter0, filter1, filter2, filter3); + + src0 = __lsx_vld(_src, 0); + DUP2_ARG2(__lsx_vldx, _src, src_stride, _src, src_stride2, src1, src2); + src3 = __lsx_vldx(_src, src_stride3); + _src += src_stride4; + DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128, src0, + src1, src2, src3); + HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, mask3, + filter0, filter1, filter2, filter3, out0, out1); + src0 = __lsx_vld(_src, 0); + DUP2_ARG2(__lsx_vldx, _src, src_stride, _src, src_stride2, src1, src2); + src3 = __lsx_vldx(_src, src_stride3); + DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128, src0, + src1, src2, src3); + HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, mask3, + filter0, filter1, filter2, filter3, out2, out3); + DUP2_ARG3(__lsx_vssrarni_b_h, out1, out0, 7, out3, out2, 7, out0, out1); + DUP2_ARG2(__lsx_vxori_b, out0, 128, out1, 128, out0, out1); + + __lsx_vstelm_w(out0, dst, 0, 0); + dst += dst_stride; + __lsx_vstelm_w(out0, dst, 0, 1); + dst += dst_stride; + __lsx_vstelm_w(out0, dst, 0, 2); + dst += dst_stride; + __lsx_vstelm_w(out0, dst, 0, 3); + dst += dst_stride; + __lsx_vstelm_w(out1, dst, 0, 0); + dst += dst_stride; + __lsx_vstelm_w(out1, dst, 0, 1); + dst += dst_stride; + __lsx_vstelm_w(out1, dst, 0, 2); + dst += dst_stride; + __lsx_vstelm_w(out1, dst, 0, 3); +} + +static void common_hz_8t_4w_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + const int8_t *filter, int32_t height) { + if (height == 4) { + common_hz_8t_4x4_lsx(src, src_stride, dst, dst_stride, filter); + } else if (height == 8) { + common_hz_8t_4x8_lsx(src, src_stride, dst, dst_stride, filter); + } +} + +static void common_hz_8t_8x4_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + const int8_t *filter) { + __m128i src0, src1, src2, src3; + __m128i filter0, filter1, filter2, filter3; + __m128i mask0, mask1, mask2, mask3; + __m128i out0, out1, out2, out3; + + mask0 = __lsx_vld(mc_filt_mask_arr, 0); + src -= 3; + DUP2_ARG2(__lsx_vaddi_bu, mask0, 2, mask0, 4, mask1, mask2); + mask3 = __lsx_vaddi_bu(mask0, 6); + DUP4_ARG2(__lsx_vldrepl_h, filter, 0, filter, 2, filter, 4, filter, 6, + filter0, filter1, filter2, filter3); + + LSX_LD_4(src, src_stride, src0, src1, src2, src3); + DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128, src0, + src1, src2, src3); + HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, mask3, + filter0, filter1, filter2, filter3, out0, out1, + out2, out3); + DUP2_ARG3(__lsx_vssrarni_b_h, out1, out0, 7, out3, out2, 7, out0, out1); + DUP2_ARG2(__lsx_vxori_b, out0, 128, out1, 128, out0, out1); + __lsx_vstelm_d(out0, dst, 0, 0); + dst += dst_stride; + __lsx_vstelm_d(out0, dst, 0, 1); + dst += dst_stride; + __lsx_vstelm_d(out1, dst, 0, 0); + dst += dst_stride; + __lsx_vstelm_d(out1, dst, 0, 1); +} + +static void common_hz_8t_8x8mult_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + const int8_t *filter, int32_t height) { + uint32_t loop_cnt = height >> 2; + int32_t src_stride2 = src_stride << 1; + int32_t src_stride3 = src_stride + src_stride2; + int32_t src_stride4 = src_stride2 << 1; + __m128i src0, src1, src2, src3; + __m128i filter0, filter1, filter2, filter3; + __m128i mask0, mask1, mask2, mask3; + __m128i out0, out1, out2, out3; + uint8_t *_src = (uint8_t *)src - 3; + + mask0 = __lsx_vld(mc_filt_mask_arr, 0); + DUP2_ARG2(__lsx_vaddi_bu, mask0, 2, mask0, 4, mask1, mask2); + mask3 = __lsx_vaddi_bu(mask0, 6); + DUP4_ARG2(__lsx_vldrepl_h, filter, 0, filter, 2, filter, 4, filter, 6, + filter0, filter1, filter2, filter3); + + for (; loop_cnt--;) { + src0 = __lsx_vld(_src, 0); + DUP2_ARG2(__lsx_vldx, _src, src_stride, _src, src_stride2, src1, src2); + src3 = __lsx_vldx(_src, src_stride3); + _src += src_stride4; + DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128, src0, + src1, src2, src3); + HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, + mask3, filter0, filter1, filter2, filter3, out0, + out1, out2, out3); + DUP2_ARG3(__lsx_vssrarni_b_h, out1, out0, 7, out3, out2, 7, out0, out1); + DUP2_ARG2(__lsx_vxori_b, out0, 128, out1, 128, out0, out1); + __lsx_vstelm_d(out0, dst, 0, 0); + dst += dst_stride; + __lsx_vstelm_d(out0, dst, 0, 1); + dst += dst_stride; + __lsx_vstelm_d(out1, dst, 0, 0); + dst += dst_stride; + __lsx_vstelm_d(out1, dst, 0, 1); + dst += dst_stride; + } +} + +static void common_hz_8t_8w_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + const int8_t *filter, int32_t height) { + if (height == 4) { + common_hz_8t_8x4_lsx(src, src_stride, dst, dst_stride, filter); + } else { + common_hz_8t_8x8mult_lsx(src, src_stride, dst, dst_stride, filter, height); + } +} + +static void common_hz_8t_16w_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + const int8_t *filter, int32_t height) { + uint32_t loop_cnt = height >> 1; + int32_t stride = src_stride << 1; + __m128i src0, src1, src2, src3; + __m128i filter0, filter1, filter2, filter3; + __m128i mask0, mask1, mask2, mask3; + __m128i out0, out1, out2, out3; + + mask0 = __lsx_vld(mc_filt_mask_arr, 0); + src -= 3; + DUP2_ARG2(__lsx_vaddi_bu, mask0, 2, mask0, 4, mask1, mask2); + mask3 = __lsx_vaddi_bu(mask0, 6); + DUP4_ARG2(__lsx_vldrepl_h, filter, 0, filter, 2, filter, 4, filter, 6, + filter0, filter1, filter2, filter3); + + for (; loop_cnt--;) { + const uint8_t *_src = src + src_stride; + DUP2_ARG2(__lsx_vld, src, 0, _src, 0, src0, src2); + DUP2_ARG2(__lsx_vld, src, 8, _src, 8, src1, src3); + DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128, src0, + src1, src2, src3); + HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, + mask3, filter0, filter1, filter2, filter3, out0, + out1, out2, out3); + DUP2_ARG3(__lsx_vssrarni_b_h, out1, out0, 7, out3, out2, 7, out0, out1); + DUP2_ARG2(__lsx_vxori_b, out0, 128, out1, 128, out0, out1); + __lsx_vst(out0, dst, 0); + dst += dst_stride; + __lsx_vst(out1, dst, 0); + dst += dst_stride; + src += stride; + } +} + +static void common_hz_8t_32w_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + const int8_t *filter, int32_t height) { + uint32_t loop_cnt = height >> 1; + __m128i src0, src1, src2, src3; + __m128i filter0, filter1, filter2, filter3; + __m128i mask0, mask1, mask2, mask3; + __m128i out0, out1, out2, out3; + __m128i shuff = { 0x0F0E0D0C0B0A0908, 0x1716151413121110 }; + + mask0 = __lsx_vld(mc_filt_mask_arr, 0); + src -= 3; + DUP2_ARG2(__lsx_vaddi_bu, mask0, 2, mask0, 4, mask1, mask2); + mask3 = __lsx_vaddi_bu(mask0, 6); + DUP4_ARG2(__lsx_vldrepl_h, filter, 0, filter, 2, filter, 4, filter, 6, + filter0, filter1, filter2, filter3); + + for (; loop_cnt--;) { + DUP2_ARG2(__lsx_vld, src, 0, src, 16, src0, src2); + src3 = __lsx_vld(src, 24); + src1 = __lsx_vshuf_b(src2, src0, shuff); + src += src_stride; + DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128, src0, + src1, src2, src3); + HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, + mask3, filter0, filter1, filter2, filter3, out0, + out1, out2, out3); + DUP2_ARG3(__lsx_vssrarni_b_h, out1, out0, 7, out3, out2, 7, out0, out1); + DUP2_ARG2(__lsx_vxori_b, out0, 128, out1, 128, out0, out1); + __lsx_vst(out0, dst, 0); + __lsx_vst(out1, dst, 16); + + DUP2_ARG2(__lsx_vld, src, 0, src, 16, src0, src2); + src3 = __lsx_vld(src, 24); + src1 = __lsx_vshuf_b(src2, src0, shuff); + src += src_stride; + + dst += dst_stride; + DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128, src0, + src1, src2, src3); + HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, + mask3, filter0, filter1, filter2, filter3, out0, + out1, out2, out3); + DUP2_ARG3(__lsx_vssrarni_b_h, out1, out0, 7, out3, out2, 7, out0, out1); + DUP2_ARG2(__lsx_vxori_b, out0, 128, out1, 128, out0, out1); + __lsx_vst(out0, dst, 0); + __lsx_vst(out1, dst, 16); + dst += dst_stride; + } +} + +static void common_hz_8t_64w_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + const int8_t *filter, int32_t height) { + int32_t loop_cnt = height; + __m128i src0, src1, src2, src3; + __m128i filter0, filter1, filter2, filter3; + __m128i mask0, mask1, mask2, mask3; + __m128i out0, out1, out2, out3; + __m128i shuff = { 0x0F0E0D0C0B0A0908, 0x1716151413121110 }; + + mask0 = __lsx_vld(mc_filt_mask_arr, 0); + src -= 3; + DUP2_ARG2(__lsx_vaddi_bu, mask0, 2, mask0, 4, mask1, mask2); + mask3 = __lsx_vaddi_bu(mask0, 6); + DUP4_ARG2(__lsx_vldrepl_h, filter, 0, filter, 2, filter, 4, filter, 6, + filter0, filter1, filter2, filter3); + + for (; loop_cnt--;) { + DUP2_ARG2(__lsx_vld, src, 0, src, 16, src0, src2); + src3 = __lsx_vld(src, 24); + src1 = __lsx_vshuf_b(src2, src0, shuff); + DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128, src0, + src1, src2, src3); + HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, + mask3, filter0, filter1, filter2, filter3, out0, + out1, out2, out3); + DUP2_ARG3(__lsx_vssrarni_b_h, out1, out0, 7, out3, out2, 7, out0, out1); + DUP2_ARG2(__lsx_vxori_b, out0, 128, out1, 128, out0, out1); + __lsx_vst(out0, dst, 0); + __lsx_vst(out1, dst, 16); + + DUP2_ARG2(__lsx_vld, src, 32, src, 48, src0, src2); + src3 = __lsx_vld(src, 56); + src1 = __lsx_vshuf_b(src2, src0, shuff); + DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128, src0, + src1, src2, src3); + HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, + mask3, filter0, filter1, filter2, filter3, out0, + out1, out2, out3); + DUP2_ARG3(__lsx_vssrarni_b_h, out1, out0, 7, out3, out2, 7, out0, out1); + DUP2_ARG2(__lsx_vxori_b, out0, 128, out1, 128, out0, out1); + __lsx_vst(out0, dst, 32); + __lsx_vst(out1, dst, 48); + src += src_stride; + dst += dst_stride; + } +} + +static void common_hz_2t_4x4_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter) { + __m128i src0, src1, src2, src3, mask; + __m128i filt0, vec0, vec1, res0, res1; + __m128i vec2, vec3; + int32_t src_stride2 = src_stride << 1; + int32_t src_stride3 = src_stride + src_stride2; + + int32_t dst_stride2 = dst_stride << 1; + int32_t dst_stride3 = dst_stride + dst_stride2; + + mask = __lsx_vld(mc_filt_mask_arr, 16); + /* rearranging filter */ + filt0 = __lsx_vldrepl_h(filter, 0); + + src0 = __lsx_vld(src, 0); + DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride2, src1, src2); + src3 = __lsx_vldx(src, src_stride3); + DUP2_ARG3(__lsx_vshuf_b, src1, src0, mask, src3, src2, mask, vec0, vec1); + DUP2_ARG2(__lsx_vdp2_h_bu, vec0, filt0, vec1, filt0, vec2, vec3); + DUP2_ARG2(__lsx_vsrari_h, vec2, FILTER_BITS, vec3, FILTER_BITS, vec2, vec3); + DUP2_ARG2(__lsx_vpickev_b, vec2, vec2, vec3, vec3, res0, res1); + + __lsx_vstelm_w(res0, dst, 0, 0); + __lsx_vstelm_w(res0, dst + dst_stride, 0, 1); + __lsx_vstelm_w(res1, dst + dst_stride2, 0, 0); + __lsx_vstelm_w(res1, dst + dst_stride3, 0, 1); +} + +static void common_hz_2t_4x8_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter) { + __m128i vec0, vec1, vec2, vec3, filt0; + __m128i src0, src1, src2, src3, src4, src5, src6, src7, mask; + __m128i res0, res1, res2, res3; + __m128i vec4, vec5, vec6, vec7; + int32_t src_stride2 = src_stride << 1; + int32_t src_stride3 = src_stride + src_stride2; + int32_t src_stride4 = src_stride2 << 1; + + int32_t dst_stride2 = dst_stride << 1; + int32_t dst_stride3 = dst_stride + dst_stride2; + + uint8_t *src_tmp1 = src + src_stride4; + + mask = __lsx_vld(mc_filt_mask_arr, 16); + + /* rearranging filter */ + filt0 = __lsx_vldrepl_h(filter, 0); + + src0 = __lsx_vld(src, 0); + DUP4_ARG2(__lsx_vldx, src, src_stride, src, src_stride2, src, src_stride3, + src, src_stride4, src1, src2, src3, src4); + DUP2_ARG2(__lsx_vldx, src_tmp1, src_stride, src_tmp1, src_stride2, src5, + src6); + src7 = __lsx_vldx(src_tmp1, src_stride3); + + DUP4_ARG3(__lsx_vshuf_b, src1, src0, mask, src3, src2, mask, src5, src4, mask, + src7, src6, mask, vec0, vec1, vec2, vec3); + DUP4_ARG2(__lsx_vdp2_h_bu, vec0, filt0, vec1, filt0, vec2, filt0, vec3, filt0, + vec4, vec5, vec6, vec7); + DUP4_ARG2(__lsx_vsrari_h, vec4, FILTER_BITS, vec5, FILTER_BITS, vec6, + FILTER_BITS, vec7, FILTER_BITS, vec4, vec5, vec6, vec7); + DUP4_ARG2(__lsx_vpickev_b, vec4, vec4, vec5, vec5, vec6, vec6, vec7, vec7, + res0, res1, res2, res3); + __lsx_vstelm_w(res0, dst, 0, 0); + dst += dst_stride; + __lsx_vstelm_w(res0, dst, 0, 1); + dst += dst_stride; + __lsx_vstelm_w(res1, dst, 0, 0); + dst += dst_stride; + __lsx_vstelm_w(res1, dst, 0, 1); + dst += dst_stride; + + __lsx_vstelm_w(res2, dst, 0, 0); + __lsx_vstelm_w(res2, dst + dst_stride, 0, 1); + __lsx_vstelm_w(res3, dst + dst_stride2, 0, 0); + __lsx_vstelm_w(res3, dst + dst_stride3, 0, 1); +} + +static void common_hz_2t_4w_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + if (height == 4) { + common_hz_2t_4x4_lsx(src, src_stride, dst, dst_stride, filter); + } else if (height == 8) { + common_hz_2t_4x8_lsx(src, src_stride, dst, dst_stride, filter); + } +} + +static void common_hz_2t_8x4_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter) { + __m128i filt0, mask; + __m128i src0, src1, src2, src3; + __m128i vec0, vec1, vec2, vec3; + + int32_t src_stride2 = src_stride << 1; + int32_t src_stride3 = src_stride2 + src_stride; + + int32_t dst_stride2 = dst_stride << 1; + int32_t dst_stride3 = dst_stride2 + dst_stride; + + mask = __lsx_vld(mc_filt_mask_arr, 0); + + /* rearranging filter */ + filt0 = __lsx_vldrepl_h(filter, 0); + + src0 = __lsx_vld(src, 0); + DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride2, src1, src2); + src3 = __lsx_vldx(src, src_stride3); + + DUP4_ARG3(__lsx_vshuf_b, src0, src0, mask, src1, src1, mask, src2, src2, mask, + src3, src3, mask, vec0, vec1, vec2, vec3); + DUP4_ARG2(__lsx_vdp2_h_bu, vec0, filt0, vec1, filt0, vec2, filt0, vec3, filt0, + vec0, vec1, vec2, vec3); + DUP4_ARG2(__lsx_vsrari_h, vec0, FILTER_BITS, vec1, FILTER_BITS, vec2, + FILTER_BITS, vec3, FILTER_BITS, vec0, vec1, vec2, vec3); + DUP2_ARG2(__lsx_vpickev_b, vec1, vec0, vec3, vec2, src0, src1); + + __lsx_vstelm_d(src0, dst, 0, 0); + __lsx_vstelm_d(src0, dst + dst_stride, 0, 1); + __lsx_vstelm_d(src1, dst + dst_stride2, 0, 0); + __lsx_vstelm_d(src1, dst + dst_stride3, 0, 1); +} + +static void common_hz_2t_8x8mult_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + __m128i filt0, mask; + __m128i src0, src1, src2, src3, out0, out1; + __m128i vec0, vec1, vec2, vec3; + + int32_t src_stride2 = src_stride << 1; + int32_t src_stride3 = src_stride2 + src_stride; + int32_t src_stride4 = src_stride2 << 1; + + int32_t dst_stride2 = dst_stride << 1; + int32_t dst_stride3 = dst_stride2 + dst_stride; + int32_t dst_stride4 = dst_stride2 << 1; + + mask = __lsx_vld(mc_filt_mask_arr, 0); + + /* rearranging filter */ + filt0 = __lsx_vldrepl_h(filter, 0); + + src0 = __lsx_vld(src, 0); + DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride2, src1, src2); + src3 = __lsx_vldx(src, src_stride3); + src += src_stride4; + + DUP4_ARG3(__lsx_vshuf_b, src0, src0, mask, src1, src1, mask, src2, src2, mask, + src3, src3, mask, vec0, vec1, vec2, vec3); + DUP4_ARG2(__lsx_vdp2_h_bu, vec0, filt0, vec1, filt0, vec2, filt0, vec3, filt0, + vec0, vec1, vec2, vec3); + DUP4_ARG2(__lsx_vsrari_h, vec0, FILTER_BITS, vec1, FILTER_BITS, vec2, + FILTER_BITS, vec3, FILTER_BITS, vec0, vec1, vec2, vec3); + + src0 = __lsx_vld(src, 0); + DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride2, src1, src2); + src3 = __lsx_vldx(src, src_stride3); + src += src_stride4; + + DUP2_ARG2(__lsx_vpickev_b, vec1, vec0, vec3, vec2, out0, out1); + __lsx_vstelm_d(out0, dst, 0, 0); + dst += dst_stride; + __lsx_vstelm_d(out0, dst, 0, 1); + dst += dst_stride; + __lsx_vstelm_d(out1, dst, 0, 0); + dst += dst_stride; + __lsx_vstelm_d(out1, dst, 0, 1); + dst += dst_stride; + + DUP4_ARG3(__lsx_vshuf_b, src0, src0, mask, src1, src1, mask, src2, src2, mask, + src3, src3, mask, vec0, vec1, vec2, vec3); + DUP4_ARG2(__lsx_vdp2_h_bu, vec0, filt0, vec1, filt0, vec2, filt0, vec3, filt0, + vec0, vec1, vec2, vec3); + DUP4_ARG2(__lsx_vsrari_h, vec0, FILTER_BITS, vec1, FILTER_BITS, vec2, + FILTER_BITS, vec3, FILTER_BITS, vec0, vec1, vec2, vec3); + DUP2_ARG2(__lsx_vpickev_b, vec1, vec0, vec3, vec2, out0, out1); + + __lsx_vstelm_d(out0, dst, 0, 0); + dst += dst_stride; + __lsx_vstelm_d(out0, dst, 0, 1); + dst += dst_stride; + __lsx_vstelm_d(out1, dst, 0, 0); + dst += dst_stride; + __lsx_vstelm_d(out1, dst, 0, 1); + dst += dst_stride; + + if (height == 16) { + uint8_t *dst_tmp1 = dst + dst_stride4; + + src0 = __lsx_vld(src, 0); + DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride2, src1, src2); + src3 = __lsx_vldx(src, src_stride3); + src += src_stride4; + + DUP4_ARG3(__lsx_vshuf_b, src0, src0, mask, src1, src1, mask, src2, src2, + mask, src3, src3, mask, vec0, vec1, vec2, vec3); + DUP4_ARG2(__lsx_vdp2_h_bu, vec0, filt0, vec1, filt0, vec2, filt0, vec3, + filt0, vec0, vec1, vec2, vec3); + DUP4_ARG2(__lsx_vsrari_h, vec0, FILTER_BITS, vec1, FILTER_BITS, vec2, + FILTER_BITS, vec3, FILTER_BITS, vec0, vec1, vec2, vec3); + + src0 = __lsx_vld(src, 0); + DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride2, src1, src2); + src3 = __lsx_vldx(src, src_stride3); + src += src_stride4; + + DUP2_ARG2(__lsx_vpickev_b, vec1, vec0, vec3, vec2, out0, out1); + __lsx_vstelm_d(out0, dst, 0, 0); + __lsx_vstelm_d(out0, dst + dst_stride, 0, 1); + __lsx_vstelm_d(out1, dst + dst_stride2, 0, 0); + __lsx_vstelm_d(out1, dst + dst_stride3, 0, 1); + + DUP4_ARG3(__lsx_vshuf_b, src0, src0, mask, src1, src1, mask, src2, src2, + mask, src3, src3, mask, vec0, vec1, vec2, vec3); + DUP4_ARG2(__lsx_vdp2_h_bu, vec0, filt0, vec1, filt0, vec2, filt0, vec3, + filt0, vec0, vec1, vec2, vec3); + DUP4_ARG2(__lsx_vsrari_h, vec0, FILTER_BITS, vec1, FILTER_BITS, vec2, + FILTER_BITS, vec3, FILTER_BITS, vec0, vec1, vec2, vec3); + DUP2_ARG2(__lsx_vpickev_b, vec1, vec0, vec3, vec2, out0, out1); + + __lsx_vstelm_d(out0, dst_tmp1, 0, 0); + __lsx_vstelm_d(out0, dst_tmp1 + dst_stride, 0, 1); + __lsx_vstelm_d(out1, dst_tmp1 + dst_stride2, 0, 0); + __lsx_vstelm_d(out1, dst_tmp1 + dst_stride3, 0, 1); + } +} + +static void common_hz_2t_8w_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + if (height == 4) { + common_hz_2t_8x4_lsx(src, src_stride, dst, dst_stride, filter); + } else { + common_hz_2t_8x8mult_lsx(src, src_stride, dst, dst_stride, filter, height); + } +} + +static void common_hz_2t_16w_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + uint32_t loop_cnt = (height >> 2) - 1; + __m128i src0, src1, src2, src3, src4, src5, src6, src7, mask; + __m128i filt0, vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7; + __m128i out0, out1, out2, out3, out4, out5, out6, out7, tmp; + + int32_t src_stride2 = src_stride << 1; + int32_t src_stride3 = src_stride2 + src_stride; + int32_t src_stride4 = src_stride2 << 1; + + uint8_t *src_tmp1 = src + 8; + mask = __lsx_vld(mc_filt_mask_arr, 0); + filt0 = __lsx_vldrepl_h(filter, 0); + + src0 = __lsx_vld(src, 0); + DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride2, src2, src4); + src6 = __lsx_vldx(src, src_stride3); + src1 = __lsx_vld(src_tmp1, 0); + DUP2_ARG2(__lsx_vldx, src_tmp1, src_stride, src_tmp1, src_stride2, src3, + src5); + src7 = __lsx_vldx(src_tmp1, src_stride3); + src += src_stride4; + + DUP4_ARG3(__lsx_vshuf_b, src0, src0, mask, src1, src1, mask, src2, src2, mask, + src3, src3, mask, vec0, vec1, vec2, vec3); + DUP4_ARG3(__lsx_vshuf_b, src4, src4, mask, src5, src5, mask, src6, src6, mask, + src7, src7, mask, vec4, vec5, vec6, vec7); + DUP4_ARG2(__lsx_vdp2_h_bu, vec0, filt0, vec1, filt0, vec2, filt0, vec3, filt0, + out0, out1, out2, out3); + DUP4_ARG2(__lsx_vdp2_h_bu, vec4, filt0, vec5, filt0, vec6, filt0, vec7, filt0, + out4, out5, out6, out7); + DUP4_ARG2(__lsx_vsrari_h, out0, FILTER_BITS, out1, FILTER_BITS, out2, + FILTER_BITS, out3, FILTER_BITS, out0, out1, out2, out3); + DUP4_ARG2(__lsx_vsrari_h, out4, FILTER_BITS, out5, FILTER_BITS, out6, + FILTER_BITS, out7, FILTER_BITS, out4, out5, out6, out7); + + tmp = __lsx_vpickev_b(out1, out0); + __lsx_vst(tmp, dst, 0); + dst += dst_stride; + tmp = __lsx_vpickev_b(out3, out2); + __lsx_vst(tmp, dst, 0); + dst += dst_stride; + tmp = __lsx_vpickev_b(out5, out4); + __lsx_vst(tmp, dst, 0); + dst += dst_stride; + tmp = __lsx_vpickev_b(out7, out6); + __lsx_vst(tmp, dst, 0); + dst += dst_stride; + + for (; loop_cnt--;) { + src_tmp1 += src_stride4; + + src0 = __lsx_vld(src, 0); + DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride2, src2, src4); + src6 = __lsx_vldx(src, src_stride3); + + src1 = __lsx_vld(src_tmp1, 0); + DUP2_ARG2(__lsx_vldx, src_tmp1, src_stride, src_tmp1, src_stride2, src3, + src5); + src7 = __lsx_vldx(src_tmp1, src_stride3); + src += src_stride4; + + DUP4_ARG3(__lsx_vshuf_b, src0, src0, mask, src1, src1, mask, src2, src2, + mask, src3, src3, mask, vec0, vec1, vec2, vec3); + DUP4_ARG3(__lsx_vshuf_b, src4, src4, mask, src5, src5, mask, src6, src6, + mask, src7, src7, mask, vec4, vec5, vec6, vec7); + DUP4_ARG2(__lsx_vdp2_h_bu, vec0, filt0, vec1, filt0, vec2, filt0, vec3, + filt0, out0, out1, out2, out3); + DUP4_ARG2(__lsx_vdp2_h_bu, vec4, filt0, vec5, filt0, vec6, filt0, vec7, + filt0, out4, out5, out6, out7); + DUP4_ARG2(__lsx_vsrari_h, out0, FILTER_BITS, out1, FILTER_BITS, out2, + FILTER_BITS, out3, FILTER_BITS, out0, out1, out2, out3); + DUP4_ARG2(__lsx_vsrari_h, out4, FILTER_BITS, out5, FILTER_BITS, out6, + FILTER_BITS, out7, FILTER_BITS, out4, out5, out6, out7); + + tmp = __lsx_vpickev_b(out1, out0); + __lsx_vst(tmp, dst, 0); + dst += dst_stride; + tmp = __lsx_vpickev_b(out3, out2); + __lsx_vst(tmp, dst, 0); + dst += dst_stride; + tmp = __lsx_vpickev_b(out5, out4); + __lsx_vst(tmp, dst, 0); + dst += dst_stride; + tmp = __lsx_vpickev_b(out7, out6); + __lsx_vst(tmp, dst, 0); + dst += dst_stride; + } +} + +static void common_hz_2t_32w_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + uint32_t loop_cnt = (height >> 1); + __m128i src0, src1, src2, src3, src4, src5, src6, src7, mask; + __m128i filt0, vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7; + __m128i out0, out1, out2, out3, out4, out5, out6, out7, tmp; + __m128i shuff = { 0x0F0E0D0C0B0A0908, 0x1716151413121110 }; + + mask = __lsx_vld(mc_filt_mask_arr, 0); + /* rearranging filter */ + filt0 = __lsx_vldrepl_h(filter, 0); + + for (; loop_cnt--;) { + DUP2_ARG2(__lsx_vld, src, 0, src, 16, src0, src2); + src3 = __lsx_vld(src, 24); + src1 = __lsx_vshuf_b(src2, src0, shuff); + src += src_stride; + DUP2_ARG2(__lsx_vld, src, 0, src, 16, src4, src6); + src7 = __lsx_vld(src, 24); + src5 = __lsx_vshuf_b(src6, src4, shuff); + src += src_stride; + + DUP4_ARG3(__lsx_vshuf_b, src0, src0, mask, src1, src1, mask, src2, src2, + mask, src3, src3, mask, vec0, vec1, vec2, vec3); + DUP4_ARG3(__lsx_vshuf_b, src4, src4, mask, src5, src5, mask, src6, src6, + mask, src7, src7, mask, vec4, vec5, vec6, vec7); + DUP4_ARG2(__lsx_vdp2_h_bu, vec0, filt0, vec1, filt0, vec2, filt0, vec3, + filt0, out0, out1, out2, out3); + DUP4_ARG2(__lsx_vdp2_h_bu, vec4, filt0, vec5, filt0, vec6, filt0, vec7, + filt0, out4, out5, out6, out7); + DUP4_ARG2(__lsx_vsrari_h, out0, FILTER_BITS, out1, FILTER_BITS, out2, + FILTER_BITS, out3, FILTER_BITS, out0, out1, out2, out3); + DUP4_ARG2(__lsx_vsrari_h, out4, FILTER_BITS, out5, FILTER_BITS, out6, + FILTER_BITS, out7, FILTER_BITS, out4, out5, out6, out7); + + tmp = __lsx_vpickev_b(out1, out0); + __lsx_vst(tmp, dst, 0); + tmp = __lsx_vpickev_b(out3, out2); + __lsx_vst(tmp, dst, 16); + dst += dst_stride; + + tmp = __lsx_vpickev_b(out5, out4); + __lsx_vst(tmp, dst, 0); + tmp = __lsx_vpickev_b(out7, out6); + __lsx_vst(tmp, dst, 16); + dst += dst_stride; + } +} + +static void common_hz_2t_64w_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + uint32_t loop_cnt = height; + __m128i src0, src1, src2, src3, src4, src5, src6, src7, mask; + __m128i filt0, vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7; + __m128i out0, out1, out2, out3, out4, out5, out6, out7, tmp; + __m128i shuff = { 0x0F0E0D0C0B0A0908, 0x1716151413121110 }; + + mask = __lsx_vld(mc_filt_mask_arr, 0); + + /* rearranging filter */ + filt0 = __lsx_vldrepl_h(filter, 0); + + for (; loop_cnt--;) { + DUP4_ARG2(__lsx_vld, src, 0, src, 16, src, 32, src, 48, src0, src2, src4, + src6); + src7 = __lsx_vld(src, 56); + DUP2_ARG3(__lsx_vshuf_b, src2, src0, shuff, src4, src2, shuff, src1, src3); + src5 = __lsx_vshuf_b(src6, src4, shuff); + src += src_stride; + + DUP4_ARG3(__lsx_vshuf_b, src0, src0, mask, src1, src1, mask, src2, src2, + mask, src3, src3, mask, vec0, vec1, vec2, vec3); + DUP4_ARG3(__lsx_vshuf_b, src4, src4, mask, src5, src5, mask, src6, src6, + mask, src7, src7, mask, vec4, vec5, vec6, vec7); + + DUP4_ARG2(__lsx_vdp2_h_bu, vec0, filt0, vec1, filt0, vec2, filt0, vec3, + filt0, out0, out1, out2, out3); + DUP4_ARG2(__lsx_vdp2_h_bu, vec4, filt0, vec5, filt0, vec6, filt0, vec7, + filt0, out4, out5, out6, out7); + DUP4_ARG2(__lsx_vsrari_h, out0, FILTER_BITS, out1, FILTER_BITS, out2, + FILTER_BITS, out3, FILTER_BITS, out0, out1, out2, out3); + DUP4_ARG2(__lsx_vsrari_h, out4, FILTER_BITS, out5, FILTER_BITS, out6, + FILTER_BITS, out7, FILTER_BITS, out4, out5, out6, out7); + + tmp = __lsx_vpickev_b(out1, out0); + __lsx_vst(tmp, dst, 0); + tmp = __lsx_vpickev_b(out3, out2); + __lsx_vst(tmp, dst, 16); + tmp = __lsx_vpickev_b(out5, out4); + __lsx_vst(tmp, dst, 32); + tmp = __lsx_vpickev_b(out7, out6); + __lsx_vst(tmp, dst, 48); + dst += dst_stride; + } +} + +void vpx_convolve8_horiz_lsx(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const InterpKernel *filter, int x0_q4, + int x_step_q4, int y0_q4, int y_step_q4, int w, + int h) { + const int16_t *const filter_x = filter[x0_q4]; + int8_t cnt, filt_hor[8]; + + assert(x_step_q4 == 16); + assert(((const int32_t *)filter_x)[1] != 0x800000); + + for (cnt = 0; cnt < 8; ++cnt) { + filt_hor[cnt] = filter_x[cnt]; + } + if (vpx_get_filter_taps(filter_x) == 2) { + switch (w) { + case 4: + common_hz_2t_4w_lsx(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + &filt_hor[3], h); + break; + case 8: + common_hz_2t_8w_lsx(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + &filt_hor[3], h); + break; + case 16: + common_hz_2t_16w_lsx(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + &filt_hor[3], h); + break; + case 32: + common_hz_2t_32w_lsx(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + &filt_hor[3], h); + break; + case 64: + common_hz_2t_64w_lsx(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + &filt_hor[3], h); + break; + default: + vpx_convolve8_horiz_c(src, src_stride, dst, dst_stride, filter, x0_q4, + x_step_q4, y0_q4, y_step_q4, w, h); + break; + } + } else { + switch (w) { + case 4: + common_hz_8t_4w_lsx(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + filt_hor, h); + break; + case 8: + common_hz_8t_8w_lsx(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + filt_hor, h); + break; + + case 16: + common_hz_8t_16w_lsx(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + filt_hor, h); + break; + + case 32: + common_hz_8t_32w_lsx(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + filt_hor, h); + break; + + case 64: + common_hz_8t_64w_lsx(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + filt_hor, h); + break; + default: + vpx_convolve8_horiz_c(src, src_stride, dst, dst_stride, filter, x0_q4, + x_step_q4, y0_q4, y_step_q4, w, h); + break; + } + } +} diff --git a/vpx_dsp/loongarch/vpx_convolve8_lsx.c b/vpx_dsp/loongarch/vpx_convolve8_lsx.c new file mode 100644 index 0000000..51a162b --- /dev/null +++ b/vpx_dsp/loongarch/vpx_convolve8_lsx.c @@ -0,0 +1,718 @@ +/* + * 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 +#include "./vpx_dsp_rtcd.h" +#include "vpx_dsp/loongarch/vpx_convolve_lsx.h" + +#define LSX_LD_4(_src, _stride, _src0, _src1, _src2, _src3) \ + { \ + _src0 = __lsx_vld(_src, 0); \ + _src += _stride; \ + _src1 = __lsx_vld(_src, 0); \ + _src += _stride; \ + _src2 = __lsx_vld(_src, 0); \ + _src += _stride; \ + _src3 = __lsx_vld(_src, 0); \ + } + +static const uint8_t mc_filt_mask_arr[16 * 3] = { + /* 8 width cases */ + 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, + /* 4 width cases */ + 0, 1, 1, 2, 2, 3, 3, 4, 16, 17, 17, 18, 18, 19, 19, 20, + /* 4 width cases */ + 8, 9, 9, 10, 10, 11, 11, 12, 24, 25, 25, 26, 26, 27, 27, 28 +}; + +static void common_hv_8ht_8vt_4w_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter_horiz, int8_t *filter_vert, + int32_t height) { + uint32_t loop_cnt = (height >> 2); + __m128i src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10; + __m128i filt_hz0, filt_hz1, filt_hz2, filt_hz3; + __m128i filt_vt0, filt_vt1, filt_vt2, filt_vt3; + __m128i mask0, mask1, mask2, mask3; + __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5; + __m128i out0, out1; + __m128i shuff = { 0x0F0E0D0C0B0A0908, 0x1716151413121110 }; + + mask0 = __lsx_vld(mc_filt_mask_arr, 16); + src -= (3 + 3 * src_stride); + DUP4_ARG2(__lsx_vldrepl_h, filter_horiz, 0, filter_horiz, 2, filter_horiz, 4, + filter_horiz, 6, filt_hz0, filt_hz1, filt_hz2, filt_hz3); + DUP2_ARG2(__lsx_vaddi_bu, mask0, 2, mask0, 4, mask1, mask2); + mask3 = __lsx_vaddi_bu(mask0, 6); + + LSX_LD_4(src, src_stride, src0, src1, src2, src3); + src += src_stride; + src4 = __lsx_vld(src, 0); + src += src_stride; + src5 = __lsx_vld(src, 0); + src += src_stride; + src6 = __lsx_vld(src, 0); + src += src_stride; + DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128, src0, + src1, src2, src3); + DUP2_ARG2(__lsx_vxori_b, src4, 128, src5, 128, src4, src5); + src6 = __lsx_vxori_b(src6, 128); + + tmp0 = HORIZ_8TAP_FILT(src0, src1, mask0, mask1, mask2, mask3, filt_hz0, + filt_hz1, filt_hz2, filt_hz3); + tmp2 = HORIZ_8TAP_FILT(src2, src3, mask0, mask1, mask2, mask3, filt_hz0, + filt_hz1, filt_hz2, filt_hz3); + tmp4 = HORIZ_8TAP_FILT(src4, src5, mask0, mask1, mask2, mask3, filt_hz0, + filt_hz1, filt_hz2, filt_hz3); + tmp5 = HORIZ_8TAP_FILT(src5, src6, mask0, mask1, mask2, mask3, filt_hz0, + filt_hz1, filt_hz2, filt_hz3); + DUP2_ARG3(__lsx_vshuf_b, tmp2, tmp0, shuff, tmp4, tmp2, shuff, tmp1, tmp3); + DUP4_ARG2(__lsx_vldrepl_h, filter_vert, 0, filter_vert, 2, filter_vert, 4, + filter_vert, 6, filt_vt0, filt_vt1, filt_vt2, filt_vt3); + DUP2_ARG2(__lsx_vpackev_b, tmp1, tmp0, tmp3, tmp2, tmp0, tmp1); + tmp2 = __lsx_vpackev_b(tmp5, tmp4); + + for (; loop_cnt--;) { + LSX_LD_4(src, src_stride, src7, src8, src9, src10); + src += src_stride; + DUP4_ARG2(__lsx_vxori_b, src7, 128, src8, 128, src9, 128, src10, 128, src7, + src8, src9, src10); + tmp3 = HORIZ_8TAP_FILT(src7, src8, mask0, mask1, mask2, mask3, filt_hz0, + filt_hz1, filt_hz2, filt_hz3); + tmp4 = __lsx_vshuf_b(tmp3, tmp5, shuff); + tmp4 = __lsx_vpackev_b(tmp3, tmp4); + out0 = FILT_8TAP_DPADD_S_H(tmp0, tmp1, tmp2, tmp4, filt_vt0, filt_vt1, + filt_vt2, filt_vt3); + src1 = HORIZ_8TAP_FILT(src9, src10, mask0, mask1, mask2, mask3, filt_hz0, + filt_hz1, filt_hz2, filt_hz3); + src0 = __lsx_vshuf_b(src1, tmp3, shuff); + src0 = __lsx_vpackev_b(src1, src0); + out1 = FILT_8TAP_DPADD_S_H(tmp1, tmp2, tmp4, src0, filt_vt0, filt_vt1, + filt_vt2, filt_vt3); + out0 = __lsx_vssrarni_b_h(out1, out0, 7); + out0 = __lsx_vxori_b(out0, 128); + __lsx_vstelm_w(out0, dst, 0, 0); + dst += dst_stride; + __lsx_vstelm_w(out0, dst, 0, 1); + dst += dst_stride; + __lsx_vstelm_w(out0, dst, 0, 2); + dst += dst_stride; + __lsx_vstelm_w(out0, dst, 0, 3); + dst += dst_stride; + + tmp5 = src1; + tmp0 = tmp2; + tmp1 = tmp4; + tmp2 = src0; + } +} + +static void common_hv_8ht_8vt_8w_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter_horiz, int8_t *filter_vert, + int32_t height) { + uint32_t loop_cnt = (height >> 2); + __m128i src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10; + __m128i filt_hz0, filt_hz1, filt_hz2, filt_hz3; + __m128i filt_vt0, filt_vt1, filt_vt2, filt_vt3; + __m128i mask0, mask1, mask2, mask3; + __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6; + __m128i out0, out1; + + mask0 = __lsx_vld(mc_filt_mask_arr, 0); + src -= (3 + 3 * src_stride); + DUP4_ARG2(__lsx_vldrepl_h, filter_horiz, 0, filter_horiz, 2, filter_horiz, 4, + filter_horiz, 6, filt_hz0, filt_hz1, filt_hz2, filt_hz3); + DUP2_ARG2(__lsx_vaddi_bu, mask0, 2, mask0, 4, mask1, mask2); + mask3 = __lsx_vaddi_bu(mask0, 6); + + LSX_LD_4(src, src_stride, src0, src1, src2, src3); + src += src_stride; + src4 = __lsx_vld(src, 0); + src += src_stride; + src5 = __lsx_vld(src, 0); + src += src_stride; + src6 = __lsx_vld(src, 0); + src += src_stride; + DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128, src0, + src1, src2, src3); + DUP2_ARG2(__lsx_vxori_b, src4, 128, src5, 128, src4, src5); + src6 = __lsx_vxori_b(src6, 128); + + src0 = HORIZ_8TAP_FILT(src0, src0, mask0, mask1, mask2, mask3, filt_hz0, + filt_hz1, filt_hz2, filt_hz3); + src1 = HORIZ_8TAP_FILT(src1, src1, mask0, mask1, mask2, mask3, filt_hz0, + filt_hz1, filt_hz2, filt_hz3); + src2 = HORIZ_8TAP_FILT(src2, src2, mask0, mask1, mask2, mask3, filt_hz0, + filt_hz1, filt_hz2, filt_hz3); + src3 = HORIZ_8TAP_FILT(src3, src3, mask0, mask1, mask2, mask3, filt_hz0, + filt_hz1, filt_hz2, filt_hz3); + src4 = HORIZ_8TAP_FILT(src4, src4, mask0, mask1, mask2, mask3, filt_hz0, + filt_hz1, filt_hz2, filt_hz3); + src5 = HORIZ_8TAP_FILT(src5, src5, mask0, mask1, mask2, mask3, filt_hz0, + filt_hz1, filt_hz2, filt_hz3); + src6 = HORIZ_8TAP_FILT(src6, src6, mask0, mask1, mask2, mask3, filt_hz0, + filt_hz1, filt_hz2, filt_hz3); + + DUP4_ARG2(__lsx_vldrepl_h, filter_vert, 0, filter_vert, 2, filter_vert, 4, + filter_vert, 6, filt_vt0, filt_vt1, filt_vt2, filt_vt3); + DUP4_ARG2(__lsx_vpackev_b, src1, src0, src3, src2, src5, src4, src2, src1, + tmp0, tmp1, tmp2, tmp4); + DUP2_ARG2(__lsx_vpackev_b, src4, src3, src6, src5, tmp5, tmp6); + + for (; loop_cnt--;) { + LSX_LD_4(src, src_stride, src7, src8, src9, src10); + src += src_stride; + DUP4_ARG2(__lsx_vxori_b, src7, 128, src8, 128, src9, 128, src10, 128, src7, + src8, src9, src10); + src7 = HORIZ_8TAP_FILT(src7, src7, mask0, mask1, mask2, mask3, filt_hz0, + filt_hz1, filt_hz2, filt_hz3); + tmp3 = __lsx_vpackev_b(src7, src6); + out0 = FILT_8TAP_DPADD_S_H(tmp0, tmp1, tmp2, tmp3, filt_vt0, filt_vt1, + filt_vt2, filt_vt3); + src8 = HORIZ_8TAP_FILT(src8, src8, mask0, mask1, mask2, mask3, filt_hz0, + filt_hz1, filt_hz2, filt_hz3); + src0 = __lsx_vpackev_b(src8, src7); + out1 = FILT_8TAP_DPADD_S_H(tmp4, tmp5, tmp6, src0, filt_vt0, filt_vt1, + filt_vt2, filt_vt3); + src9 = HORIZ_8TAP_FILT(src9, src9, mask0, mask1, mask2, mask3, filt_hz0, + filt_hz1, filt_hz2, filt_hz3); + src1 = __lsx_vpackev_b(src9, src8); + src3 = FILT_8TAP_DPADD_S_H(tmp1, tmp2, tmp3, src1, filt_vt0, filt_vt1, + filt_vt2, filt_vt3); + src10 = HORIZ_8TAP_FILT(src10, src10, mask0, mask1, mask2, mask3, filt_hz0, + filt_hz1, filt_hz2, filt_hz3); + src2 = __lsx_vpackev_b(src10, src9); + src4 = FILT_8TAP_DPADD_S_H(tmp5, tmp6, src0, src2, filt_vt0, filt_vt1, + filt_vt2, filt_vt3); + DUP2_ARG3(__lsx_vssrarni_b_h, out1, out0, 7, src4, src3, 7, out0, out1); + DUP2_ARG2(__lsx_vxori_b, out0, 128, out1, 128, out0, out1); + __lsx_vstelm_d(out0, dst, 0, 0); + dst += dst_stride; + __lsx_vstelm_d(out0, dst, 0, 1); + dst += dst_stride; + __lsx_vstelm_d(out1, dst, 0, 0); + dst += dst_stride; + __lsx_vstelm_d(out1, dst, 0, 1); + dst += dst_stride; + + src6 = src10; + tmp0 = tmp2; + tmp1 = tmp3; + tmp2 = src1; + tmp4 = tmp6; + tmp5 = src0; + tmp6 = src2; + } +} + +static void common_hv_8ht_8vt_16w_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter_horiz, int8_t *filter_vert, + int32_t height) { + common_hv_8ht_8vt_8w_lsx(src, src_stride, dst, dst_stride, filter_horiz, + filter_vert, height); + src += 8; + dst += 8; + + common_hv_8ht_8vt_8w_lsx(src, src_stride, dst, dst_stride, filter_horiz, + filter_vert, height); + src += 8; + dst += 8; +} + +static void common_hv_8ht_8vt_32w_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter_horiz, int8_t *filter_vert, + int32_t height) { + int32_t multiple8_cnt; + for (multiple8_cnt = 4; multiple8_cnt--;) { + common_hv_8ht_8vt_8w_lsx(src, src_stride, dst, dst_stride, filter_horiz, + filter_vert, height); + src += 8; + dst += 8; + } +} + +static void common_hv_8ht_8vt_64w_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter_horiz, int8_t *filter_vert, + int32_t height) { + int32_t multiple8_cnt; + for (multiple8_cnt = 8; multiple8_cnt--;) { + common_hv_8ht_8vt_8w_lsx(src, src_stride, dst, dst_stride, filter_horiz, + filter_vert, height); + src += 8; + dst += 8; + } +} + +static void common_hv_2ht_2vt_4x4_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter_horiz, + int8_t *filter_vert) { + __m128i src0, src1, src2, src3, src4, mask; + __m128i filt_vt, filt_hz, vec0, vec1, res0, res1; + __m128i hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, tmp0, tmp1; + __m128i shuff = { 0x0F0E0D0C0B0A0908, 0x1716151413121110 }; + + int32_t src_stride2 = src_stride << 1; + int32_t src_stride3 = src_stride + src_stride2; + int32_t src_stride4 = src_stride2 << 1; + + int32_t dst_stride2 = dst_stride << 1; + int32_t dst_stride3 = dst_stride2 + dst_stride; + mask = __lsx_vld(mc_filt_mask_arr, 16); + + /* rearranging filter */ + filt_hz = __lsx_vldrepl_h(filter_horiz, 0); + filt_vt = __lsx_vldrepl_h(filter_vert, 0); + + src0 = __lsx_vld(src, 0); + DUP4_ARG2(__lsx_vldx, src, src_stride, src, src_stride2, src, src_stride3, + src, src_stride4, src1, src2, src3, src4); + hz_out0 = HORIZ_2TAP_FILT_UH(src0, src1, mask, filt_hz, FILTER_BITS); + hz_out2 = HORIZ_2TAP_FILT_UH(src2, src3, mask, filt_hz, FILTER_BITS); + hz_out4 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS); + + hz_out1 = __lsx_vshuf_b(hz_out2, hz_out0, shuff); + hz_out3 = __lsx_vpickod_d(hz_out4, hz_out2); + + DUP2_ARG2(__lsx_vpackev_b, hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1); + DUP2_ARG2(__lsx_vdp2_h_bu, vec0, filt_vt, vec1, filt_vt, tmp0, tmp1); + DUP2_ARG2(__lsx_vsrari_h, tmp0, FILTER_BITS, tmp1, FILTER_BITS, tmp0, tmp1); + DUP2_ARG2(__lsx_vpickev_b, tmp0, tmp0, tmp1, tmp1, res0, res1); + + __lsx_vstelm_w(res0, dst, 0, 0); + __lsx_vstelm_w(res0, dst + dst_stride, 0, 1); + __lsx_vstelm_w(res1, dst + dst_stride2, 0, 0); + __lsx_vstelm_w(res1, dst + dst_stride3, 0, 1); +} + +static void common_hv_2ht_2vt_4x8_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter_horiz, + int8_t *filter_vert) { + __m128i src0, src1, src2, src3, src4, src5, src6, src7, src8, mask; + __m128i res0, res1, res2, res3; + __m128i filt_hz, filt_vt, vec0, vec1, vec2, vec3; + __m128i hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, hz_out5, hz_out6; + __m128i hz_out7, hz_out8, vec4, vec5, vec6, vec7; + __m128i shuff = { 0x0F0E0D0C0B0A0908, 0x1716151413121110 }; + + int32_t src_stride2 = src_stride << 1; + int32_t src_stride3 = src_stride2 + src_stride; + int32_t src_stride4 = src_stride2 << 1; + + int32_t dst_stride2 = dst_stride << 1; + int32_t dst_stride3 = dst_stride2 + dst_stride; + int32_t dst_stride4 = dst_stride2 << 1; + + mask = __lsx_vld(mc_filt_mask_arr, 16); + + /* rearranging filter */ + DUP2_ARG2(__lsx_vldrepl_h, filter_horiz, 0, filter_vert, 0, filt_hz, filt_vt); + + src0 = __lsx_vld(src, 0); + DUP4_ARG2(__lsx_vldx, src, src_stride, src, src_stride2, src, src_stride3, + src, src_stride4, src1, src2, src3, src4); + src += src_stride4; + DUP4_ARG2(__lsx_vldx, src, src_stride, src, src_stride2, src, src_stride3, + src, src_stride4, src5, src6, src7, src8); + src += src_stride4; + + hz_out0 = HORIZ_2TAP_FILT_UH(src0, src1, mask, filt_hz, FILTER_BITS); + hz_out2 = HORIZ_2TAP_FILT_UH(src2, src3, mask, filt_hz, FILTER_BITS); + hz_out4 = HORIZ_2TAP_FILT_UH(src4, src5, mask, filt_hz, FILTER_BITS); + hz_out6 = HORIZ_2TAP_FILT_UH(src6, src7, mask, filt_hz, FILTER_BITS); + hz_out8 = HORIZ_2TAP_FILT_UH(src8, src8, mask, filt_hz, FILTER_BITS); + + DUP2_ARG3(__lsx_vshuf_b, hz_out2, hz_out0, shuff, hz_out4, hz_out2, shuff, + hz_out1, hz_out3); + hz_out5 = __lsx_vshuf_b(hz_out6, hz_out4, shuff); + hz_out7 = __lsx_vpickod_d(hz_out8, hz_out6); + DUP4_ARG2(__lsx_vpackev_b, hz_out1, hz_out0, hz_out3, hz_out2, hz_out5, + hz_out4, hz_out7, hz_out6, vec0, vec1, vec2, vec3); + DUP4_ARG2(__lsx_vdp2_h_bu, vec0, filt_vt, vec1, filt_vt, vec2, filt_vt, vec3, + filt_vt, vec4, vec5, vec6, vec7); + DUP4_ARG2(__lsx_vsrari_h, vec4, FILTER_BITS, vec5, FILTER_BITS, vec6, + FILTER_BITS, vec7, FILTER_BITS, vec4, vec5, vec6, vec7); + DUP4_ARG2(__lsx_vpickev_b, vec4, vec4, vec5, vec5, vec6, vec6, vec7, vec7, + res0, res1, res2, res3); + + __lsx_vstelm_w(res0, dst, 0, 0); + __lsx_vstelm_w(res0, dst + dst_stride, 0, 1); + __lsx_vstelm_w(res1, dst + dst_stride2, 0, 0); + __lsx_vstelm_w(res1, dst + dst_stride3, 0, 1); + dst += dst_stride4; + __lsx_vstelm_w(res2, dst, 0, 0); + __lsx_vstelm_w(res2, dst + dst_stride, 0, 1); + __lsx_vstelm_w(res3, dst + dst_stride2, 0, 0); + __lsx_vstelm_w(res3, dst + dst_stride3, 0, 1); +} + +static void common_hv_2ht_2vt_4w_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter_horiz, int8_t *filter_vert, + int32_t height) { + if (height == 4) { + common_hv_2ht_2vt_4x4_lsx(src, src_stride, dst, dst_stride, filter_horiz, + filter_vert); + } else if (height == 8) { + common_hv_2ht_2vt_4x8_lsx(src, src_stride, dst, dst_stride, filter_horiz, + filter_vert); + } +} + +static void common_hv_2ht_2vt_8x4_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter_horiz, + int8_t *filter_vert) { + __m128i src0, src1, src2, src3, src4, mask, out0, out1; + __m128i filt_hz, filt_vt, vec0, vec1, vec2, vec3; + __m128i hz_out0, hz_out1, tmp0, tmp1, tmp2, tmp3; + + int32_t src_stride2 = src_stride << 1; + int32_t src_stride3 = src_stride2 + src_stride; + int32_t src_stride4 = src_stride2 << 1; + + int32_t dst_stride2 = dst_stride << 1; + int32_t dst_stride3 = dst_stride2 + dst_stride; + + mask = __lsx_vld(mc_filt_mask_arr, 0); + + /* rearranging filter */ + DUP2_ARG2(__lsx_vldrepl_h, filter_horiz, 0, filter_vert, 0, filt_hz, filt_vt); + + src0 = __lsx_vld(src, 0); + DUP4_ARG2(__lsx_vldx, src, src_stride, src, src_stride2, src, src_stride3, + src, src_stride4, src1, src2, src3, src4); + + hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS); + hz_out1 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS); + vec0 = __lsx_vpackev_b(hz_out1, hz_out0); + tmp0 = __lsx_vdp2_h_bu(vec0, filt_vt); + + hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS); + vec1 = __lsx_vpackev_b(hz_out0, hz_out1); + tmp1 = __lsx_vdp2_h_bu(vec1, filt_vt); + + hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS); + vec2 = __lsx_vpackev_b(hz_out1, hz_out0); + tmp2 = __lsx_vdp2_h_bu(vec2, filt_vt); + + hz_out0 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS); + vec3 = __lsx_vpackev_b(hz_out0, hz_out1); + tmp3 = __lsx_vdp2_h_bu(vec3, filt_vt); + + DUP4_ARG2(__lsx_vsrari_h, tmp0, FILTER_BITS, tmp1, FILTER_BITS, tmp2, + FILTER_BITS, tmp3, FILTER_BITS, tmp0, tmp1, tmp2, tmp3); + DUP2_ARG2(__lsx_vpickev_b, tmp1, tmp0, tmp3, tmp2, out0, out1); + + __lsx_vstelm_d(out0, dst, 0, 0); + __lsx_vstelm_d(out0, dst + dst_stride, 0, 1); + __lsx_vstelm_d(out1, dst + dst_stride2, 0, 0); + __lsx_vstelm_d(out1, dst + dst_stride3, 0, 1); +} + +static void common_hv_2ht_2vt_8x8mult_lsx(const uint8_t *src, + int32_t src_stride, uint8_t *dst, + int32_t dst_stride, + int8_t *filter_horiz, + int8_t *filter_vert, int32_t height) { + uint32_t loop_cnt = (height >> 3); + __m128i src0, src1, src2, src3, src4, mask, out0, out1; + __m128i filt_hz, filt_vt, vec0; + __m128i hz_out0, hz_out1, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8; + + int32_t src_stride2 = src_stride << 1; + int32_t src_stride3 = src_stride2 + src_stride; + int32_t src_stride4 = src_stride2 << 1; + + mask = __lsx_vld(mc_filt_mask_arr, 0); + + /* rearranging filter */ + DUP2_ARG2(__lsx_vldrepl_h, filter_horiz, 0, filter_vert, 0, filt_hz, filt_vt); + + src0 = __lsx_vld(src, 0); + src += src_stride; + + hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS); + + for (; loop_cnt--;) { + src1 = __lsx_vld(src, 0); + DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride2, src2, src3); + src4 = __lsx_vldx(src, src_stride3); + src += src_stride4; + + hz_out1 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS); + vec0 = __lsx_vpackev_b(hz_out1, hz_out0); + tmp1 = __lsx_vdp2_h_bu(vec0, filt_vt); + + hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS); + vec0 = __lsx_vpackev_b(hz_out0, hz_out1); + tmp2 = __lsx_vdp2_h_bu(vec0, filt_vt); + + DUP2_ARG2(__lsx_vsrari_h, tmp1, FILTER_BITS, tmp2, FILTER_BITS, tmp1, tmp2); + + hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS); + vec0 = __lsx_vpackev_b(hz_out1, hz_out0); + tmp3 = __lsx_vdp2_h_bu(vec0, filt_vt); + + hz_out0 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS); + src1 = __lsx_vld(src, 0); + DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride2, src2, src3); + src4 = __lsx_vldx(src, src_stride3); + src += src_stride4; + vec0 = __lsx_vpackev_b(hz_out0, hz_out1); + tmp4 = __lsx_vdp2_h_bu(vec0, filt_vt); + + DUP2_ARG2(__lsx_vsrari_h, tmp3, FILTER_BITS, tmp4, FILTER_BITS, tmp3, tmp4); + DUP2_ARG2(__lsx_vpickev_b, tmp2, tmp1, tmp4, tmp3, out0, out1); + __lsx_vstelm_d(out0, dst, 0, 0); + dst += dst_stride; + __lsx_vstelm_d(out0, dst, 0, 1); + dst += dst_stride; + __lsx_vstelm_d(out1, dst, 0, 0); + dst += dst_stride; + __lsx_vstelm_d(out1, dst, 0, 1); + dst += dst_stride; + + hz_out1 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS); + vec0 = __lsx_vpackev_b(hz_out1, hz_out0); + tmp5 = __lsx_vdp2_h_bu(vec0, filt_vt); + + hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS); + vec0 = __lsx_vpackev_b(hz_out0, hz_out1); + tmp6 = __lsx_vdp2_h_bu(vec0, filt_vt); + + hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS); + vec0 = __lsx_vpackev_b(hz_out1, hz_out0); + tmp7 = __lsx_vdp2_h_bu(vec0, filt_vt); + + hz_out0 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS); + vec0 = __lsx_vpackev_b(hz_out0, hz_out1); + tmp8 = __lsx_vdp2_h_bu(vec0, filt_vt); + + DUP4_ARG2(__lsx_vsrari_h, tmp5, FILTER_BITS, tmp6, FILTER_BITS, tmp7, + FILTER_BITS, tmp8, FILTER_BITS, tmp5, tmp6, tmp7, tmp8); + DUP2_ARG2(__lsx_vpickev_b, tmp6, tmp5, tmp8, tmp7, out0, out1); + __lsx_vstelm_d(out0, dst, 0, 0); + dst += dst_stride; + __lsx_vstelm_d(out0, dst, 0, 1); + dst += dst_stride; + __lsx_vstelm_d(out1, dst, 0, 0); + dst += dst_stride; + __lsx_vstelm_d(out1, dst, 0, 1); + dst += dst_stride; + } +} + +static void common_hv_2ht_2vt_8w_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter_horiz, int8_t *filter_vert, + int32_t height) { + if (height == 4) { + common_hv_2ht_2vt_8x4_lsx(src, src_stride, dst, dst_stride, filter_horiz, + filter_vert); + } else { + common_hv_2ht_2vt_8x8mult_lsx(src, src_stride, dst, dst_stride, + filter_horiz, filter_vert, height); + } +} + +static void common_hv_2ht_2vt_16w_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter_horiz, int8_t *filter_vert, + int32_t height) { + uint32_t loop_cnt = (height >> 2); + __m128i src0, src1, src2, src3, src4, src5, src6, src7, mask; + __m128i filt_hz, filt_vt, vec0, vec1; + __m128i tmp, tmp1, tmp2, hz_out0, hz_out1, hz_out2, hz_out3; + + int32_t src_stride2 = src_stride << 1; + int32_t src_stride3 = src_stride2 + src_stride; + int32_t src_stride4 = src_stride2 << 1; + + mask = __lsx_vld(mc_filt_mask_arr, 0); + + /* rearranging filter */ + DUP2_ARG2(__lsx_vldrepl_h, filter_horiz, 0, filter_vert, 0, filt_hz, filt_vt); + + DUP2_ARG2(__lsx_vld, src, 0, src, 8, src0, src1); + src += src_stride; + + hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS); + hz_out2 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS); + + for (; loop_cnt--;) { + uint8_t *src_tmp0 = src + 8; + + DUP2_ARG2(__lsx_vld, src, 0, src_tmp0, 0, src0, src1); + DUP4_ARG2(__lsx_vldx, src, src_stride, src_tmp0, src_stride, src, + src_stride2, src_tmp0, src_stride2, src2, src3, src4, src5); + DUP2_ARG2(__lsx_vldx, src, src_stride3, src_tmp0, src_stride3, src6, src7); + src += src_stride4; + + hz_out1 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS); + hz_out3 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS); + DUP2_ARG2(__lsx_vpackev_b, hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1); + DUP2_ARG2(__lsx_vdp2_h_bu, vec0, filt_vt, vec1, filt_vt, tmp1, tmp2); + DUP2_ARG2(__lsx_vsrari_h, tmp1, FILTER_BITS, tmp2, FILTER_BITS, tmp1, tmp2); + tmp = __lsx_vpickev_b(tmp2, tmp1); + __lsx_vst(tmp, dst, 0); + dst += dst_stride; + + hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS); + hz_out2 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS); + DUP2_ARG2(__lsx_vpackev_b, hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1); + DUP2_ARG2(__lsx_vdp2_h_bu, vec0, filt_vt, vec1, filt_vt, tmp1, tmp2); + DUP2_ARG2(__lsx_vsrari_h, tmp1, FILTER_BITS, tmp2, FILTER_BITS, tmp1, tmp2); + tmp = __lsx_vpickev_b(tmp2, tmp1); + __lsx_vst(tmp, dst, 0); + dst += dst_stride; + + hz_out1 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS); + hz_out3 = HORIZ_2TAP_FILT_UH(src5, src5, mask, filt_hz, FILTER_BITS); + DUP2_ARG2(__lsx_vpackev_b, hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1); + DUP2_ARG2(__lsx_vdp2_h_bu, vec0, filt_vt, vec1, filt_vt, tmp1, tmp2); + DUP2_ARG2(__lsx_vsrari_h, tmp1, FILTER_BITS, tmp2, FILTER_BITS, tmp1, tmp2); + tmp = __lsx_vpickev_b(tmp2, tmp1); + __lsx_vst(tmp, dst, 0); + dst += dst_stride; + + hz_out0 = HORIZ_2TAP_FILT_UH(src6, src6, mask, filt_hz, FILTER_BITS); + hz_out2 = HORIZ_2TAP_FILT_UH(src7, src7, mask, filt_hz, FILTER_BITS); + DUP2_ARG2(__lsx_vpackev_b, hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1); + DUP2_ARG2(__lsx_vdp2_h_bu, vec0, filt_vt, vec1, filt_vt, tmp1, tmp2); + DUP2_ARG2(__lsx_vsrari_h, tmp1, FILTER_BITS, tmp2, FILTER_BITS, tmp1, tmp2); + tmp = __lsx_vpickev_b(tmp2, tmp1); + __lsx_vst(tmp, dst, 0); + dst += dst_stride; + } +} + +static void common_hv_2ht_2vt_32w_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter_horiz, int8_t *filter_vert, + int32_t height) { + common_hv_2ht_2vt_16w_lsx(src, src_stride, dst, dst_stride, filter_horiz, + filter_vert, height); + src += 16; + dst += 16; + + common_hv_2ht_2vt_16w_lsx(src, src_stride, dst, dst_stride, filter_horiz, + filter_vert, height); + src += 16; + dst += 16; +} + +static void common_hv_2ht_2vt_64w_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter_horiz, int8_t *filter_vert, + int32_t height) { + int32_t multiple8_cnt; + for (multiple8_cnt = 4; multiple8_cnt--;) { + common_hv_2ht_2vt_16w_lsx(src, src_stride, dst, dst_stride, filter_horiz, + filter_vert, height); + src += 16; + dst += 16; + } +} + +void vpx_convolve8_lsx(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, + ptrdiff_t dst_stride, const InterpKernel *filter, + int x0_q4, int32_t x_step_q4, int y0_q4, + int32_t y_step_q4, int32_t w, int32_t h) { + const int16_t *const filter_x = filter[x0_q4]; + const int16_t *const filter_y = filter[y0_q4]; + int8_t cnt, filt_hor[8], filt_ver[8]; + + assert(x_step_q4 == 16); + assert(y_step_q4 == 16); + assert(((const int32_t *)filter_x)[1] != 0x800000); + assert(((const int32_t *)filter_y)[1] != 0x800000); + + for (cnt = 0; cnt < 8; ++cnt) { + filt_hor[cnt] = filter_x[cnt]; + filt_ver[cnt] = filter_y[cnt]; + } + + if (vpx_get_filter_taps(filter_x) == 2 && + vpx_get_filter_taps(filter_y) == 2) { + switch (w) { + case 4: + common_hv_2ht_2vt_4w_lsx(src, (int32_t)src_stride, dst, + (int32_t)dst_stride, &filt_hor[3], + &filt_ver[3], (int32_t)h); + break; + case 8: + common_hv_2ht_2vt_8w_lsx(src, (int32_t)src_stride, dst, + (int32_t)dst_stride, &filt_hor[3], + &filt_ver[3], (int32_t)h); + break; + case 16: + common_hv_2ht_2vt_16w_lsx(src, (int32_t)src_stride, dst, + (int32_t)dst_stride, &filt_hor[3], + &filt_ver[3], (int32_t)h); + break; + case 32: + common_hv_2ht_2vt_32w_lsx(src, (int32_t)src_stride, dst, + (int32_t)dst_stride, &filt_hor[3], + &filt_ver[3], (int32_t)h); + break; + case 64: + common_hv_2ht_2vt_64w_lsx(src, (int32_t)src_stride, dst, + (int32_t)dst_stride, &filt_hor[3], + &filt_ver[3], (int32_t)h); + break; + default: + vpx_convolve8_c(src, src_stride, dst, dst_stride, filter, x0_q4, + x_step_q4, y0_q4, y_step_q4, w, h); + break; + } + } else if (vpx_get_filter_taps(filter_x) == 2 || + vpx_get_filter_taps(filter_y) == 2) { + vpx_convolve8_c(src, src_stride, dst, dst_stride, filter, x0_q4, x_step_q4, + y0_q4, y_step_q4, w, h); + } else { + switch (w) { + case 4: + common_hv_8ht_8vt_4w_lsx(src, (int32_t)src_stride, dst, + (int32_t)dst_stride, filt_hor, filt_ver, + (int32_t)h); + break; + case 8: + common_hv_8ht_8vt_8w_lsx(src, (int32_t)src_stride, dst, + (int32_t)dst_stride, filt_hor, filt_ver, + (int32_t)h); + break; + case 16: + common_hv_8ht_8vt_16w_lsx(src, (int32_t)src_stride, dst, + (int32_t)dst_stride, filt_hor, filt_ver, + (int32_t)h); + break; + case 32: + common_hv_8ht_8vt_32w_lsx(src, (int32_t)src_stride, dst, + (int32_t)dst_stride, filt_hor, filt_ver, + (int32_t)h); + break; + case 64: + common_hv_8ht_8vt_64w_lsx(src, (int32_t)src_stride, dst, + (int32_t)dst_stride, filt_hor, filt_ver, + (int32_t)h); + break; + default: + vpx_convolve8_c(src, src_stride, dst, dst_stride, filter, x0_q4, + x_step_q4, y0_q4, y_step_q4, w, h); + break; + } + } +} diff --git a/vpx_dsp/loongarch/vpx_convolve8_vert_lsx.c b/vpx_dsp/loongarch/vpx_convolve8_vert_lsx.c new file mode 100644 index 0000000..c0bb10f --- /dev/null +++ b/vpx_dsp/loongarch/vpx_convolve8_vert_lsx.c @@ -0,0 +1,870 @@ +/* + * 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 +#include "./vpx_dsp_rtcd.h" +#include "vpx_dsp/loongarch/vpx_convolve_lsx.h" + +static void common_vt_8t_4w_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + uint32_t loop_cnt = height >> 2; + int32_t src_stride2 = src_stride << 1; + int32_t src_stride3 = src_stride + src_stride2; + int32_t src_stride4 = src_stride2 << 1; + __m128i src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10; + __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5; + __m128i reg0, reg1, reg2, reg3, reg4; + __m128i filter0, filter1, filter2, filter3; + __m128i out0, out1; + uint8_t *_src = (uint8_t *)src - src_stride3; + + DUP4_ARG2(__lsx_vldrepl_h, filter, 0, filter, 2, filter, 4, filter, 6, + filter0, filter1, filter2, filter3); + src0 = __lsx_vld(_src, 0); + DUP2_ARG2(__lsx_vldx, _src, src_stride, _src, src_stride2, src1, src2); + src3 = __lsx_vldx(_src, src_stride3); + _src += src_stride4; + src4 = __lsx_vld(_src, 0); + DUP2_ARG2(__lsx_vldx, _src, src_stride, _src, src_stride2, src5, src6); + _src += src_stride3; + DUP4_ARG2(__lsx_vilvl_b, src1, src0, src3, src2, src5, src4, src2, src1, tmp0, + tmp1, tmp2, tmp3); + DUP2_ARG2(__lsx_vilvl_b, src4, src3, src6, src5, tmp4, tmp5); + DUP2_ARG2(__lsx_vilvl_d, tmp3, tmp0, tmp4, tmp1, reg0, reg1); + reg2 = __lsx_vilvl_d(tmp5, tmp2); + DUP2_ARG2(__lsx_vxori_b, reg0, 128, reg1, 128, reg0, reg1); + reg2 = __lsx_vxori_b(reg2, 128); + + for (; loop_cnt--;) { + src7 = __lsx_vld(_src, 0); + DUP2_ARG2(__lsx_vldx, _src, src_stride, _src, src_stride2, src8, src9); + src10 = __lsx_vldx(_src, src_stride3); + _src += src_stride4; + DUP4_ARG2(__lsx_vilvl_b, src7, src6, src8, src7, src9, src8, src10, src9, + tmp0, tmp1, tmp2, tmp3); + DUP2_ARG2(__lsx_vilvl_d, tmp1, tmp0, tmp3, tmp2, reg3, reg4); + DUP2_ARG2(__lsx_vxori_b, reg3, 128, reg4, 128, reg3, reg4); + out0 = FILT_8TAP_DPADD_S_H(reg0, reg1, reg2, reg3, filter0, filter1, + filter2, filter3); + out1 = FILT_8TAP_DPADD_S_H(reg1, reg2, reg3, reg4, filter0, filter1, + filter2, filter3); + out0 = __lsx_vssrarni_b_h(out1, out0, 7); + out0 = __lsx_vxori_b(out0, 128); + __lsx_vstelm_w(out0, dst, 0, 0); + dst += dst_stride; + __lsx_vstelm_w(out0, dst, 0, 1); + dst += dst_stride; + __lsx_vstelm_w(out0, dst, 0, 2); + dst += dst_stride; + __lsx_vstelm_w(out0, dst, 0, 3); + dst += dst_stride; + + reg0 = reg2; + reg1 = reg3; + reg2 = reg4; + src6 = src10; + } +} + +static void common_vt_8t_8w_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + uint32_t loop_cnt = height >> 2; + __m128i src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10; + __m128i tmp0, tmp1, tmp2, tmp3; + __m128i reg0, reg1, reg2, reg3, reg4, reg5; + __m128i filter0, filter1, filter2, filter3; + __m128i out0, out1, out2, out3; + int32_t src_stride2 = src_stride << 1; + int32_t src_stride3 = src_stride + src_stride2; + int32_t src_stride4 = src_stride2 << 1; + src = src - src_stride3; + + DUP4_ARG2(__lsx_vldrepl_h, filter, 0, filter, 2, filter, 4, filter, 6, + filter0, filter1, filter2, filter3); + + src0 = __lsx_vld(src, 0); + DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride2, src1, src2); + src3 = __lsx_vldx(src, src_stride3); + src += src_stride4; + src4 = __lsx_vld(src, 0); + DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride2, src5, src6); + src += src_stride3; + + DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128, src0, + src1, src2, src3); + DUP2_ARG2(__lsx_vxori_b, src4, 128, src5, 128, src4, src5); + src6 = __lsx_vxori_b(src6, 128); + DUP4_ARG2(__lsx_vilvl_b, src1, src0, src3, src2, src5, src4, src2, src1, reg0, + reg1, reg2, reg3); + DUP2_ARG2(__lsx_vilvl_b, src4, src3, src6, src5, reg4, reg5); + + for (; loop_cnt--;) { + src7 = __lsx_vld(src, 0); + DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride2, src8, src9); + src10 = __lsx_vldx(src, src_stride3); + src += src_stride4; + DUP4_ARG2(__lsx_vxori_b, src7, 128, src8, 128, src9, 128, src10, 128, src7, + src8, src9, src10); + DUP4_ARG2(__lsx_vilvl_b, src7, src6, src8, src7, src9, src8, src10, src9, + tmp0, tmp1, tmp2, tmp3); + out0 = FILT_8TAP_DPADD_S_H(reg0, reg1, reg2, tmp0, filter0, filter1, + filter2, filter3); + out1 = FILT_8TAP_DPADD_S_H(reg3, reg4, reg5, tmp1, filter0, filter1, + filter2, filter3); + out2 = FILT_8TAP_DPADD_S_H(reg1, reg2, tmp0, tmp2, filter0, filter1, + filter2, filter3); + out3 = FILT_8TAP_DPADD_S_H(reg4, reg5, tmp1, tmp3, filter0, filter1, + filter2, filter3); + DUP2_ARG3(__lsx_vssrarni_b_h, out1, out0, 7, out3, out2, 7, out0, out1); + DUP2_ARG2(__lsx_vxori_b, out0, 128, out1, 128, out0, out1); + __lsx_vstelm_d(out0, dst, 0, 0); + dst += dst_stride; + __lsx_vstelm_d(out0, dst, 0, 1); + dst += dst_stride; + __lsx_vstelm_d(out1, dst, 0, 0); + dst += dst_stride; + __lsx_vstelm_d(out1, dst, 0, 1); + dst += dst_stride; + + reg0 = reg2; + reg1 = tmp0; + reg2 = tmp2; + reg3 = reg5; + reg4 = tmp1; + reg5 = tmp3; + src6 = src10; + } +} + +static void common_vt_8t_16w_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + uint32_t loop_cnt = height >> 2; + __m128i src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10; + __m128i filter0, filter1, filter2, filter3; + __m128i reg0, reg1, reg2, reg3, reg4, reg5; + __m128i reg6, reg7, reg8, reg9, reg10, reg11; + __m128i tmp0, tmp1, tmp2, tmp3; + int32_t src_stride2 = src_stride << 1; + int32_t src_stride3 = src_stride + src_stride2; + int32_t src_stride4 = src_stride2 << 1; + // uint8_t *_src = (uint8_t *)src - src_stride3; + src -= src_stride3; + DUP4_ARG2(__lsx_vldrepl_h, filter, 0, filter, 2, filter, 4, filter, 6, + filter0, filter1, filter2, filter3); + + src0 = __lsx_vld(src, 0); + DUP4_ARG2(__lsx_vldx, src, src_stride, src, src_stride2, src, src_stride3, + src, src_stride4, src1, src2, src3, src4); + src += src_stride4; + DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride2, src5, src6); + src += src_stride3; + + DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128, src0, + src1, src2, src3); + DUP2_ARG2(__lsx_vxori_b, src4, 128, src5, 128, src4, src5); + src6 = __lsx_vxori_b(src6, 128); + DUP4_ARG2(__lsx_vilvl_b, src1, src0, src3, src2, src5, src4, src2, src1, reg0, + reg1, reg2, reg3); + DUP2_ARG2(__lsx_vilvl_b, src4, src3, src6, src5, reg4, reg5); + DUP4_ARG2(__lsx_vilvh_b, src1, src0, src3, src2, src5, src4, src2, src1, reg6, + reg7, reg8, reg9); + DUP2_ARG2(__lsx_vilvh_b, src4, src3, src6, src5, reg10, reg11); + + for (; loop_cnt--;) { + src7 = __lsx_vld(src, 0); + DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride2, src8, src9); + src10 = __lsx_vldx(src, src_stride3); + src += src_stride4; + + DUP4_ARG2(__lsx_vxori_b, src7, 128, src8, 128, src9, 128, src10, 128, src7, + src8, src9, src10); + DUP4_ARG2(__lsx_vilvl_b, src7, src6, src8, src7, src9, src8, src10, src9, + src0, src1, src2, src3); + DUP4_ARG2(__lsx_vilvh_b, src7, src6, src8, src7, src9, src8, src10, src9, + src4, src5, src7, src8); + tmp0 = FILT_8TAP_DPADD_S_H(reg0, reg1, reg2, src0, filter0, filter1, + filter2, filter3); + tmp1 = FILT_8TAP_DPADD_S_H(reg3, reg4, reg5, src1, filter0, filter1, + filter2, filter3); + tmp2 = FILT_8TAP_DPADD_S_H(reg6, reg7, reg8, src4, filter0, filter1, + filter2, filter3); + tmp3 = FILT_8TAP_DPADD_S_H(reg9, reg10, reg11, src5, filter0, filter1, + filter2, filter3); + DUP2_ARG3(__lsx_vssrarni_b_h, tmp2, tmp0, 7, tmp3, tmp1, 7, tmp0, tmp1); + DUP2_ARG2(__lsx_vxori_b, tmp0, 128, tmp1, 128, tmp0, tmp1); + __lsx_vst(tmp0, dst, 0); + dst += dst_stride; + __lsx_vst(tmp1, dst, 0); + dst += dst_stride; + tmp0 = FILT_8TAP_DPADD_S_H(reg1, reg2, src0, src2, filter0, filter1, + filter2, filter3); + tmp1 = FILT_8TAP_DPADD_S_H(reg4, reg5, src1, src3, filter0, filter1, + filter2, filter3); + tmp2 = FILT_8TAP_DPADD_S_H(reg7, reg8, src4, src7, filter0, filter1, + filter2, filter3); + tmp3 = FILT_8TAP_DPADD_S_H(reg10, reg11, src5, src8, filter0, filter1, + filter2, filter3); + DUP2_ARG3(__lsx_vssrarni_b_h, tmp2, tmp0, 7, tmp3, tmp1, 7, tmp0, tmp1); + DUP2_ARG2(__lsx_vxori_b, tmp0, 128, tmp1, 128, tmp0, tmp1); + __lsx_vst(tmp0, dst, 0); + dst += dst_stride; + __lsx_vst(tmp1, dst, 0); + dst += dst_stride; + + reg0 = reg2; + reg1 = src0; + reg2 = src2; + reg3 = reg5; + reg4 = src1; + reg5 = src3; + reg6 = reg8; + reg7 = src4; + reg8 = src7; + reg9 = reg11; + reg10 = src5; + reg11 = src8; + src6 = src10; + } +} + +static void common_vt_8t_16w_mult_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height, + int32_t width) { + uint8_t *src_tmp; + uint8_t *dst_tmp; + uint32_t cnt = width >> 4; + __m128i src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10; + __m128i filter0, filter1, filter2, filter3; + __m128i reg0, reg1, reg2, reg3, reg4, reg5; + __m128i reg6, reg7, reg8, reg9, reg10, reg11; + __m128i tmp0, tmp1, tmp2, tmp3; + int32_t src_stride2 = src_stride << 1; + int32_t src_stride3 = src_stride + src_stride2; + int32_t src_stride4 = src_stride2 << 1; + int32_t dst_stride2 = dst_stride << 1; + int32_t dst_stride3 = dst_stride2 + dst_stride; + int32_t dst_stride4 = dst_stride2 << 1; + src -= src_stride3; + DUP4_ARG2(__lsx_vldrepl_h, filter, 0, filter, 2, filter, 4, filter, 6, + filter0, filter1, filter2, filter3); + + for (; cnt--;) { + uint32_t loop_cnt = height >> 2; + + src_tmp = src; + dst_tmp = dst; + + src0 = __lsx_vld(src_tmp, 0); + DUP2_ARG2(__lsx_vldx, src_tmp, src_stride, src_tmp, src_stride2, src1, + src2); + src3 = __lsx_vldx(src_tmp, src_stride3); + src_tmp += src_stride4; + src4 = __lsx_vld(src_tmp, 0); + DUP2_ARG2(__lsx_vldx, src_tmp, src_stride, src_tmp, src_stride2, src5, + src6); + src_tmp += src_stride3; + + DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128, src0, + src1, src2, src3); + DUP2_ARG2(__lsx_vxori_b, src4, 128, src5, 128, src4, src5); + src6 = __lsx_vxori_b(src6, 128); + DUP4_ARG2(__lsx_vilvl_b, src1, src0, src3, src2, src5, src4, src2, src1, + reg0, reg1, reg2, reg3); + DUP2_ARG2(__lsx_vilvl_b, src4, src3, src6, src5, reg4, reg5); + DUP4_ARG2(__lsx_vilvh_b, src1, src0, src3, src2, src5, src4, src2, src1, + reg6, reg7, reg8, reg9); + DUP2_ARG2(__lsx_vilvh_b, src4, src3, src6, src5, reg10, reg11); + + for (; loop_cnt--;) { + src7 = __lsx_vld(src_tmp, 0); + DUP2_ARG2(__lsx_vldx, src_tmp, src_stride, src_tmp, src_stride2, src8, + src9); + src10 = __lsx_vldx(src_tmp, src_stride3); + src_tmp += src_stride4; + DUP4_ARG2(__lsx_vxori_b, src7, 128, src8, 128, src9, 128, src10, 128, + src7, src8, src9, src10); + DUP4_ARG2(__lsx_vilvl_b, src7, src6, src8, src7, src9, src8, src10, src9, + src0, src1, src2, src3); + DUP4_ARG2(__lsx_vilvh_b, src7, src6, src8, src7, src9, src8, src10, src9, + src4, src5, src7, src8); + tmp0 = FILT_8TAP_DPADD_S_H(reg0, reg1, reg2, src0, filter0, filter1, + filter2, filter3); + tmp1 = FILT_8TAP_DPADD_S_H(reg3, reg4, reg5, src1, filter0, filter1, + filter2, filter3); + tmp2 = FILT_8TAP_DPADD_S_H(reg6, reg7, reg8, src4, filter0, filter1, + filter2, filter3); + tmp3 = FILT_8TAP_DPADD_S_H(reg9, reg10, reg11, src5, filter0, filter1, + filter2, filter3); + DUP2_ARG3(__lsx_vssrarni_b_h, tmp2, tmp0, 7, tmp3, tmp1, 7, tmp0, tmp1); + DUP2_ARG2(__lsx_vxori_b, tmp0, 128, tmp1, 128, tmp0, tmp1); + __lsx_vst(tmp0, dst_tmp, 0); + __lsx_vstx(tmp1, dst_tmp, dst_stride); + tmp0 = FILT_8TAP_DPADD_S_H(reg1, reg2, src0, src2, filter0, filter1, + filter2, filter3); + tmp1 = FILT_8TAP_DPADD_S_H(reg4, reg5, src1, src3, filter0, filter1, + filter2, filter3); + tmp2 = FILT_8TAP_DPADD_S_H(reg7, reg8, src4, src7, filter0, filter1, + filter2, filter3); + tmp3 = FILT_8TAP_DPADD_S_H(reg10, reg11, src5, src8, filter0, filter1, + filter2, filter3); + DUP2_ARG3(__lsx_vssrarni_b_h, tmp2, tmp0, 7, tmp3, tmp1, 7, tmp0, tmp1); + DUP2_ARG2(__lsx_vxori_b, tmp0, 128, tmp1, 128, tmp0, tmp1); + __lsx_vstx(tmp0, dst_tmp, dst_stride2); + __lsx_vstx(tmp1, dst_tmp, dst_stride3); + dst_tmp += dst_stride4; + + reg0 = reg2; + reg1 = src0; + reg2 = src2; + reg3 = reg5; + reg4 = src1; + reg5 = src3; + reg6 = reg8; + reg7 = src4; + reg8 = src7; + reg9 = reg11; + reg10 = src5; + reg11 = src8; + src6 = src10; + } + src += 16; + dst += 16; + } +} + +static void common_vt_8t_32w_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + common_vt_8t_16w_mult_lsx(src, src_stride, dst, dst_stride, filter, height, + 32); +} + +static void common_vt_8t_64w_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + common_vt_8t_16w_mult_lsx(src, src_stride, dst, dst_stride, filter, height, + 64); +} + +static void common_vt_2t_4x4_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter) { + __m128i src0, src1, src2, src3, src4; + __m128i src10_l, src32_l, src21_l, src43_l, src2110, src4332; + __m128i filt0, tmp0, tmp1; + + int32_t src_stride2 = src_stride << 1; + int32_t src_stride3 = src_stride2 + src_stride; + int32_t src_stride4 = src_stride2 << 1; + + int32_t dst_stride2 = dst_stride << 1; + int32_t dst_stride3 = dst_stride2 + dst_stride; + + filt0 = __lsx_vldrepl_h(filter, 0); + + src0 = __lsx_vld(src, 0); + DUP4_ARG2(__lsx_vldx, src, src_stride, src, src_stride2, src, src_stride3, + src, src_stride4, src1, src2, src3, src4); + src += (src_stride4 + src_stride); + + DUP4_ARG2(__lsx_vilvl_b, src1, src0, src2, src1, src3, src2, src4, src3, + src10_l, src21_l, src32_l, src43_l); + DUP2_ARG2(__lsx_vilvl_d, src21_l, src10_l, src43_l, src32_l, src2110, + src4332); + DUP2_ARG2(__lsx_vdp2_h_bu, src2110, filt0, src4332, filt0, tmp0, tmp1); + DUP2_ARG2(__lsx_vsrari_h, tmp0, FILTER_BITS, tmp1, FILTER_BITS, tmp0, tmp1); + src2110 = __lsx_vpickev_b(tmp1, tmp0); + + __lsx_vstelm_w(src2110, dst, 0, 0); + __lsx_vstelm_w(src2110, dst + dst_stride, 0, 1); + __lsx_vstelm_w(src2110, dst + dst_stride2, 0, 2); + __lsx_vstelm_w(src2110, dst + dst_stride3, 0, 3); +} + +static void common_vt_2t_4x8_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter) { + __m128i src0, src1, src2, src3, src4, src5, src6, src7, src8; + __m128i src10_l, src32_l, src54_l, src76_l, src21_l, src43_l; + __m128i src65_l, src87_l, src2110, src4332, src6554, src8776; + __m128i tmp0, tmp1, tmp2, tmp3; + __m128i filt0; + + int32_t src_stride2 = src_stride << 1; + int32_t src_stride3 = src_stride2 + src_stride; + int32_t src_stride4 = src_stride2 << 1; + + int32_t dst_stride2 = dst_stride << 1; + int32_t dst_stride3 = dst_stride2 + dst_stride; + int32_t dst_stride4 = dst_stride2 << 1; + + uint8_t *dst_tmp1 = dst + dst_stride4; + + filt0 = __lsx_vldrepl_h(filter, 0); + src0 = __lsx_vld(src, 0); + DUP4_ARG2(__lsx_vldx, src, src_stride, src, src_stride2, src, src_stride3, + src, src_stride4, src1, src2, src3, src4); + src += src_stride4; + DUP4_ARG2(__lsx_vldx, src, src_stride, src, src_stride2, src, src_stride3, + src, src_stride4, src5, src6, src7, src8); + src += (src_stride4 + src_stride); + + DUP4_ARG2(__lsx_vilvl_b, src1, src0, src2, src1, src3, src2, src4, src3, + src10_l, src21_l, src32_l, src43_l); + DUP4_ARG2(__lsx_vilvl_b, src5, src4, src6, src5, src7, src6, src8, src7, + src54_l, src65_l, src76_l, src87_l); + DUP4_ARG2(__lsx_vilvl_d, src21_l, src10_l, src43_l, src32_l, src65_l, src54_l, + src87_l, src76_l, src2110, src4332, src6554, src8776); + DUP4_ARG2(__lsx_vdp2_h_bu, src2110, filt0, src4332, filt0, src6554, filt0, + src8776, filt0, tmp0, tmp1, tmp2, tmp3); + DUP4_ARG2(__lsx_vsrari_h, tmp0, FILTER_BITS, tmp1, FILTER_BITS, tmp2, + FILTER_BITS, tmp3, FILTER_BITS, tmp0, tmp1, tmp2, tmp3); + DUP2_ARG2(__lsx_vpickev_b, tmp1, tmp0, tmp3, tmp2, src2110, src4332); + + __lsx_vstelm_w(src2110, dst, 0, 0); + __lsx_vstelm_w(src2110, dst + dst_stride, 0, 1); + __lsx_vstelm_w(src2110, dst + dst_stride2, 0, 2); + __lsx_vstelm_w(src2110, dst + dst_stride3, 0, 3); + + __lsx_vstelm_w(src4332, dst_tmp1, 0, 0); + __lsx_vstelm_w(src4332, dst_tmp1 + dst_stride, 0, 1); + __lsx_vstelm_w(src4332, dst_tmp1 + dst_stride2, 0, 2); + __lsx_vstelm_w(src4332, dst_tmp1 + dst_stride3, 0, 3); +} + +static void common_vt_2t_4w_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + if (height == 4) { + common_vt_2t_4x4_lsx(src, src_stride, dst, dst_stride, filter); + } else if (height == 8) { + common_vt_2t_4x8_lsx(src, src_stride, dst, dst_stride, filter); + } +} + +static void common_vt_2t_8x4_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter) { + __m128i src0, src1, src2, src3, src4, vec0, vec1, vec2, vec3, filt0; + __m128i out0, out1; + __m128i tmp0, tmp1, tmp2, tmp3; + + int32_t src_stride2 = src_stride << 1; + int32_t src_stride3 = src_stride2 + src_stride; + int32_t src_stride4 = src_stride2 << 1; + + int32_t dst_stride2 = dst_stride << 1; + int32_t dst_stride3 = dst_stride2 + dst_stride; + + /* rearranging filter_y */ + filt0 = __lsx_vldrepl_h(filter, 0); + + src0 = __lsx_vld(src, 0); + DUP4_ARG2(__lsx_vldx, src, src_stride, src, src_stride2, src, src_stride3, + src, src_stride4, src1, src2, src3, src4); + + DUP4_ARG2(__lsx_vilvl_b, src1, src0, src2, src1, src3, src2, src4, src3, vec0, + vec1, vec2, vec3); + DUP4_ARG2(__lsx_vdp2_h_bu, vec0, filt0, vec1, filt0, vec2, filt0, vec3, filt0, + tmp0, tmp1, tmp2, tmp3); + DUP4_ARG2(__lsx_vsrari_h, tmp0, FILTER_BITS, tmp1, FILTER_BITS, tmp2, + FILTER_BITS, tmp3, FILTER_BITS, tmp0, tmp1, tmp2, tmp3); + DUP2_ARG2(__lsx_vpickev_b, tmp1, tmp0, tmp3, tmp2, out0, out1); + + __lsx_vstelm_d(out0, dst, 0, 0); + __lsx_vstelm_d(out0, dst + dst_stride, 0, 1); + __lsx_vstelm_d(out1, dst + dst_stride2, 0, 0); + __lsx_vstelm_d(out1, dst + dst_stride3, 0, 1); +} + +static void common_vt_2t_8x8mult_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + uint32_t loop_cnt = (height >> 3); + __m128i src0, src1, src2, src3, src4, src5, src6, src7, src8; + __m128i vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, filt0; + __m128i out0, out1; + __m128i tmp0, tmp1, tmp2, tmp3; + + int32_t src_stride2 = src_stride << 1; + int32_t src_stride3 = src_stride2 + src_stride; + int32_t src_stride4 = src_stride2 << 1; + + int32_t dst_stride2 = dst_stride << 1; + int32_t dst_stride3 = dst_stride2 + dst_stride; + int32_t dst_stride4 = dst_stride2 << 1; + + filt0 = __lsx_vldrepl_h(filter, 0); + src0 = __lsx_vld(src, 0); + src += src_stride; + + for (; loop_cnt--;) { + src1 = __lsx_vld(src, 0); + DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride2, src2, src3); + src4 = __lsx_vldx(src, src_stride3); + src += src_stride4; + src5 = __lsx_vld(src, 0); + DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride2, src6, src7) + src8 = __lsx_vldx(src, src_stride3); + src += src_stride4; + + DUP4_ARG2(__lsx_vilvl_b, src1, src0, src2, src1, src3, src2, src4, src3, + vec0, vec1, vec2, vec3); + DUP4_ARG2(__lsx_vilvl_b, src5, src4, src6, src5, src7, src6, src8, src7, + vec4, vec5, vec6, vec7); + DUP4_ARG2(__lsx_vdp2_h_bu, vec0, filt0, vec1, filt0, vec2, filt0, vec3, + filt0, tmp0, tmp1, tmp2, tmp3); + DUP4_ARG2(__lsx_vsrari_h, tmp0, FILTER_BITS, tmp1, FILTER_BITS, tmp2, + FILTER_BITS, tmp3, FILTER_BITS, tmp0, tmp1, tmp2, tmp3); + DUP2_ARG2(__lsx_vpickev_b, tmp1, tmp0, tmp3, tmp2, out0, out1); + __lsx_vstelm_d(out0, dst, 0, 0); + __lsx_vstelm_d(out0, dst + dst_stride, 0, 1); + __lsx_vstelm_d(out1, dst + dst_stride2, 0, 0); + __lsx_vstelm_d(out1, dst + dst_stride3, 0, 1); + dst += dst_stride4; + + DUP4_ARG2(__lsx_vdp2_h_bu, vec4, filt0, vec5, filt0, vec6, filt0, vec7, + filt0, tmp0, tmp1, tmp2, tmp3); + DUP4_ARG2(__lsx_vsrari_h, tmp0, FILTER_BITS, tmp1, FILTER_BITS, tmp2, + FILTER_BITS, tmp3, FILTER_BITS, tmp0, tmp1, tmp2, tmp3); + DUP2_ARG2(__lsx_vpickev_b, tmp1, tmp0, tmp3, tmp2, out0, out1); + __lsx_vstelm_d(out0, dst, 0, 0); + __lsx_vstelm_d(out0, dst + dst_stride, 0, 1); + __lsx_vstelm_d(out1, dst + dst_stride2, 0, 0); + __lsx_vstelm_d(out1, dst + dst_stride3, 0, 1); + dst += dst_stride4; + + src0 = src8; + } +} + +static void common_vt_2t_8w_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + if (height == 4) { + common_vt_2t_8x4_lsx(src, src_stride, dst, dst_stride, filter); + } else { + common_vt_2t_8x8mult_lsx(src, src_stride, dst, dst_stride, filter, height); + } +} + +static void common_vt_2t_16w_lsx(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const InterpKernel *filter, int y0_q4, + int y_step_q4, int w, int height) { + uint32_t loop_cnt = (height >> 2); + __m128i src0, src1, src2, src3, src4; + __m128i vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, filt0; + __m128i tmp0, tmp1, tmp2, tmp3, tmp4; + int32_t src_stride2 = src_stride << 1; + int32_t src_stride3 = src_stride2 + src_stride; + int32_t src_stride4 = src_stride2 << 1; + + const int16_t *const filter_y = filter[y0_q4]; + int8_t cnt, filt_ver[8]; + + assert(y_step_q4 == 16); + assert(((const int32_t *)filter_y)[1] != 0x800000); + + for (cnt = 8; cnt--;) { + filt_ver[cnt] = filter_y[cnt]; + } + + filt0 = __lsx_vldrepl_h(&filt_ver[3], 0); + + src0 = __lsx_vld(src, 0); + src += src_stride; + + for (; loop_cnt--;) { + src1 = __lsx_vld(src, 0); + DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride2, src2, src3); + src4 = __lsx_vldx(src, src_stride3); + src += src_stride4; + + DUP2_ARG2(__lsx_vilvl_b, src1, src0, src2, src1, vec0, vec2); + DUP2_ARG2(__lsx_vilvh_b, src1, src0, src2, src1, vec1, vec3); + DUP2_ARG2(__lsx_vdp2_h_bu, vec0, filt0, vec1, filt0, tmp0, tmp1); + DUP2_ARG2(__lsx_vsrari_h, tmp0, FILTER_BITS, tmp1, FILTER_BITS, tmp0, tmp1); + tmp4 = __lsx_vpickev_b(tmp1, tmp0); + __lsx_vst(tmp4, dst, 0); + dst += dst_stride; + + DUP2_ARG2(__lsx_vilvl_b, src3, src2, src4, src3, vec4, vec6); + DUP2_ARG2(__lsx_vilvh_b, src3, src2, src4, src3, vec5, vec7); + DUP2_ARG2(__lsx_vdp2_h_bu, vec2, filt0, vec3, filt0, tmp2, tmp3); + DUP2_ARG2(__lsx_vsrari_h, tmp2, FILTER_BITS, tmp3, FILTER_BITS, tmp2, tmp3); + tmp4 = __lsx_vpickev_b(tmp3, tmp2); + __lsx_vst(tmp4, dst, 0); + dst += dst_stride; + + DUP2_ARG2(__lsx_vdp2_h_bu, vec4, filt0, vec5, filt0, tmp0, tmp1); + DUP2_ARG2(__lsx_vsrari_h, tmp0, FILTER_BITS, tmp1, FILTER_BITS, tmp0, tmp1); + tmp4 = __lsx_vpickev_b(tmp1, tmp0); + __lsx_vst(tmp4, dst, 0); + dst += dst_stride; + + DUP2_ARG2(__lsx_vdp2_h_bu, vec6, filt0, vec7, filt0, tmp2, tmp3); + DUP2_ARG2(__lsx_vsrari_h, tmp2, FILTER_BITS, tmp3, FILTER_BITS, tmp2, tmp3); + tmp4 = __lsx_vpickev_b(tmp3, tmp2); + __lsx_vst(tmp4, dst, 0); + dst += dst_stride; + + src0 = src4; + } +} + +static void common_vt_2t_32w_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + uint32_t loop_cnt = (height >> 2); + __m128i src0, src1, src2, src3, src4, src5, src6, src7, src8, src9; + __m128i vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, filt0; + __m128i tmp0, tmp1, tmp2, tmp3, tmp4; + + int32_t src_stride2 = src_stride << 1; + int32_t src_stride3 = src_stride2 + src_stride; + int32_t src_stride4 = src_stride2 << 1; + + int32_t dst_stride2 = dst_stride << 1; + int32_t dst_stride3 = dst_stride2 + dst_stride; + + uint8_t *src_tmp; + filt0 = __lsx_vldrepl_h(filter, 0); + + src0 = __lsx_vld(src, 0); + src5 = __lsx_vld(src, 16); + src += src_stride; + src_tmp = src + 16; + + for (; loop_cnt--;) { + DUP2_ARG2(__lsx_vld, src, 0, src_tmp, 0, src1, src6); + DUP4_ARG2(__lsx_vldx, src, src_stride, src_tmp, src_stride, src, + src_stride2, src_tmp, src_stride2, src2, src7, src3, src8); + DUP2_ARG2(__lsx_vldx, src, src_stride3, src_tmp, src_stride3, src4, src9); + DUP2_ARG2(__lsx_vilvl_b, src1, src0, src2, src1, vec0, vec2); + DUP2_ARG2(__lsx_vilvh_b, src1, src0, src2, src1, vec1, vec3); + src += src_stride4; + src_tmp += src_stride4; + + DUP2_ARG2(__lsx_vdp2_h_bu, vec0, filt0, vec1, filt0, tmp0, tmp1); + DUP2_ARG2(__lsx_vsrari_h, tmp0, FILTER_BITS, tmp1, FILTER_BITS, tmp0, tmp1); + tmp4 = __lsx_vpickev_b(tmp1, tmp0); + __lsx_vst(tmp4, dst, 0); + + DUP2_ARG2(__lsx_vdp2_h_bu, vec2, filt0, vec3, filt0, tmp2, tmp3); + DUP2_ARG2(__lsx_vsrari_h, tmp2, FILTER_BITS, tmp3, FILTER_BITS, tmp2, tmp3); + tmp4 = __lsx_vpickev_b(tmp3, tmp2); + __lsx_vstx(tmp4, dst, dst_stride); + + DUP2_ARG2(__lsx_vilvl_b, src3, src2, src4, src3, vec4, vec6); + DUP2_ARG2(__lsx_vilvh_b, src3, src2, src4, src3, vec5, vec7); + DUP2_ARG2(__lsx_vdp2_h_bu, vec4, filt0, vec5, filt0, tmp0, tmp1); + DUP2_ARG2(__lsx_vsrari_h, tmp0, FILTER_BITS, tmp1, FILTER_BITS, tmp0, tmp1); + tmp4 = __lsx_vpickev_b(tmp1, tmp0); + __lsx_vstx(tmp4, dst, dst_stride2); + + DUP2_ARG2(__lsx_vdp2_h_bu, vec6, filt0, vec7, filt0, tmp2, tmp3); + DUP2_ARG2(__lsx_vsrari_h, tmp2, FILTER_BITS, tmp3, FILTER_BITS, tmp2, tmp3); + tmp4 = __lsx_vpickev_b(tmp3, tmp2); + __lsx_vstx(tmp4, dst, dst_stride3); + + DUP2_ARG2(__lsx_vilvl_b, src6, src5, src7, src6, vec0, vec2); + DUP2_ARG2(__lsx_vilvh_b, src6, src5, src7, src6, vec1, vec3); + DUP2_ARG2(__lsx_vdp2_h_bu, vec0, filt0, vec1, filt0, tmp0, tmp1); + DUP2_ARG2(__lsx_vsrari_h, tmp0, FILTER_BITS, tmp1, FILTER_BITS, tmp0, tmp1); + tmp4 = __lsx_vpickev_b(tmp1, tmp0); + __lsx_vst(tmp4, dst, 16); + + DUP2_ARG2(__lsx_vdp2_h_bu, vec2, filt0, vec3, filt0, tmp2, tmp3); + DUP2_ARG2(__lsx_vsrari_h, tmp2, FILTER_BITS, tmp3, FILTER_BITS, tmp2, tmp3); + dst += dst_stride; + tmp4 = __lsx_vpickev_b(tmp3, tmp2); + __lsx_vst(tmp4, dst, 16); + + DUP2_ARG2(__lsx_vilvl_b, src8, src7, src9, src8, vec4, vec6); + DUP2_ARG2(__lsx_vilvh_b, src8, src7, src9, src8, vec5, vec7); + DUP2_ARG2(__lsx_vdp2_h_bu, vec4, filt0, vec5, filt0, tmp0, tmp1); + DUP2_ARG2(__lsx_vsrari_h, tmp0, FILTER_BITS, tmp1, FILTER_BITS, tmp0, tmp1); + dst += dst_stride; + tmp4 = __lsx_vpickev_b(tmp1, tmp0); + __lsx_vst(tmp4, dst, 16); + + DUP2_ARG2(__lsx_vdp2_h_bu, vec6, filt0, vec7, filt0, tmp2, tmp3); + DUP2_ARG2(__lsx_vsrari_h, tmp2, FILTER_BITS, tmp3, FILTER_BITS, tmp2, tmp3); + dst += dst_stride; + tmp4 = __lsx_vpickev_b(tmp3, tmp2); + __lsx_vst(tmp4, dst, 16); + + dst += dst_stride; + + src0 = src4; + src5 = src9; + } +} + +static void common_vt_2t_64w_lsx(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride, + int8_t *filter, int32_t height) { + uint32_t loop_cnt = (height >> 1); + __m128i src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10; + __m128i src11, vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, filt0; + __m128i tmp, tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; + + int32_t src_stride2 = src_stride << 1; + int32_t dst_stride2 = dst_stride << 1; + uint8_t *dst_tmp1 = dst + dst_stride; + + filt0 = __lsx_vldrepl_h(filter, 0); + + DUP4_ARG2(__lsx_vld, src, 0, src, 16, src, 32, src, 48, src0, src3, src6, + src9); + src += src_stride; + + for (; loop_cnt--;) { + uint8_t *src_tmp0 = src + src_stride; + + DUP4_ARG2(__lsx_vld, src, 0, src, 16, src, 32, src, 48, src1, src4, src7, + src10); + DUP4_ARG2(__lsx_vld, src_tmp0, 0, src_tmp0, 16, src_tmp0, 32, src_tmp0, 48, + src2, src5, src8, src11); + src += src_stride2; + + DUP2_ARG2(__lsx_vilvl_b, src1, src0, src2, src1, vec0, vec2); + DUP2_ARG2(__lsx_vilvh_b, src1, src0, src2, src1, vec1, vec3); + DUP2_ARG2(__lsx_vdp2_h_bu, vec0, filt0, vec1, filt0, tmp0, tmp1); + DUP2_ARG2(__lsx_vsrari_h, tmp0, FILTER_BITS, tmp1, FILTER_BITS, tmp0, tmp1); + tmp = __lsx_vpickev_b(tmp1, tmp0); + __lsx_vst(tmp, dst, 0); + + DUP2_ARG2(__lsx_vdp2_h_bu, vec2, filt0, vec3, filt0, tmp2, tmp3); + DUP2_ARG2(__lsx_vsrari_h, tmp2, FILTER_BITS, tmp3, FILTER_BITS, tmp2, tmp3); + tmp = __lsx_vpickev_b(tmp3, tmp2); + __lsx_vst(tmp, dst_tmp1, 0); + + DUP2_ARG2(__lsx_vilvl_b, src4, src3, src5, src4, vec4, vec6); + DUP2_ARG2(__lsx_vilvh_b, src4, src3, src5, src4, vec5, vec7); + DUP2_ARG2(__lsx_vdp2_h_bu, vec4, filt0, vec5, filt0, tmp4, tmp5); + DUP2_ARG2(__lsx_vsrari_h, tmp4, FILTER_BITS, tmp5, FILTER_BITS, tmp4, tmp5); + tmp = __lsx_vpickev_b(tmp5, tmp4); + __lsx_vst(tmp, dst, 16); + + DUP2_ARG2(__lsx_vdp2_h_bu, vec6, filt0, vec7, filt0, tmp6, tmp7); + DUP2_ARG2(__lsx_vsrari_h, tmp6, FILTER_BITS, tmp7, FILTER_BITS, tmp6, tmp7); + tmp = __lsx_vpickev_b(tmp7, tmp6); + __lsx_vst(tmp, dst_tmp1, 16); + + DUP2_ARG2(__lsx_vilvl_b, src7, src6, src8, src7, vec0, vec2); + DUP2_ARG2(__lsx_vilvh_b, src7, src6, src8, src7, vec1, vec3); + DUP2_ARG2(__lsx_vdp2_h_bu, vec0, filt0, vec1, filt0, tmp0, tmp1); + DUP2_ARG2(__lsx_vsrari_h, tmp0, FILTER_BITS, tmp1, FILTER_BITS, tmp0, tmp1); + tmp = __lsx_vpickev_b(tmp1, tmp0); + __lsx_vst(tmp, dst, 32); + + DUP2_ARG2(__lsx_vdp2_h_bu, vec2, filt0, vec3, filt0, tmp2, tmp3); + DUP2_ARG2(__lsx_vsrari_h, tmp2, FILTER_BITS, tmp3, FILTER_BITS, tmp2, tmp3); + tmp = __lsx_vpickev_b(tmp3, tmp2); + __lsx_vst(tmp, dst_tmp1, 32); + + DUP2_ARG2(__lsx_vilvl_b, src10, src9, src11, src10, vec4, vec6); + DUP2_ARG2(__lsx_vilvh_b, src10, src9, src11, src10, vec5, vec7); + DUP2_ARG2(__lsx_vdp2_h_bu, vec4, filt0, vec5, filt0, tmp4, tmp5); + DUP2_ARG2(__lsx_vsrari_h, tmp4, FILTER_BITS, tmp5, FILTER_BITS, tmp4, tmp5); + tmp = __lsx_vpickev_b(tmp5, tmp4); + __lsx_vst(tmp, dst, 48); + + DUP2_ARG2(__lsx_vdp2_h_bu, vec6, filt0, vec7, filt0, tmp6, tmp7); + DUP2_ARG2(__lsx_vsrari_h, tmp6, FILTER_BITS, tmp7, FILTER_BITS, tmp6, tmp7); + tmp = __lsx_vpickev_b(tmp7, tmp6); + __lsx_vst(tmp, dst_tmp1, 48); + dst += dst_stride2; + dst_tmp1 += dst_stride2; + + src0 = src2; + src3 = src5; + src6 = src8; + src9 = src11; + } +} + +void vpx_convolve8_vert_lsx(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const InterpKernel *filter, int x0_q4, + int32_t x_step_q4, int y0_q4, int y_step_q4, int w, + int h) { + const int16_t *const filter_y = filter[y0_q4]; + int8_t cnt, filt_ver[8]; + + assert(y_step_q4 == 16); + assert(((const int32_t *)filter_y)[1] != 0x800000); + + for (cnt = 8; cnt--;) { + filt_ver[cnt] = filter_y[cnt]; + } + + if (vpx_get_filter_taps(filter_y) == 2) { + switch (w) { + case 4: + common_vt_2t_4w_lsx(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + &filt_ver[3], h); + break; + case 8: + common_vt_2t_8w_lsx(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + &filt_ver[3], h); + break; + case 16: + common_vt_2t_16w_lsx(src, src_stride, dst, dst_stride, filter, y0_q4, + y_step_q4, w, h); + break; + case 32: + common_vt_2t_32w_lsx(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + &filt_ver[3], h); + break; + case 64: + common_vt_2t_64w_lsx(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + &filt_ver[3], h); + break; + default: + vpx_convolve8_vert_c(src, src_stride, dst, dst_stride, filter, x0_q4, + x_step_q4, y0_q4, y_step_q4, w, h); + break; + } + } else { + switch (w) { + case 4: + common_vt_8t_4w_lsx(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + filt_ver, h); + break; + case 8: + common_vt_8t_8w_lsx(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + filt_ver, h); + break; + case 16: + common_vt_8t_16w_lsx(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + filt_ver, h); + break; + case 32: + common_vt_8t_32w_lsx(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + filt_ver, h); + break; + case 64: + common_vt_8t_64w_lsx(src, (int32_t)src_stride, dst, (int32_t)dst_stride, + filt_ver, h); + break; + default: + vpx_convolve8_vert_c(src, src_stride, dst, dst_stride, filter, x0_q4, + x_step_q4, y0_q4, y_step_q4, w, h); + break; + } + } +} diff --git a/vpx_dsp/loongarch/vpx_convolve_lsx.h b/vpx_dsp/loongarch/vpx_convolve_lsx.h new file mode 100644 index 0000000..2fdb93d --- /dev/null +++ b/vpx_dsp/loongarch/vpx_convolve_lsx.h @@ -0,0 +1,117 @@ +/* + * 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. + */ + +#ifndef VPX_VPX_DSP_LOONGARCH_VPX_CONVOLVE_LSX_H_ +#define VPX_VPX_DSP_LOONGARCH_VPX_CONVOLVE_LSX_H_ + +#include "vpx_util/loongson_intrinsics.h" +#include "vpx_dsp/vpx_filter.h" + +#define FILT_8TAP_DPADD_S_H(_reg0, _reg1, _reg2, _reg3, _filter0, _filter1, \ + _filter2, _filter3) \ + ({ \ + __m128i _vec0, _vec1; \ + \ + _vec0 = __lsx_vdp2_h_b(_reg0, _filter0); \ + _vec0 = __lsx_vdp2add_h_b(_vec0, _reg1, _filter1); \ + _vec1 = __lsx_vdp2_h_b(_reg2, _filter2); \ + _vec1 = __lsx_vdp2add_h_b(_vec1, _reg3, _filter3); \ + _vec0 = __lsx_vsadd_h(_vec0, _vec1); \ + \ + _vec0; \ + }) + +#define HORIZ_8TAP_FILT(_src0, _src1, _mask0, _mask1, _mask2, _mask3, \ + _filt_h0, _filt_h1, _filt_h2, _filt_h3) \ + ({ \ + __m128i _tmp0, _tmp1, _tmp2, _tmp3; \ + __m128i _out; \ + \ + DUP4_ARG3(__lsx_vshuf_b, _src1, _src0, _mask0, _src1, _src0, _mask1, \ + _src1, _src0, _mask2, _src1, _src0, _mask3, _tmp0, _tmp1, _tmp2, \ + _tmp3); \ + _out = FILT_8TAP_DPADD_S_H(_tmp0, _tmp1, _tmp2, _tmp3, _filt_h0, _filt_h1, \ + _filt_h2, _filt_h3); \ + _out = __lsx_vsrari_h(_out, FILTER_BITS); \ + _out = __lsx_vsat_h(_out, 7); \ + \ + _out; \ + }) + +#define HORIZ_8TAP_4WID_4VECS_FILT(_src0, _src1, _src2, _src3, _mask0, _mask1, \ + _mask2, _mask3, _filter0, _filter1, \ + _filter2, _filter3, _out0, _out1) \ + { \ + __m128i _tmp0, _tmp1, _tmp2, _tmp3, _tmp4, _tmp5, _tmp6, _tmp7; \ + __m128i _reg0, _reg1, _reg2, _reg3; \ + \ + DUP2_ARG3(__lsx_vshuf_b, _src1, _src0, _mask0, _src3, _src2, _mask0, \ + _tmp0, _tmp1); \ + DUP2_ARG2(__lsx_vdp2_h_b, _tmp0, _filter0, _tmp1, _filter0, _reg0, _reg1); \ + DUP2_ARG3(__lsx_vshuf_b, _src1, _src0, _mask1, _src3, _src2, _mask1, \ + _tmp2, _tmp3); \ + DUP2_ARG3(__lsx_vdp2add_h_b, _reg0, _tmp2, _filter1, _reg1, _tmp3, \ + _filter1, _reg0, _reg1); \ + DUP2_ARG3(__lsx_vshuf_b, _src1, _src0, _mask2, _src3, _src2, _mask2, \ + _tmp4, _tmp5); \ + DUP2_ARG2(__lsx_vdp2_h_b, _tmp4, _filter2, _tmp5, _filter2, _reg2, _reg3); \ + DUP2_ARG3(__lsx_vshuf_b, _src1, _src0, _mask3, _src3, _src2, _mask3, \ + _tmp6, _tmp7); \ + DUP2_ARG3(__lsx_vdp2add_h_b, _reg2, _tmp6, _filter3, _reg3, _tmp7, \ + _filter3, _reg2, _reg3); \ + DUP2_ARG2(__lsx_vsadd_h, _reg0, _reg2, _reg1, _reg3, _out0, _out1); \ + } + +#define HORIZ_8TAP_8WID_4VECS_FILT( \ + _src0, _src1, _src2, _src3, _mask0, _mask1, _mask2, _mask3, _filter0, \ + _filter1, _filter2, _filter3, _out0, _out1, _out2, _out3) \ + { \ + __m128i _tmp0, _tmp1, _tmp2, _tmp3, _tmp4, _tmp5, _tmp6, _tmp7; \ + __m128i _reg0, _reg1, _reg2, _reg3, _reg4, _reg5, _reg6, _reg7; \ + \ + DUP4_ARG3(__lsx_vshuf_b, _src0, _src0, _mask0, _src1, _src1, _mask0, \ + _src2, _src2, _mask0, _src3, _src3, _mask0, _tmp0, _tmp1, _tmp2, \ + _tmp3); \ + DUP4_ARG2(__lsx_vdp2_h_b, _tmp0, _filter0, _tmp1, _filter0, _tmp2, \ + _filter0, _tmp3, _filter0, _reg0, _reg1, _reg2, _reg3); \ + DUP4_ARG3(__lsx_vshuf_b, _src0, _src0, _mask2, _src1, _src1, _mask2, \ + _src2, _src2, _mask2, _src3, _src3, _mask2, _tmp0, _tmp1, _tmp2, \ + _tmp3); \ + DUP4_ARG2(__lsx_vdp2_h_b, _tmp0, _filter2, _tmp1, _filter2, _tmp2, \ + _filter2, _tmp3, _filter2, _reg4, _reg5, _reg6, _reg7); \ + DUP4_ARG3(__lsx_vshuf_b, _src0, _src0, _mask1, _src1, _src1, _mask1, \ + _src2, _src2, _mask1, _src3, _src3, _mask1, _tmp4, _tmp5, _tmp6, \ + _tmp7); \ + DUP4_ARG3(__lsx_vdp2add_h_b, _reg0, _tmp4, _filter1, _reg1, _tmp5, \ + _filter1, _reg2, _tmp6, _filter1, _reg3, _tmp7, _filter1, _reg0, \ + _reg1, _reg2, _reg3); \ + DUP4_ARG3(__lsx_vshuf_b, _src0, _src0, _mask3, _src1, _src1, _mask3, \ + _src2, _src2, _mask3, _src3, _src3, _mask3, _tmp4, _tmp5, _tmp6, \ + _tmp7); \ + DUP4_ARG3(__lsx_vdp2add_h_b, _reg4, _tmp4, _filter3, _reg5, _tmp5, \ + _filter3, _reg6, _tmp6, _filter3, _reg7, _tmp7, _filter3, _reg4, \ + _reg5, _reg6, _reg7); \ + DUP4_ARG2(__lsx_vsadd_h, _reg0, _reg4, _reg1, _reg5, _reg2, _reg6, _reg3, \ + _reg7, _out0, _out1, _out2, _out3); \ + } + +#define HORIZ_2TAP_FILT_UH(in0, in1, mask, coeff, shift) \ + ({ \ + __m128i tmp0_m; \ + __m128i tmp1_m; \ + \ + tmp0_m = __lsx_vshuf_b(in1, in0, mask); \ + tmp1_m = __lsx_vdp2_h_bu(tmp0_m, coeff); \ + tmp1_m = __lsx_vsrari_h(tmp1_m, shift); \ + \ + tmp1_m; \ + }) + +#endif // VPX_VPX_DSP_LOONGARCH_VPX_CONVOLVE_LSX_H_ diff --git a/vpx_dsp/vpx_dsp.mk b/vpx_dsp/vpx_dsp.mk index 976c652..91a983d 100644 --- a/vpx_dsp/vpx_dsp.mk +++ b/vpx_dsp/vpx_dsp.mk @@ -163,6 +163,12 @@ DSP_SRCS-$(HAVE_DSPR2) += mips/convolve8_vert_dspr2.c DSP_SRCS-$(HAVE_VSX) += ppc/vpx_convolve_vsx.c +# common (lsx) +DSP_SRCS-$(HAVE_LSX) += loongarch/vpx_convolve8_horiz_lsx.c +DSP_SRCS-$(HAVE_LSX) += loongarch/vpx_convolve8_vert_lsx.c +DSP_SRCS-$(HAVE_LSX) += loongarch/vpx_convolve8_lsx.c +DSP_SRCS-$(HAVE_LSX) += loongarch/vpx_convolve_lsx.h + # loop filters DSP_SRCS-yes += loopfilter.c diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl index ce0780f..49eea44 100644 --- a/vpx_dsp/vpx_dsp_rtcd_defs.pl +++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl @@ -374,13 +374,13 @@ add_proto qw/void vpx_convolve_avg/, "const uint8_t *src, ptrdiff_t src_stride, specialize qw/vpx_convolve_avg neon dspr2 msa sse2 vsx mmi/; add_proto qw/void vpx_convolve8/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h"; -specialize qw/vpx_convolve8 sse2 ssse3 avx2 neon dspr2 msa vsx mmi/; +specialize qw/vpx_convolve8 sse2 ssse3 avx2 neon dspr2 msa vsx mmi lsx/; add_proto qw/void vpx_convolve8_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h"; -specialize qw/vpx_convolve8_horiz sse2 ssse3 avx2 neon dspr2 msa vsx mmi/; +specialize qw/vpx_convolve8_horiz sse2 ssse3 avx2 neon dspr2 msa vsx mmi lsx/; add_proto qw/void vpx_convolve8_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h"; -specialize qw/vpx_convolve8_vert sse2 ssse3 avx2 neon dspr2 msa vsx mmi/; +specialize qw/vpx_convolve8_vert sse2 ssse3 avx2 neon dspr2 msa vsx mmi lsx/; add_proto qw/void vpx_convolve8_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h"; specialize qw/vpx_convolve8_avg sse2 ssse3 avx2 neon dspr2 msa vsx mmi/; -- 2.7.4