/* * Copyright (c) 2026, Alliance for Open Media. All rights reserved. * * This source code is subject to the terms of the BSD 2 Clause License and * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License * was not distributed with this source code in the LICENSE file, you can * obtain it at www.aomedia.org/license/software. If the Alliance for Open * Media Patent License 1.0 was not distributed with this source code in the * PATENTS file, you can obtain it at www.aomedia.org/license/patent. */ #include #include #include "config/aom_config.h" #include "config/av1_rtcd.h" #include "av1/common/convolve.h" #include "av1/common/filter.h" #include "av1/common/riscv/highbd_compound_convolve_rvv.h" static inline vuint16mf2_t highbd_convolve6_4_rvv( const vint16mf2_t s0, const vint16mf2_t s1, const vint16mf2_t s2, const vint16mf2_t s3, const vint16mf2_t s4, const vint16mf2_t s5, const int16_t *filter, const vint32m1_t offset_vec, const int32_t round_bits, const size_t vl) { // Values at indices 0 and 7 of filter are zero. vint32m1_t sum = __riscv_vwmacc_vx_i32m1(offset_vec, filter[1], s0, vl); sum = __riscv_vwmacc_vx_i32m1(sum, filter[2], s1, vl); sum = __riscv_vwmacc_vx_i32m1(sum, filter[3], s2, vl); sum = __riscv_vwmacc_vx_i32m1(sum, filter[4], s3, vl); sum = __riscv_vwmacc_vx_i32m1(sum, filter[5], s4, vl); sum = __riscv_vwmacc_vx_i32m1(sum, filter[6], s5, vl); vuint32m1_t sum_u32 = __riscv_vreinterpret_v_i32m1_u32m1(__riscv_vmax_vx_i32m1(sum, 0, vl)); #if __riscv_v_intrinsic >= 12000 vuint16mf2_t res = __riscv_vnclipu_wx_u16mf2(sum_u32, round_bits, __RISCV_VXRM_RDN, vl); #else vuint16mf2_t res = __riscv_vnsrl_wx_u16mf2(sum_u32, round_bits, vl); #endif return res; } static inline vuint16m1_t highbd_convolve6_8_rvv( const vint16m1_t s0, const vint16m1_t s1, const vint16m1_t s2, const vint16m1_t s3, const vint16m1_t s4, const vint16m1_t s5, const int16_t *filter, const vint32m2_t offset_vec, const int32_t round_bits, const size_t vl) { // Values at indices 0 and 7 of filter are zero. vint32m2_t sum = __riscv_vwmacc_vx_i32m2(offset_vec, filter[1], s0, vl); sum = __riscv_vwmacc_vx_i32m2(sum, filter[2], s1, vl); sum = __riscv_vwmacc_vx_i32m2(sum, filter[3], s2, vl); sum = __riscv_vwmacc_vx_i32m2(sum, filter[4], s3, vl); sum = __riscv_vwmacc_vx_i32m2(sum, filter[5], s4, vl); sum = __riscv_vwmacc_vx_i32m2(sum, filter[6], s5, vl); vuint32m2_t sum_u32 = __riscv_vreinterpret_v_i32m2_u32m2(__riscv_vmax_vx_i32m2(sum, 0, vl)); #if __riscv_v_intrinsic >= 12000 vuint16m1_t res = __riscv_vnclipu_wx_u16m1(sum_u32, round_bits, __RISCV_VXRM_RDN, vl); #else vuint16m1_t res = __riscv_vnsrl_wx_u16m1(sum_u32, round_bits, vl); #endif return res; } static inline void highbd_dist_wtd_convolve_x_6tap_rvv( const uint16_t *src_ptr, int src_stride, uint16_t *dst_ptr, int dst_stride, int w, int h, const int16_t *x_filter_ptr, const int offset, const int round_bits) { int height = h; const size_t vl = __riscv_vsetvl_e16m1(w); vint32m2_t offset_vec = __riscv_vmv_v_x_i32m2(offset, vl); do { int width = w; const int16_t *s = (const int16_t *)src_ptr; uint16_t *d = dst_ptr; do { vint16m1_t s00, s01, s02, s03, s04, s05; vint16m1_t s10, s11, s12, s13, s14, s15; vint16m1_t s20, s21, s22, s23, s24, s25; vint16m1_t s30, s31, s32, s33, s34, s35; // Load 6 consecutive 8-element vectors from 4 rows load_s16_8x6(s + 0 * src_stride, 1, &s00, &s01, &s02, &s03, &s04, &s05, vl); load_s16_8x6(s + 1 * src_stride, 1, &s10, &s11, &s12, &s13, &s14, &s15, vl); load_s16_8x6(s + 2 * src_stride, 1, &s20, &s21, &s22, &s23, &s24, &s25, vl); load_s16_8x6(s + 3 * src_stride, 1, &s30, &s31, &s32, &s33, &s34, &s35, vl); // Convolve each row vuint16m1_t d0 = highbd_convolve6_8_rvv(s00, s01, s02, s03, s04, s05, x_filter_ptr, offset_vec, round_bits, vl); vuint16m1_t d1 = highbd_convolve6_8_rvv(s10, s11, s12, s13, s14, s15, x_filter_ptr, offset_vec, round_bits, vl); vuint16m1_t d2 = highbd_convolve6_8_rvv(s20, s21, s22, s23, s24, s25, x_filter_ptr, offset_vec, round_bits, vl); vuint16m1_t d3 = highbd_convolve6_8_rvv(s30, s31, s32, s33, s34, s35, x_filter_ptr, offset_vec, round_bits, vl); // Store results with stride store_u16_8x4(d, dst_stride, d0, d1, d2, d3, vl); s += vl; d += vl; width -= vl; } while (width > 0); src_ptr += 4 * src_stride; dst_ptr += 4 * dst_stride; height -= 4; } while (height != 0); } static inline vuint16mf2_t highbd_convolve8_4_rvv( const vint16mf2_t s0, const vint16mf2_t s1, const vint16mf2_t s2, const vint16mf2_t s3, const vint16mf2_t s4, const vint16mf2_t s5, const vint16mf2_t s6, const vint16mf2_t s7, const int16_t *filter, const vint32m1_t offset_vec, const int32_t round_bits, const size_t vl) { vint32m1_t sum = __riscv_vwmacc_vx_i32m1(offset_vec, filter[0], s0, vl); sum = __riscv_vwmacc_vx_i32m1(sum, filter[1], s1, vl); sum = __riscv_vwmacc_vx_i32m1(sum, filter[2], s2, vl); sum = __riscv_vwmacc_vx_i32m1(sum, filter[3], s3, vl); sum = __riscv_vwmacc_vx_i32m1(sum, filter[4], s4, vl); sum = __riscv_vwmacc_vx_i32m1(sum, filter[5], s5, vl); sum = __riscv_vwmacc_vx_i32m1(sum, filter[6], s6, vl); sum = __riscv_vwmacc_vx_i32m1(sum, filter[7], s7, vl); vuint32m1_t sum_u32 = __riscv_vreinterpret_v_i32m1_u32m1(__riscv_vmax_vx_i32m1(sum, 0, vl)); #if __riscv_v_intrinsic >= 12000 vuint16mf2_t res = __riscv_vnclipu_wx_u16mf2(sum_u32, round_bits, __RISCV_VXRM_RDN, vl); #else vuint16mf2_t res = __riscv_vnsrl_wx_u16mf2(sum_u32, round_bits, vl); #endif return res; } static inline vuint16m1_t highbd_convolve8_8_rvv( const vint16m1_t s0, const vint16m1_t s1, const vint16m1_t s2, const vint16m1_t s3, const vint16m1_t s4, const vint16m1_t s5, const vint16m1_t s6, const vint16m1_t s7, const int16_t *filter, const vint32m2_t offset_vec, const int32_t round_bits, const size_t vl) { vint32m2_t sum = __riscv_vwmacc_vx_i32m2(offset_vec, filter[0], s0, vl); sum = __riscv_vwmacc_vx_i32m2(sum, filter[1], s1, vl); sum = __riscv_vwmacc_vx_i32m2(sum, filter[2], s2, vl); sum = __riscv_vwmacc_vx_i32m2(sum, filter[3], s3, vl); sum = __riscv_vwmacc_vx_i32m2(sum, filter[4], s4, vl); sum = __riscv_vwmacc_vx_i32m2(sum, filter[5], s5, vl); sum = __riscv_vwmacc_vx_i32m2(sum, filter[6], s6, vl); sum = __riscv_vwmacc_vx_i32m2(sum, filter[7], s7, vl); vuint32m2_t sum_u32 = __riscv_vreinterpret_v_i32m2_u32m2(__riscv_vmax_vx_i32m2(sum, 0, vl)); #if __riscv_v_intrinsic >= 12000 vuint16m1_t res = __riscv_vnclipu_wx_u16m1(sum_u32, round_bits, __RISCV_VXRM_RDN, vl); #else vuint16m1_t res = __riscv_vnsrl_wx_u16m1(sum_u32, round_bits, vl); #endif return res; } static inline vuint16mf2_t highbd_convolve4_4_x_rvv( const vint16mf2_t s0, const vint16mf2_t s1, const vint16mf2_t s2, const vint16mf2_t s3, const int16_t *filter, const vint32m1_t offset_vec, const int round_bits, const size_t vl) { vint32m1_t sum = __riscv_vwmacc_vx_i32m1(offset_vec, filter[0], s0, vl); sum = __riscv_vwmacc_vx_i32m1(sum, filter[1], s1, vl); sum = __riscv_vwmacc_vx_i32m1(sum, filter[2], s2, vl); sum = __riscv_vwmacc_vx_i32m1(sum, filter[3], s3, vl); vuint32m1_t sum_u32 = __riscv_vreinterpret_v_i32m1_u32m1(__riscv_vmax_vx_i32m1(sum, 0, vl)); #if __riscv_v_intrinsic >= 12000 vuint16mf2_t res = __riscv_vnclipu_wx_u16mf2(sum_u32, round_bits, __RISCV_VXRM_RDN, vl); #else vuint16mf2_t res = __riscv_vnsrl_wx_u16mf2(sum_u32, round_bits, vl); #endif return res; } static inline void highbd_dist_wtd_convolve_x_rvv( const uint16_t *src_ptr, int src_stride, uint16_t *dst_ptr, int dst_stride, int w, int h, const int16_t *x_filter_ptr, const int offset, const int round_bits) { size_t vl = __riscv_vsetvl_e16m1(w); if (w == 4) { // 4-tap filters are used for blocks having width == 4. const int16_t *filter = x_filter_ptr + 2; // Skip first 2 taps const int16_t *s = (const int16_t *)(src_ptr + 2); uint16_t *d = dst_ptr; vint32m1_t offset_vec = __riscv_vmv_v_x_i32m1(offset, vl); do { // Load 4 taps for 4 rows vint16mf2_t s00, s01, s02, s03; vint16mf2_t s10, s11, s12, s13; vint16mf2_t s20, s21, s22, s23; vint16mf2_t s30, s31, s32, s33; load_s16_4x4(s + 0 * src_stride, 1, &s00, &s01, &s02, &s03, vl); load_s16_4x4(s + 1 * src_stride, 1, &s10, &s11, &s12, &s13, vl); load_s16_4x4(s + 2 * src_stride, 1, &s20, &s21, &s22, &s23, vl); load_s16_4x4(s + 3 * src_stride, 1, &s30, &s31, &s32, &s33, vl); // Convolve each row vuint16mf2_t d0 = highbd_convolve4_4_x_rvv(s00, s01, s02, s03, filter, offset_vec, round_bits, vl); vuint16mf2_t d1 = highbd_convolve4_4_x_rvv(s10, s11, s12, s13, filter, offset_vec, round_bits, vl); vuint16mf2_t d2 = highbd_convolve4_4_x_rvv(s20, s21, s22, s23, filter, offset_vec, round_bits, vl); vuint16mf2_t d3 = highbd_convolve4_4_x_rvv(s30, s31, s32, s33, filter, offset_vec, round_bits, vl); // Store results store_u16_4x4(d, dst_stride, d0, d1, d2, d3, vl); s += 4 * src_stride; d += 4 * dst_stride; h -= 4; } while (h != 0); } else { // 8-tap filter path const int16_t *filter = x_filter_ptr; int height = h; vint32m2_t offset_vec = __riscv_vmv_v_x_i32m2(offset, vl); do { int width = w; const int16_t *s = (const int16_t *)src_ptr; uint16_t *d = dst_ptr; do { vint16m1_t s00, s01, s02, s03, s04, s05, s06, s07; vint16m1_t s10, s11, s12, s13, s14, s15, s16, s17; vint16m1_t s20, s21, s22, s23, s24, s25, s26, s27; vint16m1_t s30, s31, s32, s33, s34, s35, s36, s37; // Load elements for each of 4 rows load_s16_8x8(s + 0 * src_stride, 1, &s00, &s01, &s02, &s03, &s04, &s05, &s06, &s07, vl); load_s16_8x8(s + 1 * src_stride, 1, &s10, &s11, &s12, &s13, &s14, &s15, &s16, &s17, vl); load_s16_8x8(s + 2 * src_stride, 1, &s20, &s21, &s22, &s23, &s24, &s25, &s26, &s27, vl); load_s16_8x8(s + 3 * src_stride, 1, &s30, &s31, &s32, &s33, &s34, &s35, &s36, &s37, vl); // Convolve each row vuint16m1_t d0 = highbd_convolve8_8_rvv(s00, s01, s02, s03, s04, s05, s06, s07, filter, offset_vec, round_bits, vl); vuint16m1_t d1 = highbd_convolve8_8_rvv(s10, s11, s12, s13, s14, s15, s16, s17, filter, offset_vec, round_bits, vl); vuint16m1_t d2 = highbd_convolve8_8_rvv(s20, s21, s22, s23, s24, s25, s26, s27, filter, offset_vec, round_bits, vl); vuint16m1_t d3 = highbd_convolve8_8_rvv(s30, s31, s32, s33, s34, s35, s36, s37, filter, offset_vec, round_bits, vl); // Store results store_u16_8x4(d, dst_stride, d0, d1, d2, d3, vl); s += vl; d += vl; width -= vl; } while (width > 0); src_ptr += 4 * src_stride; dst_ptr += 4 * dst_stride; height -= 4; } while (height != 0); } } void av1_highbd_dist_wtd_convolve_x_rvv( const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd) { DECLARE_ALIGNED(16, uint16_t, im_block[(MAX_SB_SIZE + MAX_FILTER_TAP) * MAX_SB_SIZE]); CONV_BUF_TYPE *dst16 = conv_params->dst; const int x_filter_taps = get_filter_tap(filter_params_x, subpel_x_qn); int dst16_stride = conv_params->dst_stride; const int im_stride = MAX_SB_SIZE; const int horiz_offset = filter_params_x->taps / 2 - 1; assert(FILTER_BITS == COMPOUND_ROUND1_BITS); const int offset_convolve = (1 << (conv_params->round_0 - 1)) + (1 << (bd + FILTER_BITS)) + (1 << (bd + FILTER_BITS - 1)); const int16_t *x_filter_ptr = av1_get_interp_filter_subpel_kernel( filter_params_x, subpel_x_qn & SUBPEL_MASK); src -= horiz_offset; int round_bits, shift_bits, offset_bits; // horizontal filter if (bd == 12) { round_bits = ROUND0_BITS + 2; shift_bits = ROUND_SHIFT - 2; offset_bits = bd + 2 * FILTER_BITS - ROUND0_BITS - 2; } else { round_bits = ROUND0_BITS; shift_bits = ROUND_SHIFT; offset_bits = bd + 2 * FILTER_BITS - ROUND0_BITS; } if (conv_params->do_average) { if (x_filter_taps <= 6 && w != 4) { highbd_dist_wtd_convolve_x_6tap_rvv(src + 1, src_stride, im_block, im_stride, w, h, x_filter_ptr, offset_convolve, round_bits); } else { highbd_dist_wtd_convolve_x_rvv(src, src_stride, im_block, im_stride, w, h, x_filter_ptr, offset_convolve, round_bits); } if (conv_params->use_dist_wtd_comp_avg) { highbd_dist_wtd_comp_avg_rvv(im_block, im_stride, dst, dst_stride, w, h, conv_params, bd, shift_bits, offset_bits); } else { highbd_comp_avg_rvv(im_block, im_stride, dst, dst_stride, w, h, conv_params, bd, shift_bits, offset_bits); } } else { if (x_filter_taps <= 6 && w != 4) { highbd_dist_wtd_convolve_x_6tap_rvv(src + 1, src_stride, dst16, dst16_stride, w, h, x_filter_ptr, offset_convolve, round_bits); } else { highbd_dist_wtd_convolve_x_rvv(src, src_stride, dst16, dst16_stride, w, h, x_filter_ptr, offset_convolve, round_bits); } } } static inline void highbd_dist_wtd_convolve_y_6tap_rvv( const uint16_t *src_ptr, int src_stride, uint16_t *dst_ptr, int dst_stride, int w, int h, const int16_t *y_filter_ptr, const int offset, const int round_bits) { const int16_t *filter = y_filter_ptr; size_t vl = __riscv_vsetvl_e16m1(w); if (w == 4) { const int16_t *s = (const int16_t *)src_ptr; uint16_t *d = dst_ptr; vint32m1_t offset_vec = __riscv_vmv_v_x_i32m1(offset, vl); // Load initial 5 rows for 6-tap filter vint16mf2_t s0, s1, s2, s3, s4; load_s16_4x5(s, src_stride, &s0, &s1, &s2, &s3, &s4, vl); s += 5 * src_stride; do { // Load next 4 rows vint16mf2_t s5, s6, s7, s8; load_s16_4x4(s, src_stride, &s5, &s6, &s7, &s8, vl); // Convolve 4 output rows vuint16mf2_t d0 = highbd_convolve6_4_rvv(s0, s1, s2, s3, s4, s5, filter, offset_vec, round_bits, vl); vuint16mf2_t d1 = highbd_convolve6_4_rvv(s1, s2, s3, s4, s5, s6, filter, offset_vec, round_bits, vl); vuint16mf2_t d2 = highbd_convolve6_4_rvv(s2, s3, s4, s5, s6, s7, filter, offset_vec, round_bits, vl); vuint16mf2_t d3 = highbd_convolve6_4_rvv(s3, s4, s5, s6, s7, s8, filter, offset_vec, round_bits, vl); store_u16_4x4(d, dst_stride, d0, d1, d2, d3, vl); // Shift window: s0-s4 become s4-s8 for next iteration s0 = s4; s1 = s5; s2 = s6; s3 = s7; s4 = s8; s += 4 * src_stride; d += 4 * dst_stride; h -= 4; } while (h != 0); } else { vint32m2_t offset_vec = __riscv_vmv_v_x_i32m2(offset, vl); do { int height = h; const int16_t *s = (const int16_t *)src_ptr; uint16_t *d = dst_ptr; // Load initial 5 rows for 6-tap filter vint16m1_t s0, s1, s2, s3, s4; load_s16_8x5(s, src_stride, &s0, &s1, &s2, &s3, &s4, vl); s += 5 * src_stride; do { // Load next 4 rows vint16m1_t s5, s6, s7, s8; load_s16_8x4(s, src_stride, &s5, &s6, &s7, &s8, vl); // Convolve 4 output rows vuint16m1_t d0 = highbd_convolve6_8_rvv(s0, s1, s2, s3, s4, s5, filter, offset_vec, round_bits, vl); vuint16m1_t d1 = highbd_convolve6_8_rvv(s1, s2, s3, s4, s5, s6, filter, offset_vec, round_bits, vl); vuint16m1_t d2 = highbd_convolve6_8_rvv(s2, s3, s4, s5, s6, s7, filter, offset_vec, round_bits, vl); vuint16m1_t d3 = highbd_convolve6_8_rvv(s3, s4, s5, s6, s7, s8, filter, offset_vec, round_bits, vl); store_u16_8x4(d, dst_stride, d0, d1, d2, d3, vl); // Shift window: s0-s4 become s4-s8 for next iteration s0 = s4; s1 = s5; s2 = s6; s3 = s7; s4 = s8; s += 4 * src_stride; d += 4 * dst_stride; height -= 4; } while (height != 0); src_ptr += vl; dst_ptr += vl; w -= vl; } while (w > 0); } } static inline vuint16mf2_t highbd_convolve4_4_rvv( const vint16mf2_t s0, const vint16mf2_t s1, const vint16mf2_t s2, const vint16mf2_t s3, const int16_t *filter, const int32_t offset, const int32_t round_bits, size_t vl) { vint32m1_t sum = __riscv_vwmul_vx_i32m1(s0, filter[0], vl); sum = __riscv_vwmacc_vx_i32m1(sum, filter[1], s1, vl); sum = __riscv_vwmacc_vx_i32m1(sum, filter[2], s2, vl); sum = __riscv_vwmacc_vx_i32m1(sum, filter[3], s3, vl); sum = __riscv_vadd_vx_i32m1(sum, offset, vl); vuint32m1_t sum_u32 = __riscv_vreinterpret_v_i32m1_u32m1(__riscv_vmax_vx_i32m1(sum, 0, vl)); #if __riscv_v_intrinsic >= 12000 vuint16mf2_t res = __riscv_vnclipu_wx_u16mf2(sum_u32, round_bits, __RISCV_VXRM_RDN, vl); #else vuint16mf2_t res = __riscv_vnsrl_wx_u16mf2(sum_u32, round_bits, vl); #endif return res; } static inline vuint16m1_t highbd_convolve4_8_rvv( const vint16m1_t s0, const vint16m1_t s1, const vint16m1_t s2, const vint16m1_t s3, const int16_t *filter, const int32_t offset, const int32_t round_bits, size_t vl) { vint32m2_t sum = __riscv_vwmul_vx_i32m2(s0, filter[0], vl); sum = __riscv_vwmacc_vx_i32m2(sum, filter[1], s1, vl); sum = __riscv_vwmacc_vx_i32m2(sum, filter[2], s2, vl); sum = __riscv_vwmacc_vx_i32m2(sum, filter[3], s3, vl); sum = __riscv_vadd_vx_i32m2(sum, offset, vl); vuint32m2_t sum_u32 = __riscv_vreinterpret_v_i32m2_u32m2(__riscv_vmax_vx_i32m2(sum, 0, vl)); #if __riscv_v_intrinsic >= 12000 vuint16m1_t res = __riscv_vnclipu_wx_u16m1(sum_u32, round_bits, __RISCV_VXRM_RDN, vl); #else vuint16m1_t res = __riscv_vnsrl_wx_u16m1(sum_u32, round_bits, vl); #endif return res; } static inline void highbd_dist_wtd_convolve_y_4tap_rvv( const uint16_t *src_ptr, int src_stride, uint16_t *dst_ptr, int dst_stride, int w, int h, const int16_t *y_filter_ptr, const int offset, const int round_bits) { const int16_t *filter = y_filter_ptr + 2; size_t vl = __riscv_vsetvl_e16m1(w); if (w == 4) { const int16_t *s = (const int16_t *)src_ptr; uint16_t *d = dst_ptr; // Load initial 3 rows vint16mf2_t s0, s1, s2; load_s16_4x3(s, src_stride, &s0, &s1, &s2, vl); s += 3 * src_stride; do { // Load next 4 rows vint16mf2_t s3, s4, s5, s6; load_s16_4x4(s, src_stride, &s3, &s4, &s5, &s6, vl); // Convolve 4 output rows vuint16mf2_t d0 = highbd_convolve4_4_rvv(s0, s1, s2, s3, filter, offset, round_bits, vl); vuint16mf2_t d1 = highbd_convolve4_4_rvv(s1, s2, s3, s4, filter, offset, round_bits, vl); vuint16mf2_t d2 = highbd_convolve4_4_rvv(s2, s3, s4, s5, filter, offset, round_bits, vl); vuint16mf2_t d3 = highbd_convolve4_4_rvv(s3, s4, s5, s6, filter, offset, round_bits, vl); store_u16_4x4(d, dst_stride, d0, d1, d2, d3, vl); // Shift window s0 = s4; s1 = s5; s2 = s6; s += 4 * src_stride; d += 4 * dst_stride; h -= 4; } while (h != 0); } else { do { int height = h; const int16_t *s = (const int16_t *)src_ptr; uint16_t *d = dst_ptr; // Load initial 3 rows vint16m1_t s0, s1, s2; load_s16_8x3(s, src_stride, &s0, &s1, &s2, vl); s += 3 * src_stride; do { // Load next 4 rows vint16m1_t s3, s4, s5, s6; load_s16_8x4(s, src_stride, &s3, &s4, &s5, &s6, vl); // Convolve 4 output rows vuint16m1_t d0 = highbd_convolve4_8_rvv(s0, s1, s2, s3, filter, offset, round_bits, vl); vuint16m1_t d1 = highbd_convolve4_8_rvv(s1, s2, s3, s4, filter, offset, round_bits, vl); vuint16m1_t d2 = highbd_convolve4_8_rvv(s2, s3, s4, s5, filter, offset, round_bits, vl); vuint16m1_t d3 = highbd_convolve4_8_rvv(s3, s4, s5, s6, filter, offset, round_bits, vl); store_u16_8x4(d, dst_stride, d0, d1, d2, d3, vl); // Shift window s0 = s4; s1 = s5; s2 = s6; s += 4 * src_stride; d += 4 * dst_stride; height -= 4; } while (height != 0); src_ptr += vl; dst_ptr += vl; w -= vl; } while (w > 0); } } static inline void highbd_dist_wtd_convolve_y_8tap_rvv( const uint16_t *src_ptr, int src_stride, uint16_t *dst_ptr, int dst_stride, int w, int h, const int16_t *y_filter_ptr, const int offset, const int32_t round_bits) { const int16_t *filter = y_filter_ptr; size_t vl = __riscv_vsetvl_e16m1(w); if (w == 4) { const int16_t *s = (const int16_t *)src_ptr; uint16_t *d = dst_ptr; vint32m1_t offset_vec = __riscv_vmv_v_x_i32m1(offset, vl); // Load initial 7 rows vint16mf2_t s0, s1, s2, s3, s4, s5, s6; load_s16_4x7(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6, vl); s += 7 * src_stride; do { // Load next 4 rows vint16mf2_t s7, s8, s9, s10; load_s16_4x4(s, src_stride, &s7, &s8, &s9, &s10, vl); // Convolve 4 output rows vuint16mf2_t d0 = highbd_convolve8_4_rvv( s0, s1, s2, s3, s4, s5, s6, s7, filter, offset_vec, round_bits, vl); vuint16mf2_t d1 = highbd_convolve8_4_rvv( s1, s2, s3, s4, s5, s6, s7, s8, filter, offset_vec, round_bits, vl); vuint16mf2_t d2 = highbd_convolve8_4_rvv( s2, s3, s4, s5, s6, s7, s8, s9, filter, offset_vec, round_bits, vl); vuint16mf2_t d3 = highbd_convolve8_4_rvv( s3, s4, s5, s6, s7, s8, s9, s10, filter, offset_vec, round_bits, vl); store_u16_4x4(d, dst_stride, d0, d1, d2, d3, vl); // Shift window s0 = s4; s1 = s5; s2 = s6; s3 = s7; s4 = s8; s5 = s9; s6 = s10; s += 4 * src_stride; d += 4 * dst_stride; h -= 4; } while (h != 0); } else { vint32m2_t offset_vec = __riscv_vmv_v_x_i32m2(offset, vl); do { int height = h; const int16_t *s = (const int16_t *)src_ptr; uint16_t *d = dst_ptr; // Load initial 7 rows vint16m1_t s0, s1, s2, s3, s4, s5, s6; load_s16_8x7(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6, vl); s += 7 * src_stride; do { // Load next 4 rows vint16m1_t s7, s8, s9, s10; load_s16_8x4(s, src_stride, &s7, &s8, &s9, &s10, vl); // Convolve 4 output rows vuint16m1_t d0 = highbd_convolve8_8_rvv( s0, s1, s2, s3, s4, s5, s6, s7, filter, offset_vec, round_bits, vl); vuint16m1_t d1 = highbd_convolve8_8_rvv( s1, s2, s3, s4, s5, s6, s7, s8, filter, offset_vec, round_bits, vl); vuint16m1_t d2 = highbd_convolve8_8_rvv( s2, s3, s4, s5, s6, s7, s8, s9, filter, offset_vec, round_bits, vl); vuint16m1_t d3 = highbd_convolve8_8_rvv(s3, s4, s5, s6, s7, s8, s9, s10, filter, offset_vec, round_bits, vl); store_u16_8x4(d, dst_stride, d0, d1, d2, d3, vl); // Shift window s0 = s4; s1 = s5; s2 = s6; s3 = s7; s4 = s8; s5 = s9; s6 = s10; s += 4 * src_stride; d += 4 * dst_stride; height -= 4; } while (height != 0); src_ptr += vl; dst_ptr += vl; w -= vl; } while (w > 0); } } void av1_highbd_dist_wtd_convolve_y_rvv( const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd) { DECLARE_ALIGNED(16, uint16_t, im_block[(MAX_SB_SIZE + MAX_FILTER_TAP) * MAX_SB_SIZE]); CONV_BUF_TYPE *dst16 = conv_params->dst; const int y_filter_taps = get_filter_tap(filter_params_y, subpel_y_qn); int dst16_stride = conv_params->dst_stride; const int im_stride = MAX_SB_SIZE; const int vert_offset = filter_params_y->taps / 2 - 1; assert(FILTER_BITS == COMPOUND_ROUND1_BITS); const int round_offset_conv = (1 << (conv_params->round_0 - 1)) + (1 << (bd + FILTER_BITS)) + (1 << (bd + FILTER_BITS - 1)); const int16_t *y_filter_ptr = av1_get_interp_filter_subpel_kernel( filter_params_y, subpel_y_qn & SUBPEL_MASK); src -= vert_offset * src_stride; int round_bits, shift_bits, offset_bits; if (bd == 12) { round_bits = ROUND0_BITS + 2; shift_bits = ROUND_SHIFT - 2; offset_bits = bd + 2 * FILTER_BITS - ROUND0_BITS - 2; } else { round_bits = ROUND0_BITS; shift_bits = ROUND_SHIFT; offset_bits = bd + 2 * FILTER_BITS - ROUND0_BITS; } if (conv_params->do_average) { if (y_filter_taps <= 4) { highbd_dist_wtd_convolve_y_4tap_rvv( src + 2 * src_stride, src_stride, im_block, im_stride, w, h, y_filter_ptr, round_offset_conv, round_bits); } else if (y_filter_taps == 6) { highbd_dist_wtd_convolve_y_6tap_rvv( src + src_stride, src_stride, im_block, im_stride, w, h, y_filter_ptr, round_offset_conv, round_bits); } else { highbd_dist_wtd_convolve_y_8tap_rvv(src, src_stride, im_block, im_stride, w, h, y_filter_ptr, round_offset_conv, round_bits); } if (conv_params->use_dist_wtd_comp_avg) { highbd_dist_wtd_comp_avg_rvv(im_block, im_stride, dst, dst_stride, w, h, conv_params, bd, shift_bits, offset_bits); } else { highbd_comp_avg_rvv(im_block, im_stride, dst, dst_stride, w, h, conv_params, bd, shift_bits, offset_bits); } } else { if (y_filter_taps <= 4) { highbd_dist_wtd_convolve_y_4tap_rvv( src + 2 * src_stride, src_stride, dst16, dst16_stride, w, h, y_filter_ptr, round_offset_conv, round_bits); } else if (y_filter_taps == 6) { highbd_dist_wtd_convolve_y_6tap_rvv(src + src_stride, src_stride, dst16, dst16_stride, w, h, y_filter_ptr, round_offset_conv, round_bits); } else { highbd_dist_wtd_convolve_y_8tap_rvv(src, src_stride, dst16, dst16_stride, w, h, y_filter_ptr, round_offset_conv, round_bits); } } } static inline void highbd_2d_copy_rvv(const uint16_t *src_ptr, int src_stride, uint16_t *dst_ptr, int dst_stride, int w, int h, int round_bits, const int offset) { assert(h % 4 == 0); const uint16_t round_val = (1 << round_bits); int height = h; if (w <= 4) { const size_t vl = __riscv_vsetvl_e16mf2(w); vuint16mf2_t vec_offset = __riscv_vmv_v_x_u16mf2(offset, vl); do { vuint16mf2_t s0, s1, s2, s3; load_u16_4x4(src_ptr, src_stride, &s0, &s1, &s2, &s3, vl); vuint16mf2_t d0 = __riscv_vmacc_vx_u16mf2(vec_offset, round_val, s0, vl); vuint16mf2_t d1 = __riscv_vmacc_vx_u16mf2(vec_offset, round_val, s1, vl); vuint16mf2_t d2 = __riscv_vmacc_vx_u16mf2(vec_offset, round_val, s2, vl); vuint16mf2_t d3 = __riscv_vmacc_vx_u16mf2(vec_offset, round_val, s3, vl); store_u16_4x4(dst_ptr, dst_stride, d0, d1, d2, d3, vl); src_ptr += 4 * src_stride; dst_ptr += 4 * dst_stride; height -= 4; } while (height > 0); } else { size_t vl_max = __riscv_vsetvlmax_e16m1(); vuint16m1_t vec_offset = __riscv_vmv_v_x_u16m1(offset, vl_max); do { const uint16_t *s = src_ptr; uint16_t *d = dst_ptr; int width = w; do { size_t vl = __riscv_vsetvl_e16m1(width); vuint16m1_t s0, s1, s2, s3; load_u16_8x4(s, src_stride, &s0, &s1, &s2, &s3, vl); // apply shift and add round offset vuint16m1_t d0 = __riscv_vmacc_vx_u16m1(vec_offset, round_val, s0, vl); vuint16m1_t d1 = __riscv_vmacc_vx_u16m1(vec_offset, round_val, s1, vl); vuint16m1_t d2 = __riscv_vmacc_vx_u16m1(vec_offset, round_val, s2, vl); vuint16m1_t d3 = __riscv_vmacc_vx_u16m1(vec_offset, round_val, s3, vl); // store results store_u16_8x4(d, dst_stride, d0, d1, d2, d3, vl); s += vl; d += vl; width -= vl; } while (width > 0); src_ptr += 4 * src_stride; dst_ptr += 4 * dst_stride; height -= 4; } while (height > 0); } } void av1_highbd_dist_wtd_convolve_2d_copy_rvv(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd) { DECLARE_ALIGNED(16, uint16_t, im_block[(MAX_SB_SIZE + MAX_FILTER_TAP) * MAX_SB_SIZE]); const int im_stride = MAX_SB_SIZE; CONV_BUF_TYPE *dst16 = conv_params->dst; int dst16_stride = conv_params->dst_stride; const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0; const int round_offset = (1 << (offset_bits - conv_params->round_1)) + (1 << (offset_bits - conv_params->round_1 - 1)); const int round_bits = 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1; assert(round_bits >= 0); if (conv_params->do_average) { highbd_2d_copy_rvv(src, src_stride, im_block, im_stride, w, h, round_bits, round_offset); } else { highbd_2d_copy_rvv(src, src_stride, dst16, dst16_stride, w, h, round_bits, round_offset); } if (conv_params->do_average) { int shift_bits; if (bd == 12) { shift_bits = ROUND_SHIFT - 2; } else { shift_bits = ROUND_SHIFT; } if (conv_params->use_dist_wtd_comp_avg) { highbd_dist_wtd_comp_avg_rvv(im_block, im_stride, dst, dst_stride, w, h, conv_params, bd, shift_bits, offset_bits); } else { highbd_comp_avg_rvv(im_block, im_stride, dst, dst_stride, w, h, conv_params, bd, shift_bits, offset_bits); } } } static inline vuint16mf2_t highbd_convolve6_4_2d_v_rvv( const vint16mf2_t s0, const vint16mf2_t s1, const vint16mf2_t s2, const vint16mf2_t s3, const vint16mf2_t s4, const vint16mf2_t s5, const int16_t *y_filter, const vint32m1_t offset, size_t vl) { // Values at indices 0 and 7 of y_filter are zero. // Accumulate widening multiplications into the offset vector. vint32m1_t sum = __riscv_vwmacc_vx_i32m1(offset, y_filter[1], s0, vl); sum = __riscv_vwmacc_vx_i32m1(sum, y_filter[2], s1, vl); sum = __riscv_vwmacc_vx_i32m1(sum, y_filter[3], s2, vl); sum = __riscv_vwmacc_vx_i32m1(sum, y_filter[4], s3, vl); sum = __riscv_vwmacc_vx_i32m1(sum, y_filter[5], s4, vl); sum = __riscv_vwmacc_vx_i32m1(sum, y_filter[6], s5, vl); vuint32m1_t sum_u32 = __riscv_vreinterpret_v_i32m1_u32m1(__riscv_vmax_vx_i32m1(sum, 0, vl)); #if __riscv_v_intrinsic >= 12000 return __riscv_vnclipu_wx_u16mf2(sum_u32, COMPOUND_ROUND1_BITS, __RISCV_VXRM_RNU, vl); #else return __riscv_vnclipu_wx_u16mf2(sum_u32, COMPOUND_ROUND1_BITS, vl); #endif } static inline vuint16m1_t highbd_convolve6_8_2d_v_rvv( const vint16m1_t s0, const vint16m1_t s1, const vint16m1_t s2, const vint16m1_t s3, const vint16m1_t s4, const vint16m1_t s5, const int16_t *y_filter, const vint32m2_t offset, size_t vl) { // Values at indices 0 and 7 of y_filter are zero. // Perform widening multiply-accumulate starting with the offset vector. vint32m2_t sum = __riscv_vwmacc_vx_i32m2(offset, y_filter[1], s0, vl); sum = __riscv_vwmacc_vx_i32m2(sum, y_filter[2], s1, vl); sum = __riscv_vwmacc_vx_i32m2(sum, y_filter[3], s2, vl); sum = __riscv_vwmacc_vx_i32m2(sum, y_filter[4], s3, vl); sum = __riscv_vwmacc_vx_i32m2(sum, y_filter[5], s4, vl); sum = __riscv_vwmacc_vx_i32m2(sum, y_filter[6], s5, vl); vuint32m2_t sum_u32 = __riscv_vreinterpret_v_i32m2_u32m2(__riscv_vmax_vx_i32m2(sum, 0, vl)); #if __riscv_v_intrinsic >= 12000 return __riscv_vnclipu_wx_u16m1(sum_u32, COMPOUND_ROUND1_BITS, __RISCV_VXRM_RNU, vl); #else return __riscv_vnclipu_wx_u16m1(sum_u32, COMPOUND_ROUND1_BITS, vl); #endif } static inline void highbd_dist_wtd_convolve_2d_vert_6tap_rvv( const uint16_t *src_ptr, int src_stride, uint16_t *dst_ptr, int dst_stride, int w, int h, const int16_t *y_filter_ptr, int offset) { size_t vl = __riscv_vsetvl_e16m1(w); if (w == 4) { // Broadcast offset to a 32-bit vector vint32m1_t offset_vec = __riscv_vmv_v_x_i32m1(offset, vl); const int16_t *s = (const int16_t *)src_ptr; uint16_t *d = dst_ptr; // Load initial 5 rows vint16mf2_t s0, s1, s2, s3, s4; load_s16_4x5(s, src_stride, &s0, &s1, &s2, &s3, &s4, vl); s += 5 * src_stride; do { vint16mf2_t s5, s6, s7, s8; load_s16_4x4(s, src_stride, &s5, &s6, &s7, &s8, vl); // Perform 6-tap convolution for 4 rows vuint16mf2_t d0 = highbd_convolve6_4_2d_v_rvv( s0, s1, s2, s3, s4, s5, y_filter_ptr, offset_vec, vl); vuint16mf2_t d1 = highbd_convolve6_4_2d_v_rvv( s1, s2, s3, s4, s5, s6, y_filter_ptr, offset_vec, vl); vuint16mf2_t d2 = highbd_convolve6_4_2d_v_rvv( s2, s3, s4, s5, s6, s7, y_filter_ptr, offset_vec, vl); vuint16mf2_t d3 = highbd_convolve6_4_2d_v_rvv( s3, s4, s5, s6, s7, s8, y_filter_ptr, offset_vec, vl); store_u16_4x4(d, dst_stride, d0, d1, d2, d3, vl); // Update sliding window state s0 = s4; s1 = s5; s2 = s6; s3 = s7; s4 = s8; s += 4 * src_stride; d += 4 * dst_stride; h -= 4; } while (h != 0); } else { // w > 4 case do { vint32m2_t offset_vec = __riscv_vmv_v_x_i32m2(offset, vl); const int16_t *s = (const int16_t *)src_ptr; uint16_t *d = dst_ptr; int height = h; vint16m1_t s0, s1, s2, s3, s4; load_s16_8x5(s, src_stride, &s0, &s1, &s2, &s3, &s4, vl); s += 5 * src_stride; do { vint16m1_t s5, s6, s7, s8; load_s16_8x4(s, src_stride, &s5, &s6, &s7, &s8, vl); vuint16m1_t d0 = highbd_convolve6_8_2d_v_rvv( s0, s1, s2, s3, s4, s5, y_filter_ptr, offset_vec, vl); vuint16m1_t d1 = highbd_convolve6_8_2d_v_rvv( s1, s2, s3, s4, s5, s6, y_filter_ptr, offset_vec, vl); vuint16m1_t d2 = highbd_convolve6_8_2d_v_rvv( s2, s3, s4, s5, s6, s7, y_filter_ptr, offset_vec, vl); vuint16m1_t d3 = highbd_convolve6_8_2d_v_rvv( s3, s4, s5, s6, s7, s8, y_filter_ptr, offset_vec, vl); store_u16_8x4(d, dst_stride, d0, d1, d2, d3, vl); s0 = s4; s1 = s5; s2 = s6; s3 = s7; s4 = s8; s += 4 * src_stride; d += 4 * dst_stride; height -= 4; } while (height != 0); src_ptr += vl; dst_ptr += vl; w -= (int)vl; } while (w > 0); } } static inline vuint16mf2_t highbd_convolve8_4_2d_v_rvv( const vint16mf2_t s0, const vint16mf2_t s1, const vint16mf2_t s2, const vint16mf2_t s3, const vint16mf2_t s4, const vint16mf2_t s5, const vint16mf2_t s6, const vint16mf2_t s7, const int16_t *y_filter, const vint32m1_t offset, size_t vl) { // Perform widening multiply-accumulate for all 8 taps. // Accumulate directly into the offset vector. vint32m1_t sum = __riscv_vwmacc_vx_i32m1(offset, y_filter[0], s0, vl); sum = __riscv_vwmacc_vx_i32m1(sum, y_filter[1], s1, vl); sum = __riscv_vwmacc_vx_i32m1(sum, y_filter[2], s2, vl); sum = __riscv_vwmacc_vx_i32m1(sum, y_filter[3], s3, vl); sum = __riscv_vwmacc_vx_i32m1(sum, y_filter[4], s4, vl); sum = __riscv_vwmacc_vx_i32m1(sum, y_filter[5], s5, vl); sum = __riscv_vwmacc_vx_i32m1(sum, y_filter[6], s6, vl); sum = __riscv_vwmacc_vx_i32m1(sum, y_filter[7], s7, vl); // NEON vqrshrun_n_s32 -> RVV vnclipu_wx (with default RNU rounding) vuint32m1_t sum_u32 = __riscv_vreinterpret_v_i32m1_u32m1(__riscv_vmax_vx_i32m1(sum, 0, vl)); #if __riscv_v_intrinsic >= 12000 return __riscv_vnclipu_wx_u16mf2(sum_u32, COMPOUND_ROUND1_BITS, __RISCV_VXRM_RNU, vl); #else return __riscv_vnclipu_wx_u16mf2(sum_u32, COMPOUND_ROUND1_BITS, vl); #endif } static inline vuint16m1_t highbd_convolve8_8_2d_v_rvv( const vint16m1_t s0, const vint16m1_t s1, const vint16m1_t s2, const vint16m1_t s3, const vint16m1_t s4, const vint16m1_t s5, const vint16m1_t s6, const vint16m1_t s7, const int16_t *y_filter, const vint32m2_t offset, size_t vl) { // Perform widening multiply-accumulate for all 8 taps. vint32m2_t sum = __riscv_vwmacc_vx_i32m2(offset, y_filter[0], s0, vl); sum = __riscv_vwmacc_vx_i32m2(sum, y_filter[1], s1, vl); sum = __riscv_vwmacc_vx_i32m2(sum, y_filter[2], s2, vl); sum = __riscv_vwmacc_vx_i32m2(sum, y_filter[3], s3, vl); sum = __riscv_vwmacc_vx_i32m2(sum, y_filter[4], s4, vl); sum = __riscv_vwmacc_vx_i32m2(sum, y_filter[5], s5, vl); sum = __riscv_vwmacc_vx_i32m2(sum, y_filter[6], s6, vl); sum = __riscv_vwmacc_vx_i32m2(sum, y_filter[7], s7, vl); // NEON vqrshrun_n_s32 -> RVV vnclipu_wx (with default RNU rounding) vuint32m2_t sum_u32 = __riscv_vreinterpret_v_i32m2_u32m2(__riscv_vmax_vx_i32m2(sum, 0, vl)); #if __riscv_v_intrinsic >= 12000 return __riscv_vnclipu_wx_u16m1(sum_u32, COMPOUND_ROUND1_BITS, __RISCV_VXRM_RNU, vl); #else return __riscv_vnclipu_wx_u16m1(sum_u32, COMPOUND_ROUND1_BITS, vl); #endif } static inline void highbd_dist_wtd_convolve_2d_vert_8tap_rvv( const uint16_t *src_ptr, int src_stride, uint16_t *dst_ptr, int dst_stride, int w, int h, const int16_t *y_filter_ptr, int offset) { size_t vl = __riscv_vsetvl_e16m1(w); if (w == 4) { vint32m1_t offset_vec = __riscv_vmv_v_x_i32m1(offset, vl); const int16_t *s = (const int16_t *)src_ptr; uint16_t *d = dst_ptr; // Load initial 7 rows for 8-tap filter vint16mf2_t s0, s1, s2, s3, s4, s5, s6; load_s16_4x7(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6, vl); s += 7 * src_stride; do { vint16mf2_t s7, s8, s9, s10; load_s16_4x4(s, src_stride, &s7, &s8, &s9, &s10, vl); // Perform 8-tap convolution for 4 rows vuint16mf2_t d0 = highbd_convolve8_4_2d_v_rvv( s0, s1, s2, s3, s4, s5, s6, s7, y_filter_ptr, offset_vec, vl); vuint16mf2_t d1 = highbd_convolve8_4_2d_v_rvv( s1, s2, s3, s4, s5, s6, s7, s8, y_filter_ptr, offset_vec, vl); vuint16mf2_t d2 = highbd_convolve8_4_2d_v_rvv( s2, s3, s4, s5, s6, s7, s8, s9, y_filter_ptr, offset_vec, vl); vuint16mf2_t d3 = highbd_convolve8_4_2d_v_rvv( s3, s4, s5, s6, s7, s8, s9, s10, y_filter_ptr, offset_vec, vl); store_u16_4x4(d, dst_stride, d0, d1, d2, d3, vl); // Update sliding window state s0 = s4; s1 = s5; s2 = s6; s3 = s7; s4 = s8; s5 = s9; s6 = s10; s += 4 * src_stride; d += 4 * dst_stride; h -= 4; } while (h != 0); } else { // w > 4 case vint32m2_t offset_vec = __riscv_vmv_v_x_i32m2(offset, vl); do { const int16_t *s = (const int16_t *)src_ptr; uint16_t *d = dst_ptr; int height = h; vint16m1_t s0, s1, s2, s3, s4, s5, s6; load_s16_8x7(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6, vl); s += 7 * src_stride; do { vint16m1_t s7, s8, s9, s10; load_s16_8x4(s, src_stride, &s7, &s8, &s9, &s10, vl); vuint16m1_t d0 = highbd_convolve8_8_2d_v_rvv( s0, s1, s2, s3, s4, s5, s6, s7, y_filter_ptr, offset_vec, vl); vuint16m1_t d1 = highbd_convolve8_8_2d_v_rvv( s1, s2, s3, s4, s5, s6, s7, s8, y_filter_ptr, offset_vec, vl); vuint16m1_t d2 = highbd_convolve8_8_2d_v_rvv( s2, s3, s4, s5, s6, s7, s8, s9, y_filter_ptr, offset_vec, vl); vuint16m1_t d3 = highbd_convolve8_8_2d_v_rvv( s3, s4, s5, s6, s7, s8, s9, s10, y_filter_ptr, offset_vec, vl); store_u16_8x4(d, dst_stride, d0, d1, d2, d3, vl); s0 = s4; s1 = s5; s2 = s6; s3 = s7; s4 = s8; s5 = s9; s6 = s10; s += 4 * src_stride; d += 4 * dst_stride; height -= 4; } while (height != 0); src_ptr += vl; dst_ptr += vl; w -= (int)vl; } while (w > 0); } } static inline void highbd_dist_wtd_convolve_2d_horiz_6tap_rvv( const uint16_t *src_ptr, int src_stride, uint16_t *dst_ptr, int dst_stride, int w, int h, const int16_t *x_filter_ptr, const int offset, const int32_t round_bits) { const int16_t *filter = x_filter_ptr; size_t vl = __riscv_vsetvl_e16m1(w); vint32m2_t offset_vec = __riscv_vmv_v_x_i32m2(offset, vl); do { const int16_t *s = (const int16_t *)src_ptr; uint16_t *d = dst_ptr; int height = h; // Process 4 rows at a time to reduce loop overhead while (height >= 4) { vint16m1_t s00, s01, s02, s03, s04, s05; vint16m1_t s10, s11, s12, s13, s14, s15; vint16m1_t s20, s21, s22, s23, s24, s25; vint16m1_t s30, s31, s32, s33, s34, s35; // Load 6 vectors per row with stride 1 (horizontal overlap) load_s16_8x6(s + 0 * src_stride, 1, &s00, &s01, &s02, &s03, &s04, &s05, vl); load_s16_8x6(s + 1 * src_stride, 1, &s10, &s11, &s12, &s13, &s14, &s15, vl); load_s16_8x6(s + 2 * src_stride, 1, &s20, &s21, &s22, &s23, &s24, &s25, vl); load_s16_8x6(s + 3 * src_stride, 1, &s30, &s31, &s32, &s33, &s34, &s35, vl); vuint16m1_t d0 = highbd_convolve6_8_rvv( s00, s01, s02, s03, s04, s05, filter, offset_vec, round_bits, vl); vuint16m1_t d1 = highbd_convolve6_8_rvv( s10, s11, s12, s13, s14, s15, filter, offset_vec, round_bits, vl); vuint16m1_t d2 = highbd_convolve6_8_rvv( s20, s21, s22, s23, s24, s25, filter, offset_vec, round_bits, vl); vuint16m1_t d3 = highbd_convolve6_8_rvv( s30, s31, s32, s33, s34, s35, filter, offset_vec, round_bits, vl); store_u16_8x4(d, dst_stride, d0, d1, d2, d3, vl); s += 4 * src_stride; d += 4 * dst_stride; height -= 4; } // Handle remaining rows while (height > 0) { vint16m1_t s00, s01, s02, s03, s04, s05; load_s16_8x6(s, 1, &s00, &s01, &s02, &s03, &s04, &s05, vl); vuint16m1_t d0 = highbd_convolve6_8_rvv( s00, s01, s02, s03, s04, s05, filter, offset_vec, round_bits, vl); __riscv_vse16_v_u16m1(d, d0, vl); s += src_stride; d += dst_stride; height--; } src_ptr += vl; dst_ptr += vl; w -= (int)vl; } while (w > 0); } static inline void highbd_dist_wtd_convolve_2d_horiz_rvv( const uint16_t *src_ptr, int src_stride, uint16_t *dst_ptr, int dst_stride, int w, int h, const int16_t *x_filter_ptr, const int offset, const int32_t round_bits) { assert(h >= 5); if (w == 4) { size_t vl = __riscv_vsetvl_e16mf2(4); vint32m1_t offset_vec = __riscv_vmv_v_x_i32m1(offset, vl); const int16_t *filter = x_filter_ptr + 2; // Use middle 4 taps const int16_t *s = (const int16_t *)(src_ptr + 1); uint16_t *d = dst_ptr; while (h >= 4) { vint16mf2_t s00, s01, s02, s03; vint16mf2_t s10, s11, s12, s13; vint16mf2_t s20, s21, s22, s23; vint16mf2_t s30, s31, s32, s33; load_s16_4x4(s + 0 * src_stride, 1, &s00, &s01, &s02, &s03, vl); load_s16_4x4(s + 1 * src_stride, 1, &s10, &s11, &s12, &s13, vl); load_s16_4x4(s + 2 * src_stride, 1, &s20, &s21, &s22, &s23, vl); load_s16_4x4(s + 3 * src_stride, 1, &s30, &s31, &s32, &s33, vl); vuint16mf2_t d0 = highbd_convolve4_4_x_rvv(s00, s01, s02, s03, filter, offset_vec, round_bits, vl); vuint16mf2_t d1 = highbd_convolve4_4_x_rvv(s10, s11, s12, s13, filter, offset_vec, round_bits, vl); vuint16mf2_t d2 = highbd_convolve4_4_x_rvv(s20, s21, s22, s23, filter, offset_vec, round_bits, vl); vuint16mf2_t d3 = highbd_convolve4_4_x_rvv(s30, s31, s32, s33, filter, offset_vec, round_bits, vl); store_u16_4x4(d, dst_stride, d0, d1, d2, d3, vl); s += 4 * src_stride; d += 4 * dst_stride; h -= 4; } while (h > 0) { vint16mf2_t s00, s01, s02, s03; load_s16_4x4(s, 1, &s00, &s01, &s02, &s03, vl); vuint16mf2_t d0 = highbd_convolve4_4_x_rvv(s00, s01, s02, s03, filter, offset_vec, round_bits, vl); __riscv_vse16_v_u16mf2(d, d0, vl); s += src_stride; d += dst_stride; h--; } } else { // w > 4 case using 8-tap filter const int16_t *filter = x_filter_ptr; size_t vl = __riscv_vsetvl_e16m1(w); vint32m2_t offset_vec = __riscv_vmv_v_x_i32m2(offset, vl); do { const int16_t *s_base = (const int16_t *)src_ptr; uint16_t *d_base = dst_ptr; int height = h; while (height >= 4) { vint16m1_t s00, s01, s02, s03, s04, s05, s06, s07; vint16m1_t s10, s11, s12, s13, s14, s15, s16, s17; vint16m1_t s20, s21, s22, s23, s24, s25, s26, s27; vint16m1_t s30, s31, s32, s33, s34, s35, s36, s37; load_s16_8x8(s_base + 0 * src_stride, 1, &s00, &s01, &s02, &s03, &s04, &s05, &s06, &s07, vl); load_s16_8x8(s_base + 1 * src_stride, 1, &s10, &s11, &s12, &s13, &s14, &s15, &s16, &s17, vl); load_s16_8x8(s_base + 2 * src_stride, 1, &s20, &s21, &s22, &s23, &s24, &s25, &s26, &s27, vl); load_s16_8x8(s_base + 3 * src_stride, 1, &s30, &s31, &s32, &s33, &s34, &s35, &s36, &s37, vl); vuint16m1_t d0 = highbd_convolve8_8_rvv(s00, s01, s02, s03, s04, s05, s06, s07, filter, offset_vec, round_bits, vl); vuint16m1_t d1 = highbd_convolve8_8_rvv(s10, s11, s12, s13, s14, s15, s16, s17, filter, offset_vec, round_bits, vl); vuint16m1_t d2 = highbd_convolve8_8_rvv(s20, s21, s22, s23, s24, s25, s26, s27, filter, offset_vec, round_bits, vl); vuint16m1_t d3 = highbd_convolve8_8_rvv(s30, s31, s32, s33, s34, s35, s36, s37, filter, offset_vec, round_bits, vl); store_u16_8x4(d_base, dst_stride, d0, d1, d2, d3, vl); s_base += 4 * src_stride; d_base += 4 * dst_stride; height -= 4; } while (height > 0) { vint16m1_t s00, s01, s02, s03, s04, s05, s06, s07; load_s16_8x8(s_base, 1, &s00, &s01, &s02, &s03, &s04, &s05, &s06, &s07, vl); vuint16m1_t d0 = highbd_convolve8_8_rvv(s00, s01, s02, s03, s04, s05, s06, s07, filter, offset_vec, round_bits, vl); __riscv_vse16_v_u16m1(d_base, d0, vl); s_base += src_stride; d_base += dst_stride; height--; } src_ptr += vl; dst_ptr += vl; w -= (int)vl; } while (w > 0); } } void av1_highbd_dist_wtd_convolve_2d_rvv( const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd) { DECLARE_ALIGNED(16, uint16_t, im_block[(MAX_SB_SIZE + MAX_FILTER_TAP) * MAX_SB_SIZE]); DECLARE_ALIGNED(16, uint16_t, im_block2[(MAX_SB_SIZE + MAX_FILTER_TAP) * MAX_SB_SIZE]); CONV_BUF_TYPE *dst16 = conv_params->dst; int dst16_stride = conv_params->dst_stride; const int x_filter_taps = get_filter_tap(filter_params_x, subpel_x_qn); const int clamped_x_taps = x_filter_taps < 6 ? 6 : x_filter_taps; const int y_filter_taps = get_filter_tap(filter_params_y, subpel_y_qn); const int clamped_y_taps = y_filter_taps < 6 ? 6 : y_filter_taps; const int im_h = h + clamped_y_taps - 1; const int im_stride = MAX_SB_SIZE; const int vert_offset = clamped_y_taps / 2 - 1; const int horiz_offset = clamped_x_taps / 2 - 1; const int round_offset_conv_x = (1 << (bd + FILTER_BITS - 1)) + (1 << (conv_params->round_0 - 1)); const int y_offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0; const int round_offset_conv_y = (1 << y_offset_bits); const uint16_t *src_ptr = src - vert_offset * src_stride - horiz_offset; const int16_t *x_filter_ptr = av1_get_interp_filter_subpel_kernel( filter_params_x, subpel_x_qn & SUBPEL_MASK); const int16_t *y_filter_ptr = av1_get_interp_filter_subpel_kernel( filter_params_y, subpel_y_qn & SUBPEL_MASK); // 1. Horizontal Filter Pass int round_bits; if (bd == 12) { round_bits = ROUND0_BITS + 2; } else { round_bits = ROUND0_BITS; } if (x_filter_taps <= 6 && w != 4) { highbd_dist_wtd_convolve_2d_horiz_6tap_rvv(src_ptr, src_stride, im_block, im_stride, w, im_h, x_filter_ptr, round_offset_conv_x, round_bits); } else { highbd_dist_wtd_convolve_2d_horiz_rvv(src_ptr, src_stride, im_block, im_stride, w, im_h, x_filter_ptr, round_offset_conv_x, round_bits); } // 2. Vertical Filter Pass if (y_filter_taps <= 6) { if (conv_params->do_average) { highbd_dist_wtd_convolve_2d_vert_6tap_rvv(im_block, im_stride, im_block2, im_stride, w, h, y_filter_ptr, round_offset_conv_y); } else { highbd_dist_wtd_convolve_2d_vert_6tap_rvv( im_block, im_stride, dst16, dst16_stride, w, h, y_filter_ptr, round_offset_conv_y); } } else { if (conv_params->do_average) { highbd_dist_wtd_convolve_2d_vert_8tap_rvv(im_block, im_stride, im_block2, im_stride, w, h, y_filter_ptr, round_offset_conv_y); } else { highbd_dist_wtd_convolve_2d_vert_8tap_rvv( im_block, im_stride, dst16, dst16_stride, w, h, y_filter_ptr, round_offset_conv_y); } } // 3. Compound Averaging if (conv_params->do_average) { int shift_bits, offset_bits; if (bd == 12) { shift_bits = ROUND_SHIFT - 2; offset_bits = bd + 2 * FILTER_BITS - ROUND0_BITS - 2; } else { shift_bits = ROUND_SHIFT; offset_bits = bd + 2 * FILTER_BITS - ROUND0_BITS; } if (conv_params->use_dist_wtd_comp_avg) { highbd_dist_wtd_comp_avg_rvv(im_block2, im_stride, dst, dst_stride, w, h, conv_params, bd, shift_bits, offset_bits); } else { highbd_comp_avg_rvv(im_block2, im_stride, dst, dst_stride, w, h, conv_params, bd, shift_bits, offset_bits); } } }