| /* |
| * 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 <assert.h> |
| #include <riscv_vector.h> |
| |
| #include "av1/common/riscv/compound_convolve_rvv.h" |
| #include "config/aom_config.h" |
| #include "config/av1_rtcd.h" |
| |
| static uint16_t trans_data[64] = { |
| 0, 8, 16, 24, 32, 40, 48, 56, 1, 9, 17, 25, 33, 41, 49, 57, |
| 2, 10, 18, 26, 34, 42, 50, 58, 3, 11, 19, 27, 35, 43, 51, 59, |
| 4, 12, 20, 28, 36, 44, 52, 60, 5, 13, 21, 29, 37, 45, 53, 61, |
| 6, 14, 22, 30, 38, 46, 54, 62, 7, 15, 23, 31, 39, 47, 55, 63, |
| }; |
| |
| static inline void transpose_elems_inplace_s16_8x8_rvv( |
| vint16m1_t *d0, vint16m1_t *d1, vint16m1_t *d2, vint16m1_t *d3, |
| vint16m1_t *d4, vint16m1_t *d5, vint16m1_t *d6, vint16m1_t *d7, |
| const vuint16m8_t trans_index) { |
| vint16m8_t matrix_8x8 = __riscv_vundefined_i16m8(); |
| matrix_8x8 = __riscv_vset_v_i16m1_i16m8(matrix_8x8, 0, *d0); |
| matrix_8x8 = __riscv_vset_v_i16m1_i16m8(matrix_8x8, 1, *d1); |
| matrix_8x8 = __riscv_vset_v_i16m1_i16m8(matrix_8x8, 2, *d2); |
| matrix_8x8 = __riscv_vset_v_i16m1_i16m8(matrix_8x8, 3, *d3); |
| matrix_8x8 = __riscv_vset_v_i16m1_i16m8(matrix_8x8, 4, *d4); |
| matrix_8x8 = __riscv_vset_v_i16m1_i16m8(matrix_8x8, 5, *d5); |
| matrix_8x8 = __riscv_vset_v_i16m1_i16m8(matrix_8x8, 6, *d6); |
| matrix_8x8 = __riscv_vset_v_i16m1_i16m8(matrix_8x8, 7, *d7); |
| |
| // transpose_elems_inplace_s16_8x8 |
| vint16m8_t trans_matrix = |
| __riscv_vrgather_vv_i16m8(matrix_8x8, trans_index, 64); |
| |
| *d0 = __riscv_vget_v_i16m8_i16m1(trans_matrix, 0); |
| *d1 = __riscv_vget_v_i16m8_i16m1(trans_matrix, 1); |
| *d2 = __riscv_vget_v_i16m8_i16m1(trans_matrix, 2); |
| *d3 = __riscv_vget_v_i16m8_i16m1(trans_matrix, 3); |
| *d4 = __riscv_vget_v_i16m8_i16m1(trans_matrix, 4); |
| *d5 = __riscv_vget_v_i16m8_i16m1(trans_matrix, 5); |
| *d6 = __riscv_vget_v_i16m8_i16m1(trans_matrix, 6); |
| *d7 = __riscv_vget_v_i16m8_i16m1(trans_matrix, 7); |
| } |
| |
| static inline vint16m1_t convolve4_4_2d_h_rvv( |
| const vint16m1_t s0, const vint16m1_t s1, const vint16m1_t s2, |
| const vint16m1_t s3, const int16_t *filter, const int16_t horiz_const, |
| size_t vl) { |
| vint16m1_t sum = __riscv_vmul_vx_i16m1(s0, filter[0], vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[1], s1, vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[2], s2, vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[3], s3, vl); |
| sum = __riscv_vadd_vx_i16m1(sum, horiz_const, vl); |
| |
| // We halved the convolution filter values so -1 from the right shift. |
| return __riscv_vsra_vx_i16m1(sum, ROUND0_BITS - 1, vl); |
| } |
| |
| static inline vint16m1_t convolve8_8_2d_h_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 int16_t horiz_const, size_t vl) { |
| vint16m1_t sum = __riscv_vmul_vx_i16m1(s0, filter[0], vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[1], s1, vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[2], s2, vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[3], s3, vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[4], s4, vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[5], s5, vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[6], s6, vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[7], s7, vl); |
| sum = __riscv_vadd_vx_i16m1(sum, horiz_const, vl); |
| |
| // We halved the convolution filter values so -1 from the right shift. |
| return __riscv_vsra_vx_i16m1(sum, ROUND0_BITS - 1, vl); |
| } |
| |
| static inline void dist_wtd_convolve_2d_horiz_rvv( |
| const uint8_t *src, int src_stride, int16_t *im_block, const int im_stride, |
| const int16_t *x_filter_ptr, const int im_h, int w) { |
| const int bd = 8; |
| // A shim of 1 << ((ROUND0_BITS - 1) - 1) enables us to use non-rounding |
| // shifts - which are generally faster than rounding shifts on modern CPUs. |
| // (The extra -1 is needed because we halved the filter values.) |
| const int16_t horiz_const = |
| ((1 << (bd + FILTER_BITS - 2)) + (1 << ((ROUND0_BITS - 1) - 1))); |
| |
| const uint8_t *src_ptr = src; |
| int16_t *dst_ptr = im_block; |
| int dst_stride = im_stride; |
| int height = im_h; |
| size_t vl = __riscv_vsetvl_e16m1(w); |
| |
| if (w == 4) { |
| // 4-tap filters are used for blocks having width <= 4. |
| // Filter values are even, so halve to reduce intermediate precision reqs. |
| const int16_t *x_filter = x_filter_ptr + 2; |
| int16_t filter[4]; |
| for (int i = 0; i < 4; i++) filter[i] = *x_filter++ >> 1; |
| |
| src_ptr += 2; |
| |
| do { |
| // load |
| vuint8mf2_t t0 = __riscv_vle8_v_u8mf2(src_ptr + 0 * src_stride, |
| 8); // a0 a1 a2 a3 a4 a5 a6 a7 |
| vuint8mf2_t t1 = __riscv_vle8_v_u8mf2(src_ptr + 1 * src_stride, |
| 8); // b0 b1 b2 b3 b4 b5 b6 b7 |
| |
| __builtin_prefetch(dst_ptr + 0 * dst_stride); |
| __builtin_prefetch(dst_ptr + 1 * dst_stride); |
| |
| // widen to 16-bit |
| vint16m1_t s00 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t0, 8)); |
| vint16m1_t s01 = |
| __riscv_vslide1down_vx_i16m1(s00, 0, 8); // a1 a2 a3 a4 a5 a6 a7 0 |
| vint16m1_t s02 = |
| __riscv_vslide1down_vx_i16m1(s01, 0, 8); // a2 a3 a4 a5 a6 a7 0 0 |
| vint16m1_t s03 = |
| __riscv_vslide1down_vx_i16m1(s02, 0, 8); // a3 a4 a5 a6 a7 0 0 0 |
| |
| vint16m1_t s10 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t1, 8)); |
| vint16m1_t s11 = |
| __riscv_vslide1down_vx_i16m1(s10, 0, 8); // b1 b2 b3 b4 b5 b6 b7 0 |
| vint16m1_t s12 = |
| __riscv_vslide1down_vx_i16m1(s11, 0, 8); // b2 b3 b4 b5 b6 b7 0 0 |
| vint16m1_t s13 = |
| __riscv_vslide1down_vx_i16m1(s12, 0, 8); // b3 b4 b5 b6 b7 0 0 0 |
| |
| // perform convolution |
| vint16m1_t d0 = |
| convolve4_4_2d_h_rvv(s00, s01, s02, s03, filter, horiz_const, vl); |
| vint16m1_t d1 = |
| convolve4_4_2d_h_rvv(s10, s11, s12, s13, filter, horiz_const, vl); |
| |
| // store result |
| __riscv_vse16_v_i16m1(dst_ptr + 0 * dst_stride, d0, vl); |
| __riscv_vse16_v_i16m1(dst_ptr + 1 * dst_stride, d1, vl); |
| src_ptr += 2 * src_stride; |
| dst_ptr += 2 * dst_stride; |
| height -= 2; |
| } while (height >= 2); |
| |
| if (height > 0) { |
| vuint8mf2_t t0 = __riscv_vle8_v_u8mf2(src_ptr + 0 * src_stride, |
| 8); // a0 a1 a2 a3 a4 a5 a6 a7 |
| |
| // widen to 16-bit |
| vint16m1_t s00 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t0, 8)); |
| vint16m1_t s01 = |
| __riscv_vslide1down_vx_i16m1(s00, 0, 8); // a1 a2 a3 a4 a5 a6 a7 0 |
| vint16m1_t s02 = |
| __riscv_vslide1down_vx_i16m1(s01, 0, 8); // a2 a3 a4 a5 a6 a7 0 0 |
| vint16m1_t s03 = |
| __riscv_vslide1down_vx_i16m1(s02, 0, 8); // a3 a4 a5 a6 a7 0 0 0 |
| |
| // perform convolution |
| vint16m1_t d0 = |
| convolve4_4_2d_h_rvv(s00, s01, s02, s03, filter, horiz_const, vl); |
| |
| // store result |
| __riscv_vse16_v_i16m1(dst_ptr + 0 * dst_stride, d0, vl); |
| } |
| } else { |
| // Filter values are even, so halve to reduce intermediate precision reqs. |
| int16_t filter[8]; |
| for (int i = 0; i < 8; i++) filter[i] = *x_filter_ptr++ >> 1; |
| |
| // for VLEN = 128 case |
| if ((vl == 8) && (w > 8) && (height >= 8)) { |
| vuint16m8_t trans_index = __riscv_vle16_v_u16m8(trans_data, 64); |
| |
| while (height >= 8) { |
| const uint8_t *s = src_ptr; |
| int16_t *d = dst_ptr; |
| int width = w; |
| |
| // Load |
| vuint8mf2_t t0, t1, t2, t3, t4, t5, t6, t7; |
| load_stride_u8_8x8(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, |
| &t7, vl); |
| |
| vint16m1_t s0 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t0, vl)); |
| vint16m1_t s1 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t1, vl)); |
| vint16m1_t s2 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t2, vl)); |
| vint16m1_t s3 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t3, vl)); |
| vint16m1_t s4 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t4, vl)); |
| vint16m1_t s5 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t5, vl)); |
| vint16m1_t s6 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t6, vl)); |
| |
| __builtin_prefetch(d + 0 * dst_stride); |
| __builtin_prefetch(d + 1 * dst_stride); |
| __builtin_prefetch(d + 2 * dst_stride); |
| __builtin_prefetch(d + 3 * dst_stride); |
| __builtin_prefetch(d + 4 * dst_stride); |
| __builtin_prefetch(d + 5 * dst_stride); |
| __builtin_prefetch(d + 6 * dst_stride); |
| __builtin_prefetch(d + 7 * dst_stride); |
| |
| s += 7; |
| |
| do { |
| load_stride_u8_8x8(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, |
| &t7, vl); |
| |
| vint16m1_t s7 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t0, vl)); |
| vint16m1_t s8 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t1, vl)); |
| vint16m1_t s9 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t2, vl)); |
| vint16m1_t s10 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t3, vl)); |
| vint16m1_t s11 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t4, vl)); |
| vint16m1_t s12 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t5, vl)); |
| vint16m1_t s13 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t6, vl)); |
| vint16m1_t s14 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t7, vl)); |
| |
| // Perform convolution |
| vint16m1_t d0 = convolve8_8_2d_h_rvv(s0, s1, s2, s3, s4, s5, s6, s7, |
| filter, horiz_const, vl); |
| vint16m1_t d1 = convolve8_8_2d_h_rvv(s1, s2, s3, s4, s5, s6, s7, s8, |
| filter, horiz_const, vl); |
| vint16m1_t d2 = convolve8_8_2d_h_rvv(s2, s3, s4, s5, s6, s7, s8, s9, |
| filter, horiz_const, vl); |
| vint16m1_t d3 = convolve8_8_2d_h_rvv(s3, s4, s5, s6, s7, s8, s9, s10, |
| filter, horiz_const, vl); |
| vint16m1_t d4 = convolve8_8_2d_h_rvv(s4, s5, s6, s7, s8, s9, s10, s11, |
| filter, horiz_const, vl); |
| vint16m1_t d5 = convolve8_8_2d_h_rvv(s5, s6, s7, s8, s9, s10, s11, |
| s12, filter, horiz_const, vl); |
| vint16m1_t d6 = convolve8_8_2d_h_rvv(s6, s7, s8, s9, s10, s11, s12, |
| s13, filter, horiz_const, vl); |
| vint16m1_t d7 = convolve8_8_2d_h_rvv(s7, s8, s9, s10, s11, s12, s13, |
| s14, filter, horiz_const, vl); |
| |
| // transpose |
| transpose_elems_inplace_s16_8x8_rvv(&d0, &d1, &d2, &d3, &d4, &d5, &d6, |
| &d7, trans_index); |
| |
| // Store result |
| store_s16_8x8(d, dst_stride, d0, d1, d2, d3, d4, d5, d6, d7, vl); |
| |
| // Update sliding window |
| s0 = __riscv_vmv_v_v_i16m1(s8, vl); |
| s1 = __riscv_vmv_v_v_i16m1(s9, vl); |
| s2 = __riscv_vmv_v_v_i16m1(s10, vl); |
| s3 = __riscv_vmv_v_v_i16m1(s11, vl); |
| s4 = __riscv_vmv_v_v_i16m1(s12, vl); |
| s5 = __riscv_vmv_v_v_i16m1(s13, vl); |
| s6 = __riscv_vmv_v_v_i16m1(s14, vl); |
| s += 8; |
| d += 8; |
| width -= 8; |
| } while (width != 0); |
| src_ptr += 8 * src_stride; |
| dst_ptr += 8 * dst_stride; |
| height -= 8; |
| } |
| } |
| |
| while (height > 0) { |
| const uint8_t *s = src_ptr; |
| int16_t *d = dst_ptr; |
| int width = w; |
| |
| __builtin_prefetch(d); |
| |
| do { |
| // Load |
| vuint8mf2_t t0, t1, t2, t3, t4, t5, t6, t7; |
| load_u8_8x8(s, 1, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7, vl); |
| |
| // Convert to 16-bit integers |
| vint16m1_t s0 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t0, vl)); |
| vint16m1_t s1 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t1, vl)); |
| vint16m1_t s2 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t2, vl)); |
| vint16m1_t s3 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t3, vl)); |
| vint16m1_t s4 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t4, vl)); |
| vint16m1_t s5 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t5, vl)); |
| vint16m1_t s6 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t6, vl)); |
| vint16m1_t s7 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t7, vl)); |
| |
| // Perform convolution |
| vint16m1_t d0 = convolve8_8_2d_h_rvv(s0, s1, s2, s3, s4, s5, s6, s7, |
| filter, horiz_const, vl); |
| |
| // Store result |
| __riscv_vse16_v_i16m1(d, d0, vl); |
| |
| s += vl; |
| d += vl; |
| width -= vl; |
| } while (width > 0); |
| src_ptr += src_stride; |
| dst_ptr += dst_stride; |
| height--; |
| } |
| } |
| } |
| |
| void av1_dist_wtd_convolve_2d_rvv(const uint8_t *src, int src_stride, |
| uint8_t *dst8, int dst8_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) { |
| assert(w % 4 == 0); |
| assert(h % 4 == 0); |
| |
| DECLARE_ALIGNED(16, int16_t, |
| im_block[(MAX_SB_SIZE + SUBPEL_TAPS - 1) * MAX_SB_SIZE]); |
| |
| 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 = filter_params_x->taps / 2 - 1; |
| const uint8_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); |
| |
| dist_wtd_convolve_2d_horiz_rvv(src_ptr, src_stride, im_block, im_stride, |
| x_filter_ptr, im_h, w); |
| |
| if (clamped_y_taps == 6) { |
| if (conv_params->do_average) { |
| if (UNLIKELY(conv_params->use_dist_wtd_comp_avg)) { |
| dist_wtd_convolve_2d_vert_6tap_dist_wtd_avg_rvv( |
| im_block, im_stride, dst8, dst8_stride, conv_params, y_filter_ptr, |
| h, w); |
| } else { |
| dist_wtd_convolve_2d_vert_6tap_avg_rvv(im_block, im_stride, dst8, |
| dst8_stride, conv_params, |
| y_filter_ptr, h, w); |
| } |
| } else { |
| dist_wtd_convolve_2d_vert_6tap_rvv(im_block, im_stride, conv_params, |
| y_filter_ptr, h, w); |
| } |
| } else { |
| if (conv_params->do_average) { |
| if (UNLIKELY(conv_params->use_dist_wtd_comp_avg)) { |
| dist_wtd_convolve_2d_vert_8tap_dist_wtd_avg_rvv( |
| im_block, im_stride, dst8, dst8_stride, conv_params, y_filter_ptr, |
| h, w); |
| } else { |
| dist_wtd_convolve_2d_vert_8tap_avg_rvv(im_block, im_stride, dst8, |
| dst8_stride, conv_params, |
| y_filter_ptr, h, w); |
| } |
| } else { |
| dist_wtd_convolve_2d_vert_8tap_rvv(im_block, im_stride, conv_params, |
| y_filter_ptr, h, w); |
| } |
| } |
| } |
| |
| static inline void dist_wtd_convolve_2d_copy_dist_wtd_avg_rvv( |
| const uint8_t *src, int src_stride, uint8_t *dst8, int dst8_stride, int w, |
| int h, ConvolveParams *conv_params) { |
| assert(w % 4 == 0); |
| assert(h % 4 == 0); |
| |
| const int bd = 8; |
| const int offset_bits = bd + 2 * FILTER_BITS - ROUND0_BITS; |
| const uint16_t round_offset = (1 << (offset_bits - COMPOUND_ROUND1_BITS)) + |
| (1 << (offset_bits - COMPOUND_ROUND1_BITS - 1)); |
| const uint8_t shift_by_bits = 1 << (FILTER_BITS - ROUND0_BITS); |
| |
| const uint16_t fwd_offset = conv_params->fwd_offset; |
| const uint16_t bck_offset = conv_params->bck_offset; |
| |
| CONV_BUF_TYPE *dst = conv_params->dst; |
| const int dst_stride = conv_params->dst_stride; |
| int height = h; |
| size_t vl = __riscv_vsetvl_e16m1(w); |
| |
| if (w == 4) { |
| vuint16mf2_t round_offset_u16 = __riscv_vmv_v_x_u16mf2(round_offset, vl); |
| do { |
| // load 4 block of uint8_t |
| vuint8mf4_t s0 = __riscv_vle8_v_u8mf4(src + 0 * src_stride, vl); |
| vuint8mf4_t s1 = __riscv_vle8_v_u8mf4(src + 1 * src_stride, vl); |
| vuint8mf4_t s2 = __riscv_vle8_v_u8mf4(src + 2 * src_stride, vl); |
| vuint8mf4_t s3 = __riscv_vle8_v_u8mf4(src + 3 * src_stride, vl); |
| |
| // apply shift and add round offset |
| vuint16mf2_t d0 = |
| __riscv_vwmaccu_vx_u16mf2(round_offset_u16, shift_by_bits, s0, vl); |
| vuint16mf2_t d1 = |
| __riscv_vwmaccu_vx_u16mf2(round_offset_u16, shift_by_bits, s1, vl); |
| vuint16mf2_t d2 = |
| __riscv_vwmaccu_vx_u16mf2(round_offset_u16, shift_by_bits, s2, vl); |
| vuint16mf2_t d3 = |
| __riscv_vwmaccu_vx_u16mf2(round_offset_u16, shift_by_bits, s3, vl); |
| |
| // load existing dst values |
| vuint16mf2_t dd0, dd1, dd2, dd3; |
| load_u16_4x4(dst, dst_stride, &dd0, &dd1, &dd2, &dd3, vl); |
| |
| // compute distance-weighted average |
| vuint8mf4_t d0_u8, d1_u8, d2_u8, d3_u8; |
| compute_dist_wtd_avg_4x4_rvv(dd0, dd1, dd2, dd3, d0, d1, d2, d3, |
| fwd_offset, bck_offset, round_offset, &d0_u8, |
| &d1_u8, &d2_u8, &d3_u8, vl); |
| |
| // store results |
| __riscv_vse8_v_u8mf4(dst8 + 0 * dst8_stride, d0_u8, vl); |
| __riscv_vse8_v_u8mf4(dst8 + 1 * dst8_stride, d1_u8, vl); |
| __riscv_vse8_v_u8mf4(dst8 + 2 * dst8_stride, d2_u8, vl); |
| __riscv_vse8_v_u8mf4(dst8 + 3 * dst8_stride, d3_u8, vl); |
| |
| // update pointers |
| src += 4 * src_stride; |
| dst += 4 * dst_stride; |
| dst8 += 4 * dst8_stride; |
| height -= 4; |
| } while (height != 0); |
| } else { |
| vuint16m1_t round_offset_u16 = __riscv_vmv_v_x_u16m1(round_offset, vl); |
| do { |
| const uint8_t *s = src; |
| CONV_BUF_TYPE *d = dst; |
| uint8_t *d_u8 = dst8; |
| int width = w; |
| |
| do { |
| // load 4 block of uint8_t |
| vuint8mf2_t s0, s1, s2, s3; |
| load_u8_8x4(s, src_stride, &s0, &s1, &s2, &s3, vl); |
| |
| // apply shift and add round offset |
| vuint16m1_t d0 = |
| __riscv_vwmaccu_vx_u16m1(round_offset_u16, shift_by_bits, s0, vl); |
| vuint16m1_t d1 = |
| __riscv_vwmaccu_vx_u16m1(round_offset_u16, shift_by_bits, s1, vl); |
| vuint16m1_t d2 = |
| __riscv_vwmaccu_vx_u16m1(round_offset_u16, shift_by_bits, s2, vl); |
| vuint16m1_t d3 = |
| __riscv_vwmaccu_vx_u16m1(round_offset_u16, shift_by_bits, s3, vl); |
| |
| // load existing dst values |
| vuint16m1_t dd0, dd1, dd2, dd3; |
| load_u16_8x4(d, dst_stride, &dd0, &dd1, &dd2, &dd3, vl); |
| |
| // compute distance-weighted average |
| vuint8mf2_t d0_u8, d1_u8, d2_u8, d3_u8; |
| compute_dist_wtd_avg_8x4_rvv(dd0, dd1, dd2, dd3, d0, d1, d2, d3, |
| fwd_offset, bck_offset, round_offset, |
| &d0_u8, &d1_u8, &d2_u8, &d3_u8, vl); |
| |
| // store results |
| store_u8_8x4(d_u8, dst8_stride, d0_u8, d1_u8, d2_u8, d3_u8, vl); |
| |
| // update pointers |
| s += vl; |
| d += vl; |
| d_u8 += vl; |
| width -= vl; |
| } while (width > 0); |
| src += 4 * src_stride; |
| dst += 4 * dst_stride; |
| dst8 += 4 * dst8_stride; |
| height -= 4; |
| } while (height != 0); |
| } |
| } |
| |
| static inline void dist_wtd_convolve_2d_copy_avg_rvv( |
| const uint8_t *src, int src_stride, uint8_t *dst8, int dst8_stride, int w, |
| int h, ConvolveParams *conv_params) { |
| assert(w % 4 == 0); |
| assert(h % 4 == 0); |
| |
| const int bd = 8; |
| const int offset_bits = bd + 2 * FILTER_BITS - ROUND0_BITS; |
| const uint16_t round_offset = (1 << (offset_bits - COMPOUND_ROUND1_BITS)) + |
| (1 << (offset_bits - COMPOUND_ROUND1_BITS - 1)); |
| const uint8_t shift_by_bits = 1 << (FILTER_BITS - ROUND0_BITS); |
| |
| CONV_BUF_TYPE *dst = conv_params->dst; |
| const int dst_stride = conv_params->dst_stride; |
| int height = h; |
| size_t vl = __riscv_vsetvl_e16m1(w); |
| |
| if (w == 4) { |
| vuint16mf2_t round_offset_u16 = __riscv_vmv_v_x_u16mf2(round_offset, vl); |
| do { |
| // load 4 block of uint8_t |
| vuint8mf4_t s0 = __riscv_vle8_v_u8mf4(src + 0 * src_stride, vl); |
| vuint8mf4_t s1 = __riscv_vle8_v_u8mf4(src + 1 * src_stride, vl); |
| vuint8mf4_t s2 = __riscv_vle8_v_u8mf4(src + 2 * src_stride, vl); |
| vuint8mf4_t s3 = __riscv_vle8_v_u8mf4(src + 3 * src_stride, vl); |
| |
| // apply shift and add round offset |
| vuint16mf2_t d0 = |
| __riscv_vwmaccu_vx_u16mf2(round_offset_u16, shift_by_bits, s0, vl); |
| vuint16mf2_t d1 = |
| __riscv_vwmaccu_vx_u16mf2(round_offset_u16, shift_by_bits, s1, vl); |
| vuint16mf2_t d2 = |
| __riscv_vwmaccu_vx_u16mf2(round_offset_u16, shift_by_bits, s2, vl); |
| vuint16mf2_t d3 = |
| __riscv_vwmaccu_vx_u16mf2(round_offset_u16, shift_by_bits, s3, vl); |
| |
| // load existing dst values |
| vuint16mf2_t dd0, dd1, dd2, dd3; |
| load_u16_4x4(dst, dst_stride, &dd0, &dd1, &dd2, &dd3, vl); |
| |
| // compute basic average |
| vuint8mf4_t d0_u8, d1_u8, d2_u8, d3_u8; |
| compute_basic_avg_4x4_rvv(dd0, dd1, dd2, dd3, d0, d1, d2, d3, |
| round_offset, &d0_u8, &d1_u8, &d2_u8, &d3_u8, |
| vl); |
| |
| // store result |
| __riscv_vse8_v_u8mf4(dst8 + 0 * dst8_stride, d0_u8, vl); |
| __riscv_vse8_v_u8mf4(dst8 + 1 * dst8_stride, d1_u8, vl); |
| __riscv_vse8_v_u8mf4(dst8 + 2 * dst8_stride, d2_u8, vl); |
| __riscv_vse8_v_u8mf4(dst8 + 3 * dst8_stride, d3_u8, vl); |
| |
| // update pointers |
| src += 4 * src_stride; |
| dst += 4 * dst_stride; |
| dst8 += 4 * dst8_stride; |
| height -= 4; |
| } while (height != 0); |
| } else { |
| vuint16m1_t round_offset_u16 = __riscv_vmv_v_x_u16m1(round_offset, vl); |
| do { |
| const uint8_t *s = src; |
| CONV_BUF_TYPE *d = dst; |
| uint8_t *d_u8 = dst8; |
| int width = w; |
| |
| do { |
| // load 4 block of uint8_t |
| vuint8mf2_t s0, s1, s2, s3; |
| load_u8_8x4(s, src_stride, &s0, &s1, &s2, &s3, vl); |
| |
| // apply shift and add round offset |
| vuint16m1_t d0 = |
| __riscv_vwmaccu_vx_u16m1(round_offset_u16, shift_by_bits, s0, vl); |
| vuint16m1_t d1 = |
| __riscv_vwmaccu_vx_u16m1(round_offset_u16, shift_by_bits, s1, vl); |
| vuint16m1_t d2 = |
| __riscv_vwmaccu_vx_u16m1(round_offset_u16, shift_by_bits, s2, vl); |
| vuint16m1_t d3 = |
| __riscv_vwmaccu_vx_u16m1(round_offset_u16, shift_by_bits, s3, vl); |
| |
| // load existing dst values |
| vuint16m1_t dd0, dd1, dd2, dd3; |
| load_u16_8x4(d, dst_stride, &dd0, &dd1, &dd2, &dd3, vl); |
| |
| // compute basic average |
| vuint8mf2_t d0_u8, d1_u8, d2_u8, d3_u8; |
| compute_basic_avg_8x4_rvv(dd0, dd1, dd2, dd3, d0, d1, d2, d3, |
| round_offset, &d0_u8, &d1_u8, &d2_u8, &d3_u8, |
| vl); |
| |
| // store results |
| store_u8_8x4(d_u8, dst8_stride, d0_u8, d1_u8, d2_u8, d3_u8, vl); |
| |
| // update pointers |
| s += vl; |
| d += vl; |
| d_u8 += vl; |
| width -= vl; |
| } while (width != 0); |
| src += 4 * src_stride; |
| dst += 4 * dst_stride; |
| dst8 += 4 * dst8_stride; |
| height -= 4; |
| } while (height != 0); |
| } |
| } |
| |
| static inline void dist_wtd_convolve_2d_copy_rvv(const uint8_t *src, |
| int src_stride, int w, int h, |
| ConvolveParams *conv_params) { |
| assert(w % 4 == 0); |
| assert(h % 4 == 0); |
| |
| const int bd = 8; |
| const int offset_bits = bd + 2 * FILTER_BITS - ROUND0_BITS; |
| const uint16_t round_offset = (1 << (offset_bits - COMPOUND_ROUND1_BITS)) + |
| (1 << (offset_bits - COMPOUND_ROUND1_BITS - 1)); |
| const uint8_t shift_by_bits = 1 << (FILTER_BITS - ROUND0_BITS); |
| |
| CONV_BUF_TYPE *dst = conv_params->dst; |
| const int dst_stride = conv_params->dst_stride; |
| int height = h; |
| size_t vl = __riscv_vsetvl_e16m1(w); |
| |
| if (w == 4) { |
| vuint16mf2_t round_offset_u16 = __riscv_vmv_v_x_u16mf2(round_offset, vl); |
| do { |
| // load 4 block of uint8_t |
| vuint8mf4_t s0 = __riscv_vle8_v_u8mf4(src + 0 * src_stride, vl); |
| vuint8mf4_t s1 = __riscv_vle8_v_u8mf4(src + 1 * src_stride, vl); |
| vuint8mf4_t s2 = __riscv_vle8_v_u8mf4(src + 2 * src_stride, vl); |
| vuint8mf4_t s3 = __riscv_vle8_v_u8mf4(src + 3 * src_stride, vl); |
| |
| // apply shift and add round offset |
| vuint16mf2_t d0 = |
| __riscv_vwmaccu_vx_u16mf2(round_offset_u16, shift_by_bits, s0, vl); |
| vuint16mf2_t d1 = |
| __riscv_vwmaccu_vx_u16mf2(round_offset_u16, shift_by_bits, s1, vl); |
| vuint16mf2_t d2 = |
| __riscv_vwmaccu_vx_u16mf2(round_offset_u16, shift_by_bits, s2, vl); |
| vuint16mf2_t d3 = |
| __riscv_vwmaccu_vx_u16mf2(round_offset_u16, shift_by_bits, s3, vl); |
| |
| // store result |
| store_u16_4x4(dst, dst_stride, d0, d1, d2, d3, vl); |
| |
| // update pointers |
| src += 4 * src_stride; |
| dst += 4 * dst_stride; |
| height -= 4; |
| } while (height != 0); |
| } else { |
| vuint16m1_t round_offset_u16 = __riscv_vmv_v_x_u16m1(round_offset, vl); |
| do { |
| const uint8_t *s = src; |
| CONV_BUF_TYPE *d = dst; |
| int width = w; |
| |
| do { |
| // load 4 block of uint8_t |
| vuint8mf2_t s0, s1, s2, s3; |
| load_u8_8x4(s, src_stride, &s0, &s1, &s2, &s3, vl); |
| |
| // apply shift and add round offset |
| vuint16m1_t d0 = |
| __riscv_vwmaccu_vx_u16m1(round_offset_u16, shift_by_bits, s0, vl); |
| vuint16m1_t d1 = |
| __riscv_vwmaccu_vx_u16m1(round_offset_u16, shift_by_bits, s1, vl); |
| vuint16m1_t d2 = |
| __riscv_vwmaccu_vx_u16m1(round_offset_u16, shift_by_bits, s2, vl); |
| vuint16m1_t d3 = |
| __riscv_vwmaccu_vx_u16m1(round_offset_u16, shift_by_bits, s3, vl); |
| |
| // store results |
| store_u16_8x4(d, dst_stride, d0, d1, d2, d3, vl); |
| |
| // update pointers |
| s += vl; |
| d += vl; |
| width -= vl; |
| } while (width != 0); |
| src += 4 * src_stride; |
| dst += 4 * dst_stride; |
| height -= 4; |
| } while (height != 0); |
| } |
| } |
| |
| void av1_dist_wtd_convolve_2d_copy_rvv(const uint8_t *src, int src_stride, |
| uint8_t *dst8, int dst8_stride, int w, |
| int h, ConvolveParams *conv_params) { |
| if (conv_params->do_average) { |
| if (UNLIKELY(conv_params->use_dist_wtd_comp_avg)) { |
| dist_wtd_convolve_2d_copy_dist_wtd_avg_rvv( |
| src, src_stride, dst8, dst8_stride, w, h, conv_params); |
| } else { |
| dist_wtd_convolve_2d_copy_avg_rvv(src, src_stride, dst8, dst8_stride, w, |
| h, conv_params); |
| } |
| } else { |
| dist_wtd_convolve_2d_copy_rvv(src, src_stride, w, h, conv_params); |
| } |
| } |
| |
| static inline vuint16m1_t convolve4_4_x_rvv( |
| const vint16m1_t s0, const vint16m1_t s1, const vint16m1_t s2, |
| const vint16m1_t s3, const int16_t *filter, const int16_t round_offset, |
| size_t vl) { |
| vint16m1_t sum = __riscv_vmul_vx_i16m1(s0, filter[0], vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[1], s1, vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[2], s2, vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[3], s3, vl); |
| |
| // Right shift with rounding: vrsra_n_s16(round_offset, sum, ROUND0_BITS - 1) |
| // vrsra_n_s16(a, b, n) = a + ((b + (1 << (n-1))) >> n) |
| // We halved the convolution filter values so -1 from the right shift. |
| vint16m1_t res = __riscv_vadd_vx_i16m1(sum, 1 << ((ROUND0_BITS - 1) - 1), vl); |
| res = __riscv_vsra_vx_i16m1(res, ROUND0_BITS - 1, vl); |
| res = __riscv_vadd_vx_i16m1(res, round_offset, vl); |
| |
| // Reinterpret as uint16 |
| return __riscv_vreinterpret_v_i16m1_u16m1(res); |
| } |
| |
| static inline vuint16m1_t convolve8_8_x_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 int16_t round_offset, size_t vl) { |
| vint16m1_t sum = __riscv_vmul_vx_i16m1(s0, filter[0], vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[1], s1, vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[2], s2, vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[3], s3, vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[4], s4, vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[5], s5, vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[6], s6, vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[7], s7, vl); |
| |
| vint16m1_t res = __riscv_vadd_vx_i16m1(sum, 1 << ((ROUND0_BITS - 1) - 1), vl); |
| res = __riscv_vsra_vx_i16m1(res, ROUND0_BITS - 1, vl); |
| res = __riscv_vadd_vx_i16m1(res, round_offset, vl); |
| |
| // Reinterpret as uint16 |
| return __riscv_vreinterpret_v_i16m1_u16m1(res); |
| } |
| |
| static inline void transpose_elems_inplace_u16_8x8_rvv( |
| vuint16m1_t *d0, vuint16m1_t *d1, vuint16m1_t *d2, vuint16m1_t *d3, |
| vuint16m1_t *d4, vuint16m1_t *d5, vuint16m1_t *d6, vuint16m1_t *d7, |
| const vuint16m8_t trans_index) { |
| vuint16m8_t matrix_8x8 = __riscv_vundefined_u16m8(); |
| matrix_8x8 = __riscv_vset_v_u16m1_u16m8(matrix_8x8, 0, *d0); |
| matrix_8x8 = __riscv_vset_v_u16m1_u16m8(matrix_8x8, 1, *d1); |
| matrix_8x8 = __riscv_vset_v_u16m1_u16m8(matrix_8x8, 2, *d2); |
| matrix_8x8 = __riscv_vset_v_u16m1_u16m8(matrix_8x8, 3, *d3); |
| matrix_8x8 = __riscv_vset_v_u16m1_u16m8(matrix_8x8, 4, *d4); |
| matrix_8x8 = __riscv_vset_v_u16m1_u16m8(matrix_8x8, 5, *d5); |
| matrix_8x8 = __riscv_vset_v_u16m1_u16m8(matrix_8x8, 6, *d6); |
| matrix_8x8 = __riscv_vset_v_u16m1_u16m8(matrix_8x8, 7, *d7); |
| |
| // transpose_elems_inplace_u16_8x8 |
| vuint16m8_t trans_matrix = |
| __riscv_vrgather_vv_u16m8(matrix_8x8, trans_index, 64); |
| |
| *d0 = __riscv_vget_v_u16m8_u16m1(trans_matrix, 0); |
| *d1 = __riscv_vget_v_u16m8_u16m1(trans_matrix, 1); |
| *d2 = __riscv_vget_v_u16m8_u16m1(trans_matrix, 2); |
| *d3 = __riscv_vget_v_u16m8_u16m1(trans_matrix, 3); |
| *d4 = __riscv_vget_v_u16m8_u16m1(trans_matrix, 4); |
| *d5 = __riscv_vget_v_u16m8_u16m1(trans_matrix, 5); |
| *d6 = __riscv_vget_v_u16m8_u16m1(trans_matrix, 6); |
| *d7 = __riscv_vget_v_u16m8_u16m1(trans_matrix, 7); |
| } |
| |
| static inline void dist_wtd_convolve_x_dist_wtd_avg_rvv( |
| const uint8_t *src, int src_stride, uint8_t *dst8, int dst8_stride, int w, |
| int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, |
| ConvolveParams *conv_params) { |
| assert(w % 4 == 0); |
| assert(h % 4 == 0); |
| |
| const int bd = 8; |
| const int offset_bits = bd + 2 * FILTER_BITS - ROUND0_BITS; |
| const int16_t round_offset = (1 << (offset_bits - COMPOUND_ROUND1_BITS)) + |
| (1 << (offset_bits - COMPOUND_ROUND1_BITS - 1)); |
| const uint16_t fwd_offset = conv_params->fwd_offset; |
| const uint16_t bck_offset = conv_params->bck_offset; |
| |
| // Horizontal filter. |
| const int16_t *x_filter_ptr = av1_get_interp_filter_subpel_kernel( |
| filter_params_x, subpel_x_qn & SUBPEL_MASK); |
| |
| const int horiz_offset = filter_params_x->taps / 2 - 1; |
| const uint8_t *src_ptr = src - horiz_offset; |
| CONV_BUF_TYPE *dst_ptr = conv_params->dst; |
| uint8_t *dst8_ptr = dst8; |
| int dst_stride = conv_params->dst_stride; |
| int height = h; |
| size_t vl = __riscv_vsetvl_e16m1(w); |
| |
| if (w == 4) { |
| // 4-tap filters are used for blocks having width <= 4. |
| // Filter values are even, so halve to reduce intermediate precision reqs. |
| const int16_t *x_filter = x_filter_ptr + 2; |
| int16_t filter[4]; |
| for (int i = 0; i < 4; i++) filter[i] = *x_filter++ >> 1; |
| |
| src_ptr += 2; |
| |
| do { |
| // Load |
| vuint8mf2_t t0 = __riscv_vle8_v_u8mf2(src_ptr + 0 * src_stride, |
| 8); // a0 a1 a2 a3 a4 a5 a6 a7 |
| vuint8mf2_t t1 = __riscv_vle8_v_u8mf2(src_ptr + 1 * src_stride, |
| 8); // b0 b1 b2 b3 b4 b5 b6 b7 |
| |
| __builtin_prefetch(dst_ptr); |
| __builtin_prefetch(dst8_ptr); |
| |
| // Widen to 16-bit |
| vint16m1_t s00 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t0, 8)); |
| vint16m1_t s01 = |
| __riscv_vslide1down_vx_i16m1(s00, 0, 8); // a1 a2 a3 a4 a5 a6 a7 0 |
| vint16m1_t s02 = |
| __riscv_vslide1down_vx_i16m1(s01, 0, 8); // a2 a3 a4 a5 a6 a7 0 0 |
| vint16m1_t s03 = |
| __riscv_vslide1down_vx_i16m1(s02, 0, 8); // a3 a4 a5 a6 a7 0 0 0 |
| |
| vint16m1_t s10 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t1, 8)); |
| vint16m1_t s11 = |
| __riscv_vslide1down_vx_i16m1(s10, 0, 8); // b1 b2 b3 b4 b5 b6 b7 0 |
| vint16m1_t s12 = |
| __riscv_vslide1down_vx_i16m1(s11, 0, 8); // b2 b3 b4 b5 b6 b7 0 0 |
| vint16m1_t s13 = |
| __riscv_vslide1down_vx_i16m1(s12, 0, 8); // b3 b4 b5 b6 b7 0 0 0 |
| |
| // Perform convolution |
| vuint8mf2_t d01, d11; |
| vuint16m1_t d0 = |
| convolve4_4_x_rvv(s00, s01, s02, s03, filter, round_offset, vl); |
| vuint16m1_t d1 = |
| convolve4_4_x_rvv(s10, s11, s12, s13, filter, round_offset, vl); |
| |
| // weighted average |
| vuint16m1_t dd0 = __riscv_vle16_v_u16m1(dst_ptr + 0 * dst_stride, vl); |
| vuint16m1_t dd1 = __riscv_vle16_v_u16m1(dst_ptr + 1 * dst_stride, vl); |
| compute_dist_wtd_avg_4x1_rvv(dd0, d0, fwd_offset, bck_offset, |
| round_offset, &d01, vl); |
| compute_dist_wtd_avg_4x1_rvv(dd1, d1, fwd_offset, bck_offset, |
| round_offset, &d11, vl); |
| |
| // Store result |
| __riscv_vse8_v_u8mf2(dst8_ptr + 0 * dst8_stride, d01, vl); |
| __riscv_vse8_v_u8mf2(dst8_ptr + 1 * dst8_stride, d11, vl); |
| |
| src_ptr += 2 * src_stride; |
| dst_ptr += 2 * dst_stride; |
| dst8_ptr += 2 * dst8_stride; |
| height -= 2; |
| } while (height >= 2); |
| } else { |
| // Filter values are even, so halve to reduce intermediate precision reqs. |
| int16_t filter[8]; |
| for (int i = 0; i < 8; i++) filter[i] = *x_filter_ptr++ >> 1; |
| |
| // for VLEN = 128 case |
| if ((w > 8) && (vl == 8) && (height >= 8)) { |
| vuint16m8_t trans_index = __riscv_vle16_v_u16m8(trans_data, 64); |
| |
| while (height >= 8) { |
| const uint8_t *s = src_ptr; |
| CONV_BUF_TYPE *d = dst_ptr; |
| uint8_t *d_u8 = dst8_ptr; |
| int width = w; |
| |
| // Load |
| vuint8mf2_t t0, t1, t2, t3, t4, t5, t6, t7; |
| load_stride_u8_8x8(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, |
| &t7, vl); |
| |
| vint16m1_t s0 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t0, vl)); |
| vint16m1_t s1 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t1, vl)); |
| vint16m1_t s2 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t2, vl)); |
| vint16m1_t s3 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t3, vl)); |
| vint16m1_t s4 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t4, vl)); |
| vint16m1_t s5 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t5, vl)); |
| vint16m1_t s6 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t6, vl)); |
| |
| __builtin_prefetch(d + 0 * dst_stride); |
| __builtin_prefetch(d + 1 * dst_stride); |
| __builtin_prefetch(d + 2 * dst_stride); |
| __builtin_prefetch(d + 3 * dst_stride); |
| __builtin_prefetch(d + 4 * dst_stride); |
| __builtin_prefetch(d + 5 * dst_stride); |
| __builtin_prefetch(d + 6 * dst_stride); |
| __builtin_prefetch(d + 7 * dst_stride); |
| |
| s += 7; |
| |
| do { |
| load_stride_u8_8x8(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, |
| &t7, vl); |
| |
| vint16m1_t s7 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t0, vl)); |
| vint16m1_t s8 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t1, vl)); |
| vint16m1_t s9 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t2, vl)); |
| vint16m1_t s10 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t3, vl)); |
| vint16m1_t s11 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t4, vl)); |
| vint16m1_t s12 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t5, vl)); |
| vint16m1_t s13 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t6, vl)); |
| vint16m1_t s14 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t7, vl)); |
| |
| // Perform convolution |
| vuint16m1_t d0 = convolve8_8_x_rvv(s0, s1, s2, s3, s4, s5, s6, s7, |
| filter, round_offset, vl); |
| vuint16m1_t d1 = convolve8_8_x_rvv(s1, s2, s3, s4, s5, s6, s7, s8, |
| filter, round_offset, vl); |
| vuint16m1_t d2 = convolve8_8_x_rvv(s2, s3, s4, s5, s6, s7, s8, s9, |
| filter, round_offset, vl); |
| vuint16m1_t d3 = convolve8_8_x_rvv(s3, s4, s5, s6, s7, s8, s9, s10, |
| filter, round_offset, vl); |
| vuint16m1_t d4 = convolve8_8_x_rvv(s4, s5, s6, s7, s8, s9, s10, s11, |
| filter, round_offset, vl); |
| vuint16m1_t d5 = convolve8_8_x_rvv(s5, s6, s7, s8, s9, s10, s11, s12, |
| filter, round_offset, vl); |
| vuint16m1_t d6 = convolve8_8_x_rvv(s6, s7, s8, s9, s10, s11, s12, s13, |
| filter, round_offset, vl); |
| vuint16m1_t d7 = convolve8_8_x_rvv(s7, s8, s9, s10, s11, s12, s13, |
| s14, filter, round_offset, vl); |
| |
| // transpose |
| transpose_elems_inplace_u16_8x8_rvv(&d0, &d1, &d2, &d3, &d4, &d5, &d6, |
| &d7, trans_index); |
| |
| vuint16m1_t dd0, dd1, dd2, dd3; |
| load_u16_8x4(d, dst_stride, &dd0, &dd1, &dd2, &dd3, vl); |
| |
| vuint8mf2_t d0_u8, d1_u8, d2_u8, d3_u8; |
| compute_dist_wtd_avg_8x4_rvv(dd0, dd1, dd2, dd3, d0, d1, d2, d3, |
| fwd_offset, bck_offset, round_offset, |
| &d0_u8, &d1_u8, &d2_u8, &d3_u8, vl); |
| store_u8_8x4(d_u8, dst8_stride, d0_u8, d1_u8, d2_u8, d3_u8, vl); |
| |
| vuint16m1_t dd4, dd5, dd6, dd7; |
| load_u16_8x4(d + 4 * dst_stride, dst_stride, &dd4, &dd5, &dd6, &dd7, |
| vl); |
| |
| vuint8mf2_t d4_u8, d5_u8, d6_u8, d7_u8; |
| compute_dist_wtd_avg_8x4_rvv(dd4, dd5, dd6, dd7, d4, d5, d6, d7, |
| fwd_offset, bck_offset, round_offset, |
| &d4_u8, &d5_u8, &d6_u8, &d7_u8, vl); |
| store_u8_8x4(d_u8 + 4 * dst8_stride, dst8_stride, d4_u8, d5_u8, d6_u8, |
| d7_u8, vl); |
| |
| // Update sliding window |
| s0 = __riscv_vmv_v_v_i16m1(s8, vl); |
| s1 = __riscv_vmv_v_v_i16m1(s9, vl); |
| s2 = __riscv_vmv_v_v_i16m1(s10, vl); |
| s3 = __riscv_vmv_v_v_i16m1(s11, vl); |
| s4 = __riscv_vmv_v_v_i16m1(s12, vl); |
| s5 = __riscv_vmv_v_v_i16m1(s13, vl); |
| s6 = __riscv_vmv_v_v_i16m1(s14, vl); |
| s += 8; |
| d += 8; |
| d_u8 += 8; |
| width -= 8; |
| } while (width != 0); |
| src_ptr += 8 * src_stride; |
| dst_ptr += 8 * dst_stride; |
| dst8_ptr += 8 * dst8_stride; |
| height -= 8; |
| } |
| } |
| |
| while (height > 0) { |
| const uint8_t *s = src_ptr; |
| CONV_BUF_TYPE *d = dst_ptr; |
| uint8_t *d_u8 = dst8_ptr; |
| int width = w; |
| |
| do { |
| // Load |
| vuint8mf2_t t0, t1, t2, t3, t4, t5, t6, t7; |
| load_u8_8x8(s, 1, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7, vl); |
| |
| // Convert to 16-bit integers |
| vint16m1_t s0 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t0, vl)); |
| vint16m1_t s1 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t1, vl)); |
| vint16m1_t s2 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t2, vl)); |
| vint16m1_t s3 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t3, vl)); |
| vint16m1_t s4 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t4, vl)); |
| vint16m1_t s5 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t5, vl)); |
| vint16m1_t s6 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t6, vl)); |
| vint16m1_t s7 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t7, vl)); |
| |
| // Perform convolution |
| vuint16m1_t d0 = convolve8_8_x_rvv(s0, s1, s2, s3, s4, s5, s6, s7, |
| filter, round_offset, vl); |
| vuint16m1_t dd0 = __riscv_vle16_v_u16m1(d, vl); |
| |
| vuint8mf2_t d0_u8; |
| compute_dist_wtd_avg_8x1_rvv(dd0, d0, fwd_offset, bck_offset, |
| round_offset, &d0_u8, vl); |
| |
| // Store result |
| __riscv_vse8_v_u8mf2(d_u8, d0_u8, vl); |
| |
| s += vl; |
| d += vl; |
| d_u8 += vl; |
| width -= vl; |
| } while (width > 0); |
| src_ptr += src_stride; |
| dst_ptr += dst_stride; |
| dst8_ptr += dst8_stride; |
| height--; |
| } |
| } |
| } |
| |
| static inline void dist_wtd_convolve_x_avg_rvv( |
| const uint8_t *src, int src_stride, uint8_t *dst8, int dst8_stride, int w, |
| int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, |
| ConvolveParams *conv_params) { |
| const int bd = 8; |
| const int offset_bits = bd + 2 * FILTER_BITS - ROUND0_BITS; |
| const int16_t round_offset = (1 << (offset_bits - COMPOUND_ROUND1_BITS)) + |
| (1 << (offset_bits - COMPOUND_ROUND1_BITS - 1)); |
| // Horizontal filter. |
| const int16_t *x_filter_ptr = av1_get_interp_filter_subpel_kernel( |
| filter_params_x, subpel_x_qn & SUBPEL_MASK); |
| |
| const int horiz_offset = filter_params_x->taps / 2 - 1; |
| const uint8_t *src_ptr = src - horiz_offset; |
| CONV_BUF_TYPE *dst_ptr = conv_params->dst; |
| uint8_t *dst8_ptr = dst8; |
| int dst_stride = conv_params->dst_stride; |
| int height = h; |
| |
| size_t vl = __riscv_vsetvl_e16m1(w); |
| if (w == 4) { |
| // 4-tap filters are used for blocks having width <= 4. |
| // Filter values are even, so halve to reduce intermediate precision reqs. |
| const int16_t *x_filter = x_filter_ptr + 2; |
| int16_t filter[4]; |
| for (int i = 0; i < 4; i++) filter[i] = *x_filter++ >> 1; |
| |
| src_ptr += 2; |
| |
| do { |
| // Load |
| vuint8mf2_t t0 = __riscv_vle8_v_u8mf2(src_ptr + 0 * src_stride, |
| 8); // a0 a1 a2 a3 a4 a5 a6 a7 |
| vuint8mf2_t t1 = __riscv_vle8_v_u8mf2(src_ptr + 1 * src_stride, |
| 8); // b0 b1 b2 b3 b4 b5 b6 b7 |
| |
| __builtin_prefetch(dst_ptr); |
| __builtin_prefetch(dst8_ptr); |
| |
| // Widen to 16-bit |
| vint16m1_t s00 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t0, 8)); |
| vint16m1_t s01 = |
| __riscv_vslide1down_vx_i16m1(s00, 0, 8); // a1 a2 a3 a4 a5 a6 a7 0 |
| vint16m1_t s02 = |
| __riscv_vslide1down_vx_i16m1(s01, 0, 8); // a2 a3 a4 a5 a6 a7 0 0 |
| vint16m1_t s03 = |
| __riscv_vslide1down_vx_i16m1(s02, 0, 8); // a3 a4 a5 a6 a7 0 0 0 |
| |
| vint16m1_t s10 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t1, 8)); |
| vint16m1_t s11 = |
| __riscv_vslide1down_vx_i16m1(s10, 0, 8); // b1 b2 b3 b4 b5 b6 b7 0 |
| vint16m1_t s12 = |
| __riscv_vslide1down_vx_i16m1(s11, 0, 8); // b2 b3 b4 b5 b6 b7 0 0 |
| vint16m1_t s13 = |
| __riscv_vslide1down_vx_i16m1(s12, 0, 8); // b3 b4 b5 b6 b7 0 0 0 |
| |
| // Perform convolution |
| vuint8mf2_t d01, d11; |
| vuint16m1_t d0 = |
| convolve4_4_x_rvv(s00, s01, s02, s03, filter, round_offset, vl); |
| vuint16m1_t d1 = |
| convolve4_4_x_rvv(s10, s11, s12, s13, filter, round_offset, vl); |
| |
| // average |
| vuint16m1_t dd0 = __riscv_vle16_v_u16m1(dst_ptr + 0 * dst_stride, vl); |
| vuint16m1_t dd1 = __riscv_vle16_v_u16m1(dst_ptr + 1 * dst_stride, vl); |
| compute_basic_avg_4x1_rvv(dd0, d0, round_offset, &d01, vl); |
| compute_basic_avg_4x1_rvv(dd1, d1, round_offset, &d11, vl); |
| |
| // Store result |
| __riscv_vse8_v_u8mf2(dst8_ptr + 0 * dst8_stride, d01, vl); |
| __riscv_vse8_v_u8mf2(dst8_ptr + 1 * dst8_stride, d11, vl); |
| |
| src_ptr += 2 * src_stride; |
| dst_ptr += 2 * dst_stride; |
| dst8_ptr += 2 * dst8_stride; |
| height -= 2; |
| } while (height >= 2); |
| } else { |
| // Filter values are even, so halve to reduce intermediate precision reqs. |
| int16_t filter[8]; |
| for (int i = 0; i < 8; i++) filter[i] = *x_filter_ptr++ >> 1; |
| |
| // for VLEN = 128 case |
| if ((w > 8) && (vl == 8) && (height >= 8)) { |
| vuint16m8_t trans_index = __riscv_vle16_v_u16m8(trans_data, 64); |
| |
| while (height >= 8) { |
| const uint8_t *s = src_ptr; |
| CONV_BUF_TYPE *d = dst_ptr; |
| uint8_t *d_u8 = dst8_ptr; |
| int width = w; |
| |
| // Load |
| vuint8mf2_t t0, t1, t2, t3, t4, t5, t6, t7; |
| load_stride_u8_8x8(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, |
| &t7, vl); |
| |
| vint16m1_t s0 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t0, vl)); |
| vint16m1_t s1 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t1, vl)); |
| vint16m1_t s2 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t2, vl)); |
| vint16m1_t s3 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t3, vl)); |
| vint16m1_t s4 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t4, vl)); |
| vint16m1_t s5 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t5, vl)); |
| vint16m1_t s6 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t6, vl)); |
| |
| __builtin_prefetch(d + 0 * dst_stride); |
| __builtin_prefetch(d + 1 * dst_stride); |
| __builtin_prefetch(d + 2 * dst_stride); |
| __builtin_prefetch(d + 3 * dst_stride); |
| __builtin_prefetch(d + 4 * dst_stride); |
| __builtin_prefetch(d + 5 * dst_stride); |
| __builtin_prefetch(d + 6 * dst_stride); |
| __builtin_prefetch(d + 7 * dst_stride); |
| |
| s += 7; |
| |
| do { |
| load_stride_u8_8x8(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, |
| &t7, vl); |
| |
| vint16m1_t s7 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t0, vl)); |
| vint16m1_t s8 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t1, vl)); |
| vint16m1_t s9 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t2, vl)); |
| vint16m1_t s10 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t3, vl)); |
| vint16m1_t s11 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t4, vl)); |
| vint16m1_t s12 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t5, vl)); |
| vint16m1_t s13 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t6, vl)); |
| vint16m1_t s14 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t7, vl)); |
| |
| // Perform convolution |
| vuint16m1_t d0 = convolve8_8_x_rvv(s0, s1, s2, s3, s4, s5, s6, s7, |
| filter, round_offset, vl); |
| vuint16m1_t d1 = convolve8_8_x_rvv(s1, s2, s3, s4, s5, s6, s7, s8, |
| filter, round_offset, vl); |
| vuint16m1_t d2 = convolve8_8_x_rvv(s2, s3, s4, s5, s6, s7, s8, s9, |
| filter, round_offset, vl); |
| vuint16m1_t d3 = convolve8_8_x_rvv(s3, s4, s5, s6, s7, s8, s9, s10, |
| filter, round_offset, vl); |
| vuint16m1_t d4 = convolve8_8_x_rvv(s4, s5, s6, s7, s8, s9, s10, s11, |
| filter, round_offset, vl); |
| vuint16m1_t d5 = convolve8_8_x_rvv(s5, s6, s7, s8, s9, s10, s11, s12, |
| filter, round_offset, vl); |
| vuint16m1_t d6 = convolve8_8_x_rvv(s6, s7, s8, s9, s10, s11, s12, s13, |
| filter, round_offset, vl); |
| vuint16m1_t d7 = convolve8_8_x_rvv(s7, s8, s9, s10, s11, s12, s13, |
| s14, filter, round_offset, vl); |
| |
| // transpose |
| transpose_elems_inplace_u16_8x8_rvv(&d0, &d1, &d2, &d3, &d4, &d5, &d6, |
| &d7, trans_index); |
| |
| vuint16m1_t dd0, dd1, dd2, dd3; |
| load_u16_8x4(d, dst_stride, &dd0, &dd1, &dd2, &dd3, vl); |
| |
| vuint8mf2_t d0_u8, d1_u8, d2_u8, d3_u8; |
| compute_basic_avg_8x4_rvv(dd0, dd1, dd2, dd3, d0, d1, d2, d3, |
| round_offset, &d0_u8, &d1_u8, &d2_u8, |
| &d3_u8, vl); |
| store_u8_8x4(d_u8, dst8_stride, d0_u8, d1_u8, d2_u8, d3_u8, vl); |
| |
| vuint16m1_t dd4, dd5, dd6, dd7; |
| load_u16_8x4(d + 4 * dst_stride, dst_stride, &dd4, &dd5, &dd6, &dd7, |
| vl); |
| |
| vuint8mf2_t d4_u8, d5_u8, d6_u8, d7_u8; |
| compute_basic_avg_8x4_rvv(dd4, dd5, dd6, dd7, d4, d5, d6, d7, |
| round_offset, &d4_u8, &d5_u8, &d6_u8, |
| &d7_u8, vl); |
| store_u8_8x4(d_u8 + 4 * dst8_stride, dst8_stride, d4_u8, d5_u8, d6_u8, |
| d7_u8, vl); |
| |
| // Update sliding window |
| s0 = __riscv_vmv_v_v_i16m1(s8, vl); |
| s1 = __riscv_vmv_v_v_i16m1(s9, vl); |
| s2 = __riscv_vmv_v_v_i16m1(s10, vl); |
| s3 = __riscv_vmv_v_v_i16m1(s11, vl); |
| s4 = __riscv_vmv_v_v_i16m1(s12, vl); |
| s5 = __riscv_vmv_v_v_i16m1(s13, vl); |
| s6 = __riscv_vmv_v_v_i16m1(s14, vl); |
| s += 8; |
| d += 8; |
| d_u8 += 8; |
| width -= 8; |
| } while (width != 0); |
| src_ptr += 8 * src_stride; |
| dst_ptr += 8 * dst_stride; |
| dst8_ptr += 8 * dst8_stride; |
| height -= 8; |
| } |
| } |
| |
| while (height > 0) { |
| const uint8_t *s = src_ptr; |
| CONV_BUF_TYPE *d = dst_ptr; |
| uint8_t *d_u8 = dst8_ptr; |
| int width = w; |
| |
| do { |
| // Load |
| vuint8mf2_t t0, t1, t2, t3, t4, t5, t6, t7; |
| load_u8_8x8(s, 1, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7, vl); |
| |
| // Convert to 16-bit integers |
| vint16m1_t s0 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t0, vl)); |
| vint16m1_t s1 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t1, vl)); |
| vint16m1_t s2 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t2, vl)); |
| vint16m1_t s3 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t3, vl)); |
| vint16m1_t s4 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t4, vl)); |
| vint16m1_t s5 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t5, vl)); |
| vint16m1_t s6 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t6, vl)); |
| vint16m1_t s7 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t7, vl)); |
| |
| // Perform convolution |
| vuint16m1_t d0 = convolve8_8_x_rvv(s0, s1, s2, s3, s4, s5, s6, s7, |
| filter, round_offset, vl); |
| vuint16m1_t dd0 = __riscv_vle16_v_u16m1(d, vl); |
| |
| vuint8mf2_t d0_u8; |
| compute_basic_avg_8x1_rvv(dd0, d0, round_offset, &d0_u8, vl); |
| |
| // Store result |
| __riscv_vse8_v_u8mf2(d_u8, d0_u8, vl); |
| |
| s += vl; |
| d += vl; |
| d_u8 += vl; |
| width -= vl; |
| } while (width > 0); |
| src_ptr += src_stride; |
| dst_ptr += dst_stride; |
| dst8_ptr += dst8_stride; |
| height--; |
| } |
| } |
| } |
| |
| static inline void dist_wtd_convolve_x_rvv( |
| const uint8_t *src, int src_stride, int w, int h, |
| const InterpFilterParams *filter_params_x, const int subpel_x_qn, |
| ConvolveParams *conv_params) { |
| const int bd = 8; |
| const int offset_bits = bd + 2 * FILTER_BITS - ROUND0_BITS; |
| const int16_t round_offset = (1 << (offset_bits - COMPOUND_ROUND1_BITS)) + |
| (1 << (offset_bits - COMPOUND_ROUND1_BITS - 1)); |
| // Horizontal filter. |
| const int16_t *x_filter_ptr = av1_get_interp_filter_subpel_kernel( |
| filter_params_x, subpel_x_qn & SUBPEL_MASK); |
| |
| const int horiz_offset = filter_params_x->taps / 2 - 1; |
| const uint8_t *src_ptr = src - horiz_offset; |
| CONV_BUF_TYPE *dst_ptr = conv_params->dst; |
| int dst_stride = conv_params->dst_stride; |
| int height = h; |
| |
| size_t vl = __riscv_vsetvl_e16m1(w); |
| if (w == 4) { |
| // 4-tap filters are used for blocks having width <= 4. |
| // Filter values are even, so halve to reduce intermediate precision reqs. |
| const int16_t *x_filter = x_filter_ptr + 2; |
| int16_t filter[4]; |
| for (int i = 0; i < 4; i++) filter[i] = *(x_filter + i) >> 1; |
| |
| src_ptr += 2; |
| |
| do { |
| // Load |
| vuint8mf2_t t0 = __riscv_vle8_v_u8mf2(src_ptr + 0 * src_stride, |
| 8); // a0 a1 a2 a3 a4 a5 a6 a7 |
| vuint8mf2_t t1 = __riscv_vle8_v_u8mf2(src_ptr + 1 * src_stride, |
| 8); // b0 b1 b2 b3 b4 b5 b6 b7 |
| |
| __builtin_prefetch(dst_ptr); |
| |
| // Widen to 16-bit |
| vint16m1_t s00 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t0, 8)); |
| vint16m1_t s01 = |
| __riscv_vslide1down_vx_i16m1(s00, 0, 8); // a1 a2 a3 a4 a5 a6 a7 0 |
| vint16m1_t s02 = |
| __riscv_vslide1down_vx_i16m1(s01, 0, 8); // a2 a3 a4 a5 a6 a7 0 0 |
| vint16m1_t s03 = |
| __riscv_vslide1down_vx_i16m1(s02, 0, 8); // a3 a4 a5 a6 a7 0 0 0 |
| |
| vint16m1_t s10 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t1, 8)); |
| vint16m1_t s11 = |
| __riscv_vslide1down_vx_i16m1(s10, 0, 8); // b1 b2 b3 b4 b5 b6 b7 0 |
| vint16m1_t s12 = |
| __riscv_vslide1down_vx_i16m1(s11, 0, 8); // b2 b3 b4 b5 b6 b7 0 0 |
| vint16m1_t s13 = |
| __riscv_vslide1down_vx_i16m1(s12, 0, 8); // b3 b4 b5 b6 b7 0 0 0 |
| |
| // Perform convolution |
| vuint16m1_t d0 = |
| convolve4_4_x_rvv(s00, s01, s02, s03, filter, round_offset, vl); |
| vuint16m1_t d1 = |
| convolve4_4_x_rvv(s10, s11, s12, s13, filter, round_offset, vl); |
| |
| // Store result |
| __riscv_vse16_v_u16m1(dst_ptr + 0 * dst_stride, d0, vl); |
| __riscv_vse16_v_u16m1(dst_ptr + 1 * dst_stride, d1, vl); |
| |
| src_ptr += 2 * src_stride; |
| dst_ptr += 2 * dst_stride; |
| height -= 2; |
| } while (height >= 2); |
| } else { |
| // Filter values are even, so halve to reduce intermediate precision reqs. |
| int16_t filter[8]; |
| for (int i = 0; i < 8; i++) filter[i] = *x_filter_ptr++ >> 1; |
| |
| // for VLEN = 128 case |
| if ((w > 8) && (vl == 8) && (height >= 8)) { |
| vuint16m8_t trans_index = __riscv_vle16_v_u16m8(trans_data, 64); |
| |
| while (height >= 8) { |
| const uint8_t *s = src_ptr; |
| CONV_BUF_TYPE *d = dst_ptr; |
| int width = w; |
| |
| // Load |
| vuint8mf2_t t0, t1, t2, t3, t4, t5, t6, t7; |
| load_stride_u8_8x8(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, |
| &t7, vl); |
| |
| vint16m1_t s0 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t0, vl)); |
| vint16m1_t s1 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t1, vl)); |
| vint16m1_t s2 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t2, vl)); |
| vint16m1_t s3 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t3, vl)); |
| vint16m1_t s4 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t4, vl)); |
| vint16m1_t s5 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t5, vl)); |
| vint16m1_t s6 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t6, vl)); |
| |
| __builtin_prefetch(d + 0 * dst_stride); |
| __builtin_prefetch(d + 1 * dst_stride); |
| __builtin_prefetch(d + 2 * dst_stride); |
| __builtin_prefetch(d + 3 * dst_stride); |
| __builtin_prefetch(d + 4 * dst_stride); |
| __builtin_prefetch(d + 5 * dst_stride); |
| __builtin_prefetch(d + 6 * dst_stride); |
| __builtin_prefetch(d + 7 * dst_stride); |
| |
| s += 7; |
| |
| do { |
| load_stride_u8_8x8(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, |
| &t7, vl); |
| |
| vint16m1_t s7 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t0, vl)); |
| vint16m1_t s8 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t1, vl)); |
| vint16m1_t s9 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t2, vl)); |
| vint16m1_t s10 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t3, vl)); |
| vint16m1_t s11 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t4, vl)); |
| vint16m1_t s12 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t5, vl)); |
| vint16m1_t s13 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t6, vl)); |
| vint16m1_t s14 = __riscv_vreinterpret_v_u16m1_i16m1( |
| __riscv_vzext_vf2_u16m1(t7, vl)); |
| |
| // Perform convolution |
| vuint16m1_t d0 = convolve8_8_x_rvv(s0, s1, s2, s3, s4, s5, s6, s7, |
| filter, round_offset, vl); |
| vuint16m1_t d1 = convolve8_8_x_rvv(s1, s2, s3, s4, s5, s6, s7, s8, |
| filter, round_offset, vl); |
| vuint16m1_t d2 = convolve8_8_x_rvv(s2, s3, s4, s5, s6, s7, s8, s9, |
| filter, round_offset, vl); |
| vuint16m1_t d3 = convolve8_8_x_rvv(s3, s4, s5, s6, s7, s8, s9, s10, |
| filter, round_offset, vl); |
| vuint16m1_t d4 = convolve8_8_x_rvv(s4, s5, s6, s7, s8, s9, s10, s11, |
| filter, round_offset, vl); |
| vuint16m1_t d5 = convolve8_8_x_rvv(s5, s6, s7, s8, s9, s10, s11, s12, |
| filter, round_offset, vl); |
| vuint16m1_t d6 = convolve8_8_x_rvv(s6, s7, s8, s9, s10, s11, s12, s13, |
| filter, round_offset, vl); |
| vuint16m1_t d7 = convolve8_8_x_rvv(s7, s8, s9, s10, s11, s12, s13, |
| s14, filter, round_offset, vl); |
| |
| // transpose |
| transpose_elems_inplace_u16_8x8_rvv(&d0, &d1, &d2, &d3, &d4, &d5, &d6, |
| &d7, trans_index); |
| |
| // Store result |
| store_u16_8x8(d, dst_stride, d0, d1, d2, d3, d4, d5, d6, d7, vl); |
| |
| // Update sliding window |
| s0 = __riscv_vmv_v_v_i16m1(s8, vl); |
| s1 = __riscv_vmv_v_v_i16m1(s9, vl); |
| s2 = __riscv_vmv_v_v_i16m1(s10, vl); |
| s3 = __riscv_vmv_v_v_i16m1(s11, vl); |
| s4 = __riscv_vmv_v_v_i16m1(s12, vl); |
| s5 = __riscv_vmv_v_v_i16m1(s13, vl); |
| s6 = __riscv_vmv_v_v_i16m1(s14, vl); |
| s += 8; |
| d += 8; |
| width -= 8; |
| } while (width != 0); |
| src_ptr += 8 * src_stride; |
| dst_ptr += 8 * dst_stride; |
| height -= 8; |
| } |
| } |
| |
| while (height > 0) { |
| const uint8_t *s = src_ptr; |
| CONV_BUF_TYPE *d = dst_ptr; |
| int width = w; |
| |
| __builtin_prefetch(d); |
| |
| do { |
| // Load |
| vuint8mf2_t t0, t1, t2, t3, t4, t5, t6, t7; |
| load_u8_8x8(s, 1, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7, vl); |
| |
| // Convert to 16-bit integers |
| vint16m1_t s0 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t0, vl)); |
| vint16m1_t s1 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t1, vl)); |
| vint16m1_t s2 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t2, vl)); |
| vint16m1_t s3 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t3, vl)); |
| vint16m1_t s4 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t4, vl)); |
| vint16m1_t s5 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t5, vl)); |
| vint16m1_t s6 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t6, vl)); |
| vint16m1_t s7 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t7, vl)); |
| |
| // Perform convolution |
| vuint16m1_t d0 = convolve8_8_x_rvv(s0, s1, s2, s3, s4, s5, s6, s7, |
| filter, round_offset, vl); |
| |
| // Store result |
| __riscv_vse16_v_u16m1(d, d0, vl); |
| |
| s += vl; |
| d += vl; |
| width -= vl; |
| } while (width > 0); |
| src_ptr += src_stride; |
| dst_ptr += dst_stride; |
| height--; |
| } |
| } |
| } |
| |
| void av1_dist_wtd_convolve_x_rvv(const uint8_t *src, int src_stride, |
| uint8_t *dst8, int dst8_stride, int w, int h, |
| const InterpFilterParams *filter_params_x, |
| const int subpel_x_qn, |
| ConvolveParams *conv_params) { |
| if (conv_params->do_average) { |
| if (UNLIKELY(conv_params->use_dist_wtd_comp_avg)) { |
| dist_wtd_convolve_x_dist_wtd_avg_rvv(src, src_stride, dst8, dst8_stride, |
| w, h, filter_params_x, subpel_x_qn, |
| conv_params); |
| } else { |
| dist_wtd_convolve_x_avg_rvv(src, src_stride, dst8, dst8_stride, w, h, |
| filter_params_x, subpel_x_qn, conv_params); |
| } |
| } else { |
| dist_wtd_convolve_x_rvv(src, src_stride, w, h, filter_params_x, subpel_x_qn, |
| conv_params); |
| } |
| } |
| |
| static inline vuint16mf2_t convolve6_4_y_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 int16_t round_offset, size_t vl) { |
| // Filter values at indices 0 and 7 are 0. |
| vint16mf2_t sum = __riscv_vmul_vx_i16mf2(s0, filter[1], vl); |
| sum = __riscv_vmacc_vx_i16mf2(sum, filter[2], s1, vl); |
| sum = __riscv_vmacc_vx_i16mf2(sum, filter[3], s2, vl); |
| sum = __riscv_vmacc_vx_i16mf2(sum, filter[4], s3, vl); |
| sum = __riscv_vmacc_vx_i16mf2(sum, filter[5], s4, vl); |
| sum = __riscv_vmacc_vx_i16mf2(sum, filter[6], s5, vl); |
| |
| // Right shift with rounding: vrsra_n_s16(round_offset, sum, ROUND0_BITS - 1) |
| // vrsra_n_s16(a, b, n) = a + ((b + (1 << (n-1))) >> n) |
| // We halved the convolution filter values so -1 from the right shift. |
| vint16mf2_t res = |
| __riscv_vadd_vx_i16mf2(sum, 1 << ((ROUND0_BITS - 1) - 1), vl); |
| res = __riscv_vsra_vx_i16mf2(res, ROUND0_BITS - 1, vl); |
| res = __riscv_vadd_vx_i16mf2(res, round_offset, vl); |
| |
| // Reinterpret as uint16 |
| return __riscv_vreinterpret_v_i16mf2_u16mf2(res); |
| } |
| |
| static inline vuint16m1_t convolve6_8_y_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 int16_t round_offset, size_t vl) { |
| // Filter values at indices 0 and 7 are 0. |
| vint16m1_t sum = __riscv_vmul_vx_i16m1(s0, filter[1], vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[2], s1, vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[3], s2, vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[4], s3, vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[5], s4, vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[6], s5, vl); |
| |
| // Right shift with rounding: vrsra_n_s16(round_offset, sum, ROUND0_BITS - 1) |
| // vrsra_n_s16(a, b, n) = a + ((b + (1 << (n-1))) >> n) |
| // We halved the convolution filter values so -1 from the right shift. |
| vint16m1_t res = __riscv_vadd_vx_i16m1(sum, 1 << ((ROUND0_BITS - 1) - 1), vl); |
| res = __riscv_vsra_vx_i16m1(res, ROUND0_BITS - 1, vl); |
| res = __riscv_vadd_vx_i16m1(res, round_offset, vl); |
| |
| // Reinterpret as uint16 |
| return __riscv_vreinterpret_v_i16m1_u16m1(res); |
| } |
| |
| static inline void dist_wtd_convolve_y_6tap_dist_wtd_avg_rvv( |
| const uint8_t *src_ptr, int src_stride, uint8_t *dst8_ptr, |
| const int dst8_stride, int w, int h, const int16_t *y_filter, |
| ConvolveParams *conv_params) { |
| const int bd = 8; |
| const int offset_bits = bd + 2 * FILTER_BITS - ROUND0_BITS; |
| const int16_t round_offset = (1 << (offset_bits - COMPOUND_ROUND1_BITS)) + |
| (1 << (offset_bits - COMPOUND_ROUND1_BITS - 1)); |
| const uint16_t fwd_offset = conv_params->fwd_offset; |
| const uint16_t bck_offset = conv_params->bck_offset; |
| CONV_BUF_TYPE *dst_ptr = conv_params->dst; |
| const int dst_stride = conv_params->dst_stride; |
| int width = w; |
| |
| if (w == 4 || h == 4) { |
| size_t vl = __riscv_vsetvl_e16m1(4); |
| do { |
| const uint8_t *s = src_ptr; |
| CONV_BUF_TYPE *d = dst_ptr; |
| uint8_t *d_u8 = dst8_ptr; |
| int height = h; |
| |
| // Load initial 5 rows of data |
| vuint8mf4_t t0, t1, t2, t3, t4; |
| load_u8_4x5(s, src_stride, &t0, &t1, &t2, &t3, &t4, vl); |
| |
| // Convert to 16-bit |
| vint16mf2_t s0 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t0, vl)); |
| vint16mf2_t s1 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t1, vl)); |
| vint16mf2_t s2 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t2, vl)); |
| vint16mf2_t s3 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t3, vl)); |
| vint16mf2_t s4 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t4, vl)); |
| |
| s += 5 * src_stride; |
| |
| do { |
| // Load next rows of data |
| vuint8mf4_t t5 = __riscv_vle8_v_u8mf4(s + 0 * src_stride, vl); |
| vuint8mf4_t t6 = __riscv_vle8_v_u8mf4(s + 1 * src_stride, vl); |
| vuint8mf4_t t7 = __riscv_vle8_v_u8mf4(s + 2 * src_stride, vl); |
| vuint8mf4_t t8 = __riscv_vle8_v_u8mf4(s + 3 * src_stride, vl); |
| |
| // Convert to 16-bit |
| vint16mf2_t s5 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t5, vl)); |
| vint16mf2_t s6 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t6, vl)); |
| vint16mf2_t s7 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t7, vl)); |
| vint16mf2_t s8 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t8, vl)); |
| |
| // Perform convolution |
| vuint16mf2_t d0 = convolve6_4_y_rvv(s0, s1, s2, s3, s4, s5, y_filter, |
| round_offset, vl); |
| vuint16mf2_t d1 = convolve6_4_y_rvv(s1, s2, s3, s4, s5, s6, y_filter, |
| round_offset, vl); |
| vuint16mf2_t d2 = convolve6_4_y_rvv(s2, s3, s4, s5, s6, s7, y_filter, |
| round_offset, vl); |
| vuint16mf2_t d3 = convolve6_4_y_rvv(s3, s4, s5, s6, s7, s8, y_filter, |
| round_offset, vl); |
| |
| vuint16mf2_t dd0, dd1, dd2, dd3; |
| load_u16_4x4(d, dst_stride, &dd0, &dd1, &dd2, &dd3, vl); |
| |
| vuint8mf4_t d0_u8, d1_u8, d2_u8, d3_u8; |
| compute_dist_wtd_avg_4x4_rvv(dd0, dd1, dd2, dd3, d0, d1, d2, d3, |
| fwd_offset, bck_offset, round_offset, |
| &d0_u8, &d1_u8, &d2_u8, &d3_u8, vl); |
| |
| // Store result |
| __riscv_vse8_v_u8mf4(d_u8 + 0 * dst8_stride, d0_u8, vl); |
| __riscv_vse8_v_u8mf4(d_u8 + 1 * dst8_stride, d1_u8, vl); |
| __riscv_vse8_v_u8mf4(d_u8 + 2 * dst8_stride, d2_u8, vl); |
| __riscv_vse8_v_u8mf4(d_u8 + 3 * dst8_stride, d3_u8, vl); |
| |
| // Update sliding window |
| s0 = __riscv_vmv_v_v_i16mf2(s4, vl); |
| s1 = __riscv_vmv_v_v_i16mf2(s5, vl); |
| s2 = __riscv_vmv_v_v_i16mf2(s6, vl); |
| s3 = __riscv_vmv_v_v_i16mf2(s7, vl); |
| s4 = __riscv_vmv_v_v_i16mf2(s8, vl); |
| s += 4 * src_stride; |
| d += 4 * dst_stride; |
| d_u8 += 4 * dst8_stride; |
| height -= 4; |
| } while (height != 0); |
| src_ptr += 4; |
| dst_ptr += 4; |
| dst8_ptr += 4; |
| width -= 4; |
| } while (width != 0); |
| } else { |
| size_t vl = __riscv_vsetvl_e16m1(w); |
| do { |
| const uint8_t *s = src_ptr; |
| CONV_BUF_TYPE *d = dst_ptr; |
| uint8_t *d_u8 = dst8_ptr; |
| int height = h; |
| |
| // Load initial 5 rows of data |
| vuint8mf2_t t0, t1, t2, t3, t4; |
| load_u8_8x5(s, src_stride, &t0, &t1, &t2, &t3, &t4, vl); |
| |
| // Convert to 16-bit |
| vint16m1_t s0 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t0, vl)); |
| vint16m1_t s1 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t1, vl)); |
| vint16m1_t s2 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t2, vl)); |
| vint16m1_t s3 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t3, vl)); |
| vint16m1_t s4 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t4, vl)); |
| |
| s += 5 * src_stride; |
| |
| do { |
| vuint8mf2_t t5, t6, t7; |
| load_u8_8x8(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7, vl); |
| |
| // Convert to 16-bit |
| vint16m1_t s5 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t0, vl)); |
| vint16m1_t s6 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t1, vl)); |
| vint16m1_t s7 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t2, vl)); |
| vint16m1_t s8 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t3, vl)); |
| vint16m1_t s9 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t4, vl)); |
| vint16m1_t s10 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t5, vl)); |
| vint16m1_t s11 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t6, vl)); |
| vint16m1_t s12 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t7, vl)); |
| |
| // Perform convolution |
| vuint16m1_t d0 = convolve6_8_y_rvv(s0, s1, s2, s3, s4, s5, y_filter, |
| round_offset, vl); |
| vuint16m1_t d1 = convolve6_8_y_rvv(s1, s2, s3, s4, s5, s6, y_filter, |
| round_offset, vl); |
| vuint16m1_t d2 = convolve6_8_y_rvv(s2, s3, s4, s5, s6, s7, y_filter, |
| round_offset, vl); |
| vuint16m1_t d3 = convolve6_8_y_rvv(s3, s4, s5, s6, s7, s8, y_filter, |
| round_offset, vl); |
| vuint16m1_t d4 = convolve6_8_y_rvv(s4, s5, s6, s7, s8, s9, y_filter, |
| round_offset, vl); |
| vuint16m1_t d5 = convolve6_8_y_rvv(s5, s6, s7, s8, s9, s10, y_filter, |
| round_offset, vl); |
| vuint16m1_t d6 = convolve6_8_y_rvv(s6, s7, s8, s9, s10, s11, y_filter, |
| round_offset, vl); |
| vuint16m1_t d7 = convolve6_8_y_rvv(s7, s8, s9, s10, s11, s12, y_filter, |
| round_offset, vl); |
| |
| vuint16m1_t dd0, dd1, dd2, dd3; |
| load_u16_8x4(d, dst_stride, &dd0, &dd1, &dd2, &dd3, vl); |
| |
| vuint8mf2_t d0_u8, d1_u8, d2_u8, d3_u8; |
| compute_dist_wtd_avg_8x4_rvv(dd0, dd1, dd2, dd3, d0, d1, d2, d3, |
| fwd_offset, bck_offset, round_offset, |
| &d0_u8, &d1_u8, &d2_u8, &d3_u8, vl); |
| store_u8_8x4(d_u8, dst8_stride, d0_u8, d1_u8, d2_u8, d3_u8, vl); |
| d_u8 += 4 * dst8_stride; |
| |
| vuint16m1_t dd4, dd5, dd6, dd7; |
| load_u16_8x4(d + 4 * dst_stride, dst_stride, &dd4, &dd5, &dd6, &dd7, |
| vl); |
| |
| vuint8mf2_t d4_u8, d5_u8, d6_u8, d7_u8; |
| compute_dist_wtd_avg_8x4_rvv(dd4, dd5, dd6, dd7, d4, d5, d6, d7, |
| fwd_offset, bck_offset, round_offset, |
| &d4_u8, &d5_u8, &d6_u8, &d7_u8, vl); |
| store_u8_8x4(d_u8, dst8_stride, d4_u8, d5_u8, d6_u8, d7_u8, vl); |
| d_u8 += 4 * dst8_stride; |
| |
| // Update sliding window |
| s0 = __riscv_vmv_v_v_i16m1(s8, vl); |
| s1 = __riscv_vmv_v_v_i16m1(s9, vl); |
| s2 = __riscv_vmv_v_v_i16m1(s10, vl); |
| s3 = __riscv_vmv_v_v_i16m1(s11, vl); |
| s4 = __riscv_vmv_v_v_i16m1(s12, vl); |
| s += 8 * src_stride; |
| d += 8 * dst_stride; |
| height -= 8; |
| } while (height != 0); |
| src_ptr += vl; |
| dst_ptr += vl; |
| dst8_ptr += vl; |
| width -= vl; |
| } while (width > 0); |
| } |
| } |
| |
| static inline void dist_wtd_convolve_y_6tap_avg_rvv( |
| const uint8_t *src_ptr, int src_stride, uint8_t *dst8_ptr, |
| const int dst8_stride, int w, int h, const int16_t *y_filter, |
| ConvolveParams *conv_params) { |
| const int bd = 8; |
| const int offset_bits = bd + 2 * FILTER_BITS - ROUND0_BITS; |
| const int16_t round_offset = (1 << (offset_bits - COMPOUND_ROUND1_BITS)) + |
| (1 << (offset_bits - COMPOUND_ROUND1_BITS - 1)); |
| CONV_BUF_TYPE *dst_ptr = conv_params->dst; |
| const int dst_stride = conv_params->dst_stride; |
| int width = w; |
| |
| if (w == 4 || h == 4) { |
| size_t vl = __riscv_vsetvl_e16m1(4); |
| do { |
| const uint8_t *s = src_ptr; |
| CONV_BUF_TYPE *d = dst_ptr; |
| uint8_t *d_u8 = dst8_ptr; |
| int height = h; |
| |
| // Load initial 5 rows of data |
| vuint8mf4_t t0, t1, t2, t3, t4; |
| load_u8_4x5(s, src_stride, &t0, &t1, &t2, &t3, &t4, vl); |
| |
| // Convert to 16-bit |
| vint16mf2_t s0 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t0, vl)); |
| vint16mf2_t s1 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t1, vl)); |
| vint16mf2_t s2 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t2, vl)); |
| vint16mf2_t s3 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t3, vl)); |
| vint16mf2_t s4 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t4, vl)); |
| |
| s += 5 * src_stride; |
| |
| do { |
| // Load next rows of data |
| vuint8mf4_t t5 = __riscv_vle8_v_u8mf4(s + 0 * src_stride, vl); |
| vuint8mf4_t t6 = __riscv_vle8_v_u8mf4(s + 1 * src_stride, vl); |
| vuint8mf4_t t7 = __riscv_vle8_v_u8mf4(s + 2 * src_stride, vl); |
| vuint8mf4_t t8 = __riscv_vle8_v_u8mf4(s + 3 * src_stride, vl); |
| |
| // Convert to 16-bit |
| vint16mf2_t s5 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t5, vl)); |
| vint16mf2_t s6 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t6, vl)); |
| vint16mf2_t s7 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t7, vl)); |
| vint16mf2_t s8 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t8, vl)); |
| |
| // Perform convolution |
| vuint16mf2_t d0 = convolve6_4_y_rvv(s0, s1, s2, s3, s4, s5, y_filter, |
| round_offset, vl); |
| vuint16mf2_t d1 = convolve6_4_y_rvv(s1, s2, s3, s4, s5, s6, y_filter, |
| round_offset, vl); |
| vuint16mf2_t d2 = convolve6_4_y_rvv(s2, s3, s4, s5, s6, s7, y_filter, |
| round_offset, vl); |
| vuint16mf2_t d3 = convolve6_4_y_rvv(s3, s4, s5, s6, s7, s8, y_filter, |
| round_offset, vl); |
| |
| vuint16mf2_t dd0, dd1, dd2, dd3; |
| load_u16_4x4(d, dst_stride, &dd0, &dd1, &dd2, &dd3, vl); |
| |
| vuint8mf4_t d0_u8, d1_u8, d2_u8, d3_u8; |
| compute_basic_avg_4x4_rvv(dd0, dd1, dd2, dd3, d0, d1, d2, d3, |
| round_offset, &d0_u8, &d1_u8, &d2_u8, &d3_u8, |
| vl); |
| |
| // Store result |
| __riscv_vse8_v_u8mf4(d_u8 + 0 * dst8_stride, d0_u8, vl); |
| __riscv_vse8_v_u8mf4(d_u8 + 1 * dst8_stride, d1_u8, vl); |
| __riscv_vse8_v_u8mf4(d_u8 + 2 * dst8_stride, d2_u8, vl); |
| __riscv_vse8_v_u8mf4(d_u8 + 3 * dst8_stride, d3_u8, vl); |
| |
| // Update sliding window |
| s0 = __riscv_vmv_v_v_i16mf2(s4, vl); |
| s1 = __riscv_vmv_v_v_i16mf2(s5, vl); |
| s2 = __riscv_vmv_v_v_i16mf2(s6, vl); |
| s3 = __riscv_vmv_v_v_i16mf2(s7, vl); |
| s4 = __riscv_vmv_v_v_i16mf2(s8, vl); |
| s += 4 * src_stride; |
| d += 4 * dst_stride; |
| d_u8 += 4 * dst8_stride; |
| height -= 4; |
| } while (height != 0); |
| src_ptr += 4; |
| dst_ptr += 4; |
| dst8_ptr += 4; |
| width -= 4; |
| } while (width != 0); |
| } else { |
| size_t vl = __riscv_vsetvl_e16m1(w); |
| do { |
| const uint8_t *s = src_ptr; |
| CONV_BUF_TYPE *d = dst_ptr; |
| uint8_t *d_u8 = dst8_ptr; |
| int height = h; |
| |
| // Load initial 5 rows of data |
| vuint8mf2_t t0, t1, t2, t3, t4; |
| load_u8_8x5(s, src_stride, &t0, &t1, &t2, &t3, &t4, vl); |
| |
| // Convert to 16-bit |
| vint16m1_t s0 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t0, vl)); |
| vint16m1_t s1 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t1, vl)); |
| vint16m1_t s2 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t2, vl)); |
| vint16m1_t s3 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t3, vl)); |
| vint16m1_t s4 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t4, vl)); |
| |
| s += 5 * src_stride; |
| |
| do { |
| vuint8mf2_t t5, t6, t7; |
| load_u8_8x8(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7, vl); |
| |
| // Convert to 16-bit |
| vint16m1_t s5 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t0, vl)); |
| vint16m1_t s6 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t1, vl)); |
| vint16m1_t s7 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t2, vl)); |
| vint16m1_t s8 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t3, vl)); |
| vint16m1_t s9 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t4, vl)); |
| vint16m1_t s10 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t5, vl)); |
| vint16m1_t s11 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t6, vl)); |
| vint16m1_t s12 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t7, vl)); |
| |
| // Perform convolution |
| vuint16m1_t d0 = convolve6_8_y_rvv(s0, s1, s2, s3, s4, s5, y_filter, |
| round_offset, vl); |
| vuint16m1_t d1 = convolve6_8_y_rvv(s1, s2, s3, s4, s5, s6, y_filter, |
| round_offset, vl); |
| vuint16m1_t d2 = convolve6_8_y_rvv(s2, s3, s4, s5, s6, s7, y_filter, |
| round_offset, vl); |
| vuint16m1_t d3 = convolve6_8_y_rvv(s3, s4, s5, s6, s7, s8, y_filter, |
| round_offset, vl); |
| vuint16m1_t d4 = convolve6_8_y_rvv(s4, s5, s6, s7, s8, s9, y_filter, |
| round_offset, vl); |
| vuint16m1_t d5 = convolve6_8_y_rvv(s5, s6, s7, s8, s9, s10, y_filter, |
| round_offset, vl); |
| vuint16m1_t d6 = convolve6_8_y_rvv(s6, s7, s8, s9, s10, s11, y_filter, |
| round_offset, vl); |
| vuint16m1_t d7 = convolve6_8_y_rvv(s7, s8, s9, s10, s11, s12, y_filter, |
| round_offset, vl); |
| |
| vuint16m1_t dd0, dd1, dd2, dd3; |
| load_u16_8x4(d, dst_stride, &dd0, &dd1, &dd2, &dd3, vl); |
| |
| vuint8mf2_t d0_u8, d1_u8, d2_u8, d3_u8; |
| compute_basic_avg_8x4_rvv(dd0, dd1, dd2, dd3, d0, d1, d2, d3, |
| round_offset, &d0_u8, &d1_u8, &d2_u8, &d3_u8, |
| vl); |
| store_u8_8x4(d_u8, dst8_stride, d0_u8, d1_u8, d2_u8, d3_u8, vl); |
| d_u8 += 4 * dst8_stride; |
| |
| vuint16m1_t dd4, dd5, dd6, dd7; |
| load_u16_8x4(d + 4 * dst_stride, dst_stride, &dd4, &dd5, &dd6, &dd7, |
| vl); |
| |
| vuint8mf2_t d4_u8, d5_u8, d6_u8, d7_u8; |
| compute_basic_avg_8x4_rvv(dd4, dd5, dd6, dd7, d4, d5, d6, d7, |
| round_offset, &d4_u8, &d5_u8, &d6_u8, &d7_u8, |
| vl); |
| store_u8_8x4(d_u8, dst8_stride, d4_u8, d5_u8, d6_u8, d7_u8, vl); |
| d_u8 += 4 * dst8_stride; |
| |
| // Update sliding window |
| s0 = __riscv_vmv_v_v_i16m1(s8, vl); |
| s1 = __riscv_vmv_v_v_i16m1(s9, vl); |
| s2 = __riscv_vmv_v_v_i16m1(s10, vl); |
| s3 = __riscv_vmv_v_v_i16m1(s11, vl); |
| s4 = __riscv_vmv_v_v_i16m1(s12, vl); |
| s += 8 * src_stride; |
| d += 8 * dst_stride; |
| height -= 8; |
| } while (height != 0); |
| src_ptr += vl; |
| dst_ptr += vl; |
| dst8_ptr += vl; |
| width -= vl; |
| } while (width > 0); |
| } |
| } |
| |
| static inline void dist_wtd_convolve_y_6tap_rvv(const uint8_t *src_ptr, |
| int src_stride, int w, int h, |
| const int16_t *y_filter, |
| ConvolveParams *conv_params) { |
| const int bd = 8; |
| const int offset_bits = bd + 2 * FILTER_BITS - ROUND0_BITS; |
| const int16_t round_offset = (1 << (offset_bits - COMPOUND_ROUND1_BITS)) + |
| (1 << (offset_bits - COMPOUND_ROUND1_BITS - 1)); |
| CONV_BUF_TYPE *dst_ptr = conv_params->dst; |
| const int dst_stride = conv_params->dst_stride; |
| int width = w; |
| |
| if (w == 4 || h == 4) { |
| size_t vl = __riscv_vsetvl_e16m1(4); |
| do { |
| const uint8_t *s = src_ptr; |
| CONV_BUF_TYPE *d = dst_ptr; |
| int height = h; |
| |
| // Load initial 5 rows of data |
| vuint8mf4_t t0, t1, t2, t3, t4; |
| load_u8_4x5(s, src_stride, &t0, &t1, &t2, &t3, &t4, vl); |
| |
| // Convert to 16-bit |
| vint16mf2_t s0 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t0, vl)); |
| vint16mf2_t s1 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t1, vl)); |
| vint16mf2_t s2 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t2, vl)); |
| vint16mf2_t s3 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t3, vl)); |
| vint16mf2_t s4 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t4, vl)); |
| |
| s += 5 * src_stride; |
| |
| do { |
| // Load next rows of data |
| vuint8mf4_t t5 = __riscv_vle8_v_u8mf4(s + 0 * src_stride, vl); |
| vuint8mf4_t t6 = __riscv_vle8_v_u8mf4(s + 1 * src_stride, vl); |
| vuint8mf4_t t7 = __riscv_vle8_v_u8mf4(s + 2 * src_stride, vl); |
| vuint8mf4_t t8 = __riscv_vle8_v_u8mf4(s + 3 * src_stride, vl); |
| |
| // Convert to 16-bit |
| vint16mf2_t s5 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t5, vl)); |
| vint16mf2_t s6 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t6, vl)); |
| vint16mf2_t s7 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t7, vl)); |
| vint16mf2_t s8 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t8, vl)); |
| |
| // Perform convolution |
| vuint16mf2_t d0 = convolve6_4_y_rvv(s0, s1, s2, s3, s4, s5, y_filter, |
| round_offset, vl); |
| vuint16mf2_t d1 = convolve6_4_y_rvv(s1, s2, s3, s4, s5, s6, y_filter, |
| round_offset, vl); |
| vuint16mf2_t d2 = convolve6_4_y_rvv(s2, s3, s4, s5, s6, s7, y_filter, |
| round_offset, vl); |
| vuint16mf2_t d3 = convolve6_4_y_rvv(s3, s4, s5, s6, s7, s8, y_filter, |
| round_offset, vl); |
| |
| // Store result |
| store_u16_4x4(d, dst_stride, d0, d1, d2, d3, vl); |
| |
| // Update sliding window |
| s0 = __riscv_vmv_v_v_i16mf2(s4, vl); |
| s1 = __riscv_vmv_v_v_i16mf2(s5, vl); |
| s2 = __riscv_vmv_v_v_i16mf2(s6, vl); |
| s3 = __riscv_vmv_v_v_i16mf2(s7, vl); |
| s4 = __riscv_vmv_v_v_i16mf2(s8, vl); |
| s += 4 * src_stride; |
| d += 4 * dst_stride; |
| height -= 4; |
| } while (height != 0); |
| src_ptr += 4; |
| dst_ptr += 4; |
| width -= 4; |
| } while (width != 0); |
| } else { |
| size_t vl = __riscv_vsetvl_e16m1(w); |
| do { |
| const uint8_t *s = src_ptr; |
| CONV_BUF_TYPE *d = dst_ptr; |
| int height = h; |
| |
| // Load initial 5 rows of data |
| vuint8mf2_t t0, t1, t2, t3, t4; |
| load_u8_8x5(s, src_stride, &t0, &t1, &t2, &t3, &t4, vl); |
| |
| // Convert to 16-bit |
| vint16m1_t s0 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t0, vl)); |
| vint16m1_t s1 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t1, vl)); |
| vint16m1_t s2 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t2, vl)); |
| vint16m1_t s3 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t3, vl)); |
| vint16m1_t s4 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t4, vl)); |
| |
| s += 5 * src_stride; |
| |
| do { |
| vuint8mf2_t t5, t6, t7; |
| load_u8_8x8(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7, vl); |
| |
| // Convert to 16-bit |
| vint16m1_t s5 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t0, vl)); |
| vint16m1_t s6 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t1, vl)); |
| vint16m1_t s7 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t2, vl)); |
| vint16m1_t s8 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t3, vl)); |
| vint16m1_t s9 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t4, vl)); |
| vint16m1_t s10 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t5, vl)); |
| vint16m1_t s11 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t6, vl)); |
| vint16m1_t s12 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t7, vl)); |
| |
| // Perform convolution |
| vuint16m1_t d0 = convolve6_8_y_rvv(s0, s1, s2, s3, s4, s5, y_filter, |
| round_offset, vl); |
| vuint16m1_t d1 = convolve6_8_y_rvv(s1, s2, s3, s4, s5, s6, y_filter, |
| round_offset, vl); |
| vuint16m1_t d2 = convolve6_8_y_rvv(s2, s3, s4, s5, s6, s7, y_filter, |
| round_offset, vl); |
| vuint16m1_t d3 = convolve6_8_y_rvv(s3, s4, s5, s6, s7, s8, y_filter, |
| round_offset, vl); |
| vuint16m1_t d4 = convolve6_8_y_rvv(s4, s5, s6, s7, s8, s9, y_filter, |
| round_offset, vl); |
| vuint16m1_t d5 = convolve6_8_y_rvv(s5, s6, s7, s8, s9, s10, y_filter, |
| round_offset, vl); |
| vuint16m1_t d6 = convolve6_8_y_rvv(s6, s7, s8, s9, s10, s11, y_filter, |
| round_offset, vl); |
| vuint16m1_t d7 = convolve6_8_y_rvv(s7, s8, s9, s10, s11, s12, y_filter, |
| round_offset, vl); |
| |
| // Store result |
| store_u16_8x8(d, dst_stride, d0, d1, d2, d3, d4, d5, d6, d7, vl); |
| |
| // Update sliding window |
| s0 = __riscv_vmv_v_v_i16m1(s8, vl); |
| s1 = __riscv_vmv_v_v_i16m1(s9, vl); |
| s2 = __riscv_vmv_v_v_i16m1(s10, vl); |
| s3 = __riscv_vmv_v_v_i16m1(s11, vl); |
| s4 = __riscv_vmv_v_v_i16m1(s12, vl); |
| s += 8 * src_stride; |
| d += 8 * dst_stride; |
| height -= 8; |
| } while (height != 0); |
| src_ptr += vl; |
| dst_ptr += vl; |
| width -= vl; |
| } while (width > 0); |
| } |
| } |
| |
| static inline vuint16mf2_t convolve8_4_y_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 int16_t round_offset, size_t vl) { |
| // Filter values at indices 0 and 7 are 0. |
| vint16mf2_t sum = __riscv_vmul_vx_i16mf2(s0, filter[0], vl); |
| sum = __riscv_vmacc_vx_i16mf2(sum, filter[1], s1, vl); |
| sum = __riscv_vmacc_vx_i16mf2(sum, filter[2], s2, vl); |
| sum = __riscv_vmacc_vx_i16mf2(sum, filter[3], s3, vl); |
| sum = __riscv_vmacc_vx_i16mf2(sum, filter[4], s4, vl); |
| sum = __riscv_vmacc_vx_i16mf2(sum, filter[5], s5, vl); |
| sum = __riscv_vmacc_vx_i16mf2(sum, filter[6], s6, vl); |
| sum = __riscv_vmacc_vx_i16mf2(sum, filter[7], s7, vl); |
| |
| // Right shift with rounding: vrsra_n_s16(round_offset, sum, ROUND0_BITS - 1) |
| // vrsra_n_s16(a, b, n) = a + ((b + (1 << (n-1))) >> n) |
| // We halved the convolution filter values so -1 from the right shift. |
| vint16mf2_t res = |
| __riscv_vadd_vx_i16mf2(sum, 1 << ((ROUND0_BITS - 1) - 1), vl); |
| res = __riscv_vsra_vx_i16mf2(res, ROUND0_BITS - 1, vl); |
| res = __riscv_vadd_vx_i16mf2(res, round_offset, vl); |
| |
| // Reinterpret as uint16 |
| return __riscv_vreinterpret_v_i16mf2_u16mf2(res); |
| } |
| |
| static inline vuint16m1_t convolve8_8_y_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 int16_t round_offset, size_t vl) { |
| // Filter values at indices 0 and 7 are 0. |
| vint16m1_t sum = __riscv_vmul_vx_i16m1(s0, filter[0], vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[1], s1, vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[2], s2, vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[3], s3, vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[4], s4, vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[5], s5, vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[6], s6, vl); |
| sum = __riscv_vmacc_vx_i16m1(sum, filter[7], s7, vl); |
| |
| // Right shift with rounding: vrsra_n_s16(round_offset, sum, ROUND0_BITS - 1) |
| // vrsra_n_s16(a, b, n) = a + ((b + (1 << (n-1))) >> n) |
| // We halved the convolution filter values so -1 from the right shift. |
| vint16m1_t res = __riscv_vadd_vx_i16m1(sum, 1 << ((ROUND0_BITS - 1) - 1), vl); |
| res = __riscv_vsra_vx_i16m1(res, ROUND0_BITS - 1, vl); |
| res = __riscv_vadd_vx_i16m1(res, round_offset, vl); |
| |
| // Reinterpret as uint16 |
| return __riscv_vreinterpret_v_i16m1_u16m1(res); |
| } |
| |
| static inline void dist_wtd_convolve_y_8tap_dist_wtd_avg_rvv( |
| const uint8_t *src_ptr, int src_stride, uint8_t *dst8_ptr, |
| const int dst8_stride, int w, int h, const int16_t *y_filter, |
| ConvolveParams *conv_params) { |
| const int bd = 8; |
| const int offset_bits = bd + 2 * FILTER_BITS - ROUND0_BITS; |
| const int16_t round_offset = (1 << (offset_bits - COMPOUND_ROUND1_BITS)) + |
| (1 << (offset_bits - COMPOUND_ROUND1_BITS - 1)); |
| const uint16_t fwd_offset = conv_params->fwd_offset; |
| const uint16_t bck_offset = conv_params->bck_offset; |
| CONV_BUF_TYPE *dst_ptr = conv_params->dst; |
| const int dst_stride = conv_params->dst_stride; |
| int width = w; |
| |
| if (w == 4 || h == 4) { |
| size_t vl = __riscv_vsetvl_e16m1(4); |
| do { |
| const uint8_t *s = src_ptr; |
| CONV_BUF_TYPE *d = dst_ptr; |
| uint8_t *d_u8 = dst8_ptr; |
| int height = h; |
| |
| __builtin_prefetch(s + 0 * src_stride); |
| __builtin_prefetch(s + 1 * src_stride); |
| __builtin_prefetch(s + 2 * src_stride); |
| __builtin_prefetch(s + 3 * src_stride); |
| |
| // Load initial 7 rows of data |
| vuint8mf4_t t0, t1, t2, t3, t4, t5, t6; |
| load_u8_4x7(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, vl); |
| |
| // Convert to 16-bit |
| vint16mf2_t s0 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t0, vl)); |
| vint16mf2_t s1 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t1, vl)); |
| vint16mf2_t s2 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t2, vl)); |
| vint16mf2_t s3 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t3, vl)); |
| vint16mf2_t s4 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t4, vl)); |
| vint16mf2_t s5 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t5, vl)); |
| vint16mf2_t s6 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t6, vl)); |
| |
| s += 7 * src_stride; |
| |
| do { |
| // Load next rows of data |
| t0 = __riscv_vle8_v_u8mf4(s + 0 * src_stride, vl); |
| t1 = __riscv_vle8_v_u8mf4(s + 1 * src_stride, vl); |
| t2 = __riscv_vle8_v_u8mf4(s + 2 * src_stride, vl); |
| t3 = __riscv_vle8_v_u8mf4(s + 3 * src_stride, vl); |
| |
| // Convert to 16-bit |
| vint16mf2_t s7 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t0, vl)); |
| vint16mf2_t s8 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t1, vl)); |
| vint16mf2_t s9 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t2, vl)); |
| vint16mf2_t s10 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t3, vl)); |
| |
| // Perform convolution |
| vuint16mf2_t d0 = convolve8_4_y_rvv(s0, s1, s2, s3, s4, s5, s6, s7, |
| y_filter, round_offset, vl); |
| vuint16mf2_t d1 = convolve8_4_y_rvv(s1, s2, s3, s4, s5, s6, s7, s8, |
| y_filter, round_offset, vl); |
| vuint16mf2_t d2 = convolve8_4_y_rvv(s2, s3, s4, s5, s6, s7, s8, s9, |
| y_filter, round_offset, vl); |
| vuint16mf2_t d3 = convolve8_4_y_rvv(s3, s4, s5, s6, s7, s8, s9, s10, |
| y_filter, round_offset, vl); |
| |
| __builtin_prefetch(d + 0 * dst_stride); |
| __builtin_prefetch(d + 1 * dst_stride); |
| __builtin_prefetch(d + 2 * dst_stride); |
| __builtin_prefetch(d + 3 * dst_stride); |
| |
| __builtin_prefetch(d_u8 + 0 * dst8_stride); |
| __builtin_prefetch(d_u8 + 1 * dst8_stride); |
| __builtin_prefetch(d_u8 + 2 * dst8_stride); |
| __builtin_prefetch(d_u8 + 3 * dst8_stride); |
| |
| vuint16mf2_t dd0, dd1, dd2, dd3; |
| load_u16_4x4(d, dst_stride, &dd0, &dd1, &dd2, &dd3, vl); |
| |
| vuint8mf4_t d0_u8, d1_u8, d2_u8, d3_u8; |
| compute_dist_wtd_avg_4x4_rvv(dd0, dd1, dd2, dd3, d0, d1, d2, d3, |
| fwd_offset, bck_offset, round_offset, |
| &d0_u8, &d1_u8, &d2_u8, &d3_u8, vl); |
| |
| // Store result |
| __riscv_vse8_v_u8mf4(d_u8 + 0 * dst8_stride, d0_u8, vl); |
| __riscv_vse8_v_u8mf4(d_u8 + 1 * dst8_stride, d1_u8, vl); |
| __riscv_vse8_v_u8mf4(d_u8 + 2 * dst8_stride, d2_u8, vl); |
| __riscv_vse8_v_u8mf4(d_u8 + 3 * dst8_stride, d3_u8, vl); |
| |
| // Update sliding window |
| s0 = __riscv_vmv_v_v_i16mf2(s4, vl); |
| s1 = __riscv_vmv_v_v_i16mf2(s5, vl); |
| s2 = __riscv_vmv_v_v_i16mf2(s6, vl); |
| s3 = __riscv_vmv_v_v_i16mf2(s7, vl); |
| s4 = __riscv_vmv_v_v_i16mf2(s8, vl); |
| s5 = __riscv_vmv_v_v_i16mf2(s9, vl); |
| s6 = __riscv_vmv_v_v_i16mf2(s10, vl); |
| s += 4 * src_stride; |
| d += 4 * dst_stride; |
| d_u8 += 4 * dst8_stride; |
| height -= 4; |
| } while (height != 0); |
| src_ptr += 4; |
| dst_ptr += 4; |
| dst8_ptr += 4; |
| width -= 4; |
| } while (width != 0); |
| } else { |
| size_t vl = __riscv_vsetvl_e16m1(w); |
| do { |
| const uint8_t *s = src_ptr; |
| CONV_BUF_TYPE *d = dst_ptr; |
| uint8_t *d_u8 = dst8_ptr; |
| int height = h; |
| |
| __builtin_prefetch(s + 0 * src_stride); |
| __builtin_prefetch(s + 1 * src_stride); |
| __builtin_prefetch(s + 2 * src_stride); |
| __builtin_prefetch(s + 3 * src_stride); |
| __builtin_prefetch(s + 4 * src_stride); |
| __builtin_prefetch(s + 5 * src_stride); |
| __builtin_prefetch(s + 6 * src_stride); |
| |
| // Load initial 7 rows of data |
| vuint8mf2_t t0, t1, t2, t3, t4, t5, t6; |
| load_u8_8x7(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, vl); |
| |
| // Convert to 16-bit |
| vint16m1_t s0 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t0, vl)); |
| vint16m1_t s1 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t1, vl)); |
| vint16m1_t s2 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t2, vl)); |
| vint16m1_t s3 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t3, vl)); |
| vint16m1_t s4 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t4, vl)); |
| vint16m1_t s5 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t5, vl)); |
| vint16m1_t s6 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t6, vl)); |
| |
| s += 7 * src_stride; |
| |
| do { |
| vuint8mf2_t t7; |
| load_u8_8x8(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7, vl); |
| |
| // Convert to 16-bit |
| vint16m1_t s7 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t0, vl)); |
| vint16m1_t s8 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t1, vl)); |
| vint16m1_t s9 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t2, vl)); |
| vint16m1_t s10 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t3, vl)); |
| vint16m1_t s11 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t4, vl)); |
| vint16m1_t s12 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t5, vl)); |
| vint16m1_t s13 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t6, vl)); |
| vint16m1_t s14 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t7, vl)); |
| |
| __builtin_prefetch(d + 0 * dst_stride); |
| __builtin_prefetch(d + 1 * dst_stride); |
| __builtin_prefetch(d + 2 * dst_stride); |
| __builtin_prefetch(d + 3 * dst_stride); |
| |
| // Perform convolution |
| vuint16m1_t d0 = convolve8_8_y_rvv(s0, s1, s2, s3, s4, s5, s6, s7, |
| y_filter, round_offset, vl); |
| vuint16m1_t d1 = convolve8_8_y_rvv(s1, s2, s3, s4, s5, s6, s7, s8, |
| y_filter, round_offset, vl); |
| vuint16m1_t d2 = convolve8_8_y_rvv(s2, s3, s4, s5, s6, s7, s8, s9, |
| y_filter, round_offset, vl); |
| vuint16m1_t d3 = convolve8_8_y_rvv(s3, s4, s5, s6, s7, s8, s9, s10, |
| y_filter, round_offset, vl); |
| vuint16m1_t d4 = convolve8_8_y_rvv(s4, s5, s6, s7, s8, s9, s10, s11, |
| y_filter, round_offset, vl); |
| vuint16m1_t d5 = convolve8_8_y_rvv(s5, s6, s7, s8, s9, s10, s11, s12, |
| y_filter, round_offset, vl); |
| vuint16m1_t d6 = convolve8_8_y_rvv(s6, s7, s8, s9, s10, s11, s12, s13, |
| y_filter, round_offset, vl); |
| vuint16m1_t d7 = convolve8_8_y_rvv(s7, s8, s9, s10, s11, s12, s13, s14, |
| y_filter, round_offset, vl); |
| |
| __builtin_prefetch(d_u8 + 0 * dst8_stride); |
| __builtin_prefetch(d_u8 + 1 * dst8_stride); |
| __builtin_prefetch(d_u8 + 2 * dst8_stride); |
| __builtin_prefetch(d_u8 + 3 * dst8_stride); |
| |
| vuint16m1_t dd0, dd1, dd2, dd3; |
| load_u16_8x4(d, dst_stride, &dd0, &dd1, &dd2, &dd3, vl); |
| |
| vuint8mf2_t d0_u8, d1_u8, d2_u8, d3_u8; |
| compute_dist_wtd_avg_8x4_rvv(dd0, dd1, dd2, dd3, d0, d1, d2, d3, |
| fwd_offset, bck_offset, round_offset, |
| &d0_u8, &d1_u8, &d2_u8, &d3_u8, vl); |
| store_u8_8x4(d_u8, dst8_stride, d0_u8, d1_u8, d2_u8, d3_u8, vl); |
| d_u8 += 4 * dst8_stride; |
| |
| vuint16m1_t dd4, dd5, dd6, dd7; |
| load_u16_8x4(d + 4 * dst_stride, dst_stride, &dd4, &dd5, &dd6, &dd7, |
| vl); |
| |
| vuint8mf2_t d4_u8, d5_u8, d6_u8, d7_u8; |
| compute_dist_wtd_avg_8x4_rvv(dd4, dd5, dd6, dd7, d4, d5, d6, d7, |
| fwd_offset, bck_offset, round_offset, |
| &d4_u8, &d5_u8, &d6_u8, &d7_u8, vl); |
| store_u8_8x4(d_u8, dst8_stride, d4_u8, d5_u8, d6_u8, d7_u8, vl); |
| d_u8 += 4 * dst8_stride; |
| |
| // Update sliding window |
| s0 = __riscv_vmv_v_v_i16m1(s8, vl); |
| s1 = __riscv_vmv_v_v_i16m1(s9, vl); |
| s2 = __riscv_vmv_v_v_i16m1(s10, vl); |
| s3 = __riscv_vmv_v_v_i16m1(s11, vl); |
| s4 = __riscv_vmv_v_v_i16m1(s12, vl); |
| s5 = __riscv_vmv_v_v_i16m1(s13, vl); |
| s6 = __riscv_vmv_v_v_i16m1(s14, vl); |
| s += 8 * src_stride; |
| d += 8 * dst_stride; |
| height -= 8; |
| } while (height != 0); |
| src_ptr += vl; |
| dst_ptr += vl; |
| dst8_ptr += vl; |
| width -= vl; |
| } while (width > 0); |
| } |
| } |
| |
| static inline void dist_wtd_convolve_y_8tap_avg_rvv( |
| const uint8_t *src_ptr, int src_stride, uint8_t *dst8_ptr, |
| const int dst8_stride, int w, int h, const int16_t *y_filter, |
| ConvolveParams *conv_params) { |
| const int bd = 8; |
| const int offset_bits = bd + 2 * FILTER_BITS - ROUND0_BITS; |
| const int16_t round_offset = (1 << (offset_bits - COMPOUND_ROUND1_BITS)) + |
| (1 << (offset_bits - COMPOUND_ROUND1_BITS - 1)); |
| CONV_BUF_TYPE *dst_ptr = conv_params->dst; |
| const int dst_stride = conv_params->dst_stride; |
| int width = w; |
| |
| if (w == 4 || h == 4) { |
| size_t vl = __riscv_vsetvl_e16m1(4); |
| do { |
| const uint8_t *s = src_ptr; |
| CONV_BUF_TYPE *d = dst_ptr; |
| uint8_t *d_u8 = dst8_ptr; |
| int height = h; |
| |
| __builtin_prefetch(s + 0 * src_stride); |
| __builtin_prefetch(s + 1 * src_stride); |
| __builtin_prefetch(s + 2 * src_stride); |
| __builtin_prefetch(s + 3 * src_stride); |
| |
| // Load initial 7 rows of data |
| vuint8mf4_t t0, t1, t2, t3, t4, t5, t6; |
| load_u8_4x7(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, vl); |
| |
| // Convert to 16-bit |
| vint16mf2_t s0 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t0, vl)); |
| vint16mf2_t s1 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t1, vl)); |
| vint16mf2_t s2 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t2, vl)); |
| vint16mf2_t s3 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t3, vl)); |
| vint16mf2_t s4 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t4, vl)); |
| vint16mf2_t s5 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t5, vl)); |
| vint16mf2_t s6 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t6, vl)); |
| |
| s += 7 * src_stride; |
| |
| do { |
| // Load next rows of data |
| t0 = __riscv_vle8_v_u8mf4(s + 0 * src_stride, vl); |
| t1 = __riscv_vle8_v_u8mf4(s + 1 * src_stride, vl); |
| t2 = __riscv_vle8_v_u8mf4(s + 2 * src_stride, vl); |
| t3 = __riscv_vle8_v_u8mf4(s + 3 * src_stride, vl); |
| |
| // Convert to 16-bit |
| vint16mf2_t s7 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t0, vl)); |
| vint16mf2_t s8 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t1, vl)); |
| vint16mf2_t s9 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t2, vl)); |
| vint16mf2_t s10 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t3, vl)); |
| |
| // Perform convolution |
| vuint16mf2_t d0 = convolve8_4_y_rvv(s0, s1, s2, s3, s4, s5, s6, s7, |
| y_filter, round_offset, vl); |
| vuint16mf2_t d1 = convolve8_4_y_rvv(s1, s2, s3, s4, s5, s6, s7, s8, |
| y_filter, round_offset, vl); |
| vuint16mf2_t d2 = convolve8_4_y_rvv(s2, s3, s4, s5, s6, s7, s8, s9, |
| y_filter, round_offset, vl); |
| vuint16mf2_t d3 = convolve8_4_y_rvv(s3, s4, s5, s6, s7, s8, s9, s10, |
| y_filter, round_offset, vl); |
| |
| __builtin_prefetch(d + 0 * dst_stride); |
| __builtin_prefetch(d + 1 * dst_stride); |
| __builtin_prefetch(d + 2 * dst_stride); |
| __builtin_prefetch(d + 3 * dst_stride); |
| |
| __builtin_prefetch(d_u8 + 0 * dst8_stride); |
| __builtin_prefetch(d_u8 + 1 * dst8_stride); |
| __builtin_prefetch(d_u8 + 2 * dst8_stride); |
| __builtin_prefetch(d_u8 + 3 * dst8_stride); |
| |
| vuint16mf2_t dd0, dd1, dd2, dd3; |
| load_u16_4x4(d, dst_stride, &dd0, &dd1, &dd2, &dd3, vl); |
| |
| vuint8mf4_t d0_u8, d1_u8, d2_u8, d3_u8; |
| compute_basic_avg_4x4_rvv(dd0, dd1, dd2, dd3, d0, d1, d2, d3, |
| round_offset, &d0_u8, &d1_u8, &d2_u8, &d3_u8, |
| vl); |
| |
| // Store result |
| __riscv_vse8_v_u8mf4(d_u8 + 0 * dst8_stride, d0_u8, vl); |
| __riscv_vse8_v_u8mf4(d_u8 + 1 * dst8_stride, d1_u8, vl); |
| __riscv_vse8_v_u8mf4(d_u8 + 2 * dst8_stride, d2_u8, vl); |
| __riscv_vse8_v_u8mf4(d_u8 + 3 * dst8_stride, d3_u8, vl); |
| |
| // Update sliding window |
| s0 = __riscv_vmv_v_v_i16mf2(s4, vl); |
| s1 = __riscv_vmv_v_v_i16mf2(s5, vl); |
| s2 = __riscv_vmv_v_v_i16mf2(s6, vl); |
| s3 = __riscv_vmv_v_v_i16mf2(s7, vl); |
| s4 = __riscv_vmv_v_v_i16mf2(s8, vl); |
| s5 = __riscv_vmv_v_v_i16mf2(s9, vl); |
| s6 = __riscv_vmv_v_v_i16mf2(s10, vl); |
| s += 4 * src_stride; |
| d += 4 * dst_stride; |
| d_u8 += 4 * dst8_stride; |
| height -= 4; |
| } while (height != 0); |
| src_ptr += 4; |
| dst_ptr += 4; |
| dst8_ptr += 4; |
| width -= 4; |
| } while (width != 0); |
| } else { |
| size_t vl = __riscv_vsetvl_e16m1(w); |
| do { |
| const uint8_t *s = src_ptr; |
| CONV_BUF_TYPE *d = dst_ptr; |
| uint8_t *d_u8 = dst8_ptr; |
| int height = h; |
| |
| __builtin_prefetch(s + 0 * src_stride); |
| __builtin_prefetch(s + 1 * src_stride); |
| __builtin_prefetch(s + 2 * src_stride); |
| __builtin_prefetch(s + 3 * src_stride); |
| __builtin_prefetch(s + 4 * src_stride); |
| __builtin_prefetch(s + 5 * src_stride); |
| __builtin_prefetch(s + 6 * src_stride); |
| |
| // Load initial 7 rows of data |
| vuint8mf2_t t0, t1, t2, t3, t4, t5, t6; |
| load_u8_8x7(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, vl); |
| |
| // Convert to 16-bit |
| vint16m1_t s0 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t0, vl)); |
| vint16m1_t s1 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t1, vl)); |
| vint16m1_t s2 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t2, vl)); |
| vint16m1_t s3 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t3, vl)); |
| vint16m1_t s4 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t4, vl)); |
| vint16m1_t s5 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t5, vl)); |
| vint16m1_t s6 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t6, vl)); |
| |
| s += 7 * src_stride; |
| |
| do { |
| vuint8mf2_t t7; |
| load_u8_8x8(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7, vl); |
| |
| // Convert to 16-bit |
| vint16m1_t s7 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t0, vl)); |
| vint16m1_t s8 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t1, vl)); |
| vint16m1_t s9 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t2, vl)); |
| vint16m1_t s10 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t3, vl)); |
| vint16m1_t s11 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t4, vl)); |
| vint16m1_t s12 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t5, vl)); |
| vint16m1_t s13 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t6, vl)); |
| vint16m1_t s14 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t7, vl)); |
| |
| __builtin_prefetch(d + 0 * dst_stride); |
| __builtin_prefetch(d + 1 * dst_stride); |
| __builtin_prefetch(d + 2 * dst_stride); |
| __builtin_prefetch(d + 3 * dst_stride); |
| |
| // Perform convolution |
| vuint16m1_t d0 = convolve8_8_y_rvv(s0, s1, s2, s3, s4, s5, s6, s7, |
| y_filter, round_offset, vl); |
| vuint16m1_t d1 = convolve8_8_y_rvv(s1, s2, s3, s4, s5, s6, s7, s8, |
| y_filter, round_offset, vl); |
| vuint16m1_t d2 = convolve8_8_y_rvv(s2, s3, s4, s5, s6, s7, s8, s9, |
| y_filter, round_offset, vl); |
| vuint16m1_t d3 = convolve8_8_y_rvv(s3, s4, s5, s6, s7, s8, s9, s10, |
| y_filter, round_offset, vl); |
| vuint16m1_t d4 = convolve8_8_y_rvv(s4, s5, s6, s7, s8, s9, s10, s11, |
| y_filter, round_offset, vl); |
| vuint16m1_t d5 = convolve8_8_y_rvv(s5, s6, s7, s8, s9, s10, s11, s12, |
| y_filter, round_offset, vl); |
| vuint16m1_t d6 = convolve8_8_y_rvv(s6, s7, s8, s9, s10, s11, s12, s13, |
| y_filter, round_offset, vl); |
| vuint16m1_t d7 = convolve8_8_y_rvv(s7, s8, s9, s10, s11, s12, s13, s14, |
| y_filter, round_offset, vl); |
| |
| __builtin_prefetch(d_u8 + 0 * dst8_stride); |
| __builtin_prefetch(d_u8 + 1 * dst8_stride); |
| __builtin_prefetch(d_u8 + 2 * dst8_stride); |
| __builtin_prefetch(d_u8 + 3 * dst8_stride); |
| |
| vuint16m1_t dd0, dd1, dd2, dd3; |
| load_u16_8x4(d, dst_stride, &dd0, &dd1, &dd2, &dd3, vl); |
| |
| vuint8mf2_t d0_u8, d1_u8, d2_u8, d3_u8; |
| compute_basic_avg_8x4_rvv(dd0, dd1, dd2, dd3, d0, d1, d2, d3, |
| round_offset, &d0_u8, &d1_u8, &d2_u8, &d3_u8, |
| vl); |
| store_u8_8x4(d_u8, dst8_stride, d0_u8, d1_u8, d2_u8, d3_u8, vl); |
| d_u8 += 4 * dst8_stride; |
| |
| vuint16m1_t dd4, dd5, dd6, dd7; |
| load_u16_8x4(d + 4 * dst_stride, dst_stride, &dd4, &dd5, &dd6, &dd7, |
| vl); |
| |
| vuint8mf2_t d4_u8, d5_u8, d6_u8, d7_u8; |
| compute_basic_avg_8x4_rvv(dd4, dd5, dd6, dd7, d4, d5, d6, d7, |
| round_offset, &d4_u8, &d5_u8, &d6_u8, &d7_u8, |
| vl); |
| store_u8_8x4(d_u8, dst8_stride, d4_u8, d5_u8, d6_u8, d7_u8, vl); |
| d_u8 += 4 * dst8_stride; |
| |
| // Update sliding window |
| s0 = __riscv_vmv_v_v_i16m1(s8, vl); |
| s1 = __riscv_vmv_v_v_i16m1(s9, vl); |
| s2 = __riscv_vmv_v_v_i16m1(s10, vl); |
| s3 = __riscv_vmv_v_v_i16m1(s11, vl); |
| s4 = __riscv_vmv_v_v_i16m1(s12, vl); |
| s5 = __riscv_vmv_v_v_i16m1(s13, vl); |
| s6 = __riscv_vmv_v_v_i16m1(s14, vl); |
| s += 8 * src_stride; |
| d += 8 * dst_stride; |
| height -= 8; |
| } while (height != 0); |
| src_ptr += vl; |
| dst_ptr += vl; |
| dst8_ptr += vl; |
| width -= vl; |
| } while (width > 0); |
| } |
| } |
| |
| static inline void dist_wtd_convolve_y_8tap_rvv(const uint8_t *src_ptr, |
| int src_stride, int w, int h, |
| const int16_t *y_filter, |
| ConvolveParams *conv_params) { |
| const int bd = 8; |
| const int offset_bits = bd + 2 * FILTER_BITS - ROUND0_BITS; |
| const int16_t round_offset = (1 << (offset_bits - COMPOUND_ROUND1_BITS)) + |
| (1 << (offset_bits - COMPOUND_ROUND1_BITS - 1)); |
| CONV_BUF_TYPE *dst_ptr = conv_params->dst; |
| const int dst_stride = conv_params->dst_stride; |
| int width = w; |
| |
| if (w == 4 || h == 4) { |
| size_t vl = __riscv_vsetvl_e16m1(4); |
| do { |
| const uint8_t *s = src_ptr; |
| CONV_BUF_TYPE *d = dst_ptr; |
| int height = h; |
| |
| __builtin_prefetch(s + 0 * src_stride); |
| __builtin_prefetch(s + 1 * src_stride); |
| __builtin_prefetch(s + 2 * src_stride); |
| __builtin_prefetch(s + 3 * src_stride); |
| |
| // Load initial 7 rows of data |
| vuint8mf4_t t0, t1, t2, t3, t4, t5, t6; |
| load_u8_4x7(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, vl); |
| |
| // Convert to 16-bit |
| vint16mf2_t s0 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t0, vl)); |
| vint16mf2_t s1 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t1, vl)); |
| vint16mf2_t s2 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t2, vl)); |
| vint16mf2_t s3 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t3, vl)); |
| vint16mf2_t s4 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t4, vl)); |
| vint16mf2_t s5 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t5, vl)); |
| vint16mf2_t s6 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t6, vl)); |
| |
| s += 7 * src_stride; |
| |
| do { |
| // Load next rows of data |
| t0 = __riscv_vle8_v_u8mf4(s + 0 * src_stride, vl); |
| t1 = __riscv_vle8_v_u8mf4(s + 1 * src_stride, vl); |
| t2 = __riscv_vle8_v_u8mf4(s + 2 * src_stride, vl); |
| t3 = __riscv_vle8_v_u8mf4(s + 3 * src_stride, vl); |
| |
| // Convert to 16-bit |
| vint16mf2_t s7 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t0, vl)); |
| vint16mf2_t s8 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t1, vl)); |
| vint16mf2_t s9 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t2, vl)); |
| vint16mf2_t s10 = __riscv_vreinterpret_v_u16mf2_i16mf2( |
| __riscv_vzext_vf2_u16mf2(t3, vl)); |
| |
| // Perform convolution |
| vuint16mf2_t d0 = convolve8_4_y_rvv(s0, s1, s2, s3, s4, s5, s6, s7, |
| y_filter, round_offset, vl); |
| vuint16mf2_t d1 = convolve8_4_y_rvv(s1, s2, s3, s4, s5, s6, s7, s8, |
| y_filter, round_offset, vl); |
| vuint16mf2_t d2 = convolve8_4_y_rvv(s2, s3, s4, s5, s6, s7, s8, s9, |
| y_filter, round_offset, vl); |
| vuint16mf2_t d3 = convolve8_4_y_rvv(s3, s4, s5, s6, s7, s8, s9, s10, |
| y_filter, round_offset, vl); |
| |
| // Store result |
| store_u16_4x4(d, dst_stride, d0, d1, d2, d3, vl); |
| |
| // Update sliding window |
| s0 = __riscv_vmv_v_v_i16mf2(s4, vl); |
| s1 = __riscv_vmv_v_v_i16mf2(s5, vl); |
| s2 = __riscv_vmv_v_v_i16mf2(s6, vl); |
| s3 = __riscv_vmv_v_v_i16mf2(s7, vl); |
| s4 = __riscv_vmv_v_v_i16mf2(s8, vl); |
| s5 = __riscv_vmv_v_v_i16mf2(s9, vl); |
| s6 = __riscv_vmv_v_v_i16mf2(s10, vl); |
| s += 4 * src_stride; |
| d += 4 * dst_stride; |
| height -= 4; |
| } while (height != 0); |
| src_ptr += 4; |
| dst_ptr += 4; |
| width -= 4; |
| } while (width != 0); |
| } else { |
| size_t vl = __riscv_vsetvl_e16m1(w); |
| do { |
| const uint8_t *s = src_ptr; |
| CONV_BUF_TYPE *d = dst_ptr; |
| int height = h; |
| |
| __builtin_prefetch(s + 0 * src_stride); |
| __builtin_prefetch(s + 1 * src_stride); |
| __builtin_prefetch(s + 2 * src_stride); |
| __builtin_prefetch(s + 3 * src_stride); |
| __builtin_prefetch(s + 4 * src_stride); |
| __builtin_prefetch(s + 5 * src_stride); |
| __builtin_prefetch(s + 6 * src_stride); |
| |
| // Load initial 7 rows of data |
| vuint8mf2_t t0, t1, t2, t3, t4, t5, t6; |
| load_u8_8x7(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, vl); |
| |
| // Convert to 16-bit |
| vint16m1_t s0 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t0, vl)); |
| vint16m1_t s1 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t1, vl)); |
| vint16m1_t s2 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t2, vl)); |
| vint16m1_t s3 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t3, vl)); |
| vint16m1_t s4 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t4, vl)); |
| vint16m1_t s5 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t5, vl)); |
| vint16m1_t s6 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t6, vl)); |
| |
| s += 7 * src_stride; |
| |
| do { |
| vuint8mf2_t t7; |
| load_u8_8x8(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7, vl); |
| |
| // Convert to 16-bit |
| vint16m1_t s7 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t0, vl)); |
| vint16m1_t s8 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t1, vl)); |
| vint16m1_t s9 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t2, vl)); |
| vint16m1_t s10 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t3, vl)); |
| vint16m1_t s11 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t4, vl)); |
| vint16m1_t s12 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t5, vl)); |
| vint16m1_t s13 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t6, vl)); |
| vint16m1_t s14 = |
| __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t7, vl)); |
| |
| __builtin_prefetch(d + 0 * dst_stride); |
| __builtin_prefetch(d + 1 * dst_stride); |
| __builtin_prefetch(d + 2 * dst_stride); |
| __builtin_prefetch(d + 3 * dst_stride); |
| |
| // Perform convolution |
| vuint16m1_t d0 = convolve8_8_y_rvv(s0, s1, s2, s3, s4, s5, s6, s7, |
| y_filter, round_offset, vl); |
| vuint16m1_t d1 = convolve8_8_y_rvv(s1, s2, s3, s4, s5, s6, s7, s8, |
| y_filter, round_offset, vl); |
| vuint16m1_t d2 = convolve8_8_y_rvv(s2, s3, s4, s5, s6, s7, s8, s9, |
| y_filter, round_offset, vl); |
| vuint16m1_t d3 = convolve8_8_y_rvv(s3, s4, s5, s6, s7, s8, s9, s10, |
| y_filter, round_offset, vl); |
| vuint16m1_t d4 = convolve8_8_y_rvv(s4, s5, s6, s7, s8, s9, s10, s11, |
| y_filter, round_offset, vl); |
| vuint16m1_t d5 = convolve8_8_y_rvv(s5, s6, s7, s8, s9, s10, s11, s12, |
| y_filter, round_offset, vl); |
| vuint16m1_t d6 = convolve8_8_y_rvv(s6, s7, s8, s9, s10, s11, s12, s13, |
| y_filter, round_offset, vl); |
| vuint16m1_t d7 = convolve8_8_y_rvv(s7, s8, s9, s10, s11, s12, s13, s14, |
| y_filter, round_offset, vl); |
| |
| // Store result |
| store_u16_8x8(d, dst_stride, d0, d1, d2, d3, d4, d5, d6, d7, vl); |
| |
| // Update sliding window |
| s0 = __riscv_vmv_v_v_i16m1(s8, vl); |
| s1 = __riscv_vmv_v_v_i16m1(s9, vl); |
| s2 = __riscv_vmv_v_v_i16m1(s10, vl); |
| s3 = __riscv_vmv_v_v_i16m1(s11, vl); |
| s4 = __riscv_vmv_v_v_i16m1(s12, vl); |
| s5 = __riscv_vmv_v_v_i16m1(s13, vl); |
| s6 = __riscv_vmv_v_v_i16m1(s14, vl); |
| s += 8 * src_stride; |
| d += 8 * dst_stride; |
| height -= 8; |
| } while (height != 0); |
| src_ptr += vl; |
| dst_ptr += vl; |
| width -= vl; |
| } while (width > 0); |
| } |
| } |
| void av1_dist_wtd_convolve_y_rvv(const uint8_t *src, int src_stride, |
| uint8_t *dst8, int dst8_stride, int w, int h, |
| const InterpFilterParams *filter_params_y, |
| const int subpel_y_qn, |
| ConvolveParams *conv_params) { |
| assert(w % 4 == 0); |
| assert(h % 4 == 0); |
| |
| // Vertical filter. |
| const int16_t *y_filter_ptr = av1_get_interp_filter_subpel_kernel( |
| filter_params_y, subpel_y_qn & SUBPEL_MASK); |
| // Filter values are even, so downshift by 1 to reduce intermediate |
| // precision requirements. |
| int16_t y_filter[8]; |
| for (int i = 0; i < 8; i++) y_filter[i] = *y_filter_ptr++ >> 1; |
| |
| const int vert_offset = filter_params_y->taps / 2 - 1; |
| const uint8_t *src_ptr = src - (vert_offset * src_stride); |
| |
| if (get_filter_tap(filter_params_y, subpel_y_qn) <= 6) { |
| if (conv_params->do_average) { |
| if (UNLIKELY(conv_params->use_dist_wtd_comp_avg)) { |
| dist_wtd_convolve_y_6tap_dist_wtd_avg_rvv(src_ptr + src_stride, |
| src_stride, dst8, dst8_stride, |
| w, h, y_filter, conv_params); |
| } else { |
| dist_wtd_convolve_y_6tap_avg_rvv(src_ptr + src_stride, src_stride, dst8, |
| dst8_stride, w, h, y_filter, |
| conv_params); |
| } |
| } else { |
| dist_wtd_convolve_y_6tap_rvv(src_ptr + src_stride, src_stride, w, h, |
| y_filter, conv_params); |
| } |
| } else { |
| if (conv_params->do_average) { |
| if (UNLIKELY(conv_params->use_dist_wtd_comp_avg)) { |
| dist_wtd_convolve_y_8tap_dist_wtd_avg_rvv(src_ptr, src_stride, dst8, |
| dst8_stride, w, h, y_filter, |
| conv_params); |
| } else { |
| dist_wtd_convolve_y_8tap_avg_rvv(src_ptr, src_stride, dst8, dst8_stride, |
| w, h, y_filter, conv_params); |
| } |
| } else { |
| dist_wtd_convolve_y_8tap_rvv(src_ptr, src_stride, w, h, y_filter, |
| conv_params); |
| } |
| } |
| } |
