add wiener convolve optimization Change-Id: I38a5fbd8bb0c326e3b9b68fd5917789438d09ee3
diff --git a/av1/av1.cmake b/av1/av1.cmake index c440720..5e730f9 100644 --- a/av1/av1.cmake +++ b/av1/av1.cmake
@@ -482,7 +482,8 @@ "${AOM_ROOT}/av1/common/riscv/compound_convolve_rvv.c" "${AOM_ROOT}/av1/common/riscv/convolve_rvv.c" "${AOM_ROOT}/av1/common/riscv/highbd_compound_convolve_rvv.c" - "${AOM_ROOT}/av1/common/riscv/highbd_convolve_rvv.c") + "${AOM_ROOT}/av1/common/riscv/highbd_convolve_rvv.c" + "${AOM_ROOT}/av1/common/riscv/wiener_convolve_rvv.c") if(CONFIG_THREE_PASS) list(APPEND AOM_AV1_ENCODER_SOURCES "${AOM_ROOT}/av1/encoder/thirdpass.c"
diff --git a/av1/common/av1_rtcd_defs.pl b/av1/common/av1_rtcd_defs.pl index a3d0d12..d5d6370 100644 --- a/av1/common/av1_rtcd_defs.pl +++ b/av1/common/av1_rtcd_defs.pl
@@ -113,7 +113,7 @@ if ((aom_config("CONFIG_REALTIME_ONLY") ne "yes") || (aom_config("CONFIG_AV1_DECODER") eq "yes")) { add_proto qw/void av1_wiener_convolve_add_src/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params"; - specialize qw/av1_wiener_convolve_add_src sse2 avx2 neon/; + specialize qw/av1_wiener_convolve_add_src sse2 avx2 neon rvv/; } # directional intra predictor functions
diff --git a/av1/common/riscv/wiener_convolve_rvv.c b/av1/common/riscv/wiener_convolve_rvv.c new file mode 100644 index 0000000..513b1a7 --- /dev/null +++ b/av1/common/riscv/wiener_convolve_rvv.c
@@ -0,0 +1,518 @@ +/* + * 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 "config/aom_config.h" +#include "config/av1_rtcd.h" + +#include "aom_dsp/riscv/mem_rvv.h" +#include "av1/common/convolve.h" +#include "av1/common/restoration.h" + +static inline vuint16m1_t wiener_convolve5_8_2d_h_rvv( + vuint8mf2_t t0, vuint8mf2_t t1, vuint8mf2_t t2, vuint8mf2_t t3, + vuint8mf2_t t4, vint16m1_t x_f1, vint16m1_t x_f2, vint16m1_t x_f3, + vint32m2_t round_vec, uint16_t im_max_val, size_t vl) { + // Since the Wiener filter is symmetric about the middle tap (tap 2) add + // mirrored source elements before multiplying filter coefficients. + vint16m1_t s04 = + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vwaddu_vv_u16m1(t0, t4, vl)); + vint16m1_t s13 = + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vwaddu_vv_u16m1(t1, t3, vl)); + vint16m1_t s2 = + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t2, vl)); + + // x_filter[0] = 0. (5-tap filters are 0-padded to 7 taps.) + vint32m2_t sum = round_vec; + sum = __riscv_vwmacc_vv_i32m2(sum, x_f1, s04, vl); + sum = __riscv_vwmacc_vv_i32m2(sum, x_f2, s13, vl); + sum = __riscv_vwmacc_vv_i32m2(sum, x_f3, s2, vl); + +#if __riscv_v_intrinsic >= 12000 + vuint16m1_t res = __riscv_vnclipu_wx_u16m1( + __riscv_vreinterpret_v_i32m2_u32m2(__riscv_vmax_vx_i32m2(sum, 0, vl)), + WIENER_ROUND0_BITS, __RISCV_VXRM_RNU, vl); +#else + vuint16m1_t res = __riscv_vnclipu_wx_u16m1( + __riscv_vreinterpret_v_i32m2_u32m2(__riscv_vmax_vx_i32m2(sum, 0, vl)), + WIENER_ROUND0_BITS, vl); +#endif + return __riscv_vminu_vx_u16m1(res, im_max_val, vl); +} + +static inline void convolve_add_src_horiz_5tap_rvv( + const uint8_t *src_ptr, ptrdiff_t src_stride, uint16_t *dst_ptr, + ptrdiff_t dst_stride, int w, int h, const int16_t *x_filter, + int32_t round_val, uint16_t im_max_val) { + const size_t vl_max = __riscv_vsetvlmax_e16m1(); + const vint16m1_t x_f1 = __riscv_vmv_v_x_i16m1(x_filter[1], vl_max); + const vint16m1_t x_f2 = __riscv_vmv_v_x_i16m1(x_filter[2], vl_max); + const vint16m1_t x_f3 = __riscv_vmv_v_x_i16m1(x_filter[3], vl_max); + const vint32m2_t round_vec = __riscv_vmv_v_x_i32m2(round_val, vl_max); + + int row = 0; + for (; row <= h - 2; row += 2) { + const uint8_t *sa = src_ptr; + const uint8_t *sb = src_ptr + src_stride; + uint16_t *da = dst_ptr; + uint16_t *db = dst_ptr + dst_stride; + int width = w; + + while (width > 0) { + const size_t vl = __riscv_vsetvl_e16m1((size_t)width); + vuint8mf2_t a0, a1, a2, a3, a4; + load_u8_8x5(sa, 1, &a0, &a1, &a2, &a3, &a4, vl); + vuint8mf2_t b0, b1, b2, b3, b4; + load_u8_8x5(sb, 1, &b0, &b1, &b2, &b3, &b4, vl); + + const vuint16m1_t ra = wiener_convolve5_8_2d_h_rvv( + a0, a1, a2, a3, a4, x_f1, x_f2, x_f3, round_vec, im_max_val, vl); + const vuint16m1_t rb = wiener_convolve5_8_2d_h_rvv( + b0, b1, b2, b3, b4, x_f1, x_f2, x_f3, round_vec, im_max_val, vl); + __riscv_vse16_v_u16m1(da, ra, vl); + __riscv_vse16_v_u16m1(db, rb, vl); + + sa += vl; + sb += vl; + da += vl; + db += vl; + width -= (int)vl; + } + src_ptr += 2 * src_stride; + dst_ptr += 2 * dst_stride; + } + for (; row < h; row++) { + const uint8_t *s = src_ptr; + uint16_t *d = dst_ptr; + int width = w; + while (width > 0) { + const size_t vl = __riscv_vsetvl_e16m1((size_t)width); + vuint8mf2_t s0, s1, s2, s3, s4; + load_u8_8x5(s, 1, &s0, &s1, &s2, &s3, &s4, vl); + const vuint16m1_t d0 = wiener_convolve5_8_2d_h_rvv( + s0, s1, s2, s3, s4, x_f1, x_f2, x_f3, round_vec, im_max_val, vl); + __riscv_vse16_v_u16m1(d, d0, vl); + s += vl; + d += vl; + width -= (int)vl; + } + src_ptr += src_stride; + dst_ptr += dst_stride; + } +} + +static inline vuint16m1_t wiener_convolve7_8_2d_h_rvv( + vuint8mf2_t t0, vuint8mf2_t t1, vuint8mf2_t t2, vuint8mf2_t t3, + vuint8mf2_t t4, vuint8mf2_t t5, vuint8mf2_t t6, vint16m1_t x_f0, + vint16m1_t x_f1, vint16m1_t x_f2, vint16m1_t x_f3, vint32m2_t round_vec, + uint16_t im_max_val, size_t vl) { + // Since the Wiener filter is symmetric about the middle tap (tap 3) add + // mirrored source elements before multiplying by filter coefficients. + vint16m1_t s06 = + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vwaddu_vv_u16m1(t0, t6, vl)); + vint16m1_t s15 = + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vwaddu_vv_u16m1(t1, t5, vl)); + vint16m1_t s24 = + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vwaddu_vv_u16m1(t2, t4, vl)); + vint16m1_t s3 = + __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(t3, vl)); + + vint32m2_t sum = round_vec; + sum = __riscv_vwmacc_vv_i32m2(sum, x_f0, s06, vl); + sum = __riscv_vwmacc_vv_i32m2(sum, x_f1, s15, vl); + sum = __riscv_vwmacc_vv_i32m2(sum, x_f2, s24, vl); + sum = __riscv_vwmacc_vv_i32m2(sum, x_f3, s3, vl); + +#if __riscv_v_intrinsic >= 12000 + vuint16m1_t res = __riscv_vnclipu_wx_u16m1( + __riscv_vreinterpret_v_i32m2_u32m2(__riscv_vmax_vx_i32m2(sum, 0, vl)), + WIENER_ROUND0_BITS, __RISCV_VXRM_RNU, vl); +#else + vuint16m1_t res = __riscv_vnclipu_wx_u16m1( + __riscv_vreinterpret_v_i32m2_u32m2(__riscv_vmax_vx_i32m2(sum, 0, vl)), + WIENER_ROUND0_BITS, vl); +#endif + return __riscv_vminu_vx_u16m1(res, im_max_val, vl); +} + +static inline void convolve_add_src_horiz_7tap_rvv( + const uint8_t *src_ptr, ptrdiff_t src_stride, uint16_t *dst_ptr, + ptrdiff_t dst_stride, int w, int h, const int16_t *x_filter, + int32_t round_val, uint16_t im_max_val) { + const size_t vl_max = __riscv_vsetvlmax_e16m1(); + const vint16m1_t x_f0 = __riscv_vmv_v_x_i16m1(x_filter[0], vl_max); + const vint16m1_t x_f1 = __riscv_vmv_v_x_i16m1(x_filter[1], vl_max); + const vint16m1_t x_f2 = __riscv_vmv_v_x_i16m1(x_filter[2], vl_max); + const vint16m1_t x_f3 = __riscv_vmv_v_x_i16m1(x_filter[3], vl_max); + const vint32m2_t round_vec = __riscv_vmv_v_x_i32m2(round_val, vl_max); + + int row = 0; + for (; row <= h - 2; row += 2) { + const uint8_t *sa = src_ptr; + const uint8_t *sb = src_ptr + src_stride; + uint16_t *da = dst_ptr; + uint16_t *db = dst_ptr + dst_stride; + int width = w; + + while (width > 0) { + const size_t vl = __riscv_vsetvl_e16m1((size_t)width); + vuint8mf2_t a0, a1, a2, a3, a4, a5, a6; + load_u8_8x7(sa, 1, &a0, &a1, &a2, &a3, &a4, &a5, &a6, vl); + vuint8mf2_t b0, b1, b2, b3, b4, b5, b6; + load_u8_8x7(sb, 1, &b0, &b1, &b2, &b3, &b4, &b5, &b6, vl); + + const vuint16m1_t ra = + wiener_convolve7_8_2d_h_rvv(a0, a1, a2, a3, a4, a5, a6, x_f0, x_f1, + x_f2, x_f3, round_vec, im_max_val, vl); + const vuint16m1_t rb = + wiener_convolve7_8_2d_h_rvv(b0, b1, b2, b3, b4, b5, b6, x_f0, x_f1, + x_f2, x_f3, round_vec, im_max_val, vl); + __riscv_vse16_v_u16m1(da, ra, vl); + __riscv_vse16_v_u16m1(db, rb, vl); + + sa += vl; + sb += vl; + da += vl; + db += vl; + width -= (int)vl; + } + src_ptr += 2 * src_stride; + dst_ptr += 2 * dst_stride; + } + for (; row < h; row++) { + const uint8_t *s = src_ptr; + uint16_t *d = dst_ptr; + int width = w; + while (width > 0) { + const size_t vl = __riscv_vsetvl_e16m1((size_t)width); + vuint8mf2_t s0, s1, s2, s3, s4, s5, s6; + load_u8_8x7(s, 1, &s0, &s1, &s2, &s3, &s4, &s5, &s6, vl); + const vuint16m1_t d0 = + wiener_convolve7_8_2d_h_rvv(s0, s1, s2, s3, s4, s5, s6, x_f0, x_f1, + x_f2, x_f3, round_vec, im_max_val, vl); + __riscv_vse16_v_u16m1(d, d0, vl); + s += vl; + d += vl; + width -= (int)vl; + } + src_ptr += src_stride; + dst_ptr += dst_stride; + } +} + +static inline vuint8mf2_t wiener_convolve5_8_2d_v_rvv( + vint16m1_t s0, vint16m1_t s1, vint16m1_t s2, vint16m1_t s3, vint16m1_t s4, + vint16m1_t y_f1, vint16m1_t y_f2, vint16m1_t y_f3, vint32m2_t round_vec, + size_t vl) { + // Since the Wiener filter is symmetric about the middle tap (tap 2) add + // mirrored source elements before multiplying by filter coefficients. + vint16m1_t s04 = __riscv_vadd_vv_i16m1(s0, s4, vl); + vint16m1_t s13 = __riscv_vadd_vv_i16m1(s1, s3, vl); + + vint32m2_t sum = round_vec; + sum = __riscv_vwmacc_vv_i32m2(sum, y_f1, s04, vl); + sum = __riscv_vwmacc_vv_i32m2(sum, y_f2, s13, vl); + sum = __riscv_vwmacc_vv_i32m2(sum, y_f3, s2, vl); + +#if __riscv_v_intrinsic >= 12000 + vint16m1_t res_16 = __riscv_vnclip_wx_i16m1( + sum, 2 * FILTER_BITS - WIENER_ROUND0_BITS, __RISCV_VXRM_RDN, vl); + return __riscv_vnclipu_wx_u8mf2( + __riscv_vreinterpret_v_i16m1_u16m1(__riscv_vmax_vx_i16m1(res_16, 0, vl)), + 0, __RISCV_VXRM_RNU, vl); +#else + vint16m1_t res_16 = + __riscv_vnsra_wx_i16m1(sum, 2 * FILTER_BITS - WIENER_ROUND0_BITS, vl); + return __riscv_vnclipu_wx_u8mf2( + __riscv_vreinterpret_v_i16m1_u16m1(__riscv_vmax_vx_i16m1(res_16, 0, vl)), + 0, vl); +#endif +} + +static inline void convolve_add_src_vert_5tap_rvv( + const uint16_t *src, ptrdiff_t src_stride, uint8_t *dst, + ptrdiff_t dst_stride, int w, int h, const int16_t *y_filter, + int32_t round_val) { + const size_t vl_max = __riscv_vsetvlmax_e16m1(); + const vint16m1_t y_f1 = __riscv_vmv_v_x_i16m1(y_filter[1], vl_max); + const vint16m1_t y_f2 = __riscv_vmv_v_x_i16m1(y_filter[2], vl_max); + const vint16m1_t y_f3 = __riscv_vmv_v_x_i16m1(y_filter[3], vl_max); + const vint32m2_t round_vec = __riscv_vmv_v_x_i32m2(round_val, vl_max); + + int width = w; + const int16_t *s_base = (const int16_t *)src; + uint8_t *d_base = dst; + + while (width > 0) { + const size_t vl = __riscv_vsetvl_e16m1((size_t)width); + + const int16_t *s = s_base; + vint16m1_t s0 = __riscv_vle16_v_i16m1(s, vl); + s += src_stride; + vint16m1_t s1 = __riscv_vle16_v_i16m1(s, vl); + s += src_stride; + vint16m1_t s2 = __riscv_vle16_v_i16m1(s, vl); + s += src_stride; + vint16m1_t s3 = __riscv_vle16_v_i16m1(s, vl); + s += src_stride; + + uint8_t *d = d_base; + int i = 0; + for (; i <= h - 4; i += 4) { + vint16m1_t s4 = __riscv_vle16_v_i16m1(s, vl); + s += src_stride; + vint16m1_t s5 = __riscv_vle16_v_i16m1(s, vl); + s += src_stride; + vint16m1_t s6 = __riscv_vle16_v_i16m1(s, vl); + s += src_stride; + vint16m1_t s7 = __riscv_vle16_v_i16m1(s, vl); + s += src_stride; + + vuint8mf2_t d0 = wiener_convolve5_8_2d_v_rvv(s0, s1, s2, s3, s4, y_f1, + y_f2, y_f3, round_vec, vl); + vuint8mf2_t d1 = wiener_convolve5_8_2d_v_rvv(s1, s2, s3, s4, s5, y_f1, + y_f2, y_f3, round_vec, vl); + vuint8mf2_t d2 = wiener_convolve5_8_2d_v_rvv(s2, s3, s4, s5, s6, y_f1, + y_f2, y_f3, round_vec, vl); + vuint8mf2_t d3 = wiener_convolve5_8_2d_v_rvv(s3, s4, s5, s6, s7, y_f1, + y_f2, y_f3, round_vec, vl); + + __riscv_vse8_v_u8mf2(d, d0, vl); + d += dst_stride; + __riscv_vse8_v_u8mf2(d, d1, vl); + d += dst_stride; + __riscv_vse8_v_u8mf2(d, d2, vl); + d += dst_stride; + __riscv_vse8_v_u8mf2(d, d3, vl); + d += dst_stride; + + s0 = s4; + s1 = s5; + s2 = s6; + s3 = s7; + } + + for (; i < h; i++) { + vint16m1_t s4 = __riscv_vle16_v_i16m1(s, vl); + s += src_stride; + vuint8mf2_t d0 = wiener_convolve5_8_2d_v_rvv(s0, s1, s2, s3, s4, y_f1, + y_f2, y_f3, round_vec, vl); + __riscv_vse8_v_u8mf2(d, d0, vl); + d += dst_stride; + s0 = s1; + s1 = s2; + s2 = s3; + s3 = s4; + } + + s_base += vl; + d_base += vl; + width -= (int)vl; + } +} + +static inline vuint8mf2_t wiener_convolve7_8_2d_v_rvv( + vint16m1_t s0, vint16m1_t s1, vint16m1_t s2, vint16m1_t s3, vint16m1_t s4, + vint16m1_t s5, vint16m1_t s6, vint16m1_t y_f0, vint16m1_t y_f1, + vint16m1_t y_f2, vint16m1_t y_f3, vint32m2_t round_vec, size_t vl) { + // Since the Wiener filter is symmetric about the middle tap (tap 3) add + // mirrored source elements before multiplying by filter coefficients. + vint16m1_t s06 = __riscv_vadd_vv_i16m1(s0, s6, vl); + vint16m1_t s15 = __riscv_vadd_vv_i16m1(s1, s5, vl); + vint16m1_t s24 = __riscv_vadd_vv_i16m1(s2, s4, vl); + + vint32m2_t sum = round_vec; + sum = __riscv_vwmacc_vv_i32m2(sum, y_f0, s06, vl); + sum = __riscv_vwmacc_vv_i32m2(sum, y_f1, s15, vl); + sum = __riscv_vwmacc_vv_i32m2(sum, y_f2, s24, vl); + sum = __riscv_vwmacc_vv_i32m2(sum, y_f3, s3, vl); + +#if __riscv_v_intrinsic >= 12000 + vint16m1_t res_16 = __riscv_vnclip_wx_i16m1( + sum, 2 * FILTER_BITS - WIENER_ROUND0_BITS, __RISCV_VXRM_RDN, vl); + return __riscv_vnclipu_wx_u8mf2( + __riscv_vreinterpret_v_i16m1_u16m1(__riscv_vmax_vx_i16m1(res_16, 0, vl)), + 0, __RISCV_VXRM_RNU, vl); +#else + vint16m1_t res_16 = + __riscv_vnsra_wx_i16m1(sum, 2 * FILTER_BITS - WIENER_ROUND0_BITS, vl); + return __riscv_vnclipu_wx_u8mf2( + __riscv_vreinterpret_v_i16m1_u16m1(__riscv_vmax_vx_i16m1(res_16, 0, vl)), + 0, vl); +#endif +} + +static inline void convolve_add_src_vert_7tap_rvv( + const uint16_t *src, ptrdiff_t src_stride, uint8_t *dst, + ptrdiff_t dst_stride, int w, int h, const int16_t *y_filter, + int32_t round_val) { + const size_t vl_max = __riscv_vsetvlmax_e16m1(); + const vint16m1_t y_f0 = __riscv_vmv_v_x_i16m1(y_filter[0], vl_max); + const vint16m1_t y_f1 = __riscv_vmv_v_x_i16m1(y_filter[1], vl_max); + const vint16m1_t y_f2 = __riscv_vmv_v_x_i16m1(y_filter[2], vl_max); + const vint16m1_t y_f3 = __riscv_vmv_v_x_i16m1(y_filter[3], vl_max); + const vint32m2_t round_vec = __riscv_vmv_v_x_i32m2(round_val, vl_max); + + int width = w; + const int16_t *s_base = (const int16_t *)src; + uint8_t *d_base = dst; + + while (width > 0) { + const size_t vl = __riscv_vsetvl_e16m1((size_t)width); + + const int16_t *s = s_base; + vint16m1_t s0 = __riscv_vle16_v_i16m1(s, vl); + s += src_stride; + vint16m1_t s1 = __riscv_vle16_v_i16m1(s, vl); + s += src_stride; + vint16m1_t s2 = __riscv_vle16_v_i16m1(s, vl); + s += src_stride; + vint16m1_t s3 = __riscv_vle16_v_i16m1(s, vl); + s += src_stride; + vint16m1_t s4 = __riscv_vle16_v_i16m1(s, vl); + s += src_stride; + vint16m1_t s5 = __riscv_vle16_v_i16m1(s, vl); + s += src_stride; + + uint8_t *d = d_base; + int i = 0; + for (; i <= h - 4; i += 4) { + vint16m1_t s6 = __riscv_vle16_v_i16m1(s, vl); + s += src_stride; + vint16m1_t s7 = __riscv_vle16_v_i16m1(s, vl); + s += src_stride; + vint16m1_t s8 = __riscv_vle16_v_i16m1(s, vl); + s += src_stride; + vint16m1_t s9 = __riscv_vle16_v_i16m1(s, vl); + s += src_stride; + + vuint8mf2_t d0 = wiener_convolve7_8_2d_v_rvv( + s0, s1, s2, s3, s4, s5, s6, y_f0, y_f1, y_f2, y_f3, round_vec, vl); + vuint8mf2_t d1 = wiener_convolve7_8_2d_v_rvv( + s1, s2, s3, s4, s5, s6, s7, y_f0, y_f1, y_f2, y_f3, round_vec, vl); + vuint8mf2_t d2 = wiener_convolve7_8_2d_v_rvv( + s2, s3, s4, s5, s6, s7, s8, y_f0, y_f1, y_f2, y_f3, round_vec, vl); + vuint8mf2_t d3 = wiener_convolve7_8_2d_v_rvv( + s3, s4, s5, s6, s7, s8, s9, y_f0, y_f1, y_f2, y_f3, round_vec, vl); + + __riscv_vse8_v_u8mf2(d, d0, vl); + d += dst_stride; + __riscv_vse8_v_u8mf2(d, d1, vl); + d += dst_stride; + __riscv_vse8_v_u8mf2(d, d2, vl); + d += dst_stride; + __riscv_vse8_v_u8mf2(d, d3, vl); + d += dst_stride; + + s0 = s4; + s1 = s5; + s2 = s6; + s3 = s7; + s4 = s8; + s5 = s9; + } + + for (; i < h; i++) { + vint16m1_t s6 = __riscv_vle16_v_i16m1(s, vl); + s += src_stride; + vuint8mf2_t d0 = wiener_convolve7_8_2d_v_rvv( + s0, s1, s2, s3, s4, s5, s6, y_f0, y_f1, y_f2, y_f3, round_vec, vl); + __riscv_vse8_v_u8mf2(d, d0, vl); + d += dst_stride; + s0 = s1; + s1 = s2; + s2 = s3; + s3 = s4; + s4 = s5; + s5 = s6; + } + + s_base += vl; + d_base += vl; + width -= (int)vl; + } +} + +static inline int get_wiener_filter_taps(const int16_t *filter) { + assert(filter[7] == 0); + if (filter[0] == 0 && filter[6] == 0) { + return WIENER_WIN_REDUCED; + } + return WIENER_WIN; +} + +// Wiener filter 2D +// Apply horizontal filter and store in a temporary buffer. When applying +// vertical filter, overwrite the original pixel values. +void av1_wiener_convolve_add_src_rvv(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const int16_t *x_filter, int x_step_q4, + const int16_t *y_filter, int y_step_q4, + int w, int h, + const WienerConvolveParams *conv_params) { + (void)x_step_q4; + (void)y_step_q4; + (void)conv_params; + + assert(w % 8 == 0); + assert(w <= MAX_SB_SIZE && h <= MAX_SB_SIZE); + assert(x_step_q4 == 16 && y_step_q4 == 16); + assert(x_filter[7] == 0 && y_filter[7] == 0); + // For bd == 8, assert horizontal filtering output will not exceed 15-bit: + assert(8 + 1 + FILTER_BITS - conv_params->round_0 <= 15); + + const int x_filter_taps = get_wiener_filter_taps(x_filter); + const int y_filter_taps = get_wiener_filter_taps(y_filter); + + DECLARE_ALIGNED(16, uint16_t, + im_block[(MAX_SB_SIZE + WIENER_WIN - 1) * MAX_SB_SIZE]); + + const int bd = 8; + const uint16_t im_max_val = + (1 << (bd + 1 + FILTER_BITS - WIENER_ROUND0_BITS)) - 1; + const int32_t horiz_round_val = 1 << (bd + FILTER_BITS - 1); + const int32_t vert_round_val = + (1 << (2 * FILTER_BITS - WIENER_ROUND0_BITS - 1)) - + (1 << (bd + (2 * FILTER_BITS - WIENER_ROUND0_BITS) - 1)); + + const int im_stride = (w <= 32) ? w : MAX_SB_SIZE; + const int im_h = h + y_filter_taps - 1; + const int horiz_offset = x_filter_taps / 2; + const int vert_offset = (y_filter_taps / 2) * (int)src_stride; + + int16_t x_filter_5[8]; + int16_t y_filter_5[8]; + // Add 128 to tap 3. (Needed for rounding.) + for (int i = 0; i < 8; ++i) x_filter_5[i] = x_filter[i]; + x_filter_5[3] += 128; + for (int i = 0; i < 8; ++i) y_filter_5[i] = y_filter[i]; + y_filter_5[3] += 128; + + if (x_filter_taps == WIENER_WIN_REDUCED) { + convolve_add_src_horiz_5tap_rvv(src - horiz_offset - vert_offset, + src_stride, im_block, im_stride, w, im_h, + x_filter_5, horiz_round_val, im_max_val); + } else { + convolve_add_src_horiz_7tap_rvv(src - horiz_offset - vert_offset, + src_stride, im_block, im_stride, w, im_h, + x_filter_5, horiz_round_val, im_max_val); + } + + if (y_filter_taps == WIENER_WIN_REDUCED) { + convolve_add_src_vert_5tap_rvv(im_block, im_stride, dst, dst_stride, w, h, + y_filter_5, vert_round_val); + } else { + convolve_add_src_vert_7tap_rvv(im_block, im_stride, dst, dst_stride, w, h, + y_filter_5, vert_round_val); + } +}
diff --git a/test/hiprec_convolve_test.cc b/test/hiprec_convolve_test.cc index 2dde9da..c74bc76 100644 --- a/test/hiprec_convolve_test.cc +++ b/test/hiprec_convolve_test.cc
@@ -45,6 +45,11 @@ libaom_test::AV1HiprecConvolve::BuildParams( av1_wiener_convolve_add_src_neon)); #endif +#if HAVE_RVV +INSTANTIATE_TEST_SUITE_P(RVV, AV1HiprecConvolveTest, + libaom_test::AV1HiprecConvolve::BuildParams( + av1_wiener_convolve_add_src_rvv)); +#endif #if CONFIG_AV1_HIGHBITDEPTH #if HAVE_SSSE3 || HAVE_AVX2 || HAVE_NEON