[FFmpeg-devel] [PATCH v3 2/7] avcodec: [loongarch] Optimize h264_chroma_mc with LASX.
Hao Chen
chenhao at loongson.cn
Wed Dec 15 05:51:05 EET 2021
From: Shiyou Yin <yinshiyou-hf at loongson.cn>
./ffmpeg -i ../1_h264_1080p_30fps_3Mbps.mp4 -f rawvideo -y /dev/null -an
before:170
after :183
Change-Id: I42ff23cc2dc7c32bd1b7e4274da9d9ec87065f20
---
libavcodec/h264chroma.c | 2 +
libavcodec/h264chroma.h | 1 +
libavcodec/loongarch/Makefile | 2 +
.../loongarch/h264chroma_init_loongarch.c | 37 +
libavcodec/loongarch/h264chroma_lasx.c | 1280 +++++++++++
libavcodec/loongarch/h264chroma_lasx.h | 36 +
libavutil/loongarch/loongson_intrinsics.h | 1877 +++++++++++++++++
7 files changed, 3235 insertions(+)
create mode 100644 libavcodec/loongarch/Makefile
create mode 100644 libavcodec/loongarch/h264chroma_init_loongarch.c
create mode 100644 libavcodec/loongarch/h264chroma_lasx.c
create mode 100644 libavcodec/loongarch/h264chroma_lasx.h
create mode 100644 libavutil/loongarch/loongson_intrinsics.h
diff --git a/libavcodec/h264chroma.c b/libavcodec/h264chroma.c
index c2f1f30f5a..0ae6c793e1 100644
--- a/libavcodec/h264chroma.c
+++ b/libavcodec/h264chroma.c
@@ -56,4 +56,6 @@ av_cold void ff_h264chroma_init(H264ChromaContext *c, int bit_depth)
ff_h264chroma_init_x86(c, bit_depth);
if (ARCH_MIPS)
ff_h264chroma_init_mips(c, bit_depth);
+ if (ARCH_LOONGARCH64)
+ ff_h264chroma_init_loongarch(c, bit_depth);
}
diff --git a/libavcodec/h264chroma.h b/libavcodec/h264chroma.h
index 5c89fd12df..3259b4935f 100644
--- a/libavcodec/h264chroma.h
+++ b/libavcodec/h264chroma.h
@@ -36,5 +36,6 @@ void ff_h264chroma_init_arm(H264ChromaContext *c, int bit_depth);
void ff_h264chroma_init_ppc(H264ChromaContext *c, int bit_depth);
void ff_h264chroma_init_x86(H264ChromaContext *c, int bit_depth);
void ff_h264chroma_init_mips(H264ChromaContext *c, int bit_depth);
+void ff_h264chroma_init_loongarch(H264ChromaContext *c, int bit_depth);
#endif /* AVCODEC_H264CHROMA_H */
diff --git a/libavcodec/loongarch/Makefile b/libavcodec/loongarch/Makefile
new file mode 100644
index 0000000000..f8fb54c925
--- /dev/null
+++ b/libavcodec/loongarch/Makefile
@@ -0,0 +1,2 @@
+OBJS-$(CONFIG_H264CHROMA) += loongarch/h264chroma_init_loongarch.o
+LASX-OBJS-$(CONFIG_H264CHROMA) += loongarch/h264chroma_lasx.o
diff --git a/libavcodec/loongarch/h264chroma_init_loongarch.c b/libavcodec/loongarch/h264chroma_init_loongarch.c
new file mode 100644
index 0000000000..0ca24ecc47
--- /dev/null
+++ b/libavcodec/loongarch/h264chroma_init_loongarch.c
@@ -0,0 +1,37 @@
+/*
+ * Copyright (c) 2020 Loongson Technology Corporation Limited
+ * Contributed by Shiyou Yin <yinshiyou-hf at loongson.cn>
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include "h264chroma_lasx.h"
+#include "libavutil/attributes.h"
+#include "libavutil/loongarch/cpu.h"
+#include "libavcodec/h264chroma.h"
+
+av_cold void ff_h264chroma_init_loongarch(H264ChromaContext *c, int bit_depth)
+{
+ int cpu_flags = av_get_cpu_flags();
+ if (have_lasx(cpu_flags)) {
+ if (bit_depth <= 8) {
+ c->put_h264_chroma_pixels_tab[0] = ff_put_h264_chroma_mc8_lasx;
+ c->avg_h264_chroma_pixels_tab[0] = ff_avg_h264_chroma_mc8_lasx;
+ c->put_h264_chroma_pixels_tab[1] = ff_put_h264_chroma_mc4_lasx;
+ }
+ }
+}
diff --git a/libavcodec/loongarch/h264chroma_lasx.c b/libavcodec/loongarch/h264chroma_lasx.c
new file mode 100644
index 0000000000..824a78dfc8
--- /dev/null
+++ b/libavcodec/loongarch/h264chroma_lasx.c
@@ -0,0 +1,1280 @@
+/*
+ * Loongson LASX optimized h264chroma
+ *
+ * Copyright (c) 2020 Loongson Technology Corporation Limited
+ * Contributed by Shiyou Yin <yinshiyou-hf at loongson.cn>
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include "h264chroma_lasx.h"
+#include "libavutil/attributes.h"
+#include "libavutil/avassert.h"
+#include "libavutil/loongarch/loongson_intrinsics.h"
+
+static const uint8_t chroma_mask_arr[64] = {
+ 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8,
+ 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8,
+ 0, 1, 1, 2, 2, 3, 3, 4, 16, 17, 17, 18, 18, 19, 19, 20,
+ 0, 1, 1, 2, 2, 3, 3, 4, 16, 17, 17, 18, 18, 19, 19, 20
+};
+
+static av_always_inline void avc_chroma_hv_8x4_lasx(uint8_t *src, uint8_t *dst,
+ ptrdiff_t stride, uint32_t coef_hor0,
+ uint32_t coef_hor1, uint32_t coef_ver0,
+ uint32_t coef_ver1)
+{
+ ptrdiff_t stride_2x = stride << 1;
+ ptrdiff_t stride_3x = stride_2x + stride;
+ ptrdiff_t stride_4x = stride_2x << 1;
+ __m256i src0, src1, src2, src3, src4, out;
+ __m256i res_hz0, res_hz1, res_hz2, res_vt0, res_vt1;
+ __m256i mask;
+ __m256i coeff_hz_vec0 = __lasx_xvreplgr2vr_b(coef_hor0);
+ __m256i coeff_hz_vec1 = __lasx_xvreplgr2vr_b(coef_hor1);
+ __m256i coeff_hz_vec = __lasx_xvilvl_b(coeff_hz_vec0, coeff_hz_vec1);
+ __m256i coeff_vt_vec0 = __lasx_xvreplgr2vr_h(coef_ver0);
+ __m256i coeff_vt_vec1 = __lasx_xvreplgr2vr_h(coef_ver1);
+
+ DUP2_ARG2(__lasx_xvld, chroma_mask_arr, 0, src, 0, mask, src0);
+ DUP4_ARG2(__lasx_xvldx, src, stride, src, stride_2x, src, stride_3x, src, stride_4x,
+ src1, src2, src3, src4);
+ DUP2_ARG3(__lasx_xvpermi_q, src2, src1, 0x20, src4, src3, 0x20, src1, src3);
+ src0 = __lasx_xvshuf_b(src0, src0, mask);
+ DUP2_ARG3(__lasx_xvshuf_b, src1, src1, mask, src3, src3, mask, src1, src3);
+ DUP2_ARG2(__lasx_xvdp2_h_bu, src0, coeff_hz_vec, src1, coeff_hz_vec, res_hz0, res_hz1);
+ res_hz2 = __lasx_xvdp2_h_bu(src3, coeff_hz_vec);
+ res_vt0 = __lasx_xvmul_h(res_hz1, coeff_vt_vec0);
+ res_vt1 = __lasx_xvmul_h(res_hz2, coeff_vt_vec0);
+ res_hz0 = __lasx_xvpermi_q(res_hz1, res_hz0, 0x20);
+ res_hz1 = __lasx_xvpermi_q(res_hz1, res_hz2, 0x3);
+ res_vt0 = __lasx_xvmadd_h(res_vt0, res_hz0, coeff_vt_vec1);
+ res_vt1 = __lasx_xvmadd_h(res_vt1, res_hz1, coeff_vt_vec1);
+ out = __lasx_xvssrarni_bu_h(res_vt1, res_vt0, 6);
+ __lasx_xvstelm_d(out, dst, 0, 0);
+ __lasx_xvstelm_d(out, dst + stride, 0, 2);
+ __lasx_xvstelm_d(out, dst + stride_2x, 0, 1);
+ __lasx_xvstelm_d(out, dst + stride_3x, 0, 3);
+}
+
+static av_always_inline void avc_chroma_hv_8x8_lasx(uint8_t *src, uint8_t *dst,
+ ptrdiff_t stride, uint32_t coef_hor0,
+ uint32_t coef_hor1, uint32_t coef_ver0,
+ uint32_t coef_ver1)
+{
+ ptrdiff_t stride_2x = stride << 1;
+ ptrdiff_t stride_3x = stride_2x + stride;
+ ptrdiff_t stride_4x = stride << 2;
+ __m256i src0, src1, src2, src3, src4, src5, src6, src7, src8;
+ __m256i out0, out1;
+ __m256i res_hz0, res_hz1, res_hz2, res_hz3, res_hz4;
+ __m256i res_vt0, res_vt1, res_vt2, res_vt3;
+ __m256i mask;
+ __m256i coeff_hz_vec0 = __lasx_xvreplgr2vr_b(coef_hor0);
+ __m256i coeff_hz_vec1 = __lasx_xvreplgr2vr_b(coef_hor1);
+ __m256i coeff_hz_vec = __lasx_xvilvl_b(coeff_hz_vec0, coeff_hz_vec1);
+ __m256i coeff_vt_vec0 = __lasx_xvreplgr2vr_h(coef_ver0);
+ __m256i coeff_vt_vec1 = __lasx_xvreplgr2vr_h(coef_ver1);
+
+ DUP2_ARG2(__lasx_xvld, chroma_mask_arr, 0, src, 0, mask, src0);
+ DUP4_ARG2(__lasx_xvldx, src, stride, src, stride_2x, src, stride_3x, src, stride_4x,
+ src1, src2, src3, src4);
+ src += stride_4x;
+ DUP4_ARG2(__lasx_xvldx, src, stride, src, stride_2x, src, stride_3x, src, stride_4x,
+ src5, src6, src7, src8);
+ DUP4_ARG3(__lasx_xvpermi_q, src2, src1, 0x20, src4, src3, 0x20, src6, src5, 0x20,
+ src8, src7, 0x20, src1, src3, src5, src7);
+ src0 = __lasx_xvshuf_b(src0, src0, mask);
+ DUP4_ARG3(__lasx_xvshuf_b, src1, src1, mask, src3, src3, mask, src5, src5, mask, src7,
+ src7, mask, src1, src3, src5, src7);
+ DUP4_ARG2(__lasx_xvdp2_h_bu, src0, coeff_hz_vec, src1, coeff_hz_vec, src3,
+ coeff_hz_vec, src5, coeff_hz_vec, res_hz0, res_hz1, res_hz2, res_hz3);
+ res_hz4 = __lasx_xvdp2_h_bu(src7, coeff_hz_vec);
+ res_vt0 = __lasx_xvmul_h(res_hz1, coeff_vt_vec0);
+ res_vt1 = __lasx_xvmul_h(res_hz2, coeff_vt_vec0);
+ res_vt2 = __lasx_xvmul_h(res_hz3, coeff_vt_vec0);
+ res_vt3 = __lasx_xvmul_h(res_hz4, coeff_vt_vec0);
+ res_hz0 = __lasx_xvpermi_q(res_hz1, res_hz0, 0x20);
+ res_hz1 = __lasx_xvpermi_q(res_hz1, res_hz2, 0x3);
+ res_hz2 = __lasx_xvpermi_q(res_hz2, res_hz3, 0x3);
+ res_hz3 = __lasx_xvpermi_q(res_hz3, res_hz4, 0x3);
+ DUP4_ARG3(__lasx_xvmadd_h, res_vt0, res_hz0, coeff_vt_vec1, res_vt1, res_hz1, coeff_vt_vec1,
+ res_vt2, res_hz2, coeff_vt_vec1, res_vt3, res_hz3, coeff_vt_vec1,
+ res_vt0, res_vt1, res_vt2, res_vt3);
+ DUP2_ARG3(__lasx_xvssrarni_bu_h, res_vt1, res_vt0, 6, res_vt3, res_vt2, 6, out0, out1);
+ __lasx_xvstelm_d(out0, dst, 0, 0);
+ __lasx_xvstelm_d(out0, dst + stride, 0, 2);
+ __lasx_xvstelm_d(out0, dst + stride_2x, 0, 1);
+ __lasx_xvstelm_d(out0, dst + stride_3x, 0, 3);
+ dst += stride_4x;
+ __lasx_xvstelm_d(out1, dst, 0, 0);
+ __lasx_xvstelm_d(out1, dst + stride, 0, 2);
+ __lasx_xvstelm_d(out1, dst + stride_2x, 0, 1);
+ __lasx_xvstelm_d(out1, dst + stride_3x, 0, 3);
+}
+
+static av_always_inline void avc_chroma_hz_8x4_lasx(uint8_t *src, uint8_t *dst,
+ ptrdiff_t stride, uint32_t coeff0, uint32_t coeff1)
+{
+ ptrdiff_t stride_2x = stride << 1;
+ ptrdiff_t stride_3x = stride_2x + stride;
+ __m256i src0, src1, src2, src3, out;
+ __m256i res0, res1;
+ __m256i mask;
+ __m256i coeff_vec0 = __lasx_xvreplgr2vr_b(coeff0);
+ __m256i coeff_vec1 = __lasx_xvreplgr2vr_b(coeff1);
+ __m256i coeff_vec = __lasx_xvilvl_b(coeff_vec0, coeff_vec1);
+
+ coeff_vec = __lasx_xvslli_b(coeff_vec, 3);
+ DUP2_ARG2(__lasx_xvld, chroma_mask_arr, 0, src, 0, mask, src0);
+ DUP2_ARG2(__lasx_xvldx, src, stride, src, stride_2x, src1, src2);
+ src3 = __lasx_xvldx(src, stride_3x);
+ DUP2_ARG3(__lasx_xvpermi_q, src1, src0, 0x20, src3, src2, 0x20, src0, src2);
+ DUP2_ARG3(__lasx_xvshuf_b, src0, src0, mask, src2, src2, mask, src0, src2);
+ DUP2_ARG2(__lasx_xvdp2_h_bu, src0, coeff_vec, src2, coeff_vec, res0, res1);
+ out = __lasx_xvssrarni_bu_h(res1, res0, 6);
+ __lasx_xvstelm_d(out, dst, 0, 0);
+ __lasx_xvstelm_d(out, dst + stride, 0, 2);
+ __lasx_xvstelm_d(out, dst + stride_2x, 0, 1);
+ __lasx_xvstelm_d(out, dst + stride_3x, 0, 3);
+
+}
+
+static av_always_inline void avc_chroma_hz_8x8_lasx(uint8_t *src, uint8_t *dst,
+ ptrdiff_t stride, uint32_t coeff0, uint32_t coeff1)
+{
+ ptrdiff_t stride_2x = stride << 1;
+ ptrdiff_t stride_3x = stride_2x + stride;
+ ptrdiff_t stride_4x = stride << 2;
+ __m256i src0, src1, src2, src3, src4, src5, src6, src7;
+ __m256i out0, out1;
+ __m256i res0, res1, res2, res3;
+ __m256i mask;
+ __m256i coeff_vec0 = __lasx_xvreplgr2vr_b(coeff0);
+ __m256i coeff_vec1 = __lasx_xvreplgr2vr_b(coeff1);
+ __m256i coeff_vec = __lasx_xvilvl_b(coeff_vec0, coeff_vec1);
+
+ coeff_vec = __lasx_xvslli_b(coeff_vec, 3);
+ DUP2_ARG2(__lasx_xvld, chroma_mask_arr, 0, src, 0, mask, src0);
+ DUP4_ARG2(__lasx_xvldx, src, stride, src, stride_2x, src, stride_3x, src, stride_4x,
+ src1, src2, src3, src4);
+ src += stride_4x;
+ DUP2_ARG2(__lasx_xvldx, src, stride, src, stride_2x, src5, src6);
+ src7 = __lasx_xvldx(src, stride_3x);
+ DUP4_ARG3(__lasx_xvpermi_q, src1, src0, 0x20, src3, src2, 0x20, src5, src4, 0x20,
+ src7, src6, 0x20, src0, src2, src4, src6);
+ DUP4_ARG3(__lasx_xvshuf_b, src0, src0, mask, src2, src2, mask, src4, src4, mask,
+ src6, src6, mask, src0, src2, src4, src6);
+ DUP4_ARG2(__lasx_xvdp2_h_bu, src0, coeff_vec, src2, coeff_vec, src4, coeff_vec, src6,
+ coeff_vec, res0, res1, res2, res3);
+ DUP2_ARG3(__lasx_xvssrarni_bu_h, res1, res0, 6, res3, res2, 6, out0, out1);
+ __lasx_xvstelm_d(out0, dst, 0, 0);
+ __lasx_xvstelm_d(out0, dst + stride, 0, 2);
+ __lasx_xvstelm_d(out0, dst + stride_2x, 0, 1);
+ __lasx_xvstelm_d(out0, dst + stride_3x, 0, 3);
+ dst += stride_4x;
+ __lasx_xvstelm_d(out1, dst, 0, 0);
+ __lasx_xvstelm_d(out1, dst + stride, 0, 2);
+ __lasx_xvstelm_d(out1, dst + stride_2x, 0, 1);
+ __lasx_xvstelm_d(out1, dst + stride_3x, 0, 3);
+}
+
+static av_always_inline void avc_chroma_hz_nonmult_lasx(uint8_t *src,
+ uint8_t *dst, ptrdiff_t stride, uint32_t coeff0,
+ uint32_t coeff1, int32_t height)
+{
+ uint32_t row;
+ ptrdiff_t stride_2x = stride << 1;
+ ptrdiff_t stride_3x = stride_2x + stride;
+ ptrdiff_t stride_4x = stride << 2;
+ __m256i src0, src1, src2, src3, out;
+ __m256i res0, res1;
+ __m256i mask;
+ __m256i coeff_vec0 = __lasx_xvreplgr2vr_b(coeff0);
+ __m256i coeff_vec1 = __lasx_xvreplgr2vr_b(coeff1);
+ __m256i coeff_vec = __lasx_xvilvl_b(coeff_vec0, coeff_vec1);
+
+ mask = __lasx_xvld(chroma_mask_arr, 0);
+ coeff_vec = __lasx_xvslli_b(coeff_vec, 3);
+
+ for (row = height >> 2; row--;) {
+ DUP4_ARG2(__lasx_xvldx, src, 0, src, stride, src, stride_2x, src, stride_3x,
+ src0, src1, src2, src3);
+ src += stride_4x;
+ DUP2_ARG3(__lasx_xvpermi_q, src1, src0, 0x20, src3, src2, 0x20, src0, src2);
+ DUP2_ARG3(__lasx_xvshuf_b, src0, src0, mask, src2, src2, mask, src0, src2);
+ DUP2_ARG2(__lasx_xvdp2_h_bu, src0, coeff_vec, src2, coeff_vec, res0, res1);
+ out = __lasx_xvssrarni_bu_h(res1, res0, 6);
+ __lasx_xvstelm_d(out, dst, 0, 0);
+ __lasx_xvstelm_d(out, dst + stride, 0, 2);
+ __lasx_xvstelm_d(out, dst + stride_2x, 0, 1);
+ __lasx_xvstelm_d(out, dst + stride_3x, 0, 3);
+ dst += stride_4x;
+ }
+
+ if ((height & 3)) {
+ src0 = __lasx_xvld(src, 0);
+ src1 = __lasx_xvldx(src, stride);
+ src1 = __lasx_xvpermi_q(src1, src0, 0x20);
+ src0 = __lasx_xvshuf_b(src1, src1, mask);
+ res0 = __lasx_xvdp2_h_bu(src0, coeff_vec);
+ out = __lasx_xvssrarni_bu_h(res0, res0, 6);
+ __lasx_xvstelm_d(out, dst, 0, 0);
+ dst += stride;
+ __lasx_xvstelm_d(out, dst, 0, 2);
+ }
+}
+
+static av_always_inline void avc_chroma_vt_8x4_lasx(uint8_t *src, uint8_t *dst,
+ ptrdiff_t stride, uint32_t coeff0, uint32_t coeff1)
+{
+ ptrdiff_t stride_2x = stride << 1;
+ ptrdiff_t stride_3x = stride_2x + stride;
+ __m256i src0, src1, src2, src3, src4, out;
+ __m256i res0, res1;
+ __m256i coeff_vec0 = __lasx_xvreplgr2vr_b(coeff0);
+ __m256i coeff_vec1 = __lasx_xvreplgr2vr_b(coeff1);
+ __m256i coeff_vec = __lasx_xvilvl_b(coeff_vec0, coeff_vec1);
+
+ coeff_vec = __lasx_xvslli_b(coeff_vec, 3);
+ src0 = __lasx_xvld(src, 0);
+ src += stride;
+ DUP4_ARG2(__lasx_xvldx, src, 0, src, stride, src, stride_2x, src, stride_3x,
+ src1, src2, src3, src4);
+ DUP4_ARG3(__lasx_xvpermi_q, src1, src0, 0x20, src2, src1, 0x20, src3, src2, 0x20,
+ src4, src3, 0x20, src0, src1, src2, src3);
+ DUP2_ARG2(__lasx_xvilvl_b, src1, src0, src3, src2, src0, src2);
+ DUP2_ARG2(__lasx_xvdp2_h_bu, src0, coeff_vec, src2, coeff_vec, res0, res1);
+ out = __lasx_xvssrarni_bu_h(res1, res0, 6);
+ __lasx_xvstelm_d(out, dst, 0, 0);
+ __lasx_xvstelm_d(out, dst + stride, 0, 2);
+ __lasx_xvstelm_d(out, dst + stride_2x, 0, 1);
+ __lasx_xvstelm_d(out, dst + stride_3x, 0, 3);
+}
+
+static av_always_inline void avc_chroma_vt_8x8_lasx(uint8_t *src, uint8_t *dst,
+ ptrdiff_t stride, uint32_t coeff0, uint32_t coeff1)
+{
+ ptrdiff_t stride_2x = stride << 1;
+ ptrdiff_t stride_3x = stride_2x + stride;
+ ptrdiff_t stride_4x = stride << 2;
+ __m256i src0, src1, src2, src3, src4, src5, src6, src7, src8;
+ __m256i out0, out1;
+ __m256i res0, res1, res2, res3;
+ __m256i coeff_vec0 = __lasx_xvreplgr2vr_b(coeff0);
+ __m256i coeff_vec1 = __lasx_xvreplgr2vr_b(coeff1);
+ __m256i coeff_vec = __lasx_xvilvl_b(coeff_vec0, coeff_vec1);
+
+ coeff_vec = __lasx_xvslli_b(coeff_vec, 3);
+ src0 = __lasx_xvld(src, 0);
+ src += stride;
+ DUP4_ARG2(__lasx_xvldx, src, 0, src, stride, src, stride_2x, src, stride_3x,
+ src1, src2, src3, src4);
+ src += stride_4x;
+ DUP4_ARG2(__lasx_xvldx, src, 0, src, stride, src, stride_2x, src, stride_3x,
+ src5, src6, src7, src8);
+ DUP4_ARG3(__lasx_xvpermi_q, src1, src0, 0x20, src2, src1, 0x20, src3, src2, 0x20,
+ src4, src3, 0x20, src0, src1, src2, src3);
+ DUP4_ARG3(__lasx_xvpermi_q, src5, src4, 0x20, src6, src5, 0x20, src7, src6, 0x20,
+ src8, src7, 0x20, src4, src5, src6, src7);
+ DUP4_ARG2(__lasx_xvilvl_b, src1, src0, src3, src2, src5, src4, src7, src6,
+ src0, src2, src4, src6);
+ DUP4_ARG2(__lasx_xvdp2_h_bu, src0, coeff_vec, src2, coeff_vec, src4, coeff_vec,
+ src6, coeff_vec, res0, res1, res2, res3);
+ DUP2_ARG3(__lasx_xvssrarni_bu_h, res1, res0, 6, res3, res2, 6, out0, out1);
+ __lasx_xvstelm_d(out0, dst, 0, 0);
+ __lasx_xvstelm_d(out0, dst + stride, 0, 2);
+ __lasx_xvstelm_d(out0, dst + stride_2x, 0, 1);
+ __lasx_xvstelm_d(out0, dst + stride_3x, 0, 3);
+ dst += stride_4x;
+ __lasx_xvstelm_d(out1, dst, 0, 0);
+ __lasx_xvstelm_d(out1, dst + stride, 0, 2);
+ __lasx_xvstelm_d(out1, dst + stride_2x, 0, 1);
+ __lasx_xvstelm_d(out1, dst + stride_3x, 0, 3);
+}
+
+static av_always_inline void copy_width8x8_lasx(uint8_t *src, uint8_t *dst,
+ ptrdiff_t stride)
+{
+ uint64_t tmp[8];
+ ptrdiff_t stride_2, stride_3, stride_4;
+ __asm__ volatile (
+ "slli.d %[stride_2], %[stride], 1 \n\t"
+ "add.d %[stride_3], %[stride_2], %[stride] \n\t"
+ "slli.d %[stride_4], %[stride_2], 1 \n\t"
+ "ld.d %[tmp0], %[src], 0x0 \n\t"
+ "ldx.d %[tmp1], %[src], %[stride] \n\t"
+ "ldx.d %[tmp2], %[src], %[stride_2] \n\t"
+ "ldx.d %[tmp3], %[src], %[stride_3] \n\t"
+ "add.d %[src], %[src], %[stride_4] \n\t"
+ "ld.d %[tmp4], %[src], 0x0 \n\t"
+ "ldx.d %[tmp5], %[src], %[stride] \n\t"
+ "ldx.d %[tmp6], %[src], %[stride_2] \n\t"
+ "ldx.d %[tmp7], %[src], %[stride_3] \n\t"
+
+ "st.d %[tmp0], %[dst], 0x0 \n\t"
+ "stx.d %[tmp1], %[dst], %[stride] \n\t"
+ "stx.d %[tmp2], %[dst], %[stride_2] \n\t"
+ "stx.d %[tmp3], %[dst], %[stride_3] \n\t"
+ "add.d %[dst], %[dst], %[stride_4] \n\t"
+ "st.d %[tmp4], %[dst], 0x0 \n\t"
+ "stx.d %[tmp5], %[dst], %[stride] \n\t"
+ "stx.d %[tmp6], %[dst], %[stride_2] \n\t"
+ "stx.d %[tmp7], %[dst], %[stride_3] \n\t"
+ : [tmp0]"=&r"(tmp[0]), [tmp1]"=&r"(tmp[1]),
+ [tmp2]"=&r"(tmp[2]), [tmp3]"=&r"(tmp[3]),
+ [tmp4]"=&r"(tmp[4]), [tmp5]"=&r"(tmp[5]),
+ [tmp6]"=&r"(tmp[6]), [tmp7]"=&r"(tmp[7]),
+ [dst]"+&r"(dst), [src]"+&r"(src),
+ [stride_2]"=&r"(stride_2), [stride_3]"=&r"(stride_3),
+ [stride_4]"=&r"(stride_4)
+ : [stride]"r"(stride)
+ : "memory"
+ );
+}
+
+static av_always_inline void copy_width8x4_lasx(uint8_t *src, uint8_t *dst,
+ ptrdiff_t stride)
+{
+ uint64_t tmp[4];
+ ptrdiff_t stride_2, stride_3;
+ __asm__ volatile (
+ "slli.d %[stride_2], %[stride], 1 \n\t"
+ "add.d %[stride_3], %[stride_2], %[stride] \n\t"
+ "ld.d %[tmp0], %[src], 0x0 \n\t"
+ "ldx.d %[tmp1], %[src], %[stride] \n\t"
+ "ldx.d %[tmp2], %[src], %[stride_2] \n\t"
+ "ldx.d %[tmp3], %[src], %[stride_3] \n\t"
+
+ "st.d %[tmp0], %[dst], 0x0 \n\t"
+ "stx.d %[tmp1], %[dst], %[stride] \n\t"
+ "stx.d %[tmp2], %[dst], %[stride_2] \n\t"
+ "stx.d %[tmp3], %[dst], %[stride_3] \n\t"
+ : [tmp0]"=&r"(tmp[0]), [tmp1]"=&r"(tmp[1]),
+ [tmp2]"=&r"(tmp[2]), [tmp3]"=&r"(tmp[3]),
+ [stride_2]"=&r"(stride_2), [stride_3]"=&r"(stride_3)
+ : [stride]"r"(stride), [dst]"r"(dst), [src]"r"(src)
+ : "memory"
+ );
+}
+
+static void avc_chroma_hv_8w_lasx(uint8_t *src, uint8_t *dst, ptrdiff_t stride,
+ uint32_t coef_hor0, uint32_t coef_hor1,
+ uint32_t coef_ver0, uint32_t coef_ver1,
+ int32_t height)
+{
+ if (4 == height) {
+ avc_chroma_hv_8x4_lasx(src, dst, stride, coef_hor0, coef_hor1, coef_ver0,
+ coef_ver1);
+ } else if (8 == height) {
+ avc_chroma_hv_8x8_lasx(src, dst, stride, coef_hor0, coef_hor1, coef_ver0,
+ coef_ver1);
+ }
+}
+
+static void avc_chroma_hv_4x2_lasx(uint8_t *src, uint8_t *dst, ptrdiff_t stride,
+ uint32_t coef_hor0, uint32_t coef_hor1,
+ uint32_t coef_ver0, uint32_t coef_ver1)
+{
+ ptrdiff_t stride_2 = stride << 1;
+ __m256i src0, src1, src2;
+ __m256i res_hz, res_vt;
+ __m256i mask;
+ __m256i coeff_hz_vec0 = __lasx_xvreplgr2vr_b(coef_hor0);
+ __m256i coeff_hz_vec1 = __lasx_xvreplgr2vr_b(coef_hor1);
+ __m256i coeff_hz_vec = __lasx_xvilvl_b(coeff_hz_vec0, coeff_hz_vec1);
+ __m256i coeff_vt_vec0 = __lasx_xvreplgr2vr_h(coef_ver0);
+ __m256i coeff_vt_vec1 = __lasx_xvreplgr2vr_h(coef_ver1);
+ __m256i coeff_vt_vec = __lasx_xvpermi_q(coeff_vt_vec1, coeff_vt_vec0, 0x02);
+
+ DUP2_ARG2(__lasx_xvld, chroma_mask_arr, 32, src, 0, mask, src0);
+ DUP2_ARG2(__lasx_xvldx, src, stride, src, stride_2, src1, src2);
+ DUP2_ARG3(__lasx_xvshuf_b, src1, src0, mask, src2, src1, mask, src0, src1);
+ src0 = __lasx_xvpermi_q(src0, src1, 0x02);
+ res_hz = __lasx_xvdp2_h_bu(src0, coeff_hz_vec);
+ res_vt = __lasx_xvmul_h(res_hz, coeff_vt_vec);
+ res_hz = __lasx_xvpermi_q(res_hz, res_vt, 0x01);
+ res_vt = __lasx_xvadd_h(res_hz, res_vt);
+ res_vt = __lasx_xvssrarni_bu_h(res_vt, res_vt, 6);
+ __lasx_xvstelm_w(res_vt, dst, 0, 0);
+ __lasx_xvstelm_w(res_vt, dst + stride, 0, 1);
+}
+
+static void avc_chroma_hv_4x4_lasx(uint8_t *src, uint8_t *dst, ptrdiff_t stride,
+ uint32_t coef_hor0, uint32_t coef_hor1,
+ uint32_t coef_ver0, uint32_t coef_ver1)
+{
+ ptrdiff_t stride_2 = stride << 1;
+ ptrdiff_t stride_3 = stride_2 + stride;
+ ptrdiff_t stride_4 = stride_2 << 1;
+ __m256i src0, src1, src2, src3, src4;
+ __m256i res_hz0, res_hz1, res_vt0, res_vt1;
+ __m256i mask;
+ __m256i coeff_hz_vec0 = __lasx_xvreplgr2vr_b(coef_hor0);
+ __m256i coeff_hz_vec1 = __lasx_xvreplgr2vr_b(coef_hor1);
+ __m256i coeff_hz_vec = __lasx_xvilvl_b(coeff_hz_vec0, coeff_hz_vec1);
+ __m256i coeff_vt_vec0 = __lasx_xvreplgr2vr_h(coef_ver0);
+ __m256i coeff_vt_vec1 = __lasx_xvreplgr2vr_h(coef_ver1);
+
+ DUP2_ARG2(__lasx_xvld, chroma_mask_arr, 32, src, 0, mask, src0);
+ DUP4_ARG2(__lasx_xvldx, src, stride, src, stride_2, src, stride_3,
+ src, stride_4, src1, src2, src3, src4);
+ DUP4_ARG3(__lasx_xvshuf_b, src1, src0, mask, src2, src1, mask, src3, src2, mask,
+ src4, src3, mask, src0, src1, src2, src3);
+ DUP2_ARG3(__lasx_xvpermi_q, src0, src2, 0x02, src1, src3, 0x02, src0, src1);
+ DUP2_ARG2(__lasx_xvdp2_h_bu, src0, coeff_hz_vec, src1, coeff_hz_vec, res_hz0, res_hz1);
+ DUP2_ARG2(__lasx_xvmul_h, res_hz0, coeff_vt_vec1, res_hz1, coeff_vt_vec0, res_vt0, res_vt1);
+ res_hz0 = __lasx_xvadd_h(res_vt0, res_vt1);
+ res_hz0 = __lasx_xvssrarni_bu_h(res_hz0, res_hz0, 6);
+ __lasx_xvstelm_w(res_hz0, dst, 0, 0);
+ __lasx_xvstelm_w(res_hz0, dst + stride, 0, 1);
+ __lasx_xvstelm_w(res_hz0, dst + stride_2, 0, 4);
+ __lasx_xvstelm_w(res_hz0, dst + stride_3, 0, 5);
+}
+
+static void avc_chroma_hv_4x8_lasx(uint8_t *src, uint8_t * dst, ptrdiff_t stride,
+ uint32_t coef_hor0, uint32_t coef_hor1,
+ uint32_t coef_ver0, uint32_t coef_ver1)
+{
+ ptrdiff_t stride_2 = stride << 1;
+ ptrdiff_t stride_3 = stride_2 + stride;
+ ptrdiff_t stride_4 = stride_2 << 1;
+ __m256i src0, src1, src2, src3, src4, src5, src6, src7, src8;
+ __m256i res_hz0, res_hz1, res_hz2, res_hz3;
+ __m256i res_vt0, res_vt1, res_vt2, res_vt3;
+ __m256i mask;
+ __m256i coeff_hz_vec0 = __lasx_xvreplgr2vr_b(coef_hor0);
+ __m256i coeff_hz_vec1 = __lasx_xvreplgr2vr_b(coef_hor1);
+ __m256i coeff_hz_vec = __lasx_xvilvl_b(coeff_hz_vec0, coeff_hz_vec1);
+ __m256i coeff_vt_vec0 = __lasx_xvreplgr2vr_h(coef_ver0);
+ __m256i coeff_vt_vec1 = __lasx_xvreplgr2vr_h(coef_ver1);
+
+ DUP2_ARG2(__lasx_xvld, chroma_mask_arr, 32, src, 0, mask, src0);
+ DUP4_ARG2(__lasx_xvldx, src, stride, src, stride_2, src, stride_3,
+ src, stride_4, src1, src2, src3, src4);
+ src += stride_4;
+ DUP4_ARG2(__lasx_xvldx, src, stride, src, stride_2, src, stride_3,
+ src, stride_4, src5, src6, src7, src8);
+ DUP4_ARG3(__lasx_xvshuf_b, src1, src0, mask, src2, src1, mask, src3, src2, mask,
+ src4, src3, mask, src0, src1, src2, src3);
+ DUP4_ARG3(__lasx_xvshuf_b, src5, src4, mask, src6, src5, mask, src7, src6, mask,
+ src8, src7, mask, src4, src5, src6, src7);
+ DUP4_ARG3(__lasx_xvpermi_q, src0, src2, 0x02, src1, src3, 0x02, src4, src6, 0x02,
+ src5, src7, 0x02, src0, src1, src4, src5);
+ DUP4_ARG2(__lasx_xvdp2_h_bu, src0, coeff_hz_vec, src1, coeff_hz_vec, src4, coeff_hz_vec,
+ src5, coeff_hz_vec, res_hz0, res_hz1, res_hz2, res_hz3);
+ DUP4_ARG2(__lasx_xvmul_h, res_hz0, coeff_vt_vec1, res_hz1, coeff_vt_vec0, res_hz2,
+ coeff_vt_vec1, res_hz3, coeff_vt_vec0, res_vt0, res_vt1, res_vt2, res_vt3);
+ DUP2_ARG2(__lasx_xvadd_h, res_vt0, res_vt1, res_vt2, res_vt3, res_vt0, res_vt2);
+ res_hz0 = __lasx_xvssrarni_bu_h(res_vt2, res_vt0, 6);
+ __lasx_xvstelm_w(res_hz0, dst, 0, 0);
+ __lasx_xvstelm_w(res_hz0, dst + stride, 0, 1);
+ __lasx_xvstelm_w(res_hz0, dst + stride_2, 0, 4);
+ __lasx_xvstelm_w(res_hz0, dst + stride_3, 0, 5);
+ dst += stride_4;
+ __lasx_xvstelm_w(res_hz0, dst, 0, 2);
+ __lasx_xvstelm_w(res_hz0, dst + stride, 0, 3);
+ __lasx_xvstelm_w(res_hz0, dst + stride_2, 0, 6);
+ __lasx_xvstelm_w(res_hz0, dst + stride_3, 0, 7);
+}
+
+static void avc_chroma_hv_4w_lasx(uint8_t *src, uint8_t *dst, ptrdiff_t stride,
+ uint32_t coef_hor0, uint32_t coef_hor1,
+ uint32_t coef_ver0, uint32_t coef_ver1,
+ int32_t height)
+{
+ if (8 == height) {
+ avc_chroma_hv_4x8_lasx(src, dst, stride, coef_hor0, coef_hor1, coef_ver0,
+ coef_ver1);
+ } else if (4 == height) {
+ avc_chroma_hv_4x4_lasx(src, dst, stride, coef_hor0, coef_hor1, coef_ver0,
+ coef_ver1);
+ } else if (2 == height) {
+ avc_chroma_hv_4x2_lasx(src, dst, stride, coef_hor0, coef_hor1, coef_ver0,
+ coef_ver1);
+ }
+}
+
+static void avc_chroma_hz_4x2_lasx(uint8_t *src, uint8_t *dst, ptrdiff_t stride,
+ uint32_t coeff0, uint32_t coeff1)
+{
+ __m256i src0, src1;
+ __m256i res, mask;
+ __m256i coeff_vec0 = __lasx_xvreplgr2vr_b(coeff0);
+ __m256i coeff_vec1 = __lasx_xvreplgr2vr_b(coeff1);
+ __m256i coeff_vec = __lasx_xvilvl_b(coeff_vec0, coeff_vec1);
+
+ DUP2_ARG2(__lasx_xvld, chroma_mask_arr, 32, src, 0, mask, src0);
+ src1 = __lasx_xvldx(src, stride);
+ src0 = __lasx_xvshuf_b(src1, src0, mask);
+ res = __lasx_xvdp2_h_bu(src0, coeff_vec);
+ res = __lasx_xvslli_h(res, 3);
+ res = __lasx_xvssrarni_bu_h(res, res, 6);
+ __lasx_xvstelm_w(res, dst, 0, 0);
+ __lasx_xvstelm_w(res, dst + stride, 0, 1);
+}
+
+static void avc_chroma_hz_4x4_lasx(uint8_t *src, uint8_t *dst, ptrdiff_t stride,
+ uint32_t coeff0, uint32_t coeff1)
+{
+ ptrdiff_t stride_2 = stride << 1;
+ ptrdiff_t stride_3 = stride_2 + stride;
+ __m256i src0, src1, src2, src3;
+ __m256i res, mask;
+ __m256i coeff_vec0 = __lasx_xvreplgr2vr_b(coeff0);
+ __m256i coeff_vec1 = __lasx_xvreplgr2vr_b(coeff1);
+ __m256i coeff_vec = __lasx_xvilvl_b(coeff_vec0, coeff_vec1);
+
+ DUP2_ARG2(__lasx_xvld, chroma_mask_arr, 32, src, 0, mask, src0);
+ DUP2_ARG2(__lasx_xvldx, src, stride, src, stride_2, src1, src2);
+ src3 = __lasx_xvldx(src, stride_3);
+ DUP2_ARG3(__lasx_xvshuf_b, src1, src0, mask, src3, src2, mask, src0, src2);
+ src0 = __lasx_xvpermi_q(src0, src2, 0x02);
+ res = __lasx_xvdp2_h_bu(src0, coeff_vec);
+ res = __lasx_xvslli_h(res, 3);
+ res = __lasx_xvssrarni_bu_h(res, res, 6);
+ __lasx_xvstelm_w(res, dst, 0, 0);
+ __lasx_xvstelm_w(res, dst + stride, 0, 1);
+ __lasx_xvstelm_w(res, dst + stride_2, 0, 4);
+ __lasx_xvstelm_w(res, dst + stride_3, 0, 5);
+}
+
+static void avc_chroma_hz_4x8_lasx(uint8_t *src, uint8_t *dst, ptrdiff_t stride,
+ uint32_t coeff0, uint32_t coeff1)
+{
+ ptrdiff_t stride_2 = stride << 1;
+ ptrdiff_t stride_3 = stride_2 + stride;
+ ptrdiff_t stride_4 = stride_2 << 1;
+ __m256i src0, src1, src2, src3, src4, src5, src6, src7;
+ __m256i res0, res1, mask;
+ __m256i coeff_vec0 = __lasx_xvreplgr2vr_b(coeff0);
+ __m256i coeff_vec1 = __lasx_xvreplgr2vr_b(coeff1);
+ __m256i coeff_vec = __lasx_xvilvl_b(coeff_vec0, coeff_vec1);
+
+ coeff_vec = __lasx_xvslli_b(coeff_vec, 3);
+ DUP2_ARG2(__lasx_xvld, chroma_mask_arr, 32, src, 0, mask, src0);
+ DUP4_ARG2(__lasx_xvldx, src, stride, src, stride_2, src, stride_3,
+ src, stride_4, src1, src2, src3, src4);
+ src += stride_4;
+ DUP2_ARG2(__lasx_xvldx, src, stride, src, stride_2, src5, src6);
+ src7 = __lasx_xvldx(src, stride_3);
+ DUP4_ARG3(__lasx_xvshuf_b, src1, src0, mask, src3, src2, mask, src5, src4, mask,
+ src7, src6, mask, src0, src2, src4, src6);
+ DUP2_ARG3(__lasx_xvpermi_q, src0, src2, 0x02, src4, src6, 0x02, src0, src4);
+ DUP2_ARG2(__lasx_xvdp2_h_bu, src0, coeff_vec, src4, coeff_vec, res0, res1);
+ res0 = __lasx_xvssrarni_bu_h(res1, res0, 6);
+ __lasx_xvstelm_w(res0, dst, 0, 0);
+ __lasx_xvstelm_w(res0, dst + stride, 0, 1);
+ __lasx_xvstelm_w(res0, dst + stride_2, 0, 4);
+ __lasx_xvstelm_w(res0, dst + stride_3, 0, 5);
+ dst += stride_4;
+ __lasx_xvstelm_w(res0, dst, 0, 2);
+ __lasx_xvstelm_w(res0, dst + stride, 0, 3);
+ __lasx_xvstelm_w(res0, dst + stride_2, 0, 6);
+ __lasx_xvstelm_w(res0, dst + stride_3, 0, 7);
+}
+
+static void avc_chroma_hz_4w_lasx(uint8_t *src, uint8_t *dst, ptrdiff_t stride,
+ uint32_t coeff0, uint32_t coeff1,
+ int32_t height)
+{
+ if (8 == height) {
+ avc_chroma_hz_4x8_lasx(src, dst, stride, coeff0, coeff1);
+ } else if (4 == height) {
+ avc_chroma_hz_4x4_lasx(src, dst, stride, coeff0, coeff1);
+ } else if (2 == height) {
+ avc_chroma_hz_4x2_lasx(src, dst, stride, coeff0, coeff1);
+ }
+}
+
+static void avc_chroma_hz_8w_lasx(uint8_t *src, uint8_t *dst, ptrdiff_t stride,
+ uint32_t coeff0, uint32_t coeff1,
+ int32_t height)
+{
+ if (4 == height) {
+ avc_chroma_hz_8x4_lasx(src, dst, stride, coeff0, coeff1);
+ } else if (8 == height) {
+ avc_chroma_hz_8x8_lasx(src, dst, stride, coeff0, coeff1);
+ } else {
+ avc_chroma_hz_nonmult_lasx(src, dst, stride, coeff0, coeff1, height);
+ }
+}
+
+static void avc_chroma_vt_4x2_lasx(uint8_t *src, uint8_t *dst, ptrdiff_t stride,
+ uint32_t coeff0, uint32_t coeff1)
+{
+ __m256i src0, src1, src2;
+ __m256i tmp0, tmp1;
+ __m256i res;
+ __m256i coeff_vec0 = __lasx_xvreplgr2vr_b(coeff0);
+ __m256i coeff_vec1 = __lasx_xvreplgr2vr_b(coeff1);
+ __m256i coeff_vec = __lasx_xvilvl_b(coeff_vec0, coeff_vec1);
+
+ src0 = __lasx_xvld(src, 0);
+ DUP2_ARG2(__lasx_xvldx, src, stride, src, stride << 1, src1, src2);
+ DUP2_ARG2(__lasx_xvilvl_b, src1, src0, src2, src1, tmp0, tmp1);
+ tmp0 = __lasx_xvilvl_d(tmp1, tmp0);
+ res = __lasx_xvdp2_h_bu(tmp0, coeff_vec);
+ res = __lasx_xvslli_h(res, 3);
+ res = __lasx_xvssrarni_bu_h(res, res, 6);
+ __lasx_xvstelm_w(res, dst, 0, 0);
+ __lasx_xvstelm_w(res, dst + stride, 0, 1);
+}
+
+static void avc_chroma_vt_4x4_lasx(uint8_t *src, uint8_t *dst, ptrdiff_t stride,
+ uint32_t coeff0, uint32_t coeff1)
+{
+ ptrdiff_t stride_2 = stride << 1;
+ ptrdiff_t stride_3 = stride_2 + stride;
+ ptrdiff_t stride_4 = stride_2 << 1;
+ __m256i src0, src1, src2, src3, src4;
+ __m256i tmp0, tmp1, tmp2, tmp3;
+ __m256i res;
+ __m256i coeff_vec0 = __lasx_xvreplgr2vr_b(coeff0);
+ __m256i coeff_vec1 = __lasx_xvreplgr2vr_b(coeff1);
+ __m256i coeff_vec = __lasx_xvilvl_b(coeff_vec0, coeff_vec1);
+
+ src0 = __lasx_xvld(src, 0);
+ DUP4_ARG2(__lasx_xvldx, src, stride, src, stride_2, src, stride_3,
+ src, stride_4, src1, src2, src3, src4);
+ DUP4_ARG2(__lasx_xvilvl_b, src1, src0, src2, src1, src3, src2, src4, src3,
+ tmp0, tmp1, tmp2, tmp3);
+ DUP2_ARG2(__lasx_xvilvl_d, tmp1, tmp0, tmp3, tmp2, tmp0, tmp2);
+ tmp0 = __lasx_xvpermi_q(tmp0, tmp2, 0x02);
+ res = __lasx_xvdp2_h_bu(tmp0, coeff_vec);
+ res = __lasx_xvslli_h(res, 3);
+ res = __lasx_xvssrarni_bu_h(res, res, 6);
+ __lasx_xvstelm_w(res, dst, 0, 0);
+ __lasx_xvstelm_w(res, dst + stride, 0, 1);
+ __lasx_xvstelm_w(res, dst + stride_2, 0, 4);
+ __lasx_xvstelm_w(res, dst + stride_3, 0, 5);
+}
+
+static void avc_chroma_vt_4x8_lasx(uint8_t *src, uint8_t *dst, ptrdiff_t stride,
+ uint32_t coeff0, uint32_t coeff1)
+{
+ ptrdiff_t stride_2 = stride << 1;
+ ptrdiff_t stride_3 = stride_2 + stride;
+ ptrdiff_t stride_4 = stride_2 << 1;
+ __m256i src0, src1, src2, src3, src4, src5, src6, src7, src8;
+ __m256i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+ __m256i res0, res1;
+ __m256i coeff_vec0 = __lasx_xvreplgr2vr_b(coeff0);
+ __m256i coeff_vec1 = __lasx_xvreplgr2vr_b(coeff1);
+ __m256i coeff_vec = __lasx_xvilvl_b(coeff_vec0, coeff_vec1);
+
+ coeff_vec = __lasx_xvslli_b(coeff_vec, 3);
+ src0 = __lasx_xvld(src, 0);
+ DUP4_ARG2(__lasx_xvldx, src, stride, src, stride_2, src, stride_3,
+ src, stride_4, src1, src2, src3, src4);
+ src += stride_4;
+ DUP4_ARG2(__lasx_xvldx, src, stride, src, stride_2, src, stride_3,
+ src, stride_4, src5, src6, src7, src8);
+ DUP4_ARG2(__lasx_xvilvl_b, src1, src0, src2, src1, src3, src2, src4, src3,
+ tmp0, tmp1, tmp2, tmp3);
+ DUP4_ARG2(__lasx_xvilvl_b, src5, src4, src6, src5, src7, src6, src8, src7,
+ tmp4, tmp5, tmp6, tmp7);
+ DUP4_ARG2(__lasx_xvilvl_d, tmp1, tmp0, tmp3, tmp2, tmp5, tmp4, tmp7, tmp6,
+ tmp0, tmp2, tmp4, tmp6);
+ tmp0 = __lasx_xvpermi_q(tmp0, tmp2, 0x02);
+ tmp4 = __lasx_xvpermi_q(tmp4, tmp6, 0x02);
+ DUP2_ARG2(__lasx_xvdp2_h_bu, tmp0, coeff_vec, tmp4, coeff_vec, res0, res1);
+ res0 = __lasx_xvssrarni_bu_h(res1, res0, 6);
+ __lasx_xvstelm_w(res0, dst, 0, 0);
+ __lasx_xvstelm_w(res0, dst + stride, 0, 1);
+ __lasx_xvstelm_w(res0, dst + stride_2, 0, 4);
+ __lasx_xvstelm_w(res0, dst + stride_3, 0, 5);
+ dst += stride_4;
+ __lasx_xvstelm_w(res0, dst, 0, 2);
+ __lasx_xvstelm_w(res0, dst + stride, 0, 3);
+ __lasx_xvstelm_w(res0, dst + stride_2, 0, 6);
+ __lasx_xvstelm_w(res0, dst + stride_3, 0, 7);
+}
+
+static void avc_chroma_vt_4w_lasx(uint8_t *src, uint8_t *dst, ptrdiff_t stride,
+ uint32_t coeff0, uint32_t coeff1,
+ int32_t height)
+{
+ if (8 == height) {
+ avc_chroma_vt_4x8_lasx(src, dst, stride, coeff0, coeff1);
+ } else if (4 == height) {
+ avc_chroma_vt_4x4_lasx(src, dst, stride, coeff0, coeff1);
+ } else if (2 == height) {
+ avc_chroma_vt_4x2_lasx(src, dst, stride, coeff0, coeff1);
+ }
+}
+
+static void avc_chroma_vt_8w_lasx(uint8_t *src, uint8_t *dst, ptrdiff_t stride,
+ uint32_t coeff0, uint32_t coeff1,
+ int32_t height)
+{
+ if (4 == height) {
+ avc_chroma_vt_8x4_lasx(src, dst, stride, coeff0, coeff1);
+ } else if (8 == height) {
+ avc_chroma_vt_8x8_lasx(src, dst, stride, coeff0, coeff1);
+ }
+}
+
+static void copy_width4_lasx(uint8_t *src, uint8_t *dst, ptrdiff_t stride,
+ int32_t height)
+{
+ uint32_t tp0, tp1, tp2, tp3, tp4, tp5, tp6, tp7;
+
+ if (8 == height) {
+ ptrdiff_t stride_2, stride_3, stride_4;
+
+ __asm__ volatile (
+ "slli.d %[stride_2], %[stride], 1 \n\t"
+ "add.d %[stride_3], %[stride_2], %[stride] \n\t"
+ "slli.d %[stride_4], %[stride_2], 1 \n\t"
+ "ld.wu %[tp0], %[src], 0 \n\t"
+ "ldx.wu %[tp1], %[src], %[stride] \n\t"
+ "ldx.wu %[tp2], %[src], %[stride_2] \n\t"
+ "ldx.wu %[tp3], %[src], %[stride_3] \n\t"
+ "add.d %[src], %[src], %[stride_4] \n\t"
+ "ld.wu %[tp4], %[src], 0 \n\t"
+ "ldx.wu %[tp5], %[src], %[stride] \n\t"
+ "ldx.wu %[tp6], %[src], %[stride_2] \n\t"
+ "ldx.wu %[tp7], %[src], %[stride_3] \n\t"
+ "st.w %[tp0], %[dst], 0 \n\t"
+ "stx.w %[tp1], %[dst], %[stride] \n\t"
+ "stx.w %[tp2], %[dst], %[stride_2] \n\t"
+ "stx.w %[tp3], %[dst], %[stride_3] \n\t"
+ "add.d %[dst], %[dst], %[stride_4] \n\t"
+ "st.w %[tp4], %[dst], 0 \n\t"
+ "stx.w %[tp5], %[dst], %[stride] \n\t"
+ "stx.w %[tp6], %[dst], %[stride_2] \n\t"
+ "stx.w %[tp7], %[dst], %[stride_3] \n\t"
+ : [stride_2]"+&r"(stride_2), [stride_3]"+&r"(stride_3), [stride_4]"+&r"(stride_4),
+ [src]"+&r"(src), [dst]"+&r"(dst), [tp0]"+&r"(tp0), [tp1]"+&r"(tp1),
+ [tp2]"+&r"(tp2), [tp3]"+&r"(tp3), [tp4]"+&r"(tp4), [tp5]"+&r"(tp5),
+ [tp6]"+&r"(tp6), [tp7]"+&r"(tp7)
+ : [stride]"r"(stride)
+ : "memory"
+ );
+ } else if (4 == height) {
+ ptrdiff_t stride_2, stride_3;
+
+ __asm__ volatile (
+ "slli.d %[stride_2], %[stride], 1 \n\t"
+ "add.d %[stride_3], %[stride_2], %[stride] \n\t"
+ "ld.wu %[tp0], %[src], 0 \n\t"
+ "ldx.wu %[tp1], %[src], %[stride] \n\t"
+ "ldx.wu %[tp2], %[src], %[stride_2] \n\t"
+ "ldx.wu %[tp3], %[src], %[stride_3] \n\t"
+ "st.w %[tp0], %[dst], 0 \n\t"
+ "stx.w %[tp1], %[dst], %[stride] \n\t"
+ "stx.w %[tp2], %[dst], %[stride_2] \n\t"
+ "stx.w %[tp3], %[dst], %[stride_3] \n\t"
+ : [stride_2]"+&r"(stride_2), [stride_3]"+&r"(stride_3),
+ [src]"+&r"(src), [dst]"+&r"(dst), [tp0]"+&r"(tp0), [tp1]"+&r"(tp1),
+ [tp2]"+&r"(tp2), [tp3]"+&r"(tp3)
+ : [stride]"r"(stride)
+ : "memory"
+ );
+ } else if (2 == height) {
+ __asm__ volatile (
+ "ld.wu %[tp0], %[src], 0 \n\t"
+ "ldx.wu %[tp1], %[src], %[stride] \n\t"
+ "st.w %[tp0], %[dst], 0 \n\t"
+ "stx.w %[tp1], %[dst], %[stride] \n\t"
+ : [tp0]"+&r"(tp0), [tp1]"+&r"(tp1)
+ : [src]"r"(src), [dst]"r"(dst), [stride]"r"(stride)
+ : "memory"
+ );
+ }
+}
+
+static void copy_width8_lasx(uint8_t *src, uint8_t *dst, ptrdiff_t stride,
+ int32_t height)
+{
+ if (8 == height) {
+ copy_width8x8_lasx(src, dst, stride);
+ } else if (4 == height) {
+ copy_width8x4_lasx(src, dst, stride);
+ }
+}
+
+void ff_put_h264_chroma_mc4_lasx(uint8_t *dst, uint8_t *src, ptrdiff_t stride,
+ int height, int x, int y)
+{
+ av_assert2(x < 8 && y < 8 && x >= 0 && y >= 0);
+
+ if(x && y) {
+ avc_chroma_hv_4w_lasx(src, dst, stride, x, (8 - x), y, (8 - y), height);
+ } else if (x) {
+ avc_chroma_hz_4w_lasx(src, dst, stride, x, (8 - x), height);
+ } else if (y) {
+ avc_chroma_vt_4w_lasx(src, dst, stride, y, (8 - y), height);
+ } else {
+ copy_width4_lasx(src, dst, stride, height);
+ }
+}
+
+void ff_put_h264_chroma_mc8_lasx(uint8_t *dst, uint8_t *src, ptrdiff_t stride,
+ int height, int x, int y)
+{
+ av_assert2(x < 8 && y < 8 && x >= 0 && y >= 0);
+
+ if (!(x || y)) {
+ copy_width8_lasx(src, dst, stride, height);
+ } else if (x && y) {
+ avc_chroma_hv_8w_lasx(src, dst, stride, x, (8 - x), y, (8 - y), height);
+ } else if (x) {
+ avc_chroma_hz_8w_lasx(src, dst, stride, x, (8 - x), height);
+ } else {
+ avc_chroma_vt_8w_lasx(src, dst, stride, y, (8 - y), height);
+ }
+}
+
+static av_always_inline void avc_chroma_hv_and_aver_dst_8x4_lasx(uint8_t *src,
+ uint8_t *dst, ptrdiff_t stride, uint32_t coef_hor0,
+ uint32_t coef_hor1, uint32_t coef_ver0,
+ uint32_t coef_ver1)
+{
+ ptrdiff_t stride_2x = stride << 1;
+ ptrdiff_t stride_3x = stride_2x + stride;
+ ptrdiff_t stride_4x = stride << 2;
+ __m256i tp0, tp1, tp2, tp3;
+ __m256i src0, src1, src2, src3, src4, out;
+ __m256i res_hz0, res_hz1, res_hz2, res_vt0, res_vt1;
+ __m256i mask;
+ __m256i coeff_hz_vec0 = __lasx_xvreplgr2vr_b(coef_hor0);
+ __m256i coeff_hz_vec1 = __lasx_xvreplgr2vr_b(coef_hor1);
+ __m256i coeff_hz_vec = __lasx_xvilvl_b(coeff_hz_vec0, coeff_hz_vec1);
+ __m256i coeff_vt_vec0 = __lasx_xvreplgr2vr_h(coef_ver0);
+ __m256i coeff_vt_vec1 = __lasx_xvreplgr2vr_h(coef_ver1);
+
+ DUP2_ARG2(__lasx_xvld, chroma_mask_arr, 0, src, 0, mask, src0);
+ DUP4_ARG2(__lasx_xvldx, src, stride, src, stride_2x, src, stride_3x, src, stride_4x,
+ src1, src2, src3, src4);
+ DUP2_ARG3(__lasx_xvpermi_q, src2, src1, 0x20, src4, src3, 0x20, src1, src3);
+ src0 = __lasx_xvshuf_b(src0, src0, mask);
+ DUP2_ARG3(__lasx_xvshuf_b, src1, src1, mask, src3, src3, mask, src1, src3);
+ DUP2_ARG2(__lasx_xvdp2_h_bu, src0, coeff_hz_vec, src1, coeff_hz_vec, res_hz0, res_hz1);
+ res_hz2 = __lasx_xvdp2_h_bu(src3, coeff_hz_vec);
+ res_vt0 = __lasx_xvmul_h(res_hz1, coeff_vt_vec0);
+ res_vt1 = __lasx_xvmul_h(res_hz2, coeff_vt_vec0);
+ res_hz0 = __lasx_xvpermi_q(res_hz1, res_hz0, 0x20);
+ res_hz1 = __lasx_xvpermi_q(res_hz1, res_hz2, 0x3);
+ res_vt0 = __lasx_xvmadd_h(res_vt0, res_hz0, coeff_vt_vec1);
+ res_vt1 = __lasx_xvmadd_h(res_vt1, res_hz1, coeff_vt_vec1);
+ out = __lasx_xvssrarni_bu_h(res_vt1, res_vt0, 6);
+ DUP4_ARG2(__lasx_xvldx, dst, 0, dst, stride, dst, stride_2x, dst, stride_3x,
+ tp0, tp1, tp2, tp3);
+ DUP2_ARG2(__lasx_xvilvl_d, tp2, tp0, tp3, tp1, tp0, tp2);
+ tp0 = __lasx_xvpermi_q(tp2, tp0, 0x20);
+ out = __lasx_xvavgr_bu(out, tp0);
+ __lasx_xvstelm_d(out, dst, 0, 0);
+ __lasx_xvstelm_d(out, dst + stride, 0, 2);
+ __lasx_xvstelm_d(out, dst + stride_2x, 0, 1);
+ __lasx_xvstelm_d(out, dst + stride_3x, 0, 3);
+}
+
+static av_always_inline void avc_chroma_hv_and_aver_dst_8x8_lasx(uint8_t *src,
+ uint8_t *dst, ptrdiff_t stride, uint32_t coef_hor0,
+ uint32_t coef_hor1, uint32_t coef_ver0,
+ uint32_t coef_ver1)
+{
+ ptrdiff_t stride_2x = stride << 1;
+ ptrdiff_t stride_3x = stride_2x + stride;
+ ptrdiff_t stride_4x = stride << 2;
+ __m256i tp0, tp1, tp2, tp3, dst0, dst1;
+ __m256i src0, src1, src2, src3, src4, src5, src6, src7, src8;
+ __m256i out0, out1;
+ __m256i res_hz0, res_hz1, res_hz2, res_hz3, res_hz4;
+ __m256i res_vt0, res_vt1, res_vt2, res_vt3;
+ __m256i mask;
+ __m256i coeff_hz_vec0 = __lasx_xvreplgr2vr_b(coef_hor0);
+ __m256i coeff_hz_vec1 = __lasx_xvreplgr2vr_b(coef_hor1);
+ __m256i coeff_vt_vec0 = __lasx_xvreplgr2vr_h(coef_ver0);
+ __m256i coeff_vt_vec1 = __lasx_xvreplgr2vr_h(coef_ver1);
+ __m256i coeff_hz_vec = __lasx_xvilvl_b(coeff_hz_vec0, coeff_hz_vec1);
+
+ DUP2_ARG2(__lasx_xvld, chroma_mask_arr, 0, src, 0, mask, src0);
+ src += stride;
+ DUP4_ARG2(__lasx_xvldx, src, 0, src, stride, src, stride_2x, src, stride_3x,
+ src1, src2, src3, src4);
+ src += stride_4x;
+ DUP4_ARG2(__lasx_xvldx, src, 0, src, stride, src, stride_2x, src, stride_3x,
+ src5, src6, src7, src8);
+ DUP4_ARG3(__lasx_xvpermi_q, src2, src1, 0x20, src4, src3, 0x20, src6, src5, 0x20,
+ src8, src7, 0x20, src1, src3, src5, src7);
+ src0 = __lasx_xvshuf_b(src0, src0, mask);
+ DUP4_ARG3(__lasx_xvshuf_b, src1, src1, mask, src3, src3, mask, src5, src5, mask, src7,
+ src7, mask, src1, src3, src5, src7);
+ DUP4_ARG2(__lasx_xvdp2_h_bu, src0, coeff_hz_vec, src1, coeff_hz_vec, src3,
+ coeff_hz_vec, src5, coeff_hz_vec, res_hz0, res_hz1, res_hz2, res_hz3);
+ res_hz4 = __lasx_xvdp2_h_bu(src7, coeff_hz_vec);
+ res_vt0 = __lasx_xvmul_h(res_hz1, coeff_vt_vec0);
+ res_vt1 = __lasx_xvmul_h(res_hz2, coeff_vt_vec0);
+ res_vt2 = __lasx_xvmul_h(res_hz3, coeff_vt_vec0);
+ res_vt3 = __lasx_xvmul_h(res_hz4, coeff_vt_vec0);
+ res_hz0 = __lasx_xvpermi_q(res_hz1, res_hz0, 0x20);
+ res_hz1 = __lasx_xvpermi_q(res_hz1, res_hz2, 0x3);
+ res_hz2 = __lasx_xvpermi_q(res_hz2, res_hz3, 0x3);
+ res_hz3 = __lasx_xvpermi_q(res_hz3, res_hz4, 0x3);
+ res_vt0 = __lasx_xvmadd_h(res_vt0, res_hz0, coeff_vt_vec1);
+ res_vt1 = __lasx_xvmadd_h(res_vt1, res_hz1, coeff_vt_vec1);
+ res_vt2 = __lasx_xvmadd_h(res_vt2, res_hz2, coeff_vt_vec1);
+ res_vt3 = __lasx_xvmadd_h(res_vt3, res_hz3, coeff_vt_vec1);
+ DUP2_ARG3(__lasx_xvssrarni_bu_h, res_vt1, res_vt0, 6, res_vt3, res_vt2, 6,
+ out0, out1);
+ DUP4_ARG2(__lasx_xvldx, dst, 0, dst, stride, dst, stride_2x, dst, stride_3x,
+ tp0, tp1, tp2, tp3);
+ DUP2_ARG2(__lasx_xvilvl_d, tp2, tp0, tp3, tp1, tp0, tp2);
+ dst0 = __lasx_xvpermi_q(tp2, tp0, 0x20);
+ dst += stride_4x;
+ DUP4_ARG2(__lasx_xvldx, dst, 0, dst, stride, dst, stride_2x, dst, stride_3x,
+ tp0, tp1, tp2, tp3);
+ dst -= stride_4x;
+ DUP2_ARG2(__lasx_xvilvl_d, tp2, tp0, tp3, tp1, tp0, tp2);
+ dst1 = __lasx_xvpermi_q(tp2, tp0, 0x20);
+ out0 = __lasx_xvavgr_bu(out0, dst0);
+ out1 = __lasx_xvavgr_bu(out1, dst1);
+ __lasx_xvstelm_d(out0, dst, 0, 0);
+ __lasx_xvstelm_d(out0, dst + stride, 0, 2);
+ __lasx_xvstelm_d(out0, dst + stride_2x, 0, 1);
+ __lasx_xvstelm_d(out0, dst + stride_3x, 0, 3);
+ dst += stride_4x;
+ __lasx_xvstelm_d(out1, dst, 0, 0);
+ __lasx_xvstelm_d(out1, dst + stride, 0, 2);
+ __lasx_xvstelm_d(out1, dst + stride_2x, 0, 1);
+ __lasx_xvstelm_d(out1, dst + stride_3x, 0, 3);
+}
+
+static av_always_inline void avc_chroma_hz_and_aver_dst_8x4_lasx(uint8_t *src,
+ uint8_t *dst, ptrdiff_t stride, uint32_t coeff0,
+ uint32_t coeff1)
+{
+ ptrdiff_t stride_2x = stride << 1;
+ ptrdiff_t stride_3x = stride_2x + stride;
+ __m256i tp0, tp1, tp2, tp3;
+ __m256i src0, src1, src2, src3, out;
+ __m256i res0, res1;
+ __m256i mask;
+ __m256i coeff_vec0 = __lasx_xvreplgr2vr_b(coeff0);
+ __m256i coeff_vec1 = __lasx_xvreplgr2vr_b(coeff1);
+ __m256i coeff_vec = __lasx_xvilvl_b(coeff_vec0, coeff_vec1);
+
+ coeff_vec = __lasx_xvslli_b(coeff_vec, 3);
+ mask = __lasx_xvld(chroma_mask_arr, 0);
+ DUP4_ARG2(__lasx_xvldx, src, 0, src, stride, src, stride_2x, src, stride_3x,
+ src0, src1, src2, src3);
+ DUP2_ARG3(__lasx_xvpermi_q, src1, src0, 0x20, src3, src2, 0x20, src0, src2);
+ DUP2_ARG3(__lasx_xvshuf_b, src0, src0, mask, src2, src2, mask, src0, src2);
+ DUP2_ARG2(__lasx_xvdp2_h_bu, src0, coeff_vec, src2, coeff_vec, res0, res1);
+ out = __lasx_xvssrarni_bu_h(res1, res0, 6);
+ DUP4_ARG2(__lasx_xvldx, dst, 0, dst, stride, dst, stride_2x, dst, stride_3x,
+ tp0, tp1, tp2, tp3);
+ DUP2_ARG2(__lasx_xvilvl_d, tp2, tp0, tp3, tp1, tp0, tp2);
+ tp0 = __lasx_xvpermi_q(tp2, tp0, 0x20);
+ out = __lasx_xvavgr_bu(out, tp0);
+ __lasx_xvstelm_d(out, dst, 0, 0);
+ __lasx_xvstelm_d(out, dst + stride, 0, 2);
+ __lasx_xvstelm_d(out, dst + stride_2x, 0, 1);
+ __lasx_xvstelm_d(out, dst + stride_3x, 0, 3);
+}
+
+static av_always_inline void avc_chroma_hz_and_aver_dst_8x8_lasx(uint8_t *src,
+ uint8_t *dst, ptrdiff_t stride, uint32_t coeff0,
+ uint32_t coeff1)
+{
+ ptrdiff_t stride_2x = stride << 1;
+ ptrdiff_t stride_3x = stride_2x + stride;
+ ptrdiff_t stride_4x = stride << 2;
+ __m256i tp0, tp1, tp2, tp3, dst0, dst1;
+ __m256i src0, src1, src2, src3, src4, src5, src6, src7;
+ __m256i out0, out1;
+ __m256i res0, res1, res2, res3;
+ __m256i mask;
+ __m256i coeff_vec0 = __lasx_xvreplgr2vr_b(coeff0);
+ __m256i coeff_vec1 = __lasx_xvreplgr2vr_b(coeff1);
+ __m256i coeff_vec = __lasx_xvilvl_b(coeff_vec0, coeff_vec1);
+
+ coeff_vec = __lasx_xvslli_b(coeff_vec, 3);
+ mask = __lasx_xvld(chroma_mask_arr, 0);
+ DUP4_ARG2(__lasx_xvldx, src, 0, src, stride, src, stride_2x, src, stride_3x,
+ src0, src1, src2, src3);
+ src += stride_4x;
+ DUP4_ARG2(__lasx_xvldx, src, 0, src, stride, src, stride_2x, src, stride_3x,
+ src4, src5, src6, src7);
+ DUP4_ARG3(__lasx_xvpermi_q, src1, src0, 0x20, src3, src2, 0x20, src5, src4, 0x20,
+ src7, src6, 0x20, src0, src2, src4, src6);
+ DUP4_ARG3(__lasx_xvshuf_b, src0, src0, mask, src2, src2, mask, src4, src4,
+ mask, src6, src6, mask, src0, src2, src4, src6);
+ DUP4_ARG2(__lasx_xvdp2_h_bu, src0, coeff_vec, src2, coeff_vec, src4, coeff_vec, src6,
+ coeff_vec, res0, res1, res2, res3);
+ DUP2_ARG3(__lasx_xvssrarni_bu_h, res1, res0, 6, res3, res2, 6, out0, out1);
+ DUP4_ARG2(__lasx_xvldx, dst, 0, dst, stride, dst, stride_2x, dst, stride_3x,
+ tp0, tp1, tp2, tp3);
+ DUP2_ARG2(__lasx_xvilvl_d, tp2, tp0, tp3, tp1, tp0, tp2);
+ dst0 = __lasx_xvpermi_q(tp2, tp0, 0x20);
+ dst += stride_4x;
+ DUP4_ARG2(__lasx_xvldx, dst, 0, dst, stride, dst, stride_2x, dst, stride_3x,
+ tp0, tp1, tp2, tp3);
+ dst -= stride_4x;
+ DUP2_ARG2(__lasx_xvilvl_d, tp2, tp0, tp3, tp1, tp0, tp2);
+ dst1 = __lasx_xvpermi_q(tp2, tp0, 0x20);
+ out0 = __lasx_xvavgr_bu(out0, dst0);
+ out1 = __lasx_xvavgr_bu(out1, dst1);
+ __lasx_xvstelm_d(out0, dst, 0, 0);
+ __lasx_xvstelm_d(out0, dst + stride, 0, 2);
+ __lasx_xvstelm_d(out0, dst + stride_2x, 0, 1);
+ __lasx_xvstelm_d(out0, dst + stride_3x, 0, 3);
+ dst += stride_4x;
+ __lasx_xvstelm_d(out1, dst, 0, 0);
+ __lasx_xvstelm_d(out1, dst + stride, 0, 2);
+ __lasx_xvstelm_d(out1, dst + stride_2x, 0, 1);
+ __lasx_xvstelm_d(out1, dst + stride_3x, 0, 3);
+}
+
+static av_always_inline void avc_chroma_vt_and_aver_dst_8x4_lasx(uint8_t *src,
+ uint8_t *dst, ptrdiff_t stride, uint32_t coeff0,
+ uint32_t coeff1)
+{
+ ptrdiff_t stride_2x = stride << 1;
+ ptrdiff_t stride_3x = stride_2x + stride;
+ ptrdiff_t stride_4x = stride << 2;
+ __m256i tp0, tp1, tp2, tp3;
+ __m256i src0, src1, src2, src3, src4, out;
+ __m256i res0, res1;
+ __m256i coeff_vec0 = __lasx_xvreplgr2vr_b(coeff0);
+ __m256i coeff_vec1 = __lasx_xvreplgr2vr_b(coeff1);
+ __m256i coeff_vec = __lasx_xvilvl_b(coeff_vec0, coeff_vec1);
+
+ coeff_vec = __lasx_xvslli_b(coeff_vec, 3);
+ src0 = __lasx_xvld(src, 0);
+ DUP4_ARG2(__lasx_xvldx, src, stride, src, stride_2x, src, stride_3x, src, stride_4x,
+ src1, src2, src3, src4);
+ DUP4_ARG3(__lasx_xvpermi_q, src1, src0, 0x20, src2, src1, 0x20, src3, src2, 0x20,
+ src4, src3, 0x20, src0, src1, src2, src3);
+ DUP2_ARG2(__lasx_xvilvl_b, src1, src0, src3, src2, src0, src2);
+ DUP2_ARG2(__lasx_xvdp2_h_bu, src0, coeff_vec, src2, coeff_vec, res0, res1);
+ out = __lasx_xvssrarni_bu_h(res1, res0, 6);
+ DUP4_ARG2(__lasx_xvldx, dst, 0, dst, stride, dst, stride_2x, dst, stride_3x,
+ tp0, tp1, tp2, tp3);
+ DUP2_ARG2(__lasx_xvilvl_d, tp2, tp0, tp3, tp1, tp0, tp2);
+ tp0 = __lasx_xvpermi_q(tp2, tp0, 0x20);
+ out = __lasx_xvavgr_bu(out, tp0);
+ __lasx_xvstelm_d(out, dst, 0, 0);
+ __lasx_xvstelm_d(out, dst + stride, 0, 2);
+ __lasx_xvstelm_d(out, dst + stride_2x, 0, 1);
+ __lasx_xvstelm_d(out, dst + stride_3x, 0, 3);
+}
+
+static av_always_inline void avc_chroma_vt_and_aver_dst_8x8_lasx(uint8_t *src,
+ uint8_t *dst, ptrdiff_t stride, uint32_t coeff0,
+ uint32_t coeff1)
+{
+ ptrdiff_t stride_2x = stride << 1;
+ ptrdiff_t stride_3x = stride_2x + stride;
+ ptrdiff_t stride_4x = stride << 2;
+ __m256i tp0, tp1, tp2, tp3, dst0, dst1;
+ __m256i src0, src1, src2, src3, src4, src5, src6, src7, src8;
+ __m256i out0, out1;
+ __m256i res0, res1, res2, res3;
+ __m256i coeff_vec0 = __lasx_xvreplgr2vr_b(coeff0);
+ __m256i coeff_vec1 = __lasx_xvreplgr2vr_b(coeff1);
+ __m256i coeff_vec = __lasx_xvilvl_b(coeff_vec0, coeff_vec1);
+
+ coeff_vec = __lasx_xvslli_b(coeff_vec, 3);
+ src0 = __lasx_xvld(src, 0);
+ src += stride;
+ DUP4_ARG2(__lasx_xvldx, src, 0, src, stride, src, stride_2x, src, stride_3x,
+ src1, src2, src3, src4);
+ src += stride_4x;
+ DUP4_ARG2(__lasx_xvldx, src, 0, src, stride, src, stride_2x, src, stride_3x,
+ src5, src6, src7, src8);
+ DUP4_ARG3(__lasx_xvpermi_q, src1, src0, 0x20, src2, src1, 0x20, src3, src2, 0x20,
+ src4, src3, 0x20, src0, src1, src2, src3);
+ DUP4_ARG3(__lasx_xvpermi_q, src5, src4, 0x20, src6, src5, 0x20, src7, src6, 0x20,
+ src8, src7, 0x20, src4, src5, src6, src7);
+ DUP4_ARG2(__lasx_xvilvl_b, src1, src0, src3, src2, src5, src4, src7, src6,
+ src0, src2, src4, src6);
+ DUP4_ARG2(__lasx_xvdp2_h_bu, src0, coeff_vec, src2, coeff_vec, src4, coeff_vec, src6,
+ coeff_vec, res0, res1, res2, res3);
+ DUP2_ARG3(__lasx_xvssrarni_bu_h, res1, res0, 6, res3, res2, 6, out0, out1);
+ DUP4_ARG2(__lasx_xvldx, dst, 0, dst, stride, dst, stride_2x, dst, stride_3x,
+ tp0, tp1, tp2, tp3);
+ DUP2_ARG2(__lasx_xvilvl_d, tp2, tp0, tp3, tp1, tp0, tp2);
+ dst0 = __lasx_xvpermi_q(tp2, tp0, 0x20);
+ dst += stride_4x;
+ DUP4_ARG2(__lasx_xvldx, dst, 0, dst, stride, dst, stride_2x, dst, stride_3x,
+ tp0, tp1, tp2, tp3);
+ dst -= stride_4x;
+ DUP2_ARG2(__lasx_xvilvl_d, tp2, tp0, tp3, tp1, tp0, tp2);
+ dst1 = __lasx_xvpermi_q(tp2, tp0, 0x20);
+ out0 = __lasx_xvavgr_bu(out0, dst0);
+ out1 = __lasx_xvavgr_bu(out1, dst1);
+ __lasx_xvstelm_d(out0, dst, 0, 0);
+ __lasx_xvstelm_d(out0, dst + stride, 0, 2);
+ __lasx_xvstelm_d(out0, dst + stride_2x, 0, 1);
+ __lasx_xvstelm_d(out0, dst + stride_3x, 0, 3);
+ dst += stride_4x;
+ __lasx_xvstelm_d(out1, dst, 0, 0);
+ __lasx_xvstelm_d(out1, dst + stride, 0, 2);
+ __lasx_xvstelm_d(out1, dst + stride_2x, 0, 1);
+ __lasx_xvstelm_d(out1, dst + stride_3x, 0, 3);
+}
+
+static av_always_inline void avg_width8x8_lasx(uint8_t *src, uint8_t *dst,
+ ptrdiff_t stride)
+{
+ __m256i src0, src1, src2, src3;
+ __m256i dst0, dst1, dst2, dst3;
+ ptrdiff_t stride_2x = stride << 1;
+ ptrdiff_t stride_3x = stride_2x + stride;
+ ptrdiff_t stride_4x = stride << 2;
+
+ src0 = __lasx_xvldrepl_d(src, 0);
+ src1 = __lasx_xvldrepl_d(src + stride, 0);
+ src2 = __lasx_xvldrepl_d(src + stride_2x, 0);
+ src3 = __lasx_xvldrepl_d(src + stride_3x, 0);
+ dst0 = __lasx_xvldrepl_d(dst, 0);
+ dst1 = __lasx_xvldrepl_d(dst + stride, 0);
+ dst2 = __lasx_xvldrepl_d(dst + stride_2x, 0);
+ dst3 = __lasx_xvldrepl_d(dst + stride_3x, 0);
+ src0 = __lasx_xvpackev_d(src1,src0);
+ src2 = __lasx_xvpackev_d(src3,src2);
+ src0 = __lasx_xvpermi_q(src0, src2, 0x02);
+ dst0 = __lasx_xvpackev_d(dst1,dst0);
+ dst2 = __lasx_xvpackev_d(dst3,dst2);
+ dst0 = __lasx_xvpermi_q(dst0, dst2, 0x02);
+ dst0 = __lasx_xvavgr_bu(src0, dst0);
+ __lasx_xvstelm_d(dst0, dst, 0, 0);
+ __lasx_xvstelm_d(dst0, dst + stride, 0, 1);
+ __lasx_xvstelm_d(dst0, dst + stride_2x, 0, 2);
+ __lasx_xvstelm_d(dst0, dst + stride_3x, 0, 3);
+
+ src += stride_4x;
+ dst += stride_4x;
+ src0 = __lasx_xvldrepl_d(src, 0);
+ src1 = __lasx_xvldrepl_d(src + stride, 0);
+ src2 = __lasx_xvldrepl_d(src + stride_2x, 0);
+ src3 = __lasx_xvldrepl_d(src + stride_3x, 0);
+ dst0 = __lasx_xvldrepl_d(dst, 0);
+ dst1 = __lasx_xvldrepl_d(dst + stride, 0);
+ dst2 = __lasx_xvldrepl_d(dst + stride_2x, 0);
+ dst3 = __lasx_xvldrepl_d(dst + stride_3x, 0);
+ src0 = __lasx_xvpackev_d(src1,src0);
+ src2 = __lasx_xvpackev_d(src3,src2);
+ src0 = __lasx_xvpermi_q(src0, src2, 0x02);
+ dst0 = __lasx_xvpackev_d(dst1,dst0);
+ dst2 = __lasx_xvpackev_d(dst3,dst2);
+ dst0 = __lasx_xvpermi_q(dst0, dst2, 0x02);
+ dst0 = __lasx_xvavgr_bu(src0, dst0);
+ __lasx_xvstelm_d(dst0, dst, 0, 0);
+ __lasx_xvstelm_d(dst0, dst + stride, 0, 1);
+ __lasx_xvstelm_d(dst0, dst + stride_2x, 0, 2);
+ __lasx_xvstelm_d(dst0, dst + stride_3x, 0, 3);
+}
+
+static av_always_inline void avg_width8x4_lasx(uint8_t *src, uint8_t *dst,
+ ptrdiff_t stride)
+{
+ __m256i src0, src1, src2, src3;
+ __m256i dst0, dst1, dst2, dst3;
+ ptrdiff_t stride_2x = stride << 1;
+ ptrdiff_t stride_3x = stride_2x + stride;
+
+ src0 = __lasx_xvldrepl_d(src, 0);
+ src1 = __lasx_xvldrepl_d(src + stride, 0);
+ src2 = __lasx_xvldrepl_d(src + stride_2x, 0);
+ src3 = __lasx_xvldrepl_d(src + stride_3x, 0);
+ dst0 = __lasx_xvldrepl_d(dst, 0);
+ dst1 = __lasx_xvldrepl_d(dst + stride, 0);
+ dst2 = __lasx_xvldrepl_d(dst + stride_2x, 0);
+ dst3 = __lasx_xvldrepl_d(dst + stride_3x, 0);
+ src0 = __lasx_xvpackev_d(src1,src0);
+ src2 = __lasx_xvpackev_d(src3,src2);
+ src0 = __lasx_xvpermi_q(src0, src2, 0x02);
+ dst0 = __lasx_xvpackev_d(dst1,dst0);
+ dst2 = __lasx_xvpackev_d(dst3,dst2);
+ dst0 = __lasx_xvpermi_q(dst0, dst2, 0x02);
+ dst0 = __lasx_xvavgr_bu(src0, dst0);
+ __lasx_xvstelm_d(dst0, dst, 0, 0);
+ __lasx_xvstelm_d(dst0, dst + stride, 0, 1);
+ __lasx_xvstelm_d(dst0, dst + stride_2x, 0, 2);
+ __lasx_xvstelm_d(dst0, dst + stride_3x, 0, 3);
+}
+
+static void avc_chroma_hv_and_aver_dst_8w_lasx(uint8_t *src, uint8_t *dst,
+ ptrdiff_t stride,
+ uint32_t coef_hor0,
+ uint32_t coef_hor1,
+ uint32_t coef_ver0,
+ uint32_t coef_ver1,
+ int32_t height)
+{
+ if (4 == height) {
+ avc_chroma_hv_and_aver_dst_8x4_lasx(src, dst, stride, coef_hor0,
+ coef_hor1, coef_ver0, coef_ver1);
+ } else if (8 == height) {
+ avc_chroma_hv_and_aver_dst_8x8_lasx(src, dst, stride, coef_hor0,
+ coef_hor1, coef_ver0, coef_ver1);
+ }
+}
+
+static void avc_chroma_hz_and_aver_dst_8w_lasx(uint8_t *src, uint8_t *dst,
+ ptrdiff_t stride, uint32_t coeff0,
+ uint32_t coeff1, int32_t height)
+{
+ if (4 == height) {
+ avc_chroma_hz_and_aver_dst_8x4_lasx(src, dst, stride, coeff0, coeff1);
+ } else if (8 == height) {
+ avc_chroma_hz_and_aver_dst_8x8_lasx(src, dst, stride, coeff0, coeff1);
+ }
+}
+
+static void avc_chroma_vt_and_aver_dst_8w_lasx(uint8_t *src, uint8_t *dst,
+ ptrdiff_t stride, uint32_t coeff0,
+ uint32_t coeff1, int32_t height)
+{
+ if (4 == height) {
+ avc_chroma_vt_and_aver_dst_8x4_lasx(src, dst, stride, coeff0, coeff1);
+ } else if (8 == height) {
+ avc_chroma_vt_and_aver_dst_8x8_lasx(src, dst, stride, coeff0, coeff1);
+ }
+}
+
+static void avg_width8_lasx(uint8_t *src, uint8_t *dst, ptrdiff_t stride,
+ int32_t height)
+{
+ if (8 == height) {
+ avg_width8x8_lasx(src, dst, stride);
+ } else if (4 == height) {
+ avg_width8x4_lasx(src, dst, stride);
+ }
+}
+
+void ff_avg_h264_chroma_mc8_lasx(uint8_t *dst, uint8_t *src, ptrdiff_t stride,
+ int height, int x, int y)
+{
+ av_assert2(x < 8 && y < 8 && x >= 0 && y >= 0);
+
+ if (!(x || y)) {
+ avg_width8_lasx(src, dst, stride, height);
+ } else if (x && y) {
+ avc_chroma_hv_and_aver_dst_8w_lasx(src, dst, stride, x, (8 - x), y,
+ (8 - y), height);
+ } else if (x) {
+ avc_chroma_hz_and_aver_dst_8w_lasx(src, dst, stride, x, (8 - x), height);
+ } else {
+ avc_chroma_vt_and_aver_dst_8w_lasx(src, dst, stride, y, (8 - y), height);
+ }
+}
diff --git a/libavcodec/loongarch/h264chroma_lasx.h b/libavcodec/loongarch/h264chroma_lasx.h
new file mode 100644
index 0000000000..4aac8db8cb
--- /dev/null
+++ b/libavcodec/loongarch/h264chroma_lasx.h
@@ -0,0 +1,36 @@
+/*
+ * Copyright (c) 2020 Loongson Technology Corporation Limited
+ * Contributed by Shiyou Yin <yinshiyou-hf at loongson.cn>
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#ifndef AVCODEC_LOONGARCH_H264CHROMA_LASX_H
+#define AVCODEC_LOONGARCH_H264CHROMA_LASX_H
+
+#include <stdint.h>
+#include <stddef.h>
+#include "libavcodec/h264.h"
+
+void ff_put_h264_chroma_mc4_lasx(uint8_t *dst, uint8_t *src, ptrdiff_t stride,
+ int h, int x, int y);
+void ff_put_h264_chroma_mc8_lasx(uint8_t *dst, uint8_t *src, ptrdiff_t stride,
+ int h, int x, int y);
+void ff_avg_h264_chroma_mc8_lasx(uint8_t *dst, uint8_t *src, ptrdiff_t stride,
+ int h, int x, int y);
+
+#endif /* AVCODEC_LOONGARCH_H264CHROMA_LASX_H */
diff --git a/libavutil/loongarch/loongson_intrinsics.h b/libavutil/loongarch/loongson_intrinsics.h
new file mode 100644
index 0000000000..6e0439f829
--- /dev/null
+++ b/libavutil/loongarch/loongson_intrinsics.h
@@ -0,0 +1,1877 @@
+/*
+ * Copyright (c) 2021 Loongson Technology Corporation Limited
+ * All rights reserved.
+ * Contributed by Shiyou Yin <yinshiyou-hf at loongson.cn>
+ * Xiwei Gu <guxiwei-hf at loongson.cn>
+ * Lu Wang <wanglu at loongson.cn>
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ *
+ */
+
+#ifndef AVUTIL_LOONGARCH_LOONGSON_INTRINSICS_H
+#define AVUTIL_LOONGARCH_LOONGSON_INTRINSICS_H
+
+/*
+ * Copyright (c) 2021 Loongson Technology Corporation Limited
+ * All rights reserved.
+ * Contributed by Shiyou Yin <yinshiyou-hf at loongson.cn>
+ * Xiwei Gu <guxiwei-hf at loongson.cn>
+ * Lu Wang <wanglu at loongson.cn>
+ *
+ * This file is a header file for loongarch builtin extention.
+ *
+ */
+
+#ifndef LOONGSON_INTRINSICS_H
+#define LOONGSON_INTRINSICS_H
+
+/**
+ * MAJOR version: Macro usage changes.
+ * MINOR version: Add new functions, or bug fix.
+ * MICRO version: Comment changes or implementation changes.
+ */
+#define LSOM_VERSION_MAJOR 1
+#define LSOM_VERSION_MINOR 0
+#define LSOM_VERSION_MICRO 3
+
+#define DUP2_ARG1(_INS, _IN0, _IN1, _OUT0, _OUT1) \
+{ \
+ _OUT0 = _INS(_IN0); \
+ _OUT1 = _INS(_IN1); \
+}
+
+#define DUP2_ARG2(_INS, _IN0, _IN1, _IN2, _IN3, _OUT0, _OUT1) \
+{ \
+ _OUT0 = _INS(_IN0, _IN1); \
+ _OUT1 = _INS(_IN2, _IN3); \
+}
+
+#define DUP2_ARG3(_INS, _IN0, _IN1, _IN2, _IN3, _IN4, _IN5, _OUT0, _OUT1) \
+{ \
+ _OUT0 = _INS(_IN0, _IN1, _IN2); \
+ _OUT1 = _INS(_IN3, _IN4, _IN5); \
+}
+
+#define DUP4_ARG1(_INS, _IN0, _IN1, _IN2, _IN3, _OUT0, _OUT1, _OUT2, _OUT3) \
+{ \
+ DUP2_ARG1(_INS, _IN0, _IN1, _OUT0, _OUT1); \
+ DUP2_ARG1(_INS, _IN2, _IN3, _OUT2, _OUT3); \
+}
+
+#define DUP4_ARG2(_INS, _IN0, _IN1, _IN2, _IN3, _IN4, _IN5, _IN6, _IN7, \
+ _OUT0, _OUT1, _OUT2, _OUT3) \
+{ \
+ DUP2_ARG2(_INS, _IN0, _IN1, _IN2, _IN3, _OUT0, _OUT1); \
+ DUP2_ARG2(_INS, _IN4, _IN5, _IN6, _IN7, _OUT2, _OUT3); \
+}
+
+#define DUP4_ARG3(_INS, _IN0, _IN1, _IN2, _IN3, _IN4, _IN5, _IN6, _IN7, \
+ _IN8, _IN9, _IN10, _IN11, _OUT0, _OUT1, _OUT2, _OUT3) \
+{ \
+ DUP2_ARG3(_INS, _IN0, _IN1, _IN2, _IN3, _IN4, _IN5, _OUT0, _OUT1); \
+ DUP2_ARG3(_INS, _IN6, _IN7, _IN8, _IN9, _IN10, _IN11, _OUT2, _OUT3); \
+}
+
+#ifdef __loongarch_sx
+#include <lsxintrin.h>
+/*
+ * =============================================================================
+ * Description : Dot product & addition of byte vector elements
+ * Arguments : Inputs - in_c, in_h, in_l
+ * Outputs - out
+ * Retrun Type - halfword
+ * Details : Signed byte elements from in_h are multiplied by
+ * signed byte elements from in_l, and then added adjacent to
+ * each other to get results with the twice size of input.
+ * Then the results plus to signed half word elements from in_c.
+ * Example : out = __lsx_vdp2add_h_b(in_c, in_h, in_l)
+ * in_c : 1,2,3,4, 1,2,3,4
+ * in_h : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ * in_l : 8,7,6,5, 4,3,2,1, 8,7,6,5, 4,3,2,1
+ * out : 23,40,41,26, 23,40,41,26
+ * =============================================================================
+ */
+static inline __m128i __lsx_vdp2add_h_b(__m128i in_c, __m128i in_h, __m128i in_l)
+{
+ __m128i out;
+
+ out = __lsx_vmaddwev_h_b(in_c, in_h, in_l);
+ out = __lsx_vmaddwod_h_b(out, in_h, in_l);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product & addition of byte vector elements
+ * Arguments : Inputs - in_c, in_h, in_l
+ * Outputs - out
+ * Retrun Type - halfword
+ * Details : Unsigned byte elements from in_h are multiplied by
+ * unsigned byte elements from in_l, and then added adjacent to
+ * each other to get results with the twice size of input.
+ * The results plus to signed half word elements from in_c.
+ * Example : out = __lsx_vdp2add_h_b(in_c, in_h, in_l)
+ * in_c : 1,2,3,4, 1,2,3,4
+ * in_h : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ * in_l : 8,7,6,5, 4,3,2,1, 8,7,6,5, 4,3,2,1
+ * out : 23,40,41,26, 23,40,41,26
+ * =============================================================================
+ */
+static inline __m128i __lsx_vdp2add_h_bu(__m128i in_c, __m128i in_h, __m128i in_l)
+{
+ __m128i out;
+
+ out = __lsx_vmaddwev_h_bu(in_c, in_h, in_l);
+ out = __lsx_vmaddwod_h_bu(out, in_h, in_l);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product & addition of half word vector elements
+ * Arguments : Inputs - in_c, in_h, in_l
+ * Outputs - out
+ * Retrun Type - __m128i
+ * Details : Signed half word elements from in_h are multiplied by
+ * signed half word elements from in_l, and then added adjacent to
+ * each other to get results with the twice size of input.
+ * Then the results plus to signed word elements from in_c.
+ * Example : out = __lsx_vdp2add_h_b(in_c, in_h, in_l)
+ * in_c : 1,2,3,4
+ * in_h : 1,2,3,4, 5,6,7,8
+ * in_l : 8,7,6,5, 4,3,2,1
+ * out : 23,40,41,26
+ * =============================================================================
+ */
+static inline __m128i __lsx_vdp2add_w_h(__m128i in_c, __m128i in_h, __m128i in_l)
+{
+ __m128i out;
+
+ out = __lsx_vmaddwev_w_h(in_c, in_h, in_l);
+ out = __lsx_vmaddwod_w_h(out, in_h, in_l);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of byte vector elements
+ * Arguments : Inputs - in_h, in_l
+ * Outputs - out
+ * Retrun Type - halfword
+ * Details : Signed byte elements from in_h are multiplied by
+ * signed byte elements from in_l, and then added adjacent to
+ * each other to get results with the twice size of input.
+ * Example : out = __lsx_vdp2_h_b(in_h, in_l)
+ * in_h : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ * in_l : 8,7,6,5, 4,3,2,1, 8,7,6,5, 4,3,2,1
+ * out : 22,38,38,22, 22,38,38,22
+ * =============================================================================
+ */
+static inline __m128i __lsx_vdp2_h_b(__m128i in_h, __m128i in_l)
+{
+ __m128i out;
+
+ out = __lsx_vmulwev_h_b(in_h, in_l);
+ out = __lsx_vmaddwod_h_b(out, in_h, in_l);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of byte vector elements
+ * Arguments : Inputs - in_h, in_l
+ * Outputs - out
+ * Retrun Type - halfword
+ * Details : Unsigned byte elements from in_h are multiplied by
+ * unsigned byte elements from in_l, and then added adjacent to
+ * each other to get results with the twice size of input.
+ * Example : out = __lsx_vdp2_h_bu(in_h, in_l)
+ * in_h : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ * in_l : 8,7,6,5, 4,3,2,1, 8,7,6,5, 4,3,2,1
+ * out : 22,38,38,22, 22,38,38,22
+ * =============================================================================
+ */
+static inline __m128i __lsx_vdp2_h_bu(__m128i in_h, __m128i in_l)
+{
+ __m128i out;
+
+ out = __lsx_vmulwev_h_bu(in_h, in_l);
+ out = __lsx_vmaddwod_h_bu(out, in_h, in_l);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of byte vector elements
+ * Arguments : Inputs - in_h, in_l
+ * Outputs - out
+ * Retrun Type - halfword
+ * Details : Unsigned byte elements from in_h are multiplied by
+ * signed byte elements from in_l, and then added adjacent to
+ * each other to get results with the twice size of input.
+ * Example : out = __lsx_vdp2_h_bu_b(in_h, in_l)
+ * in_h : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ * in_l : 8,7,6,5, 4,3,2,1, 8,7,6,5, 4,3,2,-1
+ * out : 22,38,38,22, 22,38,38,6
+ * =============================================================================
+ */
+static inline __m128i __lsx_vdp2_h_bu_b(__m128i in_h, __m128i in_l)
+{
+ __m128i out;
+
+ out = __lsx_vmulwev_h_bu_b(in_h, in_l);
+ out = __lsx_vmaddwod_h_bu_b(out, in_h, in_l);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of byte vector elements
+ * Arguments : Inputs - in_h, in_l
+ * Outputs - out
+ * Retrun Type - halfword
+ * Details : Signed byte elements from in_h are multiplied by
+ * signed byte elements from in_l, and then added adjacent to
+ * each other to get results with the twice size of input.
+ * Example : out = __lsx_vdp2_w_h(in_h, in_l)
+ * in_h : 1,2,3,4, 5,6,7,8
+ * in_l : 8,7,6,5, 4,3,2,1
+ * out : 22,38,38,22
+ * =============================================================================
+ */
+static inline __m128i __lsx_vdp2_w_h(__m128i in_h, __m128i in_l)
+{
+ __m128i out;
+
+ out = __lsx_vmulwev_w_h(in_h, in_l);
+ out = __lsx_vmaddwod_w_h(out, in_h, in_l);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Clip all halfword elements of input vector between min & max
+ * out = ((_in) < (min)) ? (min) : (((_in) > (max)) ? (max) : (_in))
+ * Arguments : Inputs - _in (input vector)
+ * - min (min threshold)
+ * - max (max threshold)
+ * Outputs - out (output vector with clipped elements)
+ * Return Type - signed halfword
+ * Example : out = __lsx_vclip_h(_in)
+ * _in : -8,2,280,249, -8,255,280,249
+ * min : 1,1,1,1, 1,1,1,1
+ * max : 9,9,9,9, 9,9,9,9
+ * out : 1,2,9,9, 1,9,9,9
+ * =============================================================================
+ */
+static inline __m128i __lsx_vclip_h(__m128i _in, __m128i min, __m128i max)
+{
+ __m128i out;
+
+ out = __lsx_vmax_h(min, _in);
+ out = __lsx_vmin_h(max, out);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Set each element of vector between 0 and 255
+ * Arguments : Inputs - _in
+ * Outputs - out
+ * Retrun Type - halfword
+ * Details : Signed byte elements from _in are clamped between 0 and 255.
+ * Example : out = __lsx_vclip255_h(_in)
+ * _in : -8,255,280,249, -8,255,280,249
+ * out : 0,255,255,249, 0,255,255,249
+ * =============================================================================
+ */
+static inline __m128i __lsx_vclip255_h(__m128i _in)
+{
+ __m128i out;
+
+ out = __lsx_vmaxi_h(_in, 0);
+ out = __lsx_vsat_hu(out, 7);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Set each element of vector between 0 and 255
+ * Arguments : Inputs - _in
+ * Outputs - out
+ * Retrun Type - word
+ * Details : Signed byte elements from _in are clamped between 0 and 255.
+ * Example : out = __lsx_vclip255_w(_in)
+ * _in : -8,255,280,249
+ * out : 0,255,255,249
+ * =============================================================================
+ */
+static inline __m128i __lsx_vclip255_w(__m128i _in)
+{
+ __m128i out;
+
+ out = __lsx_vmaxi_w(_in, 0);
+ out = __lsx_vsat_wu(out, 7);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Swap two variables
+ * Arguments : Inputs - _in0, _in1
+ * Outputs - _in0, _in1 (in-place)
+ * Details : Swapping of two input variables using xor
+ * Example : LSX_SWAP(_in0, _in1)
+ * _in0 : 1,2,3,4
+ * _in1 : 5,6,7,8
+ * _in0(out) : 5,6,7,8
+ * _in1(out) : 1,2,3,4
+ * =============================================================================
+ */
+#define LSX_SWAP(_in0, _in1) \
+{ \
+ _in0 = __lsx_vxor_v(_in0, _in1); \
+ _in1 = __lsx_vxor_v(_in0, _in1); \
+ _in0 = __lsx_vxor_v(_in0, _in1); \
+} \
+
+/*
+ * =============================================================================
+ * Description : Transpose 4x4 block with word elements in vectors
+ * Arguments : Inputs - in0, in1, in2, in3
+ * Outputs - out0, out1, out2, out3
+ * Details :
+ * Example :
+ * 1, 2, 3, 4 1, 5, 9,13
+ * 5, 6, 7, 8 to 2, 6,10,14
+ * 9,10,11,12 =====> 3, 7,11,15
+ * 13,14,15,16 4, 8,12,16
+ * =============================================================================
+ */
+#define LSX_TRANSPOSE4x4_W(_in0, _in1, _in2, _in3, _out0, _out1, _out2, _out3) \
+{ \
+ __m128i _t0, _t1, _t2, _t3; \
+ \
+ _t0 = __lsx_vilvl_w(_in1, _in0); \
+ _t1 = __lsx_vilvh_w(_in1, _in0); \
+ _t2 = __lsx_vilvl_w(_in3, _in2); \
+ _t3 = __lsx_vilvh_w(_in3, _in2); \
+ _out0 = __lsx_vilvl_d(_t2, _t0); \
+ _out1 = __lsx_vilvh_d(_t2, _t0); \
+ _out2 = __lsx_vilvl_d(_t3, _t1); \
+ _out3 = __lsx_vilvh_d(_t3, _t1); \
+}
+
+/*
+ * =============================================================================
+ * Description : Transpose 8x8 block with byte elements in vectors
+ * Arguments : Inputs - _in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7
+ * Outputs - _out0, _out1, _out2, _out3, _out4, _out5, _out6, _out7
+ * Details : The rows of the matrix become columns, and the columns become rows.
+ * Example : LSX_TRANSPOSE8x8_B
+ * _in0 : 00,01,02,03,04,05,06,07, 00,00,00,00,00,00,00,00
+ * _in1 : 10,11,12,13,14,15,16,17, 00,00,00,00,00,00,00,00
+ * _in2 : 20,21,22,23,24,25,26,27, 00,00,00,00,00,00,00,00
+ * _in3 : 30,31,32,33,34,35,36,37, 00,00,00,00,00,00,00,00
+ * _in4 : 40,41,42,43,44,45,46,47, 00,00,00,00,00,00,00,00
+ * _in5 : 50,51,52,53,54,55,56,57, 00,00,00,00,00,00,00,00
+ * _in6 : 60,61,62,63,64,65,66,67, 00,00,00,00,00,00,00,00
+ * _in7 : 70,71,72,73,74,75,76,77, 00,00,00,00,00,00,00,00
+ *
+ * _ out0 : 00,10,20,30,40,50,60,70, 00,00,00,00,00,00,00,00
+ * _ out1 : 01,11,21,31,41,51,61,71, 00,00,00,00,00,00,00,00
+ * _ out2 : 02,12,22,32,42,52,62,72, 00,00,00,00,00,00,00,00
+ * _ out3 : 03,13,23,33,43,53,63,73, 00,00,00,00,00,00,00,00
+ * _ out4 : 04,14,24,34,44,54,64,74, 00,00,00,00,00,00,00,00
+ * _ out5 : 05,15,25,35,45,55,65,75, 00,00,00,00,00,00,00,00
+ * _ out6 : 06,16,26,36,46,56,66,76, 00,00,00,00,00,00,00,00
+ * _ out7 : 07,17,27,37,47,57,67,77, 00,00,00,00,00,00,00,00
+ * =============================================================================
+ */
+#define LSX_TRANSPOSE8x8_B(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7, \
+ _out0, _out1, _out2, _out3, _out4, _out5, _out6, _out7)\
+{ \
+ __m128i zero = {0}; \
+ __m128i shuf8 = {0x0F0E0D0C0B0A0908, 0x1716151413121110}; \
+ __m128i _t0, _t1, _t2, _t3, _t4, _t5, _t6, _t7; \
+ \
+ _t0 = __lsx_vilvl_b(_in2, _in0); \
+ _t1 = __lsx_vilvl_b(_in3, _in1); \
+ _t2 = __lsx_vilvl_b(_in6, _in4); \
+ _t3 = __lsx_vilvl_b(_in7, _in5); \
+ _t4 = __lsx_vilvl_b(_t1, _t0); \
+ _t5 = __lsx_vilvh_b(_t1, _t0); \
+ _t6 = __lsx_vilvl_b(_t3, _t2); \
+ _t7 = __lsx_vilvh_b(_t3, _t2); \
+ _out0 = __lsx_vilvl_w(_t6, _t4); \
+ _out2 = __lsx_vilvh_w(_t6, _t4); \
+ _out4 = __lsx_vilvl_w(_t7, _t5); \
+ _out6 = __lsx_vilvh_w(_t7, _t5); \
+ _out1 = __lsx_vshuf_b(zero, _out0, shuf8); \
+ _out3 = __lsx_vshuf_b(zero, _out2, shuf8); \
+ _out5 = __lsx_vshuf_b(zero, _out4, shuf8); \
+ _out7 = __lsx_vshuf_b(zero, _out6, shuf8); \
+}
+
+/*
+ * =============================================================================
+ * Description : Transpose 8x8 block with half word elements in vectors
+ * Arguments : Inputs - in0, in1, in2, in3, in4, in5, in6, in7
+ * Outputs - out0, out1, out2, out3, out4, out5, out6, out7
+ * Details :
+ * Example :
+ * 00,01,02,03,04,05,06,07 00,10,20,30,40,50,60,70
+ * 10,11,12,13,14,15,16,17 01,11,21,31,41,51,61,71
+ * 20,21,22,23,24,25,26,27 02,12,22,32,42,52,62,72
+ * 30,31,32,33,34,35,36,37 to 03,13,23,33,43,53,63,73
+ * 40,41,42,43,44,45,46,47 ======> 04,14,24,34,44,54,64,74
+ * 50,51,52,53,54,55,56,57 05,15,25,35,45,55,65,75
+ * 60,61,62,63,64,65,66,67 06,16,26,36,46,56,66,76
+ * 70,71,72,73,74,75,76,77 07,17,27,37,47,57,67,77
+ * =============================================================================
+ */
+#define LSX_TRANSPOSE8x8_H(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7, \
+ _out0, _out1, _out2, _out3, _out4, _out5, _out6, _out7)\
+{ \
+ __m128i _s0, _s1, _t0, _t1, _t2, _t3, _t4, _t5, _t6, _t7; \
+ \
+ _s0 = __lsx_vilvl_h(_in6, _in4); \
+ _s1 = __lsx_vilvl_h(_in7, _in5); \
+ _t0 = __lsx_vilvl_h(_s1, _s0); \
+ _t1 = __lsx_vilvh_h(_s1, _s0); \
+ _s0 = __lsx_vilvh_h(_in6, _in4); \
+ _s1 = __lsx_vilvh_h(_in7, _in5); \
+ _t2 = __lsx_vilvl_h(_s1, _s0); \
+ _t3 = __lsx_vilvh_h(_s1, _s0); \
+ _s0 = __lsx_vilvl_h(_in2, _in0); \
+ _s1 = __lsx_vilvl_h(_in3, _in1); \
+ _t4 = __lsx_vilvl_h(_s1, _s0); \
+ _t5 = __lsx_vilvh_h(_s1, _s0); \
+ _s0 = __lsx_vilvh_h(_in2, _in0); \
+ _s1 = __lsx_vilvh_h(_in3, _in1); \
+ _t6 = __lsx_vilvl_h(_s1, _s0); \
+ _t7 = __lsx_vilvh_h(_s1, _s0); \
+ \
+ _out0 = __lsx_vpickev_d(_t0, _t4); \
+ _out2 = __lsx_vpickev_d(_t1, _t5); \
+ _out4 = __lsx_vpickev_d(_t2, _t6); \
+ _out6 = __lsx_vpickev_d(_t3, _t7); \
+ _out1 = __lsx_vpickod_d(_t0, _t4); \
+ _out3 = __lsx_vpickod_d(_t1, _t5); \
+ _out5 = __lsx_vpickod_d(_t2, _t6); \
+ _out7 = __lsx_vpickod_d(_t3, _t7); \
+}
+
+/*
+ * =============================================================================
+ * Description : Transpose input 8x4 byte block into 4x8
+ * Arguments : Inputs - _in0, _in1, _in2, _in3 (input 8x4 byte block)
+ * Outputs - _out0, _out1, _out2, _out3 (output 4x8 byte block)
+ * Return Type - as per RTYPE
+ * Details : The rows of the matrix become columns, and the columns become rows.
+ * Example : LSX_TRANSPOSE8x4_B
+ * _in0 : 00,01,02,03,00,00,00,00, 00,00,00,00,00,00,00,00
+ * _in1 : 10,11,12,13,00,00,00,00, 00,00,00,00,00,00,00,00
+ * _in2 : 20,21,22,23,00,00,00,00, 00,00,00,00,00,00,00,00
+ * _in3 : 30,31,32,33,00,00,00,00, 00,00,00,00,00,00,00,00
+ * _in4 : 40,41,42,43,00,00,00,00, 00,00,00,00,00,00,00,00
+ * _in5 : 50,51,52,53,00,00,00,00, 00,00,00,00,00,00,00,00
+ * _in6 : 60,61,62,63,00,00,00,00, 00,00,00,00,00,00,00,00
+ * _in7 : 70,71,72,73,00,00,00,00, 00,00,00,00,00,00,00,00
+ *
+ * _out0 : 00,10,20,30,40,50,60,70, 00,00,00,00,00,00,00,00
+ * _out1 : 01,11,21,31,41,51,61,71, 00,00,00,00,00,00,00,00
+ * _out2 : 02,12,22,32,42,52,62,72, 00,00,00,00,00,00,00,00
+ * _out3 : 03,13,23,33,43,53,63,73, 00,00,00,00,00,00,00,00
+ * =============================================================================
+ */
+#define LSX_TRANSPOSE8x4_B(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7, \
+ _out0, _out1, _out2, _out3) \
+{ \
+ __m128i _tmp0_m, _tmp1_m, _tmp2_m, _tmp3_m; \
+ \
+ _tmp0_m = __lsx_vpackev_w(_in4, _in0); \
+ _tmp1_m = __lsx_vpackev_w(_in5, _in1); \
+ _tmp2_m = __lsx_vilvl_b(_tmp1_m, _tmp0_m); \
+ _tmp0_m = __lsx_vpackev_w(_in6, _in2); \
+ _tmp1_m = __lsx_vpackev_w(_in7, _in3); \
+ \
+ _tmp3_m = __lsx_vilvl_b(_tmp1_m, _tmp0_m); \
+ _tmp0_m = __lsx_vilvl_h(_tmp3_m, _tmp2_m); \
+ _tmp1_m = __lsx_vilvh_h(_tmp3_m, _tmp2_m); \
+ \
+ _out0 = __lsx_vilvl_w(_tmp1_m, _tmp0_m); \
+ _out2 = __lsx_vilvh_w(_tmp1_m, _tmp0_m); \
+ _out1 = __lsx_vilvh_d(_out2, _out0); \
+ _out3 = __lsx_vilvh_d(_out0, _out2); \
+}
+
+/*
+ * =============================================================================
+ * Description : Transpose 16x8 block with byte elements in vectors
+ * Arguments : Inputs - in0, in1, in2, in3, in4, in5, in6, in7, in8
+ * in9, in10, in11, in12, in13, in14, in15
+ * Outputs - out0, out1, out2, out3, out4, out5, out6, out7
+ * Details :
+ * Example :
+ * 000,001,002,003,004,005,006,007
+ * 008,009,010,011,012,013,014,015
+ * 016,017,018,019,020,021,022,023
+ * 024,025,026,027,028,029,030,031
+ * 032,033,034,035,036,037,038,039
+ * 040,041,042,043,044,045,046,047 000,008,...,112,120
+ * 048,049,050,051,052,053,054,055 001,009,...,113,121
+ * 056,057,058,059,060,061,062,063 to 002,010,...,114,122
+ * 064,068,066,067,068,069,070,071 =====> 003,011,...,115,123
+ * 072,073,074,075,076,077,078,079 004,012,...,116,124
+ * 080,081,082,083,084,085,086,087 005,013,...,117,125
+ * 088,089,090,091,092,093,094,095 006,014,...,118,126
+ * 096,097,098,099,100,101,102,103 007,015,...,119,127
+ * 104,105,106,107,108,109,110,111
+ * 112,113,114,115,116,117,118,119
+ * 120,121,122,123,124,125,126,127
+ * =============================================================================
+ */
+#define LSX_TRANSPOSE16x8_B(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7, _in8, \
+ _in9, _in10, _in11, _in12, _in13, _in14, _in15, _out0, \
+ _out1, _out2, _out3, _out4, _out5, _out6, _out7) \
+{ \
+ __m128i _tmp0, _tmp1, _tmp2, _tmp3, _tmp4, _tmp5, _tmp6, _tmp7; \
+ __m128i _t0, _t1, _t2, _t3, _t4, _t5, _t6, _t7; \
+ DUP4_ARG2(__lsx_vilvl_b, _in2, _in0, _in3, _in1, _in6, _in4, _in7, _in5, \
+ _tmp0, _tmp1, _tmp2, _tmp3); \
+ DUP4_ARG2(__lsx_vilvl_b, _in10, _in8, _in11, _in9, _in14, _in12, _in15, \
+ _in13, _tmp4, _tmp5, _tmp6, _tmp7); \
+ DUP2_ARG2(__lsx_vilvl_b, _tmp1, _tmp0, _tmp3, _tmp2, _t0, _t2); \
+ DUP2_ARG2(__lsx_vilvh_b, _tmp1, _tmp0, _tmp3, _tmp2, _t1, _t3); \
+ DUP2_ARG2(__lsx_vilvl_b, _tmp5, _tmp4, _tmp7, _tmp6, _t4, _t6); \
+ DUP2_ARG2(__lsx_vilvh_b, _tmp5, _tmp4, _tmp7, _tmp6, _t5, _t7); \
+ DUP2_ARG2(__lsx_vilvl_w, _t2, _t0, _t3, _t1, _tmp0, _tmp4); \
+ DUP2_ARG2(__lsx_vilvh_w, _t2, _t0, _t3, _t1, _tmp2, _tmp6); \
+ DUP2_ARG2(__lsx_vilvl_w, _t6, _t4, _t7, _t5, _tmp1, _tmp5); \
+ DUP2_ARG2(__lsx_vilvh_w, _t6, _t4, _t7, _t5, _tmp3, _tmp7); \
+ DUP2_ARG2(__lsx_vilvl_d, _tmp1, _tmp0, _tmp3, _tmp2, _out0, _out2); \
+ DUP2_ARG2(__lsx_vilvh_d, _tmp1, _tmp0, _tmp3, _tmp2, _out1, _out3); \
+ DUP2_ARG2(__lsx_vilvl_d, _tmp5, _tmp4, _tmp7, _tmp6, _out4, _out6); \
+ DUP2_ARG2(__lsx_vilvh_d, _tmp5, _tmp4, _tmp7, _tmp6, _out5, _out7); \
+}
+
+/*
+ * =============================================================================
+ * Description : Butterfly of 4 input vectors
+ * Arguments : Inputs - in0, in1, in2, in3
+ * Outputs - out0, out1, out2, out3
+ * Details : Butterfly operation
+ * Example :
+ * out0 = in0 + in3;
+ * out1 = in1 + in2;
+ * out2 = in1 - in2;
+ * out3 = in0 - in3;
+ * =============================================================================
+ */
+#define LSX_BUTTERFLY_4_B(_in0, _in1, _in2, _in3, _out0, _out1, _out2, _out3) \
+{ \
+ _out0 = __lsx_vadd_b(_in0, _in3); \
+ _out1 = __lsx_vadd_b(_in1, _in2); \
+ _out2 = __lsx_vsub_b(_in1, _in2); \
+ _out3 = __lsx_vsub_b(_in0, _in3); \
+}
+#define LSX_BUTTERFLY_4_H(_in0, _in1, _in2, _in3, _out0, _out1, _out2, _out3) \
+{ \
+ _out0 = __lsx_vadd_h(_in0, _in3); \
+ _out1 = __lsx_vadd_h(_in1, _in2); \
+ _out2 = __lsx_vsub_h(_in1, _in2); \
+ _out3 = __lsx_vsub_h(_in0, _in3); \
+}
+#define LSX_BUTTERFLY_4_W(_in0, _in1, _in2, _in3, _out0, _out1, _out2, _out3) \
+{ \
+ _out0 = __lsx_vadd_w(_in0, _in3); \
+ _out1 = __lsx_vadd_w(_in1, _in2); \
+ _out2 = __lsx_vsub_w(_in1, _in2); \
+ _out3 = __lsx_vsub_w(_in0, _in3); \
+}
+#define LSX_BUTTERFLY_4_D(_in0, _in1, _in2, _in3, _out0, _out1, _out2, _out3) \
+{ \
+ _out0 = __lsx_vadd_d(_in0, _in3); \
+ _out1 = __lsx_vadd_d(_in1, _in2); \
+ _out2 = __lsx_vsub_d(_in1, _in2); \
+ _out3 = __lsx_vsub_d(_in0, _in3); \
+}
+
+/*
+ * =============================================================================
+ * Description : Butterfly of 8 input vectors
+ * Arguments : Inputs - _in0, _in1, _in2, _in3, ~
+ * Outputs - _out0, _out1, _out2, _out3, ~
+ * Details : Butterfly operation
+ * Example :
+ * _out0 = _in0 + _in7;
+ * _out1 = _in1 + _in6;
+ * _out2 = _in2 + _in5;
+ * _out3 = _in3 + _in4;
+ * _out4 = _in3 - _in4;
+ * _out5 = _in2 - _in5;
+ * _out6 = _in1 - _in6;
+ * _out7 = _in0 - _in7;
+ * =============================================================================
+ */
+#define LSX_BUTTERFLY_8_B(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7, \
+ _out0, _out1, _out2, _out3, _out4, _out5, _out6, _out7)\
+{ \
+ _out0 = __lsx_vadd_b(_in0, _in7); \
+ _out1 = __lsx_vadd_b(_in1, _in6); \
+ _out2 = __lsx_vadd_b(_in2, _in5); \
+ _out3 = __lsx_vadd_b(_in3, _in4); \
+ _out4 = __lsx_vsub_b(_in3, _in4); \
+ _out5 = __lsx_vsub_b(_in2, _in5); \
+ _out6 = __lsx_vsub_b(_in1, _in6); \
+ _out7 = __lsx_vsub_b(_in0, _in7); \
+}
+
+#define LSX_BUTTERFLY_8_H(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7, \
+ _out0, _out1, _out2, _out3, _out4, _out5, _out6, _out7)\
+{ \
+ _out0 = __lsx_vadd_h(_in0, _in7); \
+ _out1 = __lsx_vadd_h(_in1, _in6); \
+ _out2 = __lsx_vadd_h(_in2, _in5); \
+ _out3 = __lsx_vadd_h(_in3, _in4); \
+ _out4 = __lsx_vsub_h(_in3, _in4); \
+ _out5 = __lsx_vsub_h(_in2, _in5); \
+ _out6 = __lsx_vsub_h(_in1, _in6); \
+ _out7 = __lsx_vsub_h(_in0, _in7); \
+}
+
+#define LSX_BUTTERFLY_8_W(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7, \
+ _out0, _out1, _out2, _out3, _out4, _out5, _out6, _out7)\
+{ \
+ _out0 = __lsx_vadd_w(_in0, _in7); \
+ _out1 = __lsx_vadd_w(_in1, _in6); \
+ _out2 = __lsx_vadd_w(_in2, _in5); \
+ _out3 = __lsx_vadd_w(_in3, _in4); \
+ _out4 = __lsx_vsub_w(_in3, _in4); \
+ _out5 = __lsx_vsub_w(_in2, _in5); \
+ _out6 = __lsx_vsub_w(_in1, _in6); \
+ _out7 = __lsx_vsub_w(_in0, _in7); \
+}
+
+#define LSX_BUTTERFLY_8_D(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7, \
+ _out0, _out1, _out2, _out3, _out4, _out5, _out6, _out7)\
+{ \
+ _out0 = __lsx_vadd_d(_in0, _in7); \
+ _out1 = __lsx_vadd_d(_in1, _in6); \
+ _out2 = __lsx_vadd_d(_in2, _in5); \
+ _out3 = __lsx_vadd_d(_in3, _in4); \
+ _out4 = __lsx_vsub_d(_in3, _in4); \
+ _out5 = __lsx_vsub_d(_in2, _in5); \
+ _out6 = __lsx_vsub_d(_in1, _in6); \
+ _out7 = __lsx_vsub_d(_in0, _in7); \
+}
+
+#endif //LSX
+
+#ifdef __loongarch_asx
+#include <lasxintrin.h>
+/*
+ * =============================================================================
+ * Description : Dot product of byte vector elements
+ * Arguments : Inputs - in_h, in_l
+ * Output - out
+ * Return Type - signed halfword
+ * Details : Unsigned byte elements from in_h are multiplied with
+ * unsigned byte elements from in_l producing a result
+ * twice the size of input i.e. signed halfword.
+ * Then this multiplied results of adjacent odd-even elements
+ * are added to the out vector
+ * Example : See out = __lasx_xvdp2_w_h(in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2_h_bu(__m256i in_h, __m256i in_l)
+{
+ __m256i out;
+
+ out = __lasx_xvmulwev_h_bu(in_h, in_l);
+ out = __lasx_xvmaddwod_h_bu(out, in_h, in_l);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of byte vector elements
+ * Arguments : Inputs - in_h, in_l
+ * Output - out
+ * Return Type - signed halfword
+ * Details : Signed byte elements from in_h are multiplied with
+ * signed byte elements from in_l producing a result
+ * twice the size of input i.e. signed halfword.
+ * Then this iniplication results of adjacent odd-even elements
+ * are added to the out vector
+ * Example : See out = __lasx_xvdp2_w_h(in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2_h_b(__m256i in_h, __m256i in_l)
+{
+ __m256i out;
+
+ out = __lasx_xvmulwev_h_b(in_h, in_l);
+ out = __lasx_xvmaddwod_h_b(out, in_h, in_l);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of halfword vector elements
+ * Arguments : Inputs - in_h, in_l
+ * Output - out
+ * Return Type - signed word
+ * Details : Signed halfword elements from in_h are multiplied with
+ * signed halfword elements from in_l producing a result
+ * twice the size of input i.e. signed word.
+ * Then this multiplied results of adjacent odd-even elements
+ * are added to the out vector.
+ * Example : out = __lasx_xvdp2_w_h(in_h, in_l)
+ * in_h : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ * in_l : 8,7,6,5, 4,3,2,1, 8,7,6,5, 4,3,2,1
+ * out : 22,38,38,22, 22,38,38,22
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2_w_h(__m256i in_h, __m256i in_l)
+{
+ __m256i out;
+
+ out = __lasx_xvmulwev_w_h(in_h, in_l);
+ out = __lasx_xvmaddwod_w_h(out, in_h, in_l);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of word vector elements
+ * Arguments : Inputs - in_h, in_l
+ * Output - out
+ * Retrun Type - signed double
+ * Details : Signed word elements from in_h are multiplied with
+ * signed word elements from in_l producing a result
+ * twice the size of input i.e. signed double word.
+ * Then this multiplied results of adjacent odd-even elements
+ * are added to the out vector.
+ * Example : See out = __lasx_xvdp2_w_h(in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2_d_w(__m256i in_h, __m256i in_l)
+{
+ __m256i out;
+
+ out = __lasx_xvmulwev_d_w(in_h, in_l);
+ out = __lasx_xvmaddwod_d_w(out, in_h, in_l);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of halfword vector elements
+ * Arguments : Inputs - in_h, in_l
+ * Output - out
+ * Return Type - signed word
+ * Details : Unsigned halfword elements from in_h are multiplied with
+ * signed halfword elements from in_l producing a result
+ * twice the size of input i.e. unsigned word.
+ * Multiplication result of adjacent odd-even elements
+ * are added to the out vector
+ * Example : See out = __lasx_xvdp2_w_h(in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2_w_hu_h(__m256i in_h, __m256i in_l)
+{
+ __m256i out;
+
+ out = __lasx_xvmulwev_w_hu_h(in_h, in_l);
+ out = __lasx_xvmaddwod_w_hu_h(out, in_h, in_l);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product & addition of byte vector elements
+ * Arguments : Inputs - in_h, in_l
+ * Output - out
+ * Retrun Type - halfword
+ * Details : Signed byte elements from in_h are multiplied with
+ * signed byte elements from in_l producing a result
+ * twice the size of input i.e. signed halfword.
+ * Then this multiplied results of adjacent odd-even elements
+ * are added to the in_c vector.
+ * Example : See out = __lasx_xvdp2add_w_h(in_c, in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2add_h_b(__m256i in_c,__m256i in_h, __m256i in_l)
+{
+ __m256i out;
+
+ out = __lasx_xvmaddwev_h_b(in_c, in_h, in_l);
+ out = __lasx_xvmaddwod_h_b(out, in_h, in_l);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of halfword vector elements
+ * Arguments : Inputs - in_c, in_h, in_l
+ * Output - out
+ * Return Type - per RTYPE
+ * Details : Signed halfword elements from in_h are multiplied with
+ * signed halfword elements from in_l producing a result
+ * twice the size of input i.e. signed word.
+ * Multiplication result of adjacent odd-even elements
+ * are added to the in_c vector.
+ * Example : out = __lasx_xvdp2add_w_h(in_c, in_h, in_l)
+ * in_c : 1,2,3,4, 1,2,3,4
+ * in_h : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8,
+ * in_l : 8,7,6,5, 4,3,2,1, 8,7,6,5, 4,3,2,1,
+ * out : 23,40,41,26, 23,40,41,26
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2add_w_h(__m256i in_c, __m256i in_h, __m256i in_l)
+{
+ __m256i out;
+
+ out = __lasx_xvmaddwev_w_h(in_c, in_h, in_l);
+ out = __lasx_xvmaddwod_w_h(out, in_h, in_l);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of halfword vector elements
+ * Arguments : Inputs - in_c, in_h, in_l
+ * Output - out
+ * Return Type - signed word
+ * Details : Unsigned halfword elements from in_h are multiplied with
+ * unsigned halfword elements from in_l producing a result
+ * twice the size of input i.e. signed word.
+ * Multiplication result of adjacent odd-even elements
+ * are added to the in_c vector.
+ * Example : See out = __lasx_xvdp2add_w_h(in_c, in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2add_w_hu(__m256i in_c, __m256i in_h, __m256i in_l)
+{
+ __m256i out;
+
+ out = __lasx_xvmaddwev_w_hu(in_c, in_h, in_l);
+ out = __lasx_xvmaddwod_w_hu(out, in_h, in_l);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of halfword vector elements
+ * Arguments : Inputs - in_c, in_h, in_l
+ * Output - out
+ * Return Type - signed word
+ * Details : Unsigned halfword elements from in_h are multiplied with
+ * signed halfword elements from in_l producing a result
+ * twice the size of input i.e. signed word.
+ * Multiplication result of adjacent odd-even elements
+ * are added to the in_c vector
+ * Example : See out = __lasx_xvdp2add_w_h(in_c, in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2add_w_hu_h(__m256i in_c, __m256i in_h, __m256i in_l)
+{
+ __m256i out;
+
+ out = __lasx_xvmaddwev_w_hu_h(in_c, in_h, in_l);
+ out = __lasx_xvmaddwod_w_hu_h(out, in_h, in_l);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Vector Unsigned Dot Product and Subtract
+ * Arguments : Inputs - in_c, in_h, in_l
+ * Output - out
+ * Return Type - signed halfword
+ * Details : Unsigned byte elements from in_h are multiplied with
+ * unsigned byte elements from in_l producing a result
+ * twice the size of input i.e. signed halfword.
+ * Multiplication result of adjacent odd-even elements
+ * are added together and subtracted from double width elements
+ * in_c vector.
+ * Example : See out = __lasx_xvdp2sub_w_h(in_c, in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2sub_h_bu(__m256i in_c, __m256i in_h, __m256i in_l)
+{
+ __m256i out;
+
+ out = __lasx_xvmulwev_h_bu(in_h, in_l);
+ out = __lasx_xvmaddwod_h_bu(out, in_h, in_l);
+ out = __lasx_xvsub_h(in_c, out);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Vector Signed Dot Product and Subtract
+ * Arguments : Inputs - in_c, in_h, in_l
+ * Output - out
+ * Return Type - signed word
+ * Details : Signed halfword elements from in_h are multiplied with
+ * Signed halfword elements from in_l producing a result
+ * twice the size of input i.e. signed word.
+ * Multiplication result of adjacent odd-even elements
+ * are added together and subtracted from double width elements
+ * in_c vector.
+ * Example : out = __lasx_xvdp2sub_w_h(in_c, in_h, in_l)
+ * in_c : 0,0,0,0, 0,0,0,0
+ * in_h : 3,1,3,0, 0,0,0,1, 0,0,1,1, 0,0,0,1
+ * in_l : 2,1,1,0, 1,0,0,0, 0,0,1,0, 1,0,0,1
+ * out : -7,-3,0,0, 0,-1,0,-1
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2sub_w_h(__m256i in_c, __m256i in_h, __m256i in_l)
+{
+ __m256i out;
+
+ out = __lasx_xvmulwev_w_h(in_h, in_l);
+ out = __lasx_xvmaddwod_w_h(out, in_h, in_l);
+ out = __lasx_xvsub_w(in_c, out);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of halfword vector elements
+ * Arguments : Inputs - in_h, in_l
+ * Output - out
+ * Return Type - signed word
+ * Details : Signed halfword elements from in_h are iniplied with
+ * signed halfword elements from in_l producing a result
+ * four times the size of input i.e. signed doubleword.
+ * Then this iniplication results of four adjacent elements
+ * are added together and stored to the out vector.
+ * Example : out = __lasx_xvdp4_d_h(in_h, in_l)
+ * in_h : 3,1,3,0, 0,0,0,1, 0,0,1,-1, 0,0,0,1
+ * in_l : -2,1,1,0, 1,0,0,0, 0,0,1, 0, 1,0,0,1
+ * out : -2,0,1,1
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp4_d_h(__m256i in_h, __m256i in_l)
+{
+ __m256i out;
+
+ out = __lasx_xvmulwev_w_h(in_h, in_l);
+ out = __lasx_xvmaddwod_w_h(out, in_h, in_l);
+ out = __lasx_xvhaddw_d_w(out, out);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : The high half of the vector elements are expanded and
+ * added after being doubled.
+ * Arguments : Inputs - in_h, in_l
+ * Output - out
+ * Details : The in_h vector and the in_l vector are added after the
+ * higher half of the two-fold sign extension (signed byte
+ * to signed halfword) and stored to the out vector.
+ * Example : See out = __lasx_xvaddwh_w_h(in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvaddwh_h_b(__m256i in_h, __m256i in_l)
+{
+ __m256i out;
+
+ out = __lasx_xvilvh_b(in_h, in_l);
+ out = __lasx_xvhaddw_h_b(out, out);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : The high half of the vector elements are expanded and
+ * added after being doubled.
+ * Arguments : Inputs - in_h, in_l
+ * Output - out
+ * Details : The in_h vector and the in_l vector are added after the
+ * higher half of the two-fold sign extension (signed halfword
+ * to signed word) and stored to the out vector.
+ * Example : out = __lasx_xvaddwh_w_h(in_h, in_l)
+ * in_h : 3, 0,3,0, 0,0,0,-1, 0,0,1,-1, 0,0,0,1
+ * in_l : 2,-1,1,2, 1,0,0, 0, 1,0,1, 0, 1,0,0,1
+ * out : 1,0,0,-1, 1,0,0, 2
+ * =============================================================================
+ */
+ static inline __m256i __lasx_xvaddwh_w_h(__m256i in_h, __m256i in_l)
+{
+ __m256i out;
+
+ out = __lasx_xvilvh_h(in_h, in_l);
+ out = __lasx_xvhaddw_w_h(out, out);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : The low half of the vector elements are expanded and
+ * added after being doubled.
+ * Arguments : Inputs - in_h, in_l
+ * Output - out
+ * Details : The in_h vector and the in_l vector are added after the
+ * lower half of the two-fold sign extension (signed byte
+ * to signed halfword) and stored to the out vector.
+ * Example : See out = __lasx_xvaddwl_w_h(in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvaddwl_h_b(__m256i in_h, __m256i in_l)
+{
+ __m256i out;
+
+ out = __lasx_xvilvl_b(in_h, in_l);
+ out = __lasx_xvhaddw_h_b(out, out);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : The low half of the vector elements are expanded and
+ * added after being doubled.
+ * Arguments : Inputs - in_h, in_l
+ * Output - out
+ * Details : The in_h vector and the in_l vector are added after the
+ * lower half of the two-fold sign extension (signed halfword
+ * to signed word) and stored to the out vector.
+ * Example : out = __lasx_xvaddwl_w_h(in_h, in_l)
+ * in_h : 3, 0,3,0, 0,0,0,-1, 0,0,1,-1, 0,0,0,1
+ * in_l : 2,-1,1,2, 1,0,0, 0, 1,0,1, 0, 1,0,0,1
+ * out : 5,-1,4,2, 1,0,2,-1
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvaddwl_w_h(__m256i in_h, __m256i in_l)
+{
+ __m256i out;
+
+ out = __lasx_xvilvl_h(in_h, in_l);
+ out = __lasx_xvhaddw_w_h(out, out);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : The low half of the vector elements are expanded and
+ * added after being doubled.
+ * Arguments : Inputs - in_h, in_l
+ * Output - out
+ * Details : The out vector and the out vector are added after the
+ * lower half of the two-fold zero extension (unsigned byte
+ * to unsigned halfword) and stored to the out vector.
+ * Example : See out = __lasx_xvaddwl_w_h(in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvaddwl_h_bu(__m256i in_h, __m256i in_l)
+{
+ __m256i out;
+
+ out = __lasx_xvilvl_b(in_h, in_l);
+ out = __lasx_xvhaddw_hu_bu(out, out);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : The low half of the vector elements are expanded and
+ * added after being doubled.
+ * Arguments : Inputs - in_h, in_l
+ * Output - out
+ * Details : The in_l vector after double zero extension (unsigned byte to
+ * signed halfword)，added to the in_h vector.
+ * Example : See out = __lasx_xvaddw_w_w_h(in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvaddw_h_h_bu(__m256i in_h, __m256i in_l)
+{
+ __m256i out;
+
+ out = __lasx_xvsllwil_hu_bu(in_l, 0);
+ out = __lasx_xvadd_h(in_h, out);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : The low half of the vector elements are expanded and
+ * added after being doubled.
+ * Arguments : Inputs - in_h, in_l
+ * Output - out
+ * Details : The in_l vector after double sign extension (signed halfword to
+ * signed word), added to the in_h vector.
+ * Example : out = __lasx_xvaddw_w_w_h(in_h, in_l)
+ * in_h : 0, 1,0,0, -1,0,0,1,
+ * in_l : 2,-1,1,2, 1,0,0,0, 0,0,1,0, 1,0,0,1,
+ * out : 2, 0,1,2, -1,0,1,1,
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvaddw_w_w_h(__m256i in_h, __m256i in_l)
+{
+ __m256i out;
+
+ out = __lasx_xvsllwil_w_h(in_l, 0);
+ out = __lasx_xvadd_w(in_h, out);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Multiplication and addition calculation after expansion
+ * of the lower half of the vector.
+ * Arguments : Inputs - in_c, in_h, in_l
+ * Output - out
+ * Details : The in_h vector and the in_l vector are multiplied after
+ * the lower half of the two-fold sign extension (signed halfword
+ * to signed word), and the result is added to the vector in_c,
+ * then stored to the out vector.
+ * Example : out = __lasx_xvmaddwl_w_h(in_c, in_h, in_l)
+ * in_c : 1,2,3,4, 5,6,7,8
+ * in_h : 1,2,3,4, 1,2,3,4, 5,6,7,8, 5,6,7,8
+ * in_l : 200, 300, 400, 500, 2000, 3000, 4000, 5000,
+ * -200,-300,-400,-500, -2000,-3000,-4000,-5000
+ * out : 201, 602,1203,2004, -995, -1794,-2793,-3992
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvmaddwl_w_h(__m256i in_c, __m256i in_h, __m256i in_l)
+{
+ __m256i tmp0, tmp1, out;
+
+ tmp0 = __lasx_xvsllwil_w_h(in_h, 0);
+ tmp1 = __lasx_xvsllwil_w_h(in_l, 0);
+ tmp0 = __lasx_xvmul_w(tmp0, tmp1);
+ out = __lasx_xvadd_w(tmp0, in_c);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Multiplication and addition calculation after expansion
+ * of the higher half of the vector.
+ * Arguments : Inputs - in_c, in_h, in_l
+ * Output - out
+ * Details : The in_h vector and the in_l vector are multiplied after
+ * the higher half of the two-fold sign extension (signed
+ * halfword to signed word), and the result is added to
+ * the vector in_c, then stored to the out vector.
+ * Example : See out = __lasx_xvmaddwl_w_h(in_c, in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvmaddwh_w_h(__m256i in_c, __m256i in_h, __m256i in_l)
+{
+ __m256i tmp0, tmp1, out;
+
+ tmp0 = __lasx_xvilvh_h(in_h, in_h);
+ tmp1 = __lasx_xvilvh_h(in_l, in_l);
+ tmp0 = __lasx_xvmulwev_w_h(tmp0, tmp1);
+ out = __lasx_xvadd_w(tmp0, in_c);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Multiplication calculation after expansion of the lower
+ * half of the vector.
+ * Arguments : Inputs - in_h, in_l
+ * Output - out
+ * Details : The in_h vector and the in_l vector are multiplied after
+ * the lower half of the two-fold sign extension (signed
+ * halfword to signed word), then stored to the out vector.
+ * Example : out = __lasx_xvmulwl_w_h(in_h, in_l)
+ * in_h : 3,-1,3,0, 0,0,0,-1, 0,0,1,-1, 0,0,0,1
+ * in_l : 2,-1,1,2, 1,0,0, 0, 0,0,1, 0, 1,0,0,1
+ * out : 6,1,3,0, 0,0,1,0
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvmulwl_w_h(__m256i in_h, __m256i in_l)
+{
+ __m256i tmp0, tmp1, out;
+
+ tmp0 = __lasx_xvsllwil_w_h(in_h, 0);
+ tmp1 = __lasx_xvsllwil_w_h(in_l, 0);
+ out = __lasx_xvmul_w(tmp0, tmp1);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Multiplication calculation after expansion of the lower
+ * half of the vector.
+ * Arguments : Inputs - in_h, in_l
+ * Output - out
+ * Details : The in_h vector and the in_l vector are multiplied after
+ * the lower half of the two-fold sign extension (signed
+ * halfword to signed word), then stored to the out vector.
+ * Example : out = __lasx_xvmulwh_w_h(in_h, in_l)
+ * in_h : 3,-1,3,0, 0,0,0,-1, 0,0,1,-1, 0,0,0,1
+ * in_l : 2,-1,1,2, 1,0,0, 0, 0,0,1, 0, 1,0,0,1
+ * out : 0,0,0,0, 0,0,0,1
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvmulwh_w_h(__m256i in_h, __m256i in_l)
+{
+ __m256i tmp0, tmp1, out;
+
+ tmp0 = __lasx_xvilvh_h(in_h, in_h);
+ tmp1 = __lasx_xvilvh_h(in_l, in_l);
+ out = __lasx_xvmulwev_w_h(tmp0, tmp1);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : The low half of the vector elements are expanded and
+ * added saturately after being doubled.
+ * Arguments : Inputs - in_h, in_l
+ * Output - out
+ * Details : The in_h vector adds the in_l vector saturately after the lower
+ * half of the two-fold zero extension (unsigned byte to unsigned
+ * halfword) and the results are stored to the out vector.
+ * Example : out = __lasx_xvsaddw_hu_hu_bu(in_h, in_l)
+ * in_h : 2,65532,1,2, 1,0,0,0, 0,0,1,0, 1,0,0,1
+ * in_l : 3,6,3,0, 0,0,0,1, 0,0,1,1, 0,0,0,1, 3,18,3,0, 0,0,0,1, 0,0,1,1, 0,0,0,1
+ * out : 5,65535,4,2, 1,0,0,1, 3,18,4,0, 1,0,0,2,
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvsaddw_hu_hu_bu(__m256i in_h, __m256i in_l)
+{
+ __m256i tmp1, out;
+ __m256i zero = {0};
+
+ tmp1 = __lasx_xvilvl_b(zero, in_l);
+ out = __lasx_xvsadd_hu(in_h, tmp1);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Clip all halfword elements of input vector between min & max
+ * out = ((in) < (min)) ? (min) : (((in) > (max)) ? (max) : (in))
+ * Arguments : Inputs - in (input vector)
+ * - min (min threshold)
+ * - max (max threshold)
+ * Outputs - in (output vector with clipped elements)
+ * Return Type - signed halfword
+ * Example : out = __lasx_xvclip_h(in, min, max)
+ * in : -8,2,280,249, -8,255,280,249, 4,4,4,4, 5,5,5,5
+ * min : 1,1,1,1, 1,1,1,1, 1,1,1,1, 1,1,1,1
+ * max : 9,9,9,9, 9,9,9,9, 9,9,9,9, 9,9,9,9
+ * out : 1,2,9,9, 1,9,9,9, 4,4,4,4, 5,5,5,5
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvclip_h(__m256i in, __m256i min, __m256i max)
+{
+ __m256i out;
+
+ out = __lasx_xvmax_h(min, in);
+ out = __lasx_xvmin_h(max, out);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Clip all signed halfword elements of input vector
+ * between 0 & 255
+ * Arguments : Inputs - in (input vector)
+ * Outputs - out (output vector with clipped elements)
+ * Return Type - signed halfword
+ * Example : See out = __lasx_xvclip255_w(in)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvclip255_h(__m256i in)
+{
+ __m256i out;
+
+ out = __lasx_xvmaxi_h(in, 0);
+ out = __lasx_xvsat_hu(out, 7);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Clip all signed word elements of input vector
+ * between 0 & 255
+ * Arguments : Inputs - in (input vector)
+ * Output - out (output vector with clipped elements)
+ * Return Type - signed word
+ * Example : out = __lasx_xvclip255_w(in)
+ * in : -8,255,280,249, -8,255,280,249
+ * out : 0,255,255,249, 0,255,255,249
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvclip255_w(__m256i in)
+{
+ __m256i out;
+
+ out = __lasx_xvmaxi_w(in, 0);
+ out = __lasx_xvsat_wu(out, 7);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Indexed halfword element values are replicated to all
+ * elements in output vector. If 'indx < 8' use xvsplati_l_*,
+ * if 'indx >= 8' use xvsplati_h_*.
+ * Arguments : Inputs - in, idx
+ * Output - out
+ * Details : Idx element value from in vector is replicated to all
+ * elements in out vector.
+ * Valid index range for halfword operation is 0-7
+ * Example : out = __lasx_xvsplati_l_h(in, idx)
+ * in : 20,10,11,12, 13,14,15,16, 0,0,2,0, 0,0,0,0
+ * idx : 0x02
+ * out : 11,11,11,11, 11,11,11,11, 11,11,11,11, 11,11,11,11
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvsplati_l_h(__m256i in, int idx)
+{
+ __m256i out;
+
+ out = __lasx_xvpermi_q(in, in, 0x02);
+ out = __lasx_xvreplve_h(out, idx);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Indexed halfword element values are replicated to all
+ * elements in output vector. If 'indx < 8' use xvsplati_l_*,
+ * if 'indx >= 8' use xvsplati_h_*.
+ * Arguments : Inputs - in, idx
+ * Output - out
+ * Details : Idx element value from in vector is replicated to all
+ * elements in out vector.
+ * Valid index range for halfword operation is 0-7
+ * Example : out = __lasx_xvsplati_h_h(in, idx)
+ * in : 20,10,11,12, 13,14,15,16, 0,2,0,0, 0,0,0,0
+ * idx : 0x09
+ * out : 2,2,2,2, 2,2,2,2, 2,2,2,2, 2,2,2,2
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvsplati_h_h(__m256i in, int idx)
+{
+ __m256i out;
+
+ out = __lasx_xvpermi_q(in, in, 0x13);
+ out = __lasx_xvreplve_h(out, idx);
+ return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Transpose 4x4 block with double word elements in vectors
+ * Arguments : Inputs - _in0, _in1, _in2, _in3
+ * Outputs - _out0, _out1, _out2, _out3
+ * Example : LASX_TRANSPOSE4x4_D
+ * _in0 : 1,2,3,4
+ * _in1 : 1,2,3,4
+ * _in2 : 1,2,3,4
+ * _in3 : 1,2,3,4
+ *
+ * _out0 : 1,1,1,1
+ * _out1 : 2,2,2,2
+ * _out2 : 3,3,3,3
+ * _out3 : 4,4,4,4
+ * =============================================================================
+ */
+#define LASX_TRANSPOSE4x4_D(_in0, _in1, _in2, _in3, _out0, _out1, _out2, _out3) \
+{ \
+ __m256i _tmp0, _tmp1, _tmp2, _tmp3; \
+ _tmp0 = __lasx_xvilvl_d(_in1, _in0); \
+ _tmp1 = __lasx_xvilvh_d(_in1, _in0); \
+ _tmp2 = __lasx_xvilvl_d(_in3, _in2); \
+ _tmp3 = __lasx_xvilvh_d(_in3, _in2); \
+ _out0 = __lasx_xvpermi_q(_tmp2, _tmp0, 0x20); \
+ _out2 = __lasx_xvpermi_q(_tmp2, _tmp0, 0x31); \
+ _out1 = __lasx_xvpermi_q(_tmp3, _tmp1, 0x20); \
+ _out3 = __lasx_xvpermi_q(_tmp3, _tmp1, 0x31); \
+}
+
+/*
+ * =============================================================================
+ * Description : Transpose 8x8 block with word elements in vectors
+ * Arguments : Inputs - _in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7
+ * Outputs - _out0, _out1, _out2, _out3, _out4, _out5, _out6, _out7
+ * Example : LASX_TRANSPOSE8x8_W
+ * _in0 : 1,2,3,4,5,6,7,8
+ * _in1 : 2,2,3,4,5,6,7,8
+ * _in2 : 3,2,3,4,5,6,7,8
+ * _in3 : 4,2,3,4,5,6,7,8
+ * _in4 : 5,2,3,4,5,6,7,8
+ * _in5 : 6,2,3,4,5,6,7,8
+ * _in6 : 7,2,3,4,5,6,7,8
+ * _in7 : 8,2,3,4,5,6,7,8
+ *
+ * _out0 : 1,2,3,4,5,6,7,8
+ * _out1 : 2,2,2,2,2,2,2,2
+ * _out2 : 3,3,3,3,3,3,3,3
+ * _out3 : 4,4,4,4,4,4,4,4
+ * _out4 : 5,5,5,5,5,5,5,5
+ * _out5 : 6,6,6,6,6,6,6,6
+ * _out6 : 7,7,7,7,7,7,7,7
+ * _out7 : 8,8,8,8,8,8,8,8
+ * =============================================================================
+ */
+#define LASX_TRANSPOSE8x8_W(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7, \
+ _out0, _out1, _out2, _out3, _out4, _out5, _out6, _out7) \
+{ \
+ __m256i _s0_m, _s1_m; \
+ __m256i _tmp0_m, _tmp1_m, _tmp2_m, _tmp3_m; \
+ __m256i _tmp4_m, _tmp5_m, _tmp6_m, _tmp7_m; \
+ \
+ _s0_m = __lasx_xvilvl_w(_in2, _in0); \
+ _s1_m = __lasx_xvilvl_w(_in3, _in1); \
+ _tmp0_m = __lasx_xvilvl_w(_s1_m, _s0_m); \
+ _tmp1_m = __lasx_xvilvh_w(_s1_m, _s0_m); \
+ _s0_m = __lasx_xvilvh_w(_in2, _in0); \
+ _s1_m = __lasx_xvilvh_w(_in3, _in1); \
+ _tmp2_m = __lasx_xvilvl_w(_s1_m, _s0_m); \
+ _tmp3_m = __lasx_xvilvh_w(_s1_m, _s0_m); \
+ _s0_m = __lasx_xvilvl_w(_in6, _in4); \
+ _s1_m = __lasx_xvilvl_w(_in7, _in5); \
+ _tmp4_m = __lasx_xvilvl_w(_s1_m, _s0_m); \
+ _tmp5_m = __lasx_xvilvh_w(_s1_m, _s0_m); \
+ _s0_m = __lasx_xvilvh_w(_in6, _in4); \
+ _s1_m = __lasx_xvilvh_w(_in7, _in5); \
+ _tmp6_m = __lasx_xvilvl_w(_s1_m, _s0_m); \
+ _tmp7_m = __lasx_xvilvh_w(_s1_m, _s0_m); \
+ _out0 = __lasx_xvpermi_q(_tmp4_m, _tmp0_m, 0x20); \
+ _out1 = __lasx_xvpermi_q(_tmp5_m, _tmp1_m, 0x20); \
+ _out2 = __lasx_xvpermi_q(_tmp6_m, _tmp2_m, 0x20); \
+ _out3 = __lasx_xvpermi_q(_tmp7_m, _tmp3_m, 0x20); \
+ _out4 = __lasx_xvpermi_q(_tmp4_m, _tmp0_m, 0x31); \
+ _out5 = __lasx_xvpermi_q(_tmp5_m, _tmp1_m, 0x31); \
+ _out6 = __lasx_xvpermi_q(_tmp6_m, _tmp2_m, 0x31); \
+ _out7 = __lasx_xvpermi_q(_tmp7_m, _tmp3_m, 0x31); \
+}
+
+/*
+ * =============================================================================
+ * Description : Transpose input 16x8 byte block
+ * Arguments : Inputs - _in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,
+ * _in8, _in9, _in10, _in11, _in12, _in13, _in14, _in15
+ * (input 16x8 byte block)
+ * Outputs - _out0, _out1, _out2, _out3, _out4, _out5, _out6, _out7
+ * (output 8x16 byte block)
+ * Details : The rows of the matrix become columns, and the columns become rows.
+ * Example : See LASX_TRANSPOSE16x8_H
+ * =============================================================================
+ */
+#define LASX_TRANSPOSE16x8_B(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7, \
+ _in8, _in9, _in10, _in11, _in12, _in13, _in14, _in15, \
+ _out0, _out1, _out2, _out3, _out4, _out5, _out6, _out7) \
+{ \
+ __m256i _tmp0_m, _tmp1_m, _tmp2_m, _tmp3_m; \
+ __m256i _tmp4_m, _tmp5_m, _tmp6_m, _tmp7_m; \
+ \
+ _tmp0_m = __lasx_xvilvl_b(_in2, _in0); \
+ _tmp1_m = __lasx_xvilvl_b(_in3, _in1); \
+ _tmp2_m = __lasx_xvilvl_b(_in6, _in4); \
+ _tmp3_m = __lasx_xvilvl_b(_in7, _in5); \
+ _tmp4_m = __lasx_xvilvl_b(_in10, _in8); \
+ _tmp5_m = __lasx_xvilvl_b(_in11, _in9); \
+ _tmp6_m = __lasx_xvilvl_b(_in14, _in12); \
+ _tmp7_m = __lasx_xvilvl_b(_in15, _in13); \
+ _out0 = __lasx_xvilvl_b(_tmp1_m, _tmp0_m); \
+ _out1 = __lasx_xvilvh_b(_tmp1_m, _tmp0_m); \
+ _out2 = __lasx_xvilvl_b(_tmp3_m, _tmp2_m); \
+ _out3 = __lasx_xvilvh_b(_tmp3_m, _tmp2_m); \
+ _out4 = __lasx_xvilvl_b(_tmp5_m, _tmp4_m); \
+ _out5 = __lasx_xvilvh_b(_tmp5_m, _tmp4_m); \
+ _out6 = __lasx_xvilvl_b(_tmp7_m, _tmp6_m); \
+ _out7 = __lasx_xvilvh_b(_tmp7_m, _tmp6_m); \
+ _tmp0_m = __lasx_xvilvl_w(_out2, _out0); \
+ _tmp2_m = __lasx_xvilvh_w(_out2, _out0); \
+ _tmp4_m = __lasx_xvilvl_w(_out3, _out1); \
+ _tmp6_m = __lasx_xvilvh_w(_out3, _out1); \
+ _tmp1_m = __lasx_xvilvl_w(_out6, _out4); \
+ _tmp3_m = __lasx_xvilvh_w(_out6, _out4); \
+ _tmp5_m = __lasx_xvilvl_w(_out7, _out5); \
+ _tmp7_m = __lasx_xvilvh_w(_out7, _out5); \
+ _out0 = __lasx_xvilvl_d(_tmp1_m, _tmp0_m); \
+ _out1 = __lasx_xvilvh_d(_tmp1_m, _tmp0_m); \
+ _out2 = __lasx_xvilvl_d(_tmp3_m, _tmp2_m); \
+ _out3 = __lasx_xvilvh_d(_tmp3_m, _tmp2_m); \
+ _out4 = __lasx_xvilvl_d(_tmp5_m, _tmp4_m); \
+ _out5 = __lasx_xvilvh_d(_tmp5_m, _tmp4_m); \
+ _out6 = __lasx_xvilvl_d(_tmp7_m, _tmp6_m); \
+ _out7 = __lasx_xvilvh_d(_tmp7_m, _tmp6_m); \
+}
+
+/*
+ * =============================================================================
+ * Description : Transpose input 16x8 byte block
+ * Arguments : Inputs - _in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,
+ * _in8, _in9, _in10, _in11, _in12, _in13, _in14, _in15
+ * (input 16x8 byte block)
+ * Outputs - _out0, _out1, _out2, _out3, _out4, _out5, _out6, _out7
+ * (output 8x16 byte block)
+ * Details : The rows of the matrix become columns, and the columns become rows.
+ * Example : LASX_TRANSPOSE16x8_H
+ * _in0 : 1,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ * _in1 : 2,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ * _in2 : 3,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ * _in3 : 4,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ * _in4 : 5,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ * _in5 : 6,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ * _in6 : 7,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ * _in7 : 8,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ * _in8 : 9,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ * _in9 : 1,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ * _in10 : 0,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ * _in11 : 2,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ * _in12 : 3,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ * _in13 : 7,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ * _in14 : 5,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ * _in15 : 6,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *
+ * _out0 : 1,2,3,4,5,6,7,8,9,1,0,2,3,7,5,6
+ * _out1 : 2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2
+ * _out2 : 3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3
+ * _out3 : 4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4
+ * _out4 : 5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5
+ * _out5 : 6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6
+ * _out6 : 7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7
+ * _out7 : 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8
+ * =============================================================================
+ */
+#define LASX_TRANSPOSE16x8_H(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7, \
+ _in8, _in9, _in10, _in11, _in12, _in13, _in14, _in15, \
+ _out0, _out1, _out2, _out3, _out4, _out5, _out6, _out7) \
+ { \
+ __m256i _tmp0_m, _tmp1_m, _tmp2_m, _tmp3_m; \
+ __m256i _tmp4_m, _tmp5_m, _tmp6_m, _tmp7_m; \
+ __m256i _t0, _t1, _t2, _t3, _t4, _t5, _t6, _t7; \
+ \
+ _tmp0_m = __lasx_xvilvl_h(_in2, _in0); \
+ _tmp1_m = __lasx_xvilvl_h(_in3, _in1); \
+ _tmp2_m = __lasx_xvilvl_h(_in6, _in4); \
+ _tmp3_m = __lasx_xvilvl_h(_in7, _in5); \
+ _tmp4_m = __lasx_xvilvl_h(_in10, _in8); \
+ _tmp5_m = __lasx_xvilvl_h(_in11, _in9); \
+ _tmp6_m = __lasx_xvilvl_h(_in14, _in12); \
+ _tmp7_m = __lasx_xvilvl_h(_in15, _in13); \
+ _t0 = __lasx_xvilvl_h(_tmp1_m, _tmp0_m); \
+ _t1 = __lasx_xvilvh_h(_tmp1_m, _tmp0_m); \
+ _t2 = __lasx_xvilvl_h(_tmp3_m, _tmp2_m); \
+ _t3 = __lasx_xvilvh_h(_tmp3_m, _tmp2_m); \
+ _t4 = __lasx_xvilvl_h(_tmp5_m, _tmp4_m); \
+ _t5 = __lasx_xvilvh_h(_tmp5_m, _tmp4_m); \
+ _t6 = __lasx_xvilvl_h(_tmp7_m, _tmp6_m); \
+ _t7 = __lasx_xvilvh_h(_tmp7_m, _tmp6_m); \
+ _tmp0_m = __lasx_xvilvl_d(_t2, _t0); \
+ _tmp2_m = __lasx_xvilvh_d(_t2, _t0); \
+ _tmp4_m = __lasx_xvilvl_d(_t3, _t1); \
+ _tmp6_m = __lasx_xvilvh_d(_t3, _t1); \
+ _tmp1_m = __lasx_xvilvl_d(_t6, _t4); \
+ _tmp3_m = __lasx_xvilvh_d(_t6, _t4); \
+ _tmp5_m = __lasx_xvilvl_d(_t7, _t5); \
+ _tmp7_m = __lasx_xvilvh_d(_t7, _t5); \
+ _out0 = __lasx_xvpermi_q(_tmp1_m, _tmp0_m, 0x20); \
+ _out1 = __lasx_xvpermi_q(_tmp3_m, _tmp2_m, 0x20); \
+ _out2 = __lasx_xvpermi_q(_tmp5_m, _tmp4_m, 0x20); \
+ _out3 = __lasx_xvpermi_q(_tmp7_m, _tmp6_m, 0x20); \
+ \
+ _tmp0_m = __lasx_xvilvh_h(_in2, _in0); \
+ _tmp1_m = __lasx_xvilvh_h(_in3, _in1); \
+ _tmp2_m = __lasx_xvilvh_h(_in6, _in4); \
+ _tmp3_m = __lasx_xvilvh_h(_in7, _in5); \
+ _tmp4_m = __lasx_xvilvh_h(_in10, _in8); \
+ _tmp5_m = __lasx_xvilvh_h(_in11, _in9); \
+ _tmp6_m = __lasx_xvilvh_h(_in14, _in12); \
+ _tmp7_m = __lasx_xvilvh_h(_in15, _in13); \
+ _t0 = __lasx_xvilvl_h(_tmp1_m, _tmp0_m); \
+ _t1 = __lasx_xvilvh_h(_tmp1_m, _tmp0_m); \
+ _t2 = __lasx_xvilvl_h(_tmp3_m, _tmp2_m); \
+ _t3 = __lasx_xvilvh_h(_tmp3_m, _tmp2_m); \
+ _t4 = __lasx_xvilvl_h(_tmp5_m, _tmp4_m); \
+ _t5 = __lasx_xvilvh_h(_tmp5_m, _tmp4_m); \
+ _t6 = __lasx_xvilvl_h(_tmp7_m, _tmp6_m); \
+ _t7 = __lasx_xvilvh_h(_tmp7_m, _tmp6_m); \
+ _tmp0_m = __lasx_xvilvl_d(_t2, _t0); \
+ _tmp2_m = __lasx_xvilvh_d(_t2, _t0); \
+ _tmp4_m = __lasx_xvilvl_d(_t3, _t1); \
+ _tmp6_m = __lasx_xvilvh_d(_t3, _t1); \
+ _tmp1_m = __lasx_xvilvl_d(_t6, _t4); \
+ _tmp3_m = __lasx_xvilvh_d(_t6, _t4); \
+ _tmp5_m = __lasx_xvilvl_d(_t7, _t5); \
+ _tmp7_m = __lasx_xvilvh_d(_t7, _t5); \
+ _out4 = __lasx_xvpermi_q(_tmp1_m, _tmp0_m, 0x20); \
+ _out5 = __lasx_xvpermi_q(_tmp3_m, _tmp2_m, 0x20); \
+ _out6 = __lasx_xvpermi_q(_tmp5_m, _tmp4_m, 0x20); \
+ _out7 = __lasx_xvpermi_q(_tmp7_m, _tmp6_m, 0x20); \
+}
+
+/*
+ * =============================================================================
+ * Description : Transpose 4x4 block with halfword elements in vectors
+ * Arguments : Inputs - _in0, _in1, _in2, _in3
+ * Outputs - _out0, _out1, _out2, _out3
+ * Return Type - signed halfword
+ * Details : The rows of the matrix become columns, and the columns become rows.
+ * Example : See LASX_TRANSPOSE8x8_H
+ * =============================================================================
+ */
+#define LASX_TRANSPOSE4x4_H(_in0, _in1, _in2, _in3, _out0, _out1, _out2, _out3) \
+{ \
+ __m256i _s0_m, _s1_m; \
+ \
+ _s0_m = __lasx_xvilvl_h(_in1, _in0); \
+ _s1_m = __lasx_xvilvl_h(_in3, _in2); \
+ _out0 = __lasx_xvilvl_w(_s1_m, _s0_m); \
+ _out2 = __lasx_xvilvh_w(_s1_m, _s0_m); \
+ _out1 = __lasx_xvilvh_d(_out0, _out0); \
+ _out3 = __lasx_xvilvh_d(_out2, _out2); \
+}
+
+/*
+ * =============================================================================
+ * Description : Transpose input 8x8 byte block
+ * Arguments : Inputs - _in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7
+ * (input 8x8 byte block)
+ * Outputs - _out0, _out1, _out2, _out3, _out4, _out5, _out6, _out7
+ * (output 8x8 byte block)
+ * Example : See LASX_TRANSPOSE8x8_H
+ * =============================================================================
+ */
+#define LASX_TRANSPOSE8x8_B(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7, _out0, \
+ _out1, _out2, _out3, _out4, _out5, _out6, _out7) \
+{ \
+ __m256i _tmp0_m, _tmp1_m, _tmp2_m, _tmp3_m; \
+ __m256i _tmp4_m, _tmp5_m, _tmp6_m, _tmp7_m; \
+ _tmp0_m = __lasx_xvilvl_b(_in2, _in0); \
+ _tmp1_m = __lasx_xvilvl_b(_in3, _in1); \
+ _tmp2_m = __lasx_xvilvl_b(_in6, _in4); \
+ _tmp3_m = __lasx_xvilvl_b(_in7, _in5); \
+ _tmp4_m = __lasx_xvilvl_b(_tmp1_m, _tmp0_m); \
+ _tmp5_m = __lasx_xvilvh_b(_tmp1_m, _tmp0_m); \
+ _tmp6_m = __lasx_xvilvl_b(_tmp3_m, _tmp2_m); \
+ _tmp7_m = __lasx_xvilvh_b(_tmp3_m, _tmp2_m); \
+ _out0 = __lasx_xvilvl_w(_tmp6_m, _tmp4_m); \
+ _out2 = __lasx_xvilvh_w(_tmp6_m, _tmp4_m); \
+ _out4 = __lasx_xvilvl_w(_tmp7_m, _tmp5_m); \
+ _out6 = __lasx_xvilvh_w(_tmp7_m, _tmp5_m); \
+ _out1 = __lasx_xvbsrl_v(_out0, 8); \
+ _out3 = __lasx_xvbsrl_v(_out2, 8); \
+ _out5 = __lasx_xvbsrl_v(_out4, 8); \
+ _out7 = __lasx_xvbsrl_v(_out6, 8); \
+}
+
+/*
+ * =============================================================================
+ * Description : Transpose 8x8 block with halfword elements in vectors.
+ * Arguments : Inputs - _in0, _in1, ~
+ * Outputs - _out0, _out1, ~
+ * Details : The rows of the matrix become columns, and the columns become rows.
+ * Example : LASX_TRANSPOSE8x8_H
+ * _in0 : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ * _in1 : 8,2,3,4, 5,6,7,8, 8,2,3,4, 5,6,7,8
+ * _in2 : 8,2,3,4, 5,6,7,8, 8,2,3,4, 5,6,7,8
+ * _in3 : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ * _in4 : 9,2,3,4, 5,6,7,8, 9,2,3,4, 5,6,7,8
+ * _in5 : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ * _in6 : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ * _in7 : 9,2,3,4, 5,6,7,8, 9,2,3,4, 5,6,7,8
+ *
+ * _out0 : 1,8,8,1, 9,1,1,9, 1,8,8,1, 9,1,1,9
+ * _out1 : 2,2,2,2, 2,2,2,2, 2,2,2,2, 2,2,2,2
+ * _out2 : 3,3,3,3, 3,3,3,3, 3,3,3,3, 3,3,3,3
+ * _out3 : 4,4,4,4, 4,4,4,4, 4,4,4,4, 4,4,4,4
+ * _out4 : 5,5,5,5, 5,5,5,5, 5,5,5,5, 5,5,5,5
+ * _out5 : 6,6,6,6, 6,6,6,6, 6,6,6,6, 6,6,6,6
+ * _out6 : 7,7,7,7, 7,7,7,7, 7,7,7,7, 7,7,7,7
+ * _out7 : 8,8,8,8, 8,8,8,8, 8,8,8,8, 8,8,8,8
+ * =============================================================================
+ */
+#define LASX_TRANSPOSE8x8_H(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7, _out0, \
+ _out1, _out2, _out3, _out4, _out5, _out6, _out7) \
+{ \
+ __m256i _s0_m, _s1_m; \
+ __m256i _tmp0_m, _tmp1_m, _tmp2_m, _tmp3_m; \
+ __m256i _tmp4_m, _tmp5_m, _tmp6_m, _tmp7_m; \
+ \
+ _s0_m = __lasx_xvilvl_h(_in6, _in4); \
+ _s1_m = __lasx_xvilvl_h(_in7, _in5); \
+ _tmp0_m = __lasx_xvilvl_h(_s1_m, _s0_m); \
+ _tmp1_m = __lasx_xvilvh_h(_s1_m, _s0_m); \
+ _s0_m = __lasx_xvilvh_h(_in6, _in4); \
+ _s1_m = __lasx_xvilvh_h(_in7, _in5); \
+ _tmp2_m = __lasx_xvilvl_h(_s1_m, _s0_m); \
+ _tmp3_m = __lasx_xvilvh_h(_s1_m, _s0_m); \
+ \
+ _s0_m = __lasx_xvilvl_h(_in2, _in0); \
+ _s1_m = __lasx_xvilvl_h(_in3, _in1); \
+ _tmp4_m = __lasx_xvilvl_h(_s1_m, _s0_m); \
+ _tmp5_m = __lasx_xvilvh_h(_s1_m, _s0_m); \
+ _s0_m = __lasx_xvilvh_h(_in2, _in0); \
+ _s1_m = __lasx_xvilvh_h(_in3, _in1); \
+ _tmp6_m = __lasx_xvilvl_h(_s1_m, _s0_m); \
+ _tmp7_m = __lasx_xvilvh_h(_s1_m, _s0_m); \
+ \
+ _out0 = __lasx_xvpickev_d(_tmp0_m, _tmp4_m); \
+ _out2 = __lasx_xvpickev_d(_tmp1_m, _tmp5_m); \
+ _out4 = __lasx_xvpickev_d(_tmp2_m, _tmp6_m); \
+ _out6 = __lasx_xvpickev_d(_tmp3_m, _tmp7_m); \
+ _out1 = __lasx_xvpickod_d(_tmp0_m, _tmp4_m); \
+ _out3 = __lasx_xvpickod_d(_tmp1_m, _tmp5_m); \
+ _out5 = __lasx_xvpickod_d(_tmp2_m, _tmp6_m); \
+ _out7 = __lasx_xvpickod_d(_tmp3_m, _tmp7_m); \
+}
+
+/*
+ * =============================================================================
+ * Description : Butterfly of 4 input vectors
+ * Arguments : Inputs - _in0, _in1, _in2, _in3
+ * Outputs - _out0, _out1, _out2, _out3
+ * Details : Butterfly operation
+ * Example : LASX_BUTTERFLY_4
+ * _out0 = _in0 + _in3;
+ * _out1 = _in1 + _in2;
+ * _out2 = _in1 - _in2;
+ * _out3 = _in0 - _in3;
+ * =============================================================================
+ */
+#define LASX_BUTTERFLY_4_B(_in0, _in1, _in2, _in3, _out0, _out1, _out2, _out3) \
+{ \
+ _out0 = __lasx_xvadd_b(_in0, _in3); \
+ _out1 = __lasx_xvadd_b(_in1, _in2); \
+ _out2 = __lasx_xvsub_b(_in1, _in2); \
+ _out3 = __lasx_xvsub_b(_in0, _in3); \
+}
+#define LASX_BUTTERFLY_4_H(_in0, _in1, _in2, _in3, _out0, _out1, _out2, _out3) \
+{ \
+ _out0 = __lasx_xvadd_h(_in0, _in3); \
+ _out1 = __lasx_xvadd_h(_in1, _in2); \
+ _out2 = __lasx_xvsub_h(_in1, _in2); \
+ _out3 = __lasx_xvsub_h(_in0, _in3); \
+}
+#define LASX_BUTTERFLY_4_W(_in0, _in1, _in2, _in3, _out0, _out1, _out2, _out3) \
+{ \
+ _out0 = __lasx_xvadd_w(_in0, _in3); \
+ _out1 = __lasx_xvadd_w(_in1, _in2); \
+ _out2 = __lasx_xvsub_w(_in1, _in2); \
+ _out3 = __lasx_xvsub_w(_in0, _in3); \
+}
+#define LASX_BUTTERFLY_4_D(_in0, _in1, _in2, _in3, _out0, _out1, _out2, _out3) \
+{ \
+ _out0 = __lasx_xvadd_d(_in0, _in3); \
+ _out1 = __lasx_xvadd_d(_in1, _in2); \
+ _out2 = __lasx_xvsub_d(_in1, _in2); \
+ _out3 = __lasx_xvsub_d(_in0, _in3); \
+}
+
+/*
+ * =============================================================================
+ * Description : Butterfly of 8 input vectors
+ * Arguments : Inputs - _in0, _in1, _in2, _in3, ~
+ * Outputs - _out0, _out1, _out2, _out3, ~
+ * Details : Butterfly operation
+ * Example : LASX_BUTTERFLY_8
+ * _out0 = _in0 + _in7;
+ * _out1 = _in1 + _in6;
+ * _out2 = _in2 + _in5;
+ * _out3 = _in3 + _in4;
+ * _out4 = _in3 - _in4;
+ * _out5 = _in2 - _in5;
+ * _out6 = _in1 - _in6;
+ * _out7 = _in0 - _in7;
+ * =============================================================================
+ */
+#define LASX_BUTTERFLY_8_B(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7, \
+ _out0, _out1, _out2, _out3, _out4, _out5, _out6, _out7)\
+{ \
+ _out0 = __lasx_xvadd_b(_in0, _in7); \
+ _out1 = __lasx_xvadd_b(_in1, _in6); \
+ _out2 = __lasx_xvadd_b(_in2, _in5); \
+ _out3 = __lasx_xvadd_b(_in3, _in4); \
+ _out4 = __lasx_xvsub_b(_in3, _in4); \
+ _out5 = __lasx_xvsub_b(_in2, _in5); \
+ _out6 = __lasx_xvsub_b(_in1, _in6); \
+ _out7 = __lasx_xvsub_b(_in0, _in7); \
+}
+
+#define LASX_BUTTERFLY_8_H(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7, \
+ _out0, _out1, _out2, _out3, _out4, _out5, _out6, _out7)\
+{ \
+ _out0 = __lasx_xvadd_h(_in0, _in7); \
+ _out1 = __lasx_xvadd_h(_in1, _in6); \
+ _out2 = __lasx_xvadd_h(_in2, _in5); \
+ _out3 = __lasx_xvadd_h(_in3, _in4); \
+ _out4 = __lasx_xvsub_h(_in3, _in4); \
+ _out5 = __lasx_xvsub_h(_in2, _in5); \
+ _out6 = __lasx_xvsub_h(_in1, _in6); \
+ _out7 = __lasx_xvsub_h(_in0, _in7); \
+}
+
+#define LASX_BUTTERFLY_8_W(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7, \
+ _out0, _out1, _out2, _out3, _out4, _out5, _out6, _out7)\
+{ \
+ _out0 = __lasx_xvadd_w(_in0, _in7); \
+ _out1 = __lasx_xvadd_w(_in1, _in6); \
+ _out2 = __lasx_xvadd_w(_in2, _in5); \
+ _out3 = __lasx_xvadd_w(_in3, _in4); \
+ _out4 = __lasx_xvsub_w(_in3, _in4); \
+ _out5 = __lasx_xvsub_w(_in2, _in5); \
+ _out6 = __lasx_xvsub_w(_in1, _in6); \
+ _out7 = __lasx_xvsub_w(_in0, _in7); \
+}
+
+#define LASX_BUTTERFLY_8_D(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7, \
+ _out0, _out1, _out2, _out3, _out4, _out5, _out6, _out7)\
+{ \
+ _out0 = __lasx_xvadd_d(_in0, _in7); \
+ _out1 = __lasx_xvadd_d(_in1, _in6); \
+ _out2 = __lasx_xvadd_d(_in2, _in5); \
+ _out3 = __lasx_xvadd_d(_in3, _in4); \
+ _out4 = __lasx_xvsub_d(_in3, _in4); \
+ _out5 = __lasx_xvsub_d(_in2, _in5); \
+ _out6 = __lasx_xvsub_d(_in1, _in6); \
+ _out7 = __lasx_xvsub_d(_in0, _in7); \
+}
+
+#endif //LASX
+
+/*
+ * =============================================================================
+ * Description : Print out elements in vector.
+ * Arguments : Inputs - RTYPE, _element_num, _in0, _enter
+ * Outputs -
+ * Details : Print out '_element_num' elements in 'RTYPE' vector '_in0', if
+ * '_enter' is TRUE, prefix "\nVP:" will be added first.
+ * Example : VECT_PRINT(v4i32,4,in0,1); // in0: 1,2,3,4
+ * VP:1,2,3,4,
+ * =============================================================================
+ */
+#define VECT_PRINT(RTYPE, element_num, in0, enter) \
+{ \
+ RTYPE _tmp0 = (RTYPE)in0; \
+ int _i = 0; \
+ if (enter) \
+ printf("\nVP:"); \
+ for(_i = 0; _i < element_num; _i++) \
+ printf("%d,",_tmp0[_i]); \
+}
+
+#endif /* LOONGSON_INTRINSICS_H */
+#endif /* AVUTIL_LOONGARCH_LOONGSON_INTRINSICS_H */
--
2.20.1
More information about the ffmpeg-devel
mailing list