[FFmpeg-devel] [PATCH 3/3] avcodec/hevc: Add wasm simd128 idct

Zhao Zhili quinkblack at foxmail.com
Wed Nov 13 16:16:36 EET 2024


From: Zhao Zhili <zhilizhao at tencent.com>

We don't use intrinsics normally for performance reasons. However,
WASM isn't machine instruction. I suspect that in this scenario,
handwritten assembly may not bring much performance compared with
intrinsics. And I don't want to dig into the WASM ABI. So this
patch is implemented with WASM simd128 intrinsics. Maybe we can get
a better performance with the upcoming relaxed SIMD proposal.

Benchmark with wasmtime:
wasmtime checkasm --test=hevc_idct --benchmark

hevc_idct_4x4_8_c:                                      20.8 ( 1.00x)
hevc_idct_4x4_8_simd128:                                10.2 ( 2.02x)
hevc_idct_4x4_10_c:                                     21.0 ( 1.00x)
hevc_idct_4x4_10_simd128:                               10.2 ( 2.05x)
hevc_idct_8x8_8_c:                                     198.0 ( 1.00x)
hevc_idct_8x8_8_simd128:                                41.5 ( 4.77x)
hevc_idct_8x8_10_c:                                    197.8 ( 1.00x)
hevc_idct_8x8_10_simd128:                               41.8 ( 4.74x)
hevc_idct_16x16_8_c:                                  1385.2 ( 1.00x)
hevc_idct_16x16_8_simd128:                             291.8 ( 4.75x)
hevc_idct_16x16_10_c:                                 1375.0 ( 1.00x)
hevc_idct_16x16_10_simd128:                            281.2 ( 4.89x)
hevc_idct_32x32_8_c:                                 13479.0 ( 1.00x)
hevc_idct_32x32_8_simd128:                            1541.5 ( 8.74x)
hevc_idct_32x32_10_c:                                13448.0 ( 1.00x)
hevc_idct_32x32_10_simd128:                           1541.5 ( 8.72x)

Decoding test:

1. Without the optimization:
$ wasm ./ffmpeg -cpuflags 0 -i 265.mp4 -an -t 120 -f null - -benchmark
frame= 2880 fps= 94 q=-0.0 Lsize=N/A time=00:02:00.00 bitrate=N/A speed=3.92x

2. With the optimization:
$ wasm ./ffmpeg -i 265.mp4 -an -t 120 -f null - -benchmark
frame= 2880 fps=110 q=-0.0 Lsize=N/A time=00:02:00.00 bitrate=N/A speed=4.58x
---
 libavcodec/Makefile             |   1 +
 libavcodec/hevc/dsp.c           |   2 +
 libavcodec/hevc/dsp.h           |   1 +
 libavcodec/wasm/hevc/Makefile   |   3 +
 libavcodec/wasm/hevc/dsp_init.c |  48 ++
 libavcodec/wasm/hevc/idct.c     | 869 ++++++++++++++++++++++++++++++++
 libavcodec/wasm/hevc/idct.h     |  36 ++
 7 files changed, 960 insertions(+)
 create mode 100644 libavcodec/wasm/hevc/Makefile
 create mode 100644 libavcodec/wasm/hevc/dsp_init.c
 create mode 100644 libavcodec/wasm/hevc/idct.c
 create mode 100644 libavcodec/wasm/hevc/idct.h

diff --git a/libavcodec/Makefile b/libavcodec/Makefile
index 676ff542af..ea9572e527 100644
--- a/libavcodec/Makefile
+++ b/libavcodec/Makefile
@@ -71,6 +71,7 @@ include $(SRC_PATH)/libavcodec/hevc/Makefile
 include $(SRC_PATH)/libavcodec/opus/Makefile
 include $(SRC_PATH)/libavcodec/vvc/Makefile
 include $(SRC_PATH)/libavcodec/vulkan/Makefile
+-include $(SRC_PATH)/libavcodec/$(ARCH)/hevc/Makefile
 -include $(SRC_PATH)/libavcodec/$(ARCH)/vvc/Makefile
 OBJS-$(CONFIG_AANDCTTABLES)            += aandcttab.o
 OBJS-$(CONFIG_AC3DSP)                  += ac3dsp.o ac3.o ac3tab.o
diff --git a/libavcodec/hevc/dsp.c b/libavcodec/hevc/dsp.c
index 60f059292c..2b347781df 100644
--- a/libavcodec/hevc/dsp.c
+++ b/libavcodec/hevc/dsp.c
@@ -265,6 +265,8 @@ int i = 0;
     ff_hevc_dsp_init_arm(hevcdsp, bit_depth);
 #elif ARCH_PPC
     ff_hevc_dsp_init_ppc(hevcdsp, bit_depth);
+#elif ARCH_WASM
+    ff_hevc_dsp_init_wasm(hevcdsp, bit_depth);
 #elif ARCH_X86
     ff_hevc_dsp_init_x86(hevcdsp, bit_depth);
 #elif ARCH_MIPS
diff --git a/libavcodec/hevc/dsp.h b/libavcodec/hevc/dsp.h
index 02b8e0e8e2..4277d695ba 100644
--- a/libavcodec/hevc/dsp.h
+++ b/libavcodec/hevc/dsp.h
@@ -133,6 +133,7 @@ extern const int8_t ff_hevc_qpel_filters[4][16];
 void ff_hevc_dsp_init_aarch64(HEVCDSPContext *c, const int bit_depth);
 void ff_hevc_dsp_init_arm(HEVCDSPContext *c, const int bit_depth);
 void ff_hevc_dsp_init_ppc(HEVCDSPContext *c, const int bit_depth);
+void ff_hevc_dsp_init_wasm(HEVCDSPContext *c, const int bit_depth);
 void ff_hevc_dsp_init_x86(HEVCDSPContext *c, const int bit_depth);
 void ff_hevc_dsp_init_mips(HEVCDSPContext *c, const int bit_depth);
 void ff_hevc_dsp_init_loongarch(HEVCDSPContext *c, const int bit_depth);
diff --git a/libavcodec/wasm/hevc/Makefile b/libavcodec/wasm/hevc/Makefile
new file mode 100644
index 0000000000..132daa3106
--- /dev/null
+++ b/libavcodec/wasm/hevc/Makefile
@@ -0,0 +1,3 @@
+OBJS-$(CONFIG_HEVC_DECODER)        += wasm/hevc/dsp_init.o
+
+SIMD128-OBJS-$(CONFIG_HEVC_DECODER) += wasm/hevc/idct.o
diff --git a/libavcodec/wasm/hevc/dsp_init.c b/libavcodec/wasm/hevc/dsp_init.c
new file mode 100644
index 0000000000..4350965925
--- /dev/null
+++ b/libavcodec/wasm/hevc/dsp_init.c
@@ -0,0 +1,48 @@
+/*
+ * Copyright (c) 2024 Zhao Zhili
+ *
+ * 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 <stdint.h>
+#include <wasm_simd128.h>
+
+#include "libavutil/cpu_internal.h"
+#include "libavcodec/hevc/dsp.h"
+#include "libavcodec/wasm/hevc/idct.h"
+
+av_cold void ff_hevc_dsp_init_wasm(HEVCDSPContext *c, const int bit_depth)
+{
+    int cpu_flags = av_get_cpu_flags();
+
+    if (!CPUEXT(cpu_flags, SIMD128))
+        return;
+
+#if HAVE_SIMD128
+    if (bit_depth == 8) {
+        c->idct[0] = ff_hevc_idct_4x4_8_simd128;
+        c->idct[1] = ff_hevc_idct_8x8_8_simd128;
+        c->idct[2] = ff_hevc_idct_16x16_8_simd128;
+        c->idct[3] = ff_hevc_idct_32x32_8_simd128;
+    } else if (bit_depth == 10) {
+        c->idct[0] = ff_hevc_idct_4x4_10_simd128;
+        c->idct[1] = ff_hevc_idct_8x8_10_simd128;
+        c->idct[2] = ff_hevc_idct_16x16_10_simd128;
+        c->idct[3] = ff_hevc_idct_32x32_10_simd128;
+    }
+#endif
+}
diff --git a/libavcodec/wasm/hevc/idct.c b/libavcodec/wasm/hevc/idct.c
new file mode 100644
index 0000000000..828fc26a49
--- /dev/null
+++ b/libavcodec/wasm/hevc/idct.c
@@ -0,0 +1,869 @@
+/*
+ * Copyright (c) 2024 Zhao Zhili
+ *
+ * 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 "libavcodec/wasm/hevc/idct.h"
+
+#include <wasm_simd128.h>
+
+#include "libavutil/mem_internal.h"
+
+static const int8_t transform[] = {
+    64, 83, 64, 36, 89, 75, 50, 18,
+    90, 87, 80, 70, 57, 43, 25, 9,
+    90, 90, 88, 85, 82, 78, 73, 67,
+    61, 54, 46, 38, 31, 22, 13, 4,
+};
+
+static inline void transpose_4x8h(v128_t *src)
+{
+    v128_t t0 = wasm_i16x8_shuffle(src[0], src[1], 0, 8, 2, 10, 4, 12, 6, 14);
+    v128_t t1 = wasm_i16x8_shuffle(src[0], src[1], 1, 9, 3, 11, 5, 13, 7, 15);
+    v128_t t2 = wasm_i16x8_shuffle(src[2], src[3], 0, 8, 2, 10, 4, 12, 6, 14);
+    v128_t t3 = wasm_i16x8_shuffle(src[2], src[3], 1, 9, 3, 11, 5, 13, 7, 15);
+
+    src[0] = wasm_i32x4_shuffle(t0, t2, 0, 4, 2, 6);
+    src[2] = wasm_i32x4_shuffle(t0, t2, 1, 5, 3, 7);
+    src[1] = wasm_i32x4_shuffle(t1, t3, 0, 4, 2, 6);
+    src[3] = wasm_i32x4_shuffle(t1, t3, 1, 5, 3, 7);
+}
+
+static inline void transpose_8x8h(v128_t *src)
+{
+    transpose_4x8h(src);
+    transpose_4x8h(&src[4]);
+}
+
+static inline void tr_4x4(v128_t *src, v128_t *trans, int shift)
+{
+    v128_t tmp[4];
+    v128_t add = wasm_i32x4_splat(1 << (shift - 1));
+    v128_t e0 = wasm_i32x4_extmul_low_i16x8(src[0], trans[0]);
+    v128_t e1 = wasm_i32x4_extmul_low_i16x8(src[0], trans[0]);
+    v128_t o0 = wasm_i32x4_extmul_low_i16x8(src[1], trans[1]);
+    v128_t o1 = wasm_i32x4_extmul_low_i16x8(src[1], trans[3]);
+
+    tmp[0] = wasm_i32x4_extmul_low_i16x8(src[2], trans[0]);
+    tmp[1] = wasm_i32x4_extmul_low_i16x8(src[2], trans[0]);
+    tmp[2] = wasm_i32x4_extmul_low_i16x8(src[3], trans[3]);
+    tmp[3] = wasm_i32x4_extmul_low_i16x8(src[3], trans[1]);
+    e0 = wasm_i32x4_add(e0, tmp[0]);
+    e1 = wasm_i32x4_sub(e1, tmp[1]);
+    o0 = wasm_i32x4_add(o0, tmp[2]);
+    o1 = wasm_i32x4_sub(o1, tmp[3]);
+
+    tmp[0] = wasm_i32x4_add(e0, o0);
+    tmp[1] = wasm_i32x4_sub(e0, o0);
+    tmp[2] = wasm_i32x4_add(e1, o1);
+    tmp[3] = wasm_i32x4_sub(e1, o1);
+
+    tmp[0] = wasm_i32x4_add(tmp[0], add);
+    tmp[1] = wasm_i32x4_add(tmp[1], add);
+    tmp[2] = wasm_i32x4_add(tmp[2], add);
+    tmp[3] = wasm_i32x4_add(tmp[3], add);
+    tmp[0] = wasm_i32x4_shr(tmp[0], shift);
+    tmp[1] = wasm_i32x4_shr(tmp[1], shift);
+    tmp[2] = wasm_i32x4_shr(tmp[2], shift);
+    tmp[3] = wasm_i32x4_shr(tmp[3], shift);
+
+    src[0] = wasm_i16x8_narrow_i32x4(tmp[0], tmp[0]);
+    src[3] = wasm_i16x8_narrow_i32x4(tmp[1], tmp[1]);
+    src[1] = wasm_i16x8_narrow_i32x4(tmp[2], tmp[2]);
+    src[2] = wasm_i16x8_narrow_i32x4(tmp[3], tmp[3]);
+}
+
+static void idct_4x4(int16_t *coeffs, int bit_depth)
+{
+    v128_t src[4];
+    v128_t trans[4];
+
+    src[0] = wasm_v128_load64_zero(&coeffs[0]);
+    src[1] = wasm_v128_load64_zero(&coeffs[4]);
+    src[2] = wasm_v128_load64_zero(&coeffs[8]);
+    src[3] = wasm_v128_load64_zero(&coeffs[12]);
+
+    trans[0] = wasm_i16x8_const_splat(transform[0]);
+    trans[1] = wasm_i16x8_const_splat(transform[1]);
+    trans[2] = wasm_i16x8_const_splat(transform[2]);
+    trans[3] = wasm_i16x8_const_splat(transform[3]);
+
+    tr_4x4(src, trans, 7);
+    transpose_4x8h(src);
+
+    tr_4x4(src, trans, 20 - bit_depth);
+    transpose_4x8h(src);
+
+    src[0] = wasm_i64x2_shuffle(src[0], src[1], 0, 2);
+    src[2] = wasm_i64x2_shuffle(src[2], src[3], 0, 2);
+    wasm_v128_store(&coeffs[0], src[0]);
+    wasm_v128_store(&coeffs[8], src[2]);
+}
+
+void ff_hevc_idct_4x4_8_simd128(int16_t *coeffs, int col_limit)
+{
+    idct_4x4(coeffs, 8);
+}
+
+void ff_hevc_idct_4x4_10_simd128(int16_t *coeffs, int col_limit)
+{
+    idct_4x4(coeffs, 10);
+}
+
+static inline void shift_narrow_low(v128_t src, v128_t *dst, v128_t add, int shift)
+{
+    src = wasm_i32x4_add(src, add);
+    src = wasm_i32x4_shr(src, shift);
+    *dst = wasm_i64x2_shuffle(wasm_i16x8_narrow_i32x4(src, src), *dst, 0, 3);
+}
+
+static inline void shift_narrow_high(v128_t src, v128_t *dst, v128_t add, int shift)
+{
+    src = wasm_i32x4_add(src, add);
+    src = wasm_i32x4_shr(src, shift);
+    *dst = wasm_i64x2_shuffle(wasm_i16x8_narrow_i32x4(src, src), *dst, 2, 0);
+}
+
+#define tr_4x4_8(in0, in1, in2, in3, dst0, dst1, dst2, dst3, trans, half0, half1)   \
+    do {                                                                            \
+        v128_t e0, e1, o0, o1;                                                      \
+        v128_t tmp[4];                                                              \
+                                                                                    \
+        e0 = wasm_i32x4_extmul_ ## half0 ## _i16x8(in0, trans[0]);                  \
+        e1 = e0;                                                                    \
+        o0 = wasm_i32x4_extmul_ ## half0 ## _i16x8(in1, trans[1]);                  \
+        o1 = wasm_i32x4_extmul_ ## half0 ## _i16x8(in1, trans[3]);                  \
+                                                                                    \
+        tmp[0] = wasm_i32x4_extmul_ ## half1 ## _i16x8(in2, trans[0]);              \
+        tmp[1] = wasm_i32x4_extmul_ ## half1 ## _i16x8(in2, trans[0]);              \
+        tmp[2] = wasm_i32x4_extmul_ ## half1 ## _i16x8(in3, trans[3]);              \
+        tmp[3] = wasm_i32x4_extmul_ ## half1 ## _i16x8(in3, trans[1]);              \
+        e0 = wasm_i32x4_add(e0, tmp[0]);                                            \
+        e1 = wasm_i32x4_sub(e1, tmp[1]);                                            \
+        o0 = wasm_i32x4_add(o0, tmp[2]);                                            \
+        o1 = wasm_i32x4_sub(o1, tmp[3]);                                            \
+        dst0 = wasm_i32x4_add(e0, o0);                                              \
+        dst1 = wasm_i32x4_add(e1, o1);                                              \
+        dst2 = wasm_i32x4_sub(e1, o1);                                              \
+        dst3 = wasm_i32x4_sub(e0, o0);                                              \
+    } while (0)
+
+#define tr_8x4(src0, src1, half0, half1, trans, shift)                                          \
+    do {                                                                                        \
+        v128_t v24, v25, v26, v27, v28, v29, v30, v31;                                          \
+        v128_t add = wasm_i32x4_splat(1 << (shift - 1));                                        \
+                                                                                                \
+        tr_4x4_8(src0[0], src0[2], src1[0], src1[2], v24, v25, v26, v27, trans, half0, half1);  \
+                                                                                                \
+        v30 = wasm_i32x4_extmul_ ## half0 ## _i16x8(src0[1], trans[6]);                         \
+        v28 = wasm_i32x4_extmul_ ## half0 ## _i16x8(src0[1], trans[4]);                         \
+        v29 = wasm_i32x4_extmul_ ## half0 ## _i16x8(src0[1], trans[5]);                         \
+        v30 = wasm_i32x4_sub(v30, wasm_i32x4_extmul_ ## half0 ## _i16x8(src0[3], trans[4]));    \
+        v28 = wasm_i32x4_add(v28, wasm_i32x4_extmul_ ## half0 ## _i16x8(src0[3], trans[5]));    \
+        v29 = wasm_i32x4_sub(v29, wasm_i32x4_extmul_ ## half0 ## _i16x8(src0[3], trans[7]));    \
+                                                                                                \
+        v30 = wasm_i32x4_add(v30, wasm_i32x4_extmul_ ## half1 ## _i16x8(src1[1], trans[7]));    \
+        v28 = wasm_i32x4_add(v28, wasm_i32x4_extmul_ ## half1 ## _i16x8(src1[1], trans[6]));    \
+        v29 = wasm_i32x4_sub(v29, wasm_i32x4_extmul_ ## half1 ## _i16x8(src1[1], trans[4]));    \
+                                                                                                \
+        v30 = wasm_i32x4_add(v30, wasm_i32x4_extmul_ ## half1 ## _i16x8(src1[3], trans[5]));    \
+        v28 = wasm_i32x4_add(v28, wasm_i32x4_extmul_ ## half1 ## _i16x8(src1[3], trans[7]));    \
+        v29 = wasm_i32x4_sub(v29, wasm_i32x4_extmul_ ## half1 ## _i16x8(src1[3], trans[6]));    \
+                                                                                                \
+        v31 = wasm_i32x4_add(v26, v30);                                                         \
+        v26 = wasm_i32x4_sub(v26, v30);                                                         \
+        shift_narrow_ ## half0 (v31, &src0[2], add, shift);                                     \
+        v31 = wasm_i32x4_extmul_ ## half0 ## _i16x8(src0[1], trans[7]);                         \
+        v31 = wasm_i32x4_sub(v31, wasm_i32x4_extmul_ ## half0 ## _i16x8(src0[3], trans[6]));    \
+        v31 = wasm_i32x4_add(v31, wasm_i32x4_extmul_ ## half1 ## _i16x8(src1[1], trans[5]));    \
+        v31 = wasm_i32x4_sub(v31, wasm_i32x4_extmul_ ## half1 ## _i16x8(src1[3], trans[4]));    \
+        shift_narrow_ ## half1 (v26, &src1[1], add, shift);                                     \
+        v26 = wasm_i32x4_add(v24, v28);                                                         \
+        v24 = wasm_i32x4_sub(v24, v28);                                                         \
+        v28 = wasm_i32x4_add(v25, v29);                                                         \
+        v25 = wasm_i32x4_sub(v25, v29);                                                         \
+        v30 = wasm_i32x4_add(v27, v31);                                                         \
+        v27 = wasm_i32x4_sub(v27, v31);                                                         \
+        shift_narrow_ ## half0 (v26, &src0[0], add, shift);                                     \
+        shift_narrow_ ## half1 (v24, &src1[3], add, shift);                                     \
+        shift_narrow_ ## half0 (v28, &src0[1], add, shift);                                     \
+        shift_narrow_ ## half1 (v25, &src1[2], add, shift);                                     \
+        shift_narrow_ ## half0 (v30, &src0[3], add, shift);                                     \
+        shift_narrow_ ## half1 (v27, &src1[0], add, shift);                                     \
+    } while (0)
+
+static void idct_8x8(int16_t *coeffs, int bit_depth)
+{
+    v128_t src[8];
+    v128_t trans[8];
+    v128_t *src1;
+    int shift1 = 7;
+    int shift2 = 20 - bit_depth;
+
+    src[0] = wasm_v128_load(coeffs + 0 * 8);
+    src[1] = wasm_v128_load(coeffs + 1 * 8);
+    src[2] = wasm_v128_load(coeffs + 2 * 8);
+    src[3] = wasm_v128_load(coeffs + 3 * 8);
+    src[4] = wasm_v128_load(coeffs + 4 * 8);
+    src[5] = wasm_v128_load(coeffs + 5 * 8);
+    src[6] = wasm_v128_load(coeffs + 6 * 8);
+    src[7] = wasm_v128_load(coeffs + 7 * 8);
+
+    trans[0] = wasm_i16x8_const_splat(transform[0]);
+    trans[1] = wasm_i16x8_const_splat(transform[1]);
+    trans[2] = wasm_i16x8_const_splat(transform[2]);
+    trans[3] = wasm_i16x8_const_splat(transform[3]);
+    trans[4] = wasm_i16x8_const_splat(transform[4]);
+    trans[5] = wasm_i16x8_const_splat(transform[5]);
+    trans[6] = wasm_i16x8_const_splat(transform[6]);
+    trans[7] = wasm_i16x8_const_splat(transform[7]);
+
+    src1 = &src[4];
+    tr_8x4(src, src1, low, low, trans, shift1);
+    tr_8x4(src, src1, high, high, trans, shift1);
+    transpose_8x8h(src);
+    tr_8x4(src, src, low, high, trans, shift2);
+    tr_8x4(src1, src1, low, high, trans, shift2);
+    transpose_8x8h(src);
+
+    wasm_v128_store(&coeffs[0 * 8], src[0]);
+    wasm_v128_store(&coeffs[1 * 8], src[1]);
+    wasm_v128_store(&coeffs[2 * 8], src[2]);
+    wasm_v128_store(&coeffs[3 * 8], src[3]);
+    wasm_v128_store(&coeffs[4 * 8], src[4]);
+    wasm_v128_store(&coeffs[5 * 8], src[5]);
+    wasm_v128_store(&coeffs[6 * 8], src[6]);
+    wasm_v128_store(&coeffs[7 * 8], src[7]);
+}
+
+void ff_hevc_idct_8x8_8_simd128(int16_t *coeffs, int col_limit)
+{
+    idct_8x8(coeffs, 8);
+}
+
+void ff_hevc_idct_8x8_10_simd128(int16_t *coeffs, int col_limit)
+{
+    idct_8x8(coeffs, 10);
+}
+
+#define load16(x1, x3, x2, in0, in1, in2, in3)  \
+    in0 = wasm_v128_load64_zero(x1);            \
+    in0 = wasm_v128_load64_lane(x3, in0, 1);    \
+    x1 += x2;                                   \
+    x3 += x2;                                   \
+    in1 = wasm_v128_load64_zero(x1);            \
+    in1 = wasm_v128_load64_lane(x3, in1, 1);    \
+    x1 += x2;                                   \
+    x3 += x2;                                   \
+    in2 = wasm_v128_load64_zero(x1);            \
+    in2 = wasm_v128_load64_lane(x3, in2, 1);    \
+    x1 += x2;                                   \
+    x3 += x2;                                   \
+    in3 = wasm_v128_load64_zero(x1);            \
+    in3 = wasm_v128_load64_lane(x3, in3, 1);    \
+    x1 += x2;                                   \
+    x3 += x2;                                   \
+
+#define bufferfly(e, o, p, m)   \
+    p = wasm_i32x4_add(e, o);   \
+    m = wasm_i32x4_sub(e, o);   \
+
+static void tr16_8x4(v128_t in0, v128_t in1, v128_t in2, v128_t in3,
+                     const v128_t *trans, char *sp, int offset)
+{
+    v128_t v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31;
+
+    tr_4x4_8(in0, in1, in2, in3, v24, v25, v26, v27, trans, low, low);
+
+    v28 = wasm_i32x4_extmul_high_i16x8(in0, trans[4]);
+    v29 = wasm_i32x4_extmul_high_i16x8(in0, trans[5]);
+    v30 = wasm_i32x4_extmul_high_i16x8(in0, trans[6]);
+    v31 = wasm_i32x4_extmul_high_i16x8(in0, trans[7]);
+    v28 = wasm_i32x4_add(v28, wasm_i32x4_extmul_high_i16x8(in1, trans[5]));
+    v29 = wasm_i32x4_sub(v29, wasm_i32x4_extmul_high_i16x8(in1, trans[7]));
+    v30 = wasm_i32x4_sub(v30, wasm_i32x4_extmul_high_i16x8(in1, trans[4]));
+    v31 = wasm_i32x4_sub(v31, wasm_i32x4_extmul_high_i16x8(in1, trans[6]));
+
+    v28 = wasm_i32x4_add(v28, wasm_i32x4_extmul_high_i16x8(in2, trans[6]));
+    v29 = wasm_i32x4_sub(v29, wasm_i32x4_extmul_high_i16x8(in2, trans[4]));
+    v30 = wasm_i32x4_add(v30, wasm_i32x4_extmul_high_i16x8(in2, trans[7]));
+    v31 = wasm_i32x4_add(v31, wasm_i32x4_extmul_high_i16x8(in2, trans[5]));
+
+    v28 = wasm_i32x4_add(v28, wasm_i32x4_extmul_high_i16x8(in3, trans[7]));
+    v29 = wasm_i32x4_sub(v29, wasm_i32x4_extmul_high_i16x8(in3, trans[6]));
+    v30 = wasm_i32x4_add(v30, wasm_i32x4_extmul_high_i16x8(in3, trans[5]));
+    v31 = wasm_i32x4_sub(v31, wasm_i32x4_extmul_high_i16x8(in3, trans[4]));
+
+    bufferfly(v24, v28, v16, v23);
+    bufferfly(v25, v29, v17, v22);
+    bufferfly(v26, v30, v18, v21);
+    bufferfly(v27, v31, v19, v20);
+
+    sp += offset;
+    wasm_v128_store(sp, v16); sp += 16;
+    wasm_v128_store(sp, v17); sp += 16;
+    wasm_v128_store(sp, v18); sp += 16;
+    wasm_v128_store(sp, v19); sp += 16;
+    wasm_v128_store(sp, v20); sp += 16;
+    wasm_v128_store(sp, v21); sp += 16;
+    wasm_v128_store(sp, v22); sp += 16;
+    wasm_v128_store(sp, v23);
+}
+
+static void scale(v128_t *out0, v128_t *out1, v128_t *out2, v128_t *out3,
+        v128_t in0, v128_t in1, v128_t in2, v128_t in3,
+        v128_t in4, v128_t in5, v128_t in6, v128_t in7,
+        int shift)
+{
+    v128_t add = wasm_i32x4_splat(1 << (shift - 1));
+
+    in0 = wasm_i32x4_add(in0, add);
+    in1 = wasm_i32x4_add(in1, add);
+    in2 = wasm_i32x4_add(in2, add);
+    in3 = wasm_i32x4_add(in3, add);
+    in4 = wasm_i32x4_add(in4, add);
+    in5 = wasm_i32x4_add(in5, add);
+    in6 = wasm_i32x4_add(in6, add);
+    in7 = wasm_i32x4_add(in7, add);
+
+    in0 = wasm_i32x4_shr(in0, shift);
+    in1 = wasm_i32x4_shr(in1, shift);
+    in2 = wasm_i32x4_shr(in2, shift);
+    in3 = wasm_i32x4_shr(in3, shift);
+    in4 = wasm_i32x4_shr(in4, shift);
+    in5 = wasm_i32x4_shr(in5, shift);
+    in6 = wasm_i32x4_shr(in6, shift);
+    in7 = wasm_i32x4_shr(in7, shift);
+
+    *out0 = wasm_i16x8_narrow_i32x4(in0, in1);
+    *out1 = wasm_i16x8_narrow_i32x4(in2, in3);
+    *out2 = wasm_i16x8_narrow_i32x4(in4, in5);
+    *out3 = wasm_i16x8_narrow_i32x4(in6, in7);
+}
+
+static void transpose16_4x4_2(v128_t *r0, v128_t *r1, v128_t *r2, v128_t *r3)
+{
+    v128_t t0, t1, t2, t3, t4, t5;
+
+    t0 = wasm_i16x8_shuffle(*r0, *r1, 0, 8, 2, 10, 4, 12, 6, 14);
+    t1 = wasm_i16x8_shuffle(*r0, *r1, 1, 9, 3, 11, 5, 13, 7, 15);
+    t2 = wasm_i16x8_shuffle(*r2, *r3, 0, 8, 2, 10, 4, 12, 6, 14);
+    t3 = wasm_i16x8_shuffle(*r2, *r3, 1, 9, 3, 11, 5, 13, 7, 15);
+    t4 = wasm_i32x4_shuffle(t0, t2, 0, 4, 2, 6);
+    t5 = wasm_i32x4_shuffle(t0, t2, 1, 5, 3, 7);
+    t0 = wasm_i32x4_shuffle(t1, t3, 0, 4, 2, 6);
+    t2 = wasm_i32x4_shuffle(t1, t3, 1, 5, 3, 7);
+    *r0 = wasm_i64x2_shuffle(t4, *r0, 0, 3);
+    *r2 = wasm_i64x2_shuffle(t5, *r2, 0, 3);
+    *r1 = wasm_i64x2_shuffle(t0, *r1, 0, 3);
+    *r3 = wasm_i64x2_shuffle(t2, *r3, 0, 3);
+
+    t0 = wasm_i16x8_shuffle(*r3, *r2, 0, 8, 2, 10, 4, 12, 6, 14);
+    t1 = wasm_i16x8_shuffle(*r3, *r2, 1, 9, 3, 11, 5, 13, 7, 15);
+    t2 = wasm_i16x8_shuffle(*r1, *r0, 0, 8, 2, 10, 4, 12, 6, 14);
+    t3 = wasm_i16x8_shuffle(*r1, *r0, 1, 9, 3, 11, 5, 13, 7, 15);
+    t4 = wasm_i32x4_shuffle(t0, t2, 0, 4, 2, 6);
+    t5 = wasm_i32x4_shuffle(t0, t2, 1, 5, 3, 7);
+    t0 = wasm_i32x4_shuffle(t1, t3, 0, 4, 2, 6);
+    t2 = wasm_i32x4_shuffle(t1, t3, 1, 5, 3, 7);
+    *r3 = wasm_i64x2_shuffle(*r3, t4, 0, 3);
+    *r1 = wasm_i64x2_shuffle(*r1, t5, 0, 3);
+    *r2 = wasm_i64x2_shuffle(*r2, t0, 0, 3);
+    *r0 = wasm_i64x2_shuffle(*r0, t2, 0, 3);
+}
+
+static void store16(v128_t in0, v128_t in1, v128_t in2, v128_t in3,
+                    char *x1, char *x3, int x1_step, int x3_step)
+{
+    wasm_v128_store64_lane(x1, in0, 0);
+    wasm_v128_store64_lane(x3, in0, 1);
+    x1 += x1_step;
+    x3 += x3_step;
+
+    wasm_v128_store64_lane(x1, in1, 0);
+    wasm_v128_store64_lane(x3, in1, 1);
+    x1 += x1_step;
+    x3 += x3_step;
+
+    wasm_v128_store64_lane(x1, in2, 0);
+    wasm_v128_store64_lane(x3, in2, 1);
+    x1 += x1_step;
+    x3 += x3_step;
+
+    wasm_v128_store64_lane(x1, in3, 0);
+    wasm_v128_store64_lane(x3, in3, 1);
+}
+
+
+static void store_to_stack(char *sp, int off1, int off2,
+                           v128_t in0, v128_t in2, v128_t in4, v128_t in6,
+                           v128_t in7, v128_t in5, v128_t in3, v128_t in1)
+{
+    char *x1 = sp + off1;
+    char *x3 = sp + off2;
+
+    wasm_v128_store(x1, in0);
+    wasm_v128_store(x3, in1);
+    x1 += 16;
+    x3 -= 16;
+    wasm_v128_store(x1, in2);
+    wasm_v128_store(x3, in3);
+    x1 += 16;
+    x3 -= 16;
+    wasm_v128_store(x1, in4);
+    wasm_v128_store(x3, in5);
+    x1 += 16;
+    x3 -= 16;
+    wasm_v128_store(x1, in6);
+    wasm_v128_store(x3, in7);
+}
+
+#define sum_sub(out, in0, in1, operation, half) \
+    out = wasm_i32x4_ ## operation (out, wasm_i32x4_extmul_ ## half ## _i16x8(in0, in1));
+
+#define add_member(in, t0, t1, t2, t3, t4, t5, t6, t7, op0, op1, op2, op3, op4, op5, op6, op7, half) \
+    do {                                \
+        sum_sub(v21, in, t0, op0, half) \
+        sum_sub(v22, in, t1, op1, half) \
+        sum_sub(v23, in, t2, op2, half) \
+        sum_sub(v24, in, t3, op3, half) \
+        sum_sub(v25, in, t4, op4, half) \
+        sum_sub(v26, in, t5, op5, half) \
+        sum_sub(v27, in, t6, op6, half) \
+        sum_sub(v28, in, t7, op7, half) \
+    } while (0)
+
+#define butterfly16(in0, in1, in2, in3, in4, in5, in6, in7) \
+    do {                                                    \
+        v20 = wasm_i32x4_add(in0, in1);                     \
+        in0 = wasm_i32x4_sub(in0, in1);                     \
+        in1 = wasm_i32x4_add(in2, in3);                     \
+        in2 = wasm_i32x4_sub(in2, in3);                     \
+        in3 = wasm_i32x4_add(in4, in5);                     \
+        in4 = wasm_i32x4_sub(in4, in5);                     \
+        in5 = wasm_i32x4_add(in6, in7);                     \
+        in6 = wasm_i32x4_sub(in6, in7);                     \
+    } while (0)
+
+static void tr_16x4(char *src, char *buf, char *sp,
+                    int shift, int offset, int step)
+{
+    char *x1, *x3, *x4;
+    int x2;
+    v128_t trans[8];
+    v128_t v16, v17, v18, v19, v20, v21, v22, v23,
+           v24, v25, v26, v27, v28, v29, v30, v31;
+
+    trans[0] = wasm_i16x8_const_splat(transform[0]);
+    trans[1] = wasm_i16x8_const_splat(transform[1]);
+    trans[2] = wasm_i16x8_const_splat(transform[2]);
+    trans[3] = wasm_i16x8_const_splat(transform[3]);
+    trans[4] = wasm_i16x8_const_splat(transform[4]);
+    trans[5] = wasm_i16x8_const_splat(transform[5]);
+    trans[6] = wasm_i16x8_const_splat(transform[6]);
+    trans[7] = wasm_i16x8_const_splat(transform[7]);
+
+    x1 = src;
+    x3 = src + step * 64;
+    x2 = step * 128;
+    load16(x1, x3, x2, v16, v17, v18, v19);
+    tr16_8x4(v16, v17, v18, v19, trans, sp, offset);
+
+    x1 = src + step * 32;
+    x3 = src + step * 3 * 32;
+    x2 = step * 128;
+    load16(x1, x3, x2, v20, v17, v18, v19);
+
+    trans[0] = wasm_i16x8_const_splat(transform[0 + 8]);
+    trans[1] = wasm_i16x8_const_splat(transform[1 + 8]);
+    trans[2] = wasm_i16x8_const_splat(transform[2 + 8]);
+    trans[3] = wasm_i16x8_const_splat(transform[3 + 8]);
+    trans[4] = wasm_i16x8_const_splat(transform[4 + 8]);
+    trans[5] = wasm_i16x8_const_splat(transform[5 + 8]);
+    trans[6] = wasm_i16x8_const_splat(transform[6 + 8]);
+    trans[7] = wasm_i16x8_const_splat(transform[7 + 8]);
+
+    v21 = wasm_i32x4_extmul_low_i16x8(v20, trans[0]);
+    v22 = wasm_i32x4_extmul_low_i16x8(v20, trans[1]);
+    v23 = wasm_i32x4_extmul_low_i16x8(v20, trans[2]);
+    v24 = wasm_i32x4_extmul_low_i16x8(v20, trans[3]);
+    v25 = wasm_i32x4_extmul_low_i16x8(v20, trans[4]);
+    v26 = wasm_i32x4_extmul_low_i16x8(v20, trans[5]);
+    v27 = wasm_i32x4_extmul_low_i16x8(v20, trans[6]);
+    v28 = wasm_i32x4_extmul_low_i16x8(v20, trans[7]);
+
+    add_member(v20, trans[1], trans[4], trans[7], trans[5],
+        trans[2], trans[0], trans[3], trans[6],
+        add, add, add, sub, sub, sub, sub, sub, high);
+    add_member(v17, trans[2], trans[7], trans[3], trans[1],
+        trans[6], trans[4], trans[0], trans[5],
+        add, add, sub, sub, sub, add, add, add, low);
+    add_member(v17, trans[3], trans[5], trans[1], trans[7],
+        trans[0], trans[6], trans[2], trans[4],
+        add, sub, sub, add, add, add, sub, sub, high);
+    add_member(v18, trans[4], trans[2], trans[6], trans[0],
+        trans[7], trans[1], trans[5], trans[3],
+        add, sub, sub, add, sub, sub, add, add, low);
+    add_member(v18, trans[5], trans[0], trans[4], trans[6],
+        trans[1], trans[3], trans[7], trans[2],
+        add, sub, add, add, sub, add, add, sub, high);
+    add_member(v19, trans[6], trans[3], trans[0], trans[2],
+        trans[5], trans[7], trans[4], trans[1],
+        add, sub, add, sub, add, add, sub, add, low);
+    add_member(v19, trans[7], trans[6], trans[5], trans[4],
+        trans[3], trans[2], trans[1], trans[0],
+        add, sub, add, sub, add, sub, add, sub, high);
+
+    x4 = &sp[offset];
+    v16 = wasm_v128_load(x4);
+    x4 += 16;
+    v17 = wasm_v128_load(x4);
+    x4 += 16;
+    v18 = wasm_v128_load(x4);
+    x4 += 16;
+    v19 = wasm_v128_load(x4);
+    butterfly16(v16, v21, v17, v22, v18, v23, v19, v24);
+
+    if (shift > 0) {
+        scale(&v29, &v30, &v31, &v24,
+              v20, v16, v21, v17, v22, v18, v23, v19,
+              shift);
+        transpose16_4x4_2(&v29, &v30, &v31, &v24);
+        x1 = buf;
+        x3 = &buf[24 + 3 * 32];
+        store16(v29, v30, v31, v24, x1, x3, 32, -32);
+    } else {
+        store_to_stack(sp, offset, offset + 240,
+                v20, v21, v22, v23, v19, v18, v17, v16);
+    }
+
+    x4 = &sp[offset + 64];
+    v16 = wasm_v128_load(x4);
+    x4 += 16;
+    v17 = wasm_v128_load(x4);
+    x4 += 16;
+    v18 = wasm_v128_load(x4);
+    x4 += 16;
+    v19 = wasm_v128_load(x4);
+    butterfly16(v16, v25, v17, v26, v18, v27, v19, v28);
+
+    if (shift > 0) {
+        scale(&v29, &v30, &v31, &v20,
+              v20, v16, v25, v17, v26, v18, v27, v19,
+              shift);
+        transpose16_4x4_2(&v29, &v30, &v31, &v20);
+        x1 = &buf[8];
+        x3 = &buf[16 + 3 * 32];
+        store16(v29, v30, v31, v20, x1, x3, 32, -32);
+    } else {
+        store_to_stack(sp, offset + 64, offset + 176,
+                v20, v25, v26, v27, v19, v18, v17, v16);
+    }
+}
+
+static void idct_16x16(char *coeffs, int bit_depth)
+{
+    DECLARE_ALIGNED(16, char, sp)[640];
+
+    for (int i = 0; i < 4; i++) {
+        char *x5 = &coeffs[8 * i];
+        char *x6 = &sp[8 * i * 16];
+        tr_16x4(x5, x6, sp, 7, 512, 1);
+    }
+
+    for (int i = 0; i < 4; i++) {
+        char *x5 = &sp[8 * i];
+        char *x6 = &coeffs[8 * i * 16];
+        tr_16x4(x5, x6, sp, 20 - bit_depth, 512, 1);
+    }
+}
+
+void ff_hevc_idct_16x16_8_simd128(int16_t *coeffs, int col_limit)
+{
+    idct_16x16((char *)coeffs, 8);
+}
+
+void ff_hevc_idct_16x16_10_simd128(int16_t *coeffs, int col_limit)
+{
+    idct_16x16((char *)coeffs, 10);
+}
+
+#define add_member32(in, t0, t1, t2, t3, op0, op1, op2, op3, half) \
+    do { \
+        sum_sub(v24, in, t0, op0, half) \
+        sum_sub(v25, in, t1, op1, half) \
+        sum_sub(v26, in, t2, op2, half) \
+        sum_sub(v27, in, t3, op3, half) \
+    } while (0)
+
+#define butterfly32(in0, in1, in2, in3, out) \
+    do {                                     \
+        out = wasm_i32x4_add(in0, in1);      \
+        in0 = wasm_i32x4_sub(in0, in1);      \
+        in1 = wasm_i32x4_add(in2, in3);      \
+        in2 = wasm_i32x4_sub(in2, in3);      \
+    } while (0)
+
+static void tr_32x4(char *x5, char *x11, char *sp, int shift)
+{
+    char *x1, *x3, *x4;
+    // transform in v0 - v4
+    v128_t v0[4];
+    v128_t v1[4];
+    v128_t v2[4];
+    v128_t v3[4];
+    v128_t v4, v5, v6, v7, v16, v17, v18, v19,
+           v20, v21, v22, v23, v24, v25, v26, v27,
+           v28, v29, v30, v31, v32, v33;
+
+    tr_16x4(x5, x11, sp, 0, 2048, 4);
+
+    // load32
+    x1 = &x5[64];
+    x3 = &x1[128];
+    v4 = wasm_v128_load64_zero(x1);
+    v4 = wasm_v128_load64_lane(x3, v4, 1);
+    x1 += 256;
+    x3 += 256;
+    v5 = wasm_v128_load64_zero(x1);
+    v5 = wasm_v128_load64_lane(x3, v5, 1);
+    x1 += 256;
+    x3 += 256;
+    v6 = wasm_v128_load64_zero(x1);
+    v6 = wasm_v128_load64_lane(x3, v6, 1);
+    x1 += 256;
+    x3 += 256;
+    v7 = wasm_v128_load64_zero(x1);
+    v7 = wasm_v128_load64_lane(x3, v7, 1);
+    x1 += 256;
+    x3 += 256;
+    v16 = wasm_v128_load64_zero(x1);
+    v16 = wasm_v128_load64_lane(x3, v16, 1);
+    x1 += 256;
+    x3 += 256;
+    v17 = wasm_v128_load64_zero(x1);
+    v17 = wasm_v128_load64_lane(x3, v17, 1);
+    x1 += 256;
+    x3 += 256;
+    v18 = wasm_v128_load64_zero(x1);
+    v18 = wasm_v128_load64_lane(x3, v18, 1);
+    x1 += 256;
+    x3 += 256;
+    v19 = wasm_v128_load64_zero(x1);
+    v19 = wasm_v128_load64_lane(x3, v19, 1);
+
+    // load transform
+    v0[0] = wasm_i16x8_const_splat(transform[16 + 0]);
+    v0[1] = wasm_i16x8_const_splat(transform[16 + 1]);
+    v0[2] = wasm_i16x8_const_splat(transform[16 + 2]);
+    v0[3] = wasm_i16x8_const_splat(transform[16 + 3]);
+    v1[0] = wasm_i16x8_const_splat(transform[16 + 4]);
+    v1[1] = wasm_i16x8_const_splat(transform[16 + 5]);
+    v1[2] = wasm_i16x8_const_splat(transform[16 + 6]);
+    v1[3] = wasm_i16x8_const_splat(transform[16 + 7]);
+    v2[0] = wasm_i16x8_const_splat(transform[16 + 8]);
+    v2[1] = wasm_i16x8_const_splat(transform[16 + 9]);
+    v2[2] = wasm_i16x8_const_splat(transform[16 + 10]);
+    v2[3] = wasm_i16x8_const_splat(transform[16 + 11]);
+    v3[0] = wasm_i16x8_const_splat(transform[16 + 12]);
+    v3[1] = wasm_i16x8_const_splat(transform[16 + 13]);
+    v3[2] = wasm_i16x8_const_splat(transform[16 + 14]);
+    v3[3] = wasm_i16x8_const_splat(transform[16 + 15]);
+
+    // tr_block1
+    v24 = wasm_i32x4_extmul_low_i16x8(v4, v0[0]);
+    v25 = wasm_i32x4_extmul_low_i16x8(v4, v0[1]);
+    v26 = wasm_i32x4_extmul_low_i16x8(v4, v0[2]);
+    v27 = wasm_i32x4_extmul_low_i16x8(v4, v0[3]);
+
+    add_member32(v4, v0[1], v1[0], v1[3], v2[2], add, add, add, add, high);
+    add_member32(v5, v0[2], v1[3], v3[0], v3[2], add, add, add, sub, low);
+    add_member32(v5, v0[3], v2[2], v3[2], v1[3], add, add, sub, sub, high);
+    add_member32(v6, v1[0], v3[1], v2[1], v0[0], add, add, sub, sub, low);
+    add_member32(v6, v1[1], v3[3], v1[0], v1[2], add, sub, sub, sub, high);
+    add_member32(v7, v1[2], v3[0], v0[0], v3[1], add, sub, sub, sub, low);
+    add_member32(v7, v1[3], v2[1], v1[1], v2[3], add, sub, sub, add, high);
+    add_member32(v16, v2[0], v1[2], v2[2], v1[0], add, sub, sub, add, low);
+    add_member32(v16, v2[1], v0[3], v3[3], v0[2], add, sub, sub, add, high);
+    add_member32(v17, v2[2], v0[1], v2[3], v2[1], add, sub, add, add, low);
+    add_member32(v17, v2[3], v0[2], v1[2], v3[3], add, sub, add, sub, high);
+    add_member32(v18, v3[0], v1[1], v0[1], v2[0], add, sub, add, sub, low);
+    add_member32(v18, v3[1], v2[0], v0[3], v0[1], add, sub, add, sub, high);
+    add_member32(v19, v3[2], v2[3], v2[0], v1[1], add, sub, add, sub, low);
+    add_member32(v19, v3[3], v3[2], v3[1], v3[0], add, sub, add, sub, high);
+
+    x4 = &sp[2048];
+    // scale_store
+    v28 = wasm_v128_load(x4);
+    x4 += 16;
+    v29 = wasm_v128_load(x4);
+    x4 += 16;
+    v30 = wasm_v128_load(x4);
+    x4 += 16;
+    v31 = wasm_v128_load(x4);
+    x4 += 16;
+    butterfly32(v28, v24, v29, v25, v32);
+    butterfly32(v30, v26, v31, v27, v33);
+    scale(&v20, &v21, &v22, &v23, v32, v28, v24, v29, v33, v30, v26, v31, shift);
+    transpose16_4x4_2(&v20, &v21, &v22, &v23);
+    x1 = x11;
+    x3 = &x11[56 + 3 * 64];
+    store16(v20, v21, v22, v23, x1, x3, 64, -64);
+
+    // tr_block2
+    v24 = wasm_i32x4_extmul_low_i16x8(v4, v1[0]);
+    v25 = wasm_i32x4_extmul_low_i16x8(v4, v1[1]);
+    v26 = wasm_i32x4_extmul_low_i16x8(v4, v1[2]);
+    v27 = wasm_i32x4_extmul_low_i16x8(v4, v1[3]);
+
+    add_member32(v4,  v3[1], v3[3], v3[0], v2[1], add, sub, sub, sub, high);
+    add_member32(v5,  v2[1], v1[0], v0[0], v1[1], sub, sub, sub, sub, low);
+    add_member32(v5,  v0[0], v1[2], v3[1], v2[3], sub, sub, sub, add, high);
+    add_member32(v6,  v2[0], v3[2], v1[1], v0[3], sub, add, add, add, low);
+    add_member32(v6,  v3[2], v0[3], v1[3], v3[1], add, add, add, sub, high);
+    add_member32(v7,  v1[1], v1[3], v2[3], v0[0], add, add, sub, sub, low);
+    add_member32(v7,  v0[3], v3[1], v0[1], v3[3], add, sub, sub, add, high);
+    add_member32(v16, v3[0], v0[2], v3[2], v0[1], add, sub, sub, add, low);
+    add_member32(v16, v2[2], v2[0], v1[0], v3[2], sub, sub, add, add, high);
+    add_member32(v17, v0[1], v3[0], v2[0], v0[2], sub, add, add, sub, low);
+    add_member32(v17, v1[3], v0[1], v2[2], v3[0], sub, add, sub, sub, high);
+    add_member32(v18, v3[3], v2[1], v0[2], v1[0], add, add, sub, add, low);
+    add_member32(v18, v1[2], v2[3], v3[3], v2[2], add, sub, sub, add, high);
+    add_member32(v19, v0[2], v0[1], v0[3], v1[2], add, sub, add, sub, low);
+    add_member32(v19, v2[3], v2[2], v2[1], v2[0], add, sub, add, sub, high);
+
+    // scale_store
+    v28 = wasm_v128_load(x4);
+    x4 += 16;
+    v29 = wasm_v128_load(x4);
+    x4 += 16;
+    v30 = wasm_v128_load(x4);
+    x4 += 16;
+    v31 = wasm_v128_load(x4);
+    x4 += 16;
+    butterfly32(v28, v24, v29, v25, v32);
+    butterfly32(v30, v26, v31, v27, v33);
+    scale(&v20, &v21, &v22, &v23, v32, v28, v24, v29, v33, v30, v26, v31, shift);
+    transpose16_4x4_2(&v20, &v21, &v22, &v23);
+    x1 = &x11[8];
+    x3 = &x11[48 + 3 * 64];
+    store16(v20, v21, v22, v23, x1, x3, 64, -64);
+
+    // tr_block3
+    v24 = wasm_i32x4_extmul_low_i16x8(v4, v2[0]);
+    v25 = wasm_i32x4_extmul_low_i16x8(v4, v2[1]);
+    v26 = wasm_i32x4_extmul_low_i16x8(v4, v2[2]);
+    v27 = wasm_i32x4_extmul_low_i16x8(v4, v2[3]);
+    add_member32(v4,  v1[2], v0[3], v0[0], v0[2], sub, sub, sub, sub, high);
+    add_member32(v5,  v2[2], v3[3], v2[3], v1[2], sub, sub, add, add, low);
+    add_member32(v5,  v1[0], v0[2], v2[1], v3[3], add, add, add, sub, high);
+    add_member32(v6,  v3[0], v2[2], v0[1], v1[3], add, sub, sub, sub, low);
+    add_member32(v6,  v0[2], v2[0], v3[0], v0[0], sub, sub, add, add, high);
+    add_member32(v7,  v3[2], v1[0], v2[0], v2[2], sub, add, add, sub, low);
+    add_member32(v7,  v0[0], v3[2], v0[2], v3[0], add, add, sub, sub, high);
+    add_member32(v16, v3[3], v0[1], v3[1], v0[3], sub, sub, add, add, low);
+    add_member32(v16, v0[1], v2[3], v1[3], v1[1], sub, add, add, sub, high);
+    add_member32(v17, v3[1], v1[3], v0[3], v3[2], add, add, sub, add, low);
+    add_member32(v17, v0[3], v1[1], v3[2], v2[0], add, sub, add, add, high);
+    add_member32(v18, v2[3], v3[1], v1[2], v0[1], sub, sub, add, sub, low);
+    add_member32(v18, v1[1], v0[0], v1[0], v2[1], sub, add, sub, add, high);
+    add_member32(v19, v2[1], v3[0], v3[3], v3[1], add, sub, add, add, low);
+    add_member32(v19, v1[3], v1[2], v1[1], v1[0], add, sub, add, sub, high);
+
+    // scale_store
+    v28 = wasm_v128_load(x4);
+    x4 += 16;
+    v29 = wasm_v128_load(x4);
+    x4 += 16;
+    v30 = wasm_v128_load(x4);
+    x4 += 16;
+    v31 = wasm_v128_load(x4);
+    x4 += 16;
+    butterfly32(v28, v24, v29, v25, v32);
+    butterfly32(v30, v26, v31, v27, v33);
+    scale(&v20, &v21, &v22, &v23, v32, v28, v24, v29, v33, v30, v26, v31, shift);
+    transpose16_4x4_2(&v20, &v21, &v22, &v23);
+    x1 = &x11[16];
+    x3 = &x11[40 + 3 * 64];
+    store16(v20, v21, v22, v23, x1, x3, 64, -64);
+
+    // try_block4
+    v24 = wasm_i32x4_extmul_low_i16x8(v4, v3[0]);
+    v25 = wasm_i32x4_extmul_low_i16x8(v4, v3[1]);
+    v26 = wasm_i32x4_extmul_low_i16x8(v4, v3[2]);
+    v27 = wasm_i32x4_extmul_low_i16x8(v4, v3[3]);
+    add_member32(v4,  v1[1], v2[0], v2[3], v3[2], sub, sub, sub, sub, high);
+    add_member32(v5,  v0[0], v0[3], v2[0], v3[1], add, add, add, add, low);
+    add_member32(v5,  v2[0], v0[0], v1[1], v3[0], sub, sub, sub, sub, high);
+    add_member32(v6,  v3[3], v1[2], v0[2], v2[3], add, add, add, add, low);
+    add_member32(v6,  v2[1], v2[3], v0[0], v2[2], add, sub, sub, sub, high);
+    add_member32(v7,  v0[2], v3[3], v0[3], v2[1], sub, sub, add, add, low);
+    add_member32(v7,  v1[0], v2[2], v1[2], v2[0], add, add, sub, sub, high);
+    add_member32(v16, v2[3], v1[1], v2[1], v1[3], sub, sub, add, add, low);
+    add_member32(v16, v3[1], v0[1], v3[0], v1[2], sub, add, sub, sub, high);
+    add_member32(v17, v1[2], v1[0], v3[3], v1[1], add, sub, add, add, low);
+    add_member32(v17, v0[1], v2[1], v3[1], v1[0], sub, add, add, sub, high);
+    add_member32(v18, v1[3], v3[2], v2[2], v0[3], add, sub, sub, add, low);
+    add_member32(v18, v3[2], v3[0], v1[3], v0[2], sub, sub, add, sub, high);
+    add_member32(v19, v2[2], v1[3], v1[0], v0[1], sub, add, sub, add, low);
+    add_member32(v19, v0[3], v0[2], v0[1], v0[0], add, sub, add, sub, high);
+
+    // scale_store
+    v28 = wasm_v128_load(x4);
+    x4 += 16;
+    v29 = wasm_v128_load(x4);
+    x4 += 16;
+    v30 = wasm_v128_load(x4);
+    x4 += 16;
+    v31 = wasm_v128_load(x4);
+    butterfly32(v28, v24, v29, v25, v32);
+    butterfly32(v30, v26, v31, v27, v33);
+    scale(&v20, &v21, &v22, &v23, v32, v28, v24, v29, v33, v30, v26, v31, shift);
+    transpose16_4x4_2(&v20, &v21, &v22, &v23);
+    x1 = &x11[24];
+    x3 = &x11[32 + 3 * 64];
+    store16(v20, v21, v22, v23, x1, x3, 64, -64);
+}
+
+static void idct_32x32(char *coeffs, int bit_depth)
+{
+    DECLARE_ALIGNED(16, char, sp)[2432];
+    char *x5, *x11;
+
+    for (int i = 0; i < 8; i++) {
+        x5 = &coeffs[8 * i];
+        x11 = &sp[8 * i * 32];
+        tr_32x4(x5, x11, sp, 7);
+    }
+
+    for (int i = 0; i < 8; i++) {
+        x5 = &sp[8 * i];
+        x11 = &coeffs[8 * i * 32];
+        tr_32x4(x5, x11, sp, 20 - bit_depth);
+    }
+}
+
+void ff_hevc_idct_32x32_8_simd128(int16_t *coeffs, int col_limit)
+{
+    idct_32x32((char *)coeffs, 8);
+}
+
+void ff_hevc_idct_32x32_10_simd128(int16_t *coeffs, int col_limit)
+{
+    idct_32x32((char *)coeffs, 10);
+}
diff --git a/libavcodec/wasm/hevc/idct.h b/libavcodec/wasm/hevc/idct.h
new file mode 100644
index 0000000000..05c64ae683
--- /dev/null
+++ b/libavcodec/wasm/hevc/idct.h
@@ -0,0 +1,36 @@
+/*
+ * Copyright (c) 2024 Zhao Zhili
+ *
+ * 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_WASM_HEVC_IDCT_H
+#define AVCODEC_WASM_HEVC_IDCT_H
+
+#include <stdint.h>
+
+void ff_hevc_idct_4x4_8_simd128(int16_t *coeffs, int col_limit);
+void ff_hevc_idct_8x8_8_simd128(int16_t *coeffs, int col_limit);
+void ff_hevc_idct_16x16_8_simd128(int16_t *coeffs, int col_limit);
+void ff_hevc_idct_32x32_8_simd128(int16_t *coeffs, int col_limit);
+
+void ff_hevc_idct_4x4_10_simd128(int16_t *coeffs, int col_limit);
+void ff_hevc_idct_8x8_10_simd128(int16_t *coeffs, int col_limit);
+void ff_hevc_idct_16x16_10_simd128(int16_t *coeffs, int col_limit);
+void ff_hevc_idct_32x32_10_simd128(int16_t *coeffs, int col_limit);
+
+#endif /* AVCODEC_WASM_HEVC_IDCT_H */
-- 
2.46.0



More information about the ffmpeg-devel mailing list