Add support for v256 intrinsics
Change-Id: I1da08afaa945ca1aaf4bf9f50cf649a7feef2e60
diff --git a/aom_dsp/aom_dsp.mk b/aom_dsp/aom_dsp.mk
index c74bfe3..28e7f12 100644
--- a/aom_dsp/aom_dsp.mk
+++ b/aom_dsp/aom_dsp.mk
@@ -394,9 +394,13 @@
DSP_SRCS-yes += simd/v64_intrinsics_c.h
DSP_SRCS-yes += simd/v128_intrinsics.h
DSP_SRCS-yes += simd/v128_intrinsics_c.h
+DSP_SRCS-yes += simd/v256_intrinsics.h
+DSP_SRCS-yes += simd/v256_intrinsics_c.h
DSP_SRCS-$(HAVE_SSE2) += simd/v64_intrinsics_x86.h
DSP_SRCS-$(HAVE_SSE2) += simd/v128_intrinsics_x86.h
+DSP_SRCS-$(HAVE_SSE2) += simd/v256_intrinsics_x86.h
DSP_SRCS-$(HAVE_NEON) += simd/v64_intrinsics_arm.h
DSP_SRCS-$(HAVE_NEON) += simd/v128_intrinsics_arm.h
+DSP_SRCS-$(HAVE_NEON) += simd/v256_intrinsics_arm.h
$(eval $(call rtcd_h_template,aom_dsp_rtcd,aom_dsp/aom_dsp_rtcd_defs.pl))
diff --git a/aom_dsp/aom_simd.h b/aom_dsp/aom_simd.h
index 3879d95..ae4ff23 100644
--- a/aom_dsp/aom_simd.h
+++ b/aom_dsp/aom_simd.h
@@ -22,11 +22,11 @@
#include "./aom_simd_inline.h"
#if HAVE_NEON
-#include "simd/v128_intrinsics_arm.h"
+#include "simd/v256_intrinsics_arm.h"
#elif HAVE_SSE2
-#include "simd/v128_intrinsics_x86.h"
+#include "simd/v256_intrinsics_x86.h"
#else
-#include "simd/v128_intrinsics.h"
+#include "simd/v256_intrinsics.h"
#endif
#endif // AOM_DSP_AOM_AOM_SIMD_H_
diff --git a/aom_dsp/simd/v256_intrinsics.h b/aom_dsp/simd/v256_intrinsics.h
new file mode 100644
index 0000000..73bcd94
--- /dev/null
+++ b/aom_dsp/simd/v256_intrinsics.h
@@ -0,0 +1,274 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef _V256_INTRINSICS_H
+#define _V256_INTRINSICS_H
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "./v256_intrinsics_c.h"
+#include "./v128_intrinsics.h"
+#include "./v64_intrinsics.h"
+
+/* Fallback to plain, unoptimised C. */
+
+typedef c_v256 v256;
+
+SIMD_INLINE uint32_t v256_low_u32(v256 a) { return c_v256_low_u32(a); }
+SIMD_INLINE v64 v256_low_v64(v256 a) { return c_v256_low_v64(a); }
+SIMD_INLINE v128 v256_low_v128(v256 a) { return c_v256_low_v128(a); }
+SIMD_INLINE v128 v256_high_v128(v256 a) { return c_v256_high_v128(a); }
+SIMD_INLINE v256 v256_from_v128(v128 hi, v128 lo) {
+ return c_v256_from_v128(hi, lo);
+}
+SIMD_INLINE v256 v256_from_64(uint64_t a, uint64_t b, uint64_t c, uint64_t d) {
+ return c_v256_from_64(a, b, c, d);
+}
+SIMD_INLINE v256 v256_from_v64(v64 a, v64 b, v64 c, v64 d) {
+ return c_v256_from_v64(a, b, c, d);
+}
+
+SIMD_INLINE v256 v256_load_unaligned(const void *p) {
+ return c_v256_load_unaligned(p);
+}
+SIMD_INLINE v256 v256_load_aligned(const void *p) {
+ return c_v256_load_aligned(p);
+}
+
+SIMD_INLINE void v256_store_unaligned(void *p, v256 a) {
+ c_v256_store_unaligned(p, a);
+}
+SIMD_INLINE void v256_store_aligned(void *p, v256 a) {
+ c_v256_store_aligned(p, a);
+}
+
+SIMD_INLINE v256 v256_align(v256 a, v256 b, const unsigned int c) {
+ return c_v256_align(a, b, c);
+}
+
+SIMD_INLINE v256 v256_zero() { return c_v256_zero(); }
+SIMD_INLINE v256 v256_dup_8(uint8_t x) { return c_v256_dup_8(x); }
+SIMD_INLINE v256 v256_dup_16(uint16_t x) { return c_v256_dup_16(x); }
+SIMD_INLINE v256 v256_dup_32(uint32_t x) { return c_v256_dup_32(x); }
+
+typedef uint32_t sad256_internal;
+SIMD_INLINE sad256_internal v256_sad_u8_init() { return c_v256_sad_u8_init(); }
+SIMD_INLINE sad256_internal v256_sad_u8(sad256_internal s, v256 a, v256 b) {
+ return c_v256_sad_u8(s, a, b);
+}
+SIMD_INLINE uint32_t v256_sad_u8_sum(sad256_internal s) {
+ return c_v256_sad_u8_sum(s);
+}
+typedef uint32_t ssd256_internal;
+SIMD_INLINE ssd256_internal v256_ssd_u8_init() { return c_v256_ssd_u8_init(); }
+SIMD_INLINE ssd256_internal v256_ssd_u8(ssd256_internal s, v256 a, v256 b) {
+ return c_v256_ssd_u8(s, a, b);
+}
+SIMD_INLINE uint32_t v256_ssd_u8_sum(ssd256_internal s) {
+ return c_v256_ssd_u8_sum(s);
+}
+SIMD_INLINE int64_t v256_dotp_s16(v256 a, v256 b) {
+ return c_v256_dotp_s16(a, b);
+}
+SIMD_INLINE uint64_t v256_hadd_u8(v256 a) { return c_v256_hadd_u8(a); }
+
+SIMD_INLINE v256 v256_or(v256 a, v256 b) { return c_v256_or(a, b); }
+SIMD_INLINE v256 v256_xor(v256 a, v256 b) { return c_v256_xor(a, b); }
+SIMD_INLINE v256 v256_and(v256 a, v256 b) { return c_v256_and(a, b); }
+SIMD_INLINE v256 v256_andn(v256 a, v256 b) { return c_v256_andn(a, b); }
+
+SIMD_INLINE v256 v256_add_8(v256 a, v256 b) { return c_v256_add_8(a, b); }
+SIMD_INLINE v256 v256_add_16(v256 a, v256 b) { return c_v256_add_16(a, b); }
+SIMD_INLINE v256 v256_sadd_s16(v256 a, v256 b) { return c_v256_sadd_s16(a, b); }
+SIMD_INLINE v256 v256_add_32(v256 a, v256 b) { return c_v256_add_32(a, b); }
+SIMD_INLINE v256 v256_padd_s16(v256 a) { return c_v256_padd_s16(a); }
+SIMD_INLINE v256 v256_sub_8(v256 a, v256 b) { return c_v256_sub_8(a, b); }
+SIMD_INLINE v256 v256_ssub_u8(v256 a, v256 b) { return c_v256_ssub_u8(a, b); }
+SIMD_INLINE v256 v256_ssub_s8(v256 a, v256 b) { return c_v256_ssub_s8(a, b); }
+SIMD_INLINE v256 v256_sub_16(v256 a, v256 b) { return c_v256_sub_16(a, b); }
+SIMD_INLINE v256 v256_ssub_s16(v256 a, v256 b) { return c_v256_ssub_s16(a, b); }
+SIMD_INLINE v256 v256_sub_32(v256 a, v256 b) { return c_v256_sub_32(a, b); }
+SIMD_INLINE v256 v256_abs_s16(v256 a) { return c_v256_abs_s16(a); }
+
+SIMD_INLINE v256 v256_mul_s16(v128 a, v128 b) { return c_v256_mul_s16(a, b); }
+SIMD_INLINE v256 v256_mullo_s16(v256 a, v256 b) {
+ return c_v256_mullo_s16(a, b);
+}
+SIMD_INLINE v256 v256_mulhi_s16(v256 a, v256 b) {
+ return c_v256_mulhi_s16(a, b);
+}
+SIMD_INLINE v256 v256_mullo_s32(v256 a, v256 b) {
+ return c_v256_mullo_s32(a, b);
+}
+SIMD_INLINE v256 v256_madd_s16(v256 a, v256 b) { return c_v256_madd_s16(a, b); }
+SIMD_INLINE v256 v256_madd_us8(v256 a, v256 b) { return c_v256_madd_us8(a, b); }
+
+SIMD_INLINE v256 v256_avg_u8(v256 a, v256 b) { return c_v256_avg_u8(a, b); }
+SIMD_INLINE v256 v256_rdavg_u8(v256 a, v256 b) { return c_v256_rdavg_u8(a, b); }
+SIMD_INLINE v256 v256_avg_u16(v256 a, v256 b) { return c_v256_avg_u16(a, b); }
+SIMD_INLINE v256 v256_min_u8(v256 a, v256 b) { return c_v256_min_u8(a, b); }
+SIMD_INLINE v256 v256_max_u8(v256 a, v256 b) { return c_v256_max_u8(a, b); }
+SIMD_INLINE v256 v256_min_s8(v256 a, v256 b) { return c_v256_min_s8(a, b); }
+SIMD_INLINE v256 v256_max_s8(v256 a, v256 b) { return c_v256_max_s8(a, b); }
+SIMD_INLINE v256 v256_min_s16(v256 a, v256 b) { return c_v256_min_s16(a, b); }
+SIMD_INLINE v256 v256_max_s16(v256 a, v256 b) { return c_v256_max_s16(a, b); }
+
+SIMD_INLINE v256 v256_ziplo_8(v256 a, v256 b) { return c_v256_ziplo_8(a, b); }
+SIMD_INLINE v256 v256_ziphi_8(v256 a, v256 b) { return c_v256_ziphi_8(a, b); }
+SIMD_INLINE v256 v256_ziplo_16(v256 a, v256 b) { return c_v256_ziplo_16(a, b); }
+SIMD_INLINE v256 v256_ziphi_16(v256 a, v256 b) { return c_v256_ziphi_16(a, b); }
+SIMD_INLINE v256 v256_ziplo_32(v256 a, v256 b) { return c_v256_ziplo_32(a, b); }
+SIMD_INLINE v256 v256_ziphi_32(v256 a, v256 b) { return c_v256_ziphi_32(a, b); }
+SIMD_INLINE v256 v256_ziplo_64(v256 a, v256 b) { return c_v256_ziplo_64(a, b); }
+SIMD_INLINE v256 v256_ziphi_64(v256 a, v256 b) { return c_v256_ziphi_64(a, b); }
+SIMD_INLINE v256 v256_ziplo_128(v256 a, v256 b) {
+ return c_v256_ziplo_128(a, b);
+}
+SIMD_INLINE v256 v256_ziphi_128(v256 a, v256 b) {
+ return c_v256_ziphi_128(a, b);
+}
+SIMD_INLINE v256 v256_zip_8(v128 a, v128 b) { return c_v256_zip_8(a, b); }
+SIMD_INLINE v256 v256_zip_16(v128 a, v128 b) { return c_v256_zip_16(a, b); }
+SIMD_INLINE v256 v256_zip_32(v128 a, v128 b) { return c_v256_zip_32(a, b); }
+SIMD_INLINE v256 v256_unziplo_8(v256 a, v256 b) {
+ return c_v256_unziplo_8(a, b);
+}
+SIMD_INLINE v256 v256_unziphi_8(v256 a, v256 b) {
+ return c_v256_unziphi_8(a, b);
+}
+SIMD_INLINE v256 v256_unziplo_16(v256 a, v256 b) {
+ return c_v256_unziplo_16(a, b);
+}
+SIMD_INLINE v256 v256_unziphi_16(v256 a, v256 b) {
+ return c_v256_unziphi_16(a, b);
+}
+SIMD_INLINE v256 v256_unziplo_32(v256 a, v256 b) {
+ return c_v256_unziplo_32(a, b);
+}
+SIMD_INLINE v256 v256_unziphi_32(v256 a, v256 b) {
+ return c_v256_unziphi_32(a, b);
+}
+SIMD_INLINE v256 v256_unpack_u8_s16(v128 a) { return c_v256_unpack_u8_s16(a); }
+SIMD_INLINE v256 v256_unpacklo_u8_s16(v256 a) {
+ return c_v256_unpacklo_u8_s16(a);
+}
+SIMD_INLINE v256 v256_unpackhi_u8_s16(v256 a) {
+ return c_v256_unpackhi_u8_s16(a);
+}
+SIMD_INLINE v256 v256_pack_s32_s16(v256 a, v256 b) {
+ return c_v256_pack_s32_s16(a, b);
+}
+SIMD_INLINE v256 v256_pack_s16_u8(v256 a, v256 b) {
+ return c_v256_pack_s16_u8(a, b);
+}
+SIMD_INLINE v256 v256_pack_s16_s8(v256 a, v256 b) {
+ return c_v256_pack_s16_s8(a, b);
+}
+SIMD_INLINE v256 v256_unpack_u16_s32(v128 a) {
+ return c_v256_unpack_u16_s32(a);
+}
+SIMD_INLINE v256 v256_unpack_s16_s32(v128 a) {
+ return c_v256_unpack_s16_s32(a);
+}
+SIMD_INLINE v256 v256_unpacklo_u16_s32(v256 a) {
+ return c_v256_unpacklo_u16_s32(a);
+}
+SIMD_INLINE v256 v256_unpacklo_s16_s32(v256 a) {
+ return c_v256_unpacklo_s16_s32(a);
+}
+SIMD_INLINE v256 v256_unpackhi_u16_s32(v256 a) {
+ return c_v256_unpackhi_u16_s32(a);
+}
+SIMD_INLINE v256 v256_unpackhi_s16_s32(v256 a) {
+ return c_v256_unpackhi_s16_s32(a);
+}
+SIMD_INLINE v256 v256_shuffle_8(v256 a, v256 pattern) {
+ return c_v256_shuffle_8(a, pattern);
+}
+SIMD_INLINE v256 v256_pshuffle_8(v256 a, v256 pattern) {
+ return c_v256_pshuffle_8(a, pattern);
+}
+
+SIMD_INLINE v256 v256_cmpgt_s8(v256 a, v256 b) { return c_v256_cmpgt_s8(a, b); }
+SIMD_INLINE v256 v256_cmplt_s8(v256 a, v256 b) { return c_v256_cmplt_s8(a, b); }
+SIMD_INLINE v256 v256_cmpeq_8(v256 a, v256 b) { return c_v256_cmpeq_8(a, b); }
+SIMD_INLINE v256 v256_cmpgt_s16(v256 a, v256 b) {
+ return c_v256_cmpgt_s16(a, b);
+}
+SIMD_INLINE v256 v256_cmplt_s16(v256 a, v256 b) {
+ return c_v256_cmplt_s16(a, b);
+}
+SIMD_INLINE v256 v256_cmpeq_16(v256 a, v256 b) { return c_v256_cmpeq_16(a, b); }
+
+SIMD_INLINE v256 v256_shl_8(v256 a, unsigned int c) {
+ return c_v256_shl_8(a, c);
+}
+SIMD_INLINE v256 v256_shr_u8(v256 a, unsigned int c) {
+ return c_v256_shr_u8(a, c);
+}
+SIMD_INLINE v256 v256_shr_s8(v256 a, unsigned int c) {
+ return c_v256_shr_s8(a, c);
+}
+SIMD_INLINE v256 v256_shl_16(v256 a, unsigned int c) {
+ return c_v256_shl_16(a, c);
+}
+SIMD_INLINE v256 v256_shr_u16(v256 a, unsigned int c) {
+ return c_v256_shr_u16(a, c);
+}
+SIMD_INLINE v256 v256_shr_s16(v256 a, unsigned int c) {
+ return c_v256_shr_s16(a, c);
+}
+SIMD_INLINE v256 v256_shl_32(v256 a, unsigned int c) {
+ return c_v256_shl_32(a, c);
+}
+SIMD_INLINE v256 v256_shr_u32(v256 a, unsigned int c) {
+ return c_v256_shr_u32(a, c);
+}
+SIMD_INLINE v256 v256_shr_s32(v256 a, unsigned int c) {
+ return c_v256_shr_s32(a, c);
+}
+
+SIMD_INLINE v256 v256_shr_n_byte(v256 a, const unsigned int n) {
+ return c_v256_shr_n_byte(a, n);
+}
+SIMD_INLINE v256 v256_shl_n_byte(v256 a, const unsigned int n) {
+ return c_v256_shl_n_byte(a, n);
+}
+SIMD_INLINE v256 v256_shl_n_8(v256 a, const unsigned int n) {
+ return c_v256_shl_n_8(a, n);
+}
+SIMD_INLINE v256 v256_shl_n_16(v256 a, const unsigned int n) {
+ return c_v256_shl_n_16(a, n);
+}
+SIMD_INLINE v256 v256_shl_n_32(v256 a, const unsigned int n) {
+ return c_v256_shl_n_32(a, n);
+}
+SIMD_INLINE v256 v256_shr_n_u8(v256 a, const unsigned int n) {
+ return c_v256_shr_n_u8(a, n);
+}
+SIMD_INLINE v256 v256_shr_n_u16(v256 a, const unsigned int n) {
+ return c_v256_shr_n_u16(a, n);
+}
+SIMD_INLINE v256 v256_shr_n_u32(v256 a, const unsigned int n) {
+ return c_v256_shr_n_u32(a, n);
+}
+SIMD_INLINE v256 v256_shr_n_s8(v256 a, const unsigned int n) {
+ return c_v256_shr_n_s8(a, n);
+}
+SIMD_INLINE v256 v256_shr_n_s16(v256 a, const unsigned int n) {
+ return c_v256_shr_n_s16(a, n);
+}
+SIMD_INLINE v256 v256_shr_n_s32(v256 a, const unsigned int n) {
+ return c_v256_shr_n_s32(a, n);
+}
+
+#endif /* _V256_INTRINSICS_H */
diff --git a/aom_dsp/simd/v256_intrinsics_arm.h b/aom_dsp/simd/v256_intrinsics_arm.h
new file mode 100644
index 0000000..ba4ed71
--- /dev/null
+++ b/aom_dsp/simd/v256_intrinsics_arm.h
@@ -0,0 +1,17 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef _V256_INTRINSICS_H
+#define _V256_INTRINSICS_H
+
+#include "./v256_intrinsics_v128.h"
+
+#endif /* _V256_INTRINSICS_H */
diff --git a/aom_dsp/simd/v256_intrinsics_c.h b/aom_dsp/simd/v256_intrinsics_c.h
new file mode 100644
index 0000000..8a67f9e
--- /dev/null
+++ b/aom_dsp/simd/v256_intrinsics_c.h
@@ -0,0 +1,701 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef _V256_INTRINSICS_C_H
+#define _V256_INTRINSICS_C_H
+
+#include <stdio.h>
+#include <stdlib.h>
+#include "./v128_intrinsics_c.h"
+#include "./aom_config.h"
+
+typedef union {
+ uint8_t u8[32];
+ uint16_t u16[16];
+ uint32_t u32[8];
+ uint64_t u64[4];
+ int8_t s8[32];
+ int16_t s16[16];
+ int32_t s32[8];
+ int64_t s64[4];
+ c_v64 v64[4];
+ c_v128 v128[2];
+} c_v256;
+
+SIMD_INLINE uint32_t c_v256_low_u32(c_v256 a) { return a.u32[0]; }
+
+SIMD_INLINE c_v64 c_v256_low_v64(c_v256 a) { return a.v64[0]; }
+
+SIMD_INLINE c_v128 c_v256_low_v128(c_v256 a) { return a.v128[0]; }
+
+SIMD_INLINE c_v128 c_v256_high_v128(c_v256 a) { return a.v128[1]; }
+
+SIMD_INLINE c_v256 c_v256_from_v128(c_v128 hi, c_v128 lo) {
+ c_v256 t;
+ t.v128[1] = hi;
+ t.v128[0] = lo;
+ return t;
+}
+
+SIMD_INLINE c_v256 c_v256_from_64(uint64_t a, uint64_t b, uint64_t c,
+ uint64_t d) {
+ c_v256 t;
+ t.u64[3] = a;
+ t.u64[2] = b;
+ t.u64[1] = c;
+ t.u64[0] = d;
+ return t;
+}
+
+SIMD_INLINE c_v256 c_v256_from_v64(c_v64 a, c_v64 b, c_v64 c, c_v64 d) {
+ c_v256 t;
+ t.u64[3] = a.u64;
+ t.u64[2] = b.u64;
+ t.u64[1] = c.u64;
+ t.u64[0] = d.u64;
+ return t;
+}
+
+SIMD_INLINE c_v256 c_v256_load_unaligned(const void *p) {
+ c_v256 t;
+ uint8_t *pp = (uint8_t *)p;
+ uint8_t *q = (uint8_t *)&t;
+ int c;
+ for (c = 0; c < 32; c++) q[c] = pp[c];
+ return t;
+}
+
+SIMD_INLINE c_v256 c_v256_load_aligned(const void *p) {
+ if (simd_check && (uintptr_t)p & 31) {
+ fprintf(stderr, "Error: unaligned v256 load at %p\n", p);
+ abort();
+ }
+ return c_v256_load_unaligned(p);
+}
+
+SIMD_INLINE void c_v256_store_unaligned(void *p, c_v256 a) {
+ uint8_t *pp = (uint8_t *)p;
+ uint8_t *q = (uint8_t *)&a;
+ int c;
+ for (c = 0; c < 32; c++) pp[c] = q[c];
+}
+
+SIMD_INLINE void c_v256_store_aligned(void *p, c_v256 a) {
+ if (simd_check && (uintptr_t)p & 31) {
+ fprintf(stderr, "Error: unaligned v256 store at %p\n", p);
+ abort();
+ }
+ c_v256_store_unaligned(p, a);
+}
+
+SIMD_INLINE c_v256 c_v256_zero() {
+ c_v256 t;
+ t.u64[3] = t.u64[2] = t.u64[1] = t.u64[0] = 0;
+ return t;
+}
+
+SIMD_INLINE c_v256 c_v256_dup_8(uint8_t x) {
+ c_v256 t;
+ t.v64[3] = t.v64[2] = t.v64[1] = t.v64[0] = c_v64_dup_8(x);
+ return t;
+}
+
+SIMD_INLINE c_v256 c_v256_dup_16(uint16_t x) {
+ c_v256 t;
+ t.v64[3] = t.v64[2] = t.v64[1] = t.v64[0] = c_v64_dup_16(x);
+ return t;
+}
+
+SIMD_INLINE c_v256 c_v256_dup_32(uint32_t x) {
+ c_v256 t;
+ t.v64[3] = t.v64[2] = t.v64[1] = t.v64[0] = c_v64_dup_32(x);
+ return t;
+}
+
+SIMD_INLINE int64_t c_v256_dotp_s16(c_v256 a, c_v256 b) {
+ return c_v128_dotp_s16(a.v128[1], b.v128[1]) +
+ c_v128_dotp_s16(a.v128[0], b.v128[0]);
+}
+
+SIMD_INLINE uint64_t c_v256_hadd_u8(c_v256 a) {
+ return c_v128_hadd_u8(a.v128[1]) + c_v128_hadd_u8(a.v128[0]);
+}
+
+typedef uint32_t c_sad256_internal;
+
+SIMD_INLINE c_sad128_internal c_v256_sad_u8_init() { return 0; }
+
+/* Implementation dependent return value. Result must be finalised with
+ v256_sad_u8_sum().
+ The result for more than 16 v256_sad_u8() calls is undefined. */
+SIMD_INLINE c_sad128_internal c_v256_sad_u8(c_sad256_internal s, c_v256 a,
+ c_v256 b) {
+ int c;
+ for (c = 0; c < 32; c++)
+ s += a.u8[c] > b.u8[c] ? a.u8[c] - b.u8[c] : b.u8[c] - a.u8[c];
+ return s;
+}
+
+SIMD_INLINE uint32_t c_v256_sad_u8_sum(c_sad256_internal s) { return s; }
+
+typedef uint32_t c_ssd256_internal;
+
+SIMD_INLINE c_ssd256_internal c_v256_ssd_u8_init() { return 0; }
+
+/* Implementation dependent return value. Result must be finalised with
+ * v256_ssd_u8_sum(). */
+SIMD_INLINE c_ssd256_internal c_v256_ssd_u8(c_ssd256_internal s, c_v256 a,
+ c_v256 b) {
+ int c;
+ for (c = 0; c < 32; c++) s += (a.u8[c] - b.u8[c]) * (a.u8[c] - b.u8[c]);
+ return s;
+}
+
+SIMD_INLINE uint32_t c_v256_ssd_u8_sum(c_ssd256_internal s) { return s; }
+
+SIMD_INLINE c_v256 c_v256_or(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_or(a.v128[1], b.v128[1]),
+ c_v128_or(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_xor(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_xor(a.v128[1], b.v128[1]),
+ c_v128_xor(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_and(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_and(a.v128[1], b.v128[1]),
+ c_v128_and(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_andn(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_andn(a.v128[1], b.v128[1]),
+ c_v128_andn(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_add_8(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_add_8(a.v128[1], b.v128[1]),
+ c_v128_add_8(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_add_16(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_add_16(a.v128[1], b.v128[1]),
+ c_v128_add_16(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_sadd_s16(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_sadd_s16(a.v128[1], b.v128[1]),
+ c_v128_sadd_s16(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_add_32(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_add_32(a.v128[1], b.v128[1]),
+ c_v128_add_32(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_padd_s16(c_v256 a) {
+ c_v256 t;
+ t.s32[0] = (int32_t)a.s16[0] + (int32_t)a.s16[1];
+ t.s32[1] = (int32_t)a.s16[2] + (int32_t)a.s16[3];
+ t.s32[2] = (int32_t)a.s16[4] + (int32_t)a.s16[5];
+ t.s32[3] = (int32_t)a.s16[6] + (int32_t)a.s16[7];
+ t.s32[4] = (int32_t)a.s16[8] + (int32_t)a.s16[9];
+ t.s32[5] = (int32_t)a.s16[10] + (int32_t)a.s16[11];
+ t.s32[6] = (int32_t)a.s16[12] + (int32_t)a.s16[13];
+ t.s32[7] = (int32_t)a.s16[14] + (int32_t)a.s16[15];
+ return t;
+}
+
+SIMD_INLINE c_v256 c_v256_sub_8(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_sub_8(a.v128[1], b.v128[1]),
+ c_v128_sub_8(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_ssub_u8(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_ssub_u8(a.v128[1], b.v128[1]),
+ c_v128_ssub_u8(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_ssub_s8(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_ssub_s8(a.v128[1], b.v128[1]),
+ c_v128_ssub_s8(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_sub_16(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_sub_16(a.v128[1], b.v128[1]),
+ c_v128_sub_16(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_ssub_s16(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_ssub_s16(a.v128[1], b.v128[1]),
+ c_v128_ssub_s16(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_sub_32(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_sub_32(a.v128[1], b.v128[1]),
+ c_v128_sub_32(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_abs_s16(c_v256 a) {
+ return c_v256_from_v128(c_v128_abs_s16(a.v128[1]), c_v128_abs_s16(a.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_mul_s16(c_v128 a, c_v128 b) {
+ c_v128 lo_bits = c_v128_mullo_s16(a, b);
+ c_v128 hi_bits = c_v128_mulhi_s16(a, b);
+ return c_v256_from_v128(c_v128_ziphi_16(hi_bits, lo_bits),
+ c_v128_ziplo_16(hi_bits, lo_bits));
+}
+
+SIMD_INLINE c_v256 c_v256_mullo_s16(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_mullo_s16(a.v128[1], b.v128[1]),
+ c_v128_mullo_s16(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_mulhi_s16(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_mulhi_s16(a.v128[1], b.v128[1]),
+ c_v128_mulhi_s16(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_mullo_s32(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_mullo_s32(a.v128[1], b.v128[1]),
+ c_v128_mullo_s32(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_madd_s16(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_madd_s16(a.v128[1], b.v128[1]),
+ c_v128_madd_s16(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_madd_us8(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_madd_us8(a.v128[1], b.v128[1]),
+ c_v128_madd_us8(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_avg_u8(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_avg_u8(a.v128[1], b.v128[1]),
+ c_v128_avg_u8(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_rdavg_u8(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_rdavg_u8(a.v128[1], b.v128[1]),
+ c_v128_rdavg_u8(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_avg_u16(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_avg_u16(a.v128[1], b.v128[1]),
+ c_v128_avg_u16(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_min_u8(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_min_u8(a.v128[1], b.v128[1]),
+ c_v128_min_u8(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_max_u8(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_max_u8(a.v128[1], b.v128[1]),
+ c_v128_max_u8(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_min_s8(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_min_s8(a.v128[1], b.v128[1]),
+ c_v128_min_s8(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_max_s8(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_max_s8(a.v128[1], b.v128[1]),
+ c_v128_max_s8(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_min_s16(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_min_s16(a.v128[1], b.v128[1]),
+ c_v128_min_s16(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_max_s16(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_max_s16(a.v128[1], b.v128[1]),
+ c_v128_max_s16(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_ziplo_8(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_ziphi_8(a.v128[0], b.v128[0]),
+ c_v128_ziplo_8(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_ziphi_8(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_ziphi_8(a.v128[1], b.v128[1]),
+ c_v128_ziplo_8(a.v128[1], b.v128[1]));
+}
+
+SIMD_INLINE c_v256 c_v256_ziplo_16(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_ziphi_16(a.v128[0], b.v128[0]),
+ c_v128_ziplo_16(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_ziphi_16(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_ziphi_16(a.v128[1], b.v128[1]),
+ c_v128_ziplo_16(a.v128[1], b.v128[1]));
+}
+
+SIMD_INLINE c_v256 c_v256_ziplo_32(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_ziphi_32(a.v128[0], b.v128[0]),
+ c_v128_ziplo_32(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_ziphi_32(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_ziphi_32(a.v128[1], b.v128[1]),
+ c_v128_ziplo_32(a.v128[1], b.v128[1]));
+}
+
+SIMD_INLINE c_v256 c_v256_ziplo_64(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_ziphi_64(a.v128[0], b.v128[0]),
+ c_v128_ziplo_64(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_ziphi_64(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_ziphi_64(a.v128[1], b.v128[1]),
+ c_v128_ziplo_64(a.v128[1], b.v128[1]));
+}
+
+SIMD_INLINE c_v256 c_v256_ziplo_128(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(a.v128[0], b.v128[0]);
+}
+
+SIMD_INLINE c_v256 c_v256_ziphi_128(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(a.v128[1], b.v128[1]);
+}
+
+SIMD_INLINE c_v256 c_v256_zip_8(c_v128 a, c_v128 b) {
+ return c_v256_from_v128(c_v128_ziphi_8(a, b), c_v128_ziplo_8(a, b));
+}
+
+SIMD_INLINE c_v256 c_v256_zip_16(c_v128 a, c_v128 b) {
+ return c_v256_from_v128(c_v128_ziphi_16(a, b), c_v128_ziplo_16(a, b));
+}
+
+SIMD_INLINE c_v256 c_v256_zip_32(c_v128 a, c_v128 b) {
+ return c_v256_from_v128(c_v128_ziphi_32(a, b), c_v128_ziplo_32(a, b));
+}
+
+SIMD_INLINE c_v256 _c_v256_unzip_8(c_v256 a, c_v256 b, int mode) {
+ c_v256 t;
+ int i;
+ if (mode) {
+ for (i = 0; i < 16; i++) {
+ t.u8[i] = a.u8[i * 2 + 1];
+ t.u8[i + 16] = b.u8[i * 2 + 1];
+ }
+ } else {
+ for (i = 0; i < 16; i++) {
+ t.u8[i] = b.u8[i * 2];
+ t.u8[i + 16] = a.u8[i * 2];
+ }
+ }
+ return t;
+}
+
+SIMD_INLINE c_v256 c_v256_unziplo_8(c_v256 a, c_v256 b) {
+ return CONFIG_BIG_ENDIAN ? _c_v256_unzip_8(a, b, 1)
+ : _c_v256_unzip_8(a, b, 0);
+}
+
+SIMD_INLINE c_v256 c_v256_unziphi_8(c_v256 a, c_v256 b) {
+ return CONFIG_BIG_ENDIAN ? _c_v256_unzip_8(b, a, 0)
+ : _c_v256_unzip_8(b, a, 1);
+}
+
+SIMD_INLINE c_v256 _c_v256_unzip_16(c_v256 a, c_v256 b, int mode) {
+ c_v256 t;
+ int i;
+ if (mode) {
+ for (i = 0; i < 8; i++) {
+ t.u16[i] = a.u16[i * 2 + 1];
+ t.u16[i + 8] = b.u16[i * 2 + 1];
+ }
+ } else {
+ for (i = 0; i < 8; i++) {
+ t.u16[i] = b.u16[i * 2];
+ t.u16[i + 8] = a.u16[i * 2];
+ }
+ }
+ return t;
+}
+
+SIMD_INLINE c_v256 c_v256_unziplo_16(c_v256 a, c_v256 b) {
+ return CONFIG_BIG_ENDIAN ? _c_v256_unzip_16(a, b, 1)
+ : _c_v256_unzip_16(a, b, 0);
+}
+
+SIMD_INLINE c_v256 c_v256_unziphi_16(c_v256 a, c_v256 b) {
+ return CONFIG_BIG_ENDIAN ? _c_v256_unzip_16(b, a, 0)
+ : _c_v256_unzip_16(b, a, 1);
+}
+
+SIMD_INLINE c_v256 _c_v256_unzip_32(c_v256 a, c_v256 b, int mode) {
+ c_v256 t;
+ if (mode) {
+ t.u32[7] = b.u32[7];
+ t.u32[6] = b.u32[5];
+ t.u32[5] = b.u32[3];
+ t.u32[4] = b.u32[1];
+ t.u32[3] = a.u32[7];
+ t.u32[2] = a.u32[5];
+ t.u32[1] = a.u32[3];
+ t.u32[0] = a.u32[1];
+ } else {
+ t.u32[7] = a.u32[6];
+ t.u32[6] = a.u32[4];
+ t.u32[5] = a.u32[2];
+ t.u32[4] = a.u32[0];
+ t.u32[3] = b.u32[6];
+ t.u32[2] = b.u32[4];
+ t.u32[1] = b.u32[2];
+ t.u32[0] = b.u32[0];
+ }
+ return t;
+}
+
+SIMD_INLINE c_v256 c_v256_unziplo_32(c_v256 a, c_v256 b) {
+ return CONFIG_BIG_ENDIAN ? _c_v256_unzip_32(a, b, 1)
+ : _c_v256_unzip_32(a, b, 0);
+}
+
+SIMD_INLINE c_v256 c_v256_unziphi_32(c_v256 a, c_v256 b) {
+ return CONFIG_BIG_ENDIAN ? _c_v256_unzip_32(b, a, 0)
+ : _c_v256_unzip_32(b, a, 1);
+}
+
+SIMD_INLINE c_v256 c_v256_unpack_u8_s16(c_v128 a) {
+ return c_v256_from_v128(c_v128_unpackhi_u8_s16(a), c_v128_unpacklo_u8_s16(a));
+}
+
+SIMD_INLINE c_v256 c_v256_unpacklo_u8_s16(c_v256 a) {
+ return c_v256_from_v128(c_v128_unpackhi_u8_s16(a.v128[0]),
+ c_v128_unpacklo_u8_s16(a.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_unpackhi_u8_s16(c_v256 a) {
+ return c_v256_from_v128(c_v128_unpackhi_u8_s16(a.v128[1]),
+ c_v128_unpacklo_u8_s16(a.v128[1]));
+}
+
+SIMD_INLINE c_v256 c_v256_pack_s32_s16(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_pack_s32_s16(a.v128[1], a.v128[0]),
+ c_v128_pack_s32_s16(b.v128[1], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_pack_s16_u8(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_pack_s16_u8(a.v128[1], a.v128[0]),
+ c_v128_pack_s16_u8(b.v128[1], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_pack_s16_s8(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_pack_s16_s8(a.v128[1], a.v128[0]),
+ c_v128_pack_s16_s8(b.v128[1], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_unpack_u16_s32(c_v128 a) {
+ return c_v256_from_v128(c_v128_unpackhi_u16_s32(a),
+ c_v128_unpacklo_u16_s32(a));
+}
+
+SIMD_INLINE c_v256 c_v256_unpack_s16_s32(c_v128 a) {
+ return c_v256_from_v128(c_v128_unpackhi_s16_s32(a),
+ c_v128_unpacklo_s16_s32(a));
+}
+
+SIMD_INLINE c_v256 c_v256_unpacklo_u16_s32(c_v256 a) {
+ return c_v256_from_v128(c_v128_unpackhi_u16_s32(a.v128[0]),
+ c_v128_unpacklo_u16_s32(a.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_unpacklo_s16_s32(c_v256 a) {
+ return c_v256_from_v128(c_v128_unpackhi_s16_s32(a.v128[0]),
+ c_v128_unpacklo_s16_s32(a.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_unpackhi_u16_s32(c_v256 a) {
+ return c_v256_from_v128(c_v128_unpackhi_u16_s32(a.v128[1]),
+ c_v128_unpacklo_u16_s32(a.v128[1]));
+}
+
+SIMD_INLINE c_v256 c_v256_unpackhi_s16_s32(c_v256 a) {
+ return c_v256_from_v128(c_v128_unpackhi_s16_s32(a.v128[1]),
+ c_v128_unpacklo_s16_s32(a.v128[1]));
+}
+
+SIMD_INLINE c_v256 c_v256_shuffle_8(c_v256 a, c_v256 pattern) {
+ c_v256 t;
+ int c;
+ for (c = 0; c < 32; c++) {
+ if (pattern.u8[c] & ~31) {
+ fprintf(stderr, "Undefined v256_shuffle_8 index %d/%d\n", pattern.u8[c],
+ c);
+ abort();
+ }
+ t.u8[c] = a.u8[CONFIG_BIG_ENDIAN ? 31 - (pattern.u8[c] & 31)
+ : pattern.u8[c] & 31];
+ }
+ return t;
+}
+
+// Pairwise / dual-lane shuffle: shuffle two 128 bit lates.
+SIMD_INLINE c_v256 c_v256_pshuffle_8(c_v256 a, c_v256 pattern) {
+ return c_v256_from_v128(
+ c_v128_shuffle_8(c_v256_high_v128(a), c_v256_high_v128(pattern)),
+ c_v128_shuffle_8(c_v256_low_v128(a), c_v256_low_v128(pattern)));
+}
+
+SIMD_INLINE c_v256 c_v256_cmpgt_s8(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_cmpgt_s8(a.v128[1], b.v128[1]),
+ c_v128_cmpgt_s8(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_cmplt_s8(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_cmplt_s8(a.v128[1], b.v128[1]),
+ c_v128_cmplt_s8(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_cmpeq_8(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_cmpeq_8(a.v128[1], b.v128[1]),
+ c_v128_cmpeq_8(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_cmpgt_s16(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_cmpgt_s16(a.v128[1], b.v128[1]),
+ c_v128_cmpgt_s16(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_cmplt_s16(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_cmplt_s16(a.v128[1], b.v128[1]),
+ c_v128_cmplt_s16(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_cmpeq_16(c_v256 a, c_v256 b) {
+ return c_v256_from_v128(c_v128_cmpeq_16(a.v128[1], b.v128[1]),
+ c_v128_cmpeq_16(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_shl_n_byte(c_v256 a, const unsigned int n) {
+ if (n < 16)
+ return c_v256_from_v128(c_v128_or(c_v128_shl_n_byte(a.v128[1], n),
+ c_v128_shr_n_byte(a.v128[0], 16 - n)),
+ c_v128_shl_n_byte(a.v128[0], n));
+ else if (n > 16)
+ return c_v256_from_v128(c_v128_shl_n_byte(a.v128[0], n - 16),
+ c_v128_zero());
+ else
+ return c_v256_from_v128(c_v256_low_v128(a), c_v128_zero());
+}
+
+SIMD_INLINE c_v256 c_v256_shr_n_byte(c_v256 a, const unsigned int n) {
+ if (n < 16)
+ return c_v256_from_v128(c_v128_shr_n_byte(a.v128[1], n),
+ c_v128_or(c_v128_shr_n_byte(a.v128[0], n),
+ c_v128_shl_n_byte(a.v128[1], 16 - n)));
+ else if (n > 16)
+ return c_v256_from_v128(c_v128_zero(),
+ c_v128_shr_n_byte(a.v128[1], n - 16));
+ else
+ return c_v256_from_v128(c_v128_zero(), c_v256_high_v128(a));
+}
+
+SIMD_INLINE c_v256 c_v256_align(c_v256 a, c_v256 b, const unsigned int c) {
+ if (simd_check && c > 31) {
+ fprintf(stderr, "Error: undefined alignment %d\n", c);
+ abort();
+ }
+ return c ? c_v256_or(c_v256_shr_n_byte(b, c), c_v256_shl_n_byte(a, 32 - c))
+ : b;
+}
+
+SIMD_INLINE c_v256 c_v256_shl_8(c_v256 a, const unsigned int c) {
+ return c_v256_from_v128(c_v128_shl_8(a.v128[1], c),
+ c_v128_shl_8(a.v128[0], c));
+}
+
+SIMD_INLINE c_v256 c_v256_shr_u8(c_v256 a, const unsigned int c) {
+ return c_v256_from_v128(c_v128_shr_u8(a.v128[1], c),
+ c_v128_shr_u8(a.v128[0], c));
+}
+
+SIMD_INLINE c_v256 c_v256_shr_s8(c_v256 a, const unsigned int c) {
+ return c_v256_from_v128(c_v128_shr_s8(a.v128[1], c),
+ c_v128_shr_s8(a.v128[0], c));
+}
+
+SIMD_INLINE c_v256 c_v256_shl_16(c_v256 a, const unsigned int c) {
+ return c_v256_from_v128(c_v128_shl_16(a.v128[1], c),
+ c_v128_shl_16(a.v128[0], c));
+}
+
+SIMD_INLINE c_v256 c_v256_shr_u16(c_v256 a, const unsigned int c) {
+ return c_v256_from_v128(c_v128_shr_u16(a.v128[1], c),
+ c_v128_shr_u16(a.v128[0], c));
+}
+
+SIMD_INLINE c_v256 c_v256_shr_s16(c_v256 a, const unsigned int c) {
+ return c_v256_from_v128(c_v128_shr_s16(a.v128[1], c),
+ c_v128_shr_s16(a.v128[0], c));
+}
+
+SIMD_INLINE c_v256 c_v256_shl_32(c_v256 a, const unsigned int c) {
+ return c_v256_from_v128(c_v128_shl_32(a.v128[1], c),
+ c_v128_shl_32(a.v128[0], c));
+}
+
+SIMD_INLINE c_v256 c_v256_shr_u32(c_v256 a, const unsigned int c) {
+ return c_v256_from_v128(c_v128_shr_u32(a.v128[1], c),
+ c_v128_shr_u32(a.v128[0], c));
+}
+
+SIMD_INLINE c_v256 c_v256_shr_s32(c_v256 a, const unsigned int c) {
+ return c_v256_from_v128(c_v128_shr_s32(a.v128[1], c),
+ c_v128_shr_s32(a.v128[0], c));
+}
+
+SIMD_INLINE c_v256 c_v256_shl_n_8(c_v256 a, const unsigned int n) {
+ return c_v256_shl_8(a, n);
+}
+
+SIMD_INLINE c_v256 c_v256_shl_n_16(c_v256 a, const unsigned int n) {
+ return c_v256_shl_16(a, n);
+}
+
+SIMD_INLINE c_v256 c_v256_shl_n_32(c_v256 a, const unsigned int n) {
+ return c_v256_shl_32(a, n);
+}
+
+SIMD_INLINE c_v256 c_v256_shr_n_u8(c_v256 a, const unsigned int n) {
+ return c_v256_shr_u8(a, n);
+}
+
+SIMD_INLINE c_v256 c_v256_shr_n_u16(c_v256 a, const unsigned int n) {
+ return c_v256_shr_u16(a, n);
+}
+
+SIMD_INLINE c_v256 c_v256_shr_n_u32(c_v256 a, const unsigned int n) {
+ return c_v256_shr_u32(a, n);
+}
+
+SIMD_INLINE c_v256 c_v256_shr_n_s8(c_v256 a, const unsigned int n) {
+ return c_v256_shr_s8(a, n);
+}
+
+SIMD_INLINE c_v256 c_v256_shr_n_s16(c_v256 a, const unsigned int n) {
+ return c_v256_shr_s16(a, n);
+}
+
+SIMD_INLINE c_v256 c_v256_shr_n_s32(c_v256 a, const unsigned int n) {
+ return c_v256_shr_s32(a, n);
+}
+
+#endif /* _V256_INTRINSICS_C_H */
diff --git a/aom_dsp/simd/v256_intrinsics_v128.h b/aom_dsp/simd/v256_intrinsics_v128.h
new file mode 100644
index 0000000..93cccce
--- /dev/null
+++ b/aom_dsp/simd/v256_intrinsics_v128.h
@@ -0,0 +1,525 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef _V256_INTRINSICS_V128_H
+#define _V256_INTRINSICS_V128_H
+
+#if HAVE_NEON
+#include "./v128_intrinsics_arm.h"
+#elif HAVE_SSE2
+#include "./v128_intrinsics_x86.h"
+#else
+#include "./v128_intrinsics.h"
+#endif
+
+typedef struct { v128 lo, hi; } v256;
+
+SIMD_INLINE uint32_t v256_low_u32(v256 a) { return v128_low_u32(a.lo); }
+
+SIMD_INLINE v64 v256_low_v64(v256 a) { return v128_low_v64(a.lo); }
+
+SIMD_INLINE v128 v256_low_v128(v256 a) { return a.lo; }
+
+SIMD_INLINE v128 v256_high_v128(v256 a) { return a.hi; }
+
+SIMD_INLINE v256 v256_from_v128(v128 hi, v128 lo) {
+ v256 t;
+ t.hi = hi;
+ t.lo = lo;
+ return t;
+}
+
+SIMD_INLINE v256 v256_from_64(uint64_t a, uint64_t b, uint64_t c, uint64_t d) {
+ return v256_from_v128(v128_from_64(a, b), v128_from_64(c, d));
+}
+
+SIMD_INLINE v256 v256_from_v64(v64 a, v64 b, v64 c, v64 d) {
+ return v256_from_v128(v128_from_v64(a, b), v128_from_v64(c, d));
+}
+
+SIMD_INLINE v256 v256_load_unaligned(const void *p) {
+ return v256_from_v128(v128_load_unaligned((uint8_t *)p + 16),
+ v128_load_unaligned(p));
+}
+
+SIMD_INLINE v256 v256_load_aligned(const void *p) {
+ return v256_from_v128(v128_load_aligned((uint8_t *)p + 16),
+ v128_load_aligned(p));
+}
+
+SIMD_INLINE void v256_store_unaligned(void *p, v256 a) {
+ v128_store_unaligned(p, a.lo);
+ v128_store_unaligned((uint8_t *)p + 16, a.hi);
+}
+
+SIMD_INLINE void v256_store_aligned(void *p, v256 a) {
+ v128_store_aligned(p, a.lo);
+ v128_store_aligned((uint8_t *)p + 16, a.hi);
+}
+
+SIMD_INLINE v256 v256_zero() {
+ return v256_from_v128(v128_zero(), v128_zero());
+}
+
+SIMD_INLINE v256 v256_dup_8(uint8_t x) {
+ v128 t = v128_dup_8(x);
+ return v256_from_v128(t, t);
+}
+
+SIMD_INLINE v256 v256_dup_16(uint16_t x) {
+ v128 t = v128_dup_16(x);
+ return v256_from_v128(t, t);
+}
+
+SIMD_INLINE v256 v256_dup_32(uint32_t x) {
+ v128 t = v128_dup_32(x);
+ return v256_from_v128(t, t);
+}
+
+SIMD_INLINE int64_t v256_dotp_s16(v256 a, v256 b) {
+ return v128_dotp_s16(a.hi, b.hi) + v128_dotp_s16(a.lo, b.lo);
+}
+
+SIMD_INLINE uint64_t v256_hadd_u8(v256 a) {
+ return v128_hadd_u8(a.hi) + v128_hadd_u8(a.lo);
+}
+
+typedef struct {
+ sad128_internal hi;
+ sad128_internal lo;
+} sad256_internal;
+
+SIMD_INLINE sad256_internal v256_sad_u8_init() {
+ sad256_internal t;
+ t.hi = v128_sad_u8_init();
+ t.lo = v128_sad_u8_init();
+ return t;
+}
+
+/* Implementation dependent return value. Result must be finalised with
+ v256_sad_u8_sum().
+ The result for more than 16 v256_sad_u8() calls is undefined. */
+SIMD_INLINE sad256_internal v256_sad_u8(sad256_internal s, v256 a, v256 b) {
+ sad256_internal t;
+ t.hi = v128_sad_u8(s.hi, a.hi, b.hi);
+ t.lo = v128_sad_u8(s.lo, a.lo, b.lo);
+ return t;
+}
+
+SIMD_INLINE uint32_t v256_sad_u8_sum(sad256_internal s) {
+ return v128_sad_u8_sum(s.hi) + v128_sad_u8_sum(s.lo);
+}
+
+typedef struct {
+ ssd128_internal hi;
+ ssd128_internal lo;
+} ssd256_internal;
+
+SIMD_INLINE ssd256_internal v256_ssd_u8_init() {
+ ssd256_internal t;
+ t.hi = v128_ssd_u8_init();
+ t.lo = v128_ssd_u8_init();
+ return t;
+}
+
+/* Implementation dependent return value. Result must be finalised with
+ * v256_ssd_u8_sum(). */
+SIMD_INLINE ssd256_internal v256_ssd_u8(ssd256_internal s, v256 a, v256 b) {
+ ssd256_internal t;
+ t.hi = v128_ssd_u8(s.hi, a.hi, b.hi);
+ t.lo = v128_ssd_u8(s.lo, a.lo, b.lo);
+ return t;
+}
+
+SIMD_INLINE uint32_t v256_ssd_u8_sum(ssd256_internal s) {
+ return v128_ssd_u8_sum(s.hi) + v128_ssd_u8_sum(s.lo);
+}
+
+SIMD_INLINE v256 v256_or(v256 a, v256 b) {
+ return v256_from_v128(v128_or(a.hi, b.hi), v128_or(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_xor(v256 a, v256 b) {
+ return v256_from_v128(v128_xor(a.hi, b.hi), v128_xor(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_and(v256 a, v256 b) {
+ return v256_from_v128(v128_and(a.hi, b.hi), v128_and(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_andn(v256 a, v256 b) {
+ return v256_from_v128(v128_andn(a.hi, b.hi), v128_andn(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_add_8(v256 a, v256 b) {
+ return v256_from_v128(v128_add_8(a.hi, b.hi), v128_add_8(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_add_16(v256 a, v256 b) {
+ return v256_from_v128(v128_add_16(a.hi, b.hi), v128_add_16(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_sadd_s16(v256 a, v256 b) {
+ return v256_from_v128(v128_sadd_s16(a.hi, b.hi), v128_sadd_s16(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_add_32(v256 a, v256 b) {
+ return v256_from_v128(v128_add_32(a.hi, b.hi), v128_add_32(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_padd_s16(v256 a) {
+ return v256_from_v128(v128_padd_s16(a.hi), v128_padd_s16(a.lo));
+}
+
+SIMD_INLINE v256 v256_sub_8(v256 a, v256 b) {
+ return v256_from_v128(v128_sub_8(a.hi, b.hi), v128_sub_8(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_ssub_u8(v256 a, v256 b) {
+ return v256_from_v128(v128_ssub_u8(a.hi, b.hi), v128_ssub_u8(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_ssub_s8(v256 a, v256 b) {
+ return v256_from_v128(v128_ssub_s8(a.hi, b.hi), v128_ssub_s8(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_sub_16(v256 a, v256 b) {
+ return v256_from_v128(v128_sub_16(a.hi, b.hi), v128_sub_16(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_ssub_s16(v256 a, v256 b) {
+ return v256_from_v128(v128_ssub_s16(a.hi, b.hi), v128_ssub_s16(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_sub_32(v256 a, v256 b) {
+ return v256_from_v128(v128_sub_32(a.hi, b.hi), v128_sub_32(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_abs_s16(v256 a) {
+ return v256_from_v128(v128_abs_s16(a.hi), v128_abs_s16(a.lo));
+}
+
+SIMD_INLINE v256 v256_mul_s16(v128 a, v128 b) {
+ v128 lo_bits = v128_mullo_s16(a, b);
+ v128 hi_bits = v128_mulhi_s16(a, b);
+ return v256_from_v128(v128_ziphi_16(hi_bits, lo_bits),
+ v128_ziplo_16(hi_bits, lo_bits));
+}
+
+SIMD_INLINE v256 v256_mullo_s16(v256 a, v256 b) {
+ return v256_from_v128(v128_mullo_s16(a.hi, b.hi), v128_mullo_s16(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_mulhi_s16(v256 a, v256 b) {
+ return v256_from_v128(v128_mulhi_s16(a.hi, b.hi), v128_mulhi_s16(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_mullo_s32(v256 a, v256 b) {
+ return v256_from_v128(v128_mullo_s32(a.hi, b.hi), v128_mullo_s32(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_madd_s16(v256 a, v256 b) {
+ return v256_from_v128(v128_madd_s16(a.hi, b.hi), v128_madd_s16(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_madd_us8(v256 a, v256 b) {
+ return v256_from_v128(v128_madd_us8(a.hi, b.hi), v128_madd_us8(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_avg_u8(v256 a, v256 b) {
+ return v256_from_v128(v128_avg_u8(a.hi, b.hi), v128_avg_u8(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_rdavg_u8(v256 a, v256 b) {
+ return v256_from_v128(v128_rdavg_u8(a.hi, b.hi), v128_rdavg_u8(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_avg_u16(v256 a, v256 b) {
+ return v256_from_v128(v128_avg_u16(a.hi, b.hi), v128_avg_u16(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_min_u8(v256 a, v256 b) {
+ return v256_from_v128(v128_min_u8(a.hi, b.hi), v128_min_u8(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_max_u8(v256 a, v256 b) {
+ return v256_from_v128(v128_max_u8(a.hi, b.hi), v128_max_u8(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_min_s8(v256 a, v256 b) {
+ return v256_from_v128(v128_min_s8(a.hi, b.hi), v128_min_s8(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_max_s8(v256 a, v256 b) {
+ return v256_from_v128(v128_max_s8(a.hi, b.hi), v128_max_s8(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_min_s16(v256 a, v256 b) {
+ return v256_from_v128(v128_min_s16(a.hi, b.hi), v128_min_s16(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_max_s16(v256 a, v256 b) {
+ return v256_from_v128(v128_max_s16(a.hi, b.hi), v128_max_s16(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_ziplo_8(v256 a, v256 b) {
+ return v256_from_v128(v128_ziphi_8(a.lo, b.lo), v128_ziplo_8(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_ziphi_8(v256 a, v256 b) {
+ return v256_from_v128(v128_ziphi_8(a.hi, b.hi), v128_ziplo_8(a.hi, b.hi));
+}
+
+SIMD_INLINE v256 v256_ziplo_16(v256 a, v256 b) {
+ return v256_from_v128(v128_ziphi_16(a.lo, b.lo), v128_ziplo_16(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_ziphi_16(v256 a, v256 b) {
+ return v256_from_v128(v128_ziphi_16(a.hi, b.hi), v128_ziplo_16(a.hi, b.hi));
+}
+
+SIMD_INLINE v256 v256_ziplo_32(v256 a, v256 b) {
+ return v256_from_v128(v128_ziphi_32(a.lo, b.lo), v128_ziplo_32(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_ziphi_32(v256 a, v256 b) {
+ return v256_from_v128(v128_ziphi_32(a.hi, b.hi), v128_ziplo_32(a.hi, b.hi));
+}
+
+SIMD_INLINE v256 v256_ziplo_64(v256 a, v256 b) {
+ return v256_from_v128(v128_ziphi_64(a.lo, b.lo), v128_ziplo_64(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_ziphi_64(v256 a, v256 b) {
+ return v256_from_v128(v128_ziphi_64(a.hi, b.hi), v128_ziplo_64(a.hi, b.hi));
+}
+
+SIMD_INLINE v256 v256_ziplo_128(v256 a, v256 b) {
+ return v256_from_v128(a.lo, b.lo);
+}
+
+SIMD_INLINE v256 v256_ziphi_128(v256 a, v256 b) {
+ return v256_from_v128(a.hi, b.hi);
+}
+
+SIMD_INLINE v256 v256_zip_8(v128 a, v128 b) {
+ return v256_from_v128(v128_ziphi_8(a, b), v128_ziplo_8(a, b));
+}
+
+SIMD_INLINE v256 v256_zip_16(v128 a, v128 b) {
+ return v256_from_v128(v128_ziphi_16(a, b), v128_ziplo_16(a, b));
+}
+
+SIMD_INLINE v256 v256_zip_32(v128 a, v128 b) {
+ return v256_from_v128(v128_ziphi_32(a, b), v128_ziplo_32(a, b));
+}
+
+SIMD_INLINE v256 v256_unziplo_8(v256 a, v256 b) {
+ return v256_from_v128(v128_unziplo_8(a.hi, a.lo), v128_unziplo_8(b.hi, b.lo));
+}
+
+SIMD_INLINE v256 v256_unziphi_8(v256 a, v256 b) {
+ return v256_from_v128(v128_unziphi_8(a.hi, a.lo), v128_unziphi_8(b.hi, b.lo));
+}
+
+SIMD_INLINE v256 v256_unziplo_16(v256 a, v256 b) {
+ return v256_from_v128(v128_unziplo_16(a.hi, a.lo),
+ v128_unziplo_16(b.hi, b.lo));
+}
+
+SIMD_INLINE v256 v256_unziphi_16(v256 a, v256 b) {
+ return v256_from_v128(v128_unziphi_16(a.hi, a.lo),
+ v128_unziphi_16(b.hi, b.lo));
+}
+
+SIMD_INLINE v256 v256_unziplo_32(v256 a, v256 b) {
+ return v256_from_v128(v128_unziplo_32(a.hi, a.lo),
+ v128_unziplo_32(b.hi, b.lo));
+}
+
+SIMD_INLINE v256 v256_unziphi_32(v256 a, v256 b) {
+ return v256_from_v128(v128_unziphi_32(a.hi, a.lo),
+ v128_unziphi_32(b.hi, b.lo));
+}
+
+SIMD_INLINE v256 v256_unpack_u8_s16(v128 a) {
+ return v256_from_v128(v128_unpackhi_u8_s16(a), v128_unpacklo_u8_s16(a));
+}
+
+SIMD_INLINE v256 v256_unpacklo_u8_s16(v256 a) {
+ return v256_from_v128(v128_unpackhi_u8_s16(a.lo), v128_unpacklo_u8_s16(a.lo));
+}
+
+SIMD_INLINE v256 v256_unpackhi_u8_s16(v256 a) {
+ return v256_from_v128(v128_unpackhi_u8_s16(a.hi), v128_unpacklo_u8_s16(a.hi));
+}
+
+SIMD_INLINE v256 v256_pack_s32_s16(v256 a, v256 b) {
+ return v256_from_v128(v128_pack_s32_s16(a.hi, a.lo),
+ v128_pack_s32_s16(b.hi, b.lo));
+}
+
+SIMD_INLINE v256 v256_pack_s16_u8(v256 a, v256 b) {
+ return v256_from_v128(v128_pack_s16_u8(a.hi, a.lo),
+ v128_pack_s16_u8(b.hi, b.lo));
+}
+
+SIMD_INLINE v256 v256_pack_s16_s8(v256 a, v256 b) {
+ return v256_from_v128(v128_pack_s16_s8(a.hi, a.lo),
+ v128_pack_s16_s8(b.hi, b.lo));
+}
+
+SIMD_INLINE v256 v256_unpack_u16_s32(v128 a) {
+ return v256_from_v128(v128_unpackhi_u16_s32(a), v128_unpacklo_u16_s32(a));
+}
+
+SIMD_INLINE v256 v256_unpack_s16_s32(v128 a) {
+ return v256_from_v128(v128_unpackhi_s16_s32(a), v128_unpacklo_s16_s32(a));
+}
+
+SIMD_INLINE v256 v256_unpacklo_u16_s32(v256 a) {
+ return v256_from_v128(v128_unpackhi_u16_s32(a.lo),
+ v128_unpacklo_u16_s32(a.lo));
+}
+
+SIMD_INLINE v256 v256_unpacklo_s16_s32(v256 a) {
+ return v256_from_v128(v128_unpackhi_s16_s32(a.lo),
+ v128_unpacklo_s16_s32(a.lo));
+}
+
+SIMD_INLINE v256 v256_unpackhi_u16_s32(v256 a) {
+ return v256_from_v128(v128_unpackhi_u16_s32(a.hi),
+ v128_unpacklo_u16_s32(a.hi));
+}
+
+SIMD_INLINE v256 v256_unpackhi_s16_s32(v256 a) {
+ return v256_from_v128(v128_unpackhi_s16_s32(a.hi),
+ v128_unpacklo_s16_s32(a.hi));
+}
+
+SIMD_INLINE v256 v256_shuffle_8(v256 a, v256 pattern) {
+ v128 c16 = v128_dup_8(16);
+ v128 maskhi = v128_cmplt_s8(pattern.hi, c16);
+ v128 masklo = v128_cmplt_s8(pattern.lo, c16);
+ return v256_from_v128(
+ v128_or(
+ v128_and(v128_shuffle_8(a.lo, pattern.hi), maskhi),
+ v128_andn(v128_shuffle_8(a.hi, v128_sub_8(pattern.hi, c16)), maskhi)),
+ v128_or(v128_and(v128_shuffle_8(a.lo, pattern.lo), masklo),
+ v128_andn(v128_shuffle_8(a.hi, v128_sub_8(pattern.lo, c16)),
+ masklo)));
+}
+
+SIMD_INLINE v256 v256_pshuffle_8(v256 a, v256 pattern) {
+ return v256_from_v128(
+ v128_shuffle_8(v256_high_v128(a), v256_high_v128(pattern)),
+ v128_shuffle_8(v256_low_v128(a), v256_low_v128(pattern)));
+}
+
+SIMD_INLINE v256 v256_cmpgt_s8(v256 a, v256 b) {
+ return v256_from_v128(v128_cmpgt_s8(a.hi, b.hi), v128_cmpgt_s8(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_cmplt_s8(v256 a, v256 b) {
+ return v256_from_v128(v128_cmplt_s8(a.hi, b.hi), v128_cmplt_s8(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_cmpeq_8(v256 a, v256 b) {
+ return v256_from_v128(v128_cmpeq_8(a.hi, b.hi), v128_cmpeq_8(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_cmpgt_s16(v256 a, v256 b) {
+ return v256_from_v128(v128_cmpgt_s16(a.hi, b.hi), v128_cmpgt_s16(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_cmplt_s16(v256 a, v256 b) {
+ return v256_from_v128(v128_cmplt_s16(a.hi, b.hi), v128_cmplt_s16(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_cmpeq_16(v256 a, v256 b) {
+ return v256_from_v128(v128_cmpeq_16(a.hi, b.hi), v128_cmpeq_16(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_shl_8(v256 a, const unsigned int c) {
+ return v256_from_v128(v128_shl_8(a.hi, c), v128_shl_8(a.lo, c));
+}
+
+SIMD_INLINE v256 v256_shr_u8(v256 a, const unsigned int c) {
+ return v256_from_v128(v128_shr_u8(a.hi, c), v128_shr_u8(a.lo, c));
+}
+
+SIMD_INLINE v256 v256_shr_s8(v256 a, const unsigned int c) {
+ return v256_from_v128(v128_shr_s8(a.hi, c), v128_shr_s8(a.lo, c));
+}
+
+SIMD_INLINE v256 v256_shl_16(v256 a, const unsigned int c) {
+ return v256_from_v128(v128_shl_16(a.hi, c), v128_shl_16(a.lo, c));
+}
+
+SIMD_INLINE v256 v256_shr_u16(v256 a, const unsigned int c) {
+ return v256_from_v128(v128_shr_u16(a.hi, c), v128_shr_u16(a.lo, c));
+}
+
+SIMD_INLINE v256 v256_shr_s16(v256 a, const unsigned int c) {
+ return v256_from_v128(v128_shr_s16(a.hi, c), v128_shr_s16(a.lo, c));
+}
+
+SIMD_INLINE v256 v256_shl_32(v256 a, const unsigned int c) {
+ return v256_from_v128(v128_shl_32(a.hi, c), v128_shl_32(a.lo, c));
+}
+
+SIMD_INLINE v256 v256_shr_u32(v256 a, const unsigned int c) {
+ return v256_from_v128(v128_shr_u32(a.hi, c), v128_shr_u32(a.lo, c));
+}
+
+SIMD_INLINE v256 v256_shr_s32(v256 a, const unsigned int c) {
+ return v256_from_v128(v128_shr_s32(a.hi, c), v128_shr_s32(a.lo, c));
+}
+
+/* These intrinsics require immediate values, so we must use #defines
+ to enforce that. */
+#define v256_shl_n_byte(a, n) \
+ ((n) < 16 ? v256_from_v128(v128_or(v128_shl_n_byte(a.hi, n), \
+ v128_shr_n_byte(a.lo, 16 - (n))), \
+ v128_shl_n_byte(a.lo, (n))) \
+ : v256_from_v128((n) > 16 ? v128_shl_n_byte(a.lo, (n)-16) : a.lo, \
+ v128_zero()))
+
+#define v256_shr_n_byte(a, n) \
+ ((n) < 16 ? v256_from_v128(v128_shr_n_byte(a.hi, n), \
+ v128_or(v128_shr_n_byte(a.lo, n), \
+ v128_shl_n_byte(a.hi, 16 - (n)))) \
+ : v256_from_v128(v128_zero(), \
+ (n) > 16 ? v128_shr_n_byte(a.hi, (n)-16) : a.hi))
+
+#define v256_align(a, b, c) \
+ ((c) ? v256_or(v256_shr_n_byte(b, c), v256_shl_n_byte(a, 32 - (c))) : b)
+
+#define v256_shl_n_8(a, n) \
+ v256_from_v128(v128_shl_n_8(a.hi, n), v128_shl_n_8(a.lo, n))
+#define v256_shl_n_16(a, n) \
+ v256_from_v128(v128_shl_n_16(a.hi, n), v128_shl_n_16(a.lo, n))
+#define v256_shl_n_32(a, n) \
+ v256_from_v128(v128_shl_n_32(a.hi, n), v128_shl_n_32(a.lo, n))
+#define v256_shr_n_u8(a, n) \
+ v256_from_v128(v128_shr_n_u8(a.hi, n), v128_shr_n_u8(a.lo, n))
+#define v256_shr_n_u16(a, n) \
+ v256_from_v128(v128_shr_n_u16(a.hi, n), v128_shr_n_u16(a.lo, n))
+#define v256_shr_n_u32(a, n) \
+ v256_from_v128(v128_shr_n_u32(a.hi, n), v128_shr_n_u32(a.lo, n))
+#define v256_shr_n_s8(a, n) \
+ v256_from_v128(v128_shr_n_s8(a.hi, n), v128_shr_n_s8(a.lo, n))
+#define v256_shr_n_s16(a, n) \
+ v256_from_v128(v128_shr_n_s16(a.hi, n), v128_shr_n_s16(a.lo, n))
+#define v256_shr_n_s32(a, n) \
+ v256_from_v128(v128_shr_n_s32(a.hi, n), v128_shr_n_s32(a.lo, n))
+
+#endif /* _V256_INTRINSICS_V128_H */
diff --git a/aom_dsp/simd/v256_intrinsics_x86.h b/aom_dsp/simd/v256_intrinsics_x86.h
new file mode 100644
index 0000000..b5bdb53
--- /dev/null
+++ b/aom_dsp/simd/v256_intrinsics_x86.h
@@ -0,0 +1,528 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef _V256_INTRINSICS_H
+#define _V256_INTRINSICS_H
+
+#if !defined(__AVX2__)
+
+#include "./v256_intrinsics_v128.h"
+
+#else
+
+// The _m256i type seems to cause problems for g++'s mangling prior to
+// version 5, but adding -fabi-version=0 fixes this.
+#if !defined(__clang__) && __GNUC__ < 5 && defined(__AVX2__) && \
+ defined(__cplusplus)
+#pragma GCC optimize "-fabi-version=0"
+#endif
+
+#include <immintrin.h>
+#include "./v128_intrinsics_x86.h"
+
+typedef __m256i v256;
+
+SIMD_INLINE uint32_t v256_low_u32(v256 a) {
+ return (uint32_t)_mm_cvtsi128_si32(_mm256_extracti128_si256(a, 0));
+}
+
+SIMD_INLINE v64 v256_low_v64(v256 a) {
+ return _mm_unpacklo_epi64(_mm256_extracti128_si256(a, 0), v64_zero());
+}
+
+SIMD_INLINE v128 v256_low_v128(v256 a) {
+ return _mm256_extracti128_si256(a, 0);
+}
+
+SIMD_INLINE v128 v256_high_v128(v256 a) {
+ return _mm256_extracti128_si256(a, 1);
+}
+
+SIMD_INLINE v256 v256_from_v128(v128 a, v128 b) {
+ // gcc seems to be missing _mm256_set_m128i()
+ return _mm256_insertf128_si256(
+ _mm256_insertf128_si256(_mm256_setzero_si256(), b, 0), a, 1);
+}
+
+SIMD_INLINE v256 v256_from_v64(v64 a, v64 b, v64 c, v64 d) {
+ return v256_from_v128(v128_from_v64(a, b), v128_from_v64(c, d));
+}
+
+SIMD_INLINE v256 v256_from_64(uint64_t a, uint64_t b, uint64_t c, uint64_t d) {
+ return v256_from_v128(v128_from_64(a, b), v128_from_64(c, d));
+}
+
+SIMD_INLINE v256 v256_load_aligned(const void *p) {
+ return _mm256_load_si256((const __m256i *)p);
+}
+
+SIMD_INLINE v256 v256_load_unaligned(const void *p) {
+ return _mm256_loadu_si256((const __m256i *)p);
+}
+
+SIMD_INLINE void v256_store_aligned(void *p, v256 a) {
+ _mm256_store_si256((__m256i *)p, a);
+}
+
+SIMD_INLINE void v256_store_unaligned(void *p, v256 a) {
+ _mm256_storeu_si256((__m256i *)p, a);
+}
+
+SIMD_INLINE v256 v256_zero() { return _mm256_setzero_si256(); }
+
+SIMD_INLINE v256 v256_dup_8(uint8_t x) { return _mm256_set1_epi8(x); }
+
+SIMD_INLINE v256 v256_dup_16(uint16_t x) { return _mm256_set1_epi16(x); }
+
+SIMD_INLINE v256 v256_dup_32(uint32_t x) { return _mm256_set1_epi32(x); }
+
+SIMD_INLINE v256 v256_add_8(v256 a, v256 b) { return _mm256_add_epi8(a, b); }
+
+SIMD_INLINE v256 v256_add_16(v256 a, v256 b) { return _mm256_add_epi16(a, b); }
+
+SIMD_INLINE v256 v256_sadd_s16(v256 a, v256 b) {
+ return _mm256_adds_epi16(a, b);
+}
+
+SIMD_INLINE v256 v256_add_32(v256 a, v256 b) { return _mm256_add_epi32(a, b); }
+
+SIMD_INLINE v256 v256_padd_s16(v256 a) {
+ return _mm256_madd_epi16(a, _mm256_set1_epi16(1));
+}
+
+SIMD_INLINE v256 v256_sub_8(v256 a, v256 b) { return _mm256_sub_epi8(a, b); }
+
+SIMD_INLINE v256 v256_ssub_u8(v256 a, v256 b) { return _mm256_subs_epu8(a, b); }
+
+SIMD_INLINE v256 v256_ssub_s8(v256 a, v256 b) { return _mm256_subs_epi8(a, b); }
+
+SIMD_INLINE v256 v256_sub_16(v256 a, v256 b) { return _mm256_sub_epi16(a, b); }
+
+SIMD_INLINE v256 v256_ssub_s16(v256 a, v256 b) {
+ return _mm256_subs_epi16(a, b);
+}
+
+SIMD_INLINE v256 v256_sub_32(v256 a, v256 b) { return _mm256_sub_epi32(a, b); }
+
+SIMD_INLINE v256 v256_abs_s16(v256 a) { return _mm256_abs_epi16(a); }
+
+// AVX doesn't have the direct intrinsics to zip/unzip 8, 16, 32 bit
+// lanes of lower or upper halves of a 256bit vector because the
+// unpack/pack intrinsics operate on the 256 bit input vector as 2
+// independent 128 bit vectors.
+SIMD_INLINE v256 v256_ziplo_8(v256 a, v256 b) {
+ return v256_from_v128(v128_ziphi_8(v256_low_v128(a), v256_low_v128(b)),
+ v128_ziplo_8(v256_low_v128(a), v256_low_v128(b)));
+}
+
+SIMD_INLINE v256 v256_ziphi_8(v256 a, v256 b) {
+ return v256_from_v128(v128_ziphi_8(v256_high_v128(a), v256_high_v128(b)),
+ v128_ziplo_8(v256_high_v128(a), v256_high_v128(b)));
+}
+
+SIMD_INLINE v256 v256_ziplo_16(v256 a, v256 b) {
+ return v256_from_v128(v128_ziphi_16(v256_low_v128(a), v256_low_v128(b)),
+ v128_ziplo_16(v256_low_v128(a), v256_low_v128(b)));
+}
+
+SIMD_INLINE v256 v256_ziphi_16(v256 a, v256 b) {
+ return v256_from_v128(v128_ziphi_16(v256_high_v128(a), v256_high_v128(b)),
+ v128_ziplo_16(v256_high_v128(a), v256_high_v128(b)));
+}
+
+SIMD_INLINE v256 v256_ziplo_32(v256 a, v256 b) {
+ return v256_from_v128(v128_ziphi_32(v256_low_v128(a), v256_low_v128(b)),
+ v128_ziplo_32(v256_low_v128(a), v256_low_v128(b)));
+}
+
+SIMD_INLINE v256 v256_ziphi_32(v256 a, v256 b) {
+ return v256_from_v128(v128_ziphi_32(v256_high_v128(a), v256_high_v128(b)),
+ v128_ziplo_32(v256_high_v128(a), v256_high_v128(b)));
+}
+
+SIMD_INLINE v256 v256_ziplo_64(v256 a, v256 b) {
+ return v256_from_v128(v128_ziphi_64(v256_low_v128(a), v256_low_v128(b)),
+ v128_ziplo_64(v256_low_v128(a), v256_low_v128(b)));
+}
+
+SIMD_INLINE v256 v256_ziphi_64(v256 a, v256 b) {
+ return v256_from_v128(v128_ziphi_64(v256_high_v128(a), v256_high_v128(b)),
+ v128_ziplo_64(v256_high_v128(a), v256_high_v128(b)));
+}
+
+SIMD_INLINE v256 v256_ziplo_128(v256 a, v256 b) {
+ return v256_from_v128(v256_low_v128(a), v256_low_v128(b));
+}
+
+SIMD_INLINE v256 v256_ziphi_128(v256 a, v256 b) {
+ return v256_from_v128(v256_high_v128(a), v256_high_v128(b));
+}
+
+SIMD_INLINE v256 v256_zip_8(v128 a, v128 b) {
+ return v256_from_v128(v128_ziphi_8(a, b), v128_ziplo_8(a, b));
+}
+
+SIMD_INLINE v256 v256_zip_16(v128 a, v128 b) {
+ return v256_from_v128(v128_ziphi_16(a, b), v128_ziplo_16(a, b));
+}
+
+SIMD_INLINE v256 v256_zip_32(v128 a, v128 b) {
+ return v256_from_v128(v128_ziphi_32(a, b), v128_ziplo_32(a, b));
+}
+
+SIMD_INLINE v256 v256_unziplo_8(v256 a, v256 b) {
+ return v256_from_v128(v128_unziplo_8(v256_high_v128(a), v256_low_v128(a)),
+ v128_unziplo_8(v256_high_v128(b), v256_low_v128(b)));
+}
+
+SIMD_INLINE v256 v256_unziphi_8(v256 a, v256 b) {
+ return v256_from_v128(v128_unziphi_8(v256_high_v128(a), v256_low_v128(a)),
+ v128_unziphi_8(v256_high_v128(b), v256_low_v128(b)));
+}
+
+SIMD_INLINE v256 v256_unziplo_16(v256 a, v256 b) {
+ return v256_from_v128(v128_unziplo_16(v256_high_v128(a), v256_low_v128(a)),
+ v128_unziplo_16(v256_high_v128(b), v256_low_v128(b)));
+}
+
+SIMD_INLINE v256 v256_unziphi_16(v256 a, v256 b) {
+ return v256_from_v128(v128_unziphi_16(v256_high_v128(a), v256_low_v128(a)),
+ v128_unziphi_16(v256_high_v128(b), v256_low_v128(b)));
+}
+
+SIMD_INLINE v256 v256_unziplo_32(v256 a, v256 b) {
+ return v256_from_v128(v128_unziplo_32(v256_high_v128(a), v256_low_v128(a)),
+ v128_unziplo_32(v256_high_v128(b), v256_low_v128(b)));
+}
+
+SIMD_INLINE v256 v256_unziphi_32(v256 a, v256 b) {
+ return v256_from_v128(v128_unziphi_32(v256_high_v128(a), v256_low_v128(a)),
+ v128_unziphi_32(v256_high_v128(b), v256_low_v128(b)));
+}
+
+SIMD_INLINE v256 v256_unpack_u8_s16(v128 a) {
+ return v256_from_v128(v128_unpackhi_u8_s16(a), v128_unpacklo_u8_s16(a));
+}
+
+SIMD_INLINE v256 v256_unpacklo_u8_s16(v256 a) {
+ return v256_from_v128(v128_unpackhi_u8_s16(v256_low_v128(a)),
+ v128_unpacklo_u8_s16(v256_low_v128(a)));
+}
+
+SIMD_INLINE v256 v256_unpackhi_u8_s16(v256 a) {
+ return v256_from_v128(v128_unpackhi_u8_s16(v256_high_v128(a)),
+ v128_unpacklo_u8_s16(v256_high_v128(a)));
+}
+
+SIMD_INLINE v256 v256_pack_s32_s16(v256 a, v256 b) {
+ return v256_from_v128(v128_pack_s32_s16(v256_high_v128(a), v256_low_v128(a)),
+ v128_pack_s32_s16(v256_high_v128(b), v256_low_v128(b)));
+}
+
+SIMD_INLINE v256 v256_pack_s16_u8(v256 a, v256 b) {
+ return v256_from_v128(v128_pack_s16_u8(v256_high_v128(a), v256_low_v128(a)),
+ v128_pack_s16_u8(v256_high_v128(b), v256_low_v128(b)));
+}
+
+SIMD_INLINE v256 v256_pack_s16_s8(v256 a, v256 b) {
+ return v256_from_v128(v128_pack_s16_s8(v256_high_v128(a), v256_low_v128(a)),
+ v128_pack_s16_s8(v256_high_v128(b), v256_low_v128(b)));
+}
+
+SIMD_INLINE v256 v256_unpack_u16_s32(v128 a) {
+ return v256_from_v128(v128_unpackhi_u16_s32(a), v128_unpacklo_u16_s32(a));
+}
+
+SIMD_INLINE v256 v256_unpack_s16_s32(v128 a) {
+ return v256_from_v128(v128_unpackhi_s16_s32(a), v128_unpacklo_s16_s32(a));
+}
+
+SIMD_INLINE v256 v256_unpacklo_u16_s32(v256 a) {
+ return v256_from_v128(v128_unpackhi_u16_s32(v256_low_v128(a)),
+ v128_unpacklo_u16_s32(v256_low_v128(a)));
+}
+
+SIMD_INLINE v256 v256_unpacklo_s16_s32(v256 a) {
+ return v256_from_v128(v128_unpackhi_s16_s32(v256_low_v128(a)),
+ v128_unpacklo_s16_s32(v256_low_v128(a)));
+}
+
+SIMD_INLINE v256 v256_unpackhi_u16_s32(v256 a) {
+ return v256_from_v128(v128_unpackhi_u16_s32(v256_high_v128(a)),
+ v128_unpacklo_u16_s32(v256_high_v128(a)));
+}
+
+SIMD_INLINE v256 v256_unpackhi_s16_s32(v256 a) {
+ return v256_from_v128(v128_unpackhi_s16_s32(v256_high_v128(a)),
+ v128_unpacklo_s16_s32(v256_high_v128(a)));
+}
+SIMD_INLINE v256 v256_shuffle_8(v256 a, v256 pattern) {
+ v128 c16 = v128_dup_8(16);
+ v128 hi = v256_high_v128(pattern);
+ v128 lo = v256_low_v128(pattern);
+ v128 maskhi = v128_cmplt_s8(hi, c16);
+ v128 masklo = v128_cmplt_s8(lo, c16);
+ return v256_from_v128(
+ v128_or(v128_and(v128_shuffle_8(v256_low_v128(a), hi), maskhi),
+ v128_andn(v128_shuffle_8(v256_high_v128(a), v128_sub_8(hi, c16)),
+ maskhi)),
+ v128_or(v128_and(v128_shuffle_8(v256_low_v128(a), lo), masklo),
+ v128_andn(v128_shuffle_8(v256_high_v128(a), v128_sub_8(lo, c16)),
+ masklo)));
+}
+
+SIMD_INLINE v256 v256_pshuffle_8(v256 a, v256 pattern) {
+ return _mm256_shuffle_epi8(a, pattern);
+}
+
+SIMD_INLINE int64_t v256_dotp_s16(v256 a, v256 b) {
+ v256 r = _mm256_madd_epi16(a, b);
+#if defined(__x86_64__)
+ v128 t;
+ r = _mm256_add_epi64(_mm256_cvtepi32_epi64(v256_high_v128(r)),
+ _mm256_cvtepi32_epi64(v256_low_v128(r)));
+ t = v256_low_v128(_mm256_add_epi64(
+ r, _mm256_permute2x128_si256(r, r, _MM_SHUFFLE(2, 0, 0, 1))));
+ return _mm_cvtsi128_si64(_mm_add_epi64(t, _mm_srli_si128(t, 8)));
+#else
+ v128 l = v256_low_v128(r);
+ v128 h = v256_high_v128(r);
+ return (int64_t)_mm_cvtsi128_si32(l) +
+ (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(l, 4)) +
+ (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(l, 8)) +
+ (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(l, 12)) +
+ (int64_t)_mm_cvtsi128_si32(h) +
+ (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(h, 4)) +
+ (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(h, 8)) +
+ (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(h, 12));
+#endif
+}
+
+SIMD_INLINE uint64_t v256_hadd_u8(v256 a) {
+ v256 t = _mm256_sad_epu8(a, _mm256_setzero_si256());
+ v128 lo = v256_low_v128(t);
+ v128 hi = v256_high_v128(t);
+ lo = v128_add_32(lo, hi);
+ return v64_low_u32(v128_low_v64(lo)) + v128_low_u32(v128_high_v64(lo));
+}
+
+typedef v256 sad256_internal;
+
+SIMD_INLINE sad256_internal v256_sad_u8_init() {
+ return _mm256_setzero_si256();
+}
+
+/* Implementation dependent return value. Result must be finalised with
+ v256_sad_sum().
+ The result for more than 32 v256_sad_u8() calls is undefined. */
+SIMD_INLINE sad256_internal v256_sad_u8(sad256_internal s, v256 a, v256 b) {
+ return _mm256_add_epi64(s, _mm256_sad_epu8(a, b));
+}
+
+SIMD_INLINE uint32_t v256_sad_u8_sum(sad256_internal s) {
+ v256 t = _mm256_add_epi32(s, _mm256_unpackhi_epi64(s, s));
+ return v128_low_u32(_mm_add_epi32(v256_high_v128(t), v256_low_v128(t)));
+}
+
+typedef v256 ssd256_internal;
+
+SIMD_INLINE ssd256_internal v256_ssd_u8_init() {
+ return _mm256_setzero_si256();
+}
+
+/* Implementation dependent return value. Result must be finalised with
+ * v256_ssd_sum(). */
+SIMD_INLINE ssd256_internal v256_ssd_u8(ssd256_internal s, v256 a, v256 b) {
+ v256 l = _mm256_sub_epi16(_mm256_unpacklo_epi8(a, _mm256_setzero_si256()),
+ _mm256_unpacklo_epi8(b, _mm256_setzero_si256()));
+ v256 h = _mm256_sub_epi16(_mm256_unpackhi_epi8(a, _mm256_setzero_si256()),
+ _mm256_unpackhi_epi8(b, _mm256_setzero_si256()));
+ v256 rl = _mm256_madd_epi16(l, l);
+ v256 rh = _mm256_madd_epi16(h, h);
+ v128 c = _mm_cvtsi32_si128(32);
+ rl = _mm256_add_epi32(rl, _mm256_srli_si256(rl, 8));
+ rl = _mm256_add_epi32(rl, _mm256_srli_si256(rl, 4));
+ rh = _mm256_add_epi32(rh, _mm256_srli_si256(rh, 8));
+ rh = _mm256_add_epi32(rh, _mm256_srli_si256(rh, 4));
+ return _mm256_add_epi64(
+ s,
+ _mm256_srl_epi64(_mm256_sll_epi64(_mm256_unpacklo_epi64(rl, rh), c), c));
+}
+
+SIMD_INLINE uint32_t v256_ssd_u8_sum(ssd256_internal s) {
+ v256 t = _mm256_add_epi32(s, _mm256_unpackhi_epi64(s, s));
+ return v128_low_u32(_mm_add_epi32(v256_high_v128(t), v256_low_v128(t)));
+}
+
+SIMD_INLINE v256 v256_or(v256 a, v256 b) { return _mm256_or_si256(a, b); }
+
+SIMD_INLINE v256 v256_xor(v256 a, v256 b) { return _mm256_xor_si256(a, b); }
+
+SIMD_INLINE v256 v256_and(v256 a, v256 b) { return _mm256_and_si256(a, b); }
+
+SIMD_INLINE v256 v256_andn(v256 a, v256 b) { return _mm256_andnot_si256(b, a); }
+
+SIMD_INLINE v256 v256_mul_s16(v64 a, v64 b) {
+ v128 lo_bits = v128_mullo_s16(a, b);
+ v128 hi_bits = v128_mulhi_s16(a, b);
+ return v256_from_v128(v128_ziphi_16(hi_bits, lo_bits),
+ v128_ziplo_16(hi_bits, lo_bits));
+}
+
+SIMD_INLINE v256 v256_mullo_s16(v256 a, v256 b) {
+ return _mm256_mullo_epi16(a, b);
+}
+
+SIMD_INLINE v256 v256_mulhi_s16(v256 a, v256 b) {
+ return _mm256_mulhi_epi16(a, b);
+}
+
+SIMD_INLINE v256 v256_mullo_s32(v256 a, v256 b) {
+ return _mm256_mullo_epi32(a, b);
+}
+
+SIMD_INLINE v256 v256_madd_s16(v256 a, v256 b) {
+ return _mm256_madd_epi16(a, b);
+}
+
+SIMD_INLINE v256 v256_madd_us8(v256 a, v256 b) {
+ return _mm256_maddubs_epi16(a, b);
+}
+
+SIMD_INLINE v256 v256_avg_u8(v256 a, v256 b) { return _mm256_avg_epu8(a, b); }
+
+SIMD_INLINE v256 v256_rdavg_u8(v256 a, v256 b) {
+ return _mm256_sub_epi8(
+ _mm256_avg_epu8(a, b),
+ _mm256_and_si256(_mm256_xor_si256(a, b), v256_dup_8(1)));
+}
+
+SIMD_INLINE v256 v256_avg_u16(v256 a, v256 b) { return _mm256_avg_epu16(a, b); }
+
+SIMD_INLINE v256 v256_min_u8(v256 a, v256 b) { return _mm256_min_epu8(a, b); }
+
+SIMD_INLINE v256 v256_max_u8(v256 a, v256 b) { return _mm256_max_epu8(a, b); }
+
+SIMD_INLINE v256 v256_min_s8(v256 a, v256 b) { return _mm256_min_epi8(a, b); }
+
+SIMD_INLINE v256 v256_max_s8(v256 a, v256 b) { return _mm256_max_epi8(a, b); }
+
+SIMD_INLINE v256 v256_min_s16(v256 a, v256 b) { return _mm256_min_epi16(a, b); }
+
+SIMD_INLINE v256 v256_max_s16(v256 a, v256 b) { return _mm256_max_epi16(a, b); }
+
+SIMD_INLINE v256 v256_cmpgt_s8(v256 a, v256 b) {
+ return _mm256_cmpgt_epi8(a, b);
+}
+
+SIMD_INLINE v256 v256_cmplt_s8(v256 a, v256 b) {
+ return v256_andn(_mm256_cmpgt_epi8(b, a), _mm256_cmpeq_epi8(b, a));
+}
+
+SIMD_INLINE v256 v256_cmpeq_8(v256 a, v256 b) {
+ return _mm256_cmpeq_epi8(a, b);
+}
+
+SIMD_INLINE v256 v256_cmpgt_s16(v256 a, v256 b) {
+ return _mm256_cmpgt_epi16(a, b);
+}
+
+SIMD_INLINE v256 v256_cmplt_s16(v256 a, v256 b) {
+ return v256_andn(_mm256_cmpgt_epi16(b, a), _mm256_cmpeq_epi16(b, a));
+}
+
+SIMD_INLINE v256 v256_cmpeq_16(v256 a, v256 b) {
+ return _mm256_cmpeq_epi16(a, b);
+}
+
+SIMD_INLINE v256 v256_shl_8(v256 a, unsigned int c) {
+ return _mm256_and_si256(_mm256_set1_epi8((uint8_t)(0xff << c)),
+ _mm256_sll_epi16(a, _mm_cvtsi32_si128(c)));
+}
+
+SIMD_INLINE v256 v256_shr_u8(v256 a, unsigned int c) {
+ return _mm256_and_si256(_mm256_set1_epi8(0xff >> c),
+ _mm256_srl_epi16(a, _mm_cvtsi32_si128(c)));
+}
+
+SIMD_INLINE v256 v256_shr_s8(v256 a, unsigned int c) {
+ __m128i x = _mm_cvtsi32_si128(c + 8);
+ return _mm256_packs_epi16(_mm256_sra_epi16(_mm256_unpacklo_epi8(a, a), x),
+ _mm256_sra_epi16(_mm256_unpackhi_epi8(a, a), x));
+}
+
+SIMD_INLINE v256 v256_shl_16(v256 a, unsigned int c) {
+ return _mm256_sll_epi16(a, _mm_cvtsi32_si128(c));
+}
+
+SIMD_INLINE v256 v256_shr_u16(v256 a, unsigned int c) {
+ return _mm256_srl_epi16(a, _mm_cvtsi32_si128(c));
+}
+
+SIMD_INLINE v256 v256_shr_s16(v256 a, unsigned int c) {
+ return _mm256_sra_epi16(a, _mm_cvtsi32_si128(c));
+}
+
+SIMD_INLINE v256 v256_shl_32(v256 a, unsigned int c) {
+ return _mm256_sll_epi32(a, _mm_cvtsi32_si128(c));
+}
+
+SIMD_INLINE v256 v256_shr_u32(v256 a, unsigned int c) {
+ return _mm256_srl_epi32(a, _mm_cvtsi32_si128(c));
+}
+
+SIMD_INLINE v256 v256_shr_s32(v256 a, unsigned int c) {
+ return _mm256_sra_epi32(a, _mm_cvtsi32_si128(c));
+}
+
+/* These intrinsics require immediate values, so we must use #defines
+ to enforce that. */
+// _mm256_slli_si256 works on 128 bit lanes and can't be used
+#define v256_shl_n_byte(a, n) \
+ ((n) < 16 \
+ ? v256_from_v128(v128_or(v128_shl_n_byte(v256_high_v128(a), n), \
+ v128_shr_n_byte(v256_low_v128(a), 16 - (n))), \
+ v128_shl_n_byte(v256_low_v128(a), n)) \
+ : v256_from_v128(v128_shl_n_byte(v256_low_v128(a), (n)-16), \
+ v128_zero()))
+
+// _mm256_srli_si256 works on 128 bit lanes and can't be used
+#define v256_shr_n_byte(a, n) \
+ ((n) < 16 \
+ ? _mm256_alignr_epi8( \
+ _mm256_permute2x128_si256(a, a, _MM_SHUFFLE(2, 0, 0, 1)), a, n) \
+ : ((n) > 16 \
+ ? _mm256_srli_si256( \
+ _mm256_permute2x128_si256(a, a, _MM_SHUFFLE(2, 0, 0, 1)), \
+ (n)-16) \
+ : _mm256_permute2x128_si256(a, a, _MM_SHUFFLE(2, 0, 0, 1))))
+
+// _mm256_alignr_epi8 works on two 128 bit lanes and can't be used
+#define v256_align(a, b, c) \
+ ((c) ? v256_or(v256_shr_n_byte(b, c), v256_shl_n_byte(a, 32 - c)) : b)
+
+#define v256_shl_n_8(a, c) \
+ _mm256_and_si256(_mm256_set1_epi8((uint8_t)(0xff << (c))), \
+ _mm256_slli_epi16(a, c))
+#define v256_shr_n_u8(a, c) \
+ _mm256_and_si256(_mm256_set1_epi8(0xff >> (c)), _mm256_srli_epi16(a, c))
+#define v256_shr_n_s8(a, c) \
+ _mm256_packs_epi16(_mm256_srai_epi16(_mm256_unpacklo_epi8(a, a), (c) + 8), \
+ _mm256_srai_epi16(_mm256_unpackhi_epi8(a, a), (c) + 8))
+#define v256_shl_n_16(a, c) _mm256_slli_epi16(a, c)
+#define v256_shr_n_u16(a, c) _mm256_srli_epi16(a, c)
+#define v256_shr_n_s16(a, c) _mm256_srai_epi16(a, c)
+#define v256_shl_n_32(a, c) _mm256_slli_epi32(a, c)
+#define v256_shr_n_u32(a, c) _mm256_srli_epi32(a, c)
+#define v256_shr_n_s32(a, c) _mm256_srai_epi32(a, c)
+#endif
+
+#endif /* _V256_INTRINSICS_H */
