Optimize fixed-point celt_inner_prod() and dual_inner_prod() for ARM NEON
authorLinfeng Zhang <linfengz@google.com>
Wed, 7 Sep 2016 23:43:54 +0000 (16:43 -0700)
committerJean-Marc Valin <jmvalin@jmvalin.ca>
Tue, 6 Jun 2017 21:12:20 +0000 (17:12 -0400)
This optimization is bit exact with C functions.

Change-Id: Ia9ce6dd3c20d2f56dbd43ddc02d1a6fd6554608d

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>
celt/arm/arm_celt_map.c
celt/arm/pitch_arm.h
celt/arm/pitch_neon_intr.c [new file with mode: 0644]
celt_sources.mk

index cc6eb45..7c356b9 100644 (file)
 #if defined(OPUS_HAVE_RTCD)
 
 # if defined(FIXED_POINT)
+
+# if defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && !defined(OPUS_ARM_PRESUME_NEON_INTR)
+opus_val32 (*const CELT_INNER_PROD_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *x, const opus_val16 *y, int N) = {
+  celt_inner_prod_c,   /* ARMv4 */
+  celt_inner_prod_c,   /* EDSP */
+  celt_inner_prod_c,   /* Media */
+  celt_inner_prod_neon /* NEON */
+};
+
+void (*const DUAL_INNER_PROD_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *x, const opus_val16 *y01, const opus_val16 *y02,
+      int N, opus_val32 *xy1, opus_val32 *xy2) = {
+  dual_inner_prod_c,   /* ARMv4 */
+  dual_inner_prod_c,   /* EDSP */
+  dual_inner_prod_c,   /* Media */
+  dual_inner_prod_neon /* NEON */
+};
+# endif
+
 #  if ((defined(OPUS_ARM_MAY_HAVE_NEON) && !defined(OPUS_ARM_PRESUME_NEON)) || \
     (defined(OPUS_ARM_MAY_HAVE_MEDIA) && !defined(OPUS_ARM_PRESUME_MEDIA)) || \
     (defined(OPUS_ARM_MAY_HAVE_EDSP) && !defined(OPUS_ARM_PRESUME_EDSP)))
index d8b022e..ea2f590 100644 (file)
 
 # if defined(FIXED_POINT)
 
+# if defined(OPUS_ARM_MAY_HAVE_NEON_INTR)
+opus_val32 celt_inner_prod_neon(const opus_val16 *x, const opus_val16 *y, int N);
+void dual_inner_prod_neon(const opus_val16 *x, const opus_val16 *y01,
+        const opus_val16 *y02, int N, opus_val32 *xy1, opus_val32 *xy2);
+
+#  if !defined(OPUS_HAVE_RTCD) && defined(OPUS_ARM_PRESUME_NEON)
+#   define OVERRIDE_CELT_INNER_PROD (1)
+#   define OVERRIDE_DUAL_INNER_PROD (1)
+#   define celt_inner_prod(x, y, N, arch) ((void)(arch), PRESUME_NEON(celt_inner_prod)(x, y, N))
+#   define dual_inner_prod(x, y01, y02, N, xy1, xy2, arch) ((void)(arch), PRESUME_NEON(dual_inner_prod)(x, y01, y02, N, xy1, xy2))
+#  endif
+# endif
+
+# if !defined(OVERRIDE_CELT_INNER_PROD)
+#  if defined(OPUS_HAVE_RTCD) && (defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && !defined(OPUS_ARM_PRESUME_NEON_INTR))
+extern opus_val32 (*const CELT_INNER_PROD_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *x, const opus_val16 *y, int N);
+#   define OVERRIDE_CELT_INNER_PROD (1)
+#   define celt_inner_prod(x, y, N, arch) ((*CELT_INNER_PROD_IMPL[(arch)&OPUS_ARCHMASK])(x, y, N))
+#  elif defined(OPUS_ARM_PRESUME_NEON_INTR)
+#   define OVERRIDE_CELT_INNER_PROD (1)
+#   define celt_inner_prod(x, y, N, arch) ((void)(arch), celt_inner_prod_neon(x, y, N))
+#  endif
+# endif
+
+# if !defined(OVERRIDE_DUAL_INNER_PROD)
+#  if defined(OPUS_HAVE_RTCD) && (defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && !defined(OPUS_ARM_PRESUME_NEON_INTR))
+extern void (*const DUAL_INNER_PROD_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *x,
+        const opus_val16 *y01, const opus_val16 *y02, int N, opus_val32 *xy1, opus_val32 *xy2);
+#   define OVERRIDE_DUAL_INNER_PROD (1)
+#   define dual_inner_prod(x, y01, y02, N, xy1, xy2, arch) ((*DUAL_INNER_PROD_IMPL[(arch)&OPUS_ARCHMASK])(x, y01, y02, N, xy1, xy2))
+#  elif defined(OPUS_ARM_PRESUME_NEON_INTR)
+#   define OVERRIDE_DUAL_INNER_PROD (1)
+#   define dual_inner_prod(x, y01, y02, N, xy1, xy2, arch) ((void)(arch), dual_inner_prod_neon(x, y01, y02, N, xy1, xy2))
+#  endif
+# endif
+
 #  if defined(OPUS_ARM_MAY_HAVE_NEON)
 opus_val32 celt_pitch_xcorr_neon(const opus_val16 *_x, const opus_val16 *_y,
     opus_val32 *xcorr, int len, int max_pitch, int arch);
diff --git a/celt/arm/pitch_neon_intr.c b/celt/arm/pitch_neon_intr.c
new file mode 100644 (file)
index 0000000..ad18d2d
--- /dev/null
@@ -0,0 +1,129 @@
+/***********************************************************************
+Copyright (c) 2017 Google Inc.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+- Redistributions of source code must retain the above copyright notice,
+this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+***********************************************************************/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include <arm_neon.h>
+#include "pitch.h"
+
+#ifdef FIXED_POINT
+
+opus_val32 celt_inner_prod_neon(const opus_val16 *x, const opus_val16 *y, int N)
+{
+    int i;
+    opus_val32 xy;
+    int16x8_t x_s16x8, y_s16x8;
+    int32x4_t xy_s32x4 = vdupq_n_s32(0);
+    int64x2_t xy_s64x2;
+    int64x1_t xy_s64x1;
+
+    for (i = 0; i < N - 7; i += 8) {
+        x_s16x8  = vld1q_s16(&x[i]);
+        y_s16x8  = vld1q_s16(&y[i]);
+        xy_s32x4 = vmlal_s16(xy_s32x4, vget_low_s16 (x_s16x8), vget_low_s16 (y_s16x8));
+        xy_s32x4 = vmlal_s16(xy_s32x4, vget_high_s16(x_s16x8), vget_high_s16(y_s16x8));
+    }
+
+    if (N - i >= 4) {
+        const int16x4_t x_s16x4 = vld1_s16(&x[i]);
+        const int16x4_t y_s16x4 = vld1_s16(&y[i]);
+        xy_s32x4 = vmlal_s16(xy_s32x4, x_s16x4, y_s16x4);
+        i += 4;
+    }
+
+    xy_s64x2 = vpaddlq_s32(xy_s32x4);
+    xy_s64x1 = vadd_s64(vget_low_s64(xy_s64x2), vget_high_s64(xy_s64x2));
+    xy       = vget_lane_s32(vreinterpret_s32_s64(xy_s64x1), 0);
+
+    for (; i < N; i++) {
+        xy = MAC16_16(xy, x[i], y[i]);
+    }
+
+#ifdef OPUS_CHECK_ASM
+    celt_assert(celt_inner_prod_c(x, y, N) == xy);
+#endif
+
+    return xy;
+}
+
+void dual_inner_prod_neon(const opus_val16 *x, const opus_val16 *y01, const opus_val16 *y02,
+        int N, opus_val32 *xy1, opus_val32 *xy2)
+{
+    int i;
+    opus_val32 xy01, xy02;
+    int16x8_t x_s16x8, y01_s16x8, y02_s16x8;
+    int32x4_t xy01_s32x4 = vdupq_n_s32(0);
+    int32x4_t xy02_s32x4 = vdupq_n_s32(0);
+    int64x2_t xy01_s64x2, xy02_s64x2;
+    int64x1_t xy01_s64x1, xy02_s64x1;
+
+    for (i = 0; i < N - 7; i += 8) {
+        x_s16x8    = vld1q_s16(&x[i]);
+        y01_s16x8  = vld1q_s16(&y01[i]);
+        y02_s16x8  = vld1q_s16(&y02[i]);
+        xy01_s32x4 = vmlal_s16(xy01_s32x4, vget_low_s16 (x_s16x8), vget_low_s16 (y01_s16x8));
+        xy02_s32x4 = vmlal_s16(xy02_s32x4, vget_low_s16 (x_s16x8), vget_low_s16 (y02_s16x8));
+        xy01_s32x4 = vmlal_s16(xy01_s32x4, vget_high_s16(x_s16x8), vget_high_s16(y01_s16x8));
+        xy02_s32x4 = vmlal_s16(xy02_s32x4, vget_high_s16(x_s16x8), vget_high_s16(y02_s16x8));
+    }
+
+    if (N - i >= 4) {
+        const int16x4_t x_s16x4   = vld1_s16(&x[i]);
+        const int16x4_t y01_s16x4 = vld1_s16(&y01[i]);
+        const int16x4_t y02_s16x4 = vld1_s16(&y02[i]);
+        xy01_s32x4 = vmlal_s16(xy01_s32x4, x_s16x4, y01_s16x4);
+        xy02_s32x4 = vmlal_s16(xy02_s32x4, x_s16x4, y02_s16x4);
+        i += 4;
+    }
+
+    xy01_s64x2 = vpaddlq_s32(xy01_s32x4);
+    xy02_s64x2 = vpaddlq_s32(xy02_s32x4);
+    xy01_s64x1 = vadd_s64(vget_low_s64(xy01_s64x2), vget_high_s64(xy01_s64x2));
+    xy02_s64x1 = vadd_s64(vget_low_s64(xy02_s64x2), vget_high_s64(xy02_s64x2));
+    xy01       = vget_lane_s32(vreinterpret_s32_s64(xy01_s64x1), 0);
+    xy02       = vget_lane_s32(vreinterpret_s32_s64(xy02_s64x1), 0);
+
+    for (; i < N; i++) {
+        xy01 = MAC16_16(xy01, x[i], y01[i]);
+        xy02 = MAC16_16(xy02, x[i], y02[i]);
+    }
+    *xy1 = xy01;
+    *xy2 = xy02;
+
+#ifdef OPUS_CHECK_ASM
+    {
+        opus_val32 xy1_c, xy2_c;
+        dual_inner_prod_c(x, y01, y02, N, &xy1_c, &xy2_c);
+        celt_assert(xy1_c == *xy1);
+        celt_assert(xy2_c == *xy2);
+    }
+#endif
+}
+
+#endif
index cabc48f..b6ed111 100644 (file)
@@ -37,7 +37,8 @@ CELT_AM_SOURCES_ARM_ASM = \
 celt/arm/armopts.s.in
 
 CELT_SOURCES_ARM_NEON_INTR = \
-celt/arm/celt_neon_intr.c
+celt/arm/celt_neon_intr.c \
+celt/arm/pitch_neon_intr.c
 
 CELT_SOURCES_ARM_NE10= \
 celt/arm/celt_ne10_fft.c \