SSE2 implementation of the PVQ search

We used the SSE reciprocal square root instruction to vectorize the serch rather
than compare one at a time with multiplies. Speeds up the entire encoder by 8-10%.

celt/bands.c

@@ -1036,7 +1036,7 @@ static unsigned quant_partition(struct band_ctx *ctx, celt_norm *X,
          /* Finally do the actual quantization */
          if (encode)
          {
-            cm = alg_quant(X, N, K, spread, B, ec, gain, ctx->resynth);
+            cm = alg_quant(X, N, K, spread, B, ec, gain, ctx->resynth, ctx->arch);
          } else {
             cm = alg_unquant(X, N, K, spread, B, ec, gain);
          }

celt/tests/test_unit_mathops.c

@@ -57,6 +57,7 @@
 # endif
 # if defined(OPUS_X86_MAY_HAVE_SSE2)
 #  include "x86/pitch_sse2.c"
+#  include "x86/vq_sse2.c"
 # endif
 # if defined(OPUS_X86_MAY_HAVE_SSE4_1)
 #  include "x86/pitch_sse4_1.c"

celt/tests/test_unit_rotation.c

@@ -55,6 +55,7 @@
 # endif
 # if defined(OPUS_X86_MAY_HAVE_SSE2)
 #  include "x86/pitch_sse2.c"
+#  include "x86/vq_sse2.c"
 # endif
 # if defined(OPUS_X86_MAY_HAVE_SSE4_1)
 #  include "x86/pitch_sse4_1.c"

celt/vq.c

@@ -158,29 +158,21 @@ static unsigned extract_collapse_mask(int *iy, int N, int B)
    return collapse_mask;
 }
 
-unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc,
-      opus_val16 gain, int resynth)
+opus_val16 op_pvq_search_c(celt_norm *X, int *iy, int K, int N, int arch)
 {
    VARDECL(celt_norm, y);
-   VARDECL(int, iy);
    VARDECL(int, signx);
    int i, j;
    int pulsesLeft;
    opus_val32 sum;
    opus_val32 xy;
    opus_val16 yy;
-   unsigned collapse_mask;
    SAVE_STACK;
 
-   celt_assert2(K>0, "alg_quant() needs at least one pulse");
-   celt_assert2(N>1, "alg_quant() needs at least two dimensions");
-
+   (void)arch;
    ALLOC(y, N, celt_norm);
-   ALLOC(iy, N, int);
    ALLOC(signx, N, int);
 
-   exp_rotation(X, N, 1, B, K, spread);
-
    /* Get rid of the sign */
    sum = 0;
    j=0; do {
@@ -322,6 +314,28 @@ unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc,
          but has the same performance otherwise. */
       iy[j] = (iy[j]^-signx[j]) + signx[j];
    } while (++j<N);
+   RESTORE_STACK;
+   return yy;
+}
+
+unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc,
+      opus_val16 gain, int resynth, int arch)
+{
+   VARDECL(int, iy);
+   opus_val16 yy;
+   unsigned collapse_mask;
+   SAVE_STACK;
+
+   celt_assert2(K>0, "alg_quant() needs at least one pulse");
+   celt_assert2(N>1, "alg_quant() needs at least two dimensions");
+
+   /* Covers vectorization by up to 4. */
+   ALLOC(iy, N+3, int);
+
+   exp_rotation(X, N, 1, B, K, spread);
+
+   yy = op_pvq_search(X, iy, K, N, arch);
+
    encode_pulses(iy, N, K, enc);
 
    if (resynth)

celt/vq.h

@@ -37,10 +37,20 @@
 #include "entdec.h"
 #include "modes.h"
 
+#if (defined(OPUS_X86_MAY_HAVE_SSE2) && !defined(FIXED_POINT))
+#include "x86/vq_sse.h"
+#endif
+
 #if defined(MIPSr1_ASM)
 #include "mips/vq_mipsr1.h"
 #endif
 
+opus_val16 op_pvq_search_c(celt_norm *X, int *iy, int K, int N, int arch);
+
+#if !defined(OVERRIDE_OP_PVQ_SEARCH)
+#define op_pvq_search(x, iy, K, N, arch) \
+    (op_pvq_search_c(x, iy, K, N, arch))
+#endif
 
 /** Algebraic pulse-vector quantiser. The signal x is replaced by the sum of
   * the pitch and a combination of pulses such that its norm is still equal
@@ -52,7 +62,7 @@
  * @ret A mask indicating which blocks in the band received pulses
 */
 unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc,
-      opus_val16 gain, int resynth);
+      opus_val16 gain, int resynth, int arch);
 
 /** Algebraic pulse decoder
  * @param X Decoded normalised spectrum (returned)

celt/x86/vq_sse.h

+/* Copyright (c) 2016  Jean-Marc Valin */
+/*
+   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.
+
+   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.
+*/
+
+#ifndef VQ_SSE_H
+#define VQ_SSE_H
+
+#if defined(OPUS_X86_MAY_HAVE_SSE2) && !defined(FIXED_POINT)
+#define OVERRIDE_OP_PVQ_SEARCH
+
+opus_val16 op_pvq_search_sse2(celt_norm *_X, int *iy, int K, int N, int arch);
+
+#if defined(OPUS_X86_PRESUME_SSE2)
+#define op_pvq_search(x, iy, K, N, arch) \
+    (op_pvq_search_sse2(x, iy, K, N, arch))
+
+#else
+
+extern opus_val16 (*const OP_PVQ_SEARCH_IMPL[OPUS_ARCHMASK + 1])(
+      celt_norm *_X, int *iy, int K, int N, int arch);
+
+#  define op_pvq_search(X, iy, K, N, arch) \
+    ((*OP_PVQ_SEARCH_IMPL[(arch) & OPUS_ARCHMASK])(X, iy, K, N, arch))
+
+#endif
+#endif
+
+#endif

celt/x86/vq_sse2.c

+/* Copyright (c) 2007-2008 CSIRO
+   Copyright (c) 2007-2009 Xiph.Org Foundation
+   Copyright (c) 2007-2016 Jean-Marc Valin */
+/*
+   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.
+
+   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 <xmmintrin.h>
+#include <emmintrin.h>
+#include "celt_lpc.h"
+#include "stack_alloc.h"
+#include "mathops.h"
+#include "vq.h"
+#include "x86cpu.h"
+
+
+#ifndef FIXED_POINT
+
+opus_val16 op_pvq_search_sse2(celt_norm *_X, int *iy, int K, int N, int arch)
+{
+   int i, j;
+   int pulsesLeft;
+   float xy, yy;
+   VARDECL(celt_norm, y);
+   VARDECL(celt_norm, X);
+   VARDECL(float, signy);
+   __m128 signmask;
+   __m128 sums;
+   __m128i fours;
+   SAVE_STACK;
+
+   (void)arch;
+   /* All bits set to zero, except for the sign bit. */
+   signmask = _mm_set_ps1(-0.f);
+   fours = _mm_set_epi32(4, 4, 4, 4);
+   ALLOC(y, N+3, celt_norm);
+   ALLOC(X, N+3, celt_norm);
+   ALLOC(signy, N+3, float);
+
+   OPUS_COPY(X, _X, N);
+   X[N] = X[N+1] = X[N+2] = 0;
+   sums = _mm_setzero_ps();
+   for (j=0;j<N;j+=4)
+   {
+      __m128 x4, s4;
+      x4 = _mm_loadu_ps(&X[j]);
+      s4 = _mm_cmplt_ps(x4, _mm_setzero_ps());
+      /* Get rid of the sign */
+      x4 = _mm_andnot_ps(signmask, x4);
+      sums = _mm_add_ps(sums, x4);
+      /* Clear y and iy in case we don't do the projection. */
+      _mm_storeu_ps(&y[j], _mm_setzero_ps());
+      _mm_storeu_si128((__m128i*)&iy[j], _mm_setzero_si128());
+      _mm_storeu_ps(&X[j], x4);
+      _mm_storeu_ps(&signy[j], s4);
+   }
+   sums = _mm_add_ps(sums, _mm_shuffle_ps(sums, sums, _MM_SHUFFLE(1, 0, 3, 2)));
+   sums = _mm_add_ps(sums, _mm_shuffle_ps(sums, sums, _MM_SHUFFLE(2, 3, 0, 1)));
+
+   xy = yy = 0;
+
+   pulsesLeft = K;
+
+   /* Do a pre-search by projecting on the pyramid */
+   if (K > (N>>1))
+   {
+      __m128i pulses_sum;
+      __m128 yy4, xy4;
+      __m128 rcp4;
+      opus_val32 sum = _mm_cvtss_f32(sums);
+      /* If X is too small, just replace it with a pulse at 0 */
+      /* Prevents infinities and NaNs from causing too many pulses
+         to be allocated. 64 is an approximation of infinity here. */
+      if (!(sum > EPSILON && sum < 64))
+      {
+         X[0] = QCONST16(1.f,14);
+         j=1; do
+            X[j]=0;
+         while (++j<N);
+         sums = _mm_set_ps1(1.f);
+      }
+      rcp4 = _mm_mul_ps(_mm_set_ps1((float)(K-1)), _mm_rcp_ps(sums));
+      xy4 = yy4 = _mm_setzero_ps();
+      pulses_sum = _mm_setzero_si128();
+      for (j=0;j<N;j+=4)
+      {
+         __m128 rx4, x4, y4;
+         __m128i iy4;
+         x4 = _mm_loadu_ps(&X[j]);
+         rx4 = _mm_mul_ps(x4, rcp4);
+         iy4 = _mm_cvttps_epi32(rx4);
+         pulses_sum = _mm_add_epi32(pulses_sum, iy4);
+         _mm_storeu_si128((__m128i*)&iy[j], iy4);
+         y4 = _mm_cvtepi32_ps(iy4);
+         xy4 = _mm_add_ps(xy4, _mm_mul_ps(x4, y4));
+         yy4 = _mm_add_ps(yy4, _mm_mul_ps(y4, y4));
+         /* double the y[] vector so we don't have to do it in the search loop. */
+         _mm_storeu_ps(&y[j], _mm_add_ps(y4, y4));
+      }
+      pulses_sum = _mm_add_epi32(pulses_sum, _mm_shuffle_epi32(pulses_sum, _MM_SHUFFLE(1, 0, 3, 2)));
+      pulses_sum = _mm_add_epi32(pulses_sum, _mm_shuffle_epi32(pulses_sum, _MM_SHUFFLE(2, 3, 0, 1)));
+      pulsesLeft -= _mm_cvtsi128_si32(pulses_sum);
+      xy4 = _mm_add_ps(xy4, _mm_shuffle_ps(xy4, xy4, _MM_SHUFFLE(1, 0, 3, 2)));
+      xy4 = _mm_add_ps(xy4, _mm_shuffle_ps(xy4, xy4, _MM_SHUFFLE(2, 3, 0, 1)));
+      xy = _mm_cvtss_f32(xy4);
+      yy4 = _mm_add_ps(yy4, _mm_shuffle_ps(yy4, yy4, _MM_SHUFFLE(1, 0, 3, 2)));
+      yy4 = _mm_add_ps(yy4, _mm_shuffle_ps(yy4, yy4, _MM_SHUFFLE(2, 3, 0, 1)));
+      yy = _mm_cvtss_f32(yy4);
+   }
+   X[N] = X[N+1] = X[N+2] = -100;
+   y[N] = y[N+1] = y[N+2] = 100;
+   celt_assert2(pulsesLeft>=1, "Allocated too many pulses in the quick pass");
+
+   /* This should never happen, but just in case it does (e.g. on silence)
+      we fill the first bin with pulses. */
+   if (pulsesLeft > N+3)
+   {
+      opus_val16 tmp = (opus_val16)pulsesLeft;
+      yy = MAC16_16(yy, tmp, tmp);
+      yy = MAC16_16(yy, tmp, y[0]);
+      iy[0] += pulsesLeft;
+      pulsesLeft=0;
+   }
+
+   for (i=0;i<pulsesLeft;i++)
+   {
+      int best_id;
+      __m128 xy4, yy4;
+      __m128 max, max2;
+      __m128i count;
+      __m128i pos;
+      best_id = 0;
+      /* The squared magnitude term gets added anyway, so we might as well
+         add it outside the loop */
+      yy = ADD16(yy, 1);
+      xy4 = _mm_load1_ps(&xy);
+      yy4 = _mm_load1_ps(&yy);
+      max = _mm_setzero_ps();
+      pos = _mm_setzero_si128();
+      count = _mm_set_epi32(3, 2, 1, 0);
+      for (j=0;j<N;j+=4)
+      {
+         __m128 x4, y4, r4;
+         x4 = _mm_loadu_ps(&X[j]);
+         y4 = _mm_loadu_ps(&y[j]);
+         x4 = _mm_add_ps(x4, xy4);
+         y4 = _mm_add_ps(y4, yy4);
+         y4 = _mm_rsqrt_ps(y4);
+         r4 = _mm_mul_ps(x4, y4);
+         /* Update the index of the max. */
+         pos = _mm_max_epi16(pos, _mm_and_si128(count, _mm_castps_si128(_mm_cmpgt_ps(r4, max))));
+         /* Update the max. */
+         max = _mm_max_ps(max, r4);
+         /* Update the indices (+4) */
+         count = _mm_add_epi32(count, fours);
+      }
+      /* Horizontal max */
+      max2 = _mm_max_ps(max, _mm_shuffle_ps(max, max, _MM_SHUFFLE(1, 0, 3, 2)));
+      max2 = _mm_max_ps(max2, _mm_shuffle_ps(max2, max2, _MM_SHUFFLE(2, 3, 0, 1)));
+      /* Now that max2 contains the max at all positions, look at which value(s) of the
+         partial max is equal to the global max. */
+      pos = _mm_and_si128(pos, _mm_castps_si128(_mm_cmpeq_ps(max, max2)));
+      pos = _mm_max_epi16(pos, _mm_unpackhi_epi64(pos, pos));
+      pos = _mm_max_epi16(pos, _mm_shufflelo_epi16(pos, _MM_SHUFFLE(1, 0, 3, 2)));
+      best_id = _mm_cvtsi128_si32(pos);
+
+      /* Updating the sums of the new pulse(s) */
+      xy = ADD32(xy, EXTEND32(X[best_id]));
+      /* We're multiplying y[j] by two so we don't have to do it here */
+      yy = ADD16(yy, y[best_id]);
+
+      /* Only now that we've made the final choice, update y/iy */
+      /* Multiplying y[j] by 2 so we don't have to do it everywhere else */
+      y[best_id] += 2;
+      iy[best_id]++;
+   }
+
+   /* Put the original sign back */
+   for (j=0;j<N;j+=4)
+   {
+      __m128i y4;
+      __m128i s4;
+      y4 = _mm_loadu_si128((__m128i*)&iy[j]);
+      s4 = _mm_castps_si128(_mm_loadu_ps(&signy[j]));
+      y4 = _mm_xor_si128(_mm_add_epi32(y4, s4), s4);
+      _mm_storeu_si128((__m128i*)&iy[j], y4);
+   }
+   RESTORE_STACK;
+   return yy;
+}
+
+#endif

celt/x86/x86_celt_map.c

@@ -33,6 +33,7 @@
 #include "celt_lpc.h"
 #include "pitch.h"
 #include "pitch_sse.h"
+#include "vq.h"
 
 #if defined(OPUS_HAVE_RTCD)
 
@@ -151,5 +152,17 @@ void (*const COMB_FILTER_CONST_IMPL[OPUS_ARCHMASK + 1])(
 
 #endif
 
+#if defined(OPUS_X86_MAY_HAVE_SSE2) && !defined(OPUS_X86_PRESUME_SSE2)
+opus_val16 (*const OP_PVQ_SEARCH_IMPL[OPUS_ARCHMASK + 1])(
+      celt_norm *_X, int *iy, int K, int N, int arch
+) = {
+  op_pvq_search_c,                /* non-sse */
+  op_pvq_search_c,
+  MAY_HAVE_SSE2(op_pvq_search),
+  MAY_HAVE_SSE2(op_pvq_search),
+  MAY_HAVE_SSE2(op_pvq_search)
+};
+#endif
+
 #endif
 #endif

celt_headers.mk

@@ -49,4 +49,5 @@ celt/mips/mdct_mipsr1.h \
 celt/mips/pitch_mipsr1.h \
 celt/mips/vq_mipsr1.h \
 celt/x86/pitch_sse.h \
+celt/x86/vq_sse.h \
 celt/x86/x86cpu.h

celt_sources.mk

@@ -21,7 +21,7 @@ CELT_SOURCES_SSE = celt/x86/x86cpu.c \
 celt/x86/x86_celt_map.c \
 celt/x86/pitch_sse.c
 
-CELT_SOURCES_SSE2 = celt/x86/pitch_sse2.c
+CELT_SOURCES_SSE2 = celt/x86/pitch_sse2.c celt/x86/vq_sse2.c
 
 CELT_SOURCES_SSE4_1 = celt/x86/celt_lpc_sse.c \
 celt/x86/pitch_sse4_1.c