From b7c7653941f7add1dbf58360640b0da7ee6e6cf4 Mon Sep 17 00:00:00 2001
From: Jean-Marc Valin <jmvalin@jmvalin.ca>
Date: Mon, 15 Aug 2016 17:30:00 -0400
Subject: Speeding up PVQ search by allocating even more pulses in the
 projection.

---
 celt/vq.c          | 5 +++--
 celt/x86/vq_sse2.c | 5 +++--
 2 files changed, 6 insertions(+), 4 deletions(-)

diff --git a/celt/vq.c b/celt/vq.c
index 5a4a8afd..2f864c2a 100644
--- a/celt/vq.c
+++ b/celt/vq.c
@@ -210,7 +210,8 @@ opus_val16 op_pvq_search_c(celt_norm *X, int *iy, int K, int N, int arch)
          while (++j<N);
          sum = QCONST16(1.f,14);
       }
-      rcp = EXTRACT16(MULT16_32_Q16(K-1, celt_rcp(sum)));
+      /* Using K+e with e < 1 guarantees we cannot get more than K pulses. */
+      rcp = EXTRACT16(MULT16_32_Q16(K+0.8, celt_rcp(sum)));
       j=0; do {
 #ifdef FIXED_POINT
          /* It's really important to round *towards zero* here */
@@ -225,7 +226,7 @@ opus_val16 op_pvq_search_c(celt_norm *X, int *iy, int K, int N, int arch)
          pulsesLeft -= iy[j];
       }  while (++j<N);
    }
-   celt_assert2(pulsesLeft>=1, "Allocated too many pulses in the quick pass");
+   celt_assert2(pulsesLeft>=0, "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. */
diff --git a/celt/x86/vq_sse2.c b/celt/x86/vq_sse2.c
index 0891a5ba..c82d7e24 100644
--- a/celt/x86/vq_sse2.c
+++ b/celt/x86/vq_sse2.c
@@ -104,7 +104,8 @@ opus_val16 op_pvq_search_sse2(celt_norm *_X, int *iy, int K, int N, int arch)
          while (++j<N);
          sums = _mm_set_ps1(1.f);
       }
-      rcp4 = _mm_mul_ps(_mm_set_ps1((float)(K-1)), _mm_rcp_ps(sums));
+      /* Using K+e with e < 1 guarantees we cannot get more than K pulses. */
+      rcp4 = _mm_mul_ps(_mm_set_ps1((float)(K+.8)), _mm_rcp_ps(sums));
       xy4 = yy4 = _mm_setzero_ps();
       pulses_sum = _mm_setzero_si128();
       for (j=0;j<N;j+=4)
@@ -134,7 +135,7 @@ opus_val16 op_pvq_search_sse2(celt_norm *_X, int *iy, int K, int N, int arch)
    }
    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");
+   celt_assert2(pulsesLeft>=0, "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. */
-- 
cgit v1.2.1