Revise celt_fir_c() to not pass in argument "mem"

The "mem" in celt_fir_c() either is contained in the head of input "x"
in reverse order already, or can be easily attached to the head of "x"
before calling the function. Removing argument "mem" can eliminate the
redundant buffer copies inside.
Update celt_fir_sse4_1() accordingly.

7 files changed, 36 insertions, 83 deletions

diff --git a/celt/celt_decoder.c b/celt/celt_decoder.c
index 58e7b75a..567d7456 100644
--- a/celt/celt_decoder.c
+++ b/celt/celt_decoder.c
@@ -556,10 +556,11 @@ static void celt_decode_lost(CELTDecoder * OPUS_RESTRICT st, int N, int LM)
    } else {
       /* Pitch-based PLC */
       const opus_val16 *window;
+      opus_val16 *exc;
       opus_val16 fade = Q15ONE;
       int pitch_index;
       VARDECL(opus_val32, etmp);
-      VARDECL(opus_val16, exc);
+      VARDECL(opus_val16, _exc);
 
       if (loss_count == 0)
       {
@@ -570,7 +571,8 @@ static void celt_decode_lost(CELTDecoder * OPUS_RESTRICT st, int N, int LM)
       }
 
       ALLOC(etmp, overlap, opus_val32);
-      ALLOC(exc, MAX_PERIOD, opus_val16);
+      ALLOC(_exc, MAX_PERIOD+LPC_ORDER, opus_val16);
+      exc = _exc+LPC_ORDER;
       window = mode->window;
       c=0; do {
          opus_val16 decay;
@@ -635,15 +637,14 @@ static void celt_decode_lost(CELTDecoder * OPUS_RESTRICT st, int N, int LM)
          /* Initialize the LPC history with the samples just before the start
             of the region for which we're computing the excitation. */
          {
-            opus_val16 lpc_mem[LPC_ORDER];
             for (i=0;i<LPC_ORDER;i++)
             {
-               lpc_mem[i] =
-                     ROUND16(buf[DECODE_BUFFER_SIZE-exc_length-1-i], SIG_SHIFT);
+               exc[MAX_PERIOD-exc_length-LPC_ORDER+i] =
+                     ROUND16(buf[DECODE_BUFFER_SIZE-exc_length-LPC_ORDER+i], SIG_SHIFT);
             }
             /* Compute the excitation for exc_length samples before the loss. */
             celt_fir(exc+MAX_PERIOD-exc_length, lpc+c*LPC_ORDER,
-                  exc+MAX_PERIOD-exc_length, exc_length, LPC_ORDER, lpc_mem, st->arch);
+                  exc+MAX_PERIOD-exc_length, exc_length, LPC_ORDER, st->arch);
          }
 
          /* Check if the waveform is decaying, and if so how fast.
diff --git a/celt/celt_lpc.c b/celt/celt_lpc.c
index bc9eb2c8..0aabd192 100644
--- a/celt/celt_lpc.c
+++ b/celt/celt_lpc.c
@@ -89,56 +89,47 @@ int          p
 
 
 void celt_fir_c(
-         const opus_val16 *_x,
+         const opus_val16 *x,
          const opus_val16 *num,
-         opus_val16 *_y,
+         opus_val16 *y,
          int N,
          int ord,
-         opus_val16 *mem,
          int arch)
 {
    int i,j;
    VARDECL(opus_val16, rnum);
-   VARDECL(opus_val16, x);
    SAVE_STACK;
 
    ALLOC(rnum, ord, opus_val16);
-   ALLOC(x, N+ord, opus_val16);
    for(i=0;i<ord;i++)
       rnum[i] = num[ord-i-1];
-   for(i=0;i<ord;i++)
-      x[i] = mem[ord-i-1];
-   for (i=0;i<N;i++)
-      x[i+ord]=_x[i];
-   for(i=0;i<ord;i++)
-      mem[i] = _x[N-i-1];
 #ifdef SMALL_FOOTPRINT
    (void)arch;
    for (i=0;i<N;i++)
    {
-      opus_val32 sum = SHL32(EXTEND32(_x[i]), SIG_SHIFT);
+      opus_val32 sum = SHL32(EXTEND32(x[i]), SIG_SHIFT);
       for (j=0;j<ord;j++)
       {
-         sum = MAC16_16(sum,rnum[j],x[i+j]);
+         sum = MAC16_16(sum,rnum[j],x[i+j-ord]);
       }
-      _y[i] = SATURATE16(PSHR32(sum, SIG_SHIFT));
+      y[i] = SATURATE16(PSHR32(sum, SIG_SHIFT));
    }
 #else
    for (i=0;i<N-3;i+=4)
    {
       opus_val32 sum[4]={0,0,0,0};
-      xcorr_kernel(rnum, x+i, sum, ord, arch);
-      _y[i  ] = SATURATE16(ADD32(EXTEND32(_x[i  ]), PSHR32(sum[0], SIG_SHIFT)));
-      _y[i+1] = SATURATE16(ADD32(EXTEND32(_x[i+1]), PSHR32(sum[1], SIG_SHIFT)));
-      _y[i+2] = SATURATE16(ADD32(EXTEND32(_x[i+2]), PSHR32(sum[2], SIG_SHIFT)));
-      _y[i+3] = SATURATE16(ADD32(EXTEND32(_x[i+3]), PSHR32(sum[3], SIG_SHIFT)));
+      xcorr_kernel(rnum, x+i-ord, sum, ord, arch);
+      y[i  ] = SATURATE16(ADD32(EXTEND32(x[i  ]), PSHR32(sum[0], SIG_SHIFT)));
+      y[i+1] = SATURATE16(ADD32(EXTEND32(x[i+1]), PSHR32(sum[1], SIG_SHIFT)));
+      y[i+2] = SATURATE16(ADD32(EXTEND32(x[i+2]), PSHR32(sum[2], SIG_SHIFT)));
+      y[i+3] = SATURATE16(ADD32(EXTEND32(x[i+3]), PSHR32(sum[3], SIG_SHIFT)));
    }
    for (;i<N;i++)
    {
       opus_val32 sum = 0;
       for (j=0;j<ord;j++)
-         sum = MAC16_16(sum,rnum[j],x[i+j]);
-      _y[i] = SATURATE16(ADD32(EXTEND32(_x[i]), PSHR32(sum, SIG_SHIFT)));
+         sum = MAC16_16(sum,rnum[j],x[i+j-ord]);
+      y[i] = SATURATE16(ADD32(EXTEND32(x[i]), PSHR32(sum, SIG_SHIFT)));
    }
 #endif
    RESTORE_STACK;
diff --git a/celt/celt_lpc.h b/celt/celt_lpc.h
index 323459eb..a4c5fd6e 100644
--- a/celt/celt_lpc.h
+++ b/celt/celt_lpc.h
@@ -45,12 +45,11 @@ void celt_fir_c(
          opus_val16 *y,
          int N,
          int ord,
-         opus_val16 *mem,
          int arch);
 
 #if !defined(OVERRIDE_CELT_FIR)
-#define celt_fir(x, num, y, N, ord, mem, arch) \
-    (celt_fir_c(x, num, y, N, ord, mem, arch))
+#define celt_fir(x, num, y, N, ord, arch) \
+    (celt_fir_c(x, num, y, N, ord, arch))
 #endif
 
 void celt_iir(const opus_val32 *x,
diff --git a/celt/x86/celt_lpc_sse.c b/celt/x86/celt_lpc_sse.c
index 67e5592a..12a9b0e5 100644
--- a/celt/x86/celt_lpc_sse.c
+++ b/celt/x86/celt_lpc_sse.c
@@ -40,63 +40,32 @@
 
 #if defined(FIXED_POINT)
 
-void celt_fir_sse4_1(const opus_val16 *_x,
+void celt_fir_sse4_1(const opus_val16 *x,
          const opus_val16 *num,
-         opus_val16 *_y,
+         opus_val16 *y,
          int N,
          int ord,
-         opus_val16 *mem,
          int arch)
 {
     int i,j;
     VARDECL(opus_val16, rnum);
-    VARDECL(opus_val16, x);
 
     __m128i vecNoA;
     opus_int32 noA ;
     SAVE_STACK;
 
    ALLOC(rnum, ord, opus_val16);
-   ALLOC(x, N+ord, opus_val16);
    for(i=0;i<ord;i++)
       rnum[i] = num[ord-i-1];
-   for(i=0;i<ord;i++)
-      x[i] = mem[ord-i-1];
-
-   for (i=0;i<N-7;i+=8)
-   {
-       x[i+ord  ]=_x[i  ];
-       x[i+ord+1]=_x[i+1];
-       x[i+ord+2]=_x[i+2];
-       x[i+ord+3]=_x[i+3];
-       x[i+ord+4]=_x[i+4];
-       x[i+ord+5]=_x[i+5];
-       x[i+ord+6]=_x[i+6];
-       x[i+ord+7]=_x[i+7];
-   }
-
-   for (;i<N-3;i+=4)
-   {
-       x[i+ord  ]=_x[i  ];
-       x[i+ord+1]=_x[i+1];
-       x[i+ord+2]=_x[i+2];
-       x[i+ord+3]=_x[i+3];
-   }
-
-   for (;i<N;i++)
-         x[i+ord]=_x[i];
-
-   for(i=0;i<ord;i++)
-      mem[i] = _x[N-i-1];
 #ifdef SMALL_FOOTPRINT
    for (i=0;i<N;i++)
    {
-      opus_val32 sum = SHL32(EXTEND32(_x[i]), SIG_SHIFT);
+      opus_val32 sum = SHL32(EXTEND32(x[i]), SIG_SHIFT);
       for (j=0;j<ord;j++)
       {
-         sum = MAC16_16(sum,rnum[j],x[i+j]);
+         sum = MAC16_16(sum,rnum[j],x[i+j-ord]);
       }
-      _y[i] = SATURATE16(PSHR32(sum, SIG_SHIFT));
+      y[i] = SATURATE16(PSHR32(sum, SIG_SHIFT));
    }
 #else
    noA = EXTEND32(1) << SIG_SHIFT >> 1;
@@ -107,22 +76,22 @@ void celt_fir_sse4_1(const opus_val16 *_x,
       opus_val32 sums[4] = {0};
       __m128i vecSum, vecX;
 
-      xcorr_kernel(rnum, x+i, sums, ord, arch);
+      xcorr_kernel(rnum, x+i-ord, sums, ord, arch);
 
       vecSum = _mm_loadu_si128((__m128i *)sums);
       vecSum = _mm_add_epi32(vecSum, vecNoA);
       vecSum = _mm_srai_epi32(vecSum, SIG_SHIFT);
-      vecX = OP_CVTEPI16_EPI32_M64(_x + i);
+      vecX = OP_CVTEPI16_EPI32_M64(x + i);
       vecSum = _mm_add_epi32(vecSum, vecX);
       vecSum = _mm_packs_epi32(vecSum, vecSum);
-      _mm_storel_epi64((__m128i *)(_y + i), vecSum);
+      _mm_storel_epi64((__m128i *)(y + i), vecSum);
    }
    for (;i<N;i++)
    {
       opus_val32 sum = 0;
       for (j=0;j<ord;j++)
-         sum = MAC16_16(sum, rnum[j], x[i + j]);
-      _y[i] = SATURATE16(ADD32(EXTEND32(_x[i]), PSHR32(sum, SIG_SHIFT)));
+         sum = MAC16_16(sum, rnum[j], x[i+j-ord]);
+      y[i] = SATURATE16(ADD32(EXTEND32(x[i]), PSHR32(sum, SIG_SHIFT)));
    }
 
 #endif
diff --git a/celt/x86/celt_lpc_sse.h b/celt/x86/celt_lpc_sse.h
index c5ec796e..7d1ecf75 100644
--- a/celt/x86/celt_lpc_sse.h
+++ b/celt/x86/celt_lpc_sse.h
@@ -41,12 +41,11 @@ void celt_fir_sse4_1(
          opus_val16 *y,
          int N,
          int ord,
-         opus_val16 *mem,
          int arch);
 
 #if defined(OPUS_X86_PRESUME_SSE4_1)
-#define celt_fir(x, num, y, N, ord, mem, arch) \
-    ((void)arch, celt_fir_sse4_1(x, num, y, N, ord, mem, arch))
+#define celt_fir(x, num, y, N, ord, arch) \
+    ((void)arch, celt_fir_sse4_1(x, num, y, N, ord, arch))
 
 #else
 
@@ -56,11 +55,10 @@ extern void (*const CELT_FIR_IMPL[OPUS_ARCHMASK + 1])(
          opus_val16 *y,
          int N,
          int ord,
-         opus_val16 *mem,
          int arch);
 
-#  define celt_fir(x, num, y, N, ord, mem, arch) \
-    ((*CELT_FIR_IMPL[(arch) & OPUS_ARCHMASK])(x, num, y, N, ord, mem, arch))
+#  define celt_fir(x, num, y, N, ord, arch) \
+    ((*CELT_FIR_IMPL[(arch) & OPUS_ARCHMASK])(x, num, y, N, ord, arch))
 
 #endif
 #endif
diff --git a/celt/x86/x86_celt_map.c b/celt/x86/x86_celt_map.c
index 80559a28..d39d88ed 100644
--- a/celt/x86/x86_celt_map.c
+++ b/celt/x86/x86_celt_map.c
@@ -47,7 +47,6 @@ void (*const CELT_FIR_IMPL[OPUS_ARCHMASK + 1])(
          opus_val16       *y,
          int              N,
          int              ord,
-         opus_val16       *mem,
          int              arch
 ) = {
   celt_fir_c,                /* non-sse */
diff --git a/silk/LPC_analysis_filter.c b/silk/LPC_analysis_filter.c
index 3c616d7b..20330d54 100644
--- a/silk/LPC_analysis_filter.c
+++ b/silk/LPC_analysis_filter.c
@@ -57,7 +57,6 @@ void silk_LPC_analysis_filter(
 {
     opus_int   j;
 #if USE_CELT_FIR
-    opus_int16 mem[SILK_MAX_ORDER_LPC];
     opus_int16 num[SILK_MAX_ORDER_LPC];
 #else
     int ix;
@@ -74,10 +73,7 @@ void silk_LPC_analysis_filter(
     for ( j = 0; j < d; j++ ) {
         num[ j ] = -B[ j ];
     }
-    for (j=0;j<d;j++) {
-        mem[ j ] = in[ d - j - 1 ];
-    }
-    celt_fir( in + d, num, out + d, len - d, d, mem, arch );
+    celt_fir( in + d, num, out + d, len - d, d, arch );
     for ( j = 0; j < d; j++ ) {
         out[ j ] = 0;
     }