armv7(float): Optimize decode usecase using NE10 library

Optimize opus decode (float only) use case using ARM NE10.
Mainly effects opus_ifft and ctl_mdct_backward and related
functions.

Work based on previous Encode optimization using ARM NE10
library. See previous commit for details on how to enable
this.

Signed-off-by: Timothy B. Terriberry <tterribe@xiph.org>

5 files changed, 165 insertions, 0 deletions

diff --git a/celt/arm/arm_celt_map.c b/celt/arm/arm_celt_map.c
index 87ba3b3e..4c2d28ce 100644
--- a/celt/arm/arm_celt_map.c
+++ b/celt/arm/arm_celt_map.c
@@ -79,6 +79,15 @@ void (*const OPUS_FFT[OPUS_ARCHMASK+1])(const kiss_fft_state *cfg,
    opus_fft_float_neon           /* Neon with NE10 */
 };
 
+void (*const OPUS_IFFT[OPUS_ARCHMASK+1])(const kiss_fft_state *cfg,
+                                         const kiss_fft_cpx *fin,
+                                         kiss_fft_cpx *fout) = {
+   opus_ifft_c,                   /* ARMv4 */
+   opus_ifft_c,                   /* EDSP */
+   opus_ifft_c,                   /* Media */
+   opus_ifft_float_neon           /* Neon with NE10 */
+};
+
 void (*const CLT_MDCT_FORWARD_IMPL[OPUS_ARCHMASK+1])(const mdct_lookup *l,
                                                      kiss_fft_scalar *in,
                                                      kiss_fft_scalar * OPUS_RESTRICT out,
@@ -90,6 +99,19 @@ void (*const CLT_MDCT_FORWARD_IMPL[OPUS_ARCHMASK+1])(const mdct_lookup *l,
    clt_mdct_forward_c,           /* Media */
    clt_mdct_forward_float_neon   /* Neon with NE10 */
 };
+
+void (*const CLT_MDCT_BACKWARD_IMPL[OPUS_ARCHMASK+1])(const mdct_lookup *l,
+                                                      kiss_fft_scalar *in,
+                                                      kiss_fft_scalar * OPUS_RESTRICT out,
+                                                      const opus_val16 *window,
+                                                      int overlap, int shift,
+                                                      int stride, int arch) = {
+   clt_mdct_backward_c,           /* ARMv4 */
+   clt_mdct_backward_c,           /* EDSP */
+   clt_mdct_backward_c,           /* Media */
+   clt_mdct_backward_float_neon   /* Neon with NE10 */
+};
+
 #   endif /* HAVE_ARM_NE10 */
 #  endif /* OPUS_ARM_MAY_HAVE_NEON_INTR */
 # endif /* FIXED_POINT */
diff --git a/celt/arm/celt_ne10_fft.c b/celt/arm/celt_ne10_fft.c
index fc4b0da0..2ba8c559 100644
--- a/celt/arm/celt_ne10_fft.c
+++ b/celt/arm/celt_ne10_fft.c
@@ -118,4 +118,30 @@ void opus_fft_float_neon(const kiss_fft_state *st,
    }
    RESTORE_STACK;
 }
+
+void opus_ifft_float_neon(const kiss_fft_state *st,
+                          const kiss_fft_cpx *fin,
+                          kiss_fft_cpx *fout)
+{
+   ne10_fft_state_float32_t state;
+   ne10_fft_cfg_float32_t cfg = &state;
+   VARDECL(ne10_fft_cpx_float32_t, buffer);
+   SAVE_STACK;
+   ALLOC(buffer, st->nfft, ne10_fft_cpx_float32_t);
+
+   if (!st->arch_fft->is_supported) {
+      /* This nfft length (scaled fft) not supported in NE10 */
+      opus_ifft_c(st, fin, fout);
+   }
+   else {
+      memcpy((void *)cfg, st->arch_fft->priv, sizeof(ne10_fft_state_float32_t));
+      state.buffer = (ne10_fft_cpx_float32_t *)&buffer[0];
+      state.is_backward_scaled = 0;
+
+      ne10_fft_c2c_1d_float32_neon((ne10_fft_cpx_float32_t *)fout,
+                                   (ne10_fft_cpx_float32_t *)fin,
+                                   cfg, 1);
+   }
+   RESTORE_STACK;
+}
 #endif /* !defined(FIXED_POINT) */
diff --git a/celt/arm/celt_ne10_mdct.c b/celt/arm/celt_ne10_mdct.c
index 1c6e9158..a68264e2 100644
--- a/celt/arm/celt_ne10_mdct.c
+++ b/celt/arm/celt_ne10_mdct.c
@@ -156,4 +156,106 @@ void clt_mdct_forward_float_neon(const mdct_lookup *l,
    }
    RESTORE_STACK;
 }
+
+void clt_mdct_backward_float_neon(const mdct_lookup *l,
+                                  kiss_fft_scalar *in,
+                                  kiss_fft_scalar * OPUS_RESTRICT out,
+                                  const opus_val16 * OPUS_RESTRICT window,
+                                  int overlap, int shift, int stride, int arch)
+{
+   int i;
+   int N, N2, N4;
+   VARDECL(kiss_fft_scalar, f);
+   const kiss_twiddle_scalar *trig;
+   const kiss_fft_state *st = l->kfft[shift];
+
+   N = l->n;
+   trig = l->trig;
+   for (i=0;i<shift;i++)
+   {
+      N >>= 1;
+      trig += N;
+   }
+   N2 = N>>1;
+   N4 = N>>2;
+
+   ALLOC(f, N2, kiss_fft_scalar);
+
+   /* Pre-rotate */
+   {
+      /* Temp pointers to make it really clear to the compiler what we're doing */
+      const kiss_fft_scalar * OPUS_RESTRICT xp1 = in;
+      const kiss_fft_scalar * OPUS_RESTRICT xp2 = in+stride*(N2-1);
+      kiss_fft_scalar * OPUS_RESTRICT yp = f;
+      const kiss_twiddle_scalar * OPUS_RESTRICT t = &trig[0];
+      for(i=0;i<N4;i++)
+      {
+         kiss_fft_scalar yr, yi;
+         yr = S_MUL(*xp2, t[i]) + S_MUL(*xp1, t[N4+i]);
+         yi = S_MUL(*xp1, t[i]) - S_MUL(*xp2, t[N4+i]);
+         yp[2*i] = yr;
+         yp[2*i+1] = yi;
+         xp1+=2*stride;
+         xp2-=2*stride;
+      }
+   }
+
+   opus_ifft(st, (kiss_fft_cpx *)f, (kiss_fft_cpx*)(out+(overlap>>1)), arch);
+
+   /* Post-rotate and de-shuffle from both ends of the buffer at once to make
+      it in-place. */
+   {
+      kiss_fft_scalar * yp0 = out+(overlap>>1);
+      kiss_fft_scalar * yp1 = out+(overlap>>1)+N2-2;
+      const kiss_twiddle_scalar *t = &trig[0];
+      /* Loop to (N4+1)>>1 to handle odd N4. When N4 is odd, the
+         middle pair will be computed twice. */
+      for(i=0;i<(N4+1)>>1;i++)
+      {
+         kiss_fft_scalar re, im, yr, yi;
+         kiss_twiddle_scalar t0, t1;
+         re = yp0[0];
+         im = yp0[1];
+         t0 = t[i];
+         t1 = t[N4+i];
+         /* We'd scale up by 2 here, but instead it's done when mixing the windows */
+         yr = S_MUL(re,t0) + S_MUL(im,t1);
+         yi = S_MUL(re,t1) - S_MUL(im,t0);
+         re = yp1[0];
+         im = yp1[1];
+         yp0[0] = yr;
+         yp1[1] = yi;
+
+         t0 = t[(N4-i-1)];
+         t1 = t[(N2-i-1)];
+         /* We'd scale up by 2 here, but instead it's done when mixing the windows */
+         yr = S_MUL(re,t0) + S_MUL(im,t1);
+         yi = S_MUL(re,t1) - S_MUL(im,t0);
+         yp1[0] = yr;
+         yp0[1] = yi;
+         yp0 += 2;
+         yp1 -= 2;
+      }
+   }
+
+   /* Mirror on both sides for TDAC */
+   {
+      kiss_fft_scalar * OPUS_RESTRICT xp1 = out+overlap-1;
+      kiss_fft_scalar * OPUS_RESTRICT yp1 = out;
+      const opus_val16 * OPUS_RESTRICT wp1 = window;
+      const opus_val16 * OPUS_RESTRICT wp2 = window+overlap-1;
+
+      for(i = 0; i < overlap/2; i++)
+      {
+         kiss_fft_scalar x1, x2;
+         x1 = *xp1;
+         x2 = *yp1;
+         *yp1++ = MULT16_32_Q15(*wp2, x2) - MULT16_32_Q15(*wp1, x1);
+         *xp1-- = MULT16_32_Q15(*wp1, x2) + MULT16_32_Q15(*wp2, x1);
+         wp1++;
+         wp2--;
+      }
+   }
+   RESTORE_STACK;
+}
 #endif /* !defined(FIXED_POINT) */
diff --git a/celt/arm/fft_arm.h b/celt/arm/fft_arm.h
index e7a30d69..e57b0aa6 100644
--- a/celt/arm/fft_arm.h
+++ b/celt/arm/fft_arm.h
@@ -46,6 +46,11 @@ void opus_fft_free_arm_float_neon(kiss_fft_state *st);
 void opus_fft_float_neon(const kiss_fft_state *st,
                          const kiss_fft_cpx *fin,
                          kiss_fft_cpx *fout);
+
+void opus_ifft_float_neon(const kiss_fft_state *st,
+                         const kiss_fft_cpx *fin,
+                         kiss_fft_cpx *fout);
+
 #if !defined(OPUS_HAVE_RTCD)
 #define OVERRIDE_OPUS_FFT (1)
 
@@ -58,6 +63,9 @@ void opus_fft_float_neon(const kiss_fft_state *st,
 #define opus_fft(_st, _fin, _fout, arch) \
    ((void)(arch), opus_fft_float_neon(_st, _fin, _fout))
 
+#define opus_ifft(_st, _fin, _fout, arch) \
+   ((void)(arch), opus_ifft_float_neon(_st, _fin, _fout))
+
 #endif /* OPUS_HAVE_RTCD */
 
 #endif /* HAVE_ARM_NE10 */
diff --git a/celt/arm/mdct_arm.h b/celt/arm/mdct_arm.h
index 7d60fedc..db32efe5 100644
--- a/celt/arm/mdct_arm.h
+++ b/celt/arm/mdct_arm.h
@@ -43,10 +43,17 @@ void clt_mdct_forward_float_neon(const mdct_lookup *l, kiss_fft_scalar *in,
                                  const opus_val16 *window, int overlap,
                                  int shift, int stride, int arch);
 
+void clt_mdct_backward_float_neon(const mdct_lookup *l, kiss_fft_scalar *in,
+                                  kiss_fft_scalar * OPUS_RESTRICT out,
+                                  const opus_val16 *window, int overlap,
+                                  int shift, int stride, int arch);
+
 #if !defined(OPUS_HAVE_RTCD)
 #define OVERRIDE_OPUS_MDCT (1)
 #define clt_mdct_forward(_l, _in, _out, _window, _int, _shift, _stride, _arch) \
       clt_mdct_forward_float_neon(_l, _in, _out, _window, _int, _shift, _stride, _arch)
+#define clt_mdct_backward(_l, _in, _out, _window, _int, _shift, _stride, _arch) \
+      clt_mdct_backward_float_neon(_l, _in, _out, _window, _int, _shift, _stride, _arch)
 #endif /* OPUS_HAVE_RTCD */
 #endif /* !defined(FIXED_POINT) && defined(HAVE_ARM_NE10) */