From ef0eac497f762e1dce7f338a573c27aa6a94dea7 Mon Sep 17 00:00:00 2001
From: Jean-Marc Valin <jmvalin@jmvalin.ca>
Date: Fri, 3 Jan 2014 23:55:52 -0500
Subject: [PATCH] Moving the radix-2 to expose trivial twiddle factors

---
 celt/kiss_fft.c           | 67 +++++++++++++++++++++++++++------------
 celt/static_modes_fixed.h | 26 +++++++--------
 celt/static_modes_float.h | 26 +++++++--------
 3 files changed, 73 insertions(+), 46 deletions(-)

diff --git a/celt/kiss_fft.c b/celt/kiss_fft.c
index a7bdde81..fa6cd568 100644
--- a/celt/kiss_fft.c
+++ b/celt/kiss_fft.c
@@ -47,32 +47,55 @@
 
 static void kf_bfly2(
                      kiss_fft_cpx * Fout,
-                     const size_t fstride,
-                     const kiss_fft_state *st,
                      int m,
-                     int N,
-                     int mm
+                     int N
                     )
 {
    kiss_fft_cpx * Fout2;
-   const kiss_twiddle_cpx * tw1;
-   int i,j;
-   kiss_fft_cpx * Fout_beg = Fout;
-   for (i=0;i<N;i++)
+   int i;
+#ifdef CUSTOM_MODES
+   if (m==1)
    {
-      Fout = Fout_beg + i*mm;
-      Fout2 = Fout + m;
-      tw1 = st->twiddles;
-      /* For non-custom modes, m is guaranteed to be a multiple of 4. */
-      for(j=0;j<m;j++)
+      celt_assert(m==1);
+      for (i=0;i<N;i++)
       {
          kiss_fft_cpx t;
-         C_MUL (t,  *Fout2 , *tw1);
-         tw1 += fstride;
+         Fout2 = Fout + 1;
+         t = *Fout2;
          C_SUB( *Fout2 ,  *Fout , t );
          C_ADDTO( *Fout ,  t );
-         ++Fout2;
-         ++Fout;
+         Fout += 2;
+      }
+   } else
+#endif
+   {
+      opus_val16 tw;
+      tw = QCONST16(0.7071067812f, 15);
+      /* We know that m==4 here because the radix-2 is just after a radix-4 */
+      celt_assert(m==4);
+      for (i=0;i<N;i++)
+      {
+         kiss_fft_cpx t;
+         Fout2 = Fout + 4;
+         t = Fout2[0];
+         C_SUB( Fout2[0] ,  Fout[0] , t );
+         C_ADDTO( Fout[0] ,  t );
+
+         t.r = S_MUL(Fout2[1].r+Fout2[1].i, tw);
+         t.i = S_MUL(Fout2[1].i-Fout2[1].r, tw);
+         C_SUB( Fout2[1] ,  Fout[1] , t );
+         C_ADDTO( Fout[1] ,  t );
+
+         t.r = Fout2[2].i;
+         t.i = -Fout2[2].r;
+         C_SUB( Fout2[2] ,  Fout[2] , t );
+         C_ADDTO( Fout[2] ,  t );
+
+         t.r = S_MUL(Fout2[3].i-Fout2[3].r, tw);
+         t.i = S_MUL(-Fout2[3].i-Fout2[3].r, tw);
+         C_SUB( Fout2[3] ,  Fout[3] , t );
+         C_ADDTO( Fout[3] ,  t );
+         Fout += 8;
       }
    }
 }
@@ -119,8 +142,7 @@ static void kf_bfly4(
       {
          Fout = Fout_beg + i*mm;
          tw3 = tw2 = tw1 = st->twiddles;
-         /* For non-custom modes, m=4, otherwise m is guaranteed to be a
-            multiple of 4. */
+         /* m is guaranteed to be a multiple of 4. */
          for (j=0;j<m;j++)
          {
             C_MUL(scratch[0],Fout[m] , *tw1 );
@@ -355,6 +377,11 @@ int kf_factor(int n,opus_int16 * facbuf)
            return 0;
         }
         facbuf[2*stages] = p;
+        if (p==2 && stages > 1)
+        {
+           facbuf[2*stages] = 4;
+           facbuf[2] = 2;
+        }
         stages++;
     } while (n > 1);
     n = nbak;
@@ -503,7 +530,7 @@ void opus_fft_impl(const kiss_fft_state *st,kiss_fft_cpx *fout)
        switch (st->factors[2*i])
        {
        case 2:
-          kf_bfly2(fout,fstride[i]<<shift,st,m, fstride[i], m2);
+          kf_bfly2(fout, m, fstride[i]);
           break;
        case 4:
           kf_bfly4(fout,fstride[i]<<shift,st,m, fstride[i], m2);
diff --git a/celt/static_modes_fixed.h b/celt/static_modes_fixed.h
index 7c5750cf..c822929e 100644
--- a/celt/static_modes_fixed.h
+++ b/celt/static_modes_fixed.h
@@ -342,36 +342,36 @@ static const kiss_twiddle_cpx fft_twiddles48000_960[480] = {
 #define FFT_BITREV480
 static const opus_int16 fft_bitrev480[480] = {
 0, 96, 192, 288, 384, 32, 128, 224, 320, 416, 64, 160, 256, 352, 448,
-16, 112, 208, 304, 400, 48, 144, 240, 336, 432, 80, 176, 272, 368, 464,
-4, 100, 196, 292, 388, 36, 132, 228, 324, 420, 68, 164, 260, 356, 452,
-20, 116, 212, 308, 404, 52, 148, 244, 340, 436, 84, 180, 276, 372, 468,
 8, 104, 200, 296, 392, 40, 136, 232, 328, 424, 72, 168, 264, 360, 456,
+16, 112, 208, 304, 400, 48, 144, 240, 336, 432, 80, 176, 272, 368, 464,
 24, 120, 216, 312, 408, 56, 152, 248, 344, 440, 88, 184, 280, 376, 472,
+4, 100, 196, 292, 388, 36, 132, 228, 324, 420, 68, 164, 260, 356, 452,
 12, 108, 204, 300, 396, 44, 140, 236, 332, 428, 76, 172, 268, 364, 460,
+20, 116, 212, 308, 404, 52, 148, 244, 340, 436, 84, 180, 276, 372, 468,
 28, 124, 220, 316, 412, 60, 156, 252, 348, 444, 92, 188, 284, 380, 476,
 1, 97, 193, 289, 385, 33, 129, 225, 321, 417, 65, 161, 257, 353, 449,
-17, 113, 209, 305, 401, 49, 145, 241, 337, 433, 81, 177, 273, 369, 465,
-5, 101, 197, 293, 389, 37, 133, 229, 325, 421, 69, 165, 261, 357, 453,
-21, 117, 213, 309, 405, 53, 149, 245, 341, 437, 85, 181, 277, 373, 469,
 9, 105, 201, 297, 393, 41, 137, 233, 329, 425, 73, 169, 265, 361, 457,
+17, 113, 209, 305, 401, 49, 145, 241, 337, 433, 81, 177, 273, 369, 465,
 25, 121, 217, 313, 409, 57, 153, 249, 345, 441, 89, 185, 281, 377, 473,
+5, 101, 197, 293, 389, 37, 133, 229, 325, 421, 69, 165, 261, 357, 453,
 13, 109, 205, 301, 397, 45, 141, 237, 333, 429, 77, 173, 269, 365, 461,
+21, 117, 213, 309, 405, 53, 149, 245, 341, 437, 85, 181, 277, 373, 469,
 29, 125, 221, 317, 413, 61, 157, 253, 349, 445, 93, 189, 285, 381, 477,
 2, 98, 194, 290, 386, 34, 130, 226, 322, 418, 66, 162, 258, 354, 450,
-18, 114, 210, 306, 402, 50, 146, 242, 338, 434, 82, 178, 274, 370, 466,
-6, 102, 198, 294, 390, 38, 134, 230, 326, 422, 70, 166, 262, 358, 454,
-22, 118, 214, 310, 406, 54, 150, 246, 342, 438, 86, 182, 278, 374, 470,
 10, 106, 202, 298, 394, 42, 138, 234, 330, 426, 74, 170, 266, 362, 458,
+18, 114, 210, 306, 402, 50, 146, 242, 338, 434, 82, 178, 274, 370, 466,
 26, 122, 218, 314, 410, 58, 154, 250, 346, 442, 90, 186, 282, 378, 474,
+6, 102, 198, 294, 390, 38, 134, 230, 326, 422, 70, 166, 262, 358, 454,
 14, 110, 206, 302, 398, 46, 142, 238, 334, 430, 78, 174, 270, 366, 462,
+22, 118, 214, 310, 406, 54, 150, 246, 342, 438, 86, 182, 278, 374, 470,
 30, 126, 222, 318, 414, 62, 158, 254, 350, 446, 94, 190, 286, 382, 478,
 3, 99, 195, 291, 387, 35, 131, 227, 323, 419, 67, 163, 259, 355, 451,
-19, 115, 211, 307, 403, 51, 147, 243, 339, 435, 83, 179, 275, 371, 467,
-7, 103, 199, 295, 391, 39, 135, 231, 327, 423, 71, 167, 263, 359, 455,
-23, 119, 215, 311, 407, 55, 151, 247, 343, 439, 87, 183, 279, 375, 471,
 11, 107, 203, 299, 395, 43, 139, 235, 331, 427, 75, 171, 267, 363, 459,
+19, 115, 211, 307, 403, 51, 147, 243, 339, 435, 83, 179, 275, 371, 467,
 27, 123, 219, 315, 411, 59, 155, 251, 347, 443, 91, 187, 283, 379, 475,
+7, 103, 199, 295, 391, 39, 135, 231, 327, 423, 71, 167, 263, 359, 455,
 15, 111, 207, 303, 399, 47, 143, 239, 335, 431, 79, 175, 271, 367, 463,
+23, 119, 215, 311, 407, 55, 151, 247, 343, 439, 87, 183, 279, 375, 471,
 31, 127, 223, 319, 415, 63, 159, 255, 351, 447, 95, 191, 287, 383, 479,
 };
 #endif
@@ -429,7 +429,7 @@ static const kiss_fft_state fft_state48000_960_0 = {
 17476,	/* scale */
 8,      /* scale_shift */
 -1,     /* shift */
-{5, 96, 3, 32, 2, 16, 4, 4, 4, 1, 0, 0, 0, 0, 0, 0, },	/* factors */
+{5, 96, 3, 32, 4, 8, 2, 4, 4, 1, 0, 0, 0, 0, 0, 0, },	/* factors */
 fft_bitrev480,  /* bitrev */
 fft_twiddles48000_960,  /* bitrev */
 };
diff --git a/celt/static_modes_float.h b/celt/static_modes_float.h
index 5b88e9fb..c4a5938d 100644
--- a/celt/static_modes_float.h
+++ b/celt/static_modes_float.h
@@ -342,36 +342,36 @@ static const kiss_twiddle_cpx fft_twiddles48000_960[480] = {
 #define FFT_BITREV480
 static const opus_int16 fft_bitrev480[480] = {
 0, 96, 192, 288, 384, 32, 128, 224, 320, 416, 64, 160, 256, 352, 448,
-16, 112, 208, 304, 400, 48, 144, 240, 336, 432, 80, 176, 272, 368, 464,
-4, 100, 196, 292, 388, 36, 132, 228, 324, 420, 68, 164, 260, 356, 452,
-20, 116, 212, 308, 404, 52, 148, 244, 340, 436, 84, 180, 276, 372, 468,
 8, 104, 200, 296, 392, 40, 136, 232, 328, 424, 72, 168, 264, 360, 456,
+16, 112, 208, 304, 400, 48, 144, 240, 336, 432, 80, 176, 272, 368, 464,
 24, 120, 216, 312, 408, 56, 152, 248, 344, 440, 88, 184, 280, 376, 472,
+4, 100, 196, 292, 388, 36, 132, 228, 324, 420, 68, 164, 260, 356, 452,
 12, 108, 204, 300, 396, 44, 140, 236, 332, 428, 76, 172, 268, 364, 460,
+20, 116, 212, 308, 404, 52, 148, 244, 340, 436, 84, 180, 276, 372, 468,
 28, 124, 220, 316, 412, 60, 156, 252, 348, 444, 92, 188, 284, 380, 476,
 1, 97, 193, 289, 385, 33, 129, 225, 321, 417, 65, 161, 257, 353, 449,
-17, 113, 209, 305, 401, 49, 145, 241, 337, 433, 81, 177, 273, 369, 465,
-5, 101, 197, 293, 389, 37, 133, 229, 325, 421, 69, 165, 261, 357, 453,
-21, 117, 213, 309, 405, 53, 149, 245, 341, 437, 85, 181, 277, 373, 469,
 9, 105, 201, 297, 393, 41, 137, 233, 329, 425, 73, 169, 265, 361, 457,
+17, 113, 209, 305, 401, 49, 145, 241, 337, 433, 81, 177, 273, 369, 465,
 25, 121, 217, 313, 409, 57, 153, 249, 345, 441, 89, 185, 281, 377, 473,
+5, 101, 197, 293, 389, 37, 133, 229, 325, 421, 69, 165, 261, 357, 453,
 13, 109, 205, 301, 397, 45, 141, 237, 333, 429, 77, 173, 269, 365, 461,
+21, 117, 213, 309, 405, 53, 149, 245, 341, 437, 85, 181, 277, 373, 469,
 29, 125, 221, 317, 413, 61, 157, 253, 349, 445, 93, 189, 285, 381, 477,
 2, 98, 194, 290, 386, 34, 130, 226, 322, 418, 66, 162, 258, 354, 450,
-18, 114, 210, 306, 402, 50, 146, 242, 338, 434, 82, 178, 274, 370, 466,
-6, 102, 198, 294, 390, 38, 134, 230, 326, 422, 70, 166, 262, 358, 454,
-22, 118, 214, 310, 406, 54, 150, 246, 342, 438, 86, 182, 278, 374, 470,
 10, 106, 202, 298, 394, 42, 138, 234, 330, 426, 74, 170, 266, 362, 458,
+18, 114, 210, 306, 402, 50, 146, 242, 338, 434, 82, 178, 274, 370, 466,
 26, 122, 218, 314, 410, 58, 154, 250, 346, 442, 90, 186, 282, 378, 474,
+6, 102, 198, 294, 390, 38, 134, 230, 326, 422, 70, 166, 262, 358, 454,
 14, 110, 206, 302, 398, 46, 142, 238, 334, 430, 78, 174, 270, 366, 462,
+22, 118, 214, 310, 406, 54, 150, 246, 342, 438, 86, 182, 278, 374, 470,
 30, 126, 222, 318, 414, 62, 158, 254, 350, 446, 94, 190, 286, 382, 478,
 3, 99, 195, 291, 387, 35, 131, 227, 323, 419, 67, 163, 259, 355, 451,
-19, 115, 211, 307, 403, 51, 147, 243, 339, 435, 83, 179, 275, 371, 467,
-7, 103, 199, 295, 391, 39, 135, 231, 327, 423, 71, 167, 263, 359, 455,
-23, 119, 215, 311, 407, 55, 151, 247, 343, 439, 87, 183, 279, 375, 471,
 11, 107, 203, 299, 395, 43, 139, 235, 331, 427, 75, 171, 267, 363, 459,
+19, 115, 211, 307, 403, 51, 147, 243, 339, 435, 83, 179, 275, 371, 467,
 27, 123, 219, 315, 411, 59, 155, 251, 347, 443, 91, 187, 283, 379, 475,
+7, 103, 199, 295, 391, 39, 135, 231, 327, 423, 71, 167, 263, 359, 455,
 15, 111, 207, 303, 399, 47, 143, 239, 335, 431, 79, 175, 271, 367, 463,
+23, 119, 215, 311, 407, 55, 151, 247, 343, 439, 87, 183, 279, 375, 471,
 31, 127, 223, 319, 415, 63, 159, 255, 351, 447, 95, 191, 287, 383, 479,
 };
 #endif
@@ -428,7 +428,7 @@ static const kiss_fft_state fft_state48000_960_0 = {
 480,    /* nfft */
 0.002083333f,   /* scale */
 -1,     /* shift */
-{5, 96, 3, 32, 2, 16, 4, 4, 4, 1, 0, 0, 0, 0, 0, 0, },	/* factors */
+{5, 96, 3, 32, 4, 8, 2, 4, 4, 1, 0, 0, 0, 0, 0, 0, },   /* factors */
 fft_bitrev480,  /* bitrev */
 fft_twiddles48000_960,  /* bitrev */
 };