API tools faq
paste
Advertisement
Guest User

Untitled

a guest
Dec 25th, 2013
343
0
Never
Add comment
Not a member of Pastebin yet? Sign Up, it unlocks many cool features!
ARM 9.08 KB | None | 0 0
raw download clone embed print report
  1. diff --git a/lib/rbcodec/codecs/libopus/celt/mdct.c b/lib/rbcodec/codecs/libopus/celt/mdct.c
  2. old mode 100644
  3. new mode 100755
  4. index 72ea180..974f2c1
  5. --- a/lib/rbcodec/codecs/libopus/celt/mdct.c
  6. +++ b/lib/rbcodec/codecs/libopus/celt/mdct.c
  7. @@ -210,6 +210,194 @@ void clt_mdct_forward(const mdct_lookup *l, kiss_fft_scalar *in, kiss_fft_scalar
  8.  }
  9.  #endif
  10.  
  11. +#if 1
  12. +void pre_rotate_960_armv5(kiss_fft_scalar * out, const kiss_twiddle_scalar *t, kiss_fft_scalar *in){
  13. +
  14. +    int N=960; //deliberately do not halve this so that we will have half word offsets for ldrsh
  15. +    //int N=2;
  16. +    kiss_fft_scalar *in2 = in+(960-1);
  17. +    kiss_fft_scalar in3=in;
  18. +               int i, num=960;
  19. +   //memset(out, 0, 960*4);
  20. +   kiss_fft_scalar * temp=out;
  21. +     kiss_twiddle_scalar *t2 = t+480;
  22. +     //DEBUGF("initial values: xp1:  %u,  xp2:  %u, out:  %u t1:  %u\n", (int)in, (int)in2,(int) out, (int)t2);
  23. +      asm volatile (
  24. +
  25. +         "0:"
  26. +
  27. +         "ldr r4, [%[t1], %[n]]; "     //bottom of this contains t1[n], unfortunately we count down
  28. +         "ldr r6, [%[xp1]]; "          //n is in units of half words
  29. +         "ldr r5, [%[t2], -%[n]]; "    //contains t1[n], t1[n+1] because we count up
  30. +         "ldr r9, [%[xp2]]; "
  31. +         "add %[xp1], %[xp1], #8;"
  32. +
  33. +          //compute yr
  34. +         "smulwb r7, r6, r4;"      //r8 isn't used, stupid armv4 ops
  35. +
  36. +         "smulwb r10, r9, r5;"     //remember to put 16 bit val in last register for early term
  37. +         "sub %[xp2], %[xp2], #8;"     //decrement the pointer while we wait for the multiply to finish
  38. +
  39. +         "sub r8, r7, r10;"                //r7, r10 free
  40. +         "mov r8, r8, lsl #1;"
  41. +
  42. +
  43. +         //compute yi
  44. +         "smulwb r10, r6, r5;"     //r6, r5 free
  45. +         "smlawb r5, r9, r4, r10;"     //r4, r9 free
  46. +         "mov r6,  #13;"               //sneak a constant load into this pipeline buble
  47. +
  48. +         //"add r7, r5, r10;"  //add both then make negative and finally shift, probably not optimal
  49. +         "rsb r7, r5, #0;"
  50. +         "mov r7, r7, lsl #1;"
  51. +
  52. +
  53. +         "smulwb r10, r7, r6;"             //r10=S_MUL(yi,13)
  54. +         "smlawb r9, r8, r6, r7;"              //r4=S_MUL(yr,13)
  55. +
  56. +         "sub r5, r8, r10;"
  57. +         //"add r9, r7, r4;"
  58. +
  59. +         //*yp++ =...
  60. +         "stmia %[out]!, {r5, r9};"
  61. +
  62. +         "subs  %[n], %[n], #2;"       //decrement by 1 half word
  63. +         "bne   0b;"
  64. +
  65. +
  66. +          : [out] "+r" (out), [n] "+r" (N),[xp1] "+r" (in), [xp2] "+r" (in2)  //6 registers used
  67. +          : [t1] "r" (t), [t2] "r" (t+480)
  68. +          : "r4", "r5", "r6", "r7", "r8", "r9", "r10", "memory", "cc");    //can use 7 more before we'll run out
  69. +
  70. +/*
  71. +          DEBUGF("final values:   xp1:  %u,  xp2:  %u, out:  %u\n", (int)in, (int)in2,(int) out);
  72. +          DEBUGF("final values:   xp1:  %u (%d) for N: %d\n", (int)in, ((int)in)-((int)in3), N);
  73. +              DEBUGF("out values 1->%d\n", num);
  74. +              for(i=0; i<num;i++)
  75. +               DEBUGF("%u ", temp[i]);
  76. +           DEBUGF("\n");
  77. + exit(0);
  78. + */
  79. +
  80. +}
  81. +#elif 0
  82. +void pre_rotate_960_armv4(kiss_fft_scalar * out, const kiss_twiddle_scalar *t, kiss_fft_scalar *in){
  83. +
  84. +    int N=960; //deliberately do not halve this so that we will have half word offsets for ldrsh
  85. +    //int N=2;
  86. +    kiss_fft_scalar *in2 = in+(960-1);
  87. +    kiss_fft_scalar in3=in;
  88. +               int i, num=960;
  89. +   //memset(out, 0, 960*4);
  90. +   kiss_fft_scalar * temp=out;
  91. +     kiss_twiddle_scalar *t2 = t+480;
  92. +     DEBUGF("initial values: xp1:  %u,  xp2:  %u, out:  %u t1:  %u t2:  %u\n", (int)in, (int)in2,(int) out,(int)t, (int)t2);
  93. +      asm volatile (
  94. +
  95. +         "0:"
  96. +         //we can't use ldm for either t or xp variables
  97. +         "ldr r4, [%[t1], %[n]]; "
  98. +         "ldr r6, [%[xp1]]; "          //n is in units of half words
  99. +         "ldrh r5, [%[t2], -%[n]]; "
  100. +         "ldr r9, [%[xp2]]; "
  101. +         "add %[xp1], %[xp1], #8;"
  102. +         "mov r4, r4, lsl #16;"
  103. +         "mov r5, r5, lsl #16;"
  104. +
  105. +          //compute yr
  106. +         "smull r8, r7, r6, r4;"       //r8 isn't used, stupid armv4 ops
  107. +         "sub %[xp2], %[xp2], #8;"     //decrement the pointer while we wait for the multiply to finish
  108. +
  109. +         "smull r8, r10, r9, r5;"      //remember to put 16 bit val in last register for early term
  110. +
  111. +
  112. +         "sub r8, r7, r10;"                //r7, r10 free
  113. +         "mov r8, r8, lsl #1;"
  114. +
  115. +
  116. +         //compute yi
  117. +         "smull r7, r10, r6, r5;"      //r6, r5 free
  118. +         "smull r7, r5, r9, r4;"       //r4, r9 free
  119. +         "mov r6,  #13;"               //sneak a constant load into this pipeline buble
  120. +
  121. +         "add r7, r5, r10;"    //add both then make negative and finally shift, probably not optimal
  122. +         "rsb r7, r7, #0;"
  123. +         "mov r7, r7, lsl #1;"
  124. +
  125. +
  126. +         "smull r5, r10, r7, r6;"              //r10=S_MUL(yi,13)
  127. +         "smull r5, r4, r8, r6;"               //r4=S_MUL(yr,13)
  128. +
  129. +         "sub r5, r8, r10;"
  130. +         "add r9, r7, r4;"
  131. +
  132. +         //*yp++ =...
  133. +         "stmia %[out]!, {r5, r9};"
  134. +
  135. +         "subs  %[n], %[n], #2;"       //decrement by 1 half word
  136. +         "bne   0b;"
  137. +
  138. +
  139. +          : [out] "+r" (out), [n] "+r" (N),[xp1] "+r" (in), [xp2] "+r" (in2)  //6 registers used
  140. +          : [t1] "r" (t), [t2] "r" (t+480)
  141. +          : "r4", "r5", "r6", "r7", "r8", "r9", "r10", "memory", "cc");    //can use 7 more before we'll run out
  142. +
  143. +}
  144. +#else
  145. +
  146. +
  147. + void pre_rotate_960_armv4(kiss_fft_scalar * out, const kiss_twiddle_scalar *t, kiss_fft_scalar *in){
  148. +   //DEBUGF("N: %d overlap: %d shift: %d stride: %d\n", N, overlap, shift, stride);    //N: 1920 overlap: 120 shift: 0 stride: 1
  149. +   int i;//memset(out, 0, 960*4);
  150. +     /* Temp pointers to make it really clear to the compiler what we're doing */
  151. +     const kiss_fft_scalar * OPUS_RESTRICT xp1 = in;
  152. +     const kiss_fft_scalar * OPUS_RESTRICT xp2 = in+(960-1);
  153. +     const kiss_twiddle_scalar *t2 = t+480;
  154. +     kiss_fft_scalar * OPUS_RESTRICT yp = out;
  155. +
  156. +   //      int ii, num=960;
  157. +
  158. +
  159. +          //DEBUGF("final values:   xp1:  %u,  xp2:  %u, out:  %u\n", (int)in, (int)in2,(int) out);
  160. +          //DEBUGF("final values:   xp1:  %u (%d) for N: %d\n", (int)in, ((int)in)-((int)in3), N);
  161. +
  162. +     //const kiss_twiddle_scalar *t = &l->trig[0];
  163. +
  164. +     for(i=480 ;i>0;i--)
  165. +     {
  166. +        kiss_fft_scalar yr, yi;
  167. +        yr = -S_MUL(*xp2, t2[-i]) + S_MUL(*xp1,t[i]);
  168. +        yi =  -S_MUL(*xp2, t[i]) - S_MUL(*xp1,t2[-i]);
  169. +
  170. +        DEBUGF("yr:  %u yi: %u, t1: %u, t2: %u xp2: %u xp1: %u, SMUL: %u\n", yr, yi, (int)&(t[i]),(int)&(t2[-i]), *xp2, *xp1, S_MUL(*xp1,t[i]));
  171. +       exit(0);
  172. +        /* works because the cos is nearly one */
  173. +        *yp++ = yr - S_MUL(yi,13); //sine==13 for N=1920
  174. +        *yp++ = yi + S_MUL(yr,13);
  175. +        xp1+=2;
  176. +        xp2-=2;
  177. +     }
  178. +
  179. +/*        DEBUGF("out values 1->%d\n", num);
  180. +          for(ii=0; ii<num;ii++)
  181. +           DEBUGF("%u ", out[ii]);
  182. +           DEBUGF("\n");
  183. + exit(0);*/
  184. +#if 0
  185. +     for(i=0;i<480;i++)
  186. +     {
  187. +        kiss_fft_scalar yr, yi;
  188. +        yr = -S_MUL(*xp2, t[i]) + S_MUL(*xp1,t[(480-i)]);      //[0 479], [480 1]
  189. +        yi =  -S_MUL(*xp2, t[(480-i)]) - S_MUL(*xp1,t[i]);     //[480, 1], [0 479]
  190. +        /* works because the cos is nearly one */
  191. +        *yp++ = yr - S_MUL(yi,13); //sine==13 for N=1920
  192. +        *yp++ = yi + S_MUL(yr,13);
  193. +        xp1+=2;
  194. +        xp2-=2;
  195. +     }
  196. +#endif
  197. +}
  198. +#endif
  199. void clt_mdct_backward(const mdct_lookup *l, kiss_fft_scalar *in, kiss_fft_scalar * OPUS_RESTRICT out,
  200.       const opus_val16 * OPUS_RESTRICT window, int overlap, int shift, int stride)
  201. {
  202. @@ -231,25 +419,27 @@ void clt_mdct_backward(const mdct_lookup *l, kiss_fft_scalar *in, kiss_fft_scala
  203.    sine = (kiss_twiddle_scalar)2*PI*(.125f)/N;
  204. #endif
  205.  
  206. -   /* Pre-rotate */
  207. -   {
  208. -      /* Temp pointers to make it really clear to the compiler what we're doing */
  209. -      const kiss_fft_scalar * OPUS_RESTRICT xp1 = in;
  210. -      const kiss_fft_scalar * OPUS_RESTRICT xp2 = in+stride*(N2-1);
  211. -      kiss_fft_scalar * OPUS_RESTRICT yp = f2;
  212. -      const kiss_twiddle_scalar *t = &l->trig[0];
  213. -      for(i=0;i<N4;i++)
  214. -      {
  215. -         kiss_fft_scalar yr, yi;
  216. -         yr = -S_MUL(*xp2, t[i<<shift]) + S_MUL(*xp1,t[(N4-i)<<shift]);
  217. -         yi =  -S_MUL(*xp2, t[(N4-i)<<shift]) - S_MUL(*xp1,t[i<<shift]);
  218. -         /* works because the cos is nearly one */
  219. -         *yp++ = yr - S_MUL(yi,sine);
  220. -         *yp++ = yi + S_MUL(yr,sine);
  221. -         xp1+=2*stride;
  222. -         xp2-=2*stride;
  223. -      }
  224. -   }
  225. +   if(shift > 1)
  226. +      {
  227. +         /* Temp pointers to make it really clear to the compiler what we're doing */
  228. +         const kiss_fft_scalar * OPUS_RESTRICT xp1 = in;
  229. +         const kiss_fft_scalar * OPUS_RESTRICT xp2 = in+stride*(N2-1);
  230. +         kiss_fft_scalar * OPUS_RESTRICT yp = f2;
  231. +         const kiss_twiddle_scalar *t = &l->trig[0];
  232. +         for(i=0;i<N4;i++)
  233. +         {
  234. +            kiss_fft_scalar yr, yi;
  235. +            yr = -S_MUL(*xp2, t[i<<shift]) + S_MUL(*xp1,t[(N4-i)<<shift]);
  236. +            yi =  -S_MUL(*xp2, t[(N4-i)<<shift]) - S_MUL(*xp1,t[i<<shift]);
  237. +            /* works because the cos is nearly one */
  238. +            *yp++ = yr - S_MUL(yi,sine);
  239. +            *yp++ = yi + S_MUL(yr,sine);
  240. +
  241. +            xp1+=2*stride;
  242. +            xp2-=2*stride;
  243. +         }
  244. +   }else
  245. +       pre_rotate_960_armv4(f2, &l->trig[0], in);
  246.  
  247.    /* Inverse N/4 complex FFT. This one should *not* downscale even in fixed-point */
  248.    opus_ifft(l->kfft[shift], (kiss_fft_cpx *)f2, (kiss_fft_cpx *)(out+(overlap>>1)));
  249. diff --git a/lib/rbcodec/codecs/libopus/config.h b/lib/rbcodec/codecs/libopus/config.h
  250. old mode 100644
  251. new mode 100755
  252. index 1ce75ff..5ac2f8f
  253. --- a/lib/rbcodec/codecs/libopus/config.h
  254. +++ b/lib/rbcodec/codecs/libopus/config.h
  255. @@ -43,7 +43,7 @@
  256. #if ARM_ARCH == 4
  257. #define OPUS_ARM_INLINE_ASM
  258. #elif ARM_ARCH > 4
  259. -#define OPUS_ARM_INLINE_EDSP
  260. +#define OPUS_ARM_INLINE_ASM
  261. #endif
  262. #endif
Advertisement
Add Comment
Please, Sign In to add comment
Advertisement
Public Pastes
Advertisement
We use cookies for various purposes including analytics. By continuing to use Pastebin, you agree to our use of cookies as described in the Cookies Policy.  OK, I Understand
Not a member of Pastebin yet?
Sign Up, it unlocks many cool features!