The Tor Browser / file comparison
comparison: media/libspeex_resampler/sse-detect-runtime.patch
media/libspeex_resampler/sse-detect-runtime.patch
- changeset 0
- 6474c204b198
equal
deleted
inserted
replaced
| -1:000000000000 | 0:81011f22fdf9 |
|---|---|
| 1 --- /home/paul/workspace/repositories/opus-tools/src/resample.c 2012-11-21 11:36:59.119430163 +0100 | |
| 2 +++ media/libspeex_resampler/src/resample.c 2013-08-09 19:24:39.060236120 +0200 | |
| 3 @@ -92,18 +92,28 @@ | |
| 4 | |
| 5 #define IMAX(a,b) ((a) > (b) ? (a) : (b)) | |
| 6 #define IMIN(a,b) ((a) < (b) ? (a) : (b)) | |
| 7 | |
| 8 #ifndef NULL | |
| 9 #define NULL 0 | |
| 10 #endif | |
| 11 | |
| 12 +#include "sse_detect.h" | |
| 13 + | |
| 14 +/* We compile SSE code on x86 all the time, but we only use it if we find at | |
| 15 + * runtime that the CPU supports it. */ | |
| 16 #if defined(FLOATING_POINT) && defined(__SSE__) | |
| 17 +#if defined(_MSC_VER) | |
| 18 +#define inline __inline | |
| 19 +#endif | |
| 20 # include "resample_sse.h" | |
| 21 +#ifdef _MSC_VER | |
| 22 +#undef inline | |
| 23 +#endif | |
| 24 #endif | |
| 25 | |
| 26 /* Numer of elements to allocate on the stack */ | |
| 27 #ifdef VAR_ARRAYS | |
| 28 #define FIXED_STACK_ALLOC 8192 | |
| 29 #else | |
| 30 #define FIXED_STACK_ALLOC 1024 | |
| 31 #endif | |
| 32 @@ -340,35 +350,39 @@ | |
| 33 const spx_uint32_t den_rate = st->den_rate; | |
| 34 spx_word32_t sum; | |
| 35 | |
| 36 while (!(last_sample >= (spx_int32_t)*in_len || out_sample >= (spx_int32_t)*out_len)) | |
| 37 { | |
| 38 const spx_word16_t *sinct = & sinc_table[samp_frac_num*N]; | |
| 39 const spx_word16_t *iptr = & in[last_sample]; | |
| 40 | |
| 41 -#ifndef OVERRIDE_INNER_PRODUCT_SINGLE | |
| 42 +#ifdef OVERRIDE_INNER_PRODUCT_SINGLE | |
| 43 + if (moz_has_sse()) { | |
| 44 + sum = inner_product_single(sinct, iptr, N); | |
| 45 + } else { | |
| 46 +#endif | |
| 47 int j; | |
| 48 sum = 0; | |
| 49 for(j=0;j<N;j++) sum += MULT16_16(sinct[j], iptr[j]); | |
| 50 | |
| 51 /* This code is slower on most DSPs which have only 2 accumulators. | |
| 52 Plus this this forces truncation to 32 bits and you lose the HW guard bits. | |
| 53 I think we can trust the compiler and let it vectorize and/or unroll itself. | |
| 54 spx_word32_t accum[4] = {0,0,0,0}; | |
| 55 for(j=0;j<N;j+=4) { | |
| 56 accum[0] += MULT16_16(sinct[j], iptr[j]); | |
| 57 accum[1] += MULT16_16(sinct[j+1], iptr[j+1]); | |
| 58 accum[2] += MULT16_16(sinct[j+2], iptr[j+2]); | |
| 59 accum[3] += MULT16_16(sinct[j+3], iptr[j+3]); | |
| 60 } | |
| 61 sum = accum[0] + accum[1] + accum[2] + accum[3]; | |
| 62 */ | |
| 63 -#else | |
| 64 - sum = inner_product_single(sinct, iptr, N); | |
| 65 +#ifdef OVERRIDE_INNER_PRODUCT_SINGLE | |
| 66 + } | |
| 67 #endif | |
| 68 | |
| 69 out[out_stride * out_sample++] = SATURATE32(PSHR32(sum, 15), 32767); | |
| 70 last_sample += int_advance; | |
| 71 samp_frac_num += frac_advance; | |
| 72 if (samp_frac_num >= den_rate) | |
| 73 { | |
| 74 samp_frac_num -= den_rate; | |
| 75 @@ -397,29 +411,33 @@ | |
| 76 const spx_uint32_t den_rate = st->den_rate; | |
| 77 double sum; | |
| 78 | |
| 79 while (!(last_sample >= (spx_int32_t)*in_len || out_sample >= (spx_int32_t)*out_len)) | |
| 80 { | |
| 81 const spx_word16_t *sinct = & sinc_table[samp_frac_num*N]; | |
| 82 const spx_word16_t *iptr = & in[last_sample]; | |
| 83 | |
| 84 -#ifndef OVERRIDE_INNER_PRODUCT_DOUBLE | |
| 85 - int j; | |
| 86 - double accum[4] = {0,0,0,0}; | |
| 87 - | |
| 88 - for(j=0;j<N;j+=4) { | |
| 89 - accum[0] += sinct[j]*iptr[j]; | |
| 90 - accum[1] += sinct[j+1]*iptr[j+1]; | |
| 91 - accum[2] += sinct[j+2]*iptr[j+2]; | |
| 92 - accum[3] += sinct[j+3]*iptr[j+3]; | |
| 93 +#ifdef OVERRIDE_INNER_PRODUCT_DOUBLE | |
| 94 + if(moz_has_sse2()) { | |
| 95 + sum = inner_product_double(sinct, iptr, N); | |
| 96 + } else { | |
| 97 +#endif | |
| 98 + int j; | |
| 99 + double accum[4] = {0,0,0,0}; | |
| 100 + | |
| 101 + for(j=0;j<N;j+=4) { | |
| 102 + accum[0] += sinct[j]*iptr[j]; | |
| 103 + accum[1] += sinct[j+1]*iptr[j+1]; | |
| 104 + accum[2] += sinct[j+2]*iptr[j+2]; | |
| 105 + accum[3] += sinct[j+3]*iptr[j+3]; | |
| 106 + } | |
| 107 + sum = accum[0] + accum[1] + accum[2] + accum[3]; | |
| 108 +#ifdef OVERRIDE_INNER_PRODUCT_DOUBLE | |
| 109 } | |
| 110 - sum = accum[0] + accum[1] + accum[2] + accum[3]; | |
| 111 -#else | |
| 112 - sum = inner_product_double(sinct, iptr, N); | |
| 113 #endif | |
| 114 | |
| 115 out[out_stride * out_sample++] = PSHR32(sum, 15); | |
| 116 last_sample += int_advance; | |
| 117 samp_frac_num += frac_advance; | |
| 118 if (samp_frac_num >= den_rate) | |
| 119 { | |
| 120 samp_frac_num -= den_rate; | |
| 121 @@ -453,35 +471,38 @@ | |
| 122 #ifdef FIXED_POINT | |
| 123 const spx_word16_t frac = PDIV32(SHL32((samp_frac_num*st->oversample) % st->den_rate,15),st->den_rate); | |
| 124 #else | |
| 125 const spx_word16_t frac = ((float)((samp_frac_num*st->oversample) % st->den_rate))/st->den_rate; | |
| 126 #endif | |
| 127 spx_word16_t interp[4]; | |
| 128 | |
| 129 | |
| 130 -#ifndef OVERRIDE_INTERPOLATE_PRODUCT_SINGLE | |
| 131 - int j; | |
| 132 - spx_word32_t accum[4] = {0,0,0,0}; | |
| 133 - | |
| 134 - for(j=0;j<N;j++) { | |
| 135 - const spx_word16_t curr_in=iptr[j]; | |
| 136 - accum[0] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset-2]); | |
| 137 - accum[1] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset-1]); | |
| 138 - accum[2] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset]); | |
| 139 - accum[3] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset+1]); | |
| 140 +#ifdef OVERRIDE_INTERPOLATE_PRODUCT_SINGLE | |
| 141 + if (moz_has_sse()) { | |
| 142 + cubic_coef(frac, interp); | |
| 143 + sum = interpolate_product_single(iptr, st->sinc_table + st->oversample + 4 - offset - 2, N, st->oversample, interp); | |
| 144 + } else { | |
| 145 +#endif | |
| 146 + int j; | |
| 147 + spx_word32_t accum[4] = {0,0,0,0}; | |
| 148 + | |
| 149 + for(j=0;j<N;j++) { | |
| 150 + const spx_word16_t curr_in=iptr[j]; | |
| 151 + accum[0] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset-2]); | |
| 152 + accum[1] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset-1]); | |
| 153 + accum[2] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset]); | |
| 154 + accum[3] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset+1]); | |
| 155 + } | |
| 156 + cubic_coef(frac, interp); | |
| 157 + sum = MULT16_32_Q15(interp[0],SHR32(accum[0], 1)) + MULT16_32_Q15(interp[1],SHR32(accum[1], 1)) + MULT16_32_Q15(interp[2],SHR32(accum[2], 1)) + MULT16_32_Q15(interp[3],SHR32(accum[3], 1)); | |
| 158 +#ifdef OVERRIDE_INTERPOLATE_PRODUCT_SINGLE | |
| 159 } | |
| 160 - | |
| 161 - cubic_coef(frac, interp); | |
| 162 - sum = MULT16_32_Q15(interp[0],SHR32(accum[0], 1)) + MULT16_32_Q15(interp[1],SHR32(accum[1], 1)) + MULT16_32_Q15(interp[2],SHR32(accum[2], 1)) + MULT16_32_Q15(interp[3],SHR32(accum[3], 1)); | |
| 163 -#else | |
| 164 - cubic_coef(frac, interp); | |
| 165 - sum = interpolate_product_single(iptr, st->sinc_table + st->oversample + 4 - offset - 2, N, st->oversample, interp); | |
| 166 #endif | |
| 167 - | |
| 168 + | |
| 169 out[out_stride * out_sample++] = SATURATE32(PSHR32(sum, 14), 32767); | |
| 170 last_sample += int_advance; | |
| 171 samp_frac_num += frac_advance; | |
| 172 if (samp_frac_num >= den_rate) | |
| 173 { | |
| 174 samp_frac_num -= den_rate; | |
| 175 last_sample++; | |
| 176 } | |
| 177 @@ -515,35 +536,38 @@ | |
| 178 #ifdef FIXED_POINT | |
| 179 const spx_word16_t frac = PDIV32(SHL32((samp_frac_num*st->oversample) % st->den_rate,15),st->den_rate); | |
| 180 #else | |
| 181 const spx_word16_t frac = ((float)((samp_frac_num*st->oversample) % st->den_rate))/st->den_rate; | |
| 182 #endif | |
| 183 spx_word16_t interp[4]; | |
| 184 | |
| 185 | |
| 186 -#ifndef OVERRIDE_INTERPOLATE_PRODUCT_DOUBLE | |
| 187 +#ifdef OVERRIDE_INTERPOLATE_PRODUCT_DOUBLE | |
| 188 + if (moz_has_sse2()) { | |
| 189 + cubic_coef(frac, interp); | |
| 190 + sum = interpolate_product_double(iptr, st->sinc_table + st->oversample + 4 - offset - 2, N, st->oversample, interp); | |
| 191 + } else { | |
| 192 +#endif | |
| 193 int j; | |
| 194 double accum[4] = {0,0,0,0}; | |
| 195 | |
| 196 for(j=0;j<N;j++) { | |
| 197 const double curr_in=iptr[j]; | |
| 198 accum[0] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset-2]); | |
| 199 accum[1] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset-1]); | |
| 200 accum[2] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset]); | |
| 201 accum[3] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset+1]); | |
| 202 } | |
| 203 | |
| 204 cubic_coef(frac, interp); | |
| 205 sum = MULT16_32_Q15(interp[0],accum[0]) + MULT16_32_Q15(interp[1],accum[1]) + MULT16_32_Q15(interp[2],accum[2]) + MULT16_32_Q15(interp[3],accum[3]); | |
| 206 -#else | |
| 207 - cubic_coef(frac, interp); | |
| 208 - sum = interpolate_product_double(iptr, st->sinc_table + st->oversample + 4 - offset - 2, N, st->oversample, interp); | |
| 209 +#ifdef OVERRIDE_INNER_PRODUCT_DOUBLE | |
| 210 + } | |
| 211 #endif | |
| 212 - | |
| 213 out[out_stride * out_sample++] = PSHR32(sum,15); | |
| 214 last_sample += int_advance; | |
| 215 samp_frac_num += frac_advance; | |
| 216 if (samp_frac_num >= den_rate) | |
| 217 { | |
| 218 samp_frac_num -= den_rate; | |
| 219 last_sample++; | |
| 220 } |
