1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libspeex_resampler/sse-detect-runtime.patch Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,220 @@
1.4 +--- /home/paul/workspace/repositories/opus-tools/src/resample.c 2012-11-21 11:36:59.119430163 +0100
1.5 ++++ media/libspeex_resampler/src/resample.c 2013-08-09 19:24:39.060236120 +0200
1.6 +@@ -92,18 +92,28 @@
1.7 +
1.8 + #define IMAX(a,b) ((a) > (b) ? (a) : (b))
1.9 + #define IMIN(a,b) ((a) < (b) ? (a) : (b))
1.10 +
1.11 + #ifndef NULL
1.12 + #define NULL 0
1.13 + #endif
1.14 +
1.15 ++#include "sse_detect.h"
1.16 ++
1.17 ++/* We compile SSE code on x86 all the time, but we only use it if we find at
1.18 ++ * runtime that the CPU supports it. */
1.19 + #if defined(FLOATING_POINT) && defined(__SSE__)
1.20 ++#if defined(_MSC_VER)
1.21 ++#define inline __inline
1.22 ++#endif
1.23 + # include "resample_sse.h"
1.24 ++#ifdef _MSC_VER
1.25 ++#undef inline
1.26 ++#endif
1.27 + #endif
1.28 +
1.29 + /* Numer of elements to allocate on the stack */
1.30 + #ifdef VAR_ARRAYS
1.31 + #define FIXED_STACK_ALLOC 8192
1.32 + #else
1.33 + #define FIXED_STACK_ALLOC 1024
1.34 + #endif
1.35 +@@ -340,35 +350,39 @@
1.36 + const spx_uint32_t den_rate = st->den_rate;
1.37 + spx_word32_t sum;
1.38 +
1.39 + while (!(last_sample >= (spx_int32_t)*in_len || out_sample >= (spx_int32_t)*out_len))
1.40 + {
1.41 + const spx_word16_t *sinct = & sinc_table[samp_frac_num*N];
1.42 + const spx_word16_t *iptr = & in[last_sample];
1.43 +
1.44 +-#ifndef OVERRIDE_INNER_PRODUCT_SINGLE
1.45 ++#ifdef OVERRIDE_INNER_PRODUCT_SINGLE
1.46 ++ if (moz_has_sse()) {
1.47 ++ sum = inner_product_single(sinct, iptr, N);
1.48 ++ } else {
1.49 ++#endif
1.50 + int j;
1.51 + sum = 0;
1.52 + for(j=0;j<N;j++) sum += MULT16_16(sinct[j], iptr[j]);
1.53 +
1.54 + /* This code is slower on most DSPs which have only 2 accumulators.
1.55 + Plus this this forces truncation to 32 bits and you lose the HW guard bits.
1.56 + I think we can trust the compiler and let it vectorize and/or unroll itself.
1.57 + spx_word32_t accum[4] = {0,0,0,0};
1.58 + for(j=0;j<N;j+=4) {
1.59 + accum[0] += MULT16_16(sinct[j], iptr[j]);
1.60 + accum[1] += MULT16_16(sinct[j+1], iptr[j+1]);
1.61 + accum[2] += MULT16_16(sinct[j+2], iptr[j+2]);
1.62 + accum[3] += MULT16_16(sinct[j+3], iptr[j+3]);
1.63 + }
1.64 + sum = accum[0] + accum[1] + accum[2] + accum[3];
1.65 + */
1.66 +-#else
1.67 +- sum = inner_product_single(sinct, iptr, N);
1.68 ++#ifdef OVERRIDE_INNER_PRODUCT_SINGLE
1.69 ++ }
1.70 + #endif
1.71 +
1.72 + out[out_stride * out_sample++] = SATURATE32(PSHR32(sum, 15), 32767);
1.73 + last_sample += int_advance;
1.74 + samp_frac_num += frac_advance;
1.75 + if (samp_frac_num >= den_rate)
1.76 + {
1.77 + samp_frac_num -= den_rate;
1.78 +@@ -397,29 +411,33 @@
1.79 + const spx_uint32_t den_rate = st->den_rate;
1.80 + double sum;
1.81 +
1.82 + while (!(last_sample >= (spx_int32_t)*in_len || out_sample >= (spx_int32_t)*out_len))
1.83 + {
1.84 + const spx_word16_t *sinct = & sinc_table[samp_frac_num*N];
1.85 + const spx_word16_t *iptr = & in[last_sample];
1.86 +
1.87 +-#ifndef OVERRIDE_INNER_PRODUCT_DOUBLE
1.88 +- int j;
1.89 +- double accum[4] = {0,0,0,0};
1.90 +-
1.91 +- for(j=0;j<N;j+=4) {
1.92 +- accum[0] += sinct[j]*iptr[j];
1.93 +- accum[1] += sinct[j+1]*iptr[j+1];
1.94 +- accum[2] += sinct[j+2]*iptr[j+2];
1.95 +- accum[3] += sinct[j+3]*iptr[j+3];
1.96 ++#ifdef OVERRIDE_INNER_PRODUCT_DOUBLE
1.97 ++ if(moz_has_sse2()) {
1.98 ++ sum = inner_product_double(sinct, iptr, N);
1.99 ++ } else {
1.100 ++#endif
1.101 ++ int j;
1.102 ++ double accum[4] = {0,0,0,0};
1.103 ++
1.104 ++ for(j=0;j<N;j+=4) {
1.105 ++ accum[0] += sinct[j]*iptr[j];
1.106 ++ accum[1] += sinct[j+1]*iptr[j+1];
1.107 ++ accum[2] += sinct[j+2]*iptr[j+2];
1.108 ++ accum[3] += sinct[j+3]*iptr[j+3];
1.109 ++ }
1.110 ++ sum = accum[0] + accum[1] + accum[2] + accum[3];
1.111 ++#ifdef OVERRIDE_INNER_PRODUCT_DOUBLE
1.112 + }
1.113 +- sum = accum[0] + accum[1] + accum[2] + accum[3];
1.114 +-#else
1.115 +- sum = inner_product_double(sinct, iptr, N);
1.116 + #endif
1.117 +
1.118 + out[out_stride * out_sample++] = PSHR32(sum, 15);
1.119 + last_sample += int_advance;
1.120 + samp_frac_num += frac_advance;
1.121 + if (samp_frac_num >= den_rate)
1.122 + {
1.123 + samp_frac_num -= den_rate;
1.124 +@@ -453,35 +471,38 @@
1.125 + #ifdef FIXED_POINT
1.126 + const spx_word16_t frac = PDIV32(SHL32((samp_frac_num*st->oversample) % st->den_rate,15),st->den_rate);
1.127 + #else
1.128 + const spx_word16_t frac = ((float)((samp_frac_num*st->oversample) % st->den_rate))/st->den_rate;
1.129 + #endif
1.130 + spx_word16_t interp[4];
1.131 +
1.132 +
1.133 +-#ifndef OVERRIDE_INTERPOLATE_PRODUCT_SINGLE
1.134 +- int j;
1.135 +- spx_word32_t accum[4] = {0,0,0,0};
1.136 +-
1.137 +- for(j=0;j<N;j++) {
1.138 +- const spx_word16_t curr_in=iptr[j];
1.139 +- accum[0] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset-2]);
1.140 +- accum[1] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset-1]);
1.141 +- accum[2] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset]);
1.142 +- accum[3] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset+1]);
1.143 ++#ifdef OVERRIDE_INTERPOLATE_PRODUCT_SINGLE
1.144 ++ if (moz_has_sse()) {
1.145 ++ cubic_coef(frac, interp);
1.146 ++ sum = interpolate_product_single(iptr, st->sinc_table + st->oversample + 4 - offset - 2, N, st->oversample, interp);
1.147 ++ } else {
1.148 ++#endif
1.149 ++ int j;
1.150 ++ spx_word32_t accum[4] = {0,0,0,0};
1.151 ++
1.152 ++ for(j=0;j<N;j++) {
1.153 ++ const spx_word16_t curr_in=iptr[j];
1.154 ++ accum[0] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset-2]);
1.155 ++ accum[1] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset-1]);
1.156 ++ accum[2] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset]);
1.157 ++ accum[3] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset+1]);
1.158 ++ }
1.159 ++ cubic_coef(frac, interp);
1.160 ++ 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));
1.161 ++#ifdef OVERRIDE_INTERPOLATE_PRODUCT_SINGLE
1.162 + }
1.163 +-
1.164 +- cubic_coef(frac, interp);
1.165 +- 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));
1.166 +-#else
1.167 +- cubic_coef(frac, interp);
1.168 +- sum = interpolate_product_single(iptr, st->sinc_table + st->oversample + 4 - offset - 2, N, st->oversample, interp);
1.169 + #endif
1.170 +-
1.171 ++
1.172 + out[out_stride * out_sample++] = SATURATE32(PSHR32(sum, 14), 32767);
1.173 + last_sample += int_advance;
1.174 + samp_frac_num += frac_advance;
1.175 + if (samp_frac_num >= den_rate)
1.176 + {
1.177 + samp_frac_num -= den_rate;
1.178 + last_sample++;
1.179 + }
1.180 +@@ -515,35 +536,38 @@
1.181 + #ifdef FIXED_POINT
1.182 + const spx_word16_t frac = PDIV32(SHL32((samp_frac_num*st->oversample) % st->den_rate,15),st->den_rate);
1.183 + #else
1.184 + const spx_word16_t frac = ((float)((samp_frac_num*st->oversample) % st->den_rate))/st->den_rate;
1.185 + #endif
1.186 + spx_word16_t interp[4];
1.187 +
1.188 +
1.189 +-#ifndef OVERRIDE_INTERPOLATE_PRODUCT_DOUBLE
1.190 ++#ifdef OVERRIDE_INTERPOLATE_PRODUCT_DOUBLE
1.191 ++ if (moz_has_sse2()) {
1.192 ++ cubic_coef(frac, interp);
1.193 ++ sum = interpolate_product_double(iptr, st->sinc_table + st->oversample + 4 - offset - 2, N, st->oversample, interp);
1.194 ++ } else {
1.195 ++#endif
1.196 + int j;
1.197 + double accum[4] = {0,0,0,0};
1.198 +
1.199 + for(j=0;j<N;j++) {
1.200 + const double curr_in=iptr[j];
1.201 + accum[0] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset-2]);
1.202 + accum[1] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset-1]);
1.203 + accum[2] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset]);
1.204 + accum[3] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset+1]);
1.205 + }
1.206 +
1.207 + cubic_coef(frac, interp);
1.208 + 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]);
1.209 +-#else
1.210 +- cubic_coef(frac, interp);
1.211 +- sum = interpolate_product_double(iptr, st->sinc_table + st->oversample + 4 - offset - 2, N, st->oversample, interp);
1.212 ++#ifdef OVERRIDE_INNER_PRODUCT_DOUBLE
1.213 ++ }
1.214 + #endif
1.215 +-
1.216 + out[out_stride * out_sample++] = PSHR32(sum,15);
1.217 + last_sample += int_advance;
1.218 + samp_frac_num += frac_advance;
1.219 + if (samp_frac_num >= den_rate)
1.220 + {
1.221 + samp_frac_num -= den_rate;
1.222 + last_sample++;
1.223 + }
