michael@0: --- /home/paul/workspace/repositories/opus-tools/src/resample.c	2012-11-21 11:36:59.119430163 +0100
michael@0: +++ media/libspeex_resampler/src/resample.c	2013-08-09 19:24:39.060236120 +0200
michael@0: @@ -92,18 +92,28 @@
michael@0:                 
michael@0:  #define IMAX(a,b) ((a) > (b) ? (a) : (b))
michael@0:  #define IMIN(a,b) ((a) < (b) ? (a) : (b))
michael@0:  
michael@0:  #ifndef NULL
michael@0:  #define NULL 0
michael@0:  #endif
michael@0:  
michael@0: +#include "sse_detect.h"
michael@0: +
michael@0: +/* We compile SSE code on x86 all the time, but we only use it if we find at
michael@0: + * runtime that the CPU supports it. */
michael@0:  #if defined(FLOATING_POINT) && defined(__SSE__)
michael@0: +#if defined(_MSC_VER)
michael@0: +#define inline __inline
michael@0: +#endif
michael@0:  # include "resample_sse.h"
michael@0: +#ifdef _MSC_VER
michael@0: +#undef inline
michael@0: +#endif
michael@0:  #endif
michael@0:  
michael@0:  /* Numer of elements to allocate on the stack */
michael@0:  #ifdef VAR_ARRAYS
michael@0:  #define FIXED_STACK_ALLOC 8192
michael@0:  #else
michael@0:  #define FIXED_STACK_ALLOC 1024
michael@0:  #endif
michael@0: @@ -340,35 +350,39 @@
michael@0:     const spx_uint32_t den_rate = st->den_rate;
michael@0:     spx_word32_t sum;
michael@0:  
michael@0:     while (!(last_sample >= (spx_int32_t)*in_len || out_sample >= (spx_int32_t)*out_len))
michael@0:     {
michael@0:        const spx_word16_t *sinct = & sinc_table[samp_frac_num*N];
michael@0:        const spx_word16_t *iptr = & in[last_sample];
michael@0:  
michael@0: -#ifndef OVERRIDE_INNER_PRODUCT_SINGLE
michael@0: +#ifdef OVERRIDE_INNER_PRODUCT_SINGLE
michael@0: +    if (moz_has_sse()) {
michael@0: +      sum = inner_product_single(sinct, iptr, N);
michael@0: +    } else {
michael@0: +#endif
michael@0:        int j;
michael@0:        sum = 0;
michael@0:        for(j=0;j<N;j++) sum += MULT16_16(sinct[j], iptr[j]);
michael@0:  
michael@0:  /*    This code is slower on most DSPs which have only 2 accumulators.
michael@0:        Plus this this forces truncation to 32 bits and you lose the HW guard bits.
michael@0:        I think we can trust the compiler and let it vectorize and/or unroll itself.
michael@0:        spx_word32_t accum[4] = {0,0,0,0};
michael@0:        for(j=0;j<N;j+=4) {
michael@0:          accum[0] += MULT16_16(sinct[j], iptr[j]);
michael@0:          accum[1] += MULT16_16(sinct[j+1], iptr[j+1]);
michael@0:          accum[2] += MULT16_16(sinct[j+2], iptr[j+2]);
michael@0:          accum[3] += MULT16_16(sinct[j+3], iptr[j+3]);
michael@0:        }
michael@0:        sum = accum[0] + accum[1] + accum[2] + accum[3];
michael@0:  */
michael@0: -#else
michael@0: -      sum = inner_product_single(sinct, iptr, N);
michael@0: +#ifdef OVERRIDE_INNER_PRODUCT_SINGLE
michael@0: +    }
michael@0:  #endif
michael@0:  
michael@0:        out[out_stride * out_sample++] = SATURATE32(PSHR32(sum, 15), 32767);
michael@0:        last_sample += int_advance;
michael@0:        samp_frac_num += frac_advance;
michael@0:        if (samp_frac_num >= den_rate)
michael@0:        {
michael@0:           samp_frac_num -= den_rate;
michael@0: @@ -397,29 +411,33 @@
michael@0:     const spx_uint32_t den_rate = st->den_rate;
michael@0:     double sum;
michael@0:  
michael@0:     while (!(last_sample >= (spx_int32_t)*in_len || out_sample >= (spx_int32_t)*out_len))
michael@0:     {
michael@0:        const spx_word16_t *sinct = & sinc_table[samp_frac_num*N];
michael@0:        const spx_word16_t *iptr = & in[last_sample];
michael@0:  
michael@0: -#ifndef OVERRIDE_INNER_PRODUCT_DOUBLE
michael@0: -      int j;
michael@0: -      double accum[4] = {0,0,0,0};
michael@0: -
michael@0: -      for(j=0;j<N;j+=4) {
michael@0: -        accum[0] += sinct[j]*iptr[j];
michael@0: -        accum[1] += sinct[j+1]*iptr[j+1];
michael@0: -        accum[2] += sinct[j+2]*iptr[j+2];
michael@0: -        accum[3] += sinct[j+3]*iptr[j+3];
michael@0: +#ifdef OVERRIDE_INNER_PRODUCT_DOUBLE
michael@0: +      if(moz_has_sse2()) {
michael@0: +        sum = inner_product_double(sinct, iptr, N);
michael@0: +      } else {
michael@0: +#endif
michael@0: +        int j;
michael@0: +        double accum[4] = {0,0,0,0};
michael@0: +
michael@0: +        for(j=0;j<N;j+=4) {
michael@0: +          accum[0] += sinct[j]*iptr[j];
michael@0: +          accum[1] += sinct[j+1]*iptr[j+1];
michael@0: +          accum[2] += sinct[j+2]*iptr[j+2];
michael@0: +          accum[3] += sinct[j+3]*iptr[j+3];
michael@0: +        }
michael@0: +        sum = accum[0] + accum[1] + accum[2] + accum[3];
michael@0: +#ifdef OVERRIDE_INNER_PRODUCT_DOUBLE
michael@0:        }
michael@0: -      sum = accum[0] + accum[1] + accum[2] + accum[3];
michael@0: -#else
michael@0: -      sum = inner_product_double(sinct, iptr, N);
michael@0:  #endif
michael@0:  
michael@0:        out[out_stride * out_sample++] = PSHR32(sum, 15);
michael@0:        last_sample += int_advance;
michael@0:        samp_frac_num += frac_advance;
michael@0:        if (samp_frac_num >= den_rate)
michael@0:        {
michael@0:           samp_frac_num -= den_rate;
michael@0: @@ -453,35 +471,38 @@
michael@0:  #ifdef FIXED_POINT
michael@0:        const spx_word16_t frac = PDIV32(SHL32((samp_frac_num*st->oversample) % st->den_rate,15),st->den_rate);
michael@0:  #else
michael@0:        const spx_word16_t frac = ((float)((samp_frac_num*st->oversample) % st->den_rate))/st->den_rate;
michael@0:  #endif
michael@0:        spx_word16_t interp[4];
michael@0:  
michael@0:  
michael@0: -#ifndef OVERRIDE_INTERPOLATE_PRODUCT_SINGLE
michael@0: -      int j;
michael@0: -      spx_word32_t accum[4] = {0,0,0,0};
michael@0: -
michael@0: -      for(j=0;j<N;j++) {
michael@0: -        const spx_word16_t curr_in=iptr[j];
michael@0: -        accum[0] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset-2]);
michael@0: -        accum[1] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset-1]);
michael@0: -        accum[2] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset]);
michael@0: -        accum[3] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset+1]);
michael@0: +#ifdef OVERRIDE_INTERPOLATE_PRODUCT_SINGLE
michael@0: +      if (moz_has_sse()) {
michael@0: +        cubic_coef(frac, interp);
michael@0: +        sum = interpolate_product_single(iptr, st->sinc_table + st->oversample + 4 - offset - 2, N, st->oversample, interp);
michael@0: +      } else {
michael@0: +#endif
michael@0: +        int j;
michael@0: +        spx_word32_t accum[4] = {0,0,0,0};
michael@0: +
michael@0: +        for(j=0;j<N;j++) {
michael@0: +          const spx_word16_t curr_in=iptr[j];
michael@0: +          accum[0] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset-2]);
michael@0: +          accum[1] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset-1]);
michael@0: +          accum[2] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset]);
michael@0: +          accum[3] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset+1]);
michael@0: +        }
michael@0: +        cubic_coef(frac, interp);
michael@0: +        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));
michael@0: +#ifdef OVERRIDE_INTERPOLATE_PRODUCT_SINGLE
michael@0:        }
michael@0: -
michael@0: -      cubic_coef(frac, interp);
michael@0: -      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));
michael@0: -#else
michael@0: -      cubic_coef(frac, interp);
michael@0: -      sum = interpolate_product_single(iptr, st->sinc_table + st->oversample + 4 - offset - 2, N, st->oversample, interp);
michael@0:  #endif
michael@0: -      
michael@0: +
michael@0:        out[out_stride * out_sample++] = SATURATE32(PSHR32(sum, 14), 32767);
michael@0:        last_sample += int_advance;
michael@0:        samp_frac_num += frac_advance;
michael@0:        if (samp_frac_num >= den_rate)
michael@0:        {
michael@0:           samp_frac_num -= den_rate;
michael@0:           last_sample++;
michael@0:        }
michael@0: @@ -515,35 +536,38 @@
michael@0:  #ifdef FIXED_POINT
michael@0:        const spx_word16_t frac = PDIV32(SHL32((samp_frac_num*st->oversample) % st->den_rate,15),st->den_rate);
michael@0:  #else
michael@0:        const spx_word16_t frac = ((float)((samp_frac_num*st->oversample) % st->den_rate))/st->den_rate;
michael@0:  #endif
michael@0:        spx_word16_t interp[4];
michael@0:  
michael@0:  
michael@0: -#ifndef OVERRIDE_INTERPOLATE_PRODUCT_DOUBLE
michael@0: +#ifdef OVERRIDE_INTERPOLATE_PRODUCT_DOUBLE
michael@0: +      if (moz_has_sse2()) {
michael@0: +        cubic_coef(frac, interp);
michael@0: +        sum = interpolate_product_double(iptr, st->sinc_table + st->oversample + 4 - offset - 2, N, st->oversample, interp);
michael@0: +      } else {
michael@0: +#endif
michael@0:        int j;
michael@0:        double accum[4] = {0,0,0,0};
michael@0:  
michael@0:        for(j=0;j<N;j++) {
michael@0:          const double curr_in=iptr[j];
michael@0:          accum[0] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset-2]);
michael@0:          accum[1] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset-1]);
michael@0:          accum[2] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset]);
michael@0:          accum[3] += MULT16_16(curr_in,st->sinc_table[4+(j+1)*st->oversample-offset+1]);
michael@0:        }
michael@0:  
michael@0:        cubic_coef(frac, interp);
michael@0:        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]);
michael@0: -#else
michael@0: -      cubic_coef(frac, interp);
michael@0: -      sum = interpolate_product_double(iptr, st->sinc_table + st->oversample + 4 - offset - 2, N, st->oversample, interp);
michael@0: +#ifdef OVERRIDE_INNER_PRODUCT_DOUBLE
michael@0: +      }
michael@0:  #endif
michael@0: -      
michael@0:        out[out_stride * out_sample++] = PSHR32(sum,15);
michael@0:        last_sample += int_advance;
michael@0:        samp_frac_num += frac_advance;
michael@0:        if (samp_frac_num >= den_rate)
michael@0:        {
michael@0:           samp_frac_num -= den_rate;
michael@0:           last_sample++;
michael@0:        }