/* Copyright (c) 2013 The Chromium OS Authors. All rights reserved. * Use of this source code is governed by a BSD-style license that can be * found in the LICENSE file. */ #include <stdlib.h> #include "eq2.h" struct eq2 { int n[2]; struct biquad biquad[MAX_BIQUADS_PER_EQ2][2]; }; struct eq2 *eq2_new() { struct eq2 *eq2 = (struct eq2 *)calloc(1, sizeof(*eq2)); int i, j; /* Initialize all biquads to identity filter, so if two channels have * different numbers of biquads, it still works. */ for (i = 0; i < MAX_BIQUADS_PER_EQ2; i++) for (j = 0; j < 2; j++) biquad_set(&eq2->biquad[i][j], BQ_NONE, 0, 0, 0); return eq2; } void eq2_free(struct eq2 *eq2) { free(eq2); } int eq2_append_biquad(struct eq2 *eq2, int channel, enum biquad_type type, float freq, float Q, float gain) { if (eq2->n[channel] >= MAX_BIQUADS_PER_EQ2) return -1; biquad_set(&eq2->biquad[eq2->n[channel]++][channel], type, freq, Q, gain); return 0; } int eq2_append_biquad_direct(struct eq2 *eq2, int channel, const struct biquad *biquad) { if (eq2->n[channel] >= MAX_BIQUADS_PER_EQ2) return -1; eq2->biquad[eq2->n[channel]++][channel] = *biquad; return 0; } static inline void eq2_process_one(struct biquad (*bq)[2], float *data0, float *data1, int count) { struct biquad *qL = &bq[0][0]; struct biquad *qR = &bq[0][1]; float x1L = qL->x1; float x2L = qL->x2; float y1L = qL->y1; float y2L = qL->y2; float b0L = qL->b0; float b1L = qL->b1; float b2L = qL->b2; float a1L = qL->a1; float a2L = qL->a2; float x1R = qR->x1; float x2R = qR->x2; float y1R = qR->y1; float y2R = qR->y2; float b0R = qR->b0; float b1R = qR->b1; float b2R = qR->b2; float a1R = qR->a1; float a2R = qR->a2; int j; for (j = 0; j < count; j++) { float xL = data0[j]; float xR = data1[j]; float yL = b0L*xL + b1L*x1L + b2L*x2L - a1L*y1L - a2L*y2L; x2L = x1L; x1L = xL; y2L = y1L; y1L = yL; float yR = b0R*xR + b1R*x1R + b2R*x2R - a1R*y1R - a2R*y2R; x2R = x1R; x1R = xR; y2R = y1R; y1R = yR; data0[j] = yL; data1[j] = yR; } qL->x1 = x1L; qL->x2 = x2L; qL->y1 = y1L; qL->y2 = y2L; qR->x1 = x1R; qR->x2 = x2R; qR->y1 = y1R; qR->y2 = y2R; } #ifdef __ARM_NEON__ #include <arm_neon.h> static inline void eq2_process_two_neon(struct biquad (*bq)[2], float *data0, float *data1, int count) { struct biquad *qL = &bq[0][0]; struct biquad *rL = &bq[1][0]; struct biquad *qR = &bq[0][1]; struct biquad *rR = &bq[1][1]; float32x2_t x1 = {qL->x1, qR->x1}; float32x2_t x2 = {qL->x2, qR->x2}; float32x2_t y1 = {qL->y1, qR->y1}; float32x2_t y2 = {qL->y2, qR->y2}; float32x2_t qb0 = {qL->b0, qR->b0}; float32x2_t qb1 = {qL->b1, qR->b1}; float32x2_t qb2 = {qL->b2, qR->b2}; float32x2_t qa1 = {qL->a1, qR->a1}; float32x2_t qa2 = {qL->a2, qR->a2}; float32x2_t z1 = {rL->y1, rR->y1}; float32x2_t z2 = {rL->y2, rR->y2}; float32x2_t rb0 = {rL->b0, rR->b0}; float32x2_t rb1 = {rL->b1, rR->b1}; float32x2_t rb2 = {rL->b2, rR->b2}; float32x2_t ra1 = {rL->a1, rR->a1}; float32x2_t ra2 = {rL->a2, rR->a2}; __asm__ __volatile__( /* d0 = x, d1 = y, d2 = z */ "1: \n" "vmul.f32 d1, %P[qb1], %P[x1] \n" "vld1.32 d0[0], [%[data0]] \n" "vld1.32 d0[1], [%[data1]] \n" "subs %[count], #1 \n" "vmul.f32 d2, %P[rb1], %P[y1] \n" "vmla.f32 d1, %P[qb0], d0 \n" "vmla.f32 d1, %P[qb2], %P[x2] \n" "vmov.f32 %P[x2], %P[x1] \n" "vmov.f32 %P[x1], d0 \n" "vmls.f32 d1, %P[qa1], %P[y1] \n" "vmls.f32 d1, %P[qa2], %P[y2] \n" "vmla.f32 d2, %P[rb0], d1 \n" "vmla.f32 d2, %P[rb2], %P[y2] \n" "vmov.f32 %P[y2], %P[y1] \n" "vmov.f32 %P[y1], d1 \n" "vmls.f32 d2, %P[ra1], %P[z1] \n" "vmls.f32 d2, %P[ra2], %P[z2] \n" "vmov.f32 %P[z2], %P[z1] \n" "vmov.f32 %P[z1], d2 \n" "vst1.f32 d2[0], [%[data0]]! \n" "vst1.f32 d2[1], [%[data1]]! \n" "bne 1b \n" : /* output */ [data0]"+r"(data0), [data1]"+r"(data1), [count]"+r"(count), [x1]"+w"(x1), [x2]"+w"(x2), [y1]"+w"(y1), [y2]"+w"(y2), [z1]"+w"(z1), [z2]"+w"(z2) : /* input */ [qb0]"w"(qb0), [qb1]"w"(qb1), [qb2]"w"(qb2), [qa1]"w"(qa1), [qa2]"w"(qa2), [rb0]"w"(rb0), [rb1]"w"(rb1), [rb2]"w"(rb2), [ra1]"w"(ra1), [ra2]"w"(ra2) : /* clobber */ "d0", "d1", "d2", "memory", "cc" ); qL->x1 = x1[0]; qL->x2 = x2[0]; qL->y1 = y1[0]; qL->y2 = y2[0]; rL->y1 = z1[0]; rL->y2 = z2[0]; qR->x1 = x1[1]; qR->x2 = x2[1]; qR->y1 = y1[1]; qR->y2 = y2[1]; rR->y1 = z1[1]; rR->y2 = z2[1]; } #endif #if defined(__SSE3__) && defined(__x86_64__) #include <emmintrin.h> static inline void eq2_process_two_sse3(struct biquad (*bq)[2], float *data0, float *data1, int count) { struct biquad *qL = &bq[0][0]; struct biquad *rL = &bq[1][0]; struct biquad *qR = &bq[0][1]; struct biquad *rR = &bq[1][1]; __m128 x1 = {qL->x1, qR->x1}; __m128 x2 = {qL->x2, qR->x2}; __m128 y1 = {qL->y1, qR->y1}; __m128 y2 = {qL->y2, qR->y2}; __m128 qb0 = {qL->b0, qR->b0}; __m128 qb1 = {qL->b1, qR->b1}; __m128 qb2 = {qL->b2, qR->b2}; __m128 qa1 = {qL->a1, qR->a1}; __m128 qa2 = {qL->a2, qR->a2}; __m128 z1 = {rL->y1, rR->y1}; __m128 z2 = {rL->y2, rR->y2}; __m128 rb0 = {rL->b0, rR->b0}; __m128 rb1 = {rL->b1, rR->b1}; __m128 rb2 = {rL->b2, rR->b2}; __m128 ra1 = {rL->a1, rR->a1}; __m128 ra2 = {rL->a2, rR->a2}; __asm__ __volatile__( "1: \n" "movss (%[data0]), %%xmm2 \n" "movss (%[data1]), %%xmm1 \n" "unpcklps %%xmm1, %%xmm2 \n" "mulps %[qb2],%[x2] \n" "lddqu %[qb0],%%xmm0 \n" "mulps %[ra2],%[z2] \n" "lddqu %[qb1],%%xmm1 \n" "mulps %%xmm2,%%xmm0 \n" "mulps %[x1],%%xmm1 \n" "addps %%xmm1,%%xmm0 \n" "movaps %[qa1],%%xmm1 \n" "mulps %[y1],%%xmm1 \n" "addps %[x2],%%xmm0 \n" "movaps %[rb1],%[x2] \n" "mulps %[y1],%[x2] \n" "subps %%xmm1,%%xmm0 \n" "movaps %[qa2],%%xmm1 \n" "mulps %[y2],%%xmm1 \n" "mulps %[rb2],%[y2] \n" "subps %%xmm1,%%xmm0 \n" "movaps %[rb0],%%xmm1 \n" "mulps %%xmm0,%%xmm1 \n" "addps %[x2],%%xmm1 \n" "movaps %[x1],%[x2] \n" "movaps %%xmm2,%[x1] \n" "addps %[y2],%%xmm1 \n" "movaps %[ra1],%[y2] \n" "mulps %[z1],%[y2] \n" "subps %[y2],%%xmm1 \n" "movaps %[y1],%[y2] \n" "movaps %%xmm0,%[y1] \n" "subps %[z2],%%xmm1 \n" "movaps %[z1],%[z2] \n" "movaps %%xmm1,%[z1] \n" "movss %%xmm1, (%[data0]) \n" "shufps $1, %%xmm1, %%xmm1 \n" "movss %%xmm1, (%[data1]) \n" "add $4, %[data0] \n" "add $4, %[data1] \n" "sub $1, %[count] \n" "jnz 1b \n" : /* output */ [data0]"+r"(data0), [data1]"+r"(data1), [count]"+r"(count), [x1]"+x"(x1), [x2]"+x"(x2), [y1]"+x"(y1), [y2]"+x"(y2), [z1]"+x"(z1), [z2]"+x"(z2) : /* input */ [qb0]"m"(qb0), [qb1]"m"(qb1), [qb2]"m"(qb2), [qa1]"x"(qa1), [qa2]"x"(qa2), [rb0]"x"(rb0), [rb1]"x"(rb1), [rb2]"x"(rb2), [ra1]"x"(ra1), [ra2]"x"(ra2) : /* clobber */ "xmm0", "xmm1", "xmm2", "memory", "cc" ); qL->x1 = x1[0]; qL->x2 = x2[0]; qL->y1 = y1[0]; qL->y2 = y2[0]; rL->y1 = z1[0]; rL->y2 = z2[0]; qR->x1 = x1[1]; qR->x2 = x2[1]; qR->y1 = y1[1]; qR->y2 = y2[1]; rR->y1 = z1[1]; rR->y2 = z2[1]; } #endif void eq2_process(struct eq2 *eq2, float *data0, float *data1, int count) { int i; int n; if (!count) return; n = eq2->n[0]; if (eq2->n[1] > n) n = eq2->n[1]; for (i = 0; i < n; i += 2) { if (i + 1 == n) { eq2_process_one(&eq2->biquad[i], data0, data1, count); } else { #if defined(__ARM_NEON__) eq2_process_two_neon(&eq2->biquad[i], data0, data1, count); #elif defined(__SSE3__) && defined(__x86_64__) eq2_process_two_sse3(&eq2->biquad[i], data0, data1, count); #else eq2_process_one(&eq2->biquad[i], data0, data1, count); eq2_process_one(&eq2->biquad[i+1], data0, data1, count); #endif } } }