internal void simd_gain_ramp_stereo(f32 *dst, const f32 *src, u32 frames, f32 g0, f32 g1) { f32 step = (frames > 0) ? (g1 - g0) / (f32)frames : 0.0f; u32 frame = 0; #if SIMD_NEON || SIMD_SSE // NOTE: one vector holds two stereo frames: {L0 R0 L1 R1}, so the gain lanes are {g g g+s g+s}. // we do 4 frames per iteration (two vectors) to keep both fma pipes busy. if(g0 == g1) { f32x4 gain = f32x4_set1(g0); for(; frame + 4 <= frames; frame += 4) { f32x4 a = f32x4_load(src + frame * 2 + 0); f32x4 b = f32x4_load(src + frame * 2 + 4); f32x4_store(dst + frame * 2 + 0, f32x4_mul(a, gain)); f32x4_store(dst + frame * 2 + 4, f32x4_mul(b, gain)); } } else { f32x4 gain_a = f32x4_make(g0, g0, g0 + step, g0 + step); f32x4 gain_b = f32x4_make(g0 + 2 * step, g0 + 2 * step, g0 + 3 * step, g0 + 3 * step); f32x4 delta = f32x4_set1(4 * step); for(; frame + 4 <= frames; frame += 4) { f32x4 a = f32x4_load(src + frame * 2 + 0); f32x4 b = f32x4_load(src + frame * 2 + 4); f32x4_store(dst + frame * 2 + 0, f32x4_mul(a, gain_a)); f32x4_store(dst + frame * 2 + 4, f32x4_mul(b, gain_b)); gain_a = f32x4_add(gain_a, delta); gain_b = f32x4_add(gain_b, delta); } } #endif for(; frame < frames; ++frame) { f32 gain = g0 + step * (f32)frame; dst[frame * 2 + 0] = src[frame * 2 + 0] * gain; dst[frame * 2 + 1] = src[frame * 2 + 1] * gain; } } internal void simd_gain_ramp_copy_stereo(f32 *dst, const f32 *src, u32 frames, f32 g, f32 step) { u32 frame = 0; #if SIMD_NEON || SIMD_SSE f32x4 gain_a = f32x4_make(g, g, g + step, g + step); f32x4 gain_b = f32x4_make(g + 2 * step, g + 2 * step, g + 3 * step, g + 3 * step); f32x4 delta = f32x4_set1(4 * step); for(; frame + 4 <= frames; frame += 4) { f32x4 a = f32x4_load(src + frame * 2 + 0); f32x4 b = f32x4_load(src + frame * 2 + 4); f32x4_store(dst + frame * 2 + 0, f32x4_mul(a, gain_a)); f32x4_store(dst + frame * 2 + 4, f32x4_mul(b, gain_b)); gain_a = f32x4_add(gain_a, delta); gain_b = f32x4_add(gain_b, delta); } #endif for(; frame < frames; ++frame) { f32 gain = g + step * (f32)frame; dst[frame * 2 + 0] = src[frame * 2 + 0] * gain; dst[frame * 2 + 1] = src[frame * 2 + 1] * gain; } } internal f32 simd_peak(const f32 *samples, u32 count) { u32 index = 0; f32 peak = 0.0f; #if SIMD_NEON || SIMD_SSE f32x4 acc_a = f32x4_set1(0.0f); f32x4 acc_b = f32x4_set1(0.0f); for(; index + 8 <= count; index += 8) { acc_a = f32x4_max(acc_a, f32x4_abs(f32x4_load(samples + index + 0))); acc_b = f32x4_max(acc_b, f32x4_abs(f32x4_load(samples + index + 4))); } peak = f32x4_hmax(f32x4_max(acc_a, acc_b)); #endif for(; index < count; ++index) { f32 value = fabsf(samples[index]); peak = Max(peak, value); } return peak; } internal void simd_mono_window(f32 *out, const f32 *stereo, const f32 *window, u32 frames) { u32 frame = 0; #if SIMD_NEON // NOTE: vld2q deinterleaves for free: val[0] = 4 lefts, val[1] = 4 rights f32x4 half = f32x4_set1(0.5f); for(; frame + 4 <= frames; frame += 4) { float32x4x2_t lr = vld2q_f32(stereo + frame * 2); f32x4 mono = f32x4_mul(f32x4_add(lr.val[0], lr.val[1]), half); f32x4_store(out + frame, f32x4_mul(mono, f32x4_load(window + frame))); } #elif SIMD_SSE f32x4 half = f32x4_set1(0.5f); for(; frame + 4 <= frames; frame += 4) { f32x4 a = f32x4_load(stereo + frame * 2 + 0); // L0 R0 L1 R1 f32x4 b = f32x4_load(stereo + frame * 2 + 4); // L2 R2 L3 R3 f32x4 left = _mm_shuffle_ps(a, b, _MM_SHUFFLE(2, 0, 2, 0)); f32x4 right = _mm_shuffle_ps(a, b, _MM_SHUFFLE(3, 1, 3, 1)); f32x4 mono = f32x4_mul(f32x4_add(left, right), half); f32x4_store(out + frame, f32x4_mul(mono, f32x4_load(window + frame))); } #endif for(; frame < frames; ++frame) { out[frame] = 0.5f * (stereo[frame * 2] + stereo[frame * 2 + 1]) * window[frame]; } } internal void simd_power(f32 *out, const f32 *re, const f32 *im, u32 count) { u32 index = 0; #if SIMD_NEON || SIMD_SSE for(; index + 4 <= count; index += 4) { f32x4 r = f32x4_load(re + index); f32x4 i = f32x4_load(im + index); f32x4_store(out + index, f32x4_madd(f32x4_mul(r, r), i, i)); } #endif for(; index < count; ++index) { out[index] = re[index] * re[index] + im[index] * im[index]; } }