#ifndef AUDIO_SIMD_H #define AUDIO_SIMD_H // NOTE: 4 wide f32 vectors. neon on arm64 (this machine), sse2 on x64, scalar otherwise. // only what the audio and spectrum paths need. every kernel handles the tail scalar so callers // can pass any count. #if ARCH_ARM #include #define SIMD_NEON 1 #define SIMD_SSE 0 #elif ARCH_x86_64 #include #define SIMD_NEON 0 #define SIMD_SSE 1 #else #define SIMD_NEON 0 #define SIMD_SSE 0 #endif #if SIMD_NEON typedef float32x4_t f32x4; #define f32x4_load(pointer) vld1q_f32(pointer) #define f32x4_store(pointer, v) vst1q_f32((pointer), (v)) #define f32x4_set1(value) vdupq_n_f32(value) #define f32x4_add(a, b) vaddq_f32((a), (b)) #define f32x4_sub(a, b) vsubq_f32((a), (b)) #define f32x4_mul(a, b) vmulq_f32((a), (b)) #define f32x4_madd(acc, a, b) vfmaq_f32((acc), (a), (b)) // acc + a*b #define f32x4_msub(acc, a, b) vfmsq_f32((acc), (a), (b)) // acc - a*b #define f32x4_max(a, b) vmaxq_f32((a), (b)) #define f32x4_abs(a) vabsq_f32(a) #define f32x4_sqrt(a) vsqrtq_f32(a) #define f32x4_hmax(a) vmaxvq_f32(a) #define f32x4_hsum(a) vaddvq_f32(a) internal inline f32x4 f32x4_make(f32 a, f32 b, f32 c, f32 d) { f32 values[4] = {a, b, c, d}; return vld1q_f32(values); } #elif SIMD_SSE typedef __m128 f32x4; #define f32x4_load(pointer) _mm_loadu_ps(pointer) #define f32x4_store(pointer, v) _mm_storeu_ps((pointer), (v)) #define f32x4_set1(value) _mm_set1_ps(value) #define f32x4_add(a, b) _mm_add_ps((a), (b)) #define f32x4_sub(a, b) _mm_sub_ps((a), (b)) #define f32x4_mul(a, b) _mm_mul_ps((a), (b)) #define f32x4_madd(acc, a, b) _mm_add_ps((acc), _mm_mul_ps((a), (b))) #define f32x4_msub(acc, a, b) _mm_sub_ps((acc), _mm_mul_ps((a), (b))) #define f32x4_max(a, b) _mm_max_ps((a), (b)) #define f32x4_abs(a) _mm_andnot_ps(_mm_set1_ps(-0.0f), (a)) #define f32x4_sqrt(a) _mm_sqrt_ps(a) #define f32x4_make(a, b, c, d) _mm_setr_ps((a), (b), (c), (d)) internal inline f32 f32x4_hmax(f32x4 v) { v = _mm_max_ps(v, _mm_movehl_ps(v, v)); v = _mm_max_ss(v, _mm_shuffle_ps(v, v, 1)); return _mm_cvtss_f32(v); } internal inline f32 f32x4_hsum(f32x4 v) { v = _mm_add_ps(v, _mm_movehl_ps(v, v)); v = _mm_add_ss(v, _mm_shuffle_ps(v, v, 1)); return _mm_cvtss_f32(v); } #endif //- kernels // dst[i] = src[i] * gain, gain ramps linearly g0 -> g1 over the buffer. interleaved stereo. // dst may alias src. used for volume/fade so a change never clicks. internal void simd_gain_ramp_stereo(f32 *dst, const f32 *src, u32 frames, f32 g0, f32 g1); // dst[i] = src[i] * (g + step * i), interleaved stereo, dst and src don't overlap. the copy and // the gain in one pass: the realtime thread reads the ring and writes pipewire's buffer once. internal void simd_gain_ramp_copy_stereo(f32 *dst, const f32 *src, u32 frames, f32 g, f32 step); // peak |sample| over an interleaved stereo buffer internal f32 simd_peak(const f32 *samples, u32 count); // out[i] = 0.5 * (left + right) * window[i]. deinterleaves for the fft. internal void simd_mono_window(f32 *out, const f32 *stereo, const f32 *window, u32 frames); // out[i] = re[i]^2 + im[i]^2 internal void simd_power(f32 *out, const f32 *re, const f32 *im, u32 count); #endif // AUDIO_SIMD_H