diff --git a/pipewire-pulseaudio b/pipewire-pulseaudio index 96205687c..3d95ea4d1 160000 --- a/pipewire-pulseaudio +++ b/pipewire-pulseaudio @@ -1 +1 @@ -Subproject commit 96205687c514c0440739e3f5cf900998df3244de +Subproject commit 3d95ea4d13faa3feda2c60af1c27c4af55eb517a diff --git a/spa/plugins/audioconvert/fmt-ops-sse.c b/spa/plugins/audioconvert/fmt-ops-sse2.c similarity index 77% rename from spa/plugins/audioconvert/fmt-ops-sse.c rename to spa/plugins/audioconvert/fmt-ops-sse2.c index 674e2be4b..388a831fa 100644 --- a/spa/plugins/audioconvert/fmt-ops-sse.c +++ b/spa/plugins/audioconvert/fmt-ops-sse2.c @@ -30,16 +30,30 @@ #include static void -conv_s16_to_f32d_1_sse(void *data, int n_dst, void *dst[n_dst], const void *src, int n_samples) +conv_s16_to_f32d_1_sse2(void *data, int n_dst, void *dst[n_dst], const void *src, int n_samples) { const int16_t *s = src; float **d = (float **) dst; float *d0 = d[0]; - int n; + int n = 0, unrolled; + __m128i in; __m128 out, factor = _mm_set1_ps(1.0f / S16_SCALE); - for(n = 0; n_samples--; n++) { - out = _mm_cvtsi32_ss(out, *s); + unrolled = n_samples / 4; + n_samples = n_samples & 3; + + for(; unrolled--; n += 4) { + in = _mm_setr_epi32(s[0*n_dst], + s[1*n_dst], + s[2*n_dst], + s[3*n_dst]); + out = _mm_cvtepi32_ps(in); + out = _mm_mul_ps(out, factor); + _mm_storeu_ps(&d0[n], out); + s += 4*n_dst; + } + for(; n_samples--; n++) { + out = _mm_cvtsi32_ss(out, s[0]); out = _mm_mul_ss(out, factor); _mm_store_ss(&d0[n], out); s += n_dst; @@ -47,7 +61,7 @@ conv_s16_to_f32d_1_sse(void *data, int n_dst, void *dst[n_dst], const void *src, } static void -conv_s16_to_f32d_2_sse(void *data, int n_dst, void *dst[n_dst], const void *src, int n_samples) +conv_s16_to_f32d_2_sse2(void *data, int n_dst, void *dst[n_dst], const void *src, int n_samples) { const int16_t *s = src; float **d = (float **) dst; @@ -90,40 +104,42 @@ conv_s16_to_f32d_2_sse(void *data, int n_dst, void *dst[n_dst], const void *src, } static void -conv_s16_to_f32d_sse(void *data, int n_dst, void *dst[n_dst], int n_src, const void *src[n_src], int n_samples) +conv_s16_to_f32d_sse2(void *data, int n_dst, void *dst[n_dst], int n_src, const void *src[n_src], int n_samples) { const int16_t *s = src[0]; int i = 0; for(; i + 1 < n_dst; i += 2) - conv_s16_to_f32d_2_sse(data, n_dst, &dst[i], &s[i], n_samples); + conv_s16_to_f32d_2_sse2(data, n_dst, &dst[i], &s[i], n_samples); for(; i < n_dst; i++) - conv_s16_to_f32d_1_sse(data, n_dst, &dst[i], &s[i], n_samples); + conv_s16_to_f32d_1_sse2(data, n_dst, &dst[i], &s[i], n_samples); } static void -conv_s24_to_f32d_1_sse(void *data, int n_dst, void *dst[n_dst], const void *src, int n_samples) +conv_s24_to_f32d_1_sse2(void *data, int n_dst, void *dst[n_dst], const void *src, int n_samples) { const uint8_t *s = src; float **d = (float **) dst; float *d0 = d[0]; int n = 0, unrolled; __m128i in; - union { - __m128i in; - uint8_t b[16]; - } b; __m128 out, factor = _mm_set1_ps(1.0f / S24_SCALE); unrolled = n_samples / 4; n_samples = n_samples & 3; + if (n_samples == 0) { + n_samples += 4; + unrolled--; + } for(; unrolled--; n += 4) { - memcpy(&b.b[1], &s[0 * n_dst], 3); - memcpy(&b.b[5], &s[3 * n_dst], 3); - memcpy(&b.b[9], &s[6 * n_dst], 3); - memcpy(&b.b[13], &s[9 * n_dst], 3); - in = _mm_srai_epi32(b.in, 8); + in = _mm_setr_epi32( + *((uint32_t*)&s[0 * n_dst]), + *((uint32_t*)&s[3 * n_dst]), + *((uint32_t*)&s[6 * n_dst]), + *((uint32_t*)&s[9 * n_dst])); + in = _mm_slli_epi32(in, 8); + in = _mm_srai_epi32(in, 8); out = _mm_cvtepi32_ps(in); out = _mm_mul_ps(out, factor); _mm_storeu_ps(&d0[n], out); @@ -138,17 +154,17 @@ conv_s24_to_f32d_1_sse(void *data, int n_dst, void *dst[n_dst], const void *src, } static void -conv_s24_to_f32d_sse(void *data, int n_dst, void *dst[n_dst], int n_src, const void *src[n_src], int n_samples) +conv_s24_to_f32d_sse2(void *data, int n_dst, void *dst[n_dst], int n_src, const void *src[n_src], int n_samples) { const int8_t *s = src[0]; int i = 0; for(; i < n_dst; i++) - conv_s24_to_f32d_1_sse(data, n_dst, &dst[i], &s[3*i], n_samples); + conv_s24_to_f32d_1_sse2(data, n_dst, &dst[i], &s[3*i], n_samples); } static void -conv_f32d_to_s32_1_sse(void *data, void *dst, int n_src, const void *src[n_src], int n_samples) +conv_f32d_to_s32_1_sse2(void *data, void *dst, int n_src, const void *src[n_src], int n_samples) { const float **s = (const float **) src; const float *s0 = s[0]; @@ -187,7 +203,7 @@ conv_f32d_to_s32_1_sse(void *data, void *dst, int n_src, const void *src[n_src], } static void -conv_f32d_to_s32_2_sse(void *data, void *dst, int n_src, const void *src[n_src], int n_samples) +conv_f32d_to_s32_2_sse2(void *data, void *dst, int n_src, const void *src[n_src], int n_samples) { const float **s = (const float **) src; const float *s0 = s[0], *s1 = s[1]; @@ -237,7 +253,7 @@ conv_f32d_to_s32_2_sse(void *data, void *dst, int n_src, const void *src[n_src], } static void -conv_f32d_to_s32_4_sse(void *data, void *dst, int n_src, const void *src[n_src], int n_samples) +conv_f32d_to_s32_4_sse2(void *data, void *dst, int n_src, const void *src[n_src], int n_samples) { const float **s = (const float **) src; const float *s0 = s[0], *s1 = s[1], *s2 = s[2], *s3 = s[3]; @@ -302,44 +318,50 @@ conv_f32d_to_s32_4_sse(void *data, void *dst, int n_src, const void *src[n_src], } static void -conv_f32d_to_s32_sse(void *data, int n_dst, void *dst[n_dst], int n_src, const void *src[n_src], int n_samples) +conv_f32d_to_s32_sse2(void *data, int n_dst, void *dst[n_dst], int n_src, const void *src[n_src], int n_samples) { int32_t *d = dst[0]; int i = 0; for(; i + 3 < n_src; i += 4) - conv_f32d_to_s32_4_sse(data, &d[i], n_src, &src[i], n_samples); + conv_f32d_to_s32_4_sse2(data, &d[i], n_src, &src[i], n_samples); for(; i + 1 < n_src; i += 2) - conv_f32d_to_s32_2_sse(data, &d[i], n_src, &src[i], n_samples); + conv_f32d_to_s32_2_sse2(data, &d[i], n_src, &src[i], n_samples); for(; i < n_src; i++) - conv_f32d_to_s32_1_sse(data, &d[i], n_src, &src[i], n_samples); + conv_f32d_to_s32_1_sse2(data, &d[i], n_src, &src[i], n_samples); } static void -conv_f32d_to_s16_1_sse(void *data, void *dst, int n_src, const void *src[n_src], int n_samples) +conv_f32d_to_s16_1_sse2(void *data, void *dst, int n_src, const void *src[n_src], int n_samples) { const float **s = (const float **) src; const float *s0 = s[0]; int16_t *d = dst; int n, unrolled; - __m128 in[1]; - __m128i out[4]; + __m128 in[2]; + __m128i out[2]; __m128 int_max = _mm_set1_ps(S16_MAX_F); __m128 int_min = _mm_sub_ps(_mm_setzero_ps(), int_max); - unrolled = n_samples / 4; - n_samples = n_samples & 3; + unrolled = n_samples / 8; + n_samples = n_samples & 7; - for(n = 0; unrolled--; n += 4) { + for(n = 0; unrolled--; n += 8) { in[0] = _mm_mul_ps(_mm_loadu_ps(&s0[n]), int_max); + in[1] = _mm_mul_ps(_mm_loadu_ps(&s0[n+4]), int_max); out[0] = _mm_cvtps_epi32(in[0]); - out[0] = _mm_packs_epi32(out[0], out[0]); + out[1] = _mm_cvtps_epi32(in[1]); + out[0] = _mm_packs_epi32(out[0], out[1]); - d[0*n_src] = _mm_extract_pi16(*(__m64*)out, 0); - d[1*n_src] = _mm_extract_pi16(*(__m64*)out, 1); - d[2*n_src] = _mm_extract_pi16(*(__m64*)out, 2); - d[3*n_src] = _mm_extract_pi16(*(__m64*)out, 3); - d += 4*n_src; + d[0*n_src] = _mm_extract_epi16(out[0], 0); + d[1*n_src] = _mm_extract_epi16(out[0], 1); + d[2*n_src] = _mm_extract_epi16(out[0], 2); + d[3*n_src] = _mm_extract_epi16(out[0], 3); + d[4*n_src] = _mm_extract_epi16(out[0], 4); + d[5*n_src] = _mm_extract_epi16(out[0], 5); + d[6*n_src] = _mm_extract_epi16(out[0], 6); + d[7*n_src] = _mm_extract_epi16(out[0], 7); + d += 8*n_src; } for(; n_samples--; n++) { in[0] = _mm_mul_ss(_mm_load_ss(&s0[n]), int_max); @@ -350,7 +372,7 @@ conv_f32d_to_s16_1_sse(void *data, void *dst, int n_src, const void *src[n_src], } static void -conv_f32d_to_s16_2_sse(void *data, void *dst, int n_src, const void *src[n_src], int n_samples) +conv_f32d_to_s16_2_sse2(void *data, void *dst, int n_src, const void *src[n_src], int n_samples) { const float **s = (const float **) src; const float *s0 = s[0], *s1 = s[1]; @@ -397,13 +419,13 @@ conv_f32d_to_s16_2_sse(void *data, void *dst, int n_src, const void *src[n_src], } static void -conv_f32d_to_s16_sse(void *data, int n_dst, void *dst[n_dst], int n_src, const void *src[n_src], int n_samples) +conv_f32d_to_s16_sse2(void *data, int n_dst, void *dst[n_dst], int n_src, const void *src[n_src], int n_samples) { int16_t *d = dst[0]; int i = 0; for(; i + 1 < n_src; i += 2) - conv_f32d_to_s16_2_sse(data, &d[i], n_src, &src[i], n_samples); + conv_f32d_to_s16_2_sse2(data, &d[i], n_src, &src[i], n_samples); for(; i < n_src; i++) - conv_f32d_to_s16_1_sse(data, &d[i], n_src, &src[i], n_samples); + conv_f32d_to_s16_1_sse2(data, &d[i], n_src, &src[i], n_samples); } diff --git a/spa/plugins/audioconvert/fmt-ops.c b/spa/plugins/audioconvert/fmt-ops.c index f9b732d35..3d37de978 100644 --- a/spa/plugins/audioconvert/fmt-ops.c +++ b/spa/plugins/audioconvert/fmt-ops.c @@ -64,8 +64,8 @@ static inline int32_t read_s24(const void *src) #define READ24(s) read_s24(s) -#if defined (__SSE__) -#include "fmt-ops-sse.c" +#if defined (__SSE2__) +#include "fmt-ops-sse2.c" #endif static void @@ -678,7 +678,7 @@ typedef void (*convert_func_t) (void *data, int n_dst, void *dst[n_dst], static const struct conv_info { uint32_t src_fmt; uint32_t dst_fmt; -#define FEATURE_SSE SPA_CPU_FLAG_SSE +#define FEATURE_SSE2 SPA_CPU_FLAG_SSE2 uint32_t features; convert_func_t func; @@ -693,8 +693,8 @@ static const struct conv_info { { SPA_AUDIO_FORMAT_S16, SPA_AUDIO_FORMAT_F32, 0, conv_s16_to_f32 }, { SPA_AUDIO_FORMAT_S16P, SPA_AUDIO_FORMAT_F32P, 0, conv_s16_to_f32 }, -#if defined (__SSE__) - { SPA_AUDIO_FORMAT_S16, SPA_AUDIO_FORMAT_F32P, FEATURE_SSE, conv_s16_to_f32d_sse }, +#if defined (__SSE2__) + { SPA_AUDIO_FORMAT_S16, SPA_AUDIO_FORMAT_F32P, FEATURE_SSE2, conv_s16_to_f32d_sse2 }, #endif { SPA_AUDIO_FORMAT_S16, SPA_AUDIO_FORMAT_F32P, 0, conv_s16_to_f32d }, { SPA_AUDIO_FORMAT_S16P, SPA_AUDIO_FORMAT_F32, 0, conv_s16d_to_f32 }, @@ -711,8 +711,8 @@ static const struct conv_info { { SPA_AUDIO_FORMAT_S24, SPA_AUDIO_FORMAT_F32, 0, conv_s24_to_f32 }, { SPA_AUDIO_FORMAT_S24P, SPA_AUDIO_FORMAT_F32P, 0, conv_s24_to_f32 }, -#if defined (__SSE__) - { SPA_AUDIO_FORMAT_S24, SPA_AUDIO_FORMAT_F32P, FEATURE_SSE, conv_s24_to_f32d_sse }, +#if defined (__SSE2__) + { SPA_AUDIO_FORMAT_S24, SPA_AUDIO_FORMAT_F32P, FEATURE_SSE2, conv_s24_to_f32d_sse2 }, #endif { SPA_AUDIO_FORMAT_S24, SPA_AUDIO_FORMAT_F32P, 0, conv_s24_to_f32d }, { SPA_AUDIO_FORMAT_S24P, SPA_AUDIO_FORMAT_F32, 0, conv_s24d_to_f32 }, @@ -731,16 +731,16 @@ static const struct conv_info { { SPA_AUDIO_FORMAT_F32, SPA_AUDIO_FORMAT_S16, 0, conv_f32_to_s16 }, { SPA_AUDIO_FORMAT_F32P, SPA_AUDIO_FORMAT_S16P, 0, conv_f32_to_s16 }, { SPA_AUDIO_FORMAT_F32, SPA_AUDIO_FORMAT_S16P, 0, conv_f32_to_s16d }, -#if defined (__SSE__) - { SPA_AUDIO_FORMAT_F32P, SPA_AUDIO_FORMAT_S16, FEATURE_SSE, conv_f32d_to_s16_sse }, +#if defined (__SSE2__) + { SPA_AUDIO_FORMAT_F32P, SPA_AUDIO_FORMAT_S16, FEATURE_SSE2, conv_f32d_to_s16_sse2 }, #endif { SPA_AUDIO_FORMAT_F32P, SPA_AUDIO_FORMAT_S16, 0, conv_f32d_to_s16 }, { SPA_AUDIO_FORMAT_F32, SPA_AUDIO_FORMAT_S32, 0, conv_f32_to_s32 }, { SPA_AUDIO_FORMAT_F32P, SPA_AUDIO_FORMAT_S32P, 0, conv_f32_to_s32 }, { SPA_AUDIO_FORMAT_F32, SPA_AUDIO_FORMAT_S32P, 0, conv_f32_to_s32d }, -#if defined (__SSE__) - { SPA_AUDIO_FORMAT_F32P, SPA_AUDIO_FORMAT_S32, FEATURE_SSE, conv_f32d_to_s32_sse }, +#if defined (__SSE2__) + { SPA_AUDIO_FORMAT_F32P, SPA_AUDIO_FORMAT_S32, FEATURE_SSE2, conv_f32d_to_s32_sse2 }, #endif { SPA_AUDIO_FORMAT_F32P, SPA_AUDIO_FORMAT_S32, 0, conv_f32d_to_s32 },