From 22317da68564647634c8f999281560efe69dfa25 Mon Sep 17 00:00:00 2001 From: Wim Taymans Date: Tue, 5 Jul 2022 12:20:02 +0200 Subject: [PATCH] audioconvert: tweak conversion constants Tweak the conversion constants a bit so that they handle the extreme ranges a bit better. Align the C and vector instructions. Reactivate the unit test asserts when a conversion fails. --- spa/plugins/audioconvert/fmt-ops-avx2.c | 178 ++++++++++++------------ spa/plugins/audioconvert/fmt-ops-sse2.c | 128 ++++++++--------- spa/plugins/audioconvert/fmt-ops.h | 28 ++-- spa/plugins/audioconvert/test-fmt-ops.c | 130 +++++++++-------- 4 files changed, 237 insertions(+), 227 deletions(-) diff --git a/spa/plugins/audioconvert/fmt-ops-avx2.c b/spa/plugins/audioconvert/fmt-ops-avx2.c index a7505a131..705de7486 100644 --- a/spa/plugins/audioconvert/fmt-ops-avx2.c +++ b/spa/plugins/audioconvert/fmt-ops-avx2.c @@ -353,7 +353,7 @@ conv_s32_to_f32d_4s_avx2(void *data, void * SPA_RESTRICT dst[], const void * SPA float *d0 = dst[0], *d1 = dst[1], *d2 = dst[2], *d3 = dst[3]; uint32_t n, unrolled; __m256i in[4], t[4]; - __m256 out[4], factor = _mm256_set1_ps(1.0f / S24_SCALE); + __m256 out[4], factor = _mm256_set1_ps(1.0f / S32_SCALE); __m256i mask1 = _mm256_setr_epi64x(0*n_channels, 0*n_channels+2, 4*n_channels, 4*n_channels+2); __m256i mask2 = _mm256_setr_epi64x(1*n_channels, 1*n_channels+2, 5*n_channels, 5*n_channels+2); __m256i mask3 = _mm256_setr_epi64x(2*n_channels, 2*n_channels+2, 6*n_channels, 6*n_channels+2); @@ -373,11 +373,6 @@ conv_s32_to_f32d_4s_avx2(void *data, void * SPA_RESTRICT dst[], const void * SPA in[2] = _mm256_i64gather_epi64((long long int *)&s[0*n_channels], mask3, 4); in[3] = _mm256_i64gather_epi64((long long int *)&s[0*n_channels], mask4, 4); - in[0] = _mm256_srai_epi32(in[0], 8); /* a0 b0 c0 d0 a4 b4 c4 d4 */ - in[1] = _mm256_srai_epi32(in[1], 8); /* a1 b1 c1 d1 a5 b5 c5 d5 */ - in[2] = _mm256_srai_epi32(in[2], 8); /* a2 b2 c2 d2 a6 b6 c6 d6 */ - in[3] = _mm256_srai_epi32(in[3], 8); /* a3 b3 c3 d3 a7 b7 c7 d7 */ - t[0] = _mm256_unpacklo_epi32(in[0], in[1]); /* a0 a1 b0 b1 a4 a5 b4 b5 */ t[1] = _mm256_unpackhi_epi32(in[0], in[1]); /* c0 c1 d0 d1 c4 c5 d4 d5 */ t[2] = _mm256_unpacklo_epi32(in[2], in[3]); /* a2 a3 b2 b3 a6 a7 b6 b7 */ @@ -405,11 +400,11 @@ conv_s32_to_f32d_4s_avx2(void *data, void * SPA_RESTRICT dst[], const void * SPA s += 8*n_channels; } for(; n < n_samples; n++) { - __m128 out[4], factor = _mm_set1_ps(1.0f / S24_SCALE); - out[0] = _mm_cvtsi32_ss(factor, s[0]>>8); - out[1] = _mm_cvtsi32_ss(factor, s[1]>>8); - out[2] = _mm_cvtsi32_ss(factor, s[2]>>8); - out[3] = _mm_cvtsi32_ss(factor, s[3]>>8); + __m128 out[4], factor = _mm_set1_ps(1.0f / S32_SCALE); + out[0] = _mm_cvtsi32_ss(factor, s[0]); + out[1] = _mm_cvtsi32_ss(factor, s[1]); + out[2] = _mm_cvtsi32_ss(factor, s[2]); + out[3] = _mm_cvtsi32_ss(factor, s[3]); out[0] = _mm_mul_ss(out[0], factor); out[1] = _mm_mul_ss(out[1], factor); out[2] = _mm_mul_ss(out[2], factor); @@ -430,7 +425,7 @@ conv_s32_to_f32d_2s_avx2(void *data, void * SPA_RESTRICT dst[], const void * SPA float *d0 = dst[0], *d1 = dst[1]; uint32_t n, unrolled; __m256i in[4], t[4]; - __m256 out[4], factor = _mm256_set1_ps(1.0f / S24_SCALE); + __m256 out[4], factor = _mm256_set1_ps(1.0f / S32_SCALE); __m256i perm = _mm256_setr_epi32(0, 2, 4, 6, 1, 3, 5, 7); __m256i mask1 = _mm256_setr_epi64x(0*n_channels, 1*n_channels, 2*n_channels, 3*n_channels); __m256i mask2 = _mm256_setr_epi64x(4*n_channels, 5*n_channels, 6*n_channels, 7*n_channels); @@ -445,9 +440,6 @@ conv_s32_to_f32d_2s_avx2(void *data, void * SPA_RESTRICT dst[], const void * SPA in[0] = _mm256_i64gather_epi64((long long int *)s, mask1, 4); in[1] = _mm256_i64gather_epi64((long long int *)s, mask2, 4); - in[0] = _mm256_srai_epi32(in[0], 8); - in[1] = _mm256_srai_epi32(in[1], 8); - t[0] = _mm256_permutevar8x32_epi32(in[0], perm); t[1] = _mm256_permutevar8x32_epi32(in[1], perm); @@ -466,9 +458,9 @@ conv_s32_to_f32d_2s_avx2(void *data, void * SPA_RESTRICT dst[], const void * SPA s += 8*n_channels; } for(; n < n_samples; n++) { - __m128 out[2], factor = _mm_set1_ps(1.0f / S24_SCALE); - out[0] = _mm_cvtsi32_ss(factor, s[0]>>8); - out[1] = _mm_cvtsi32_ss(factor, s[1]>>8); + __m128 out[2], factor = _mm_set1_ps(1.0f / S32_SCALE); + out[0] = _mm_cvtsi32_ss(factor, s[0]); + out[1] = _mm_cvtsi32_ss(factor, s[1]); out[0] = _mm_mul_ss(out[0], factor); out[1] = _mm_mul_ss(out[1], factor); _mm_store_ss(&d0[n], out[0]); @@ -485,7 +477,7 @@ conv_s32_to_f32d_1s_avx2(void *data, void * SPA_RESTRICT dst[], const void * SPA float *d0 = dst[0]; uint32_t n, unrolled; __m256i in[2]; - __m256 out[2], factor = _mm256_set1_ps(1.0f / S24_SCALE); + __m256 out[2], factor = _mm256_set1_ps(1.0f / S32_SCALE); __m256i mask1 = _mm256_setr_epi64x(0*n_channels, 1*n_channels, 2*n_channels, 3*n_channels); __m256i mask2 = _mm256_setr_epi64x(4*n_channels, 5*n_channels, 6*n_channels, 7*n_channels); @@ -502,9 +494,6 @@ conv_s32_to_f32d_1s_avx2(void *data, void * SPA_RESTRICT dst[], const void * SPA _mm256_i64gather_epi32(&s[ 8*n_channels], mask1, 4), _mm256_i64gather_epi32(&s[ 8*n_channels], mask2, 4)); - in[0] = _mm256_srai_epi32(in[0], 8); - in[1] = _mm256_srai_epi32(in[1], 8); - out[0] = _mm256_cvtepi32_ps(in[0]); out[1] = _mm256_cvtepi32_ps(in[1]); @@ -517,8 +506,8 @@ conv_s32_to_f32d_1s_avx2(void *data, void * SPA_RESTRICT dst[], const void * SPA s += 16*n_channels; } for(; n < n_samples; n++) { - __m128 out, factor = _mm_set1_ps(1.0f / S24_SCALE); - out = _mm_cvtsi32_ss(factor, s[0]>>8); + __m128 out, factor = _mm_set1_ps(1.0f / S32_SCALE); + out = _mm_cvtsi32_ss(factor, s[0]); out = _mm_mul_ss(out, factor); _mm_store_ss(&d0[n], out); s += n_channels; @@ -560,7 +549,7 @@ conv_f32d_to_s32_1s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R for(n = 0; n < unrolled; n += 4) { in[0] = _mm_mul_ps(_mm_load_ps(&s0[n]), scale); in[0] = _mm_min_ps(in[0], int_max); - out[0] = _mm_cvtps_epi32(in[0]); + out[0] = _mm_cvttps_epi32(in[0]); out[1] = _mm_shuffle_epi32(out[0], _MM_SHUFFLE(0, 3, 2, 1)); out[2] = _mm_shuffle_epi32(out[0], _MM_SHUFFLE(1, 0, 3, 2)); out[3] = _mm_shuffle_epi32(out[0], _MM_SHUFFLE(2, 1, 0, 3)); @@ -605,8 +594,8 @@ conv_f32d_to_s32_2s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R in[0] = _mm256_min_ps(in[0], int_max); in[1] = _mm256_min_ps(in[1], int_max); - out[0] = _mm256_cvtps_epi32(in[0]); /* a0 a1 a2 a3 a4 a5 a6 a7 */ - out[1] = _mm256_cvtps_epi32(in[1]); /* b0 b1 b2 b3 b4 b5 b6 b7 */ + out[0] = _mm256_cvttps_epi32(in[0]); /* a0 a1 a2 a3 a4 a5 a6 a7 */ + out[1] = _mm256_cvttps_epi32(in[1]); /* b0 b1 b2 b3 b4 b5 b6 b7 */ t[0] = _mm256_unpacklo_epi32(out[0], out[1]); /* a0 b0 a1 b1 a4 b4 a5 b5 */ t[1] = _mm256_unpackhi_epi32(out[0], out[1]); /* a2 b2 a3 b3 a6 b6 a7 b7 */ @@ -645,7 +634,7 @@ conv_f32d_to_s32_2s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R in[0] = _mm_mul_ps(in[0], scale); in[0] = _mm_min_ps(in[0], int_max); - out[0] = _mm_cvtps_epi32(in[0]); + out[0] = _mm_cvttps_epi32(in[0]); _mm_storel_epi64((__m128i*)d, out[0]); d += n_channels; } @@ -682,10 +671,10 @@ conv_f32d_to_s32_4s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R in[2] = _mm256_min_ps(in[2], int_max); in[3] = _mm256_min_ps(in[3], int_max); - out[0] = _mm256_cvtps_epi32(in[0]); /* a0 a1 a2 a3 a4 a5 a6 a7 */ - out[1] = _mm256_cvtps_epi32(in[1]); /* b0 b1 b2 b3 b4 b5 b6 b7 */ - out[2] = _mm256_cvtps_epi32(in[2]); /* c0 c1 c2 c3 c4 c5 c6 c7 */ - out[3] = _mm256_cvtps_epi32(in[3]); /* d0 d1 d2 d3 d4 d5 d6 d7 */ + out[0] = _mm256_cvttps_epi32(in[0]); /* a0 a1 a2 a3 a4 a5 a6 a7 */ + out[1] = _mm256_cvttps_epi32(in[1]); /* b0 b1 b2 b3 b4 b5 b6 b7 */ + out[2] = _mm256_cvttps_epi32(in[2]); /* c0 c1 c2 c3 c4 c5 c6 c7 */ + out[3] = _mm256_cvttps_epi32(in[3]); /* d0 d1 d2 d3 d4 d5 d6 d7 */ t[0] = _mm256_unpacklo_epi32(out[0], out[1]); /* a0 b0 a1 b1 a4 b4 a5 b5 */ t[1] = _mm256_unpackhi_epi32(out[0], out[1]); /* a2 b2 a3 b3 a6 b6 a7 b7 */ @@ -724,7 +713,7 @@ conv_f32d_to_s32_4s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R in[0] = _mm_mul_ps(in[0], scale); in[0] = _mm_min_ps(in[0], int_max); - out[0] = _mm_cvtps_epi32(in[0]); + out[0] = _mm_cvttps_epi32(in[0]); _mm_storeu_si128((__m128i*)d, out[0]); d += n_channels; } @@ -754,8 +743,9 @@ conv_f32d_to_s16_1s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R uint32_t n, unrolled; __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); + __m128 int_scale = _mm_set1_ps(S16_SCALE); + __m128 int_max = _mm_set1_ps(S16_MAX); + __m128 int_min = _mm_set1_ps(S16_MIN); if (SPA_IS_ALIGNED(s0, 16)) unrolled = n_samples & ~7; @@ -763,10 +753,10 @@ conv_f32d_to_s16_1s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R unrolled = 0; for(n = 0; n < unrolled; n += 8) { - in[0] = _mm_mul_ps(_mm_load_ps(&s0[n]), int_max); - in[1] = _mm_mul_ps(_mm_load_ps(&s0[n+4]), int_max); - out[0] = _mm_cvtps_epi32(in[0]); - out[1] = _mm_cvtps_epi32(in[1]); + in[0] = _mm_mul_ps(_mm_load_ps(&s0[n]), int_scale); + in[1] = _mm_mul_ps(_mm_load_ps(&s0[n+4]), int_scale); + out[0] = _mm_cvttps_epi32(in[0]); + out[1] = _mm_cvttps_epi32(in[1]); out[0] = _mm_packs_epi32(out[0], out[1]); d[0*n_channels] = _mm_extract_epi16(out[0], 0); @@ -780,7 +770,7 @@ conv_f32d_to_s16_1s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R d += 8*n_channels; } for(; n < n_samples; n++) { - in[0] = _mm_mul_ss(_mm_load_ss(&s0[n]), int_max); + in[0] = _mm_mul_ss(_mm_load_ss(&s0[n]), int_scale); in[0] = _mm_min_ss(int_max, _mm_max_ss(in[0], int_min)); *d = _mm_cvtss_si32(in[0]); d += n_channels; @@ -796,7 +786,7 @@ conv_f32d_to_s16_2s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R uint32_t n, unrolled; __m256 in[2]; __m256i out[4], t[2]; - __m256 int_max = _mm256_set1_ps(S16_MAX_F); + __m256 int_scale = _mm256_set1_ps(S16_SCALE); if (SPA_IS_ALIGNED(s0, 32) && SPA_IS_ALIGNED(s1, 32)) @@ -805,11 +795,11 @@ conv_f32d_to_s16_2s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R unrolled = 0; for(n = 0; n < unrolled; n += 8) { - in[0] = _mm256_mul_ps(_mm256_load_ps(&s0[n+0]), int_max); - in[1] = _mm256_mul_ps(_mm256_load_ps(&s1[n+0]), int_max); + in[0] = _mm256_mul_ps(_mm256_load_ps(&s0[n+0]), int_scale); + in[1] = _mm256_mul_ps(_mm256_load_ps(&s1[n+0]), int_scale); - out[0] = _mm256_cvtps_epi32(in[0]); /* a0 a1 a2 a3 a4 a5 a6 a7 */ - out[1] = _mm256_cvtps_epi32(in[1]); /* b0 b1 b2 b3 b4 b5 b6 b7 */ + out[0] = _mm256_cvttps_epi32(in[0]); /* a0 a1 a2 a3 a4 a5 a6 a7 */ + out[1] = _mm256_cvttps_epi32(in[1]); /* b0 b1 b2 b3 b4 b5 b6 b7 */ t[0] = _mm256_unpacklo_epi32(out[0], out[1]); /* a0 b0 a1 b1 a4 b4 a5 b5 */ t[1] = _mm256_unpackhi_epi32(out[0], out[1]); /* a2 b2 a3 b3 a6 b6 a7 b7 */ @@ -829,11 +819,12 @@ conv_f32d_to_s16_2s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R } for(; n < n_samples; n++) { __m128 in[2]; - __m128 int_max = _mm_set1_ps(S16_MAX_F); - __m128 int_min = _mm_sub_ps(_mm_setzero_ps(), int_max); + __m128 int_scale = _mm_set1_ps(S16_SCALE); + __m128 int_max = _mm_set1_ps(S16_MAX); + __m128 int_min = _mm_set1_ps(S16_MIN); - in[0] = _mm_mul_ss(_mm_load_ss(&s0[n]), int_max); - in[1] = _mm_mul_ss(_mm_load_ss(&s1[n]), int_max); + in[0] = _mm_mul_ss(_mm_load_ss(&s0[n]), int_scale); + in[1] = _mm_mul_ss(_mm_load_ss(&s1[n]), int_scale); in[0] = _mm_min_ss(int_max, _mm_max_ss(in[0], int_min)); in[1] = _mm_min_ss(int_max, _mm_max_ss(in[1], int_min)); d[0] = _mm_cvtss_si32(in[0]); @@ -851,7 +842,7 @@ conv_f32d_to_s16_4s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R uint32_t n, unrolled; __m256 in[4]; __m256i out[4], t[4]; - __m256 int_max = _mm256_set1_ps(S16_MAX_F); + __m256 int_scale = _mm256_set1_ps(S16_SCALE); if (SPA_IS_ALIGNED(s0, 32) && SPA_IS_ALIGNED(s1, 32) && @@ -862,15 +853,15 @@ conv_f32d_to_s16_4s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R unrolled = 0; for(n = 0; n < unrolled; n += 8) { - in[0] = _mm256_mul_ps(_mm256_load_ps(&s0[n]), int_max); - in[1] = _mm256_mul_ps(_mm256_load_ps(&s1[n]), int_max); - in[2] = _mm256_mul_ps(_mm256_load_ps(&s2[n]), int_max); - in[3] = _mm256_mul_ps(_mm256_load_ps(&s3[n]), int_max); + in[0] = _mm256_mul_ps(_mm256_load_ps(&s0[n]), int_scale); + in[1] = _mm256_mul_ps(_mm256_load_ps(&s1[n]), int_scale); + in[2] = _mm256_mul_ps(_mm256_load_ps(&s2[n]), int_scale); + in[3] = _mm256_mul_ps(_mm256_load_ps(&s3[n]), int_scale); - t[0] = _mm256_cvtps_epi32(in[0]); /* a0 a1 a2 a3 a4 a5 a6 a7 */ - t[1] = _mm256_cvtps_epi32(in[1]); /* b0 b1 b2 b3 b4 b5 b6 b7 */ - t[2] = _mm256_cvtps_epi32(in[2]); /* c0 c1 c2 c3 c4 c5 c6 c7 */ - t[3] = _mm256_cvtps_epi32(in[3]); /* d0 d1 d2 d3 d4 d5 d6 d7 */ + t[0] = _mm256_cvttps_epi32(in[0]); /* a0 a1 a2 a3 a4 a5 a6 a7 */ + t[1] = _mm256_cvttps_epi32(in[1]); /* b0 b1 b2 b3 b4 b5 b6 b7 */ + t[2] = _mm256_cvttps_epi32(in[2]); /* c0 c1 c2 c3 c4 c5 c6 c7 */ + t[3] = _mm256_cvttps_epi32(in[3]); /* d0 d1 d2 d3 d4 d5 d6 d7 */ t[0] = _mm256_packs_epi32(t[0], t[2]); /* a0 a1 a2 a3 c0 c1 c2 c3 a4 a5 a6 a7 c4 c5 c6 c7 */ t[1] = _mm256_packs_epi32(t[1], t[3]); /* b0 b1 b2 b3 d0 d1 d2 d3 b4 b5 b6 b7 d4 d5 d6 d7 */ @@ -905,13 +896,14 @@ conv_f32d_to_s16_4s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R } for(; n < n_samples; n++) { __m128 in[4]; - __m128 int_max = _mm_set1_ps(S16_MAX_F); - __m128 int_min = _mm_sub_ps(_mm_setzero_ps(), int_max); + __m128 int_scale = _mm_set1_ps(S16_SCALE); + __m128 int_max = _mm_set1_ps(S16_MAX); + __m128 int_min = _mm_set1_ps(S16_MIN); - in[0] = _mm_mul_ss(_mm_load_ss(&s0[n]), int_max); - in[1] = _mm_mul_ss(_mm_load_ss(&s1[n]), int_max); - in[2] = _mm_mul_ss(_mm_load_ss(&s2[n]), int_max); - in[3] = _mm_mul_ss(_mm_load_ss(&s3[n]), int_max); + in[0] = _mm_mul_ss(_mm_load_ss(&s0[n]), int_scale); + in[1] = _mm_mul_ss(_mm_load_ss(&s1[n]), int_scale); + in[2] = _mm_mul_ss(_mm_load_ss(&s2[n]), int_scale); + in[3] = _mm_mul_ss(_mm_load_ss(&s3[n]), int_scale); in[0] = _mm_min_ss(int_max, _mm_max_ss(in[0], int_min)); in[1] = _mm_min_ss(int_max, _mm_max_ss(in[1], int_min)); in[2] = _mm_min_ss(int_max, _mm_max_ss(in[2], int_min)); @@ -948,7 +940,7 @@ conv_f32d_to_s16_4_avx2(struct convert *conv, void * SPA_RESTRICT dst[], const v uint32_t n, unrolled; __m256 in[4]; __m256i out[4], t[4]; - __m256 int_max = _mm256_set1_ps(S16_MAX_F); + __m256 int_scale = _mm256_set1_ps(S16_SCALE); if (SPA_IS_ALIGNED(s0, 32) && SPA_IS_ALIGNED(s1, 32) && @@ -959,15 +951,15 @@ conv_f32d_to_s16_4_avx2(struct convert *conv, void * SPA_RESTRICT dst[], const v unrolled = 0; for(n = 0; n < unrolled; n += 8) { - in[0] = _mm256_mul_ps(_mm256_load_ps(&s0[n]), int_max); - in[1] = _mm256_mul_ps(_mm256_load_ps(&s1[n]), int_max); - in[2] = _mm256_mul_ps(_mm256_load_ps(&s2[n]), int_max); - in[3] = _mm256_mul_ps(_mm256_load_ps(&s3[n]), int_max); + in[0] = _mm256_mul_ps(_mm256_load_ps(&s0[n]), int_scale); + in[1] = _mm256_mul_ps(_mm256_load_ps(&s1[n]), int_scale); + in[2] = _mm256_mul_ps(_mm256_load_ps(&s2[n]), int_scale); + in[3] = _mm256_mul_ps(_mm256_load_ps(&s3[n]), int_scale); - t[0] = _mm256_cvtps_epi32(in[0]); /* a0 a1 a2 a3 a4 a5 a6 a7 */ - t[1] = _mm256_cvtps_epi32(in[1]); /* b0 b1 b2 b3 b4 b5 b6 b7 */ - t[2] = _mm256_cvtps_epi32(in[2]); /* c0 c1 c2 c3 c4 c5 c6 c7 */ - t[3] = _mm256_cvtps_epi32(in[3]); /* d0 d1 d2 d3 d4 d5 d6 d7 */ + t[0] = _mm256_cvttps_epi32(in[0]); /* a0 a1 a2 a3 a4 a5 a6 a7 */ + t[1] = _mm256_cvttps_epi32(in[1]); /* b0 b1 b2 b3 b4 b5 b6 b7 */ + t[2] = _mm256_cvttps_epi32(in[2]); /* c0 c1 c2 c3 c4 c5 c6 c7 */ + t[3] = _mm256_cvttps_epi32(in[3]); /* d0 d1 d2 d3 d4 d5 d6 d7 */ t[0] = _mm256_packs_epi32(t[0], t[2]); /* a0 a1 a2 a3 c0 c1 c2 c3 a4 a5 a6 a7 c4 c5 c6 c7 */ t[1] = _mm256_packs_epi32(t[1], t[3]); /* b0 b1 b2 b3 d0 d1 d2 d3 b4 b5 b6 b7 d4 d5 d6 d7 */ @@ -987,13 +979,14 @@ conv_f32d_to_s16_4_avx2(struct convert *conv, void * SPA_RESTRICT dst[], const v } for(; n < n_samples; n++) { __m128 in[4]; - __m128 int_max = _mm_set1_ps(S16_MAX_F); - __m128 int_min = _mm_sub_ps(_mm_setzero_ps(), int_max); + __m128 int_scale = _mm_set1_ps(S16_SCALE); + __m128 int_max = _mm_set1_ps(S16_MAX); + __m128 int_min = _mm_set1_ps(S16_MIN); - in[0] = _mm_mul_ss(_mm_load_ss(&s0[n]), int_max); - in[1] = _mm_mul_ss(_mm_load_ss(&s1[n]), int_max); - in[2] = _mm_mul_ss(_mm_load_ss(&s2[n]), int_max); - in[3] = _mm_mul_ss(_mm_load_ss(&s3[n]), int_max); + in[0] = _mm_mul_ss(_mm_load_ss(&s0[n]), int_scale); + in[1] = _mm_mul_ss(_mm_load_ss(&s1[n]), int_scale); + in[2] = _mm_mul_ss(_mm_load_ss(&s2[n]), int_scale); + in[3] = _mm_mul_ss(_mm_load_ss(&s3[n]), int_scale); in[0] = _mm_min_ss(int_max, _mm_max_ss(in[0], int_min)); in[1] = _mm_min_ss(int_max, _mm_max_ss(in[1], int_min)); in[2] = _mm_min_ss(int_max, _mm_max_ss(in[2], int_min)); @@ -1014,7 +1007,7 @@ conv_f32d_to_s16_2_avx2(struct convert *conv, void * SPA_RESTRICT dst[], const v uint32_t n, unrolled; __m256 in[4]; __m256i out[4], t[4]; - __m256 int_max = _mm256_set1_ps(S16_MAX_F); + __m256 int_scale = _mm256_set1_ps(S16_SCALE); if (SPA_IS_ALIGNED(s0, 32) && SPA_IS_ALIGNED(s1, 32)) @@ -1023,15 +1016,15 @@ conv_f32d_to_s16_2_avx2(struct convert *conv, void * SPA_RESTRICT dst[], const v unrolled = 0; for(n = 0; n < unrolled; n += 16) { - in[0] = _mm256_mul_ps(_mm256_load_ps(&s0[n+0]), int_max); - in[1] = _mm256_mul_ps(_mm256_load_ps(&s1[n+0]), int_max); - in[2] = _mm256_mul_ps(_mm256_load_ps(&s0[n+8]), int_max); - in[3] = _mm256_mul_ps(_mm256_load_ps(&s1[n+8]), int_max); + in[0] = _mm256_mul_ps(_mm256_load_ps(&s0[n+0]), int_scale); + in[1] = _mm256_mul_ps(_mm256_load_ps(&s1[n+0]), int_scale); + in[2] = _mm256_mul_ps(_mm256_load_ps(&s0[n+8]), int_scale); + in[3] = _mm256_mul_ps(_mm256_load_ps(&s1[n+8]), int_scale); - out[0] = _mm256_cvtps_epi32(in[0]); /* a0 a1 a2 a3 a4 a5 a6 a7 */ - out[1] = _mm256_cvtps_epi32(in[1]); /* b0 b1 b2 b3 b4 b5 b6 b7 */ - out[2] = _mm256_cvtps_epi32(in[2]); /* a0 a1 a2 a3 a4 a5 a6 a7 */ - out[3] = _mm256_cvtps_epi32(in[3]); /* b0 b1 b2 b3 b4 b5 b6 b7 */ + out[0] = _mm256_cvttps_epi32(in[0]); /* a0 a1 a2 a3 a4 a5 a6 a7 */ + out[1] = _mm256_cvttps_epi32(in[1]); /* b0 b1 b2 b3 b4 b5 b6 b7 */ + out[2] = _mm256_cvttps_epi32(in[2]); /* a0 a1 a2 a3 a4 a5 a6 a7 */ + out[3] = _mm256_cvttps_epi32(in[3]); /* b0 b1 b2 b3 b4 b5 b6 b7 */ t[0] = _mm256_unpacklo_epi32(out[0], out[1]); /* a0 b0 a1 b1 a4 b4 a5 b5 */ t[1] = _mm256_unpackhi_epi32(out[0], out[1]); /* a2 b2 a3 b3 a6 b6 a7 b7 */ @@ -1048,11 +1041,12 @@ conv_f32d_to_s16_2_avx2(struct convert *conv, void * SPA_RESTRICT dst[], const v } for(; n < n_samples; n++) { __m128 in[4]; - __m128 int_max = _mm_set1_ps(S16_MAX_F); - __m128 int_min = _mm_sub_ps(_mm_setzero_ps(), int_max); + __m128 int_scale = _mm_set1_ps(S16_SCALE); + __m128 int_max = _mm_set1_ps(S16_MAX); + __m128 int_min = _mm_set1_ps(S16_MIN); - in[0] = _mm_mul_ss(_mm_load_ss(&s0[n]), int_max); - in[1] = _mm_mul_ss(_mm_load_ss(&s1[n]), int_max); + in[0] = _mm_mul_ss(_mm_load_ss(&s0[n]), int_scale); + in[1] = _mm_mul_ss(_mm_load_ss(&s1[n]), int_scale); in[0] = _mm_min_ss(int_max, _mm_max_ss(in[0], int_min)); in[1] = _mm_min_ss(int_max, _mm_max_ss(in[1], int_min)); d[0] = _mm_cvtss_si32(in[0]); diff --git a/spa/plugins/audioconvert/fmt-ops-sse2.c b/spa/plugins/audioconvert/fmt-ops-sse2.c index 0804580bd..e455acf99 100644 --- a/spa/plugins/audioconvert/fmt-ops-sse2.c +++ b/spa/plugins/audioconvert/fmt-ops-sse2.c @@ -338,7 +338,7 @@ conv_s32_to_f32d_1s_sse2(void *data, void * SPA_RESTRICT dst[], const void * SPA float *d0 = dst[0]; uint32_t n, unrolled; __m128i in; - __m128 out, factor = _mm_set1_ps(1.0f / S24_SCALE); + __m128 out, factor = _mm_set1_ps(1.0f / S32_SCALE); if (SPA_IS_ALIGNED(d0, 16)) unrolled = n_samples & ~3; @@ -350,14 +350,13 @@ conv_s32_to_f32d_1s_sse2(void *data, void * SPA_RESTRICT dst[], const void * SPA s[1*n_channels], s[2*n_channels], s[3*n_channels]); - in = _mm_srai_epi32(in, 8); out = _mm_cvtepi32_ps(in); out = _mm_mul_ps(out, factor); _mm_store_ps(&d0[n], out); s += 4*n_channels; } for(; n < n_samples; n++) { - out = _mm_cvtsi32_ss(factor, s[0]>>8); + out = _mm_cvtsi32_ss(factor, s[0]); out = _mm_mul_ss(out, factor); _mm_store_ss(&d0[n], out); s += n_channels; @@ -395,7 +394,7 @@ conv_f32d_to_s32_1s_sse2(void *data, void * SPA_RESTRICT dst, const void * SPA_R for(n = 0; n < unrolled; n += 4) { in[0] = _mm_mul_ps(_mm_load_ps(&s0[n]), scale); in[0] = _mm_min_ps(in[0], int_max); - out[0] = _mm_cvtps_epi32(in[0]); + out[0] = _mm_cvttps_epi32(in[0]); out[1] = _mm_shuffle_epi32(out[0], _MM_SHUFFLE(0, 3, 2, 1)); out[2] = _mm_shuffle_epi32(out[0], _MM_SHUFFLE(1, 0, 3, 2)); out[3] = _mm_shuffle_epi32(out[0], _MM_SHUFFLE(2, 1, 0, 3)); @@ -440,8 +439,8 @@ conv_f32d_to_s32_2s_sse2(void *data, void * SPA_RESTRICT dst, const void * SPA_R in[0] = _mm_min_ps(in[0], int_max); in[1] = _mm_min_ps(in[1], int_max); - out[0] = _mm_cvtps_epi32(in[0]); - out[1] = _mm_cvtps_epi32(in[1]); + out[0] = _mm_cvttps_epi32(in[0]); + out[1] = _mm_cvttps_epi32(in[1]); t[0] = _mm_unpacklo_epi32(out[0], out[1]); t[1] = _mm_unpackhi_epi32(out[0], out[1]); @@ -460,7 +459,7 @@ conv_f32d_to_s32_2s_sse2(void *data, void * SPA_RESTRICT dst, const void * SPA_R in[0] = _mm_mul_ps(in[0], scale); in[0] = _mm_min_ps(in[0], int_max); - out[0] = _mm_cvtps_epi32(in[0]); + out[0] = _mm_cvttps_epi32(in[0]); _mm_storel_epi64((__m128i*)d, out[0]); d += n_channels; } @@ -499,10 +498,10 @@ conv_f32d_to_s32_4s_sse2(void *data, void * SPA_RESTRICT dst, const void * SPA_R _MM_TRANSPOSE4_PS(in[0], in[1], in[2], in[3]); - out[0] = _mm_cvtps_epi32(in[0]); - out[1] = _mm_cvtps_epi32(in[1]); - out[2] = _mm_cvtps_epi32(in[2]); - out[3] = _mm_cvtps_epi32(in[3]); + out[0] = _mm_cvttps_epi32(in[0]); + out[1] = _mm_cvttps_epi32(in[1]); + out[2] = _mm_cvttps_epi32(in[2]); + out[3] = _mm_cvttps_epi32(in[3]); _mm_storeu_si128((__m128i*)(d + 0*n_channels), out[0]); _mm_storeu_si128((__m128i*)(d + 1*n_channels), out[1]); @@ -522,7 +521,7 @@ conv_f32d_to_s32_4s_sse2(void *data, void * SPA_RESTRICT dst, const void * SPA_R in[0] = _mm_mul_ps(in[0], scale); in[0] = _mm_min_ps(in[0], int_max); - out[0] = _mm_cvtps_epi32(in[0]); + out[0] = _mm_cvttps_epi32(in[0]); _mm_storeu_si128((__m128i*)d, out[0]); d += n_channels; } @@ -590,7 +589,7 @@ conv_f32d_to_s32_1s_dither_sse2(struct convert *conv, void * SPA_RESTRICT dst, c in[0] = _mm_mul_ps(_mm_load_ps(&s[n]), scale); in[0] = _mm_add_ps(in[0], _mm_load_ps(&dither[n])); in[0] = _mm_min_ps(in[0], int_max); - out[0] = _mm_cvtps_epi32(in[0]); + out[0] = _mm_cvttps_epi32(in[0]); out[1] = _mm_shuffle_epi32(out[0], _MM_SHUFFLE(0, 3, 2, 1)); out[2] = _mm_shuffle_epi32(out[0], _MM_SHUFFLE(1, 0, 3, 2)); out[3] = _mm_shuffle_epi32(out[0], _MM_SHUFFLE(2, 1, 0, 3)); @@ -986,8 +985,9 @@ conv_f32_to_s16_1_sse2(void *data, void * SPA_RESTRICT dst, const void * SPA_RES uint32_t n, unrolled; __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); + __m128 int_scale = _mm_set1_ps(S16_SCALE); + __m128 int_max = _mm_set1_ps(S16_MAX); + __m128 int_min = _mm_set1_ps(S16_MIN); if (SPA_IS_ALIGNED(s, 16)) unrolled = n_samples & ~7; @@ -995,16 +995,16 @@ conv_f32_to_s16_1_sse2(void *data, void * SPA_RESTRICT dst, const void * SPA_RES unrolled = 0; for(n = 0; n < unrolled; n += 8) { - in[0] = _mm_mul_ps(_mm_load_ps(&s[n]), int_max); - in[1] = _mm_mul_ps(_mm_load_ps(&s[n+4]), int_max); - out[0] = _mm_cvtps_epi32(in[0]); - out[1] = _mm_cvtps_epi32(in[1]); + in[0] = _mm_mul_ps(_mm_load_ps(&s[n]), int_scale); + in[1] = _mm_mul_ps(_mm_load_ps(&s[n+4]), int_scale); + out[0] = _mm_cvttps_epi32(in[0]); + out[1] = _mm_cvttps_epi32(in[1]); out[0] = _mm_packs_epi32(out[0], out[1]); _mm_storeu_si128((__m128i*)(d+0), out[0]); d += 8; } for(; n < n_samples; n++) { - in[0] = _mm_mul_ss(_mm_load_ss(&s[n]), int_max); + in[0] = _mm_mul_ss(_mm_load_ss(&s[n]), int_scale); in[0] = _mm_min_ss(int_max, _mm_max_ss(in[0], int_min)); *d++ = _mm_cvtss_si32(in[0]); } @@ -1035,8 +1035,9 @@ conv_f32d_to_s16_1s_sse2(void *data, void * SPA_RESTRICT dst, const void * SPA_R uint32_t n, unrolled; __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); + __m128 int_scale = _mm_set1_ps(S16_SCALE); + __m128 int_max = _mm_set1_ps(S16_MAX); + __m128 int_min = _mm_set1_ps(S16_MIN); if (SPA_IS_ALIGNED(s0, 16)) unrolled = n_samples & ~7; @@ -1044,10 +1045,10 @@ conv_f32d_to_s16_1s_sse2(void *data, void * SPA_RESTRICT dst, const void * SPA_R unrolled = 0; for(n = 0; n < unrolled; n += 8) { - in[0] = _mm_mul_ps(_mm_load_ps(&s0[n]), int_max); - in[1] = _mm_mul_ps(_mm_load_ps(&s0[n+4]), int_max); - out[0] = _mm_cvtps_epi32(in[0]); - out[1] = _mm_cvtps_epi32(in[1]); + in[0] = _mm_mul_ps(_mm_load_ps(&s0[n]), int_scale); + in[1] = _mm_mul_ps(_mm_load_ps(&s0[n+4]), int_scale); + out[0] = _mm_cvttps_epi32(in[0]); + out[1] = _mm_cvttps_epi32(in[1]); out[0] = _mm_packs_epi32(out[0], out[1]); d[0*n_channels] = _mm_extract_epi16(out[0], 0); @@ -1061,9 +1062,9 @@ conv_f32d_to_s16_1s_sse2(void *data, void * SPA_RESTRICT dst, const void * SPA_R d += 8*n_channels; } for(; n < n_samples; n++) { - in[0] = _mm_mul_ss(_mm_load_ss(&s0[n]), int_max); + in[0] = _mm_mul_ss(_mm_load_ss(&s0[n]), int_scale); in[0] = _mm_min_ss(int_max, _mm_max_ss(in[0], int_min)); - *d = _mm_cvtss_si32(in[0]); + *d = _mm_cvttss_si32(in[0]); d += n_channels; } } @@ -1077,8 +1078,9 @@ conv_f32d_to_s16_2s_sse2(void *data, void * SPA_RESTRICT dst, const void * SPA_R uint32_t n, unrolled; __m128 in[2]; __m128i out[4], t[2]; - __m128 int_max = _mm_set1_ps(S16_MAX_F); - __m128 int_min = _mm_sub_ps(_mm_setzero_ps(), int_max); + __m128 int_scale = _mm_set1_ps(S16_SCALE); + __m128 int_max = _mm_set1_ps(S16_MAX); + __m128 int_min = _mm_set1_ps(S16_MIN); if (SPA_IS_ALIGNED(s0, 16) && SPA_IS_ALIGNED(s1, 16)) @@ -1087,11 +1089,11 @@ conv_f32d_to_s16_2s_sse2(void *data, void * SPA_RESTRICT dst, const void * SPA_R unrolled = 0; for(n = 0; n < unrolled; n += 4) { - in[0] = _mm_mul_ps(_mm_load_ps(&s0[n]), int_max); - in[1] = _mm_mul_ps(_mm_load_ps(&s1[n]), int_max); + in[0] = _mm_mul_ps(_mm_load_ps(&s0[n]), int_scale); + in[1] = _mm_mul_ps(_mm_load_ps(&s1[n]), int_scale); - t[0] = _mm_cvtps_epi32(in[0]); - t[1] = _mm_cvtps_epi32(in[1]); + t[0] = _mm_cvttps_epi32(in[0]); + t[1] = _mm_cvttps_epi32(in[1]); t[0] = _mm_packs_epi32(t[0], t[0]); t[1] = _mm_packs_epi32(t[1], t[1]); @@ -1108,8 +1110,8 @@ conv_f32d_to_s16_2s_sse2(void *data, void * SPA_RESTRICT dst, const void * SPA_R d += 4*n_channels; } for(; n < n_samples; n++) { - in[0] = _mm_mul_ss(_mm_load_ss(&s0[n]), int_max); - in[1] = _mm_mul_ss(_mm_load_ss(&s1[n]), int_max); + in[0] = _mm_mul_ss(_mm_load_ss(&s0[n]), int_scale); + in[1] = _mm_mul_ss(_mm_load_ss(&s1[n]), int_scale); in[0] = _mm_min_ss(int_max, _mm_max_ss(in[0], int_min)); in[1] = _mm_min_ss(int_max, _mm_max_ss(in[1], int_min)); d[0] = _mm_cvtss_si32(in[0]); @@ -1127,8 +1129,9 @@ conv_f32d_to_s16_4s_sse2(void *data, void * SPA_RESTRICT dst, const void * SPA_R uint32_t n, unrolled; __m128 in[4]; __m128i out[4], t[4]; - __m128 int_max = _mm_set1_ps(S16_MAX_F); - __m128 int_min = _mm_sub_ps(_mm_setzero_ps(), int_max); + __m128 int_scale = _mm_set1_ps(S16_SCALE); + __m128 int_max = _mm_set1_ps(S16_MAX); + __m128 int_min = _mm_set1_ps(S16_MIN); if (SPA_IS_ALIGNED(s0, 16) && SPA_IS_ALIGNED(s1, 16) && @@ -1139,15 +1142,15 @@ conv_f32d_to_s16_4s_sse2(void *data, void * SPA_RESTRICT dst, const void * SPA_R unrolled = 0; for(n = 0; n < unrolled; n += 4) { - in[0] = _mm_mul_ps(_mm_load_ps(&s0[n]), int_max); - in[1] = _mm_mul_ps(_mm_load_ps(&s1[n]), int_max); - in[2] = _mm_mul_ps(_mm_load_ps(&s2[n]), int_max); - in[3] = _mm_mul_ps(_mm_load_ps(&s3[n]), int_max); + in[0] = _mm_mul_ps(_mm_load_ps(&s0[n]), int_scale); + in[1] = _mm_mul_ps(_mm_load_ps(&s1[n]), int_scale); + in[2] = _mm_mul_ps(_mm_load_ps(&s2[n]), int_scale); + in[3] = _mm_mul_ps(_mm_load_ps(&s3[n]), int_scale); - t[0] = _mm_cvtps_epi32(in[0]); - t[1] = _mm_cvtps_epi32(in[1]); - t[2] = _mm_cvtps_epi32(in[2]); - t[3] = _mm_cvtps_epi32(in[3]); + t[0] = _mm_cvttps_epi32(in[0]); + t[1] = _mm_cvttps_epi32(in[1]); + t[2] = _mm_cvttps_epi32(in[2]); + t[3] = _mm_cvttps_epi32(in[3]); t[0] = _mm_packs_epi32(t[0], t[2]); t[1] = _mm_packs_epi32(t[1], t[3]); @@ -1165,10 +1168,10 @@ conv_f32d_to_s16_4s_sse2(void *data, void * SPA_RESTRICT dst, const void * SPA_R d += 4*n_channels; } for(; n < n_samples; n++) { - in[0] = _mm_mul_ss(_mm_load_ss(&s0[n]), int_max); - in[1] = _mm_mul_ss(_mm_load_ss(&s1[n]), int_max); - in[2] = _mm_mul_ss(_mm_load_ss(&s2[n]), int_max); - in[3] = _mm_mul_ss(_mm_load_ss(&s3[n]), int_max); + in[0] = _mm_mul_ss(_mm_load_ss(&s0[n]), int_scale); + in[1] = _mm_mul_ss(_mm_load_ss(&s1[n]), int_scale); + in[2] = _mm_mul_ss(_mm_load_ss(&s2[n]), int_scale); + in[3] = _mm_mul_ss(_mm_load_ss(&s3[n]), int_scale); in[0] = _mm_min_ss(int_max, _mm_max_ss(in[0], int_min)); in[1] = _mm_min_ss(int_max, _mm_max_ss(in[1], int_min)); in[2] = _mm_min_ss(int_max, _mm_max_ss(in[2], int_min)); @@ -1205,8 +1208,9 @@ conv_f32d_to_s16_2_sse2(struct convert *conv, void * SPA_RESTRICT dst[], const v uint32_t n, unrolled; __m128 in[4]; __m128i out[4]; - __m128 int_max = _mm_set1_ps(S16_MAX_F); - __m128 int_min = _mm_sub_ps(_mm_setzero_ps(), int_max); + __m128 int_scale = _mm_set1_ps(S16_SCALE); + __m128 int_max = _mm_set1_ps(S16_MAX); + __m128 int_min = _mm_set1_ps(S16_MIN); if (SPA_IS_ALIGNED(s0, 16) && SPA_IS_ALIGNED(s1, 16)) @@ -1215,15 +1219,15 @@ conv_f32d_to_s16_2_sse2(struct convert *conv, void * SPA_RESTRICT dst[], const v unrolled = 0; for(n = 0; n < unrolled; n += 8) { - in[0] = _mm_mul_ps(_mm_load_ps(&s0[n+0]), int_max); - in[1] = _mm_mul_ps(_mm_load_ps(&s1[n+0]), int_max); - in[2] = _mm_mul_ps(_mm_load_ps(&s0[n+4]), int_max); - in[3] = _mm_mul_ps(_mm_load_ps(&s1[n+4]), int_max); + in[0] = _mm_mul_ps(_mm_load_ps(&s0[n+0]), int_scale); + in[1] = _mm_mul_ps(_mm_load_ps(&s1[n+0]), int_scale); + in[2] = _mm_mul_ps(_mm_load_ps(&s0[n+4]), int_scale); + in[3] = _mm_mul_ps(_mm_load_ps(&s1[n+4]), int_scale); - out[0] = _mm_cvtps_epi32(in[0]); - out[1] = _mm_cvtps_epi32(in[1]); - out[2] = _mm_cvtps_epi32(in[2]); - out[3] = _mm_cvtps_epi32(in[3]); + out[0] = _mm_cvttps_epi32(in[0]); + out[1] = _mm_cvttps_epi32(in[1]); + out[2] = _mm_cvttps_epi32(in[2]); + out[3] = _mm_cvttps_epi32(in[3]); out[0] = _mm_packs_epi32(out[0], out[2]); out[1] = _mm_packs_epi32(out[1], out[3]); @@ -1237,8 +1241,8 @@ conv_f32d_to_s16_2_sse2(struct convert *conv, void * SPA_RESTRICT dst[], const v d += 16; } for(; n < n_samples; n++) { - in[0] = _mm_mul_ss(_mm_load_ss(&s0[n]), int_max); - in[1] = _mm_mul_ss(_mm_load_ss(&s1[n]), int_max); + in[0] = _mm_mul_ss(_mm_load_ss(&s0[n]), int_scale); + in[1] = _mm_mul_ss(_mm_load_ss(&s1[n]), int_scale); in[0] = _mm_min_ss(int_max, _mm_max_ss(in[0], int_min)); in[1] = _mm_min_ss(int_max, _mm_max_ss(in[1], int_min)); d[0] = _mm_cvtss_si32(in[0]); diff --git a/spa/plugins/audioconvert/fmt-ops.h b/spa/plugins/audioconvert/fmt-ops.h index 9399d6c5a..b2fe70365 100644 --- a/spa/plugins/audioconvert/fmt-ops.h +++ b/spa/plugins/audioconvert/fmt-ops.h @@ -47,7 +47,6 @@ #define S8_MIN -127 #define S8_MAX 127 -#define S8_MAX_F 127.0f #define S8_SCALE 127.0f #define S8_TO_F32(v) (((int8_t)(v)) * (1.0f / S8_SCALE)) #define F32_TO_S8(v) (int8_t)SPA_CLAMP((v) * S8_SCALE, S8_MIN, S8_MAX) @@ -57,17 +56,16 @@ #define U16_MAX 65535u #define U16_SCALE 32767.5f #define U16_OFFS 32768.f -#define U16_TO_F32(v) ((((uint16_t)(v)) * (1.0f / U16_OFFS)) - 1.0) -#define U16S_TO_F32(v) (((uint16_t)bswap_16((uint16_t)(v)) * (1.0f / U16_OFFS)) - 1.0) +#define U16_TO_F32(v) ((((uint16_t)(v)) * (1.0f / U16_OFFS)) - 1.0f) +#define U16S_TO_F32(v) (((uint16_t)bswap_16((uint16_t)(v)) * (1.0f / U16_OFFS)) - 1.0f) #define F32_TO_U16(v) (uint16_t)SPA_CLAMP((v) * U16_SCALE + U16_OFFS, U16_MIN, U16_MAX) #define F32_TO_U16_D(v,d) (uint16_t)SPA_CLAMP((v) * U16_SCALE + U16_OFFS + (d), U16_MIN, U16_MAX) #define F32_TO_U16S(v) bswap_16(F32_TO_U16(v)) #define F32_TO_U16S_D(v,d) bswap_16(F32_TO_U16_D(v,d)) -#define S16_MIN -32767 +#define S16_MIN -32768 #define S16_MAX 32767 -#define S16_MAX_F 32767.0f -#define S16_SCALE 32767.0f +#define S16_SCALE 32768.0f #define S16_TO_F32(v) (((int16_t)(v)) * (1.0f / S16_SCALE)) #define S16S_TO_F32(v) (((int16_t)bswap_16(v)) * (1.0f / S16_SCALE)) #define F32_TO_S16(v) (int16_t)SPA_CLAMP((v) * S16_SCALE, S16_MIN, S16_MAX) @@ -79,14 +77,13 @@ #define U24_MAX 16777215u #define U24_SCALE 8388607.5f #define U24_OFFS 8388608.f -#define U24_TO_F32(v) ((u24_to_u32(v) * (1.0f / U24_OFFS)) - 1.0) +#define U24_TO_F32(v) ((u24_to_u32(v) * (1.0f / U24_OFFS)) - 1.0f) #define F32_TO_U24(v) u32_to_u24(SPA_CLAMP((v) * U24_SCALE + U24_OFFS, U24_MIN, U24_MAX)) #define F32_TO_U24_D(v,d) u32_to_u24(SPA_CLAMP((v) * U24_SCALE + U24_OFFS + (d), U24_MIN, U24_MAX)) -#define S24_MIN -8388607 +#define S24_MIN -8388608 #define S24_MAX 8388607 -#define S24_MAX_F 8388607.0f -#define S24_SCALE 8388607.0f +#define S24_SCALE 8388608.0f #define S24_TO_F32(v) (s24_to_s32(v) * (1.0f / S24_SCALE)) #define S24S_TO_F32(v) (s24_to_s32(bswap_s24(v)) * (1.0f / S24_SCALE)) #define F32_TO_S24(v) s32_to_s24(SPA_CLAMP((v) * S24_SCALE, S24_MIN, S24_MAX)) @@ -94,16 +91,15 @@ #define F32_TO_S24_D(v,d) s32_to_s24(SPA_CLAMP((v) * S24_SCALE + (d), S24_MIN, S24_MAX)) #define U32_MIN 0u -#define U32_MAX 4294967040u -#define U32_SCALE 2147483520.f -#define U32_OFFS 2147483520.f -#define U32_TO_F32(v) ((((uint32_t)(v)) * (1.0f / U32_OFFS)) - 1.0) +#define U32_MAX 4294967295 +#define U32_SCALE 2147483647.5f +#define U32_OFFS 2147483648.f +#define U32_TO_F32(v) ((((uint32_t)(v)) * (1.0f / U32_OFFS)) - 1.0f) #define F32_TO_U32(v) (uint32_t)SPA_CLAMP((v) * U32_SCALE + U32_OFFS, U32_MIN, U32_MAX) #define F32_TO_U32_D(v,d) (uint32_t)SPA_CLAMP((v) * U32_SCALE + U32_OFFS + (d), U32_MIN, U32_MAX) -#define S32_MIN -2147483520 +#define S32_MIN -2147483648 #define S32_MAX 2147483520 -#define S32_MAX_F 2147483520.f #define S32_SCALE 2147483648.f #define S32_TO_F32(v) (((int32_t)(v)) * (1.0f / S32_SCALE)) #define S32S_TO_F32(v) (((int32_t)bswap_32(v)) * (1.0f / S32_SCALE)) diff --git a/spa/plugins/audioconvert/test-fmt-ops.c b/spa/plugins/audioconvert/test-fmt-ops.c index ad2099fc5..148eb3f28 100644 --- a/spa/plugins/audioconvert/test-fmt-ops.c +++ b/spa/plugins/audioconvert/test-fmt-ops.c @@ -50,11 +50,11 @@ static void compare_mem(int i, int j, const void *m1, const void *m2, size_t siz { int res = memcmp(m1, m2, size); if (res != 0) { - fprintf(stderr, "%d %d:\n", i, j); + fprintf(stderr, "%d %d %zd:\n", i, j, size); spa_debug_mem(0, m1, size); spa_debug_mem(0, m2, size); } -// spa_assert_se(res == 0); + spa_assert_se(res == 0); } static void run_test(const char *name, @@ -158,7 +158,7 @@ static void test_u8_f32(void) static void test_f32_u16(void) { static const float in[] = { 0.0f, 1.0f, -1.0f, 0.5f, -0.5f, 1.1f, -1.1f }; - static const uint16_t out[] = { 32767, 65535, 0, 49150, 16383, 65535, 0 }; + static const uint16_t out[] = { 32768, 65535, 0, 49151, 16384, 65535, 0 }; run_test("test_f32_u16", in, sizeof(in[0]), out, sizeof(out[0]), SPA_N_ELEMENTS(out), true, true, conv_f32_to_u16_c); @@ -168,8 +168,8 @@ static void test_f32_u16(void) static void test_u16_f32(void) { - static const uint16_t in[] = { 32767, 65535, 0, 49150, 16383, }; - static const float out[] = { 0.0f, 1.0f, -1.0f, 0.4999847412f, -0.4999847412f }; + static const uint16_t in[] = { 32768, 65535, 0, 49152, 16384, }; + static const float out[] = { 0.0f, 0.999969482422f, -1.0f, 0.5f, -0.5f }; run_test("test_u16_f32d", in, sizeof(in[0]), out, sizeof(out[0]), SPA_N_ELEMENTS(out), true, false, conv_u16_to_f32d_c); @@ -180,7 +180,7 @@ static void test_u16_f32(void) static void test_f32_s16(void) { static const float in[] = { 0.0f, 1.0f, -1.0f, 0.5f, -0.5f, 1.1f, -1.1f }; - static const int16_t out[] = { 0, 32767, -32767, 16383, -16383, 32767, -32767 }; + static const int16_t out[] = { 0, 32767, -32768, 16384, -16384, 32767, -32768 }; run_test("test_f32_s16", in, sizeof(in[0]), out, sizeof(out[0]), SPA_N_ELEMENTS(out), true, true, conv_f32_to_s16_c); @@ -192,16 +192,26 @@ static void test_f32_s16(void) false, false, conv_f32d_to_s16d_c); #if defined(HAVE_SSE2) if (cpu_flags & SPA_CPU_FLAG_SSE2) { + run_test("test_f32_s16_sse2", in, sizeof(in[0]), out, sizeof(out[0]), SPA_N_ELEMENTS(out), + true, true, conv_f32_to_s16_sse2); run_test("test_f32d_s16_sse2", in, sizeof(in[0]), out, sizeof(out[0]), SPA_N_ELEMENTS(out), false, true, conv_f32d_to_s16_sse2); + run_test("test_f32d_s16d_sse2", in, sizeof(in[0]), out, sizeof(out[0]), SPA_N_ELEMENTS(out), + false, false, conv_f32d_to_s16d_sse2); + } +#endif +#if defined(HAVE_AVX2) + if (cpu_flags & SPA_CPU_FLAG_AVX2) { + run_test("test_f32d_s16_avx2", in, sizeof(in[0]), out, sizeof(out[0]), SPA_N_ELEMENTS(out), + false, true, conv_f32d_to_s16_avx2); } #endif } static void test_s16_f32(void) { - static const int16_t in[] = { 0, 32767, -32767, 16383, -16383, }; - static const float out[] = { 0.0f, 1.0f, -1.0f, 0.4999847412f, -0.4999847412f }; + static const int16_t in[] = { 0, 32767, -32768, 16384, -16384, }; + static const float out[] = { 0.0f, 0.999969482422f, -1.0f, 0.5f, -0.5f }; run_test("test_s16_f32d", in, sizeof(in[0]), out, sizeof(out[0]), SPA_N_ELEMENTS(out), true, false, conv_s16_to_f32d_c); @@ -217,13 +227,19 @@ static void test_s16_f32(void) true, false, conv_s16_to_f32d_sse2); } #endif +#if defined(HAVE_AVX2) + if (cpu_flags & SPA_CPU_FLAG_AVX2) { + run_test("test_s16_f32d_avx2", in, sizeof(in[0]), out, sizeof(out[0]), SPA_N_ELEMENTS(out), + true, false, conv_s16_to_f32d_avx2); + } +#endif } static void test_f32_u32(void) { static const float in[] = { 0.0f, 1.0f, -1.0f, 0.5f, -0.5f, 1.1f, -1.1f }; - static const uint32_t out[] = { 0, 0x7fffff00, 0x80000100, 0x3fffff00, 0xc0000100, - 0x7fffff00, 0x80000100 }; + static const uint32_t out[] = { 0x80000000, 0xffffffff, 0x0, 0xc0000000, 0x40000000, + 0xffffffff, 0x0 }; run_test("test_f32_u32", in, sizeof(in[0]), out, sizeof(out[0]), SPA_N_ELEMENTS(out), true, true, conv_f32_to_u32_c); @@ -233,8 +249,8 @@ static void test_f32_u32(void) static void test_u32_f32(void) { - static const uint32_t in[] = { 0, 0x7fffff00, 0x80000100, 0x3fffff00, 0xc0000100 }; - static const float out[] = { 0.0f, 1.0f, -1.0f, 0.4999999404f, -0.4999999404f, }; + static const uint32_t in[] = { 0x80000000, 0xffffffff, 0x0, 0xc0000000, 0x40000000 }; + static const float out[] = { 0.0f, 1.0f, -1.0f, 0.5f, -0.5f, }; run_test("test_u32_f32d", in, sizeof(in[0]), out, sizeof(out[0]), SPA_N_ELEMENTS(out), true, false, conv_u32_to_f32d_c); @@ -245,8 +261,8 @@ static void test_u32_f32(void) static void test_f32_s32(void) { static const float in[] = { 0.0f, 1.0f, -1.0f, 0.5f, -0.5f, 1.1f, -1.1f }; - static const int32_t out[] = { 0, 0x7fffff00, 0x80000100, 0x3fffff00, 0xc0000100, - 0x7fffff00, 0x80000100 }; + static const int32_t out[] = { 0, 0x7fffff80, 0x80000000, 0x40000000, 0xc0000000, + 0x7fffff80, 0x80000000 }; run_test("test_f32_s32", in, sizeof(in[0]), out, sizeof(out[0]), SPA_N_ELEMENTS(out), true, true, conv_f32_to_s32_c); @@ -262,12 +278,18 @@ static void test_f32_s32(void) false, true, conv_f32d_to_s32_sse2); } #endif +#if defined(HAVE_AVX2) + if (cpu_flags & SPA_CPU_FLAG_AVX2) { + run_test("test_f32d_s32_avx2", in, sizeof(in[0]), out, sizeof(out[0]), SPA_N_ELEMENTS(out), + false, true, conv_f32d_to_s32_avx2); + } +#endif } static void test_s32_f32(void) { - static const int32_t in[] = { 0, 0x7fffff00, 0x80000100, 0x3fffff00, 0xc0000100 }; - static const float out[] = { 0.0f, 1.0f, -1.0f, 0.4999999404f, -0.4999999404f, }; + static const int32_t in[] = { 0, 0x7fffff80, 0x80000000, 0x40000000, 0xc0000000 }; + static const float out[] = { 0.0f, 0.999999940395f, -1.0f, 0.5, -0.5, }; run_test("test_s32_f32d", in, sizeof(in[0]), out, sizeof(out[0]), SPA_N_ELEMENTS(out), true, false, conv_s32_to_f32d_c); @@ -283,18 +305,20 @@ static void test_s32_f32(void) true, false, conv_s32_to_f32d_sse2); } #endif +#if defined(HAVE_AVX2) + if (cpu_flags & SPA_CPU_FLAG_AVX2) { + run_test("test_s32_f32d_avx2", in, sizeof(in[0]), out, sizeof(out[0]), SPA_N_ELEMENTS(out), + true, false, conv_s32_to_f32d_avx2); + } +#endif } static void test_f32_u24(void) { static const float in[] = { 0.0f, 1.0f, -1.0f, 0.5f, -0.5f, 1.1f, -1.1f }; -#if __BYTE_ORDER == __LITTLE_ENDIAN - static const uint8_t out[] = { 0x00, 0x00, 0x00, 0xff, 0xff, 0x7f, 0x01, 0x00, 0x80, - 0xff, 0xff, 0x3f, 0x01, 0x00, 0xc0, 0xff, 0xff, 0x7f, 0x01, 0x00, 0x80 }; -#else - static const uint8_t out[] = { 0x00, 0x00, 0x00, 0x7f, 0xff, 0xff, 0x80, 0x00, 0x01, - 0x3f, 0xff, 0xff, 0xc0, 0x00, 0x01, 0x7f, 0xff, 0xff, 0x80, 0x00, 0x01 }; -#endif + static const uint24_t out[] = { U32_TO_U24(0x00800000), U32_TO_U24(0xffffff), + U32_TO_U24(0x000000), U32_TO_U24(0xc00000), U32_TO_U24(0x400000), + U32_TO_U24(0xffffff), U32_TO_U24(0x000000) }; run_test("test_f32_u24", in, sizeof(in[0]), out, 3, SPA_N_ELEMENTS(in), true, true, conv_f32_to_u24_c); @@ -304,14 +328,9 @@ static void test_f32_u24(void) static void test_u24_f32(void) { -#if __BYTE_ORDER == __LITTLE_ENDIAN - static const uint8_t in[] = { 0x00, 0x00, 0x00, 0xff, 0xff, 0x7f, 0x01, 0x00, 0x80, - 0xff, 0xff, 0x3f, 0x01, 0x00, 0xc0, }; -#else - static const uint8_t in[] = { 0x00, 0x00, 0x00, 0x7f, 0xff, 0xff, 0x80, 0x00, 0x01, - 0x3f, 0xff, 0xff, 0xc0, 0x00, 0x01, }; -#endif - static const float out[] = { 0.0f, 1.0f, -1.0f, 0.4999999404f, -0.4999999404f, }; + static const uint24_t in[] = { U32_TO_U24(0x00800000), U32_TO_U24(0xffffff), + U32_TO_U24(0x000000), U32_TO_U24(0xc00000), U32_TO_U24(0x400000) }; + static const float out[] = { 0.0f, 0.999999880791f, -1.0f, 0.5, -0.5, }; run_test("test_u24_f32d", in, 3, out, sizeof(out[0]), SPA_N_ELEMENTS(out), true, false, conv_u24_to_f32d_c); @@ -322,13 +341,9 @@ static void test_u24_f32(void) static void test_f32_s24(void) { static const float in[] = { 0.0f, 1.0f, -1.0f, 0.5f, -0.5f, 1.1f, -1.1f }; -#if __BYTE_ORDER == __LITTLE_ENDIAN - static const uint8_t out[] = { 0x00, 0x00, 0x00, 0xff, 0xff, 0x7f, 0x01, 0x00, 0x80, - 0xff, 0xff, 0x3f, 0x01, 0x00, 0xc0, 0xff, 0xff, 0x7f, 0x01, 0x00, 0x80 }; -#else - static const uint8_t out[] = { 0x00, 0x00, 0x00, 0x7f, 0xff, 0xff, 0x80, 0x00, 0x01, - 0x3f, 0xff, 0xff, 0xc0, 0x00, 0x01, 0x7f, 0xff, 0xff, 0x80, 0x00, 0x01 }; -#endif + static const int24_t out[] = { S32_TO_S24(0), S32_TO_S24(0x7fffff), + S32_TO_S24(0xff800000), S32_TO_S24(0x400000), S32_TO_S24(0xc00000), + S32_TO_S24(0x7fffff), S32_TO_S24(0xff800000) }; run_test("test_f32_s24", in, sizeof(in[0]), out, 3, SPA_N_ELEMENTS(in), true, true, conv_f32_to_s24_c); @@ -342,14 +357,9 @@ static void test_f32_s24(void) static void test_s24_f32(void) { -#if __BYTE_ORDER == __LITTLE_ENDIAN - static const uint8_t in[] = { 0x00, 0x00, 0x00, 0xff, 0xff, 0x7f, 0x01, 0x00, 0x80, - 0xff, 0xff, 0x3f, 0x01, 0x00, 0xc0, }; -#else - static const uint8_t in[] = { 0x00, 0x00, 0x00, 0x7f, 0xff, 0xff, 0x80, 0x00, 0x01, - 0x3f, 0xff, 0xff, 0xc0, 0x00, 0x01, }; -#endif - static const float out[] = { 0.0f, 1.0f, -1.0f, 0.4999999404f, -0.4999999404f, }; + static const int24_t in[] = { S32_TO_S24(0), S32_TO_S24(0x7fffff), + S32_TO_S24(0xff800000), S32_TO_S24(0x400000), S32_TO_S24(0xc00000) }; + static const float out[] = { 0.0f, 0.999999880791f, -1.0f, 0.5f, -0.5f, }; run_test("test_s24_f32d", in, 3, out, sizeof(out[0]), SPA_N_ELEMENTS(out), true, false, conv_s24_to_f32d_c); @@ -377,13 +387,19 @@ static void test_s24_f32(void) true, false, conv_s24_to_f32d_sse41); } #endif +#if defined(HAVE_AVX2) + if (cpu_flags & SPA_CPU_FLAG_AVX2) { + run_test("test_s24_f32d_avx2", in, 3, out, sizeof(out[0]), SPA_N_ELEMENTS(out), + true, false, conv_s24_to_f32d_avx2); + } +#endif } static void test_f32_u24_32(void) { static const float in[] = { 0.0f, 1.0f, -1.0f, 0.5f, -0.5f, 1.1f, -1.1f }; - static const uint32_t out[] = { 0, 0x7fffff, 0xff800001, 0x3fffff, 0xffc00001, - 0x7fffff, 0xff800001 }; + static const uint32_t out[] = { 0x800000, 0xffffff, 0x0, 0xc00000, 0x400000, + 0xffffff, 0x000000 }; run_test("test_f32_u24_32", in, sizeof(in[0]), out, sizeof(out[0]), SPA_N_ELEMENTS(out), true, true, conv_f32_to_u24_32_c); @@ -393,8 +409,8 @@ static void test_f32_u24_32(void) static void test_u24_32_f32(void) { - static const uint32_t in[] = { 0, 0x7fffff, 0xff800001, 0x3fffff, 0xffc00001 }; - static const float out[] = { 0.0f, 1.0f, -1.0f, 0.4999999404f, -0.4999999404f, }; + static const uint32_t in[] = { 0x800000, 0xffffff, 0x0, 0xc00000, 0x400000 }; + static const float out[] = { 0.0f, 0.999999880791f, -1.0f, 0.5f, -0.5f, }; run_test("test_u24_32_f32d", in, sizeof(in[0]), out, sizeof(out[0]), SPA_N_ELEMENTS(out), true, false, conv_u24_32_to_f32d_c); @@ -405,8 +421,8 @@ static void test_u24_32_f32(void) static void test_f32_s24_32(void) { static const float in[] = { 0.0f, 1.0f, -1.0f, 0.5f, -0.5f, 1.1f, -1.1f }; - static const int32_t out[] = { 0, 0x7fffff, 0xff800001, 0x3fffff, 0xffc00001, - 0x7fffff, 0xff800001 }; + static const int32_t out[] = { 0, 0x7fffff, 0xff800000, 0x400000, 0xffc00000, + 0x7fffff, 0xff800000 }; run_test("test_f32_s24_32", in, sizeof(in[0]), out, sizeof(out[0]), SPA_N_ELEMENTS(out), true, true, conv_f32_to_s24_32_c); @@ -420,8 +436,8 @@ static void test_f32_s24_32(void) static void test_s24_32_f32(void) { - static const int32_t in[] = { 0, 0x7fffff, 0xff800001, 0x3fffff, 0xffc00001 }; - static const float out[] = { 0.0f, 1.0f, -1.0f, 0.4999999404f, -0.4999999404f, }; + static const int32_t in[] = { 0, 0x7fffff, 0xff800000, 0x400000, 0xffc00000 }; + static const float out[] = { 0.0f, 0.999999880791f, -1.0f, 0.5f, -0.5f, }; run_test("test_s24_32_f32d", in, sizeof(in[0]), out, sizeof(out[0]), SPA_N_ELEMENTS(out), true, false, conv_s24_32_to_f32d_c); @@ -435,8 +451,8 @@ static void test_s24_32_f32(void) static void test_f64_f32(void) { - static const double in[] = { 0.0, 1.0, -1.0, 0.4999999404, -0.4999999404, }; - static const float out[] = { 0.0f, 1.0f, -1.0f, 0.4999999404f, -0.4999999404f, }; + static const double in[] = { 0.0, 1.0, -1.0, 0.5, -0.5, }; + static const float out[] = { 0.0, 1.0, -1.0, 0.5, -0.5, }; run_test("test_f64_f32d", in, sizeof(in[0]), out, sizeof(out[0]), SPA_N_ELEMENTS(out), true, false, conv_f64_to_f32d_c); @@ -451,7 +467,7 @@ static void test_f64_f32(void) static void test_f32_f64(void) { static const float in[] = { 0.0f, 1.0f, -1.0f, 0.5f, -0.5f, 1.1f, -1.1f }; - static const double out[] = { 0.0, 1.0, -1.0, 0.5, -0.5, 1.1, -1.1 }; + static const double out[] = { 0.0f, 1.0f, -1.0f, 0.5f, -0.5f, 1.1f, -1.1f }; run_test("test_f32_f64", in, sizeof(in[0]), out, sizeof(out[0]), SPA_N_ELEMENTS(out), true, true, conv_f32_to_f64_c);