diff --git a/spa/plugins/audioconvert/fmt-ops-avx2.c b/spa/plugins/audioconvert/fmt-ops-avx2.c index 71ba45e29..8be3960e8 100644 --- a/spa/plugins/audioconvert/fmt-ops-avx2.c +++ b/spa/plugins/audioconvert/fmt-ops-avx2.c @@ -316,7 +316,7 @@ 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]; - __m256 out[4], factor = _mm256_set1_ps(1.0f / S32_SCALE); + __m256 out[4], factor = _mm256_set1_ps(1.0f / S32_SCALE_I2F); __m256i mask1 = _mm256_setr_epi32(0*n_channels, 1*n_channels, 2*n_channels, 3*n_channels, 4*n_channels, 5*n_channels, 6*n_channels, 7*n_channels); @@ -352,7 +352,7 @@ 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 / S32_SCALE); + __m128 out[4], factor = _mm_set1_ps(1.0f / S32_SCALE_I2F); out[0] = _mm_cvtsi32_ss(factor, s[0]); out[1] = _mm_cvtsi32_ss(factor, s[1]); out[2] = _mm_cvtsi32_ss(factor, s[2]); @@ -377,7 +377,7 @@ 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]; - __m256 out[4], factor = _mm256_set1_ps(1.0f / S32_SCALE); + __m256 out[4], factor = _mm256_set1_ps(1.0f / S32_SCALE_I2F); __m256i mask1 = _mm256_setr_epi32(0*n_channels, 1*n_channels, 2*n_channels, 3*n_channels, 4*n_channels, 5*n_channels, 6*n_channels, 7*n_channels); @@ -403,7 +403,7 @@ 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 / S32_SCALE); + __m128 out[2], factor = _mm_set1_ps(1.0f / S32_SCALE_I2F); out[0] = _mm_cvtsi32_ss(factor, s[0]); out[1] = _mm_cvtsi32_ss(factor, s[1]); out[0] = _mm_mul_ss(out[0], factor); @@ -422,7 +422,7 @@ 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 / S32_SCALE); + __m256 out[2], factor = _mm256_set1_ps(1.0f / S32_SCALE_I2F); __m256i mask1 = _mm256_setr_epi32(0*n_channels, 1*n_channels, 2*n_channels, 3*n_channels, 4*n_channels, 5*n_channels, 6*n_channels, 7*n_channels); @@ -447,7 +447,7 @@ 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 / S32_SCALE); + __m128 out, factor = _mm_set1_ps(1.0f / S32_SCALE_I2F); out = _mm_cvtsi32_ss(factor, s[0]); out = _mm_mul_ss(out, factor); _mm_store_ss(&d0[n], out); @@ -479,9 +479,9 @@ conv_f32d_to_s32_1s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R uint32_t n, unrolled; __m128 in[1]; __m128i out[4]; - __m128 scale = _mm_set1_ps(S25_SCALE); - __m128 int_max = _mm_set1_ps(S25_MAX); - __m128 int_min = _mm_set1_ps(S25_MIN); + __m128 scale = _mm_set1_ps(S32_SCALE_F2I); + __m128 int_min = _mm_set1_ps(S32_MIN_F2I); + __m128 int_max = _mm_set1_ps(S32_MAX_F2I); if (SPA_IS_ALIGNED(s0, 16)) unrolled = n_samples & ~3; @@ -492,7 +492,6 @@ conv_f32d_to_s32_1s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R in[0] = _mm_mul_ps(_mm_load_ps(&s0[n]), scale); in[0] = _MM_CLAMP_PS(in[0], int_min, int_max); out[0] = _mm_cvtps_epi32(in[0]); - out[0] = _mm_slli_epi32(out[0], 7); 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)); @@ -507,7 +506,7 @@ conv_f32d_to_s32_1s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R in[0] = _mm_load_ss(&s0[n]); in[0] = _mm_mul_ss(in[0], scale); in[0] = _MM_CLAMP_SS(in[0], int_min, int_max); - *d = _mm_cvtss_si32(in[0]) << 7; + *d = _mm_cvtss_si32(in[0]); d += n_channels; } } @@ -527,9 +526,9 @@ conv_f32d_to_s32_2s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R uint32_t n, unrolled; __m256 in[2]; __m256i out[2], t[2]; - __m256 scale = _mm256_set1_ps(S25_SCALE); - __m256 int_min = _mm256_set1_ps(S25_MIN); - __m256 int_max = _mm256_set1_ps(S25_MAX); + __m256 scale = _mm256_set1_ps(S32_SCALE_F2I); + __m256 int_min = _mm256_set1_ps(S32_MIN_F2I); + __m256 int_max = _mm256_set1_ps(S32_MAX_F2I); if (SPA_IS_ALIGNED(s0, 32) && SPA_IS_ALIGNED(s1, 32)) @@ -546,8 +545,6 @@ conv_f32d_to_s32_2s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R 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_slli_epi32(out[0], 7); - out[1] = _mm256_slli_epi32(out[1], 7); 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 */ @@ -576,9 +573,9 @@ conv_f32d_to_s32_2s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R for(; n < n_samples; n++) { __m128 in[2]; __m128i out[2]; - __m128 scale = _mm_set1_ps(S25_SCALE); - __m128 int_min = _mm_set1_ps(S25_MIN); - __m128 int_max = _mm_set1_ps(S25_MAX); + __m128 scale = _mm_set1_ps(S32_SCALE_F2I); + __m128 int_min = _mm_set1_ps(S32_MIN_F2I); + __m128 int_max = _mm_set1_ps(S32_MAX_F2I); in[0] = _mm_load_ss(&s0[n]); in[1] = _mm_load_ss(&s1[n]); @@ -588,7 +585,6 @@ 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_CLAMP_PS(in[0], int_min, int_max); out[0] = _mm_cvtps_epi32(in[0]); - out[0] = _mm_slli_epi32(out[0], 7); _mm_storel_epi64((__m128i*)d, out[0]); d += n_channels; } @@ -603,9 +599,9 @@ conv_f32d_to_s32_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 scale = _mm256_set1_ps(S25_SCALE); - __m256 int_min = _mm256_set1_ps(S25_MIN); - __m256 int_max = _mm256_set1_ps(S25_MAX); + __m256 scale = _mm256_set1_ps(S32_SCALE_F2I); + __m256 int_min = _mm256_set1_ps(S32_MIN_F2I); + __m256 int_max = _mm256_set1_ps(S32_MAX_F2I); if (SPA_IS_ALIGNED(s0, 32) && SPA_IS_ALIGNED(s1, 32) && @@ -630,10 +626,6 @@ conv_f32d_to_s32_4s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R 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_slli_epi32(out[0], 7); - out[1] = _mm256_slli_epi32(out[1], 7); - out[2] = _mm256_slli_epi32(out[2], 7); - out[3] = _mm256_slli_epi32(out[3], 7); 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 */ @@ -658,9 +650,9 @@ conv_f32d_to_s32_4s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R for(; n < n_samples; n++) { __m128 in[4]; __m128i out[4]; - __m128 scale = _mm_set1_ps(S25_SCALE); - __m128 int_min = _mm_set1_ps(S25_MIN); - __m128 int_max = _mm_set1_ps(S25_MAX); + __m128 scale = _mm_set1_ps(S32_SCALE_F2I); + __m128 int_min = _mm_set1_ps(S32_MIN_F2I); + __m128 int_max = _mm_set1_ps(S32_MAX_F2I); in[0] = _mm_load_ss(&s0[n]); in[1] = _mm_load_ss(&s1[n]); @@ -674,7 +666,6 @@ 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_CLAMP_PS(in[0], int_min, int_max); out[0] = _mm_cvtps_epi32(in[0]); - out[0] = _mm_slli_epi32(out[0], 7); _mm_storeu_si128((__m128i*)d, out[0]); d += n_channels; } diff --git a/spa/plugins/audioconvert/fmt-ops-sse2.c b/spa/plugins/audioconvert/fmt-ops-sse2.c index a2a420459..957cb1669 100644 --- a/spa/plugins/audioconvert/fmt-ops-sse2.c +++ b/spa/plugins/audioconvert/fmt-ops-sse2.c @@ -335,7 +335,7 @@ 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 / S32_SCALE); + __m128 out, factor = _mm_set1_ps(1.0f / S32_SCALE_I2F); if (SPA_IS_ALIGNED(d0, 16)) unrolled = n_samples & ~3; @@ -380,9 +380,9 @@ conv_f32d_to_s32_1s_sse2(void *data, void * SPA_RESTRICT dst, const void * SPA_R uint32_t n, unrolled; __m128 in[1]; __m128i out[4]; - __m128 scale = _mm_set1_ps(S25_SCALE); - __m128 int_min = _mm_set1_ps(S25_MIN); - __m128 int_max = _mm_set1_ps(S25_MAX); + __m128 scale = _mm_set1_ps(S32_SCALE_F2I); + __m128 int_min = _mm_set1_ps(S32_MIN_F2I); + __m128 int_max = _mm_set1_ps(S32_MAX_F2I); if (SPA_IS_ALIGNED(s0, 16)) unrolled = n_samples & ~3; @@ -393,7 +393,6 @@ conv_f32d_to_s32_1s_sse2(void *data, void * SPA_RESTRICT dst, const void * SPA_R in[0] = _mm_mul_ps(_mm_load_ps(&s0[n]), scale); in[0] = _MM_CLAMP_PS(in[0], int_min, int_max); out[0] = _mm_cvtps_epi32(in[0]); - out[0] = _mm_slli_epi32(out[0], 7); 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)); @@ -408,7 +407,7 @@ conv_f32d_to_s32_1s_sse2(void *data, void * SPA_RESTRICT dst, const void * SPA_R in[0] = _mm_load_ss(&s0[n]); in[0] = _mm_mul_ss(in[0], scale); in[0] = _MM_CLAMP_SS(in[0], int_min, int_max); - *d = _mm_cvtss_si32(in[0]) << 7; + *d = _mm_cvtss_si32(in[0]); d += n_channels; } } @@ -422,9 +421,9 @@ conv_f32d_to_s32_2s_sse2(void *data, void * SPA_RESTRICT dst, const void * SPA_R uint32_t n, unrolled; __m128 in[2]; __m128i out[2], t[2]; - __m128 scale = _mm_set1_ps(S25_SCALE); - __m128 int_min = _mm_set1_ps(S25_MIN); - __m128 int_max = _mm_set1_ps(S25_MAX); + __m128 scale = _mm_set1_ps(S32_SCALE_F2I); + __m128 int_min = _mm_set1_ps(S32_MIN_F2I); + __m128 int_max = _mm_set1_ps(S32_MAX_F2I); if (SPA_IS_ALIGNED(s0, 16) && SPA_IS_ALIGNED(s1, 16)) @@ -441,8 +440,6 @@ conv_f32d_to_s32_2s_sse2(void *data, void * SPA_RESTRICT dst, const void * SPA_R out[0] = _mm_cvtps_epi32(in[0]); out[1] = _mm_cvtps_epi32(in[1]); - out[0] = _mm_slli_epi32(out[0], 7); - out[1] = _mm_slli_epi32(out[1], 7); t[0] = _mm_unpacklo_epi32(out[0], out[1]); t[1] = _mm_unpackhi_epi32(out[0], out[1]); @@ -462,7 +459,6 @@ 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_CLAMP_PS(in[0], int_min, int_max); out[0] = _mm_cvtps_epi32(in[0]); - out[0] = _mm_slli_epi32(out[0], 7); _mm_storel_epi64((__m128i*)d, out[0]); d += n_channels; } @@ -477,9 +473,9 @@ conv_f32d_to_s32_4s_sse2(void *data, void * SPA_RESTRICT dst, const void * SPA_R uint32_t n, unrolled; __m128 in[4]; __m128i out[4]; - __m128 scale = _mm_set1_ps(S25_SCALE); - __m128 int_min = _mm_set1_ps(S25_MIN); - __m128 int_max = _mm_set1_ps(S25_MAX); + __m128 scale = _mm_set1_ps(S32_SCALE_F2I); + __m128 int_min = _mm_set1_ps(S32_MIN_F2I); + __m128 int_max = _mm_set1_ps(S32_MAX_F2I); if (SPA_IS_ALIGNED(s0, 16) && SPA_IS_ALIGNED(s1, 16) && @@ -506,10 +502,6 @@ conv_f32d_to_s32_4s_sse2(void *data, void * SPA_RESTRICT dst, const void * SPA_R out[1] = _mm_cvtps_epi32(in[1]); out[2] = _mm_cvtps_epi32(in[2]); out[3] = _mm_cvtps_epi32(in[3]); - out[0] = _mm_slli_epi32(out[0], 7); - out[1] = _mm_slli_epi32(out[1], 7); - out[2] = _mm_slli_epi32(out[2], 7); - out[3] = _mm_slli_epi32(out[3], 7); _mm_storeu_si128((__m128i*)(d + 0*n_channels), out[0]); _mm_storeu_si128((__m128i*)(d + 1*n_channels), out[1]); @@ -530,7 +522,6 @@ 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_CLAMP_PS(in[0], int_min, int_max); out[0] = _mm_cvtps_epi32(in[0]); - out[0] = _mm_slli_epi32(out[0], 7); _mm_storeu_si128((__m128i*)d, out[0]); d += n_channels; } @@ -629,9 +620,9 @@ conv_f32d_to_s32_1s_noise_sse2(struct convert *conv, void * SPA_RESTRICT dst, co uint32_t n, unrolled; __m128 in[1]; __m128i out[4]; - __m128 scale = _mm_set1_ps(S25_SCALE); - __m128 int_min = _mm_set1_ps(S25_MIN); - __m128 int_max = _mm_set1_ps(S25_MAX); + __m128 scale = _mm_set1_ps(S32_SCALE_F2I); + __m128 int_min = _mm_set1_ps(S32_MIN_F2I); + __m128 int_max = _mm_set1_ps(S32_MAX_F2I); if (SPA_IS_ALIGNED(s, 16)) unrolled = n_samples & ~3; @@ -643,7 +634,6 @@ conv_f32d_to_s32_1s_noise_sse2(struct convert *conv, void * SPA_RESTRICT dst, co in[0] = _mm_add_ps(in[0], _mm_load_ps(&noise[n])); in[0] = _MM_CLAMP_PS(in[0], int_min, int_max); out[0] = _mm_cvtps_epi32(in[0]); - out[0] = _mm_slli_epi32(out[0], 7); 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)); @@ -659,7 +649,7 @@ conv_f32d_to_s32_1s_noise_sse2(struct convert *conv, void * SPA_RESTRICT dst, co in[0] = _mm_mul_ss(in[0], scale); in[0] = _mm_add_ss(in[0], _mm_load_ss(&noise[n])); in[0] = _MM_CLAMP_SS(in[0], int_min, int_max); - *d = _mm_cvtss_si32(in[0]) << 7; + *d = _mm_cvtss_si32(in[0]); d += n_channels; } } diff --git a/spa/plugins/audioconvert/fmt-ops.h b/spa/plugins/audioconvert/fmt-ops.h index da0338fd7..81434af0d 100644 --- a/spa/plugins/audioconvert/fmt-ops.h +++ b/spa/plugins/audioconvert/fmt-ops.h @@ -121,10 +121,14 @@ #define S32_MIN -2147483648 #define S32_MAX 2147483647 -#define S32_SCALE 2147483648.0f -#define S32_TO_F32(v) ITOF(int32_t, v, S32_SCALE, 0.0f) +#define S32_SCALE_I2F 2147483648.0f +#define S32_TO_F32(v) ITOF(int32_t, v, S32_SCALE_I2F, 0.0f) #define S32S_TO_F32(v) S32_TO_F32(bswap_32(v)) -#define F32_TO_S32_D(v,d) S25_32_TO_S32(F32_TO_S25_32_D(v,d)) + +#define S32_MIN_F2I ((int32_t)(((uint32_t)(S25_MIN)) << 7)) +#define S32_MAX_F2I ((S25_MAX) << 7) +#define S32_SCALE_F2I (-((float)(S32_MIN_F2I))) +#define F32_TO_S32_D(v,d) FTOI(int32_t, v, S32_SCALE_F2I, 0.0f, d, S32_MIN_F2I, S32_MAX_F2I) #define F32_TO_S32(v) F32_TO_S32_D(v, 0.0f) #define F32_TO_S32S(v) bswap_32(F32_TO_S32(v)) #define F32_TO_S32S_D(v,d) bswap_32(F32_TO_S32_D(v,d)) diff --git a/spa/plugins/audioconvert/test-fmt-ops.c b/spa/plugins/audioconvert/test-fmt-ops.c index 9c6191e0c..514f42012 100644 --- a/spa/plugins/audioconvert/test-fmt-ops.c +++ b/spa/plugins/audioconvert/test-fmt-ops.c @@ -299,11 +299,11 @@ static void test_f32_s32(void) 1.0f/0x100000000, -1.0f/0x100000000, 1.0f/0x200000000, -1.0f/0x200000000, }; static const int32_t out[] = { 0x00000000, 0x7fffff80, 0x80000000, - 0x40000000, 0xc0000000, 0x7fffff80, 0x80000000, 0x00000100, - 0xffffff00, 0x00000100, 0xffffff00, 0x00000080, 0xffffff80, - 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, - 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, - 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, + 0x40000000, 0xc0000000, 0x7fffff80, 0x80000000, 0x000000cd, + 0xffffff33, 0x00000100, 0xffffff00, 0x00000080, 0xffffff80, + 0x00000040, 0xffffffc0, 0x00000020, 0xffffffe0, 0x00000010, + 0xfffffff0, 0x00000008, 0xfffffff8, 0x00000004, 0xfffffffc, + 0x00000002, 0xfffffffe, 0x00000001, 0xffffffff, 0x00000000, 0x00000000, 0x00000000, 0x00000000, }; @@ -687,20 +687,15 @@ static void test_lossless_s32_lossless_subset(void) { int32_t i, j; - int all_lossless = 1; - int32_t max_abs_err = -1; fprintf(stderr, "test %s:\n", __func__); for (i = S25_MIN; i <= S25_MAX; i+=1) { for(j = 0; j < 8; ++j) { int32_t s = i * (1<