diff options
Diffstat (limited to 'host/lib/convert/sse2_sc16_to_fc32.cpp')
| -rw-r--r-- | host/lib/convert/sse2_sc16_to_fc32.cpp | 38 |
1 files changed, 28 insertions, 10 deletions
diff --git a/host/lib/convert/sse2_sc16_to_fc32.cpp b/host/lib/convert/sse2_sc16_to_fc32.cpp index c03e41585..0ac7f1798 100644 --- a/host/lib/convert/sse2_sc16_to_fc32.cpp +++ b/host/lib/convert/sse2_sc16_to_fc32.cpp @@ -28,6 +28,7 @@ DECLARE_CONVERTER(sc16_item32_le, 1, fc32, 1, PRIORITY_SIMD){ const __m128 scalar = _mm_set_ps1(float(scale_factor)/(1 << 16)); const __m128i zeroi = _mm_setzero_si128(); + // this macro converts values faster by using SSE intrinsics to convert 4 values at a time #define convert_item32_1_to_fc32_1_nswap_guts(_al_) \ for (; i+3 < nsamps; i+=4){ \ /* load from input */ \ @@ -50,17 +51,25 @@ DECLARE_CONVERTER(sc16_item32_le, 1, fc32, 1, PRIORITY_SIMD){ size_t i = 0; - //dispatch according to alignment + // need to dispatch according to alignment for fastest conversion switch (size_t(output) & 0xf){ - case 0x8: - item32_sc16_to_xx<uhd::htowx>(input, output, 1, scale_factor); i++; case 0x0: + // the data is 16-byte aligned, so do the fast processing of the bulk of the samples + convert_item32_1_to_fc32_1_nswap_guts(_) + break; + case 0x8: + // the first sample is 8-byte aligned - process it to align the remainder of the samples to 16-bytes + item32_sc16_to_xx<uhd::htowx>(input, output, 1, scale_factor); + i++; + // do faster processing of the bulk of the samples now that we are 16-byte aligned convert_item32_1_to_fc32_1_nswap_guts(_) break; - default: convert_item32_1_to_fc32_1_nswap_guts(u_) + default: + // we are not 8 or 16-byte aligned, so do fast processing with the unaligned load and store + convert_item32_1_to_fc32_1_nswap_guts(u_) } - //convert remainder + // convert any remaining samples item32_sc16_to_xx<uhd::htowx>(input+i, output+i, nsamps-i, scale_factor); } @@ -71,6 +80,7 @@ DECLARE_CONVERTER(sc16_item32_be, 1, fc32, 1, PRIORITY_SIMD){ const __m128 scalar = _mm_set_ps1(float(scale_factor)/(1 << 16)); const __m128i zeroi = _mm_setzero_si128(); + // this macro converts values faster by using SSE intrinsics to convert 4 values at a time #define convert_item32_1_to_fc32_1_bswap_guts(_al_) \ for (; i+3 < nsamps; i+=4){ \ /* load from input */ \ @@ -92,16 +102,24 @@ DECLARE_CONVERTER(sc16_item32_be, 1, fc32, 1, PRIORITY_SIMD){ size_t i = 0; - //dispatch according to alignment + // need to dispatch according to alignment for fastest conversion switch (size_t(output) & 0xf){ - case 0x8: - item32_sc16_to_xx<uhd::htonx>(input, output, 1, scale_factor); i++; case 0x0: + // the data is 16-byte aligned, so do the fast processing of the bulk of the samples + convert_item32_1_to_fc32_1_bswap_guts(_) + break; + case 0x8: + // the first sample is 8-byte aligned - process it to align the remainder of the samples to 16-bytes + item32_sc16_to_xx<uhd::htonx>(input, output, 1, scale_factor); + i++; + // do faster processing of the bulk of the samples now that we are 16-byte aligned convert_item32_1_to_fc32_1_bswap_guts(_) break; - default: convert_item32_1_to_fc32_1_bswap_guts(u_) + default: + // we are not 8 or 16-byte aligned, so do fast processing with the unaligned load and store + convert_item32_1_to_fc32_1_bswap_guts(u_) } - //convert remainder + // convert any remaining samples item32_sc16_to_xx<uhd::htonx>(input+i, output+i, nsamps-i, scale_factor); } |