From d46684710950a6695ee22c079327253a91eb268d Mon Sep 17 00:00:00 2001 From: Josh Blum Date: Wed, 8 Feb 2012 19:22:38 -0800 Subject: uhd: added sse2 conversions for fc64 to sc8 --- host/lib/convert/convert_fc64_with_sse2.cpp | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) (limited to 'host/lib/convert/convert_fc64_with_sse2.cpp') diff --git a/host/lib/convert/convert_fc64_with_sse2.cpp b/host/lib/convert/convert_fc64_with_sse2.cpp index 837bb584e..6e097e380 100644 --- a/host/lib/convert/convert_fc64_with_sse2.cpp +++ b/host/lib/convert/convert_fc64_with_sse2.cpp @@ -28,7 +28,7 @@ DECLARE_CONVERTER(fc64, 1, sc16_item32_le, 1, PRIORITY_SIMD){ const __m128d scalar = _mm_set1_pd(scale_factor); #define convert_fc64_1_to_item32_1_nswap_guts(_al_) \ - for (; i+4 < nsamps; i+=4){ \ + for (; i+3 < nsamps; i+=4){ \ /* load from input */ \ __m128d tmp0 = _mm_load ## _al_ ## pd(reinterpret_cast(input+i+0)); \ __m128d tmp1 = _mm_load ## _al_ ## pd(reinterpret_cast(input+i+1)); \ @@ -75,7 +75,7 @@ DECLARE_CONVERTER(fc64, 1, sc16_item32_be, 1, PRIORITY_SIMD){ const __m128d scalar = _mm_set1_pd(scale_factor); #define convert_fc64_1_to_item32_1_bswap_guts(_al_) \ - for (; i+4 < nsamps; i+=4){ \ + for (; i+3 < nsamps; i+=4){ \ /* load from input */ \ __m128d tmp0 = _mm_load ## _al_ ## pd(reinterpret_cast(input+i+0)); \ __m128d tmp1 = _mm_load ## _al_ ## pd(reinterpret_cast(input+i+1)); \ @@ -122,7 +122,7 @@ DECLARE_CONVERTER(sc16_item32_le, 1, fc64, 1, PRIORITY_SIMD){ const __m128i zeroi = _mm_setzero_si128(); #define convert_item32_1_to_fc64_1_nswap_guts(_al_) \ - for (; i+4 < nsamps; i+=4){ \ + for (; i+3 < nsamps; i+=4){ \ /* load from input */ \ __m128i tmpi = _mm_loadu_si128(reinterpret_cast(input+i)); \ \ @@ -171,7 +171,7 @@ DECLARE_CONVERTER(sc16_item32_be, 1, fc64, 1, PRIORITY_SIMD){ const __m128i zeroi = _mm_setzero_si128(); #define convert_item32_1_to_fc64_1_bswap_guts(_al_) \ - for (; i+4 < nsamps; i+=4){ \ + for (; i+3 < nsamps; i+=4){ \ /* load from input */ \ __m128i tmpi = _mm_loadu_si128(reinterpret_cast(input+i)); \ \ -- cgit v1.2.3