From 3e9a04383b1c65da9aa2253b5c96d8826a573751 Mon Sep 17 00:00:00 2001 From: Josh Blum Date: Mon, 23 May 2011 17:05:47 -0700 Subject: uhd: use aligned loads and stores for floats in sse2 conversion routines --- host/lib/convert/convert_with_sse2.cpp | 204 ++++++++++++++++++++------------- 1 file changed, 126 insertions(+), 78 deletions(-) (limited to 'host') diff --git a/host/lib/convert/convert_with_sse2.cpp b/host/lib/convert/convert_with_sse2.cpp index 52beea24a..9772028dc 100644 --- a/host/lib/convert/convert_with_sse2.cpp +++ b/host/lib/convert/convert_with_sse2.cpp @@ -25,25 +25,37 @@ DECLARE_CONVERTER(convert_fc32_1_to_item32_1_nswap, PRIORITY_CUSTOM){ const fc32_t *input = reinterpret_cast(inputs[0]); item32_t *output = reinterpret_cast(outputs[0]); - __m128 scalar = _mm_set_ps1(float(scale_factor)); - - //convert blocks of samples with intrinsics - size_t i = 0; for (; i < (nsamps & ~0x3); i+=4){ - //load from input - __m128 tmplo = _mm_loadu_ps(reinterpret_cast(input+i+0)); - __m128 tmphi = _mm_loadu_ps(reinterpret_cast(input+i+2)); - - //convert and scale - __m128i tmpilo = _mm_cvtps_epi32(_mm_mul_ps(tmplo, scalar)); - __m128i tmpihi = _mm_cvtps_epi32(_mm_mul_ps(tmphi, scalar)); - - //pack + swap 16-bit pairs - __m128i tmpi = _mm_packs_epi32(tmpilo, tmpihi); - tmpi = _mm_shufflelo_epi16(tmpi, _MM_SHUFFLE(2, 3, 0, 1)); - tmpi = _mm_shufflehi_epi16(tmpi, _MM_SHUFFLE(2, 3, 0, 1)); - - //store to output - _mm_storeu_si128(reinterpret_cast<__m128i *>(output+i), tmpi); + const __m128 scalar = _mm_set_ps1(float(scale_factor)); + + #define convert_fc32_1_to_item32_1_nswap_guts(_al_) \ + for (; i < (nsamps & ~0x3); i+=4){ \ + /* load from input */ \ + __m128 tmplo = _mm_load ## _al_ ## _ps(reinterpret_cast(input+i+0)); \ + __m128 tmphi = _mm_load ## _al_ ## _ps(reinterpret_cast(input+i+2)); \ + \ + /* convert and scale */ \ + __m128i tmpilo = _mm_cvtps_epi32(_mm_mul_ps(tmplo, scalar)); \ + __m128i tmpihi = _mm_cvtps_epi32(_mm_mul_ps(tmphi, scalar)); \ + \ + /* pack + swap 16-bit pairs */ \ + __m128i tmpi = _mm_packs_epi32(tmpilo, tmpihi); \ + tmpi = _mm_shufflelo_epi16(tmpi, _MM_SHUFFLE(2, 3, 0, 1)); \ + tmpi = _mm_shufflehi_epi16(tmpi, _MM_SHUFFLE(2, 3, 0, 1)); \ + \ + /* store to output */ \ + _mm_storeu_si128(reinterpret_cast<__m128i *>(output+i), tmpi); \ + } \ + + size_t i = 0; + + //dispatch according to alignment + switch (size_t(input) & 0xf){ + case 0x8: + output[i] = fc32_to_item32(input[i], float(scale_factor)); i++; + case 0x0: + convert_fc32_1_to_item32_1_nswap_guts() + break; + default: convert_fc32_1_to_item32_1_nswap_guts(u) } //convert remainder @@ -56,24 +68,36 @@ DECLARE_CONVERTER(convert_fc32_1_to_item32_1_bswap, PRIORITY_CUSTOM){ const fc32_t *input = reinterpret_cast(inputs[0]); item32_t *output = reinterpret_cast(outputs[0]); - __m128 scalar = _mm_set_ps1(float(scale_factor)); - - //convert blocks of samples with intrinsics - size_t i = 0; for (; i < (nsamps & ~0x3); i+=4){ - //load from input - __m128 tmplo = _mm_loadu_ps(reinterpret_cast(input+i+0)); - __m128 tmphi = _mm_loadu_ps(reinterpret_cast(input+i+2)); - - //convert and scale - __m128i tmpilo = _mm_cvtps_epi32(_mm_mul_ps(tmplo, scalar)); - __m128i tmpihi = _mm_cvtps_epi32(_mm_mul_ps(tmphi, scalar)); - - //pack + byteswap -> byteswap 16 bit words - __m128i tmpi = _mm_packs_epi32(tmpilo, tmpihi); - tmpi = _mm_or_si128(_mm_srli_epi16(tmpi, 8), _mm_slli_epi16(tmpi, 8)); - - //store to output - _mm_storeu_si128(reinterpret_cast<__m128i *>(output+i), tmpi); + const __m128 scalar = _mm_set_ps1(float(scale_factor)); + + #define convert_fc32_1_to_item32_1_bswap_guts(_al_) \ + for (; i < (nsamps & ~0x3); i+=4){ \ + /* load from input */ \ + __m128 tmplo = _mm_load ## _al_ ## _ps(reinterpret_cast(input+i+0)); \ + __m128 tmphi = _mm_load ## _al_ ## _ps(reinterpret_cast(input+i+2)); \ + \ + /* convert and scale */ \ + __m128i tmpilo = _mm_cvtps_epi32(_mm_mul_ps(tmplo, scalar)); \ + __m128i tmpihi = _mm_cvtps_epi32(_mm_mul_ps(tmphi, scalar)); \ + \ + /* pack + byteswap -> byteswap 16 bit words */ \ + __m128i tmpi = _mm_packs_epi32(tmpilo, tmpihi); \ + tmpi = _mm_or_si128(_mm_srli_epi16(tmpi, 8), _mm_slli_epi16(tmpi, 8)); \ + \ + /* store to output */ \ + _mm_storeu_si128(reinterpret_cast<__m128i *>(output+i), tmpi); \ + } \ + + size_t i = 0; + + //dispatch according to alignment + switch (size_t(input) & 0xf){ + case 0x8: + output[i] = uhd::byteswap(fc32_to_item32(input[i], float(scale_factor))); i++; + case 0x0: + convert_fc32_1_to_item32_1_bswap_guts() + break; + default: convert_fc32_1_to_item32_1_bswap_guts(u) } //convert remainder @@ -86,27 +110,39 @@ DECLARE_CONVERTER(convert_item32_1_to_fc32_1_nswap, PRIORITY_CUSTOM){ const item32_t *input = reinterpret_cast(inputs[0]); fc32_t *output = reinterpret_cast(outputs[0]); - __m128 scalar = _mm_set_ps1(float(scale_factor)/(1 << 16)); - __m128i zeroi = _mm_setzero_si128(); - - //convert blocks of samples with intrinsics - size_t i = 0; for (; i < (nsamps & ~0x3); i+=4){ - //load from input - __m128i tmpi = _mm_loadu_si128(reinterpret_cast(input+i)); - - //unpack + swap 16-bit pairs - tmpi = _mm_shufflelo_epi16(tmpi, _MM_SHUFFLE(2, 3, 0, 1)); - tmpi = _mm_shufflehi_epi16(tmpi, _MM_SHUFFLE(2, 3, 0, 1)); - __m128i tmpilo = _mm_unpacklo_epi16(zeroi, tmpi); //value in upper 16 bits - __m128i tmpihi = _mm_unpackhi_epi16(zeroi, tmpi); - - //convert and scale - __m128 tmplo = _mm_mul_ps(_mm_cvtepi32_ps(tmpilo), scalar); - __m128 tmphi = _mm_mul_ps(_mm_cvtepi32_ps(tmpihi), scalar); - - //store to output - _mm_storeu_ps(reinterpret_cast(output+i+0), tmplo); - _mm_storeu_ps(reinterpret_cast(output+i+2), tmphi); + const __m128 scalar = _mm_set_ps1(float(scale_factor)/(1 << 16)); + const __m128i zeroi = _mm_setzero_si128(); + + #define convert_item32_1_to_fc32_1_nswap_guts(_al_) \ + for (; i < (nsamps & ~0x3); i+=4){ \ + /* load from input */ \ + __m128i tmpi = _mm_loadu_si128(reinterpret_cast(input+i)); \ + \ + /* unpack + swap 16-bit pairs */ \ + tmpi = _mm_shufflelo_epi16(tmpi, _MM_SHUFFLE(2, 3, 0, 1)); \ + tmpi = _mm_shufflehi_epi16(tmpi, _MM_SHUFFLE(2, 3, 0, 1)); \ + __m128i tmpilo = _mm_unpacklo_epi16(zeroi, tmpi); /* value in upper 16 bits */ \ + __m128i tmpihi = _mm_unpackhi_epi16(zeroi, tmpi); \ + \ + /* convert and scale */ \ + __m128 tmplo = _mm_mul_ps(_mm_cvtepi32_ps(tmpilo), scalar); \ + __m128 tmphi = _mm_mul_ps(_mm_cvtepi32_ps(tmpihi), scalar); \ + \ + /* store to output */ \ + _mm_store ## _al_ ## _ps(reinterpret_cast(output+i+0), tmplo); \ + _mm_store ## _al_ ## _ps(reinterpret_cast(output+i+2), tmphi); \ + } \ + + size_t i = 0; + + //dispatch according to alignment + switch (size_t(output) & 0xf){ + case 0x8: + output[i] = item32_to_fc32(input[i], float(scale_factor)); i++; + case 0x0: + convert_item32_1_to_fc32_1_nswap_guts() + break; + default: convert_item32_1_to_fc32_1_nswap_guts(u) } //convert remainder @@ -119,26 +155,38 @@ DECLARE_CONVERTER(convert_item32_1_to_fc32_1_bswap, PRIORITY_CUSTOM){ const item32_t *input = reinterpret_cast(inputs[0]); fc32_t *output = reinterpret_cast(outputs[0]); - __m128 scalar = _mm_set_ps1(float(scale_factor)/(1 << 16)); - __m128i zeroi = _mm_setzero_si128(); - - //convert blocks of samples with intrinsics - size_t i = 0; for (; i < (nsamps & ~0x3); i+=4){ - //load from input - __m128i tmpi = _mm_loadu_si128(reinterpret_cast(input+i)); - - //byteswap + unpack -> byteswap 16 bit words - tmpi = _mm_or_si128(_mm_srli_epi16(tmpi, 8), _mm_slli_epi16(tmpi, 8)); - __m128i tmpilo = _mm_unpacklo_epi16(zeroi, tmpi); //value in upper 16 bits - __m128i tmpihi = _mm_unpackhi_epi16(zeroi, tmpi); - - //convert and scale - __m128 tmplo = _mm_mul_ps(_mm_cvtepi32_ps(tmpilo), scalar); - __m128 tmphi = _mm_mul_ps(_mm_cvtepi32_ps(tmpihi), scalar); - - //store to output - _mm_storeu_ps(reinterpret_cast(output+i+0), tmplo); - _mm_storeu_ps(reinterpret_cast(output+i+2), tmphi); + const __m128 scalar = _mm_set_ps1(float(scale_factor)/(1 << 16)); + const __m128i zeroi = _mm_setzero_si128(); + + #define convert_item32_1_to_fc32_1_bswap_guts(_al_) \ + for (; i < (nsamps & ~0x3); i+=4){ \ + /* load from input */ \ + __m128i tmpi = _mm_loadu_si128(reinterpret_cast(input+i)); \ + \ + /* byteswap + unpack -> byteswap 16 bit words */ \ + tmpi = _mm_or_si128(_mm_srli_epi16(tmpi, 8), _mm_slli_epi16(tmpi, 8)); \ + __m128i tmpilo = _mm_unpacklo_epi16(zeroi, tmpi); /* value in upper 16 bits */ \ + __m128i tmpihi = _mm_unpackhi_epi16(zeroi, tmpi); \ + \ + /* convert and scale */ \ + __m128 tmplo = _mm_mul_ps(_mm_cvtepi32_ps(tmpilo), scalar); \ + __m128 tmphi = _mm_mul_ps(_mm_cvtepi32_ps(tmpihi), scalar); \ + \ + /* store to output */ \ + _mm_store ## _al_ ## _ps(reinterpret_cast(output+i+0), tmplo); \ + _mm_store ## _al_ ## _ps(reinterpret_cast(output+i+2), tmphi); \ + } \ + + size_t i = 0; + + //dispatch according to alignment + switch (size_t(output) & 0xf){ + case 0x8: + output[i] = item32_to_fc32(uhd::byteswap(input[i]), float(scale_factor)); i++; + case 0x0: + convert_item32_1_to_fc32_1_bswap_guts() + break; + default: convert_item32_1_to_fc32_1_bswap_guts(u) } //convert remainder -- cgit v1.2.3