host: Add AVX2 support for uhd::convert #789
+1,383
−21
There are no files selected for viewing
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,193 @@ | ||
| // | ||
| // Copyright 2024 Ettus Research, a National Instruments Brand | ||
| // | ||
| // SPDX-License-Identifier: GPL-3.0-or-later | ||
| // | ||
|
|
||
| #include "convert_common.hpp" | ||
| #include <uhd/utils/byteswap.hpp> | ||
| #include <immintrin.h> | ||
|
|
||
| using namespace uhd::convert; | ||
|
|
||
| DECLARE_CONVERTER(fc32, 1, sc16_item32_le, 1, PRIORITY_SIMD) | ||
| { | ||
| const fc32_t* input = reinterpret_cast<const fc32_t*>(inputs[0]); | ||
| item32_t* output = reinterpret_cast<item32_t*>(outputs[0]); | ||
|
|
||
| const __m256 scalar = _mm256_set1_ps(float(scale_factor)); | ||
|
|
||
| // this macro converts values faster by using SSE intrinsics to convert 4 values at a time | ||
| #define convert_fc32_1_to_item32_1_nswap_guts(_al_) \ | ||
| for (; i + 7 < nsamps; i += 8) { \ | ||
| /* load from input */ \ | ||
| __m256 tmplo = \ | ||
| _mm256_load##_al_##ps(reinterpret_cast<const float*>(input + i + 0)); \ | ||
| __m256 tmphi = \ | ||
| _mm256_load##_al_##ps(reinterpret_cast<const float*>(input + i + 4)); \ | ||
| \ | ||
| /* convert and scale */ \ | ||
| __m256i tmpilo = _mm256_cvtps_epi32(_mm256_mul_ps(tmplo, scalar)); \ | ||
| __m256i tmpihi = _mm256_cvtps_epi32(_mm256_mul_ps(tmphi, scalar)); \ | ||
| \ | ||
| __m256i shuffled_lo = _mm256_permute2x128_si256( \ | ||
| tmpilo, tmpihi, 0x20); /* lower 128-bit of tmpilo and tmpihi */ \ | ||
| __m256i shuffled_hi = _mm256_permute2x128_si256( \ | ||
| tmpilo, tmpihi, 0x31); /* upper 128-bit of tmpilo and tmpihi */ \ | ||
| \ | ||
| /* now pack the shuffled data sequentially */ \ | ||
| __m256i tmpi = _mm256_packs_epi32(shuffled_lo, shuffled_hi); \ | ||
| \ | ||
| /* pack + swap 16-bit pairs */ \ | ||
| tmpi = _mm256_shufflelo_epi16(tmpi, _MM_SHUFFLE(2, 3, 0, 1)); \ | ||
| tmpi = _mm256_shufflehi_epi16(tmpi, _MM_SHUFFLE(2, 3, 0, 1)); \ | ||
| \ | ||
| /* store to output */ \ | ||
| _mm256_storeu_si256(reinterpret_cast<__m256i*>(output + i), tmpi); \ | ||
| } | ||
|
|
||
| size_t i = 0; | ||
|
|
||
| // need to dispatch according to alignment for fastest conversion | ||
| switch (size_t(input) & 0xf) { | ||
| case 0x0: | ||
| // the data is 16-byte aligned, so do the fast processing of the bulk of the | ||
| // samples | ||
| convert_fc32_1_to_item32_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 | ||
| xx_to_item32_sc16<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_fc32_1_to_item32_1_nswap_guts(_) break; | ||
| default: | ||
| // we are not 8 or 16-byte aligned, so do fast processing with the unaligned | ||
| // load | ||
| convert_fc32_1_to_item32_1_nswap_guts(u_) | ||
| } | ||
|
|
||
| // convert any remaining samples | ||
| xx_to_item32_sc16<uhd::htowx>(input + i, output + i, nsamps - i, scale_factor); | ||
| } | ||
|
|
||
| DECLARE_CONVERTER(fc32, 1, sc16_item32_be, 1, PRIORITY_SIMD) | ||
| { | ||
| const fc32_t* input = reinterpret_cast<const fc32_t*>(inputs[0]); | ||
| item32_t* output = reinterpret_cast<item32_t*>(outputs[0]); | ||
|
|
||
| const __m256 scalar = _mm256_set1_ps(float(scale_factor)); | ||
|
|
||
| // this macro converts values faster by using AVX2 intrinsics to convert 8 values at a | ||
| // time | ||
| #define convert_fc32_1_to_item32_1_bswap_guts(_al_) \ | ||
| for (; i + 7 < nsamps; i += 8) { \ | ||
| /* load from input */ \ | ||
| __m256 tmplo = \ | ||
| _mm256_load##_al_##ps(reinterpret_cast<const float*>(input + i + 0)); \ | ||
| __m256 tmphi = \ | ||
| _mm256_load##_al_##ps(reinterpret_cast<const float*>(input + i + 4)); \ | ||
| \ | ||
| /* convert and scale */ \ | ||
| __m256i tmpilo = _mm256_cvtps_epi32(_mm256_mul_ps(tmplo, scalar)); \ | ||
| __m256i tmpihi = _mm256_cvtps_epi32(_mm256_mul_ps(tmphi, scalar)); \ | ||
| \ | ||
| __m256i shuffled_lo = _mm256_permute2x128_si256( \ | ||
| tmpilo, tmpihi, 0x20); /* lower 128-bit of tmpilo and tmpihi */ \ | ||
| __m256i shuffled_hi = _mm256_permute2x128_si256( \ | ||
| tmpilo, tmpihi, 0x31); /* upper 128-bit of tmpilo and tmpihi */ \ | ||
| \ | ||
| /* Now pack the shuffled data sequentially */ \ | ||
| __m256i tmpi = _mm256_packs_epi32(shuffled_lo, shuffled_hi); \ | ||
| \ | ||
| tmpi = _mm256_or_si256(_mm256_srli_epi16(tmpi, 8), _mm256_slli_epi16(tmpi, 8)); \ | ||
| \ | ||
| /* store to output */ \ | ||
| _mm256_storeu_si256(reinterpret_cast<__m256i*>(output + i), tmpi); \ | ||
| } | ||
|
|
||
| size_t i = 0; | ||
|
|
||
| // need to dispatch according to alignment for fastest conversion | ||
| switch (size_t(input) & 0xf) { | ||
| case 0x0: | ||
| // the data is 16-byte aligned, so do the fast processing of the bulk of the | ||
| // samples | ||
| convert_fc32_1_to_item32_1_bswap_guts(_) break; | ||
| case 0x8: | ||
| // the first value is 8-byte aligned - process it and prepare the bulk of the | ||
| // data for fast conversion | ||
| xx_to_item32_sc16<uhd::htonx>(input, output, 1, scale_factor); | ||
| i++; | ||
| // do faster processing of the remaining samples now that we are 16-byte | ||
| // aligned | ||
| convert_fc32_1_to_item32_1_bswap_guts(_) break; | ||
| default: | ||
| // we are not 8 or 16-byte aligned, so do fast processing with the unaligned | ||
| // load | ||
| convert_fc32_1_to_item32_1_bswap_guts(u_) | ||
| } | ||
|
|
||
| // convert any remaining samples | ||
| xx_to_item32_sc16<uhd::htonx>(input + i, output + i, nsamps - i, scale_factor); | ||
| } | ||
|
|
||
| DECLARE_CONVERTER(fc32, 1, sc16_chdr, 1, PRIORITY_SIMD) | ||
| { | ||
| const fc32_t* input = reinterpret_cast<const fc32_t*>(inputs[0]); | ||
| sc16_t* output = reinterpret_cast<sc16_t*>(outputs[0]); | ||
|
|
||
| const __m256 scalar = _mm256_set1_ps(float(scale_factor)); | ||
|
|
||
| // this macro converts values faster by using SSE intrinsics to convert 4 values at a time | ||
| #define convert_fc32_1_to_item32_1_guts(_al_) \ | ||
| for (; i + 7 < nsamps; i += 8) { \ | ||
| /* load from input */ \ | ||
| __m256 tmplo = \ | ||
| _mm256_load##_al_##ps(reinterpret_cast<const float*>(input + i + 0)); \ | ||
| __m256 tmphi = \ | ||
| _mm256_load##_al_##ps(reinterpret_cast<const float*>(input + i + 4)); \ | ||
| \ | ||
| /* convert and scale */ \ | ||
| __m256i tmpilo = _mm256_cvtps_epi32(_mm256_mul_ps(tmplo, scalar)); \ | ||
| __m256i tmpihi = _mm256_cvtps_epi32(_mm256_mul_ps(tmphi, scalar)); \ | ||
| \ | ||
| /* mm256_packs_epi32 is not sequential, it needs to be split into m128i */ \ | ||
| __m256i shuffled_lo = _mm256_permute2x128_si256( \ | ||
| tmpilo, tmpihi, 0x20); /* lower 128-bit of tmpilo and tmpihi */ \ | ||
| __m256i shuffled_hi = _mm256_permute2x128_si256( \ | ||
| tmpilo, tmpihi, 0x31); /* upper 128-bit of tmpilo and tmpihi */ \ | ||
| \ | ||
| /* Now pack the shuffled data sequentially */ \ | ||
| __m256i tmpi = _mm256_packs_epi32(shuffled_lo, shuffled_hi); \ | ||
| \ | ||
| /* store to output */ \ | ||
| _mm256_storeu_si256(reinterpret_cast<__m256i*>(output + i), tmpi); \ | ||
| } | ||
|
|
||
| size_t i = 0; | ||
|
|
||
| // need to dispatch according to alignment for fastest conversion | ||
| switch (size_t(input) & 0xf) { | ||
| case 0x0: | ||
| // the data is 16-byte aligned, so do the fast processing of the bulk of the | ||
| // samples | ||
| convert_fc32_1_to_item32_1_guts(_) break; | ||
| case 0x8: | ||
| // the first sample is 8-byte aligned - process it to align the remainder of | ||
| // the samples to 16-bytes | ||
| xx_to_chdr_sc16(input, output, 1, scale_factor); | ||
| i++; | ||
| // do faster processing of the bulk of the samples now that we are 16-byte | ||
| // aligned | ||
| convert_fc32_1_to_item32_1_guts(_) break; | ||
| default: | ||
| // we are not 8 or 16-byte aligned, so do fast processing with the unaligned | ||
| // load | ||
| convert_fc32_1_to_item32_1_guts(u_) | ||
| } | ||
|
|
||
| // convert any remaining samples | ||
| xx_to_chdr_sc16(input + i, output + i, nsamps - i, scale_factor); | ||
| } |
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
All of this assumes the compiling machine has the same arch as the executing machine.