From 5a5d71c62382fea10b9802d577d15dd6ef3950bd Mon Sep 17 00:00:00 2001 From: aegroto Date: Sun, 8 Apr 2018 16:14:22 +0200 Subject: [PATCH] performance improvement in cryptonight x86 --- src/crypto/CryptoNight_x86.h | 40 ++++++++++++++++++++---------------- 1 file changed, 22 insertions(+), 18 deletions(-) diff --git a/src/crypto/CryptoNight_x86.h b/src/crypto/CryptoNight_x86.h index b7544dcc..f110f0b0 100644 --- a/src/crypto/CryptoNight_x86.h +++ b/src/crypto/CryptoNight_x86.h @@ -234,7 +234,9 @@ static inline void cn_explode_scratchpad(const __m128i *input, __m128i *output) xin6 = _mm_load_si128(input + 10); xin7 = _mm_load_si128(input + 11); - for (size_t i = 0; i < MEM / sizeof(__m128i); i += 8) { + __m128i *outputTmpLimit = output + (MEM / sizeof(__m128i)); + + for (__m128i *outputTmp = output; outputTmp < outputTmpLimit; outputTmp += 8) { aes_round(k0, &xin0, &xin1, &xin2, &xin3, &xin4, &xin5, &xin6, &xin7); aes_round(k1, &xin0, &xin1, &xin2, &xin3, &xin4, &xin5, &xin6, &xin7); aes_round(k2, &xin0, &xin1, &xin2, &xin3, &xin4, &xin5, &xin6, &xin7); @@ -246,14 +248,14 @@ static inline void cn_explode_scratchpad(const __m128i *input, __m128i *output) aes_round(k8, &xin0, &xin1, &xin2, &xin3, &xin4, &xin5, &xin6, &xin7); aes_round(k9, &xin0, &xin1, &xin2, &xin3, &xin4, &xin5, &xin6, &xin7); - _mm_store_si128(output + i + 0, xin0); - _mm_store_si128(output + i + 1, xin1); - _mm_store_si128(output + i + 2, xin2); - _mm_store_si128(output + i + 3, xin3); - _mm_store_si128(output + i + 4, xin4); - _mm_store_si128(output + i + 5, xin5); - _mm_store_si128(output + i + 6, xin6); - _mm_store_si128(output + i + 7, xin7); + _mm_store_si128(outputTmp, xin0); + _mm_store_si128(outputTmp + 1, xin1); + _mm_store_si128(outputTmp + 2, xin2); + _mm_store_si128(outputTmp + 3, xin3); + _mm_store_si128(outputTmp + 4, xin4); + _mm_store_si128(outputTmp + 5, xin5); + _mm_store_si128(outputTmp + 6, xin6); + _mm_store_si128(outputTmp + 7, xin7); } } @@ -275,16 +277,18 @@ static inline void cn_implode_scratchpad(const __m128i *input, __m128i *output) xout6 = _mm_load_si128(output + 10); xout7 = _mm_load_si128(output + 11); - for (size_t i = 0; i < MEM / sizeof(__m128i); i += 8) + __m128i *inputTmpLimit = (__m128i*) input + MEM / sizeof(__m128i); + + for (__m128i *inputTmp = (__m128i*) input; inputTmp < inputTmpLimit; inputTmp += 8) { - xout0 = _mm_xor_si128(_mm_load_si128(input + i + 0), xout0); - xout1 = _mm_xor_si128(_mm_load_si128(input + i + 1), xout1); - xout2 = _mm_xor_si128(_mm_load_si128(input + i + 2), xout2); - xout3 = _mm_xor_si128(_mm_load_si128(input + i + 3), xout3); - xout4 = _mm_xor_si128(_mm_load_si128(input + i + 4), xout4); - xout5 = _mm_xor_si128(_mm_load_si128(input + i + 5), xout5); - xout6 = _mm_xor_si128(_mm_load_si128(input + i + 6), xout6); - xout7 = _mm_xor_si128(_mm_load_si128(input + i + 7), xout7); + xout0 = _mm_xor_si128(_mm_load_si128(inputTmp), xout0); + xout1 = _mm_xor_si128(_mm_load_si128(inputTmp + 1), xout1); + xout2 = _mm_xor_si128(_mm_load_si128(inputTmp + 2), xout2); + xout3 = _mm_xor_si128(_mm_load_si128(inputTmp + 3), xout3); + xout4 = _mm_xor_si128(_mm_load_si128(inputTmp + 4), xout4); + xout5 = _mm_xor_si128(_mm_load_si128(inputTmp + 5), xout5); + xout6 = _mm_xor_si128(_mm_load_si128(inputTmp + 6), xout6); + xout7 = _mm_xor_si128(_mm_load_si128(inputTmp + 7), xout7); aes_round(k0, &xout0, &xout1, &xout2, &xout3, &xout4, &xout5, &xout6, &xout7); aes_round(k1, &xout0, &xout1, &xout2, &xout3, &xout4, &xout5, &xout6, &xout7);