From 04608db7e08401df4b5f91c76fff9c4dc62edaa9 Mon Sep 17 00:00:00 2001 From: kradchen Date: Tue, 9 May 2023 17:44:06 +0800 Subject: [PATCH] refactor convertfp16tofloat --- src/common/convertfp16tofloat.cpp | 155 +++++++++++++++++------------- test/Common_Test.cpp | 2 +- 2 files changed, 91 insertions(+), 66 deletions(-) diff --git a/src/common/convertfp16tofloat.cpp b/src/common/convertfp16tofloat.cpp index 6ecfa1c..1676324 100644 --- a/src/common/convertfp16tofloat.cpp +++ b/src/common/convertfp16tofloat.cpp @@ -4,74 +4,99 @@ #include #include #include +namespace { + const ushort CONVERT_AND_VALUE = 15; + // andblack + const __m128i andBlock = _mm_set_epi16(15, 15, 15, 15, 15, 15, 15, 15); + const __m128i andBlock2 = + _mm_set_epi16(2047, 2047, 2047, 2047, 2047, 2047, 2047, 2047); + const __m128i zeroBlock = _mm_set_epi16(0, 0, 0, 0, 0, 0, 0, 0); + const __m128i oneBlock = _mm_set_epi16(1, 1, 1, 1, 1, 1, 1, 1); + const __m128i twokBlock = + _mm_set_epi16(2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048); + const uint CONVERT_ADD_VALUE = UINT32_MAX - 4095; + void convert(short * ptr, double* des,bool single = false){ + // 初始化值 + auto value = _mm_set_epi16(ptr[0], ptr[1], ptr[2], ptr[3], single?ptr[0]:ptr[4], single?ptr[0]:ptr[5], + single?ptr[0]:ptr[6], single?ptr[0]:ptr[7]); + auto uvalue = _mm_set_epi16( + (ushort)ptr[0], (ushort)ptr[1], (ushort)ptr[2], (ushort)ptr[3], + (ushort)(single?ptr[0]:ptr[4]), (ushort)(single?ptr[0]:ptr[5]), + (ushort)(single?ptr[0]:ptr[6]), (ushort)(single?ptr[0]:ptr[7])); + // 位移 + auto sign_bit = _mm_srli_epi16(value, 15); // 右移16位取符号位 + auto exponent = _mm_srli_epi16(uvalue, 11); + // and + exponent = _mm_and_si128(exponent, andBlock); + // and ，then convert to int 32 bits + auto fraction3 = _mm256_cvtepi16_epi32(_mm_and_si128(uvalue, andBlock2)); + auto hidden_bit_mask = + (_mm_cmp_epi16_mask(sign_bit, oneBlock, _MM_CMPINT_EQ) & + _mm_cmp_epi16_mask(exponent, zeroBlock, _MM_CMPINT_EQ)) | + (_mm_cmp_epi16_mask(sign_bit, zeroBlock, _MM_CMPINT_EQ) & + _mm_cmp_epi16_mask(exponent, zeroBlock, _MM_CMPINT_NE)); + auto hidden_bit16 = _mm_maskz_set1_epi16(hidden_bit_mask, 2048); + auto hidden_bit32 = _mm256_cvtepi16_epi32(hidden_bit16); + auto outputBlock = _mm256_add_epi32(fraction3, hidden_bit32); + auto sign_bit_add_value = _mm256_maskz_set1_epi32( + _mm_cmp_epi16_mask(sign_bit, oneBlock, _MM_CMPINT_EQ), + CONVERT_ADD_VALUE); + outputBlock = _mm256_add_epi32(outputBlock, sign_bit_add_value); + auto exponent_mask = + _mm_cmp_epi16_mask(oneBlock, exponent, _MM_CMPINT_LT); + exponent = _mm_sub_epi16(exponent, oneBlock); + auto exponent32 = _mm256_cvtepi16_epi32(exponent); + auto zeroBlock32 = _mm256_cvtepi16_epi32(zeroBlock); + auto offsetCount = + _mm256_mask_blend_epi32(exponent_mask, zeroBlock32, exponent32); + + outputBlock = _mm256_sllv_epi32(outputBlock, offsetCount); + + des[3] = _mm256_extract_epi32(outputBlock, 4); + des[2] = _mm256_extract_epi32(outputBlock, 5); + des[1] = _mm256_extract_epi32(outputBlock, 6); + des[0] = _mm256_extract_epi32(outputBlock, 7); + if(single) return; + des[7] = _mm256_extract_epi32(outputBlock, 0); + des[6] = _mm256_extract_epi32(outputBlock, 1); + des[5] = _mm256_extract_epi32(outputBlock, 2); + des[4] = _mm256_extract_epi32(outputBlock, 3); + + } +} Aurora::Matrix Recon::convertfp16tofloat(Aurora::Matrix aMatrix) { - auto input = aMatrix.getData(); - // uint16变换为float(32位)输出大小翻倍 - auto output = Aurora::malloc(aMatrix.getDataSize() * 4); - size_t rows = aMatrix.getDataSize() * sizeof(double) / sizeof(short); - size_t total_count = aMatrix.getDataSize(); - const ushort CONVERT_AND_VALUE = 15; - // andblack - __m128i andBlock = _mm_set_epi16(15, 15, 15, 15, 15, 15, 15, 15); - __m128i andBlock2 = - _mm_set_epi16(2047, 2047, 2047, 2047, 2047, 2047, 2047, 2047); - __m128i zeroBlock = _mm_set_epi16(0, 0, 0, 0, 0, 0, 0, 0); - __m128i oneBlock = _mm_set_epi16(1, 1, 1, 1, 1, 1, 1, 1); - __m128i twokBlock = - _mm_set_epi16(2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048); - uint CONVERT_ADD_VALUE = UINT32_MAX - 4095; + auto input = aMatrix.getData(); + // uint16变换为float(32位)输出大小翻倍 + auto output = Aurora::malloc(aMatrix.getDataSize() * 4); + size_t rows = aMatrix.getDataSize() * sizeof(double) / sizeof(short); + size_t total_count = aMatrix.getDataSize(); -#pragma omp parallel for - for (size_t i = 0; i < total_count; i += 2) { - // 循环展开以避免过度的线程调用 - if (i + 2 < total_count) { - auto ptr = (short *)(input + i); - // 初始化值 - auto value = _mm_set_epi16(ptr[0], ptr[1], ptr[2], ptr[3], ptr[4], ptr[5], - ptr[6], ptr[7]); - auto uvalue = _mm_set_epi16( - (ushort)ptr[0], (ushort)ptr[1], (ushort)ptr[2], (ushort)ptr[3], - (ushort)ptr[4], (ushort)ptr[5], (ushort)ptr[6], (ushort)ptr[7]); - // 位移 - auto sign_bit = _mm_srli_epi16(value, 15); // 右移16位取符号位 - auto exponent = _mm_srli_epi16(uvalue, 11); - // and - exponent = _mm_and_si128(exponent, andBlock); - // and ，then convert to int 32 bits - auto fraction3 = _mm256_cvtepi16_epi32(_mm_and_si128(uvalue, andBlock2)); - auto hidden_bit_mask = - (_mm_cmp_epi16_mask(sign_bit, oneBlock, _MM_CMPINT_EQ) & - _mm_cmp_epi16_mask(exponent, zeroBlock, _MM_CMPINT_EQ)) | - (_mm_cmp_epi16_mask(sign_bit, zeroBlock, _MM_CMPINT_EQ) & - _mm_cmp_epi16_mask(exponent, zeroBlock, _MM_CMPINT_NE)); - auto hidden_bit16 = _mm_maskz_set1_epi16(hidden_bit_mask, 2048); - auto hidden_bit32 = _mm256_cvtepi16_epi32(hidden_bit16); - auto outputBlock = _mm256_add_epi32(fraction3, hidden_bit32); - auto sign_bit_add_value = _mm256_maskz_set1_epi32( - _mm_cmp_epi16_mask(sign_bit, oneBlock, _MM_CMPINT_EQ), - CONVERT_ADD_VALUE); - outputBlock = _mm256_add_epi32(outputBlock, sign_bit_add_value); - auto exponent_mask = - _mm_cmp_epi16_mask(oneBlock, exponent, _MM_CMPINT_LT); - exponent = _mm_sub_epi16(exponent, oneBlock); - auto exponent32 = _mm256_cvtepi16_epi32(exponent); - auto zeroBlock32 = _mm256_cvtepi16_epi32(zeroBlock); - auto offsetCount = - _mm256_mask_blend_epi32(exponent_mask, zeroBlock32, exponent32); - outputBlock = _mm256_sllv_epi32(outputBlock, offsetCount); - double *des = output + i * 4; - des[7] = (double)(int)_mm256_extract_epi32(outputBlock, 0); - des[6] = (double)(int)_mm256_extract_epi32(outputBlock, 1); - des[5] = (double)(int)_mm256_extract_epi32(outputBlock, 2); - des[4] = (double)(int)_mm256_extract_epi32(outputBlock, 3); - des[3] = (double)(int)_mm256_extract_epi32(outputBlock, 4); - des[2] = (double)(int)_mm256_extract_epi32(outputBlock, 5); - des[1] = (double)(int)_mm256_extract_epi32(outputBlock, 6); - des[0] = (double)(int)_mm256_extract_epi32(outputBlock, 7); + #pragma omp parallel for + for (size_t i = 0; i < total_count; i += 8) { + // 循环展开以避免过度的线程调用 + if (i < total_count) { + auto ptr = (short *)(input + i); + double *des = output + i * 4; + ::convert(ptr, des,i+1>total_count); + } + if (i+2 < total_count) { + auto ptr = (short *)(input + i + 2); + double *des = output + (i+2) * 4; + ::convert(ptr, des,i+3>total_count); + } + if (i+4 < total_count) { + auto ptr = (short *)(input + i + 4); + double *des = output + (i+4) * 4; + ::convert(ptr, des,i+5>total_count); + } + if (i+6 < total_count) { + auto ptr = (short *)(input + i + 6); + double *des = output + (i+6) * 4; + ::convert(ptr, des,i+7>total_count); + } } - } - return Aurora::Matrix::New(output, aMatrix.getDimSize(0), - aMatrix.getDimSize(1), aMatrix.getDimSize(2)); + return Aurora::Matrix::New(output, aMatrix.getDimSize(0), + aMatrix.getDimSize(1), aMatrix.getDimSize(2)); } \ No newline at end of file diff --git a/test/Common_Test.cpp b/test/Common_Test.cpp index 45e0c11..75d2839 100644 --- a/test/Common_Test.cpp +++ b/test/Common_Test.cpp @@ -51,7 +51,7 @@ TEST_F(Common_Test, convertfp16tofloat) { auto resultM = Recon::convertfp16tofloat(ma); auto result = resultM.getData(); auto output = m.read("output"); - for (size_t i = 0; i<10; i++) { + for (size_t i = 0; i