sunwen/UR
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

cuda2023

Browse Source
This commit is contained in:
kradchen
2023-12-26 16:29:40 +08:00
parent 34d7086b47 c40dc8e938
commit 4dce7c90ce
30 changed files with 1564 additions and 311 deletions
Download Patch File Download Diff File Expand all files Collapse all files

177
src/common/convertfp16tofloat.cpp
View File

@@ -5,98 +5,103 @@
#include <immintrin.h>
#include <sys/types.h>
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, float* 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);
// 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);
// 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(float) / sizeof(short);
size_t total_count = aMatrix.getDataSize();
#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);
float *des = output + i * 4;
::convert(ptr, des,i+1>total_count);
}
if (i+2 < total_count) {
auto ptr = (short *)(input + i + 2);
float *des = output + (i+2) * 4;
::convert(ptr, des,i+3>total_count);
}
if (i+4 < total_count) {
auto ptr = (short *)(input + i + 4);
float *des = output + (i+4) * 4;
::convert(ptr, des,i+5>total_count);
}
if (i+6 < total_count) {
auto ptr = (short *)(input + i + 6);
float *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));
// 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 += 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));
}

100
src/common/dataBlockCreation/getAScanBlockPreprocessed.cpp
View File

@@ -1,5 +1,6 @@
#include "getAScanBlockPreprocessed.h"
#include "CudaMatrix.h"
#include "Matrix.h"
#include "blockingGeometryInfo.h"
#include "removeDataFromArrays.h"
@@ -10,15 +11,36 @@
#include "src/transmissionReconstruction/dataFilter/dataFilter.h"
#include "src/reflectionReconstruction/dataFilter.h"
#include "Aurora.h"
using namespace Aurora;
using namespace Recon;
#include <sys/time.h>
#include <iostream>
void printTime()
{
struct timeval tpend;
gettimeofday(&tpend,NULL);
int secofday = (tpend.tv_sec + 3600 * 8 ) % 86400;
int hours = secofday / 3600;
int minutes = (secofday - hours * 3600 ) / 60;
int seconds = secofday % 60;
int milliseconds = tpend.tv_usec/1000;
std::cout<< hours << ":" <<minutes<<":"<<seconds<<"."<<milliseconds<<std::endl;
}
AscanBlockPreprocessed Recon::getAscanBlockPreprocessed(Parser* aParser, const Aurora::Matrix& aMp, const Aurora::Matrix& aSl, const Aurora::Matrix& aSn,
const Aurora::Matrix& aRl, const Aurora::Matrix& aRn, GeometryInfo& aGeom, const MeasurementInfo& aMeasInfo,
bool aApplyFilter, bool aTransReco)
{
//std::cout<<"strart"<<std::endl;
//printTime();
//550ms
AscanBlockPreprocessed result;
AscanBlock ascanBlock = getAscanBlock(aParser, aMp, aSl, aSn, aRl, aRn);
//printTime();
//10ms
result.gainBlock = ascanBlock.gainBlock;
result.mpBlock = ascanBlock.mpBlock;
result.rlBlock = ascanBlock.rlBlock;
@@ -26,6 +48,8 @@ AscanBlockPreprocessed Recon::getAscanBlockPreprocessed(Parser* aParser, const A
result.slBlock = ascanBlock.slBlock;
result.snBlock = ascanBlock.snBlock;
GeometryBlock geometryBlock = blockingGeometryInfos(aGeom, ascanBlock.rnBlock, ascanBlock.rlBlock, ascanBlock.snBlock, ascanBlock.slBlock, ascanBlock.mpBlock);
//printTime();
//3ms
result.receiverPositionBlock = geometryBlock.receiverPositionBlock;
result.senderPositionBlock = geometryBlock.senderPositionBlock;
if(aApplyFilter)
@@ -40,7 +64,8 @@ AscanBlockPreprocessed Recon::getAscanBlockPreprocessed(Parser* aParser, const A
{
usedData = filterReflectionData(geometryBlock.receiverPositionBlock, geometryBlock.senderPositionBlock, geometryBlock.senderNormalBlock, reflectParams::constrictReflectionAngles);
}
//printTime();
//150ms
ascanBlock.ascanBlock = removeDataFromArrays(ascanBlock.ascanBlock, usedData);
result.mpBlock = removeDataFromArrays(ascanBlock.mpBlock, usedData);
result.slBlock = removeDataFromArrays(ascanBlock.slBlock, usedData);
@@ -51,7 +76,8 @@ AscanBlockPreprocessed Recon::getAscanBlockPreprocessed(Parser* aParser, const A
result.senderPositionBlock = removeDataFromArrays(geometryBlock.senderPositionBlock, usedData);
result.receiverPositionBlock = removeDataFromArrays(geometryBlock.receiverPositionBlock, usedData);
result.gainBlock = removeDataFromArrays(ascanBlock.gainBlock, usedData);
//printTime();
//120ms
}
if (ascanBlock.ascanBlock.getDataSize() > 0)
@@ -62,6 +88,72 @@ AscanBlockPreprocessed Recon::getAscanBlockPreprocessed(Parser* aParser, const A
{
result.ascanBlockPreprocessed = ascanBlock.ascanBlock;
}
//printTime();
return result;
}
}
AscanBlockPreprocessedCuda Recon::getAscanBlockPreprocessedCuda(Parser* aParser, const Aurora::Matrix& aMp, const Aurora::Matrix& aSl, const Aurora::Matrix& aSn,
const Aurora::Matrix& aRl, const Aurora::Matrix& aRn, GeometryInfo& aGeom, const MeasurementInfo& aMeasInfo,
bool aApplyFilter, bool aTransReco)
{
//std::cout<<"strart"<<std::endl;
//printTime();
//550ms
AscanBlockPreprocessedCuda result;
AscanBlock ascanBlock = getAscanBlock(aParser, aMp, aSl, aSn, aRl, aRn);
//printTime();
//300ms
result.ascanBlockPreprocessed = ascanBlock.ascanBlock.toDeviceMatrix();
result.gainBlock = ascanBlock.gainBlock.toDeviceMatrix();
result.mpBlock = ascanBlock.mpBlock;
result.rlBlock = ascanBlock.rlBlock;
result.rnBlock = ascanBlock.rnBlock;
result.slBlock = ascanBlock.slBlock;
result.snBlock = ascanBlock.snBlock;
GeometryBlock geometryBlock = blockingGeometryInfos(aGeom, ascanBlock.rnBlock, ascanBlock.rlBlock, ascanBlock.snBlock, ascanBlock.slBlock, ascanBlock.mpBlock);
//printTime();
//3ms
result.receiverPositionBlock = geometryBlock.receiverPositionBlock;
result.senderPositionBlock = geometryBlock.senderPositionBlock;
if(aApplyFilter)
{
Matrix usedData;
if(aTransReco)
{
usedData = filterTransmissionData(ascanBlock.slBlock, ascanBlock.snBlock, ascanBlock.rlBlock, ascanBlock.rnBlock,
aGeom.sensData, geometryBlock.senderNormalBlock, geometryBlock.receiverNormalBlock);
}
else
{
usedData = filterReflectionData(geometryBlock.receiverPositionBlock, geometryBlock.senderPositionBlock, geometryBlock.senderNormalBlock, reflectParams::constrictReflectionAngles);
}
//printTime();
//40ms
CudaMatrix usedDataDevice = usedData.toDeviceMatrix();
result.ascanBlockPreprocessed = valid(result.ascanBlockPreprocessed, usedDataDevice);
result.mpBlock = removeDataFromArrays(ascanBlock.mpBlock, usedData);
result.slBlock = removeDataFromArrays(ascanBlock.slBlock, usedData);
result.snBlock = removeDataFromArrays(ascanBlock.snBlock, usedData);
result.rlBlock = removeDataFromArrays(ascanBlock.rlBlock, usedData);
result.rnBlock = removeDataFromArrays(ascanBlock.rnBlock, usedData);
result.senderPositionBlock = removeDataFromArrays(geometryBlock.senderPositionBlock, usedData);
result.receiverPositionBlock = removeDataFromArrays(geometryBlock.receiverPositionBlock, usedData);
result.gainBlock = valid(result.gainBlock, usedDataDevice);
//printTime();
//10ms
}
if (ascanBlock.ascanBlock.getDataSize() > 0)
{
result.ascanBlockPreprocessed = preprocessAscanBlockCuda(result.ascanBlockPreprocessed, aMeasInfo);
}
// else
// {
// result.ascanBlockPreprocessed = ascanBlock.ascanBlock;
// }
//printTime();
//std::cout<<"end"<<std::endl;
return result;
}

18
src/common/dataBlockCreation/getAScanBlockPreprocessed.h
View File

@@ -2,6 +2,7 @@
#define GETASCANBLOCK_PREPROCESSED_H
#include "Matrix.h"
#include "CudaMatrix.h"
#include "src/common/getGeometryInfo.h"
#include "src/common/getMeasurementMetaData.h"
@@ -25,6 +26,23 @@ namespace Recon
AscanBlockPreprocessed getAscanBlockPreprocessed(Parser* aParser, const Aurora::Matrix& aMp, const Aurora::Matrix& aSl, const Aurora::Matrix& aSn,
const Aurora::Matrix& aRl, const Aurora::Matrix& aRn, GeometryInfo& aGeom, const MeasurementInfo& aMeasInfo,
bool aApplyFilter, bool aTransReco);
struct AscanBlockPreprocessedCuda
{
Aurora::CudaMatrix ascanBlockPreprocessed;
Aurora::Matrix mpBlock;
Aurora::Matrix slBlock;
Aurora::Matrix snBlock;
Aurora::Matrix rlBlock;
Aurora::Matrix rnBlock;
Aurora::Matrix senderPositionBlock;
Aurora::Matrix receiverPositionBlock;
Aurora::CudaMatrix gainBlock;
};
AscanBlockPreprocessedCuda getAscanBlockPreprocessedCuda(Parser* aParser, const Aurora::Matrix& aMp, const Aurora::Matrix& aSl, const Aurora::Matrix& aSn,
const Aurora::Matrix& aRl, const Aurora::Matrix& aRn, GeometryInfo& aGeom, const MeasurementInfo& aMeasInfo,
bool aApplyFilter, bool aTransReco);
}

6
src/common/dataBlockCreation/getAscanBlock.cpp
View File

@@ -70,13 +70,13 @@ AscanBlock Recon::getAscanBlock(Parser* aParser, const Aurora::Matrix& aMp, cons
for(int slIndex=0; slIndex<aSl.getDataSize();++slIndex)
{
OneTasAScanData oneTasData = aParser->getOneTasAscanDataOfMotorPosition(aSl[slIndex]);
#pragma omp parallel for
for(int snIndex=0; snIndex<aSn.getDataSize();++snIndex)
{
//int mapperIndex = 0;
#pragma omp parallel for
//#pragma omp parallel for
for(int rlIndex=0; rlIndex<aRl.getDataSize();++rlIndex)
{
{
for(int rnIndex=0; rnIndex<aRn.getDataSize(); ++rnIndex)
{
size_t mapperIndex = rnIndex + rlIndex*aRn.getDataSize();

4
src/common/fileHelper.h
View File

@@ -6,8 +6,8 @@
namespace Recon
{
const std::string DEFAULT_CONFIG_PATH = "/home/UR/ConfigFiles/";
const std::string DEFAULT_OUTPUT_PATH = "/home/UR/ReconResult/USCT_Result.mat";
const std::string DEFAULT_OUTPUT_FILENAME = "USCT_Result.mat";
const std::string DEFAULT_OUTPUT_PATH = "/home/UR/ReconResult/";
const std::string DEFAULT_OUTPUT_FILENAME = "sun.mat";
std::string getPath(const std::string &aFullPath);
bool endsWithMat(const std::string &aStr);

11
src/common/preprocessAscanBlock.cpp
View File

@@ -1,5 +1,6 @@
#include "preprocessAscanBlock.h"
#include "Function1D.h"
#include "Function1D.cuh"
#include <cstddef>
Aurora::Matrix Recon::preprocessAscanBlock(const Aurora::Matrix& aAscans, const MeasurementInfo& aMeasInfo)
@@ -20,4 +21,14 @@ Aurora::Matrix Recon::preprocessAscanBlock(const Aurora::Matrix& aAscans, const
// end
return result;
}
Aurora::CudaMatrix Recon::preprocessAscanBlockCuda(const Aurora::CudaMatrix& aAscans, const MeasurementInfo& aMeasInfo)
{
if(aMeasInfo.ascanDataType == "float16")
{
return Aurora::convertfp16tofloatCuda(aAscans, aAscans.getDimSize(0), aAscans.getDimSize(1));
}
return aAscans;
}

2
src/common/preprocessAscanBlock.h
View File

@@ -7,6 +7,8 @@
namespace Recon
{
Aurora::Matrix preprocessAscanBlock(const Aurora::Matrix& aAscans, const MeasurementInfo& aMeasInfo);
Aurora::CudaMatrix preprocessAscanBlockCuda(const Aurora::CudaMatrix& aAscans, const MeasurementInfo& aMeasInfo);
}
#endif