Compare commits
4 Commits
6ec4a65117
...
35eaf08b5e
| Author | SHA1 | Date | |
|---|---|---|---|
| 35eaf08b5e | |||
| be2aaf594a | |||
| a12b9a6bec | |||
| 9d77f828cc |
2
.gitignore
vendored
Normal file
2
.gitignore
vendored
Normal file
@ -0,0 +1,2 @@
|
||||
# Assessments
|
||||
assessments/
|
||||
8
homework_1/.gitignore
vendored
8
homework_1/.gitignore
vendored
@ -9,8 +9,16 @@ mtx/
|
||||
exclude
|
||||
hpc_auth_sync.sh
|
||||
|
||||
# IDEs
|
||||
.idea/
|
||||
.clangd
|
||||
|
||||
# eclipse
|
||||
.project
|
||||
.cproject
|
||||
.settings/
|
||||
|
||||
.vs/
|
||||
.vscode/
|
||||
|
||||
|
||||
|
||||
@ -27,8 +27,8 @@ SRC_DIR_LIST := src gtest
|
||||
# Include directories list(space seperated). Makefile-relative path.
|
||||
INC_DIR_LIST := inc \
|
||||
src \
|
||||
/usr/include/hdf5/serial/ \
|
||||
gtest \
|
||||
/usr/include/hdf5/serial/
|
||||
# Libs/MATLAB/R2019b/include/ \
|
||||
|
||||
# Exclude files list(space seperated). Filenames only.
|
||||
@ -179,7 +179,7 @@ hpc-clean:
|
||||
rm hpc-results/post
|
||||
|
||||
#
|
||||
# ================ Local via docker build rules =================
|
||||
# ================ Local (and/or) via docker build rules =================
|
||||
#
|
||||
# examples:
|
||||
# make IMAGE=hpcimage v0
|
||||
@ -190,12 +190,6 @@ local_v0: CXXFLAGS := $(DEB_CXXFLAGS) -DCODE_VERSION=0
|
||||
local_v0: TARGET := local_v0
|
||||
local_v0: $(BUILD_DIR)/$(TARGET)
|
||||
cp $(BUILD_DIR)/$(TARGET) out/$(TARGET)
|
||||
|
||||
local_v0_opt: CFLAGS := $(REL_CFLAGS) -DCODE_VERSION=0
|
||||
local_v0_opt: CXXFLAGS := $(REL_CXXFLAGS) -DCODE_VERSION=0
|
||||
local_v0_opt: TARGET := local_v0_opt
|
||||
local_v0_opt: $(BUILD_DIR)/$(TARGET)
|
||||
cp $(BUILD_DIR)/$(TARGET) out/$(TARGET)
|
||||
|
||||
local_v1: CFLAGS := $(DEB_CFLAGS) -DCODE_VERSION=1
|
||||
local_v1: CXXFLAGS := $(DEB_CXXFLAGS) -DCODE_VERSION=1
|
||||
@ -203,6 +197,21 @@ local_v1: TARGET := local_v1
|
||||
local_v1: $(BUILD_DIR)/$(TARGET)
|
||||
cp $(BUILD_DIR)/$(TARGET) out/$(TARGET)
|
||||
|
||||
local_v1_omp: CFLAGS := $(DEB_CFLAGS) -fopenmp -DCODE_VERSION=1 -DOMP
|
||||
local_v1_omp: CXXFLAGS := $(DEB_CXXFLAGS) -fopenmp -DCODE_VERSION=1 -DOMP
|
||||
local_v1_omp: LDFLAGS += -fopenmp
|
||||
local_v1_omp: TARGET := local_v1_omp
|
||||
local_v1_omp: $(BUILD_DIR)/$(TARGET)
|
||||
cp $(BUILD_DIR)/$(TARGET) out/$(TARGET)
|
||||
|
||||
|
||||
local_v1_pth: CFLAGS := $(DEB_CFLAGS) -DCODE_VERSION=1 -DPTHREADS
|
||||
local_v1_pth: CXXFLAGS := $(DEB_CXXFLAGS) -DCODE_VERSION=1 -DPTHREADS
|
||||
local_v1_pth: TARGET := local_v1_pth
|
||||
local_v1_pth: $(BUILD_DIR)/$(TARGET)
|
||||
cp $(BUILD_DIR)/$(TARGET) out/$(TARGET)
|
||||
|
||||
|
||||
v0: DOCKER := $(DOCKER_CMD)
|
||||
v0: CFLAGS := $(REL_CFLAGS) -DCODE_VERSION=0
|
||||
v0: CXXFLAGS := $(REL_CXXFLAGS) -DCODE_VERSION=0
|
||||
@ -215,7 +224,7 @@ v1_cilk: CXX := /usr/local/OpenCilk-9.0.1-Linux/bin/clang++
|
||||
v1_cilk: CFLAGS := $(REL_CFLAGS) -fcilkplus -DCODE_VERSION=1 -DCILK
|
||||
v1_cilk: CXXFLAGS := $(REL_CXXFLAGS) -fcilkplus -DCODE_VERSION=1 -DCILK
|
||||
v1_cilk: LDFLAGS += -fcilkplus
|
||||
v1_cilk: TARGET := knnsearch_cilkv1
|
||||
v1_cilk: TARGET := knnsearch_v1_cilk
|
||||
v1_cilk: $(BUILD_DIR)/$(TARGET)
|
||||
cp $(BUILD_DIR)/$(TARGET) out/$(TARGET)
|
||||
|
||||
@ -223,10 +232,17 @@ v1_omp: DOCKER := $(DOCKER_CMD)
|
||||
v1_omp: CFLAGS := $(REL_CFLAGS) -fopenmp -DCODE_VERSION=1 -DOMP
|
||||
v1_omp: CXXFLAGS := $(REL_CXXFLAGS) -fopenmp -DCODE_VERSION=1 -DOMP
|
||||
v1_omp: LDFLAGS += -fopenmp
|
||||
v1_omp: TARGET := knnsearch_ompv1
|
||||
v1_omp: TARGET := knnsearch_v1_omp
|
||||
v1_omp: $(BUILD_DIR)/$(TARGET)
|
||||
cp $(BUILD_DIR)/$(TARGET) out/$(TARGET)
|
||||
|
||||
v1_pth: DOCKER := $(DOCKER_CMD)
|
||||
v1_pth: CFLAGS := $(REL_CFLAGS) -DCODE_VERSION=1 -DPTHREADS
|
||||
v1_pth: CXXFLAGS := $(REL_CXXFLAGS) -DCODE_VERSION=1 -DPTHREADS
|
||||
v1_pth: TARGET := knnsearch_v1_pth
|
||||
v1_pth: $(BUILD_DIR)/$(TARGET)
|
||||
cp $(BUILD_DIR)/$(TARGET) out/$(TARGET)
|
||||
|
||||
v1: DOCKER := $(DOCKER_CMD)
|
||||
v1: CFLAGS := $(REL_CFLAGS) -DCODE_VERSION=1
|
||||
v1: CXXFLAGS := $(REL_CXXFLAGS) -DCODE_VERSION=1
|
||||
@ -240,7 +256,12 @@ tests: TARGET := tests
|
||||
tests: $(BUILD_DIR)/$(TARGET)
|
||||
cp $(BUILD_DIR)/$(TARGET) out/$(TARGET)
|
||||
|
||||
|
||||
tests_rel: CFLAGS := $(REL_CFLAGS) -DCODE_VERSION=0 -DTESTING
|
||||
tests_rel: CXXFLAGS := $(REL_CXXFLAGS) -DCODE_VERSION=0 -DTESTING
|
||||
tests_rel: TARGET := tests
|
||||
tests_rel: $(BUILD_DIR)/$(TARGET)
|
||||
cp $(BUILD_DIR)/$(TARGET) out/$(TARGET)
|
||||
|
||||
#
|
||||
# ========= Inside CSAL Image build rules ===========
|
||||
#
|
||||
|
||||
@ -19,6 +19,8 @@
|
||||
|
||||
using matrix_t = mtx::Matrix<int>;
|
||||
|
||||
extern void loadMtx(MatrixDst& Corpus, MatrixDst& Query);
|
||||
extern void storeMtx(MatrixIdx& Idx, MatrixDst& Dst);
|
||||
|
||||
// =====================================
|
||||
// C1, Q1
|
||||
@ -140,11 +142,44 @@ TEST(Tv0_UT, pdist2_test2) {
|
||||
}
|
||||
|
||||
|
||||
TEST(Tv0_UT, pdist2_test3) {
|
||||
|
||||
mtx::Matrix<double> D2_exp(16, 16, {
|
||||
0, 0.7433, 0.6868, 0.8846, 0.6342, 0.4561, 0.5118, 0.6341, 0.5461, 0.7322, 0.6974, 0.4330, 0.7028, 0.6303, 0.6826, 0.4179,
|
||||
0.7433, 0, 0.3400, 0.4555, 0.4207, 0.9736, 0.9690, 0.7386, 1.1055, 0.5462, 0.5345, 0.6576, 0.8677, 1.0291, 0.5393, 0.8106,
|
||||
0.6868, 0.3400, 0, 0.5380, 0.6268, 0.9512, 1.0234, 0.8403, 0.9843, 0.8187, 0.3091, 0.7829, 0.5759, 0.9411, 0.7239, 0.9186,
|
||||
0.8846, 0.4555, 0.5380, 0, 0.6796, 1.1672, 1.0460, 1.1016, 1.1139, 0.7542, 0.6480, 0.9304, 1.0568, 1.3482, 0.8316, 0.9750,
|
||||
0.6342, 0.4207, 0.6268, 0.6796, 0, 0.9267, 0.8772, 0.4847, 0.9317, 0.4093, 0.8351, 0.4215, 0.9736, 0.9007, 0.5999, 0.5291,
|
||||
0.4561, 0.9736, 0.9512, 1.1672, 0.9267, 0, 0.3903, 0.7795, 0.9308, 0.8429, 0.8436, 0.5672, 0.9284, 0.7064, 0.6435, 0.5975,
|
||||
0.5118, 0.9690, 1.0234, 1.0460, 0.8772, 0.3903, 0, 0.8920, 0.9253, 0.7060, 0.9427, 0.5728, 1.1515, 0.9907, 0.6471, 0.4811,
|
||||
0.6341, 0.7386, 0.8403, 1.1016, 0.4847, 0.7795, 0.8920, 0, 0.9824, 0.6416, 0.9844, 0.3398, 0.9355, 0.5428, 0.6536, 0.5309,
|
||||
0.5461, 1.1055, 0.9843, 1.1139, 0.9317, 0.9308, 0.9253, 0.9824, 0, 1.1517, 1.0541, 0.8746, 0.8506, 0.8777, 1.2036, 0.7607,
|
||||
0.7322, 0.5462, 0.8187, 0.7542, 0.4093, 0.8429, 0.7060, 0.6416, 1.1517, 0, 0.9106, 0.4245, 1.2071, 1.0738, 0.3745, 0.5170,
|
||||
0.6974, 0.5345, 0.3091, 0.6480, 0.8351, 0.8436, 0.9427, 0.9844, 1.0541, 0.9106, 0, 0.8647, 0.5941, 0.9954, 0.7148, 0.9876,
|
||||
0.4330, 0.6576, 0.7829, 0.9304, 0.4215, 0.5672, 0.5728, 0.3398, 0.8746, 0.4245, 0.8647, 0, 0.9590, 0.6782, 0.4586, 0.2525,
|
||||
0.7028, 0.8677, 0.5759, 1.0568, 0.9736, 0.9284, 1.1515, 0.9355, 0.8506, 1.2071, 0.5941, 0.9590, 0, 0.6838, 1.0517, 1.0675,
|
||||
0.6303, 1.0291, 0.9411, 1.3482, 0.9007, 0.7064, 0.9907, 0.5428, 0.8777, 1.0738, 0.9954, 0.6782, 0.6838, 0, 0.9482, 0.7937,
|
||||
0.6826, 0.5393, 0.7239, 0.8316, 0.5999, 0.6435, 0.6471, 0.6536, 1.2036, 0.3745, 0.7148, 0.4586, 1.0517, 0.9482, 0, 0.6345,
|
||||
0.4179, 0.8106, 0.9186, 0.9750, 0.5291, 0.5975, 0.4811, 0.5309, 0.7607, 0.5170, 0.9876, 0.2525, 1.0675, 0.7937, 0.6345, 0
|
||||
});
|
||||
|
||||
mtx::Matrix<double> D (16,16);
|
||||
|
||||
v0::pdist2(C2, C2, D);
|
||||
|
||||
for (size_t i = 0 ; i< D.rows() ; ++i)
|
||||
for (size_t j = 0 ; j<D.columns() ; ++j) {
|
||||
EXPECT_EQ (D2_exp.get(i ,j) + 0.01 > D(i, j), true);
|
||||
EXPECT_EQ (D2_exp.get(i ,j) - 0.01 < D(i, j), true);
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
/*
|
||||
* ==========================================
|
||||
* v0::knn
|
||||
*/
|
||||
TEST(Tv0_UT, knn_test1) {
|
||||
TEST(Tv0_UT, knn_v0_test1) {
|
||||
size_t k = 3;
|
||||
mtx::Matrix<uint32_t> Idx_exp(5, k, {
|
||||
5, 8, 9,
|
||||
@ -177,7 +212,7 @@ TEST(Tv0_UT, knn_test1) {
|
||||
|
||||
}
|
||||
|
||||
TEST(Tv0_UT, knn_test2) {
|
||||
TEST(Tv0_UT, knn_v0_test2) {
|
||||
size_t k = 3;
|
||||
mtx::Matrix<uint32_t> Idx_exp(8, k, {
|
||||
14, 13, 1,
|
||||
@ -220,7 +255,7 @@ TEST(Tv0_UT, knn_test2) {
|
||||
* ==========================================
|
||||
* v1::knn
|
||||
*/
|
||||
TEST(Tv1_UT, knn_test1) {
|
||||
TEST(Tv1_UT, knn_v1_1slice) {
|
||||
size_t k = 3;
|
||||
mtx::Matrix<uint32_t> Idx_exp(8, k, {
|
||||
14, 13, 1,
|
||||
@ -247,7 +282,46 @@ TEST(Tv1_UT, knn_test1) {
|
||||
mtx::Matrix<uint32_t> Idx(8, k);
|
||||
mtx::Matrix<double> Dst(8, k);
|
||||
|
||||
v1::knnsearch(C2, Q2, 0, k, k, Idx, Dst);
|
||||
v1::knnsearch(C2, Q2, 1, k, k, Idx, Dst);
|
||||
|
||||
|
||||
for (size_t i = 0 ; i< Idx.rows() ; ++i)
|
||||
for (size_t j = 0 ; j<Idx.columns() ; ++j) {
|
||||
EXPECT_EQ (Idx_exp(i ,j) == Idx(i, j) + 1, true); // matlab starts from 1
|
||||
EXPECT_EQ (Dst_exp.get(i ,j) + 0.01 > Dst(i, j), true);
|
||||
EXPECT_EQ (Dst_exp.get(i ,j) - 0.01 < Dst(i, j), true);
|
||||
}
|
||||
|
||||
}
|
||||
|
||||
TEST(Tv1_UT, knn_v1_2slice) {
|
||||
size_t k = 3;
|
||||
mtx::Matrix<uint32_t> Idx_exp(8, k, {
|
||||
14, 13, 1,
|
||||
15, 10, 12,
|
||||
14, 8, 12,
|
||||
4, 1, 3,
|
||||
8, 12, 5,
|
||||
4, 3, 2,
|
||||
10, 2, 4,
|
||||
1, 11, 9
|
||||
});
|
||||
|
||||
mtx::Matrix<double> Dst_exp(8, k, {
|
||||
0.1939, 0.5768, 0.6020,
|
||||
0.2708, 0.3686, 0.3740,
|
||||
0.2684, 0.3103, 0.5277,
|
||||
0.4709, 0.6050, 0.6492,
|
||||
0.4397, 0.6842, 0.7074,
|
||||
0.3402, 0.4360, 0.4475,
|
||||
0.3708, 0.4037, 0.4417,
|
||||
0.6352, 0.6653, 0.6758
|
||||
});
|
||||
|
||||
mtx::Matrix<uint32_t> Idx(8, k);
|
||||
mtx::Matrix<double> Dst(8, k);
|
||||
|
||||
v1::knnsearch(C2, Q2, 2, k, k, Idx, Dst);
|
||||
|
||||
|
||||
for (size_t i = 0 ; i< Idx.rows() ; ++i)
|
||||
@ -260,7 +334,7 @@ TEST(Tv1_UT, knn_test1) {
|
||||
}
|
||||
|
||||
// all-to-all
|
||||
TEST(Tv1_UT, knn_test2) {
|
||||
TEST(Tv1_UT, knn_v1_4slice) {
|
||||
size_t k = 3;
|
||||
mtx::Matrix<uint32_t> Idx_exp(16, k, {
|
||||
1, 16, 12,
|
||||
@ -303,7 +377,7 @@ TEST(Tv1_UT, knn_test2) {
|
||||
mtx::Matrix<uint32_t> Idx(16, k);
|
||||
mtx::Matrix<double> Dst(16, k);
|
||||
|
||||
v1::knnsearch(C2, C2, 0, k, k, Idx, Dst);
|
||||
v1::knnsearch(C2, C2, 4, k, k, Idx, Dst);
|
||||
|
||||
|
||||
for (size_t i = 0 ; i< Idx.rows() ; ++i)
|
||||
@ -315,3 +389,130 @@ TEST(Tv1_UT, knn_test2) {
|
||||
|
||||
}
|
||||
|
||||
|
||||
|
||||
/*
|
||||
* ============== Live hdf5 tests ===============
|
||||
*
|
||||
* In order to run these test we need the followin hdf5 files in ./mtx directory:
|
||||
*
|
||||
* - fasion-mnist-784-euclidean.hdf5
|
||||
* - mnist-784-euclidean.hdf5
|
||||
* - sift-128-euclidean.hdf5
|
||||
* - gist-960-euclidean.hdf5
|
||||
*
|
||||
*/
|
||||
|
||||
TEST(Tlive_UT, knn_v0_sift_test) {
|
||||
// Instantiate matrixes
|
||||
MatrixDst Corpus;
|
||||
MatrixDst Query;
|
||||
MatrixIdx Idx;
|
||||
MatrixDst Dst;
|
||||
|
||||
// setup environment
|
||||
session.corpusMtxFile = "mtx/sift-128-euclidean.hdf5";
|
||||
session.corpusDataSet = "/test";
|
||||
session.queryMtx = false;
|
||||
session.k = 100;
|
||||
size_t m = session.k;
|
||||
session.timing = true;
|
||||
session.outMtxFile = "test/knn_v0.hdf5";
|
||||
|
||||
|
||||
loadMtx(Corpus, Query);
|
||||
|
||||
// Prepare output memory (There is no Query, so from Corpus
|
||||
Idx.resize(Corpus.rows(), session.k);
|
||||
Dst.resize(Corpus.rows(), session.k);
|
||||
|
||||
v0::knnsearch(Corpus, Corpus, 0, session.k, m, Idx, Dst);
|
||||
storeMtx(Idx, Dst);
|
||||
EXPECT_EQ(true, true);
|
||||
}
|
||||
|
||||
|
||||
TEST(Tlive_UT, knn_v1_sift_test_1slice) {
|
||||
// Instantiate matrixes
|
||||
MatrixDst Corpus;
|
||||
MatrixDst Query;
|
||||
MatrixIdx Idx;
|
||||
MatrixDst Dst;
|
||||
|
||||
// setup environment
|
||||
session.corpusMtxFile = "mtx/sift-128-euclidean.hdf5";
|
||||
session.corpusDataSet = "/test";
|
||||
session.queryMtx = false;
|
||||
session.k = 100;
|
||||
size_t m = session.k;
|
||||
session.timing = true;
|
||||
session.outMtxFile = "test/knn_v1ser.hdf5";
|
||||
|
||||
|
||||
loadMtx(Corpus, Query);
|
||||
|
||||
// Prepare output memory (There is no Query, so from Corpus
|
||||
Idx.resize(Corpus.rows(), session.k);
|
||||
Dst.resize(Corpus.rows(), session.k);
|
||||
|
||||
v1::knnsearch(Corpus, Corpus, 0, session.k, m, Idx, Dst);
|
||||
storeMtx(Idx, Dst);
|
||||
EXPECT_EQ(true, true);
|
||||
}
|
||||
|
||||
TEST(Tlive_UT, knn_v1_sift_test_2slice) {
|
||||
// Instantiate matrixes
|
||||
MatrixDst Corpus;
|
||||
MatrixDst Query;
|
||||
MatrixIdx Idx;
|
||||
MatrixDst Dst;
|
||||
|
||||
// setup environment
|
||||
session.corpusMtxFile = "mtx/sift-128-euclidean.hdf5";
|
||||
session.corpusDataSet = "/test";
|
||||
session.queryMtx = false;
|
||||
session.k = 100;
|
||||
size_t m = session.k;
|
||||
session.timing = true;
|
||||
session.outMtxFile = "test/knn_v1ser.hdf5";
|
||||
|
||||
|
||||
loadMtx(Corpus, Query);
|
||||
|
||||
// Prepare output memory (There is no Query, so from Corpus
|
||||
Idx.resize(Corpus.rows(), session.k);
|
||||
Dst.resize(Corpus.rows(), session.k);
|
||||
|
||||
v1::knnsearch(Corpus, Corpus, 2, session.k, m, Idx, Dst);
|
||||
storeMtx(Idx, Dst);
|
||||
EXPECT_EQ(true, true);
|
||||
}
|
||||
|
||||
TEST(Tlive_UT, knn_v1_sift_test_4slice) {
|
||||
// Instantiate matrixes
|
||||
MatrixDst Corpus;
|
||||
MatrixDst Query;
|
||||
MatrixIdx Idx;
|
||||
MatrixDst Dst;
|
||||
|
||||
// setup environment
|
||||
session.corpusMtxFile = "mtx/sift-128-euclidean.hdf5";
|
||||
session.corpusDataSet = "/test";
|
||||
session.queryMtx = false;
|
||||
session.k = 100;
|
||||
size_t m = session.k;
|
||||
session.timing = true;
|
||||
session.outMtxFile = "test/knn_v1ser.hdf5";
|
||||
|
||||
|
||||
loadMtx(Corpus, Query);
|
||||
|
||||
// Prepare output memory (There is no Query, so from Corpus
|
||||
Idx.resize(Corpus.rows(), session.k);
|
||||
Dst.resize(Corpus.rows(), session.k);
|
||||
|
||||
v1::knnsearch(Corpus, Corpus, 4, session.k, m, Idx, Dst);
|
||||
storeMtx(Idx, Dst);
|
||||
EXPECT_EQ(true, true);
|
||||
}
|
||||
|
||||
@ -51,11 +51,13 @@ struct session_t {
|
||||
std::string queryMtxFile {}; //!< optional query matrix file name in HDF5 format
|
||||
std::string queryDataSet {}; //!< optional query dataset name in HDF5 matrix file
|
||||
bool queryMtx {false}; //!< Flag to indicate that there is a separate query matrix
|
||||
size_t k {1}; //!< The number of nearest neighbors to find
|
||||
size_t k {1UL}; //!< The number of nearest neighbors to find
|
||||
std::string outMtxFile {"out.hdf5"}; //!< output matrix file name in HDF5 format
|
||||
std::string outMtxIdxDataSet {"/Idx"}; //!< Index output dataset name in HDF5 matrix file
|
||||
std::string outMtxDstDataSet {"/Dst"}; //!< Distance output dataset name in HDF5 matrix file
|
||||
std::size_t max_threads {}; //!< Maximum threads to use
|
||||
std::size_t max_threads {0UL}; //!< Maximum threads to use
|
||||
std::size_t slices {0UL}; //!< Slices/threads to use
|
||||
std::size_t accuracy {100UL}; //!< The neighbor finding accuracy
|
||||
bool timing {false}; //!< Enable timing prints of the program
|
||||
bool verbose {false}; //!< Flag to enable verbose output to stdout
|
||||
};
|
||||
|
||||
@ -134,6 +134,9 @@ struct Matrix {
|
||||
Matrix& operator=(Matrix&& m) noexcept { moves(std::move(m)); return *this; }
|
||||
Matrix(const Matrix& m) = delete; //!< No copy ctor
|
||||
Matrix& operator=(const Matrix& m) = delete; //!< No copy
|
||||
//Matrix(const Matrix& m);
|
||||
//Matrix& operator=(const Matrix& m) { copy(m); }
|
||||
|
||||
//! @}
|
||||
|
||||
//! \name Data exposure
|
||||
@ -233,6 +236,11 @@ struct Matrix {
|
||||
|
||||
// a basic serial iterator support
|
||||
DataType* data() noexcept { return data_; }
|
||||
DataType* begin() noexcept { return data_; }
|
||||
const DataType* begin() const noexcept { return data_; }
|
||||
DataType* end() noexcept { return data_ + capacity(rows_, cols_); }
|
||||
const DataType* end() const noexcept { return data_ + capacity(rows_, cols_); }
|
||||
|
||||
// IndexType begin_idx() noexcept { return 0; }
|
||||
// IndexType end_idx() noexcept { return capacity(rows_, cols_); }
|
||||
|
||||
@ -265,17 +273,19 @@ struct Matrix {
|
||||
std::swap(rows_, src.rows_);
|
||||
std::swap(cols_, src.cols_);
|
||||
}
|
||||
|
||||
private:
|
||||
//! move helper
|
||||
void moves(Matrix&& src) noexcept {
|
||||
data_ = std::move(src.vector_storage_);
|
||||
data_ = std::move(src.raw_storage_);
|
||||
data_ = std::move(src.data_);
|
||||
data_ = std::move(src.use_vector_);
|
||||
rows_ = std::move(src.rows_);
|
||||
cols_ = std::move(src.cols_);
|
||||
vector_storage_ = std::move(src.vector_storage_);
|
||||
raw_storage_ = std::move(src.raw_storage_);
|
||||
data_ = std::move(src.data_);
|
||||
use_vector_ = std::move(src.use_vector_);
|
||||
rows_ = std::move(src.rows_);
|
||||
cols_ = std::move(src.cols_);
|
||||
}
|
||||
|
||||
// Storage
|
||||
std::vector<DataType>
|
||||
vector_storage_; //!< Internal storage (if used).
|
||||
DataType* raw_storage_; //!< External storage (if used).
|
||||
@ -528,125 +538,6 @@ private:
|
||||
};
|
||||
|
||||
|
||||
template<typename ...> struct Matrix_view { };
|
||||
|
||||
/*!
|
||||
* @struct Matrix_view
|
||||
* @tparam MatrixType
|
||||
*/
|
||||
template<template <typename, typename, MatrixType, MatrixOrder, bool> class Matrix,
|
||||
typename DataType,
|
||||
typename IndexType,
|
||||
MatrixType Type,
|
||||
MatrixOrder Order>
|
||||
struct Matrix_view<Matrix<DataType, IndexType, Type, Order, false>> {
|
||||
using owner_t = Matrix<DataType, IndexType, Type, Order, false>;
|
||||
|
||||
using dataType = DataType; //!< meta:export of underling data type
|
||||
using indexType = IndexType; //!< meta:export of underling index type
|
||||
static constexpr MatrixOrder matrixOrder = Order; //!< meta:export of array order
|
||||
static constexpr MatrixType matrixType = Type; //!< meta:export of array type
|
||||
|
||||
/*!
|
||||
* \name Obj lifetime
|
||||
*/
|
||||
//! @{
|
||||
|
||||
//! Construct a matrix view to entire matrix
|
||||
Matrix_view(const owner_t* owner) noexcept :
|
||||
owner_(owner), m_(owner->data()), rows_(owner->rows()), cols_(owner->columns()) { }
|
||||
|
||||
Matrix_view(const owner_t* owner, IndexType begin, IndexType end) noexcept :
|
||||
owner_(owner) {
|
||||
if constexpr (Order == MatrixOrder::ROWMAJOR) {
|
||||
m_ = owner->data() + begin * owner->columns();
|
||||
rows_ = end - begin;
|
||||
cols_ = owner->columns();
|
||||
} else if (Order == MatrixOrder::COLMAJOR) {
|
||||
m_ = owner->data() + begin * owner->rows();
|
||||
rows_ = owner->rows();
|
||||
cols_ = end - begin;
|
||||
}
|
||||
}
|
||||
|
||||
Matrix_view(Matrix_view&& m) = delete; //! No move
|
||||
Matrix_view& operator=(Matrix_view&& m) = delete;
|
||||
Matrix_view(const Matrix_view& m) = delete; //!< No copy
|
||||
Matrix_view& operator=(const Matrix_view& m) = delete;
|
||||
//! @}
|
||||
|
||||
//! Get/Set the size of each dimension
|
||||
const IndexType rows() const noexcept { return rows_; }
|
||||
const IndexType columns() const noexcept { return cols_; }
|
||||
|
||||
//! Get the interface size of the Matrix (what appears to be the size)
|
||||
IndexType size() const {
|
||||
return rows_ * cols_;
|
||||
}
|
||||
|
||||
//! Actual memory capacity of the symmetric matrix
|
||||
static constexpr IndexType capacity(IndexType M, IndexType N) {
|
||||
return M*N;
|
||||
}
|
||||
/*
|
||||
* virtual 2D accessors
|
||||
*/
|
||||
const DataType get (IndexType i, IndexType j) const {
|
||||
if constexpr (Order == MatrixOrder::COLMAJOR)
|
||||
return m_[i + j*rows_];
|
||||
else
|
||||
return m_[i*cols_ + j];
|
||||
}
|
||||
|
||||
DataType set (DataType v, IndexType i, IndexType j) {
|
||||
if constexpr (Order == MatrixOrder::COLMAJOR)
|
||||
return m_[i + j*rows_] = v;
|
||||
else
|
||||
return m_[i*cols_ + j] = v;
|
||||
}
|
||||
// DataType operator()(IndexType i, IndexType j) { return get(i, j); }
|
||||
/*!
|
||||
* Return a proxy MatVal object with read and write capabilities.
|
||||
* @param i The row number
|
||||
* @param j The column number
|
||||
* @return tHE MatVal object
|
||||
*/
|
||||
MatVal<Matrix_view> operator()(IndexType i, IndexType j) noexcept {
|
||||
return MatVal<Matrix_view>(this, get(i, j), i, j);
|
||||
}
|
||||
|
||||
// a basic serial iterator support
|
||||
DataType* data() noexcept { return m_.data(); }
|
||||
// IndexType begin_idx() noexcept { return 0; }
|
||||
// IndexType end_idx() noexcept { return capacity(rows_, cols_); }
|
||||
|
||||
const DataType* data() const noexcept { return m_; }
|
||||
const IndexType begin_idx() const noexcept { return 0; }
|
||||
const IndexType end_idx() const noexcept { return capacity(rows_, cols_); }
|
||||
//! @}
|
||||
|
||||
/*!
|
||||
* \name Safe iteration API
|
||||
*
|
||||
* This api automates the iteration over the array based on
|
||||
* MatrixType
|
||||
*/
|
||||
//! @{
|
||||
template<typename F, typename... Args>
|
||||
void for_each_in (IndexType begin, IndexType end, F&& lambda, Args&&... args) {
|
||||
for (IndexType it=begin ; it<end ; ++it) {
|
||||
std::forward<F>(lambda)(std::forward<Args>(args)..., it);
|
||||
}
|
||||
}
|
||||
//! @}
|
||||
//!
|
||||
private:
|
||||
const owner_t* owner_ {nullptr}; //!< Pointer to Matrix
|
||||
DataType* m_ {nullptr}; //!< Starting address of the slice/view
|
||||
IndexType rows_{}; //!< the virtual size of rows.
|
||||
IndexType cols_{}; //!< the virtual size of columns.
|
||||
};
|
||||
|
||||
/*!
|
||||
* A view/iterator hybrid object for Matrix columns.
|
||||
*
|
||||
|
||||
@ -39,7 +39,8 @@ void pdist2(const Matrix& X, const Matrix& Y, Matrix& D2) {
|
||||
int d = X.columns();
|
||||
|
||||
// Compute the squared norms of each row in X and Y
|
||||
std::vector<DataType> X_norms(M), Y_norms(N);
|
||||
std::vector<DataType> X_norms(M);
|
||||
std::vector<DataType> Y_norms(N);
|
||||
for (int i = 0; i < M ; ++i) {
|
||||
X_norms[i] = cblas_ddot(d, X.data() + i * d, 1, X.data() + i * d, 1);
|
||||
}
|
||||
@ -54,12 +55,20 @@ void pdist2(const Matrix& X, const Matrix& Y, Matrix& D2) {
|
||||
for (int i = 0; i < M ; ++i) {
|
||||
for (int j = 0; j < N; ++j) {
|
||||
D2.set(D2.get(i, j) + X_norms[i] + Y_norms[j], i, j);
|
||||
//D2.set(std::max(D2.get(i, j), 0.0), i, j); // Ensure non-negative
|
||||
D2.set(std::max(D2.get(i, j), 0.0), i, j); // Ensure non-negative
|
||||
D2.set(std::sqrt(D2.get(i, j)), i, j); // Take the square root of each
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
/*!
|
||||
* Quick select implementation
|
||||
* \fn void quickselect(std::vector<std::pair<DataType,IndexType>>&, int)
|
||||
* \tparam DataType
|
||||
* \tparam IndexType
|
||||
* \param vec Vector of paire(distance, index) to partially sort over distance
|
||||
* \param k The number of elements to sort-select
|
||||
*/
|
||||
template<typename DataType, typename IndexType>
|
||||
void quickselect(std::vector<std::pair<DataType, IndexType>>& vec, int k) {
|
||||
std::nth_element(
|
||||
@ -75,6 +84,7 @@ void quickselect(std::vector<std::pair<DataType, IndexType>>& vec, int k) {
|
||||
/*!
|
||||
* \param C Is a MxD matrix (Corpus)
|
||||
* \param Q Is a NxD matrix (Query)
|
||||
* \param idx_offset The offset of the indexes for output (to match with the actual Corpus indexes)
|
||||
* \param k The number of nearest neighbors needed
|
||||
* \param idx Is the Nxk matrix with the k indexes of the C points, that are
|
||||
* neighbors of the nth point of Q
|
||||
@ -82,7 +92,7 @@ void quickselect(std::vector<std::pair<DataType, IndexType>>& vec, int k) {
|
||||
* point of Q
|
||||
*/
|
||||
template<typename MatrixD, typename MatrixI>
|
||||
void knnsearch(const MatrixD& C, const MatrixD& Q, size_t idx_offset, size_t k, size_t m, MatrixI& idx, MatrixD& dst) {
|
||||
void knnsearch(MatrixD& C, MatrixD& Q, size_t idx_offset, size_t k, [[maybe_unused]] size_t m, MatrixI& idx, MatrixD& dst) {
|
||||
|
||||
using DstType = typename MatrixD::dataType;
|
||||
using IdxType = typename MatrixI::dataType;
|
||||
@ -94,10 +104,10 @@ void knnsearch(const MatrixD& C, const MatrixD& Q, size_t idx_offset, size_t k,
|
||||
|
||||
pdist2(C, Q, D);
|
||||
|
||||
for (size_t j = 0; j < N; ++j) {
|
||||
for (size_t j = 0UL; j < N; ++j) {
|
||||
// Create a vector of pairs (distance, index) for the j-th query
|
||||
std::vector<std::pair<DstType, IdxType>> dst_idx(M);
|
||||
for (size_t i = 0; i < M; ++i) {
|
||||
for (size_t i = 0UL; i < M; ++i) {
|
||||
dst_idx[i] = {D.data()[i * N + j], i};
|
||||
}
|
||||
// Find the k smallest distances using quickSelectKSmallest
|
||||
@ -107,7 +117,7 @@ void knnsearch(const MatrixD& C, const MatrixD& Q, size_t idx_offset, size_t k,
|
||||
std::sort(dst_idx.begin(), dst_idx.end());
|
||||
|
||||
// Store the indices and distances
|
||||
for (size_t i = 0; i < k; ++i) {
|
||||
for (size_t i = 0UL; i < k; ++i) {
|
||||
dst.set(dst_idx[i].first, j, i);
|
||||
idx.set(dst_idx[i].second + idx_offset, j, i);
|
||||
}
|
||||
|
||||
@ -1,5 +1,5 @@
|
||||
/**
|
||||
* \file v0.hpp
|
||||
* \file v1.hpp
|
||||
* \brief
|
||||
*
|
||||
* \author
|
||||
@ -16,25 +16,61 @@
|
||||
#include "v0.hpp"
|
||||
#include "config.h"
|
||||
|
||||
#if defined CILK
|
||||
#include <cilk/cilk.h>
|
||||
#include <cilk/cilk_api.h>
|
||||
//#include <cilk/reducer_opadd.h>
|
||||
|
||||
#elif defined OMP
|
||||
#include <omp.h>
|
||||
|
||||
#elif defined PTHREADS
|
||||
#include <thread>
|
||||
#include <numeric>
|
||||
#include <functional>
|
||||
//#include <random>
|
||||
|
||||
#else
|
||||
#endif
|
||||
|
||||
|
||||
void init_workers();
|
||||
|
||||
namespace v1 {
|
||||
|
||||
|
||||
/*!
|
||||
*
|
||||
* Merge knnsearch results and select the closest neighbors
|
||||
*
|
||||
* \tparam DataType
|
||||
* \tparam IndexType
|
||||
* \param N1 Neighbors results from one knnsearch
|
||||
* \param D1 Distances results from one knnsearcs
|
||||
* \param N2 Neighbors results from second knnsearch
|
||||
* \param D2 Distances results from second knnsearch
|
||||
* \param k How many
|
||||
* \param m How accurate
|
||||
* \param N Output for Neighbors
|
||||
* \param D Output for Distances
|
||||
*/
|
||||
template <typename DataType, typename IndexType>
|
||||
void mergeResultsWithM(mtx::Matrix<IndexType>& N1, mtx::Matrix<DataType>& D1,
|
||||
mtx::Matrix<IndexType>& N2, mtx::Matrix<DataType>& D2,
|
||||
size_t k, size_t m,
|
||||
mtx::Matrix<IndexType>& N, mtx::Matrix<DataType>& D) {
|
||||
size_t numQueries = N1.rows();
|
||||
size_t maxCandidates = std::min((IndexType)m, (IndexType)(N1.columns() + N2.columns()));
|
||||
size_t maxCandidates = std::min(static_cast<IndexType>(m), static_cast<IndexType>(N1.columns() + N2.columns()));
|
||||
|
||||
for (size_t q = 0; q < numQueries; ++q) {
|
||||
for (size_t q = 0UL; q < numQueries; ++q) {
|
||||
// Combine distances and neighbors
|
||||
std::vector<std::pair<DataType, IndexType>> candidates(N1.columns() + N2.columns());
|
||||
|
||||
// Concatenate N1 and N2 rows
|
||||
for (size_t i = 0; i < N1.columns(); ++i) {
|
||||
for (size_t i = 0UL; i < N1.columns(); ++i) {
|
||||
candidates[i] = {D1.get(q, i), N1.get(q, i)};
|
||||
}
|
||||
for (size_t i = 0; i < N2.columns(); ++i) {
|
||||
for (size_t i = 0UL; i < N2.columns(); ++i) {
|
||||
candidates[i + N1.columns()] = {D2.get(q, i), N2.get(q, i)};
|
||||
}
|
||||
|
||||
@ -45,7 +81,7 @@ void mergeResultsWithM(mtx::Matrix<IndexType>& N1, mtx::Matrix<DataType>& D1,
|
||||
std::sort(candidates.begin(), candidates.begin() + maxCandidates);
|
||||
|
||||
// If m < k, pad the remaining slots with invalid values
|
||||
for (size_t i = 0; i < k; ++i) {
|
||||
for (size_t i = 0UL; i < k; ++i) {
|
||||
if (i < maxCandidates) {
|
||||
D.set(candidates[i].first, q, i);
|
||||
N.set(candidates[i].second, q, i);
|
||||
@ -57,49 +93,119 @@ void mergeResultsWithM(mtx::Matrix<IndexType>& N1, mtx::Matrix<DataType>& D1,
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
|
||||
/*!
|
||||
* The main parallelizable body
|
||||
*/
|
||||
template<typename MatrixD, typename MatrixI>
|
||||
void knnsearch(const MatrixD& C, const MatrixD& Q, size_t idx_offset, size_t k, size_t m, MatrixI& idx, MatrixD& dst) {
|
||||
|
||||
void worker_body (std::vector<MatrixD>& corpus_slices,
|
||||
std::vector<MatrixD>& query_slices,
|
||||
MatrixI& idx,
|
||||
MatrixD& dst,
|
||||
size_t slice,
|
||||
size_t num_slices, size_t corpus_slice_size, size_t query_slice_size,
|
||||
size_t k,
|
||||
size_t m) {
|
||||
// "load" types
|
||||
using DstType = typename MatrixD::dataType;
|
||||
using IdxType = typename MatrixI::dataType;
|
||||
|
||||
if (C.rows() <= 8 || Q.rows() <= 4) {
|
||||
// Base case: Call knnsearch directly
|
||||
v0::knnsearch(C, Q, idx_offset, k, m, idx, dst);
|
||||
return;
|
||||
}
|
||||
|
||||
// Divide Corpus and Query into subsets
|
||||
IdxType midC = C.rows() / 2;
|
||||
IdxType midQ = Q.rows() / 2;
|
||||
for (size_t ci = 0UL; ci < num_slices; ++ci) {
|
||||
size_t idx_offset = ci * corpus_slice_size;
|
||||
|
||||
// Slice corpus and query matrixes
|
||||
MatrixD C1((DstType*)C.data(), 0, midC, C.columns());
|
||||
MatrixD C2((DstType*)C.data(), midC, midC, C.columns());
|
||||
MatrixD Q1((DstType*)Q.data(), 0, midQ, Q.columns());
|
||||
MatrixD Q2((DstType*)Q.data(), midQ, midQ, Q.columns());
|
||||
// Intermediate matrixes for intermediate results
|
||||
MatrixI temp_idx(query_slices[slice].rows(), k);
|
||||
MatrixD temp_dst(query_slices[slice].rows(), k);
|
||||
|
||||
// Allocate temporary matrixes for all permutations
|
||||
MatrixI N1_1(midQ, k), N1_2(midQ, k), N2_1(midQ, k), N2_2(midQ, k);
|
||||
MatrixD D1_1(midQ, k), D1_2(midQ, k), D2_1(midQ, k), D2_2(midQ, k);
|
||||
// kNN for each combination
|
||||
v0::knnsearch(corpus_slices[ci], query_slices[slice], idx_offset, k, m, temp_idx, temp_dst);
|
||||
|
||||
// Recursive calls
|
||||
knnsearch(C1, Q1, idx_offset, k, m, N1_1, D1_1);
|
||||
knnsearch(C2, Q1, idx_offset + midC, k, m, N1_2, D1_2);
|
||||
knnsearch(C1, Q2, idx_offset, k, m, N2_1, D2_1);
|
||||
knnsearch(C2, Q2, idx_offset + midC, k, m, N2_2, D2_2);
|
||||
// Merge temporary results to final results
|
||||
MatrixI idx_slice(static_cast<IdxType*>(idx.data()), slice * query_slice_size, query_slices[slice].rows(), k);
|
||||
MatrixD dst_slice(static_cast<DstType*>(dst.data()), slice * query_slice_size, query_slices[slice].rows(), k);
|
||||
|
||||
// slice output matrixes
|
||||
MatrixI N1((IdxType*)idx.data(), 0, midQ, k);
|
||||
MatrixI N2((IdxType*)idx.data(), midQ, midQ, k);
|
||||
MatrixD D1((DstType*)dst.data(), 0, midQ, k);
|
||||
MatrixD D2((DstType*)dst.data(), midQ, midQ, k);
|
||||
mergeResultsWithM(idx_slice, dst_slice, temp_idx, temp_dst, k, m, idx_slice, dst_slice);
|
||||
}
|
||||
}
|
||||
|
||||
// Merge results in place
|
||||
mergeResultsWithM(N1_1, D1_1, N1_2, D1_2, k, m, N1, D1);
|
||||
mergeResultsWithM(N2_1, D2_1, N2_2, D2_2, k, m, N2, D2);
|
||||
/*!
|
||||
* \param C Is a MxD matrix (Corpus)
|
||||
* \param Q Is a NxD matrix (Query)
|
||||
* \param num_slices How many slices to Corpus-Query
|
||||
* \param k The number of nearest neighbors needed
|
||||
* \param m accuracy
|
||||
* \param idx Is the Nxk matrix with the k indexes of the C points, that are
|
||||
* neighbors of the nth point of Q
|
||||
* \param dst Is the Nxk matrix with the k distances to the C points of the nth
|
||||
* point of Q
|
||||
*/
|
||||
template<typename MatrixD, typename MatrixI>
|
||||
void knnsearch(MatrixD& C, MatrixD& Q, size_t num_slices, size_t k, size_t m, MatrixI& idx, MatrixD& dst) {
|
||||
using DstType = typename MatrixD::dataType;
|
||||
using IdxType = typename MatrixI::dataType;
|
||||
|
||||
//Slice calculations
|
||||
size_t corpus_slice_size = C.rows() / ((num_slices == 0UL)? 1UL:num_slices);
|
||||
size_t query_slice_size = Q.rows() / ((num_slices == 0UL)? 1UL:num_slices);
|
||||
|
||||
// Make slices
|
||||
std::vector<MatrixD> corpus_slices{};
|
||||
std::vector<MatrixD> query_slices{};
|
||||
|
||||
for (size_t i = 0UL; i < num_slices; ++i) {
|
||||
corpus_slices.emplace_back(
|
||||
static_cast<DstType*>(C.data()),
|
||||
i * corpus_slice_size,
|
||||
(i == num_slices - 1UL ? C.rows() - i * corpus_slice_size : corpus_slice_size),
|
||||
C.columns());
|
||||
query_slices.emplace_back(
|
||||
static_cast<DstType*>(Q.data()),
|
||||
i * query_slice_size,
|
||||
(i == num_slices - 1UL ? Q.rows() - i * query_slice_size : query_slice_size),
|
||||
Q.columns());
|
||||
}
|
||||
|
||||
// Initialize results
|
||||
for (size_t i = 0UL; i < dst.rows(); ++i) {
|
||||
for (size_t j = 0UL; j < dst.columns(); ++j) {
|
||||
dst.set(std::numeric_limits<DstType>::infinity(), i, j);
|
||||
idx.set(static_cast<IdxType>(-1), i, j);
|
||||
}
|
||||
}
|
||||
|
||||
// Main loop
|
||||
#if defined OMP
|
||||
#pragma omp parallel for
|
||||
for (size_t qi = 0UL; qi < num_slices; ++qi) {
|
||||
worker_body (corpus_slices, query_slices, idx, dst, qi, num_slices, corpus_slice_size, query_slice_size, k, m);
|
||||
}
|
||||
#elif defined CILK
|
||||
cilk_for (size_t qi = 0; qi < num_slices; ++qi) {
|
||||
worker_body (corpus_slices, query_slices, idx, dst, qi, num_slices, corpus_slice_size, query_slice_size, k, m);
|
||||
}
|
||||
#elif defined PTHREADS
|
||||
std::vector<std::thread> workers;
|
||||
workers.reserve(num_slices);
|
||||
for (size_t qi = 0; qi < num_slices; ++qi) {
|
||||
workers.push_back(
|
||||
std::thread (worker_body<MatrixD, MatrixI>,
|
||||
std::ref(corpus_slices), std::ref(query_slices),
|
||||
std::ref(idx), std::ref(dst),
|
||||
qi,
|
||||
num_slices, corpus_slice_size, query_slice_size,
|
||||
k, m)
|
||||
);
|
||||
}
|
||||
// Join threads
|
||||
std::for_each(workers.begin(), workers.end(), [](std::thread& t){
|
||||
t.join();
|
||||
});
|
||||
|
||||
#else
|
||||
for (size_t qi = 0; qi < num_slices; ++qi) {
|
||||
worker_body (corpus_slices, query_slices, idx, dst, qi, num_slices, corpus_slice_size, query_slice_size, k, m);
|
||||
}
|
||||
#endif
|
||||
|
||||
}
|
||||
|
||||
|
||||
@ -12,7 +12,10 @@ function [D2] = dist2(X, Y)
|
||||
if d1 ~= d2
|
||||
error('X,Y column dimensions must match');
|
||||
end
|
||||
%D2 = sqrt((X.^2)*ones(d,m) -2*X*Y' + ones(n,d)*(Y.^2)');
|
||||
% debug
|
||||
%X_norm = sum(X.^2, 2);
|
||||
%Y_norm = sum(Y.^2, 2)';
|
||||
%XY = 2 * X*Y';
|
||||
D2 = max(sum(X.^2, 2) - 2 * X*Y' + sum(Y.^2, 2)', 0);
|
||||
D2 = sqrt(D2);
|
||||
|
||||
|
||||
78
homework_1/matlab/measurements_over_acc.m
Normal file
78
homework_1/matlab/measurements_over_acc.m
Normal file
@ -0,0 +1,78 @@
|
||||
% Plot measurements
|
||||
|
||||
accuracy = [100 80 60 40 20 10];
|
||||
ser_sift_acc = [ 4395 4365 4384 4315 4295 4246 ];
|
||||
ser_mnist_ser_acc = [ 7936 7924 7886 7903 7844 7801 ];
|
||||
|
||||
omp_sift_acc = [
|
||||
1162 1157 1148 1138 1134 1107
|
||||
];
|
||||
omp_mnist_acc = [
|
||||
2044 2036 2028 2007 1993 1973
|
||||
];
|
||||
|
||||
cilk_sift_acc = [
|
||||
1148 1133 1127 1096 1079 1057
|
||||
];
|
||||
cilk_mnist_acc = [
|
||||
2008 2024 1974 1965 1959 1947
|
||||
];
|
||||
|
||||
pth_sift_acc = [
|
||||
1157 1159 1121 1100 1084 1075
|
||||
];
|
||||
pth_mnist_acc = [
|
||||
2050 2086 2040 2020 2004 1979
|
||||
];
|
||||
|
||||
% 1ο Διάγραμμα: OMP
|
||||
figure;
|
||||
set(gcf, 'Position', [100, 100, 1280, 720]); % Set the figure size to HD
|
||||
plot(accuracy, ser_sift_acc, '-o', 'DisplayName', 'Serial SIFT');
|
||||
hold on;
|
||||
plot(accuracy, ser_mnist_ser_acc, '-s', 'DisplayName', 'Serial MNIST');
|
||||
plot(accuracy, omp_sift_acc, '-^', 'DisplayName', 'OMP SIFT');
|
||||
plot(accuracy, omp_mnist_acc, '-d', 'DisplayName', 'OMP MNIST');
|
||||
hold off;
|
||||
title('OMP');
|
||||
xlabel('Accuracy (%)');
|
||||
ylabel('Execution Time [msec]');
|
||||
set(gca, 'XDir', 'reverse'); % reverse x
|
||||
legend('Location', 'northwest');
|
||||
grid on;
|
||||
print(gcf, 'OMP_over_accuracy.png', '-dpng', '-r300');
|
||||
|
||||
% 2ο Διάγραμμα: CILK
|
||||
figure;
|
||||
set(gcf, 'Position', [100, 100, 1280, 720]); % Set the figure size to HD
|
||||
plot(accuracy, ser_sift_acc, '-o', 'DisplayName', 'Serial SIFT');
|
||||
hold on;
|
||||
plot(accuracy, ser_mnist_ser_acc, '-s', 'DisplayName', 'Serial MNIST');
|
||||
plot(accuracy, cilk_sift_acc, '-^', 'DisplayName', 'CILK SIFT');
|
||||
plot(accuracy, cilk_mnist_acc, '-d', 'DisplayName', 'CILK MNIST');
|
||||
hold off;
|
||||
title('CILK');
|
||||
xlabel('Accuracy (%)');
|
||||
ylabel('Execution Time [msec]');
|
||||
set(gca, 'XDir', 'reverse'); % reverse x
|
||||
legend('Location', 'northwest');
|
||||
grid on;
|
||||
print(gcf, 'CILK_over_accuracy.png', '-dpng', '-r300');
|
||||
|
||||
% 3ο Διάγραμμα: Pthreads
|
||||
figure;
|
||||
set(gcf, 'Position', [100, 100, 1280, 720]); % Set the figure size to HD
|
||||
plot(accuracy, ser_sift_acc, '-o', 'DisplayName', 'Serial SIFT');
|
||||
hold on;
|
||||
plot(accuracy, ser_mnist_ser_acc, '-s', 'DisplayName', 'Serial MNIST');
|
||||
plot(accuracy, pth_sift_acc, '-^', 'DisplayName', 'Pthreads SIFT');
|
||||
plot(accuracy, pth_mnist_acc, '-d', 'DisplayName', 'Pthreads MNIST');
|
||||
hold off;
|
||||
title('Pthreads');
|
||||
xlabel('Accuracy (%)');
|
||||
ylabel('Execution Time [msec]');
|
||||
set(gca, 'XDir', 'reverse'); % reverse x
|
||||
legend('Location', 'northwest');
|
||||
grid on;
|
||||
print(gcf, 'Pthreads_over_accuracy.png', '-dpng', '-r300');
|
||||
|
||||
75
homework_1/matlab/measurements_over_th.m
Normal file
75
homework_1/matlab/measurements_over_th.m
Normal file
@ -0,0 +1,75 @@
|
||||
% Plot measurements
|
||||
|
||||
threads = [1 2 4 6 8 10 12];
|
||||
ser_sift_threads = [ 4418 4418 4418 4418 4418 4418 4418 ];
|
||||
ser_mnist_ser_threads = [ 7924 7924 7924 7924 7924 7924 7924 ];
|
||||
|
||||
omp_sift_th = [
|
||||
4374 2194 1148 751 724 637 596
|
||||
];
|
||||
omp_mnist_th = [
|
||||
7918 3941 2031 1395 1415 1330 1292
|
||||
];
|
||||
|
||||
cilk_sift_th = [
|
||||
4183 2076 1114 886 706 742 636
|
||||
];
|
||||
cilk_mnist_th = [
|
||||
7745 3883 2020 1454 1436 1342 1292
|
||||
];
|
||||
|
||||
pth_sift_th = [
|
||||
4254 2155 1133 877 724 640 682
|
||||
];
|
||||
pth_mnist_th = [
|
||||
7889 3963 2058 1445 1496 1379 1352
|
||||
];
|
||||
|
||||
% 1ο Διάγραμμα: OMP
|
||||
figure;
|
||||
set(gcf, 'Position', [100, 100, 1280, 720]); % Set the figure size to HD
|
||||
plot(threads, ser_sift_threads, '-o', 'DisplayName', 'Serial SIFT');
|
||||
hold on;
|
||||
plot(threads, ser_mnist_ser_threads, '-s', 'DisplayName', 'Serial MNIST');
|
||||
plot(threads, omp_sift_th, '-^', 'DisplayName', 'OMP SIFT');
|
||||
plot(threads, omp_mnist_th, '-d', 'DisplayName', 'OMP MNIST');
|
||||
hold off;
|
||||
title('OMP');
|
||||
xlabel('Threads');
|
||||
ylabel('Execution Time [msec]');
|
||||
legend('Location', 'northeast');
|
||||
grid on;
|
||||
print(gcf, 'OMP_over_threads.png', '-dpng', '-r300');
|
||||
|
||||
% 2ο Διάγραμμα: CILK
|
||||
figure;
|
||||
set(gcf, 'Position', [100, 100, 1280, 720]); % Set the figure size to HD
|
||||
plot(threads, ser_sift_threads, '-o', 'DisplayName', 'Serial SIFT');
|
||||
hold on;
|
||||
plot(threads, ser_mnist_ser_threads, '-s', 'DisplayName', 'Serial MNIST');
|
||||
plot(threads, cilk_sift_th, '-^', 'DisplayName', 'CILK SIFT');
|
||||
plot(threads, cilk_mnist_th, '-d', 'DisplayName', 'CILK MNIST');
|
||||
hold off;
|
||||
title('CILK');
|
||||
xlabel('Threads');
|
||||
ylabel('Execution Time [msec]');
|
||||
legend('Location', 'northeast');
|
||||
grid on;
|
||||
print(gcf, 'CILK_over_threads.png', '-dpng', '-r300');
|
||||
|
||||
% 3ο Διάγραμμα: Pthreads
|
||||
figure;
|
||||
set(gcf, 'Position', [100, 100, 1280, 720]); % Set the figure size to HD
|
||||
plot(threads, ser_sift_threads, '-o', 'DisplayName', 'Serial SIFT');
|
||||
hold on;
|
||||
plot(threads, ser_mnist_ser_threads, '-s', 'DisplayName', 'Serial MNIST');
|
||||
plot(threads, pth_sift_th, '-^', 'DisplayName', 'Pthreads SIFT');
|
||||
plot(threads, pth_mnist_th, '-d', 'DisplayName', 'Pthreads MNIST');
|
||||
hold off;
|
||||
title('Pthreads');
|
||||
xlabel('Threads');
|
||||
ylabel('Execution Time [msec]');
|
||||
legend('Location', 'northeast');
|
||||
grid on;
|
||||
print(gcf, 'Pthreads_over_threads.png', '-dpng', '-r300');
|
||||
|
||||
@ -2,6 +2,106 @@
|
||||
%
|
||||
%
|
||||
|
||||
|
||||
%
|
||||
%
|
||||
%
|
||||
C1 = [
|
||||
0.8147 0.1576;
|
||||
0.9058 0.9706;
|
||||
0.1270 0.9572;
|
||||
0.9134 0.4854;
|
||||
0.6324 0.8003;
|
||||
0.0975 0.1419;
|
||||
0.2785 0.4218;
|
||||
0.5469 0.9157;
|
||||
0.9575 0.7922;
|
||||
0.9649 0.9595 ];
|
||||
|
||||
Q1 = [
|
||||
0.6557 0.7577;
|
||||
0.0357 0.7431;
|
||||
0.8491 0.3922;
|
||||
0.9340 0.6555;
|
||||
0.6787 0.1712 ];
|
||||
|
||||
C2 = [
|
||||
0.7060 0.4456 0.5060 0.6160;
|
||||
0.0318 0.6463 0.6991 0.4733;
|
||||
0.2769 0.7094 0.8909 0.3517;
|
||||
0.0462 0.7547 0.9593 0.8308;
|
||||
0.0971 0.2760 0.5472 0.5853;
|
||||
0.8235 0.6797 0.1386 0.5497;
|
||||
0.6948 0.6551 0.1493 0.9172;
|
||||
0.3171 0.1626 0.2575 0.2858;
|
||||
0.9502 0.1190 0.8407 0.7572;
|
||||
0.0344 0.4984 0.2543 0.7537;
|
||||
0.4387 0.9597 0.8143 0.3804;
|
||||
0.3816 0.3404 0.2435 0.5678;
|
||||
0.7655 0.5853 0.9293 0.0759;
|
||||
0.7952 0.2238 0.3500 0.0540;
|
||||
0.1869 0.7513 0.1966 0.5308;
|
||||
0.4898 0.2551 0.2511 0.7792 ];
|
||||
|
||||
Q2 = [
|
||||
0.9340 0.3112 0.4505 0.0782;
|
||||
0.1299 0.5285 0.0838 0.4427;
|
||||
0.5688 0.1656 0.2290 0.1067;
|
||||
0.4694 0.6020 0.9133 0.9619;
|
||||
0.0119 0.2630 0.1524 0.0046;
|
||||
0.3371 0.6541 0.8258 0.7749;
|
||||
0.1622 0.6892 0.5383 0.8173;
|
||||
0.7943 0.7482 0.9961 0.8687 ];
|
||||
|
||||
D1_exp = [
|
||||
0.6208 0.9745 0.2371 0.5120 0.1367;
|
||||
0.3284 0.8993 0.5811 0.3164 0.8310;
|
||||
0.5651 0.2327 0.9169 0.8616 0.9603;
|
||||
0.3749 0.9147 0.1132 0.1713 0.3921;
|
||||
0.0485 0.5994 0.4621 0.3346 0.6308;
|
||||
0.8312 0.6044 0.7922 0.9815 0.5819;
|
||||
0.5052 0.4028 0.5714 0.6959 0.4722;
|
||||
0.1919 0.5395 0.6045 0.4665 0.7561;
|
||||
0.3037 0.9231 0.4144 0.1387 0.6807;
|
||||
0.3692 0.9540 0.5790 0.3056 0.8386 ];
|
||||
|
||||
|
||||
D2_exp = [
|
||||
0.6020 0.7396 0.6583 0.6050 1.0070 0.5542 0.6298 0.6352;
|
||||
1.0696 0.6348 0.9353 0.6914 0.8160 0.4475 0.4037 0.9145;
|
||||
0.9268 0.8450 0.9376 0.6492 0.9671 0.4360 0.5956 0.7400;
|
||||
1.3455 0.9876 1.2953 0.4709 1.2557 0.3402 0.4417 0.7500;
|
||||
0.9839 0.5476 0.7517 0.7216 0.7074 0.5605 0.4784 0.9954;
|
||||
0.6839 0.7200 0.7305 0.9495 1.0628 0.8718 0.8178 0.9179;
|
||||
0.9850 0.7514 0.9585 0.7996 1.2054 0.7784 0.6680 0.8591;
|
||||
0.6950 0.4730 0.3103 1.0504 0.4397 0.8967 0.8140 1.2066;
|
||||
0.8065 1.2298 0.9722 0.7153 1.3933 0.8141 1.0204 0.6758;
|
||||
1.1572 0.3686 0.9031 0.8232 0.7921 0.6656 0.3708 1.0970;
|
||||
0.9432 0.9049 1.0320 0.6905 1.1167 0.5094 0.6455 0.6653;
|
||||
0.7672 0.3740 0.5277 0.8247 0.6842 0.6945 0.5648 0.9968;
|
||||
0.5768 1.1210 0.8403 0.9345 1.1316 0.8292 1.0380 0.8127;
|
||||
0.1939 0.8703 0.2684 1.1794 0.8103 1.0683 1.1115 1.1646;
|
||||
1.0106 0.2708 0.8184 0.8954 0.7402 0.6982 0.4509 1.0594;
|
||||
0.8554 0.5878 0.6834 0.7699 0.9155 0.7161 0.6162 0.9481 ];
|
||||
|
||||
% tests
|
||||
D1 = dist2(C1, Q1);
|
||||
if norm (D1-pdist2(C1, Q1), 'fro') > 0.01
|
||||
disp('Error in dist2(C1, Q1)');
|
||||
end
|
||||
|
||||
D2 = dist2(C2, Q2);
|
||||
if norm (D2-pdist2(C2, Q2), 'fro') > 0.01
|
||||
disp('Error in dist2(C2, Q2)');
|
||||
end
|
||||
|
||||
D2 = dist2(C2, C2);
|
||||
if norm (D2-pdist2(C2, C2), 'fro') > 0.01
|
||||
disp('Error in dist2(C2, C2)');
|
||||
end
|
||||
|
||||
|
||||
|
||||
%C = rand(10000, 2); % Corpus
|
||||
%Q = rand(10000, 2); % Queries
|
||||
C = rand(20000, 2); % Δύο clusters
|
||||
|
||||
BIN
homework_1/measurements/CILK_over_accuracy.png
Normal file
BIN
homework_1/measurements/CILK_over_accuracy.png
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 93 KiB |
BIN
homework_1/measurements/CILK_over_threads.png
Normal file
BIN
homework_1/measurements/CILK_over_threads.png
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 97 KiB |
BIN
homework_1/measurements/OMP_over_accuracy.png
Normal file
BIN
homework_1/measurements/OMP_over_accuracy.png
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 97 KiB |
BIN
homework_1/measurements/OMP_over_threads.png
Normal file
BIN
homework_1/measurements/OMP_over_threads.png
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 99 KiB |
BIN
homework_1/measurements/Pthreads_over_accuracy.png
Normal file
BIN
homework_1/measurements/Pthreads_over_accuracy.png
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 100 KiB |
BIN
homework_1/measurements/Pthreads_over_threads.png
Normal file
BIN
homework_1/measurements/Pthreads_over_threads.png
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 105 KiB |
122
homework_1/measurements/over_acc.txt
Normal file
122
homework_1/measurements/over_acc.txt
Normal file
@ -0,0 +1,122 @@
|
||||
# Serial-sift over acc
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1 -c mtx/sift-128-euclidean.hdf5 /test -s 1 -a 100 -k 100 -t
|
||||
[Timing]: knnsearch: 4395 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1 -c mtx/sift-128-euclidean.hdf5 /test -s 1 -a 80 -k 100 -t
|
||||
[Timing]: knnsearch: 4365 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1 -c mtx/sift-128-euclidean.hdf5 /test -s 1 -a 60 -k 100 -t
|
||||
[Timing]: knnsearch: 4384 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1 -c mtx/sift-128-euclidean.hdf5 /test -s 1 -a 40 -k 100 -t
|
||||
[Timing]: knnsearch: 4315 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1 -c mtx/sift-128-euclidean.hdf5 /test -s 1 -a 20 -k 100 -t
|
||||
[Timing]: knnsearch: 4295 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1 -c mtx/sift-128-euclidean.hdf5 /test -s 1 -a 10 -k 100 -t
|
||||
[Timing]: knnsearch: 4246 [msec]
|
||||
|
||||
|
||||
# Serial-mnist over acc
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1 -c mtx/mnist-784-euclidean.hdf5 /test -s 1 -a 100 -k 100 -t
|
||||
[Timing]: knnsearch: 7936 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1 -c mtx/mnist-784-euclidean.hdf5 /test -s 1 -a 80 -k 100 -t
|
||||
[Timing]: knnsearch: 7924 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1 -c mtx/mnist-784-euclidean.hdf5 /test -s 1 -a 60 -k 100 -t
|
||||
[Timing]: knnsearch: 7886 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1 -c mtx/mnist-784-euclidean.hdf5 /test -s 1 -a 40 -k 100 -t
|
||||
[Timing]: knnsearch: 7903 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1 -c mtx/mnist-784-euclidean.hdf5 /test -s 1 -a 20 -k 100 -t
|
||||
[Timing]: knnsearch: 7844 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1 -c mtx/mnist-784-euclidean.hdf5 /test -s 1 -a 10 -k 100 -t
|
||||
[Timing]: knnsearch: 7801 [msec]
|
||||
|
||||
|
||||
# OMP-sift over acc
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_omp -c mtx/sift-128-euclidean.hdf5 /test -s 4 -a 100 -k 100 -t
|
||||
[Timing]: knnsearch: 1162 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_omp -c mtx/sift-128-euclidean.hdf5 /test -s 4 -a 80 -k 100 -t
|
||||
[Timing]: knnsearch: 1157 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_omp -c mtx/sift-128-euclidean.hdf5 /test -s 4 -a 60 -k 100 -t
|
||||
[Timing]: knnsearch: 1148 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_omp -c mtx/sift-128-euclidean.hdf5 /test -s 4 -a 40 -k 100 -t
|
||||
[Timing]: knnsearch: 1138 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_omp -c mtx/sift-128-euclidean.hdf5 /test -s 4 -a 20 -k 100 -t
|
||||
[Timing]: knnsearch: 1134 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_omp -c mtx/sift-128-euclidean.hdf5 /test -s 4 -a 10 -k 100 -t
|
||||
[Timing]: knnsearch: 1107 [msec]
|
||||
|
||||
# OMP-mnist over acc
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_omp -c mtx/mnist-784-euclidean.hdf5 /test -s 4 -a 100 -k 100 -t
|
||||
[Timing]: knnsearch: 2044 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_omp -c mtx/mnist-784-euclidean.hdf5 /test -s 4 -a 80 -k 100 -t
|
||||
[Timing]: knnsearch: 2036 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_omp -c mtx/mnist-784-euclidean.hdf5 /test -s 4 -a 60 -k 100 -t
|
||||
[Timing]: knnsearch: 2028 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_omp -c mtx/mnist-784-euclidean.hdf5 /test -s 4 -a 40 -k 100 -t
|
||||
[Timing]: knnsearch: 2007 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_omp -c mtx/mnist-784-euclidean.hdf5 /test -s 4 -a 20 -k 100 -t
|
||||
[Timing]: knnsearch: 1993 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_omp -c mtx/mnist-784-euclidean.hdf5 /test -s 4 -a 10 -k 100 -t
|
||||
[Timing]: knnsearch: 1973 [msec]
|
||||
|
||||
|
||||
|
||||
# CILK-sift over acc
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_cilk -c mtx/sift-128-euclidean.hdf5 /test -s 4 -a 100 -k 100 -t
|
||||
[Timing]: knnsearch: 1148 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_cilk -c mtx/sift-128-euclidean.hdf5 /test -s 4 -a 80 -k 100 -t
|
||||
[Timing]: knnsearch: 1133 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_cilk -c mtx/sift-128-euclidean.hdf5 /test -s 4 -a 60 -k 100 -t
|
||||
[Timing]: knnsearch: 1127 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_cilk -c mtx/sift-128-euclidean.hdf5 /test -s 4 -a 40 -k 100 -t
|
||||
[Timing]: knnsearch: 1096 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_cilk -c mtx/sift-128-euclidean.hdf5 /test -s 4 -a 20 -k 100 -t
|
||||
[Timing]: knnsearch: 1079 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_cilk -c mtx/sift-128-euclidean.hdf5 /test -s 4 -a 10 -k 100 -t
|
||||
[Timing]: knnsearch: 1057 [msec]
|
||||
|
||||
|
||||
# CILK-mnist over acc
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_cilk -c mtx/mnist-784-euclidean.hdf5 /test -s 4 -a 100 -k 100 -t
|
||||
[Timing]: knnsearch: 2008 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_cilk -c mtx/mnist-784-euclidean.hdf5 /test -s 4 -a 80 -k 100 -t
|
||||
[Timing]: knnsearch: 2024 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_cilk -c mtx/mnist-784-euclidean.hdf5 /test -s 4 -a 60 -k 100 -t
|
||||
[Timing]: knnsearch: 1974 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_cilk -c mtx/mnist-784-euclidean.hdf5 /test -s 4 -a 40 -k 100 -t
|
||||
[Timing]: knnsearch: 1965 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_cilk -c mtx/mnist-784-euclidean.hdf5 /test -s 4 -a 20 -k 100 -t
|
||||
[Timing]: knnsearch: 1959 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_cilk -c mtx/mnist-784-euclidean.hdf5 /test -s 4 -a 10 -k 100 -t
|
||||
[Timing]: knnsearch: 1947 [msec]
|
||||
|
||||
|
||||
# Pthreads-sift over acc
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_pth -c mtx/sift-128-euclidean.hdf5 /test -s 4 -a 100 -k 100 -t
|
||||
[Timing]: knnsearch: 1157 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_pth -c mtx/sift-128-euclidean.hdf5 /test -s 4 -a 80 -k 100 -t
|
||||
[Timing]: knnsearch: 1159 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_pth -c mtx/sift-128-euclidean.hdf5 /test -s 4 -a 60 -k 100 -t
|
||||
[Timing]: knnsearch: 1121 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_pth -c mtx/sift-128-euclidean.hdf5 /test -s 4 -a 40 -k 100 -t
|
||||
[Timing]: knnsearch: 1100 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_pth -c mtx/sift-128-euclidean.hdf5 /test -s 4 -a 20 -k 100 -t
|
||||
[Timing]: knnsearch: 1084 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_pth -c mtx/sift-128-euclidean.hdf5 /test -s 4 -a 10 -k 100 -t
|
||||
[Timing]: knnsearch: 1075 [msec]
|
||||
|
||||
|
||||
# Pthreads-mnist over acc
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_pth -c mtx/mnist-784-euclidean.hdf5 /test -s 4 -a 100 -k 100 -t
|
||||
[Timing]: knnsearch: 2050 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_pth -c mtx/mnist-784-euclidean.hdf5 /test -s 4 -a 80 -k 100 -t
|
||||
[Timing]: knnsearch: 2086 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_pth -c mtx/mnist-784-euclidean.hdf5 /test -s 4 -a 60 -k 100 -t
|
||||
[Timing]: knnsearch: 2040 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_pth -c mtx/mnist-784-euclidean.hdf5 /test -s 4 -a 40 -k 100 -t
|
||||
[Timing]: knnsearch: 2020 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_pth -c mtx/mnist-784-euclidean.hdf5 /test -s 4 -a 20 -k 100 -t
|
||||
[Timing]: knnsearch: 2004 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_pth -c mtx/mnist-784-euclidean.hdf5 /test -s 4 -a 10 -k 100 -t
|
||||
[Timing]: knnsearch: 1979 [msec]
|
||||
|
||||
|
||||
|
||||
|
||||
103
homework_1/measurements/over_threads.txt
Normal file
103
homework_1/measurements/over_threads.txt
Normal file
@ -0,0 +1,103 @@
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1 -c mtx/sift-128-euclidean.hdf5 /test -s 1 -k 100 -t
|
||||
[Timing]: knnsearch: 4418 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1 -c mtx/mnist-784-euclidean.hdf5 /test -s 1 -k 100 -t
|
||||
[Timing]: knnsearch: 7924 [msec]
|
||||
|
||||
# OMP-sift over threads
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_omp -c mtx/sift-128-euclidean.hdf5 /test -s 1 -k 100 -t
|
||||
[Timing]: knnsearch: 4374 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_omp -c mtx/sift-128-euclidean.hdf5 /test -s 2 -k 100 -t
|
||||
[Timing]: knnsearch: 2194 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_omp -c mtx/sift-128-euclidean.hdf5 /test -s 4 -k 100 -t
|
||||
[Timing]: knnsearch: 1148 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_omp -c mtx/sift-128-euclidean.hdf5 /test -s 6 -k 100 -t
|
||||
[Timing]: knnsearch: 751 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_omp -c mtx/sift-128-euclidean.hdf5 /test -s 8 -k 100 -t
|
||||
[Timing]: knnsearch: 724 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_omp -c mtx/sift-128-euclidean.hdf5 /test -s 10 -k 100 -t
|
||||
[Timing]: knnsearch: 637 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_omp -c mtx/sift-128-euclidean.hdf5 /test -s 12 -k 100 -t
|
||||
[Timing]: knnsearch: 596 [msec]
|
||||
|
||||
# OMP-mnist over threads
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_omp -c mtx/mnist-784-euclidean.hdf5 /test -s 1 -k 100 -t
|
||||
[Timing]: knnsearch: 7918 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_omp -c mtx/mnist-784-euclidean.hdf5 /test -s 2 -k 100 -t
|
||||
[Timing]: knnsearch: 3941 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_omp -c mtx/mnist-784-euclidean.hdf5 /test -s 4 -k 100 -t
|
||||
[Timing]: knnsearch: 2031 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_omp -c mtx/mnist-784-euclidean.hdf5 /test -s 6 -k 100 -t
|
||||
[Timing]: knnsearch: 1395 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_omp -c mtx/mnist-784-euclidean.hdf5 /test -s 8 -k 100 -t
|
||||
[Timing]: knnsearch: 1415 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_omp -c mtx/mnist-784-euclidean.hdf5 /test -s 10 -k 100 -t
|
||||
[Timing]: knnsearch: 1330 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_omp -c mtx/mnist-784-euclidean.hdf5 /test -s 12 -k 100 -t
|
||||
[Timing]: knnsearch: 1292 [msec]
|
||||
|
||||
|
||||
# CILK-sift over threads
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_cilk -c mtx/sift-128-euclidean.hdf5 /test -s 1 -k 100 -t
|
||||
[Timing]: knnsearch: 4183 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_cilk -c mtx/sift-128-euclidean.hdf5 /test -s 2 -k 100 -t
|
||||
[Timing]: knnsearch: 2076 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_cilk -c mtx/sift-128-euclidean.hdf5 /test -s 4 -k 100 -t
|
||||
[Timing]: knnsearch: 1114 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_cilk -c mtx/sift-128-euclidean.hdf5 /test -s 6 -k 100 -t
|
||||
[Timing]: knnsearch: 886 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_cilk -c mtx/sift-128-euclidean.hdf5 /test -s 8 -k 100 -t
|
||||
[Timing]: knnsearch: 706 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_cilk -c mtx/sift-128-euclidean.hdf5 /test -s 10 -k 100 -t
|
||||
[Timing]: knnsearch: 742 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_cilk -c mtx/sift-128-euclidean.hdf5 /test -s 12 -k 100 -t
|
||||
[Timing]: knnsearch: 636 [msec]
|
||||
|
||||
# CILK-mnist over threads
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_cilk -c mtx/mnist-784-euclidean.hdf5 /test -s 1 -k 100 -t
|
||||
[Timing]: knnsearch: 7745 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_cilk -c mtx/mnist-784-euclidean.hdf5 /test -s 2 -k 100 -t
|
||||
[Timing]: knnsearch: 3883 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_cilk -c mtx/mnist-784-euclidean.hdf5 /test -s 4 -k 100 -t
|
||||
[Timing]: knnsearch: 2020 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_cilk -c mtx/mnist-784-euclidean.hdf5 /test -s 6 -k 100 -t
|
||||
[Timing]: knnsearch: 1454 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_cilk -c mtx/mnist-784-euclidean.hdf5 /test -s 8 -k 100 -t
|
||||
[Timing]: knnsearch: 1436 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_cilk -c mtx/mnist-784-euclidean.hdf5 /test -s 10 -k 100 -t
|
||||
[Timing]: knnsearch: 1342 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_cilk -c mtx/mnist-784-euclidean.hdf5 /test -s 12 -k 100 -t
|
||||
[Timing]: knnsearch: 1292 [msec]
|
||||
|
||||
|
||||
# Pthreads-sift over threads
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_pth -c mtx/sift-128-euclidean.hdf5 /test -s 1 -k 100 -t
|
||||
[Timing]: knnsearch: 4254 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_pth -c mtx/sift-128-euclidean.hdf5 /test -s 2 -k 100 -t
|
||||
[Timing]: knnsearch: 2155 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_pth -c mtx/sift-128-euclidean.hdf5 /test -s 4 -k 100 -t
|
||||
[Timing]: knnsearch: 1133 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_pth -c mtx/sift-128-euclidean.hdf5 /test -s 6 -k 100 -t
|
||||
[Timing]: knnsearch: 877 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_pth -c mtx/sift-128-euclidean.hdf5 /test -s 8 -k 100 -t
|
||||
[Timing]: knnsearch: 724 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_pth -c mtx/sift-128-euclidean.hdf5 /test -s 10 -k 100 -t
|
||||
[Timing]: knnsearch: 640 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_pth -c mtx/sift-128-euclidean.hdf5 /test -s 12 -k 100 -t
|
||||
[Timing]: knnsearch: 682 [msec]
|
||||
|
||||
# Pthreads-mnist over threads
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_pth -c mtx/mnist-784-euclidean.hdf5 /test -s 1 -k 100 -t
|
||||
[Timing]: knnsearch: 7889 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_pth -c mtx/mnist-784-euclidean.hdf5 /test -s 2 -k 100 -t
|
||||
[Timing]: knnsearch: 3963 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_pth -c mtx/mnist-784-euclidean.hdf5 /test -s 4 -k 100 -t
|
||||
[Timing]: knnsearch: 2058 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_pth -c mtx/mnist-784-euclidean.hdf5 /test -s 6 -k 100 -t
|
||||
[Timing]: knnsearch: 1445 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_pth -c mtx/mnist-784-euclidean.hdf5 /test -s 8 -k 100 -t
|
||||
[Timing]: knnsearch: 1496 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_pth -c mtx/mnist-784-euclidean.hdf5 /test -s 10 -k 100 -t
|
||||
[Timing]: knnsearch: 1379 [msec]
|
||||
hoo2@shirka:~/Work/AUTH/PDS/homework_1$ eval $DOCK ./out/knnsearch_v1_pth -c mtx/mnist-784-euclidean.hdf5 /test -s 12 -k 100 -t
|
||||
[Timing]: knnsearch: 1352 [msec]
|
||||
|
||||
Binary file not shown.
@ -46,5 +46,197 @@
|
||||
|
||||
\section{Εισαγωγή}
|
||||
|
||||
Η παρούσα εργασία αφορά τον παραλληλισμό συστημάτων με κοινή μνήμη.
|
||||
Το αντικείμενό είναι ο παραλληλισμός τους αλγορίθμου εύρεσης κοντινότερων γειτόνων \textbf{knnseach}.
|
||||
Στην εργασία αυτή υλοποιούμε τόσο μια σειριακή προσέγγιση όσο και μια που παραλληλοποιείται με \textbf{pthreads}, με \textbf{open-cilk} και με \textbf{open-mp}.
|
||||
Στην παραλληλοποιήσιμη έκδοση δεν μας ενδιαφέρει η ακρίβεια του αλγορίθμου, κάτι που μας δίνει την δυνατότητα να βελτιώσουμε τον χρόνο εκτέλεσης.
|
||||
|
||||
|
||||
\section{Παραδοτέα}
|
||||
Τα παραδοτέα της εργασίας αποτελούνται από:
|
||||
\begin{itemize}
|
||||
\item Την παρούσα αναφορά.
|
||||
\item Το \href{https://git.hoo2.net/hoo2/PDS/src/branch/master/homework_1}{σύνδεσμο με το αποθετήριο} που περιέχει τον κώδικα για την παραγωγή των εκτελέσιμων, της αναφοράς και τις μετρήσεις.
|
||||
\item Τον \href{https://hub.docker.com/r/hoo2/hpcimage}{σύνδεσμο} με το image με το οποίο μεταγλωττίσαμε και εκτελέσαμε τις μετρήσεις.
|
||||
\end{itemize}
|
||||
|
||||
\section {Υλοποίηση}
|
||||
Πριν ξεκινήσουμε να αναλύουμε τον παραλληλισμό και τις μετρήσεις καλό θα ήταν να αναφερθούμε στην υλοποίηση.
|
||||
Για την παρούσα εργασία χρησιμοποιήσαμε τη γλώσσα \textbf{C++} και πέρα από τις βιβλιοθήκες που ζητούνται από την εκφώνηση (cilk/omp) έγινε και χρήση της \textbf{openblas}.
|
||||
Βοηθητικά έγινε επίσης χρήση του \textbf{bash} και της \textbf{matlab} κυρίως για την λήψη και οπτικοποίηση των αποτελεσμάτων.
|
||||
Στην παρούσα εργασία γίνεται ευρεία χρήση πινάκων τόσο για εισαγωγή και εξαγωγή των δεδομένων, όσο και για τους υπολογισμούς.
|
||||
Για το λόγο αυτό θεωρήσαμε ότι έπρεπε πρώτα να υλοποιήσουμε κάποιες αφαιρέσεις που θα μας επέτρεπαν να χειριστούμε ευκολότερα το πρόβλημα
|
||||
|
||||
\subsection{Ο τύπος Matrix<>}
|
||||
Η πιο βασική αφαίρεση που υλοποιήσαμε είναι ο τύπος \textbf{Matrix<>}.
|
||||
Πρόκειται για μια αναπαράσταση πινάκων, είτε μονοδιάστατων είτε δισδιάστατων.
|
||||
Η υλοποίησή του έγινε με χρήση templates και έτσι μπορούμε να κατασκευάζουμε πίνακες οποιουδήποτε τύπου.
|
||||
Η αφαίρεση εσωτερικά κάνει χρίση μονοδιάστατης αποθήκευσης των δεδομένων με αποτέλεσμα να έχουμε την δυνατότητα να διαλέγουμε εμείς το ordering του πίνακα.
|
||||
Αν είναι δηλαδή column-major ή row-major.
|
||||
Στην αφαίρεση δημιουργούμε πίνακες στους οποίους τον έλεγχο της μνήμης τον έχουμε εσωτερικά στο αντικείμενο.
|
||||
Για την αποθήκευση χρησιμοποιούμε \textbf{std::vector} κάτι που μας δίνει ευελιξία στο μέγεθος αλλά και στη χρήση.
|
||||
\par
|
||||
Η κύρια όμως λειτουργικότητα που καθιστά την αφαίρεση βολική για την συγκεκριμένη εργασία είναι η δυνατότητα να κατασκευάσουμε αντικείμενα των οποίων η μνήμη είναι και εξωτερικά του αντικειμένου.
|
||||
Να λειτουργήσει δηλαδή ο ίδιος ο τύπος ώς viewer, όπως για παράδειγμα το \href{https://en.cppreference.com/w/cpp/header/string_view}{string view}, χωρίς όμως να χρειάζεται να έχουμε διαφορετικό τύπο για αυτή τη δουλειά.
|
||||
Σε αυτή την περίπτωση στο εσωτερικό storage δεν αποθηκεύουμε παρά μόνο μια διεύθυνση στα δεδομένα του άλλου πίνακα.
|
||||
Μπορούμε όμως να χρησιμοποιήσουμε όλη την λειτουργικότητα του πίνακα.
|
||||
Αυτό μας δίνει τη δυνατότητα να δημιουργούμε αντικείμενα που “βλέπουν” σε ολόκληρους ή και σε \textbf{slices }από πίνακες.
|
||||
|
||||
Για παράδειγμα μπορούμε να γράψουμε:
|
||||
\begin{verbatim}
|
||||
Matrix<double> A(4,5); // Πίνακας 4x5
|
||||
Matrix<double> vA(A, 0, 2, 5); // View στις 2 πρώτες γραμμές του Α
|
||||
for (int i=0 ; i<vA.rows() ; ++i)
|
||||
for (int j=0 ; j<vA.columns() ; ++j)
|
||||
std::cout << A(i, j);
|
||||
\end{verbatim}
|
||||
|
||||
Στον παραπάνω παράδειγμα ο πίνακας με τα δεδομένα είναι ο \textbf{Α}.
|
||||
o \textbf{vA}, είναι ένα \textit{slice} του A, που δείχνει μόνο στις 2 πρώτες γραμμές του Α.
|
||||
|
||||
\subsection{knnsearch version 0}
|
||||
Για το πρώτο σκέλος της άσκησης είχαμε να παρουσιάσουμε την knnsearch με σειριακό τρόπο.
|
||||
Για το σκοπό αυτό υλοποιήσαμε την \textbf{v0::knnsearch}.
|
||||
Η συνάρτηση αυτή κάνει χρήση του παραπάνω τύπου (Matrix) και επομένως είναι και ή ίδια template.
|
||||
Η λειτουργία της χωρίζεται σε 3 βασικά μέρη:
|
||||
\begin{itemize}
|
||||
\item Τον υπολογισμό του \textbf{πίνακα αποστάσεων}.
|
||||
Εδώ βρίσκεται ο κύριος υπολογιστικός όγκος της συνάρτησης, λόγω του πολλαπλασιασμού πινάκων που απαιτείται.
|
||||
Για το λόγο αυτό κάνουμε χρήση της openblas.
|
||||
\item Την \textbf{αντιστοίχηση των αποστάσεων} των σημείων \textbf{με τους δείκτες (indexes)} αυτών των αποστάσεων στον αρχικό πίνακα Corpus.
|
||||
Εδώ ουσιαστικά \textit{“ζευγαρώνουμε”} τον κάθε δείκτη με τη απόσταση που του αναλογεί σε μια δομή \textbf{std::pair<>}.
|
||||
Αυτό το κάνουμε ώστε ταξινομώντας ένα διάνυσμα με τέτοια ζευγάρια ώς προς τη μικρότερη απόσταση, να ταξινομούνται αυτόματα και οι δείκτες.
|
||||
Γιαυτό το σκοπό κατασκευάζουμε για κάθε σημείο του Query, ένα διάνυσμα τέτοιων ζευγαριών.
|
||||
\item Την \textbf{μερική ταξινόμηση} του παραπάνω διανύσματος.
|
||||
Εδώ κάνουμε χρήση του αλγορίθμου quick-select.
|
||||
Ουσιαστικά ταξινομούμε μόνο ένα κομμάτι του διανύσματος όσα τα στοιχεία που θέλουμε (τον αριθμό των γειτόνων που ψάχνουμε).
|
||||
Ο αλγόριθμος αυτός είναι γραμμικός, περιορίζοντας έτσι λιγάκι τη “χασούρα”.
|
||||
Για το σκοπό αυτό χρησιμοποιήσαμε την std::nth\_element().
|
||||
\end{itemize}
|
||||
Στην υλοποίησή μας τα Corpus-Query είναι διαφορετικά, κάτι που δεν μας περιορίζει να λύσουμε το πρόβλημα όπου το Corpus == Query φυσικά.
|
||||
|
||||
\subsection{knnsearch version 1}
|
||||
Για το δεύτερο σκέλος της εργασίας είχαμε να παρουσιάσουμε την knnsearch την οποία θα παραλληλοποιήσουμε.
|
||||
Για το σκοπό αυτό υλοποιήσαμε την \textbf{v1::knnsearch}.
|
||||
Η αρχική μας προσέγγιση ήταν \textbf{αναδρομική}.
|
||||
Ουσιαστικά χωρίζαμε τα Corpus-Queries σε slices και για καθένα από αυτά καλούσαμε ξανά τον εαυτό μας.
|
||||
Η αναδρομή τερματιζόταν όταν τα slices ήταν αρκετά μικρά ώστε να τρέξει ο σειριακός αλγόριθμος.
|
||||
Για να μην χάνονται όμως γείτονες η knnsearch έπρεπε να τρέχει για όλους τους συνδυασμούς.
|
||||
Το κάθε Corpus με τα δύο διαφορετικά Queries και το κάθε Query αντίστοιχα.
|
||||
Έπειτα συνενώναμε τα αποτελέσματα.
|
||||
Αν και αυτή η προσέγγιση ήταν η ακριβείς μας έδινε όμως τη δυνατότητα να “φτηνύνουμε” τον τρόπο με τον οποίο ενώναμε τα αποτελέσματα.
|
||||
Επίσης για την περίπτωση που το Query ήταν ίδιο με το Corpus, μας προσφέρεται η δυνατότητα να αποφύγουμε έναν από τους 4 συνδυασμούς.
|
||||
\par
|
||||
Η προσέγγιση αυτή, αν και απλή, δεν βολεύει για παραλληλοποίηση, καθώς τα νήματα που πρέπει να δημιουργηθούν βρίσκονται μέσα στην αναδρομή, καθιστώντας δύσχρηστη(όχι αδύνατη) τη χρήση των εργαλείων της cilk και της openMP.
|
||||
Ακόμα και το “μέρος” στο οποίο θα έμπαινε η loop-α με τις join για την περίπτωση των pthreads ήταν περίπλοκο.
|
||||
Για το σκοπό αυτό \textbf{“ανοίξαμε” την αναδρομή} σε βρόχο επανάληψης.
|
||||
Σε αυτή την υλοποίηση τα Corpus-Queries, χωρίζονται σε ένα αριθμό από slices που μας δίνεται από την γραμμή εντολών.
|
||||
Για όλα αυτά σε ένα διπλό βρόχο καλούμε την σειριακή knnsearch.
|
||||
Ο διπλός βρόχος ουσιαστικά δημιουργεί όλους τους συνδυασμούς των slice από το Corpus με τα slices από το Query.
|
||||
Έτσι πάλι δεν χάνουμε γείτονες.
|
||||
Τα αποτελέσματα συνενώνονται όπου και κρατούνται οι κοντινότεροι γείτονες.
|
||||
Η διαδικασία επιλογής είναι ίδια με τη σειριακή.
|
||||
Γίνεται δηλαδή πάλι χρήση της quick-select.
|
||||
Για να μειώσουμε την ακρίβεια και εδώ μπορούμε να “φτωχύνουμε” το merge.
|
||||
Για το σκοπό αυτό έχουμε ένα όρισμα στη συνάρτηση το οποίο μας κόβει τον αριθμό των κοντινότερων γειτόνων.
|
||||
Σε αυτή την προσέγγιση υπάρχει ακόμα ένα ουσιαστικότερο όφελος.
|
||||
Ο χωρισμός σε slices μας δίνει την δυνατότητα να δημιουργήσουμε threads που δεν θα χρειαστεί ποτέ να χρειαστούν κομμάτια μνήμης από άλλα threads.
|
||||
Αυτό ο διαχωρισμός γίνεται στον έξω βρόχο επανάληψης της knnsearch
|
||||
\begin{verbatim}
|
||||
#pragma omp parallel for
|
||||
for (size_t qi = 0; qi < num_slices; ++qi)
|
||||
worker_body (corpus_slices, query_slices, idx, dst, qi,
|
||||
num_slices, corpus_slice_size, query_slice_size, k, m);
|
||||
\end{verbatim}
|
||||
Για τα το παραπάνω snippet (που είναι και “καρδιά” του αλγόριθμου), ο κάθε worker δουλεύει σε δικό του corpus και query slice και αποθηκεύει τα δεδομένα σε δικό του slice πίσω στους idx και dst.
|
||||
Με αυτό τον τρόπο δεν χρειαζόμαστε κανέναν συγχρονισμό για τα threads.
|
||||
|
||||
\subsection{Μετρήσεις}
|
||||
|
||||
Για την παραγωγή των εκτελέσιμων χρησιμοποιήσαμε το προσωπικό μας(μου) laptop.
|
||||
Στον κατάλογο με τον πηγαίο κώδικα υπάρχει ότι είναι απαραίτητο για την μεταγλώττιση και εκτέλεση του κώδικα.
|
||||
Για τη μεταγλώττιση κάναμε χρήση του image hpcimage που βρίσκεται \href{https://hub.docker.com/r/hoo2/hpcimage}{εδώ}.
|
||||
Για την μεταγλώττιση και εκτέλεση μπορεί κάποιος να δώσει τις παρακάτω εντολές στο root κατάλογο “homework\_1”.
|
||||
\begin{verbatim}
|
||||
DOCK="docker run --rm -v <PATH_to_homework_1>:/usr/src/PDS_homework_1 \
|
||||
-w /usr/src/PDS_homework_1/ hoo2/hpcimage"
|
||||
make -j IMAGE=hoo2/hpcimage v0
|
||||
make -j IMAGE=hoo2/hpcimage v1
|
||||
make -j IMAGE=hoo2/hpcimage v1_cilk
|
||||
make -j IMAGE=hoo2/hpcimage v1_omp
|
||||
make -j IMAGE=hoo2/hpcimage v1_pth
|
||||
# run pthreads version with 4 slices
|
||||
eval $DOCK ./out/knnsearch_v1_pth -c <path to hdf5> /test -k 100 -s 4 -t
|
||||
\end{verbatim}
|
||||
|
||||
\section{Αποτελέσματα}
|
||||
|
||||
Παρακάτω παραθέτουμε τα αποτελέσματα των μετρήσεων.
|
||||
Για αυτές:
|
||||
\begin{itemize}
|
||||
\item Εκτελέσαμε την κάθε έκδοση του προγράμματος της v1 για\textbf{ διαφορετικό αριθμό από threads} για δύο διαφορετικούς πίνακες(all-to-all) ώστε να δούμε την συμπεριφορά.
|
||||
Τα threads μπορούν να περαστούν από τη γραμμή εντολών.
|
||||
\item Εκτελέσαμε την κάθε έκδοση του προγράμματος για δύο διαφορετικούς πίνακες(all-to-all) για \textbf{διαφορετική επιθυμητή ακρίβεια}.
|
||||
\end{itemize}
|
||||
|
||||
Για αντιπαραβολή παραθέτουμε ότι ότι στο ίδιο μηχάνημα το οποίο πήραμε τις μετρήσεις η knnsearch έκδοση του \textbf{matlab} κάνει:
|
||||
\begin{itemize}
|
||||
\item sift: \textbf{3.223628 [sec]}
|
||||
\item mnist: \textbf{20.634987 [sec]}
|
||||
\end{itemize}
|
||||
|
||||
\subsection{Επίδραση των threads - OpenMP}
|
||||
|
||||
Για την έκδοση με την openMP, χρησιμοποιώντας διαφορετικό αριθμό threads έχουμε: \\
|
||||
\includegraphics[width=\textwidth]{../measurements/OMP_over_threads.png}
|
||||
\captionof{figure}{Βελτιστοποίηση από την παραλληλοποίηση [openMP].}
|
||||
\label{fig:openMP_th}
|
||||
|
||||
\subsection{Επίδραση των threads - openCilk}
|
||||
|
||||
Για την έκδοση με την Cilk έχουμε: \\
|
||||
\includegraphics[width=\textwidth]{../measurements/CILK_over_threads.png}
|
||||
\captionof{figure}{Βελτιστοποίηση από την παραλληλοποίηση [open-cilk].}
|
||||
\label{fig:cilk_th}
|
||||
|
||||
\subsection{Επίδραση των threads - pthreads}
|
||||
|
||||
Για την έκδοση με τα pthreads έχουμε: \\
|
||||
\includegraphics[width=\textwidth]{../measurements/Pthreads_over_threads.png}
|
||||
\captionof{figure}{Βελτιστοποίηση από την παραλληλοποίηση [pthreads].}
|
||||
\label{fig:pthreads_th}
|
||||
|
||||
\subsection{Επίδραση της ακρίβειας - OpenMP}
|
||||
|
||||
Για την έκδοση με την openMP όσο αλλάζουμε την ακρίβεια έχουμε: \\
|
||||
\includegraphics[width=\textwidth]{../measurements/OMP_over_accuracy.png}
|
||||
\captionof{figure}{Βελτιστοποίηση λόγο μειωμένης ακρίβειας [openMP].}
|
||||
\label{fig:openMP_acc}
|
||||
|
||||
\subsection{Επίδραση της ακρίβειας - openCilk}
|
||||
|
||||
Για την έκδοση με την Cilk όσο αλλάζουμε την ακρίβεια έχουμε: \\
|
||||
\includegraphics[width=\textwidth]{../measurements/CILK_over_accuracy.png}
|
||||
\captionof{figure}{Βελτιστοποίηση λόγο μειωμένης ακρίβειας [open-cilk].}
|
||||
\label{fig:cilk_acc}
|
||||
|
||||
\subsection{Επίδραση της ακρίβειας - pthreads}
|
||||
|
||||
Για την έκδοση με τα pthreads όσο αλλάζουμε την ακρίβεια έχουμε: \\
|
||||
\includegraphics[width=\textwidth]{../measurements/Pthreads_over_accuracy.png}
|
||||
\captionof{figure}{Βελτιστοποίηση λόγο μειωμένης ακρίβειας [pthreads].}
|
||||
\label{fig:pthreads_acc}
|
||||
|
||||
|
||||
\section{Συμπεράσματα}
|
||||
|
||||
%\justifying
|
||||
|
||||
Απ' ότι βλέπουμε παραπάνω, ενώ ο χρόνος βελτιώνεται καθώς χρησιμοποιούμε παραπάνω threads, αυτό δεν συμβαίνει όταν μειώνουμε την ακρίβεια.
|
||||
Τουλάχιστον όχι όσο θα θέλαμε.
|
||||
Φυσικά αυτό έχει να κάνει με την υλοποίηση μας, καθώς δεν επιλέξαμε καλή τεχνική συνένωσης.
|
||||
Στην υλοποίησή μας δεν έγινε πολύ επιθετική στρατηγική μείωσης της ακρίβειας των αποτελεσμάτων.
|
||||
Να πούμε επίσης ότι ο συγκεκριμένος αλγόριθμος, λόγο της ουσιαστικά “μη” επικοινωνίας που χρειάζονται τα threads, θα μπορούσε να χρησιμοποιηθεί και σε απομακρυσμένα συστήματα κάνοντας χρήση του MPI.
|
||||
|
||||
|
||||
\end{document}
|
||||
|
||||
@ -65,9 +65,15 @@ bool get_options(int argc, char* argv[]){
|
||||
else if (arg == "-k") {
|
||||
session.k = (i+1 < argc) ? std::atoi(argv[++i]) : session.k;
|
||||
}
|
||||
else if (arg == "-n" || arg == "--max_trheads") {
|
||||
else if (arg == "-n" || arg == "--max_threads") {
|
||||
session.max_threads = (i+1 < argc) ? std::atoi(argv[++i]) : session.max_threads;
|
||||
}
|
||||
else if (arg == "-s" || arg == "--slices") {
|
||||
session.slices = (i+1 < argc) ? std::atoi(argv[++i]) : session.slices;
|
||||
}
|
||||
else if (arg == "-a" || arg == "--accuracy") {
|
||||
session.accuracy = (i+1 < argc) ? std::atoi(argv[++i]) : session.accuracy;
|
||||
}
|
||||
else if (arg == "-t" || arg == "--timing")
|
||||
session.timing = true;
|
||||
else if (arg == "-v" || arg == "--verbose")
|
||||
@ -87,7 +93,12 @@ bool get_options(int argc, char* argv[]){
|
||||
std::cout << " -k <number>\n";
|
||||
std::cout << " Set the number of closest neighbors to find. \n\n";
|
||||
std::cout << " -n | --max_trheads <threads>\n";
|
||||
std::cout << " Reduce the thread number for the execution to <threads>. <threads> must be less or equal to available CPUs.\n\n";
|
||||
std::cout << " Reduce the thread number for the execution to <threads>. <threads> should be less or equal to available CPUs.\n\n";
|
||||
std::cout << " -s | --slices <slices/threads>\n";
|
||||
std::cout << " The number of slices to the Corpus matrix. In the parallel version this setting affects the number of threads\n";
|
||||
std::cout << " <threads> should be less or equal to available CPUs\n\n";
|
||||
std::cout << " -a | --accuracy <accuracy>\n";
|
||||
std::cout << " Reduce the accuracy of neighbor finding. The accuracy should be between 1-100 \n\n";
|
||||
std::cout << " -t | --timing\n";
|
||||
std::cout << " Request timing measurements output to stdout.\n\n";
|
||||
std::cout << " -v | --verbose\n";
|
||||
@ -109,6 +120,42 @@ bool get_options(int argc, char* argv[]){
|
||||
return status;
|
||||
}
|
||||
|
||||
/*!
|
||||
* Matrix load
|
||||
*
|
||||
* \fn void loadMtx(MatrixDst&, MatrixDst&)
|
||||
* \param Corpus matrix to load to
|
||||
* \param Query matrix to load to
|
||||
*/
|
||||
void loadMtx(MatrixDst& Corpus, MatrixDst& Query) {
|
||||
if (access(session.outMtxFile.c_str(), F_OK) == 0)
|
||||
std::remove(session.outMtxFile.c_str());
|
||||
|
||||
// timer.start();
|
||||
Mtx::load<MatrixDst, DstHDF5Type>(session.corpusMtxFile, session.corpusDataSet, Corpus);
|
||||
if (session.queryMtx)
|
||||
Mtx::load<MatrixDst, DstHDF5Type>(session.corpusMtxFile, session.corpusDataSet, Query);
|
||||
// timer.stop();
|
||||
// timer.print_dt("Load hdf5 files");
|
||||
}
|
||||
|
||||
|
||||
/*!
|
||||
* Matrix store
|
||||
*
|
||||
* \fn void storeMtx(MatrixIdx&, MatrixDst&)
|
||||
* \param Idx Index part(neighbors) of the matrix to store
|
||||
* \param Dst Distances part of the matrix to store
|
||||
*/
|
||||
void storeMtx(MatrixIdx& Idx, MatrixDst& Dst) {
|
||||
// timer.start();
|
||||
Mtx::store<MatrixIdx, IdxHDF5Type>(session.outMtxFile, session.outMtxIdxDataSet, Idx);
|
||||
Mtx::store<MatrixDst, DstHDF5Type>(session.outMtxFile, session.outMtxDstDataSet, Dst);
|
||||
// timer.stop();
|
||||
// timer.print_dt("Store hdf5 files");
|
||||
}
|
||||
|
||||
|
||||
#ifndef TESTING
|
||||
int main(int argc, char* argv[]) try {
|
||||
// Instantiate matrixes
|
||||
@ -127,38 +174,26 @@ int main(int argc, char* argv[]) try {
|
||||
if (!get_options(argc, argv))
|
||||
exit(1);
|
||||
|
||||
if (access(session.outMtxFile.c_str(), F_OK) == 0)
|
||||
std::remove(session.outMtxFile.c_str());
|
||||
init_workers();
|
||||
|
||||
// Load data
|
||||
timer.start();
|
||||
Mtx::load<MatrixDst, DstHDF5Type>(session.corpusMtxFile, session.corpusDataSet, Corpus);
|
||||
if (session.queryMtx)
|
||||
Mtx::load<MatrixDst, DstHDF5Type>(session.corpusMtxFile, session.corpusDataSet, Query);
|
||||
timer.stop();
|
||||
timer.print_dt("Load hdf5 files");
|
||||
loadMtx(Corpus, Query);
|
||||
|
||||
// Prepare output memory
|
||||
Idx.resize(Query.rows(), session.k);
|
||||
Dst.resize(Query.rows(), session.k);
|
||||
Idx.resize((session.queryMtx) ? Query.rows() : Corpus.rows(), session.k);
|
||||
Dst.resize((session.queryMtx) ? Query.rows() : Corpus.rows(), session.k);
|
||||
|
||||
// Do the search
|
||||
logger << "Start knnsearch ...";
|
||||
timer.start();
|
||||
if (session.queryMtx)
|
||||
knnsearch(Corpus, Query, 0, session.k, session.k, Idx, Dst);
|
||||
else
|
||||
knnsearch(Corpus, Corpus, 0, session.k, session.k, Idx, Dst);
|
||||
size_t selected_neighbors = (size_t)(session.k*(session.accuracy/100.0));
|
||||
knnsearch(Corpus, (session.queryMtx) ? Query : Corpus, session.slices, session.k, selected_neighbors, Idx, Dst);
|
||||
timer.stop();
|
||||
logger << " Done" << logger.endl;
|
||||
timer.print_dt("knnsearch");
|
||||
|
||||
// Store data
|
||||
timer.start();
|
||||
Mtx::store<MatrixIdx, IdxHDF5Type>(session.outMtxFile, session.outMtxIdxDataSet, Idx);
|
||||
Mtx::store<MatrixDst, DstHDF5Type>(session.outMtxFile, session.outMtxDstDataSet, Dst);
|
||||
timer.stop();
|
||||
timer.print_dt("Store hdf5 files");
|
||||
storeMtx(Idx, Dst);
|
||||
|
||||
return 0;
|
||||
}
|
||||
|
||||
54
homework_1/src/v1.cpp
Normal file
54
homework_1/src/v1.cpp
Normal file
@ -0,0 +1,54 @@
|
||||
/**
|
||||
* \file v1.hpp
|
||||
* \brief
|
||||
*
|
||||
* \author
|
||||
* Christos Choutouridis AEM:8997
|
||||
* <cchoutou@ece.auth.gr>
|
||||
*/
|
||||
|
||||
#include "v1.hpp"
|
||||
|
||||
/*!
|
||||
* \fn void init_workers()
|
||||
*
|
||||
* Initialize worker settings
|
||||
*/
|
||||
void init_workers() {
|
||||
#if defined CILK
|
||||
size_t cilk_w = __cilkrts_get_nworkers();
|
||||
|
||||
if (!session.max_threads)
|
||||
session.max_threads = (session.slices) ? (session.slices) : cilk_w;
|
||||
// else if (session.max_threads < cilk_w)
|
||||
// __cilkrts_set_param("nworkers", "4");
|
||||
// else ignored by cilk
|
||||
|
||||
#elif defined OMP
|
||||
// omp_set_dynamic(1);
|
||||
size_t omp_w = (size_t)omp_get_max_threads();
|
||||
|
||||
if (!session.max_threads) {
|
||||
session.max_threads = (session.slices) ? (session.slices) : omp_w;
|
||||
// omp_set_dynamic(1);
|
||||
}
|
||||
else if (session.max_threads < omp_w) {
|
||||
// omp_set_dynamic(0);
|
||||
omp_set_num_threads(session.max_threads);
|
||||
}
|
||||
// else ignored by omp
|
||||
|
||||
#elif defined PTHREADS
|
||||
size_t pth_w = std::thread::hardware_concurrency();
|
||||
|
||||
if (!session.max_threads)
|
||||
session.max_threads = (session.slices) ? (session.slices) : pth_w;
|
||||
#else
|
||||
|
||||
#endif
|
||||
if (!session.slices)
|
||||
session.slices = session.max_threads;
|
||||
|
||||
openblas_set_num_threads(1); // Limit OpenBLAS to 1 thread
|
||||
}
|
||||
|
||||
9
homework_1/test/ReadME.md
Normal file
9
homework_1/test/ReadME.md
Normal file
@ -0,0 +1,9 @@
|
||||
## NOTE ##
|
||||
This folder contains the output files from tests.
|
||||
|
||||
In order to populate it, run the tests
|
||||
|
||||
$> make tests
|
||||
$> ./tests
|
||||
|
||||
|
||||
BIN
homework_1/test/knn_v0.hdf5
Normal file
BIN
homework_1/test/knn_v0.hdf5
Normal file
Binary file not shown.
BIN
homework_1/test/knn_v1ser.hdf5
Normal file
BIN
homework_1/test/knn_v1ser.hdf5
Normal file
Binary file not shown.
Loading…
x
Reference in New Issue
Block a user