// Copyright Lin Min 2014
#include "catch/catch.hpp"
#include "dispatch/runnable.hpp"
#include "composite/layers/concat_layer.hpp"
#include "composite/layers/split_layer.hpp"
using namespace purine;
TEST_CASE("TestConcat", "[Concat]") {
Runnable run(0, 0);
Blob* data = run.create("data", { 32, 32, 32, 32 });
Blob* data_diff = run.create("data_diff", { 32, 32, 32, 32 });
Blob* data1 = run.create("data1", { 32, 32, 32, 32 });
Blob* data1_diff = run.create("data1_diff", { 32, 32, 32, 32 });
Blob* data2 = run.create("data2", { 32, 32, 32, 32 });
Blob* data2_diff = run.create("data2_diff", { 32, 32, 32, 32 });
ConcatLayer* concat = run.createGraph<ConcatLayer>("concat",
ConcatLayer::param_tuple(Split::CHANNELS));
vector<Blob*>{ data, data1, data2, data_diff, data1_diff, data2_diff }
>> *concat;
vector<Blob*> top = concat->top();
REQUIRE(top[0]->tensor()->size() == Size(32, 96, 32, 32));
REQUIRE(top[1]->tensor()->size() == Size(32, 96, 32, 32));
REQUIRE(top[0]->tensor()->stride() == Stride(98304, 1024, 32, 1));
REQUIRE(top[1]->tensor()->stride() == Stride(98304, 1024, 32, 1));
REQUIRE(data->tensor()->stride() == Stride(98304, 1024, 32, 1));
REQUIRE(data1->tensor()->stride() == Stride(98304, 1024, 32, 1));
REQUIRE(data2->tensor()->stride() == Stride(98304, 1024, 32, 1));
REQUIRE(data_diff->tensor()->stride() == Stride(98304, 1024, 32, 1));
REQUIRE(data1_diff->tensor()->stride() == Stride(98304, 1024, 32, 1));
REQUIRE(data2_diff->tensor()->stride() == Stride(98304, 1024, 32, 1));
* device: -1, 0, 1 -1, 0, 1
*/
TEST_CASE("TestCopy", "[Copy]") {
Runnable g;
SECTION("Single Copy") {
SECTION("Within rank") {
int rank = current_rank();
/**
* cpu <-> gpu
* (0, -1) -> (0, 0)
* (0, 0) -> (0, -1)
*/
SECTION("cpu <-> gpu") {
Op<Constant>* c = g.create<Constant>("constant", rank, -1, "main",
Constant::param_tuple(2.));
Blob* twos_cpu = g.create("twos_cpu", rank, -1, {1, 10, 10, 10});
Blob* twos_gpu = g.create("twos_gpu", rank, 0, {1, 10, 10, 10});
Blob* dest_cpu = g.create("dest_cpu", rank, -1, {1, 10, 10, 10});
Copy* cp = g.createAny<Copy>("cp", Copy::param_tuple());
Copy* cp2 = g.createAny<Copy>("cp2", Copy::param_tuple());
*c >> B{ twos_cpu } >> *cp >> B{ twos_gpu } >> *cp2 >> B{ dest_cpu };
REQUIRE(cp->nodes().size() == 1);
REQUIRE(cp2->nodes().size() == 1);
REQUIRE(g.nodes().size() == 6);
g.run();
REQUIRE(caffe::purine_cpu_compare(twos_cpu->tensor()->cpu_data(),
dest_cpu->tensor()->cpu_data(),
twos_cpu->tensor()->shape().Count()));
#include "operations/include/eltwise.hpp"
#include "operations/include/mem_copy.hpp"
#include "dispatch/graph_template.hpp"
#include "dispatch/op_template.hpp"
#include "dispatch/runnable.hpp"
#include "composite/layers/conv_layer.hpp"
using namespace purine;
typedef vector<Blob*> B;
TEST_CASE("TestGraph", "[Graph]") {
Runnable test_graph;
Op<Conv>* o = test_graph.create<Conv>("conv", "main",
Conv::param_tuple(2, 2, 1, 1));
Blob* bottom = test_graph.create("bottom", 0, 0, {1, 3, 10, 10});
Blob* top = test_graph.create("top", 0, 0, {1, 3, 10, 10});
Blob* weight = test_graph.create("weight", 0, 0, {3, 3, 5, 5});
B{ bottom, weight } >> (*o) >> B{ top };
SECTION("graph node number test") {
vector<Node*> nodes = test_graph.nodes();
int node_size = test_graph.nodes().size();
int source_size = test_graph.sources().size();
int sink_size = test_graph.sinks().size();
REQUIRE(node_size == 4);
REQUIRE(source_size == 2);
REQUIRE(sink_size == 1);
}
SECTION("graph sources and sinks test") {
vector<Node*>&& nodes = test_graph.nodes();
#include "operations/include/mpi.hpp"
#include "dispatch/op_template.hpp"
#include "dispatch/graph_template.hpp"
#include "dispatch/runnable.hpp"
#include "dispatch/blob.hpp"
using namespace purine;
using namespace std;
typedef vector<Blob*> B;
TEST_CASE("TestMPI", "[MPI][Thread]") {
Runnable g;
SECTION("Isend and Irecv") {
int rank = current_rank();
Blob* src = g.create("src", 0, -1, {1, 3, 10, 10});
Blob* dest = g.create("dest", 1, -1, {1, 3, 10, 10});
Op<Constant>* c = g.create<Constant>("constant", 0, -1, "main",
Constant::param_tuple(1.));
Op<Isend>* send = g.create<Isend>("send", 0, -1, "main",
Isend::param_tuple(0, dest->rank()));
Op<Irecv>* recv = g.create<Irecv>("recv", 1, -1, "main",
Irecv::param_tuple(0, src->rank()));
// connect
*c >> B{ src };
B{ src } >> *send;
*recv >> B{ dest };
// run
g.run();
Tensor* t = dest->tensor();
if (current_rank() == 1) {