gaspersic/triton
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Added GEMM benchmark

Browse Source
This commit is contained in:
Philippe Tillet
2015-01-25 18:19:19 -05:00
parent 4a9e16fefd
commit 6d7f103ab1
7 changed files with 132 additions and 75 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
bench/CMakeLists.txt
View File

@@ -17,7 +17,7 @@ include_directories(/opt/intel/system_studio_2015.1.045/mkl/include/)
set(BLAS_DEF ${BLAS_DEF} "-DBENCH_CBLAS")
set(MKLROOT /opt/intel/system_studio_2015.1.045/mkl/)
set(CCROOT /opt/intel/system_studio_2015.1.045/compiler/)
set(BLAS_LIBS ${MKLROOT}/lib/intel64/libmkl_intel_lp64.a ${MKLROOT}/lib/intel64/libmkl_intel_thread.a ${BLAS_LIBS} ${MKLROOT}/lib/intel64/libmkl_core.a ${CCROOT}/lib/intel64/libiomp5.so pthread dl m)
set(BLAS_LIBS ${MKLROOT}/lib/intel64/libmkl_intel_lp64.so ${MKLROOT}/lib/intel64/libmkl_avx.so ${MKLROOT}/lib/intel64/libmkl_intel_thread.so ${BLAS_LIBS} ${MKLROOT}/lib/intel64/libmkl_core.so ${CCROOT}/lib/intel64/libiomp5.so pthread dl m)
string(REPLACE ";" " " BLAS_DEF_STR "${BLAS_DEF}")

139
bench/blas.cpp
View File

@@ -33,62 +33,119 @@ void bench(ad::numeric_type dtype)
total_time += times.back();\
}\
float tres = median(times);\
std::cout << " " << PERF(N, tres, dtsize) << std::flush;\
std::cout << " " << PERF << std::flush;\
}
/*---------*/
/*--BLAS1--*/
/*---------*/
std::cout << "#AXPY" << std::endl;
for(std::vector<int_t>::const_iterator it = BLAS1_N.begin() ; it != BLAS1_N.end() ; ++it)
{
int_t N = *it;
std::cout << N;
/* ATIDLAS */
atidlas::array x(N, dtype), y(N, dtype);
BENCHMARK(y = x + y, bandwidth);
/* clAmdBlas */
#ifdef BENCH_CLAMDBLAS
BENCHMARK(clAmdBlasSaxpy(N, 1, x.data()(), 0, 1, y.data()(), 0, 1, 1, &atidlas::cl::get_queue(x.context(), 0)(), 0, NULL, NULL), bandwidth)
#endif
/* BLAS */
#ifdef BENCH_CBLAS
std::vector<float> cx(N), cy(N);
atidlas::copy(x, cx);
atidlas::copy(y, cy);
BENCHMARK(cblas_saxpy(N, 1, cx.data(), 1, cy.data(), 1), bandwidth);
#endif
std::cout << std::endl;
}
std::cout << "\n\n" << std::flush;
// //DOT
// BENCH(DECLARE(ad::pointed_scalar s(dtype)); DECLARE(ad::vector, x(*it, dtype), y(*it, dtype)), s = dot(x, y), BLAS1_N, bandwidth, 2*(*it), "dot");
// /*---------*/
// /*--BLAS1--*/
// /*---------*/
// std::cout << "#AXPY" << std::endl;
// for(std::vector<int_t>::const_iterator it = BLAS1_N.begin() ; it != BLAS1_N.end() ; ++it)
// {
// int_t N = *it;
// std::cout << N;
// /* ATIDLAS */
// atidlas::array x(N, dtype), y(N, dtype);
// BENCHMARK(y = x + y, bandwidth(3*N, tres, dtsize));
// /* clAmdBlas */
//#ifdef BENCH_CLAMDBLAS
// BENCHMARK(clAmdBlasSaxpy(N, 1, x.data()(), 0, 1, y.data()(), 0, 1, 1, &atidlas::cl::get_queue(x.context(), 0)(), 0, NULL, NULL), bandwidth(3*N, tres, dtsize))
//#endif
// /* BLAS */
//#ifdef BENCH_CBLAS
// std::vector<float> cx(N), cy(N);
// atidlas::copy(x, cx);
// atidlas::copy(y, cy);
// BENCHMARK(cblas_saxpy(N, 1, cx.data(), 1, cy.data(), 1), bandwidth(3*N, tres, dtsize));
//#endif
// std::cout << std::endl;
// }
// std::cout << "\n\n" << std::flush;
// std::cout << "#DOT" << std::endl;
// for(std::vector<int_t>::const_iterator it = BLAS1_N.begin() ; it != BLAS1_N.end() ; ++it)
// {
// int_t N = *it;
// std::cout << N;
// /* ATIDLAS */
// atidlas::array x(N, dtype), y(N, dtype);
// atidlas::array scratch(N, dtype);
// atidlas::scalar s(dtype);
// BENCHMARK(s = dot(x,y), bandwidth(2*N, tres, dtsize));
// /* clAmdBlas */
//#ifdef BENCH_CLAMDBLAS
// BENCHMARK(clAmdBlasSdot(N, s.data()(), 0, x.data()(), 0, 1, y.data()(), 0, 1, scratch.data()(), 1, &atidlas::cl::get_queue(x.context(), 0)(), 0, NULL, NULL), bandwidth(2*N, tres, dtsize))
//#endif
// /* BLAS */
//#ifdef BENCH_CBLAS
// std::vector<float> cx(N), cy(N);
// atidlas::copy(x, cx);
// atidlas::copy(y, cy);
// BENCHMARK(cblas_sdot(N, cx.data(), 1, cy.data(), 1), bandwidth(2*N, tres, dtsize));
//#endif
// std::cout << std::endl;
// }
// std::cout << "\n\n" << std::flush;
// /*---------*/
// /*--BLAS2--*/
// /*---------*/
// //N-layout
// for(std::vector<int>::const_iterator Mit = BLAS2_M.begin() ; Mit != BLAS2_M.end() ; ++Mit)
// {
// BENCH(DECLARE(atidlas::matrix, A(*Mit,*it)); DECLARE(atidlas::vector, y(*Mit), x(*it)),ARGS(y, viennacl::op_assign(), viennacl::linalg::prod(A,x)), BLAS2_N,
// bandwidth, (*Mit)*(*it), "row-wise-reductionN-float32");
// }
// //T-layout
// std::cout << "#GEMV-T" << std::endl;
// for(std::vector<int>::const_iterator Mit = BLAS2_M.begin() ; Mit != BLAS2_M.end() ; ++Mit)
// for(std::vector<int_t>::const_iterator Nit = BLAS2_N.begin() ; Nit != BLAS2_N.end() ; ++Nit)
// {
// BENCH(DECLARE(atidlas::matrix, A(*it,*Mit)) ; DECLARE(atidlas::vector, y(*Mit), x(*it)), ARGS(y, viennacl::op_assign(), viennacl::linalg::prod(viennacl::trans(A),x)), BLAS2_N,
// bandwidth, (*Mit)*(*it), "row-wise-reductionT-float32");
// int_t M = *Mit;
// int_t N = *Nit;
// std::cout << M << "," << N;
// /* ATIDLAS */
// atidlas::array A(N, M, dtype), y(M, dtype), x(N, dtype);
// BENCHMARK(y = dot(trans(A),x), bandwidth(M*N + M + N, tres, dtsize));
// /* clAmdBlas */
// #ifdef BENCH_CLAMDBLAS
// BENCHMARK(clAmdBlasSgemv(clAmdBlasColumnMajor, clAmdBlasTrans, N, M, 1, A.data()(), A.ld(), x.data()(), 0, 1, 0, y.data()(), 0, 1, 1, &atidlas::cl::get_queue(x.context(), 0)(),0, NULL, NULL), bandwidth(M*N + M + N, tres, dtsize))
// #endif
// /* BLAS */
// #ifdef BENCH_CBLAS
// std::vector<float> cA(N*M), cx(N), cy(M);
// atidlas::copy(x, cx);
// atidlas::copy(y, cy);
// atidlas::copy(A, cA);
// BENCHMARK(cblas_sgemv(CblasColMajor, CblasTrans, N, M, 1, cA.data(), N, cx.data(), 1, 0, cy.data(), 1), bandwidth(M*N + M + N, tres, dtsize));
// #endif
// std::cout << std::endl;
// }
// std::cout << "\n\n" << std::flush;
// /*---------*/
// /*--BLAS3--*/
// /*---------*/
std::cout << "#GEMM-NT" << std::endl;
for(std::vector<int_t>::const_iterator Mit = BLAS3_M.begin() ; Mit != BLAS3_M.end() ; ++Mit)
for(std::vector<int_t>::const_iterator Nit = BLAS3_N.begin() ; Nit != BLAS3_N.end() ; ++Nit)
for(std::vector<int_t>::const_iterator Kit = BLAS3_K.begin() ; Kit != BLAS3_K.end() ; ++Kit)
{
int_t M = *Mit, N = *Nit, K = *Kit;
std::cout << M << "," << N << "," << K;
/* ATIDLAS */
atidlas::array C(M, N, dtype), A(M, K, dtype), B(N, K, dtype);
BENCHMARK(C = dot(A,trans(B)), gflops((double)2*M*N*K, tres));
/* clAmdBlas */
#ifdef BENCH_CLAMDBLAS
BENCHMARK(clAmdBlasSgemm(clAmdBlasColumnMajor, clAmdBlasNoTrans, clAmdBlasTrans, M, N, K, 1, A.data()(), A.ld(), B.data()(), B.ld(),
0, C.data()(), C.ld(), 1, &atidlas::cl::get_queue(C.context(), 0)(),0, NULL, NULL), gflops((double)2*M*N*K, tres))
#endif
/* BLAS */
#ifdef BENCH_CBLAS
std::vector<float> cC(M*N), cA(M*K), cB(N*K);
atidlas::copy(C, cC);
atidlas::copy(A, cA);
atidlas::copy(B, cB);
BENCHMARK(cblas_sgemm(CblasColMajor, CblasNoTrans, CblasTrans, M, N, K, 1, cA.data(), M, cB.data(), N, 1, cC.data(), M), gflops((double)2*M*N*K, tres));
#endif
std::cout << std::endl;
}
}
int main(int argc, char* argv[])

11
bench/common.hpp
View File

@@ -51,13 +51,16 @@ static const std::vector<int> BLAS2_M = make_vector<int>() << 256;
static const std::vector<int> BLAS2_N = create_full_range(128, 5000, 64);
// BLAS3 Sizes
static const std::vector<int> BLAS3_N = create_full_range(128, 5000, 64);
static const std::vector<int> BLAS3_M = make_vector<int>() << 128;
static const std::vector<int> BLAS3_N = make_vector<int>() << 128;
static const std::vector<int> BLAS3_K = create_full_range(128, 5000, 64);
float bandwidth(std::size_t N, float t, unsigned int dtsize)
{
return N * dtsize * 1e-9 / t;
}
{ return N * dtsize * 1e-9 / t; }
double gflops(double nops, double t)
{ return nops * 1e-9 / t; }
template<class T>
T median(std::vector<T> x)

1
include/atidlas/model/model.h
View File

@@ -42,7 +42,6 @@ namespace atidlas
typedef std::map<std::pair<expression_type, numeric_type>, tools::shared_ptr<model> > model_map_t;
model_map_t import(std::string const & fname);
model_map_t init_models(cl::CommandQueue const & queue);
model_map_t& get_model_map(cl::CommandQueue & queue);
model& get_model(cl::CommandQueue & queue, expression_type, numeric_type);

3
lib/cl/queues.cpp
View File

@@ -49,7 +49,8 @@ std::vector<cl::CommandQueue> & get_queues(cl::Context const & ctx)
for(queues_t::iterator it = queues.begin() ; it != queues.end() ; ++it)
if(it->first()==ctx())
return it->second;
throw std::out_of_range("The context provided is not registered");
queues.push_back(std::make_pair(ctx, std::vector<cl::CommandQueue>(1, cl::CommandQueue(ctx, ctx.getInfo<CL_CONTEXT_DEVICES>()[0]))));
return queues.back().second;
}
cl::CommandQueue & get_queue(cl::Context const & ctx, std::size_t idx)

23
lib/model/model.cpp
View File

@@ -143,7 +143,7 @@ namespace detail
{
if(name=="vaxpy") return VECTOR_AXPY_TYPE;
if(name=="reduction") return REDUCTION_TYPE;
if(name=="matrix-axpy") return MATRIX_AXPY_TYPE;
if(name=="maxpy") return MATRIX_AXPY_TYPE;
if(name=="row-wise-reductionN") return ROW_WISE_REDUCTION_TYPE;
if(name=="row-wise-reductionT") return COL_WISE_REDUCTION_TYPE;
if(name=="matrix-productNN") return MATRIX_PRODUCT_NN_TYPE;
@@ -167,7 +167,7 @@ namespace detail
return tools::shared_ptr<base>(new vaxpy( vaxpy_parameters(a[0], a[1], a[2], fetch[a[3]])));
else if(template_name=="reduction")
return tools::shared_ptr<base>(new reduction( reduction_parameters(a[0], a[1], a[2], fetch[a[3]])));
else if(template_name=="matrix-axpy")
else if(template_name=="maxpy")
return tools::shared_ptr<base>(new maxpy( maxpy_parameters(a[0], a[1], a[2], a[3], a[4], fetch[a[5]])));
else if(template_name.find("row-wise-reduction")!=std::string::npos)
{
@@ -185,11 +185,9 @@ namespace detail
}
}
model_map_t import(std::string const & fname, cl::CommandQueue & queue)
void import(std::string const & fname, cl::CommandQueue & queue, model_map_t& result)
{
namespace js = rapidjson;
model_map_t result;
//Parse the JSON document
js::Document document;
std::ifstream t(fname.c_str());
@@ -201,7 +199,7 @@ model_map_t import(std::string const & fname, cl::CommandQueue & queue)
document.Parse<0>(str.c_str());
//Deserialize
std::vector<std::string> operations = tools::make_vector<std::string>() << "vaxpy" << "reduction"
<< "matrix-axpy" << "row-wise-reductionN" << "row-wise-reductionT"
<< "maxpy" << "row-wise-reductionN" << "row-wise-reductionT"
<< "matrix-productNN" << "matrix-productTN" << "matrix-productNT" << "matrix-productTT";
std::vector<std::string> dtype = tools::make_vector<std::string>() << "float32" << "float64";
for(std::vector<std::string>::iterator op = operations.begin() ; op != operations.end() ; ++op)
@@ -222,19 +220,18 @@ model_map_t import(std::string const & fname, cl::CommandQueue & queue)
js::Value const & profiles = document[opcstr][dtcstr]["profiles"];
for (js::SizeType id = 0 ; id < profiles.Size() ; ++id)
templates.push_back(detail::create(*op, tools::to_int_array<int>(profiles[id])));
if(templates.size()>1){
if(templates.size()>1)
{
// Get predictor
predictors::random_forest predictor(document[opcstr][dtcstr]["predictor"]);
result[std::make_pair(etype, dtype)] = tools::shared_ptr<model>(new model(predictor, templates, queue));
}else{
}
else
result[std::make_pair(etype, dtype)] = tools::shared_ptr<model>(new model(templates, queue));
}
}
}
}
}
return result;
}
model_map_t init_models(cl::CommandQueue & queue)
@@ -243,7 +240,7 @@ model_map_t init_models(cl::CommandQueue & queue)
typedef tools::shared_ptr<model> ptr_t;
numeric_type types[] = {CHAR_TYPE, UCHAR_TYPE, SHORT_TYPE, USHORT_TYPE, INT_TYPE, UINT_TYPE, LONG_TYPE, ULONG_TYPE, FLOAT_TYPE, DOUBLE_TYPE};
for(size_t i = 0 ; i < 1 ; ++i){
for(size_t i = 0 ; i < 10 ; ++i){
numeric_type DTYPE = types[i];
res[std::make_pair(SCALAR_AXPY_TYPE, DTYPE)] = ptr_t(new model(vaxpy(1,64,128,FETCH_FROM_GLOBAL_STRIDED), queue));
res[std::make_pair(VECTOR_AXPY_TYPE, DTYPE)] = ptr_t (new model(vaxpy(1,64,128,FETCH_FROM_GLOBAL_STRIDED), queue));
@@ -257,7 +254,7 @@ model_map_t init_models(cl::CommandQueue & queue)
res[std::make_pair(MATRIX_PRODUCT_TT_TYPE, DTYPE)] = ptr_t(new model(mproduct_tt(1, 8, 8, 8, 4, 1, 4, FETCH_FROM_LOCAL, FETCH_FROM_LOCAL, 8, 8), queue));
}
if(const char * cmodel_file = std::getenv("ATIDLAS_MODEL_DEVICE_0"))
return import(std::string(cmodel_file), queue);
import(std::string(cmodel_file), queue, res);
return res;

24
python/pyatidlas/src/_atidlas.cpp
View File

@@ -77,17 +77,17 @@ bp::tuple get_shape(atd::array const & x)
// x.reshape(size1, size2);
//}
boost::python::dict create_queues(atd::cl::queues_t queues)
{
boost::python::dict dictionary;
for (atd::cl::queues_t::iterator it = queues.begin(); it != queues.end(); ++it) {
bp::list list;
for (atd::cl::queues_t::mapped_type::iterator itt = it->second.begin(); itt != it->second.end(); ++itt)
list.append(*itt);
dictionary[it->first] = list;
}
return dictionary;
}
//boost::python::dict create_queues(atd::cl::queues_t queues)
//{
// boost::python::dict dictionary;
// for (atd::cl::queues_t::iterator it = queues.begin(); it != queues.end(); ++it) {
// bp::list list;
// for (atd::cl::queues_t::mapped_type::iterator itt = it->second.begin(); itt != it->second.end(); ++itt)
// list.append(*itt);
// dictionary[it->first] = list;
// }
// return dictionary;
//}
template<class T>
struct datatype : public atd::value_scalar
@@ -198,7 +198,7 @@ namespace detail
}
std::vector<atd::cl::CommandQueue> & get_queue(atd::cl::Context const & ctx)
{ return atd::cl::queues[ctx]; }
{ return atd::cl::get_queues(ctx); }
atd::numeric_type extract_dtype(bp::object const & odtype)
{