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

GEMV: bugfix with CUDA

Browse Source
This commit is contained in:
Philippe Tillet
2015-08-30 02:23:40 -04:00
parent b8f3e08c68
commit 836a955663
5 changed files with 51 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

57
bench/blas.cpp
View File

@@ -41,7 +41,7 @@ void bench(sc::numeric_type dtype, std::string operation)
{\
std::vector<double> times;\
double total_time = 0;\
while(total_time*1e-9 < 1e-2){\
while(total_time*1e-9 < 1e-1){\
std::list<sc::driver::Event> events;\
OP;\
queue.synchronize();\
@@ -56,7 +56,7 @@ void bench(sc::numeric_type dtype, std::string operation)
{\
std::vector<long> times;\
double total_time = 0;\
while(total_time*1e-9 < 1e-2){\
while(total_time*1e-9 < 1e-1){\
cl_event event;\
OP;\
queue.synchronize();\
@@ -72,7 +72,7 @@ void bench(sc::numeric_type dtype, std::string operation)
Timer tmr;\
long total_time = 0;\
std::vector<long> times;\
while(total_time*1e-9 < 1e-2){\
while(total_time*1e-9 < 1e-1){\
tmr.start();\
OP;\
long time = tmr.get().count();\
@@ -93,7 +93,7 @@ void bench(sc::numeric_type dtype, std::string operation)
cudaEventCreate(&stop);\
OP;\
cudaThreadSynchronize();\
while(total_time*1e-3 < 1e-2){\
while(total_time*1e-3 < 1e-1){\
cudaEventRecord(start,0);\
OP;\
cudaEventRecord(stop,0);\
@@ -103,7 +103,7 @@ void bench(sc::numeric_type dtype, std::string operation)
total_time+=time;\
}\
double t = mean(times);\
std::cout << "\t" << (int)(PERF) << std::flush;\
std::cout << " " << (int)(PERF) << std::flush;\
}
unsigned int dtsize = sc::size_of(dtype);
@@ -111,16 +111,17 @@ void bench(sc::numeric_type dtype, std::string operation)
std::map<std::string, std::string> metric{ {"axpy", "GB/s"}, {"dot", "GB/s"}, {"gemv", "GB/s"}, {"gemm", "GFLOPS"}};
sc::array flush((int)1e6, sc::FLOAT_TYPE);
std::cout << "#" << operation << " (" << metric[operation] << ")" << std::endl;
std::cout << "N";
std::cout << "\tISAAC";
std::cout << "\"N\"";
std::cout << " \"ISAAC (Pred impl.)\"";
std::cout << " \"ISAAC (Best impl.)\"";
#ifdef BENCH_CLBLAS
std::cout << "\tclBLAS";
std::cout << " \"clBLAS\"";
#endif
#ifdef BENCH_CBLAS
std::cout << "\tBLAS";
std::cout << " \"BLAS\"";
#endif
#ifdef BENCH_CUBLAS
std::cout << "\tcuBLAS";
std::cout << " \"cuBLAS\"";
#endif
std::cout << std::endl;
//
@@ -194,21 +195,23 @@ void bench(sc::numeric_type dtype, std::string operation)
if(operation.substr(0, 4)=="gemv")
{
std::vector<std::tuple<char,int_t, int_t> > MNs;
MNs.push_back(std::make_tuple('N',896,896));
MNs.push_back(std::make_tuple('N',3072,3072));
//AlexNet
MNs.push_back(std::make_tuple('N',1000,256));
MNs.push_back(std::make_tuple('N',4096,256));
//Linear System
MNs.push_back(std::make_tuple('N',153,153));
MNs.push_back(std::make_tuple('N',1024,1024));
MNs.push_back(std::make_tuple('N',2867,2867));
MNs.push_back(std::make_tuple('T',169,256));
MNs.push_back(std::make_tuple('T',169,384));
MNs.push_back(std::make_tuple('T',729,256));
MNs.push_back(std::make_tuple('T',3025,96));
//Normalization
MNs.push_back(std::make_tuple('N', 32, 60000));
MNs.push_back(std::make_tuple('N', 256, 60000));
//Householder
MNs.push_back(std::make_tuple('N', 100, 60000));
MNs.push_back(std::make_tuple('N', 90, 60000));
MNs.push_back(std::make_tuple('N', 50, 60000));
/*---------*/
/*--BLAS2--*/
/*---------*/
//T-layout
for(std::tuple<char, int_t, int_t> MN: MNs)
{
bool AT = std::get<0>(MN) == 'T';
@@ -224,6 +227,7 @@ void bench(sc::numeric_type dtype, std::string operation)
int_t lda = A.ld();
#endif
BENCHMARK_ISAAC(y = sc::control(AT?dot(A.T(),x):dot(A,x), sc::execution_options_type(0, &events)),(M*N + M + N)*dtsize/t);
BENCHMARK_ISAAC(y = sc::control(AT?dot(A.T(),x):dot(A,x), sc::execution_options_type(0, &events), sc::dispatcher_options_type(true)),(M*N + M + N)*dtsize/t);
#ifdef BENCH_CLBLAS
if(y.context().backend()==sc::driver::OPENCL)
BENCHMARK_CLBLAS(clblasSgemv(clblasColumnMajor, AT?clblasTrans:clblasNoTrans, As1, As2, 1, CL_HANDLE(A.data()), 0, lda, CL_HANDLE(x.data()), 0, 1, 0, CL_HANDLE(y.data()), 0, 1, 1, &CL_HANDLE(queue),0, NULL, &event), (M*N + M + N)*dtsize/t)
@@ -274,14 +278,14 @@ void bench(sc::numeric_type dtype, std::string operation)
// MNKs.push_back(std::make_tuple("Convolution Gradient-2 [AlexNet-2]",'N','T',729,1200,128));
// MNKs.push_back(std::make_tuple("Conv. Gradient-2 [LeNet-2]",'N','T',64,500,50));
//Covariance (e.g., ICA, 10minutes/1khz)
MNKs.push_back(std::make_tuple("ICA [32 channels]",'N','T',32,32,600000));
MNKs.push_back(std::make_tuple("ICA [256 channels]",'N','T',256,256,600000));
//Covariance (e.g., ICA, 10minutes/100Hz)
MNKs.push_back(std::make_tuple("ICA [32 channels]",'N','T',32,32,60000));
MNKs.push_back(std::make_tuple("ICA [256 channels]",'N','T',256,256,60000));
// //Bi-diagonalization
MNKs.push_back(std::make_tuple("Bidiagonalization [Iteration 1]",'N','T',4096,4096,32));
MNKs.push_back(std::make_tuple("Bidiagonalization [Iteration 10]",'N','T',3456,3456,32));
MNKs.push_back(std::make_tuple("Bidiagonalization [Iteration 50]",'N','T',896,896,32));
MNKs.push_back(std::make_tuple("Householder [Iteration 1]",'N','T',4096,4096,32));
MNKs.push_back(std::make_tuple("Householder [Iteration 10]",'N','T',3456,3456,32));
MNKs.push_back(std::make_tuple("Householder [Iteration 50]",'N','T',896,896,32));
/*---------*/
/*--BLAS3--*/
@@ -305,6 +309,7 @@ void bench(sc::numeric_type dtype, std::string operation)
#ifdef HAS_A_BLAS
int_t lda = A.ld(), ldb = B.ld(), ldc = C.ld();
#endif
BENCHMARK_ISAAC(C = sc::control(AT?(BT?dot(A.T(),B.T()):dot(A.T(),B)):(BT?dot(A,B.T()):dot(A,B)), sc::execution_options_type(0, &events), sc::dispatcher_options_type(false)), (double)2*M*N*K/t);
BENCHMARK_ISAAC(C = sc::control(AT?(BT?dot(A.T(),B.T()):dot(A.T(),B)):(BT?dot(A,B.T()):dot(A,B)), sc::execution_options_type(0, &events), sc::dispatcher_options_type(true)), (double)2*M*N*K/t);
/* clblas */
#ifdef BENCH_CLBLAS

12
lib/kernels/templates/gemv.cpp
View File

@@ -105,7 +105,7 @@ std::string gemv::generate_impl(std::string const & suffix, expressions_tuple co
for (const auto & e : dots){
std::string data_type = append_width("#scalartype",col_simd_width);
stream << e->process(data_type + " #name_acc = " + neutral_element((e)->root_op(), backend, "#scalartype") + ";") << std::endl;
stream << e->process(data_type + " #name_acc = " + InitPrefix(backend, data_type).get() + "(" + neutral_element((e)->root_op(), backend, "#scalartype") + ");") << std::endl;
}
stream << "if (r < M)" << std::endl;
@@ -122,13 +122,13 @@ std::string gemv::generate_impl(std::string const & suffix, expressions_tuple co
if(dot_type_==REDUCE_COLUMNS)
{
std::string data_type = append_width("#scalartype",row_simd_width);
accessors["array2"] = data_type + " #namereg = " + vload(row_simd_width, "#scalartype", "c*#stride", "#pointer + r*#ld", backend)+";";
accessors["repeat"] = data_type + " #namereg = " + vload(row_simd_width, "#scalartype", "(c%#tuplearg0)*#stride", "#pointer + (r%#tuplearg1)*#stride ", backend)+";";
accessors["array2"] = data_type + " #namereg = " + vload(row_simd_width, "#scalartype", "c*#stride", "#pointer + r*#ld", backend,false)+";";
accessors["repeat"] = data_type + " #namereg = " + vload(row_simd_width, "#scalartype", "(c%#tuplearg0)*#stride", "#pointer + (r%#tuplearg1)*#stride ", backend,false)+";";
}
else
{
std::string data_type = append_width("#scalartype",col_simd_width);
accessors["array2"] = data_type + " #namereg = " + vload(col_simd_width, "#scalartype", "0", "#pointer + r*#stride + c*#ld", backend) + ";";
accessors["array2"] = data_type + " #namereg = " + vload(col_simd_width, "#scalartype", "0", "#pointer + r*#stride + c*#ld", backend,false) + ";";
accessors["repeat"] = "#scalartype #namereg = $VALUE{(r%#tuplearg0)*#stride, (c%#tuplearg1)*#stride};";
}
e->process_recursive(stream, PARENT_NODE_TYPE, accessors);
@@ -206,8 +206,8 @@ std::string gemv::generate_impl(std::string const & suffix, expressions_tuple co
if(col_simd_width > 1)
stream << "if(M - r > " << col_simd_width << "){" << std::endl;
if (e->is_index_dot())
stream << e->process(vstore(col_simd_width,"uint", "#name_buf_value[lidy*" + local_size_0_ld_str + "]", "0", "#name_temp_value + r + M*" + GroupIdx0(backend).get(),backend)) << ";" << std::endl;
stream << e->process(vstore(col_simd_width,"#scalartype", "#name_buf[lidy*" + local_size_0_ld_str + "]", "0", "#name_temp + r + M*" + GroupIdx0(backend).get(),backend)) << ";" << std::endl;
stream << e->process(vstore(col_simd_width,"uint", "#name_buf_value[lidy*" + local_size_0_ld_str + "]", "0", "#name_temp_value + r + M*" + GroupIdx0(backend).get(),backend, false)) << ";" << std::endl;
stream << e->process(vstore(col_simd_width,"#scalartype", "#name_buf[lidy*" + local_size_0_ld_str + "]", "0", "#name_temp + r + M*" + GroupIdx0(backend).get(),backend, false)) << ";" << std::endl;
if(col_simd_width > 1)
{
stream << "}" << std::endl;

12
lib/kernels/templates/tools/vector_types.hpp
View File

@@ -39,7 +39,7 @@ inline std::string append_width(std::string const & str, unsigned int width)
}
inline std::string vstore(unsigned int simd_width, std::string const & dtype, std::string const & value, std::string const & offset, std::string const & ptr, driver::backend_type backend)
inline std::string vstore(unsigned int simd_width, std::string const & dtype, std::string const & value, std::string const & offset, std::string const & ptr, driver::backend_type backend, bool aligned = true)
{
std::string vdtype = append_width(dtype,simd_width);
if (simd_width==1)
@@ -49,7 +49,15 @@ inline std::string vstore(unsigned int simd_width, std::string const & dtype, st
switch(backend)
{
case driver::CUDA:
if(aligned)
return "reinterpret_cast<" + vdtype + "*>(" + ptr + ")[" + offset + "] = " + value;
else
{
std::string res;
for(unsigned int s = 0 ; s < simd_width ; ++s)
res += (s>0?";(":"(") + ptr + ")[" + offset + " + " + tools::to_string(s) + "] = " + access_vector_type(value, s);
return res;
}
case driver::OPENCL:
return append_width("vstore", simd_width) + "(" + value + ", " + offset + ", " + ptr + ")";
default:
@@ -75,7 +83,7 @@ inline std::string vload(unsigned int simd_width, std::string const & dtype, std
{
std::string res = "make_" + vdtype + "(";
for(unsigned int s = 0 ; s < simd_width ; ++s)
res += ((s>0)?",(":"(") + ptr + ")[" + offset + " + " + tools::to_string(s) + "]";
res += ((s>0)?",(":"(") + ptr + ")[" + offset + "*" + tools::to_string(simd_width) + " + " + tools::to_string(s) + "]";
res += ")";
return res;
}

4
tune/android/tune/model.py
View File

@@ -31,12 +31,12 @@ def train(X, Y, profiles):
Y = Y[p,:]
#Train the.profile
cut = int(.9*M)
cut = int(.5*M)
XTr, YTr = X[:cut,:], Y[:cut,:]
XCv, YCv = X[cut:,:], Y[cut:,:]
nrmses = {}
for N in range(1,min(M+1,10)):
for N in range(1,min(M+1,20)):
for depth in range(1,min(M+1,20)):
clf = RandomForestRegressor(N, max_depth=depth).fit(XTr, YTr)
t = np.argmax(clf.predict(XCv), axis = 1)

2
tune/android/tune/tune.py
View File

@@ -93,7 +93,7 @@ class Tuner:
(1200,128,729),
(363,96,3025)]
elif level=='full':
sizes = product(pow2range(5, 12), pow2range(5, 12), pow2range(5, 15))
sizes = product(pow2range(5, 12), pow2range(5, 12), pow2range(5, 17))
#Remove duplicates and or too small/big tuples
sizes = [x for x in sizes if 1e-4 <= tools.memory_footprint(operation, x) <= 2e-1]