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

[dnn/conv] some minor fixes

Browse Source
This commit is contained in:
Philippe Tillet
2019-05-08 10:09:30 -04:00
parent 615569287e
commit 54f888a270
2 changed files with 247 additions and 167 deletions
Download Patch File Download Diff File Expand all files Collapse all files

74
examples/cpp/conv.cpp
View File

@@ -9,22 +9,19 @@
int main() {
// initialize default compute device
auto context = triton::driver::backend::contexts::get_default();
// initialize just-in-time compiler
triton::jit jit(context);
triton::dnn::conv::type ty = triton::dnn::conv::BPROP;
// initialization
int32_t B = 4, NF = 32;
int32_t D = 1, H = 24, W = 240;
int32_t NC = 64, T = 1, R = 3, S = 3;
int32_t NC = 32, T = 1, R = 3, S = 3;
int32_t pad_d = 0, pad_h = 1, pad_w = 1;
int32_t stride_d = 1, stride_h = 1, stride_w = 1;
int32_t upsample_d = 1, upsample_h = 1, upsample_w = 1;
int32_t RD = (D*upsample_d - T + 1 + 2*pad_d + stride_d - 1)/stride_d;
int32_t RH = (H*upsample_h - R + 1 + 2*pad_h + stride_h - 1)/stride_h;
int32_t RW = (W*upsample_w - S + 1 + 2*pad_w + stride_w - 1)/stride_w;
// equivalent matmul dimensions
int32_t M = B*RD*RH*RW;
int32_t N = NF;
int32_t K = NC*T*R*S;
triton::dnn::conv configuration(B, NC, H, W, R, S, NF, 1, 1, pad_h, pad_w, ty);
// convolution configuration
std::vector<float> hc(B*RH*RW*NF);
std::vector<float> rc(B*RH*RW*NF);
@@ -36,7 +33,8 @@ int main() {
for(size_t i = 0; i < hb.size(); i++)
hb[i] = (float)rand()/RAND_MAX;
for(size_t i = 0; i < hc.size(); i++)
hc[i] = 0;
hc[i] = (float)rand()/RAND_MAX;
rc = hc;
triton::driver::buffer* dc = triton::driver::buffer::create(context, hc.size()*4);
triton::driver::buffer* da = triton::driver::buffer::create(context, ha.size()*4);
triton::driver::buffer* db = triton::driver::buffer::create(context, hb.size()*4);
@@ -45,80 +43,38 @@ int main() {
stream->write(db, true, 0, hb);
stream->write(dc, true, 0, hc);
stream->synchronize();
// memory strides for data
int32_t stride_i_w = 1;
int32_t stride_i_h = W*stride_i_w;
int32_t stride_i_d = H*stride_i_h;
int32_t stride_i_c = D*stride_i_d;
int32_t stride_i_n = NC*stride_i_c;
// memory stride for activations
int32_t stride_o_q = 1;
int32_t stride_o_p = RW*stride_o_q;
int32_t stride_o_m = RH*stride_o_p;
int32_t stride_o_k = RD*stride_o_m;
int32_t stride_o_n = NF*stride_o_k;
// look-up table
triton::dnn::conv configuration(B, NC, H, W, R, S, NF, 1, 1, 0, 0);
std::vector<int> h_delta, h_masks;
configuration.build_lut(h_delta, h_masks);
configuration.build_deltas(h_delta);
configuration.build_masks(h_masks);
// benchmark a given convolution kernel
auto benchmark = [&](triton::driver::kernel* kernel,
triton::jit::launch_information info) {
// launch info
unsigned TM = info.global_range_size[0];
unsigned TN = info.global_range_size[1];
// initialize constant memory
unsigned nthreads = info.num_threads;
std::array<size_t, 3> grid = configuration.get_grid(TM, TN);
triton::driver::buffer* delta = jit.get_buffer("delta");
triton::driver::buffer* masks = jit.get_buffer("masks");
stream->write(delta, false, 0, h_delta.size()*4, h_delta.data());
stream->write(masks, false, 0, h_masks.size()*4, h_masks.data());
stream->synchronize();
// launch info
unsigned nthreads = info.num_threads;
std::array<size_t, 3> grid = {(M + TM - 1)/TM, (N + TN - 1)/TN, 1};
// set arguments
kernel->setArg(0, da);
kernel->setArg(1, db);
kernel->setArg(2, dc);
kernel->setArg(3, M);
kernel->setArg(4, N);
kernel->setArg(5, K);
kernel->setArg(6, B);
kernel->setArg(7, H);
kernel->setArg(8, W);
kernel->setArg(9, NF);
kernel->setArg(10, RH);
kernel->setArg(11, RW);
kernel->setArg(12, NC);
kernel->setArg(13, R);
kernel->setArg(14, S);
kernel->setArg(15, stride_i_n);
kernel->setArg(16, stride_i_c);
kernel->setArg(17, stride_i_h);
kernel->setArg(18, stride_i_w);
kernel->setArg(19, stride_o_n);
kernel->setArg(20, stride_o_k);
kernel->setArg(21, stride_o_p);
kernel->setArg(22, stride_o_q);
kernel->setArg(23, pad_h);
kernel->setArg(24, pad_w);
// dry run
configuration.set_arg(kernel, da, db, dc);
stream->enqueue(kernel, grid, {nthreads, 1, 1});
stream->synchronize();
// benchmark
double ts = bench([&](){stream->enqueue(kernel, grid, {nthreads, 1, 1});},
[&](){ stream->synchronize(); }, *context->device());
return 2.*M*N*K / ts * 1e-3;
return configuration.get_nflops() / ts * 1e-3;
};
std::string src = triton::dnn::conv::src();
std::string src = configuration.src();
// jit.autotune("conv", src.c_str(), benchmark);
jit.add_module("conv", src.c_str(), triton::dnn::conv::default_params());
jit.add_module("conv", src.c_str(), configuration.default_params());
triton::driver::kernel* kernel = jit.get_function("conv");
triton::jit::launch_information info = jit.get_launch_info("conv");
std::cout << "Performance: " << benchmark(kernel, info) << " TFLOPS " << std::endl;
stream->read(dc, true, 0, hc);
cpp_conv_nchw(NC, B, NF, D, H, W, T, R, S, pad_d, pad_h, pad_w, stride_d, stride_h, stride_w, RD, RH, RW, rc, ha, hb);
for(size_t i = 0; i < M*N; i++)
configuration.cpu_ref(rc.data(), ha.data(), hb.data());
for(size_t i = 0; i < hc.size(); i++)
if(std::abs(hc[i] - rc[i])/std::max(hc[i], rc[i]) > 1e-4){
std::cout << i << " " << hc[i] << " " << rc[i] << std::endl;
exit(EXIT_FAILURE);