void getCanonicalSize(const Halide::Buffer<>& buffer, int* width, int* height,
int* channels, int* batch)
{
CV_Assert(buffer.dimensions() == 4);
*width = buffer.extent(0);
*height = buffer.extent(1);
*channels = buffer.extent(2);
*batch = buffer.extent(3);
}
void print(Halide::Buffer<T> buf) {
for (int j = 0; j < std::min(buf.height(), 10); j++) {
std::stringstream oss;
for (int i = 0; i < std::min(buf.width(), 10); i++) {
oss << " [";
for (int k = 0; k < buf.channels(); k++) {
oss << std::fixed << std::setprecision(1);
if (k > 0) {
oss << std::setw(4);
}
oss << +buf(i, j, k);
}
oss << "]";
}
LOGI("%s", oss.str().c_str());
}
}
void Closure::found_buffer_ref(const string &name, Type type,
bool read, bool written, Halide::Buffer<> image) {
if (!ignore.contains(name)) {
debug(3) << "Adding buffer " << name << " to closure\n";
Buffer &ref = buffers[name];
ref.type = type.element_of(); // TODO: Validate type is the same as existing refs?
ref.read = read;
ref.write = written;
// If reading an image/buffer, compute the size.
if (image.defined()) {
ref.size = image.size_in_bytes();
ref.dimensions = image.dimensions();
}
} else {
debug(3) << "Not adding " << name << " to closure\n";
}
}
HalideBackendWrapper::HalideBackendWrapper(const Ptr<BackendWrapper>& base,
const MatShape& shape)
: BackendWrapper(DNN_BACKEND_HALIDE, base->targetId)
{
managesDevMemory = false;
Halide::Buffer<float> baseBuffer = halideBuffer(base);
buffer = Halide::Buffer<float>((float*)baseBuffer.raw_buffer()->host,
getBufferShape(shape));
if (baseBuffer.has_device_allocation())
{
buffer.raw_buffer()->device = baseBuffer.raw_buffer()->device;
buffer.raw_buffer()->device_interface = baseBuffer.raw_buffer()->device_interface;
buffer.set_device_dirty();
}
else
{
buffer.set_host_dirty(); // Indicate that data is on CPU.
CV_Assert(targetId == DNN_TARGET_CPU);
}
}
HalideBackendWrapper::HalideBackendWrapper(const Ptr<BackendWrapper>& base,
const MatShape& shape)
: BackendWrapper(DNN_BACKEND_HALIDE, base->targetId)
{
int w, h, c, n;
getCanonicalSize(shape, &w, &h, &c, &n);
Halide::Buffer<float> baseBuffer = halideBuffer(base);
buffer = Halide::Buffer<float>((float*)baseBuffer.raw_buffer()->host,
{w, h, c, n});
if (baseBuffer.has_device_allocation())
{
buffer.raw_buffer()->device = baseBuffer.raw_buffer()->device;
buffer.raw_buffer()->device_interface = baseBuffer.raw_buffer()->device_interface;
buffer.set_device_dirty();
}
else
{
buffer.set_host_dirty(); // Indicate that data is on CPU.
CV_Assert(targetId == DNN_TARGET_CPU);
}
}
int main(int, char**)
{
std::vector<std::chrono::duration<double, std::milli>> duration_vector;
Halide::Buffer<uint8_t> input = Halide::Tools::load_image("../rgb.png");
Halide::Buffer<float> kernel(3, 3);
kernel(0,0) = 0; kernel(0,1) = 1.0f/5; kernel(0,2) = 0;
kernel(1,0) = 1.0f/5; kernel(1,1) = 1.0f/5; kernel(1,2) = 1.0f/5;
kernel(2,0) = 0; kernel(2,1) = 1; kernel(2,2) = 0;
// Small size discrepancy with Halide benchmark:
Halide::Buffer<uint8_t> output(input.width(), input.height(), input.channels());
std::cout << "Dimensions : " << std::endl;
std::cout << "input.extent(0): " << input.extent(0) << std::endl; // Rows
std::cout << "input.extent(1): " << input.extent(1) << std::endl; // Cols
std::cout << "input.extent(2): " << input.extent(2) << std::endl; // Colors
#ifdef __PROFILE_CUDA__
cudaProfilerStop();
#endif
// Warm up
pencil_convolution(input.extent(0), input.extent(1), input.extent(1),
(uint8_t *) input.raw_buffer()->host,
(float *) kernel.raw_buffer()->host,
(uint8_t *) output.raw_buffer()->host);
#ifdef __PROFILE_CUDA__
cudaProfilerStart();
#endif
// Tiramisu
for (int i = 0; i < 100; i++) {
auto start = std::chrono::high_resolution_clock::now();
pencil_convolution(input.extent(0), input.extent(1), input.extent(1),
(uint8_t *) input.raw_buffer()->host,
(float *) kernel.raw_buffer()->host,
(uint8_t *) output.raw_buffer()->host);
auto end = std::chrono::high_resolution_clock::now();
std::chrono::duration<double, std::milli> duration = end - start;
duration_vector.push_back(duration);
}
std::cout << "time: " << median(duration_vector) << std::endl;
return 0;
}
std::chrono::duration<double, std::milli> run_halide(Halide::Buffer<uint8_t> &in, Halide::Buffer<uint8_t> &y, Halide::Buffer<uint8_t> &u, Halide::Buffer<uint8_t> &v)
{
in(0, 0, 0) = in(0, 0, 0);
auto start = std::chrono::high_resolution_clock::now();
rgbyuv420gpu_ref(in.raw_buffer(), y.raw_buffer(), u.raw_buffer(), v.raw_buffer());
halide_copy_to_host(nullptr, y.raw_buffer());
halide_copy_to_host(nullptr, u.raw_buffer());
halide_copy_to_host(nullptr, v.raw_buffer());
halide_device_free(nullptr, y.raw_buffer());
halide_device_free(nullptr, u.raw_buffer());
halide_device_free(nullptr, v.raw_buffer());
halide_device_free(nullptr, in.raw_buffer());
auto end = std::chrono::high_resolution_clock::now();
return end - start;
}
int main(int, char**)
{
std::vector<std::chrono::duration<double,std::milli>> duration_vector_1;
std::vector<std::chrono::duration<double,std::milli>> duration_vector_2;
Halide::Buffer<uint8_t> input = Halide::Tools::load_image("./utils/images/rgb.png");
Halide::Buffer<int32_t> size(2);
size(0) = input.extent(0);
size(1) = input.extent(1);
Halide::Buffer<uint8_t> output_ref_y(input.width(), input.height());
Halide::Buffer<uint8_t> output_ref_u(input.width()/2, input.height()/2);
Halide::Buffer<uint8_t> output_ref_v(input.width()/2, input.height()/2);
Halide::Buffer<uint8_t> output_tiramisu_y(input.width(), input.height());
Halide::Buffer<uint8_t> output_tiramisu_u(input.width()/2, input.height()/2);
Halide::Buffer<uint8_t> output_tiramisu_v(input.width()/2, input.height()/2);
std::cout << "STARTING TEST\n";
std::cout << "y size (width, height): " << output_tiramisu_y.width() << ", " << output_tiramisu_y.height() << "\n";
std::cout << "u size (width, height): " << output_tiramisu_u.width() << ", " << output_tiramisu_u.height() << "\n";
std::cout << "v size (width, height): " << output_tiramisu_v.width() << ", " << output_tiramisu_v.height() << "\n";
// Warm up
rgbyuv420gpu_tiramisu(size.raw_buffer(), input.raw_buffer(), output_tiramisu_y.raw_buffer(), output_tiramisu_u.raw_buffer(), output_tiramisu_v.raw_buffer());
run_halide(input, output_ref_y, output_ref_u, output_ref_v);
// Tiramisu
for (int i=0; i<NB_TESTS; i++)
{
auto start1 = std::chrono::high_resolution_clock::now();
rgbyuv420gpu_tiramisu(size.raw_buffer(), input.raw_buffer(), output_tiramisu_y.raw_buffer(), output_tiramisu_u.raw_buffer(), output_tiramisu_v.raw_buffer());
auto end1 = std::chrono::high_resolution_clock::now();
std::chrono::duration<double,std::milli> duration1 = end1 - start1;
duration_vector_1.push_back(duration1);
}
// Reference
for (int i=0; i<NB_TESTS; i++)
{
duration_vector_2.push_back(run_halide(input, output_ref_y, output_ref_u, output_ref_v));
}
print_time("performance_CPU.csv", "rgbyuv420gpu",
{"Tiramisu", "Halide"},
{median(duration_vector_1), median(duration_vector_2)});
Halide::Tools::save_image(output_tiramisu_y, "./build/rgbyuv420gpu_y_tiramisu.png");
Halide::Tools::save_image(output_tiramisu_u, "./build/rgbyuv420gpu_u_tiramisu.png");
Halide::Tools::save_image(output_tiramisu_v, "./build/rgbyuv420gpu_v_tiramisu.png");
Halide::Tools::save_image(output_ref_y, "./build/rgbyuv420gpu_y_ref.png");
Halide::Tools::save_image(output_ref_u, "./build/rgbyuv420gpu_u_ref.png");
Halide::Tools::save_image(output_ref_v, "./build/rgbyuv420gpu_v_ref.png");
if (CHECK_CORRECTNESS)
{
std::cout << "Compare y buffer\n";
compare_buffers("benchmark_rgbyuv420gpu", output_tiramisu_y, output_ref_y);
std::cout << "Compare u buffer\n";
compare_buffers("benchmark_rgbyuv420gpu", output_tiramisu_u, output_ref_u);
std::cout << "Compare v buffer\n";
compare_buffers("benchmark_rgbyuv420gpu", output_tiramisu_y, output_ref_y);
}
return 0;
}
int main(int, char**)
{
std::vector<std::chrono::duration<double,std::milli>> duration_vector_1;
std::vector<std::chrono::duration<double,std::milli>> duration_vector_2;
Halide::Buffer<uint8_t> input = Halide::Tools::load_image("./utils/images/rgb.png");
Halide::Buffer<uint8_t> output_ref_f(input.width(), input.height(), input.channels());
Halide::Buffer<uint8_t> output_ref_g(input.width(), input.height(), input.channels());
Halide::Buffer<uint8_t> output_ref_h(input.width(), input.height(), input.channels());
Halide::Buffer<uint8_t> output_ref_k(input.width(), input.height(), input.channels());
Halide::Buffer<uint8_t> output_tiramisu_f(input.width(), input.height(), input.channels());
Halide::Buffer<uint8_t> output_tiramisu_g(input.width(), input.height(), input.channels());
Halide::Buffer<uint8_t> output_tiramisu_h(input.width(), input.height(), input.channels());
Halide::Buffer<uint8_t> output_tiramisu_k(input.width(), input.height(), input.channels());
// Warm up
fusion_tiramisu(input.raw_buffer(), output_tiramisu_f.raw_buffer(),
output_tiramisu_g.raw_buffer(), output_tiramisu_h.raw_buffer(),
output_tiramisu_k.raw_buffer());
fusion_ref(input.raw_buffer(), output_ref_f.raw_buffer(), output_ref_g.raw_buffer(),
output_ref_h.raw_buffer(), output_ref_k.raw_buffer());
// Tiramisu
for (int i=0; i<NB_TESTS; i++)
{
auto start1 = std::chrono::high_resolution_clock::now();
fusion_tiramisu(input.raw_buffer(), output_tiramisu_f.raw_buffer(),
output_tiramisu_g.raw_buffer(), output_tiramisu_h.raw_buffer(),
output_tiramisu_k.raw_buffer());
auto end1 = std::chrono::high_resolution_clock::now();
std::chrono::duration<double,std::milli> duration1 = end1 - start1;
duration_vector_1.push_back(duration1);
}
// Reference
for (int i=0; i<NB_TESTS; i++)
{
auto start2 = std::chrono::high_resolution_clock::now();
fusion_ref(input.raw_buffer(), output_ref_f.raw_buffer(), output_ref_g.raw_buffer(),
output_ref_h.raw_buffer(), output_ref_k.raw_buffer());
auto end2 = std::chrono::high_resolution_clock::now();
std::chrono::duration<double,std::milli> duration2 = end2 - start2;
duration_vector_2.push_back(duration2);
}
print_time("performance_CPU.csv", "fusion",
{"Tiramisu", "Halide"},
{median(duration_vector_1), median(duration_vector_2)});
Halide::Tools::save_image(output_tiramisu_h, "./build/fusion_h_tiramisu.png");
Halide::Tools::save_image(output_tiramisu_k, "./build/fusion_k_tiramisu.png");
Halide::Tools::save_image(output_ref_h, "./build/fusion_h_ref.png");
Halide::Tools::save_image(output_ref_k, "./build/fusion_k_ref.png");
if (CHECK_CORRECTNESS)
compare_buffers("Fusion", output_ref_k, output_tiramisu_k);
return 0;
}