Improve speed of style transfer demo.

Author: Iainmon

bridge/lib/bridge.cpp

@@ -67,6 +67,8 @@ extern "C" void debug_cpu_only_mode(bool_t mode) {
     } else {
         best_device = get_best_device();
     }
+    std::cout << "Debug CPU only mode: " << (debug_cpu_only ? "ON" : "OFF") << std::endl;
+    std::cout.flush();
 }
 
 extern "C" bool_t accelerator_available() {
@@ -131,20 +133,19 @@ extern "C" void free_bridge_tensor(bridge_tensor_t bt) {
 at::Tensor bridge_to_torch(bridge_tensor_t &bt) {
     std::vector<int64_t> sizes_vec(bt.sizes, bt.sizes + bt.dim);
     auto shape = torch::IntArrayRef(sizes_vec);
-    return torch::from_blob(bt.data, shape, torch::kFloat32);
+    return torch::from_blob(bt.data, shape, torch::kFloat);
 }
 
 at::Tensor bridge_to_torch(bridge_tensor_t &bt,torch::Device device, bool copy,torch::ScalarType dtype = torch::kFloat32) {
     std::vector<int64_t> sizes_vec(bt.sizes, bt.sizes + bt.dim);
     auto shape = torch::IntArrayRef(sizes_vec);
-    auto t = torch::from_blob(bt.data, shape, torch::kFloat32);
+    auto t = torch::from_blob(bt.data, shape, torch::kFloat);
     if (device != torch::kCPU)
         copy = true;
     if (copy)
         return t.to(device, dtype, /*non_blocking=*/false, /*copy=*/true);
     else
         return t.to(device, dtype, /*non_blocking=*/false, /*copy=*/false);
-    
 }
 
 extern "C" float32_t* unsafe(const float32_t* arr) {

demos/video/chapel-webcam/main.cpp

@@ -130,10 +130,10 @@ int main(int argc, char* argv[]) {
 
     int code = mirror();
 
-    std::size_t start = cv::getTickCount();
-    std::this_thread::sleep_for(std::chrono::milliseconds(2000));
-    std::size_t end = cv::getTickCount();
-    std::cout << "Total time taken: " << (end - start) / cv::getTickFrequency() << " seconds" << std::endl;
+    // std::size_t start = cv::getTickCount();
+    // std::this_thread::sleep_for(std::chrono::milliseconds(2000));
+    // std::size_t end = cv::getTickCount();
+    // std::cout << "Total time taken: " << (end - start) / cv::getTickFrequency() << " seconds" << std::endl;
 
 
     chpl_library_finalize();

demos/video/chapel-webcam/smol.chpl

@@ -12,25 +12,25 @@ const acceleratorScaleFactor = accelScale;
 
 
 export proc getScaledFrameWidth(width: int): int do
-    if Bridge.acceleratorAvailable() then
+    if Env.acceleratorAvailable() then
         return (width:real * acceleratorScaleFactor):int;
     else
         return (width:real * cpuScaleFactor):int;
 
 export proc getScaledFrameHeight(height: int): int do
-    if Bridge.acceleratorAvailable() then
+    if Env.acceleratorAvailable() then
         return (height:real * acceleratorScaleFactor):int;
     else
         return (height:real * cpuScaleFactor):int;
 
 
 // if debugCPUOnly then
 //     writeln("Debugging CPU only!");
-// Bridge.debugCpuOnlyMode(debugCPUOnly);
+// Env.debugCpuOnlyMode(debugCPUOnly);
 
 writeln("CPU Scale Factor: ", cpuScaleFactor);
 writeln("Accelerator Scale Factor: ", acceleratorScaleFactor);
-writeln("Accelerator Available: ", Bridge.acceleratorAvailable());
+writeln("Accelerator Available: ", Env.acceleratorAvailable());
 
 
 use Time;

lib/Env.chpl

@@ -15,5 +15,9 @@ module Env {
     // Maximum rank of dynamicTensor
     config param maxRank = if developmentAndTesting then minRankNeeded else maxRankNeeded;
 
-    import Bridge.acceleratorAvailable;
+    private import Bridge;
+
+    inline proc acceleratorAvailable(): bool do
+        return Bridge.acceleratorAvailable();
+
 }

lib/Layer.chpl

@@ -4,6 +4,7 @@ module Layer {
     private use Env;
     private import Utilities as util;
     private use OrderedDict;
+    private import Bridge;
 
     class ReLU : Module(?) {
 

lib/Tensor.chpl

@@ -9,5 +9,6 @@ module Tensor {
     public import Utilities as util;
     public import Utilities.Standard;
     public import Testing;
-    public import Bridge;
+    // public import Bridge;
+    public import Env;
 }