Improving bridge interface and cleaning demos up.

Author: Iainmon

CMakeLists.txt

@@ -199,10 +199,21 @@ set(LIBTORCH_LIBS_LINKER_ARGS
 )
 cmake_print_variables(LIBTORCH_LIBS_LINKER_ARGS)
 
+set(CHAI_LINKER_ARGS
+    -M ${PROJECT_ROOT_DIR}/lib
+    ${BRIDGE_DIR}/include/bridge.h
+    ${BRIDGE_OBJECT_FILES}
+    -L ${LIBTORCH_DIR}/lib
+    ${LIBTORCH_LIBS_LINKER_ARGS}
+    --ccflags "-I${BRIDGE_DIR}/include -L${PROJECT_ROOT_DIR}/build"
+    --ldflags "-L${PROJECT_ROOT_DIR}/build -Wl,-rpath,${LIBTORCH_DIR}/lib"
+)
+
 
 add_executable(TorchBridge ${BRIDGE_DIR}/lib/Bridge.chpl)
 add_dependencies(TorchBridge bridge)
 add_dependencies(TorchBridge ChAI)
+add_dependencies(TorchBridge bridge_objs)
 target_link_options(TorchBridge
     PRIVATE
         ${BRIDGE_DIR}/include/bridge.h
@@ -228,9 +239,27 @@ add_dependencies(TinyLayerTest ChAI)
 target_link_options(TinyLayerTest
     PRIVATE
         --main-module layer_test.chpl
-        # -M ${PROJECT_ROOT_DIR}/lib
+        -M ${PROJECT_ROOT_DIR}/lib
         ${CHAI_LINKER_ARGS}
 )
+
+
+add_executable(TinyBridgeSystemTest 
+    ${PROJECT_ROOT_DIR}/test/tiny/bridge_system_test.chpl
+    ${CHAI_LIB_FILES}
+    )
+add_dependencies(TinyBridgeSystemTest bridge)
+add_dependencies(TinyBridgeSystemTest ChAI)
+add_dependencies(TinyBridgeSystemTest bridge_objs)
+add_dependencies(TinyBridgeSystemTest TorchBridge)
+target_link_options(TinyBridgeSystemTest
+    PRIVATE
+        --main-module bridge_system_test.chpl
+        -M ${PROJECT_ROOT_DIR}/lib
+        ${CHAI_LINKER_ARGS}
+)
+
+
 # chpl test/tiny/layer_test.chpl -M lib bridge/include/bridge.h build/CMakeFiles/bridge.dir/bridge/lib/bridge.cpp.o -L libtorch/lib -ltorch -ltorch_cpu -lc10 -ltorch_global_deps --ldflags "-Wl,-rpath,libtorch/lib"
 
 # chpl --fast -o vgg test.chpl -M ../../lib /Users/iainmoncrief/Documents/Github/ChAI/bridge/include/bridge.h /Users/iainmoncrief/Documents/Github/ChAI/build/CMakeFiles/bridge.dir/bridge/lib/bridge.cpp.o -L /Users/iainmoncrief/Documents/Github/ChAI/libtorch/lib -ltorch -ltorch_cpu -lc10 -ltorch_global_deps --ldflags "-Wl,-rpath,/Users/iainmoncrief/Documents/Github/ChAI/libtorch/lib"
@@ -241,15 +270,7 @@ target_link_options(TinyLayerTest
 
 
 
-set(CHAI_LINKER_ARGS
-    -M ${PROJECT_ROOT_DIR}/lib
-    ${BRIDGE_DIR}/include/bridge.h
-    ${BRIDGE_OBJECT_FILES}
-    -L ${LIBTORCH_DIR}/lib
-    ${LIBTORCH_LIBS_LINKER_ARGS}
-    --ccflags "-I${BRIDGE_DIR}/include -L${PROJECT_ROOT_DIR}/build"
-    --ldflags "-L${PROJECT_ROOT_DIR}/build -Wl,-rpath,${LIBTORCH_DIR}/lib"
-)
+
 
 
 

bridge/include/bridge.h

@@ -20,12 +20,14 @@ typedef unsigned long long uint64_t;
 void debug_cpu_only_mode(bool_t mode);
 
 typedef struct bridge_tensor_t {
-    float* data;
-    int* sizes;
-    int dim;
+    float32_t* data;
+    uint32_t* sizes;
+    uint32_t dim;
     bool_t created_by_c;
+    bool_t was_freed;
 } bridge_tensor_t;
 
+void free_bridge_tensor(bridge_tensor_t bt);
 
 typedef struct bridge_pt_model_t {
     void* pt_module;

bridge/lib/bridge.cpp

@@ -14,9 +14,11 @@
 #include <sstream>
 #include <cstdlib>
 #include <vector>
-#include <cstdint>
 #include <chrono>
 #include <thread>
+#include <cstdio>
+#include <cstdint>
+#include <cstdlib>
 
 
 
@@ -81,9 +83,9 @@ torch::ScalarType get_best_dtype() {
     }
 }
 
-int bridge_tensor_elements(bridge_tensor_t &bt) {
-    int size = 1;
-    for (int i = 0; i < bt.dim; ++i) {
+std::size_t bridge_tensor_elements(bridge_tensor_t &bt) {
+    std::size_t size = 1;
+    for (std::size_t i = 0; i < bt.dim; ++i) {
         size *= bt.sizes[i];
     }
     return size;
@@ -95,24 +97,39 @@ size_t bridge_tensor_size(bridge_tensor_t &bt) {
 
 void store_tensor(at::Tensor &input, float32_t* dest) {
     float32_t * data = input.data_ptr<float32_t>();
-    size_t bytes_size = sizeof(float32_t) * input.numel();
+    std::size_t bytes_size = sizeof(float32_t) * input.numel();
     // std::memmove(dest,data,bytes_size);
     std::memcpy(dest,data,bytes_size);
 }
 
 bridge_tensor_t torch_to_bridge(at::Tensor &tensor) {
     bridge_tensor_t result;
     result.created_by_c = true;
+    result.was_freed    = false;
+
     result.dim = tensor.dim();
-    result.sizes = new int32_t[result.dim];
-    for (int i = 0; i < result.dim; ++i) {
-        result.sizes[i] = tensor.size(i);
+
+    std::size_t sizes_bytes = sizeof(uint32_t) * result.dim;
+    result.sizes = static_cast<uint32_t*>(malloc(sizes_bytes));
+    for (uint32_t i = 0; i < result.dim; ++i) {
+        result.sizes[i] = static_cast<uint32_t>(tensor.size(i));
     }
-    result.data = new float32_t[bridge_tensor_elements(result)];
+
+    std::size_t data_bytes = sizeof(float32_t) * tensor.numel();
+    result.data = static_cast<float32_t*>(malloc(data_bytes));
     store_tensor(tensor, result.data);
     return result;
 }
 
+extern "C" void free_bridge_tensor(bridge_tensor_t bt) {
+    if (bt.created_by_c && !bt.was_freed) {
+        free(bt.sizes);
+        free(bt.data);
+        bt.was_freed = true;
+    }
+}
+
+
 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);
@@ -229,7 +246,7 @@ extern "C" bridge_pt_model_t load_model(const uint8_t* model_path) {
 
 
 
-bridge_tensor_t model_forward(bridge_pt_model_t model, bridge_tensor_t input, bool is_vgg_based_model) {
+extern "C" bridge_tensor_t model_forward(bridge_pt_model_t model, bridge_tensor_t input) {
     auto tn_mps = bridge_to_torch(input,best_device,true,best_dtype);
     // tn_mps = tn_mps.permute({2, 0, 1}).contiguous();
     // tn_mps.unsqueeze_(0);//.contiguous();
@@ -243,24 +260,12 @@ bridge_tensor_t model_forward(bridge_pt_model_t model, bridge_tensor_t input, bo
     // auto tn_out = o.squeeze(0).permute({1, 2, 0}).contiguous();
     auto tn_out = o.squeeze(0).contiguous().permute({1, 2, 0}).contiguous();
 
-    if (is_vgg_based_model) {
-        tn_out.div_(255.0);
-    }
-
     auto tn_out_cpu = tn_out.to(torch::kCPU,torch::kFloat32,false,true);
 
     return torch_to_bridge(tn_out_cpu);
 
 }
 
-extern "C" bridge_tensor_t model_forward(bridge_pt_model_t model, bridge_tensor_t input) {
-    return model_forward(model, input, false);
-}
-
-extern "C" bridge_tensor_t model_forward_style_transfer(bridge_pt_model_t model, bridge_tensor_t input) {
-    return model_forward(model, input, true);
-}
-
 // std::tuple<uint64_t, uint64_t> get_cpu_frame_size(uint64_t width, uint64_t height, float32_t scale_factor) {
 //     // if (best_device == torch::kMPS || best_device == torch::kCUDA)
 //     if (accelerator_available())

demos/video/chapel-webcam/lib/smol.h

@@ -31,11 +31,8 @@ void chpl__init_ndarrayRandom(int64_t _ln,
                               int32_t _fn);
 void chpl__init_smol(int64_t _ln,
                      int32_t _fn);
-chpl_bool acceleratorAvailable(void);
 int64_t getScaledFrameWidth(int64_t width);
 int64_t getScaledFrameHeight(int64_t height);
-int64_t square(int64_t x);
-void printArray(chpl_external_array * a);
 void globalLoadModel(void);
 chpl_external_array getNewFrame(chpl_external_array * frame,
                                 int64_t height,

demos/video/chapel-webcam/main.cpp

@@ -167,20 +167,6 @@ int main(int argc, char* argv[]) {
     chpl__init_Bridge(0, 0);
     chpl__init_smol(0, 0);
 
-    square(3);
-
-    // int64_t array[4] = {1,2,3,4};
-    // chpl_external_array array_ptr = chpl_make_external_array_ptr(&array,4);
-    // int64_t sum = sumArray(&array_ptr);
-    // chpl_free_external_array(array_ptr);
-    // printf("sum: %d\n", sum);
-
-
-    int64_t matrix[2][3] = { {1, 4, 2}, {3, 6, 8} };
-    chpl_external_array matrix_ptr = chpl_make_external_array_ptr(matrix, 3 * 2);
-    printArray(&matrix_ptr);
-    chpl_free_external_array(matrix_ptr);
-
     globalLoadModel();
 
     int code = mirror();