issue/811 support cuda graph capture

Author: PanZezhong1725

include/infinicore/graph/graph.hpp

@@ -31,17 +31,21 @@ class GraphOperator {
 
 class Graph {
 public:
-    Graph() = default;
-    ~Graph() = default;
+    Graph();
+    ~Graph();
 
     void run() const;
 
 protected:
     void add_operator(std::shared_ptr<GraphOperator> op);
-
+    void instantiate();
     std::vector<std::shared_ptr<GraphOperator>> op_list_;
 
     friend class GraphManager;
+
+private:
+    struct DeviceGraph;
+    std::unique_ptr<DeviceGraph> device_graph_;
 };
 } // namespace infinicore::graph
 

include/infinirt.h

@@ -6,6 +6,9 @@
 
 typedef void *infinirtStream_t;
 typedef void *infinirtEvent_t;
+typedef void *infinirtGraph_t;
+typedef void *infinirtGraphNode_t;
+typedef void *infinirtGraphExec_t;
 
 __C __export infiniStatus_t infinirtInit();
 
@@ -63,4 +66,24 @@ __C __export infiniStatus_t infinirtMemcpyAsync(void *dst, const void *src, size
 __C __export infiniStatus_t infinirtMallocAsync(void **p_ptr, size_t size, infinirtStream_t stream);
 __C __export infiniStatus_t infinirtFreeAsync(void *ptr, infinirtStream_t stream);
 
+// Graph
+typedef enum {
+    INFINIRT_STREAM_CAPTURE_MODE_GLOBAL = 0,
+    INFINIRT_STREAM_CAPTURE_MODE_THREAD_LOCAL = 1,
+    INFINIRT_STREAM_CAPTURE_MODE_RELAXED = 2,
+
+} infinirtStreamCaptureMode_t;
+
+__C __export infiniStatus_t infinirtStreamBeginCapture(infinirtStream_t stream, infinirtStreamCaptureMode_t mode);
+__C __export infiniStatus_t infinirtStreamEndCapture(infinirtStream_t stream, infinirtGraph_t *graph_ptr);
+__C __export infiniStatus_t infinirtGraphDestroy(infinirtGraph_t graph);
+__C __export infiniStatus_t infinirtGraphInstantiate(
+    infinirtGraphExec_t *graph_exec_ptr,
+    infinirtGraph_t graph,
+    infinirtGraphNode_t *node_ptr,
+    char *log_buffer,
+    size_t buffer_size);
+__C __export infiniStatus_t infinirtGraphExecDestroy(infinirtGraphExec_t graph_exec);
+__C __export infiniStatus_t infinirtGraphLuanch(infinirtGraphExec_t graph_exec, infinirtStream_t stream);
+
 #endif // __INFINIRT_API_H__

src/infinicore/graph/graph.cc

@@ -1,6 +1,8 @@
 #include "graph_manager.hpp"
 
 #include "../utils.hpp"
+#include "infinicore/context/context.hpp"
+#include <infinirt.h>
 
 namespace infinicore::graph {
 
@@ -33,16 +35,85 @@ GraphOperator::~GraphOperator() {
  * Graph
  * ========================= */
 
+struct Graph::DeviceGraph {
+    infinirtGraph_t graph;
+    infinirtGraphExec_t exec;
+    infinirtGraphNode_t node;
+    std::vector<char> log_buffer;
+
+    DeviceGraph() {
+        log_buffer.resize(4 * 1024);
+    }
+
+    ~DeviceGraph() {
+        if (exec) {
+            infinirtGraphExecDestroy(exec);
+        }
+        if (graph) {
+            infinirtGraphDestroy(graph);
+        }
+    }
+
+    void launch() {
+        INFINICORE_CHECK_ERROR(infinirtGraphLuanch(exec, context::getStream()));
+    }
+};
+
+Graph::Graph() {
+}
+
 void Graph::run() const {
-    for (auto &op : op_list_) {
-        op->run();
+    if (device_graph_ != nullptr && device_graph_.get()->exec != nullptr) {
+        device_graph_.get()->launch();
+    } else {
+        for (auto &op : op_list_) {
+            op->run();
+        }
     }
 }
 
 void Graph::add_operator(std::shared_ptr<GraphOperator> op) {
     op_list_.push_back(op);
 }
 
+void Graph::instantiate() {
+    // Reset device graph
+    device_graph_ = std::make_unique<DeviceGraph>();
+
+    if (infinirtStreamBeginCapture(
+            context::getStream(),
+            INFINIRT_STREAM_CAPTURE_MODE_GLOBAL)
+        != INFINI_STATUS_SUCCESS) {
+        return;
+    }
+
+    // Run and record
+    this->run();
+
+    if (infinirtStreamEndCapture(
+            context::getStream(),
+            &device_graph_.get()->graph)
+        != INFINI_STATUS_SUCCESS) {
+        return;
+    }
+
+    if (infinirtGraphInstantiate(
+            &device_graph_.get()->exec,
+            device_graph_.get()->graph,
+            &device_graph_.get()->node,
+            device_graph_.get()->log_buffer.data(),
+            device_graph_.get()->log_buffer.size())
+        != INFINI_STATUS_SUCCESS) {
+        static bool warned_once = false;
+        if (!warned_once) {
+            warned_once = true;
+            spdlog::warn("Fail to instantiate device graph: {}", std::string(device_graph_.get()->log_buffer.data()));
+        }
+    }
+}
+
+Graph::~Graph() = default;
+
 /* =========================
  * GraphManager
  * ========================= */
@@ -52,19 +123,26 @@ bool GraphManager::is_recording() const {
 }
 
 void GraphManager::start_recording() {
+    if (is_recording()) {
+        spdlog::warn("Graph is already recording. Previous recording will be dropped.");
+    }
     recording_ = true;
     graph_ = std::make_shared<Graph>();
 }
 
 void GraphManager::add_operator(std::shared_ptr<GraphOperator> op) {
-    INFINICORE_ASSERT(recording_);
+    INFINICORE_ASSERT(is_recording());
 
     graph_->add_operator(op);
 }
 
 std::shared_ptr<Graph> GraphManager::stop_recording() {
-
+    if (!is_recording()) {
+        spdlog::warn("Graph is not recording. Please start recording first.");
+        return nullptr;
+    }
     recording_ = false;
+    graph_->instantiate();
     return std::exchange(graph_, nullptr);
 }
 

src/infiniop/devices/nvidia/nvidia_common.cu

@@ -23,6 +23,10 @@ Handle::Internal::Internal(int device_id) {
     _grid_size[0] = prop.maxGridSize[0];
     _grid_size[1] = prop.maxGridSize[1];
     _grid_size[2] = prop.maxGridSize[2];
+    this->useCublas(nullptr, [](cublasHandle_t handle) { return INFINI_STATUS_SUCCESS; });
+#ifdef ENABLE_CUDNN_API
+    this->useCudnn(nullptr, [](cudnnHandle_t handle) { return INFINI_STATUS_SUCCESS; });
+#endif
 }
 
 infiniStatus_t Handle::Internal::useCublas(cudaStream_t stream, const Fn<cublasHandle_t> &f) const {

src/infinirt/ascend/infinirt_ascend.cc

@@ -150,5 +150,35 @@ infiniStatus_t mallocAsync(void **p_ptr, size_t size, infinirtStream_t stream) {
 infiniStatus_t freeAsync(void *ptr, infinirtStream_t stream) {
     return freeDevice(ptr);
 }
+
+infiniStatus_t streamBeginCapture(infinirtStream_t stream, infinirtStreamCaptureMode_t mode) {
+    return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
+}
+
+infiniStatus_t streamEndCapture(infinirtStream_t stream, infinirtGraph_t *graph_ptr) {
+    return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
+}
+
+infiniStatus_t graphDestroy(infinirtGraph_t graph) {
+    return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
+}
+
+infiniStatus_t graphInstantiate(
+    infinirtGraphExec_t *graph_exec_ptr,
+    infinirtGraph_t graph,
+    infinirtGraphNode_t *node_ptr,
+    char *log_buffer,
+    size_t buffer_size) {
+    return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
+}
+
+infiniStatus_t graphExecDestroy(infinirtGraphExec_t graph_exec) {
+    return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
+}
+
+infiniStatus_t graphLuanch(infinirtGraphExec_t graph_exec, infinirtStream_t stream) {
+    return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
+}
+
 } // namespace infinirt::ascend
 #undef CHECK_ACLRT

src/infinirt/bang/infinirt_bang.cc

@@ -142,4 +142,34 @@ infiniStatus_t freeAsync(void *ptr, infinirtStream_t stream) {
     CHECK_BANGRT(cnrtFree(ptr));
     return INFINI_STATUS_SUCCESS;
 }
+
+infiniStatus_t streamBeginCapture(infinirtStream_t stream, infinirtStreamCaptureMode_t mode) {
+    return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
+}
+
+infiniStatus_t streamEndCapture(infinirtStream_t stream, infinirtGraph_t *graph_ptr) {
+    return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
+}
+
+infiniStatus_t graphDestroy(infinirtGraph_t graph) {
+    return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
+}
+
+infiniStatus_t graphInstantiate(
+    infinirtGraphExec_t *graph_exec_ptr,
+    infinirtGraph_t graph,
+    infinirtGraphNode_t *node_ptr,
+    char *log_buffer,
+    size_t buffer_size) {
+    return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
+}
+
+infiniStatus_t graphExecDestroy(infinirtGraphExec_t graph_exec) {
+    return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
+}
+
+infiniStatus_t graphLuanch(infinirtGraphExec_t graph_exec, infinirtStream_t stream) {
+    return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
+}
+
 } // namespace infinirt::bang

src/infinirt/cpu/infinirt_cpu.cc

@@ -116,4 +116,33 @@ infiniStatus_t freeAsync(void *ptr, infinirtStream_t stream) {
     return freeDevice(ptr);
 }
 
+infiniStatus_t streamBeginCapture(infinirtStream_t stream, infinirtStreamCaptureMode_t mode) {
+    return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
+}
+
+infiniStatus_t streamEndCapture(infinirtStream_t stream, infinirtGraph_t *graph_ptr) {
+    return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
+}
+
+infiniStatus_t graphDestroy(infinirtGraph_t graph) {
+    return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
+}
+
+infiniStatus_t graphInstantiate(
+    infinirtGraphExec_t *graph_exec_ptr,
+    infinirtGraph_t graph,
+    infinirtGraphNode_t *node_ptr,
+    char *log_buffer,
+    size_t buffer_size) {
+    return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
+}
+
+infiniStatus_t graphExecDestroy(infinirtGraphExec_t graph_exec) {
+    return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
+}
+
+infiniStatus_t graphLuanch(infinirtGraphExec_t graph_exec, infinirtStream_t stream) {
+    return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
+}
+
 } // namespace infinirt::cpu

src/infinirt/cuda/infinirt_cuda.cu

@@ -176,4 +176,53 @@ infiniStatus_t freeAsync(void *ptr, infinirtStream_t stream) {
     RUN_CUDART(cudaFreeAsync(ptr, (cudaStream_t)stream));
     return INFINI_STATUS_SUCCESS;
 }
+
+infiniStatus_t streamBeginCapture(infinirtStream_t stream, infinirtStreamCaptureMode_t mode) {
+    cudaStreamCaptureMode graph_mode;
+    if (mode == INFINIRT_STREAM_CAPTURE_MODE_GLOBAL) {
+        graph_mode = cudaStreamCaptureModeGlobal;
+    } else if (mode == INFINIRT_STREAM_CAPTURE_MODE_THREAD_LOCAL) {
+        graph_mode = cudaStreamCaptureModeThreadLocal;
+    } else if (mode == INFINIRT_STREAM_CAPTURE_MODE_RELAXED) {
+        graph_mode = cudaStreamCaptureModeRelaxed;
+    } else {
+        return INFINI_STATUS_BAD_PARAM;
+    }
+
+    CHECK_CUDART(cudaStreamBeginCapture((cudaStream_t)stream, graph_mode));
+
+    return INFINI_STATUS_SUCCESS;
+}
+
+infiniStatus_t streamEndCapture(infinirtStream_t stream, infinirtGraph_t *graph_ptr) {
+    cudaGraph_t graph;
+    CHECK_CUDART(cudaStreamEndCapture((cudaStream_t)stream, &graph));
+    *graph_ptr = graph;
+    return INFINI_STATUS_SUCCESS;
+}
+
+infiniStatus_t graphDestroy(infinirtGraph_t graph) {
+    RUN_CUDART(cudaGraphDestroy((cudaGraph_t)graph));
+    return INFINI_STATUS_SUCCESS;
+}
+
+infiniStatus_t graphInstantiate(
+    infinirtGraphExec_t *graph_exec_ptr,
+    infinirtGraph_t graph,
+    infinirtGraphNode_t *node_ptr,
+    char *log_buffer,
+    size_t buffer_size) {
+    CHECK_CUDART(cudaGraphInstantiate((cudaGraphExec_t *)graph_exec_ptr, (cudaGraph_t)graph, (cudaGraphNode_t *)node_ptr, log_buffer, buffer_size));
+    return INFINI_STATUS_SUCCESS;
+}
+
+infiniStatus_t graphExecDestroy(infinirtGraphExec_t graph_exec) {
+    RUN_CUDART(cudaGraphExecDestroy((cudaGraphExec_t)graph_exec));
+    return INFINI_STATUS_SUCCESS;
+}
+
+infiniStatus_t graphLuanch(infinirtGraphExec_t graph_exec, infinirtStream_t stream) {
+    CHECK_CUDART(cudaGraphLaunch((cudaGraphExec_t)graph_exec, (cudaStream_t)stream));
+    return INFINI_STATUS_SUCCESS;
+}
 }

src/infinirt/infinirt.cc

@@ -192,3 +192,32 @@ __C infiniStatus_t infinirtMallocAsync(void **p_ptr, size_t size, infinirtStream
 __C infiniStatus_t infinirtFreeAsync(void *ptr, infinirtStream_t stream) {
     INFINIRT_CALL_DEVICE_API(freeAsync, (ptr, stream));
 }
+
+__C infiniStatus_t infinirtStreamBeginCapture(infinirtStream_t stream, infinirtStreamCaptureMode_t mode) {
+    INFINIRT_CALL_DEVICE_API(streamBeginCapture, (stream, mode));
+}
+
+__C infiniStatus_t infinirtStreamEndCapture(infinirtStream_t stream, infinirtGraph_t *graph_ptr) {
+    INFINIRT_CALL_DEVICE_API(streamEndCapture, (stream, graph_ptr));
+}
+
+__C infiniStatus_t infinirtGraphDestroy(infinirtGraph_t graph) {
+    INFINIRT_CALL_DEVICE_API(graphDestroy, (graph));
+}
+
+__C infiniStatus_t infinirtGraphInstantiate(
+    infinirtGraphExec_t *graph_exec_ptr,
+    infinirtGraph_t graph,
+    infinirtGraphNode_t *node_ptr,
+    char *log_buffer,
+    size_t buffer_size) {
+    INFINIRT_CALL_DEVICE_API(graphInstantiate, (graph_exec_ptr, graph, node_ptr, log_buffer, buffer_size));
+}
+
+__C infiniStatus_t infinirtGraphExecDestroy(infinirtGraphExec_t graph_exec) {
+    INFINIRT_CALL_DEVICE_API(graphExecDestroy, (graph_exec));
+}
+
+__C infiniStatus_t infinirtGraphLuanch(infinirtGraphExec_t graph_exec, infinirtStream_t stream) {
+    INFINIRT_CALL_DEVICE_API(graphLuanch, (graph_exec, stream));
+}