Brodey/cudnn

codegen/annotations.h

@@ -1,5 +1,6 @@
 #include <nvml.h>
 #include <cuda.h>
+#include <cudnn.h>
 #include <cublas_v2.h>
 #include <cuda_runtime.h>
 
@@ -5545,6 +5546,10 @@ cudaError_t cudaGetExportTable(const void **ppExportTable, const cudaUUID_t *pEx
  * @param symbolPtr SEND_RECV
  */
 cudaError_t cudaGetFuncBySymbol(cudaFunction_t *functionPtr, const void *symbolPtr);
+/**
+ * @param handle RECV_ONLY
+ */
+cudnnStatus_t cudnnCreate(cudnnHandle_t *handle);
 /**
  * @param handle RECV_ONLY
  */
@@ -5577,3 +5582,69 @@ cublasStatus_t cublasSgemm_v2(cublasHandle_t handle,
                               const float *B, int ldb,
                               const float *beta,
                               float *C, int ldc);
+/**
+ * @param handle SEND_ONLY
+ * @param activationDesc SEND_ONLY
+ * @param alpha SEND_ONLY NULLABLE
+ * @param xDesc SEND_ONLY
+ * @param x SEND_ONLY
+ * @param beta SEND_ONLY NULLABLE
+ * @param yDesc SEND_ONLY
+ * @param y SEND_ONLY
+ */
+cudnnStatus_t cudnnActivationForward(
+    cudnnHandle_t                  handle,
+    cudnnActivationDescriptor_t    activationDesc,
+    const void                     *alpha,
+    const cudnnTensorDescriptor_t  *xDesc,
+    const void                     *x,
+    const void                     *beta,
+    const cudnnTensorDescriptor_t  *yDesc,
+    void                           *y);
+
+/**
+ * @param tensorDesc SEND_ONLY
+ * @param format SEND_ONLY
+ * @param dataType SEND_ONLY
+ * @param n SEND_ONLY
+ * @param c SEND_ONLY
+ * @param h SEND_ONLY
+ * @param w SEND_ONLY
+ */
+cudnnStatus_t cudnnSetTensor4dDescriptor(
+    cudnnTensorDescriptor_t tensorDesc,
+    cudnnTensorFormat_t     format,
+    cudnnDataType_t         dataType,
+    int                     n,
+    int                     c,
+    int                     h,
+    int                     w);
+
+/**
+ * @param tensorDesc SEND_RECV
+ */
+cudnnStatus_t cudnnCreateTensorDescriptor(
+    cudnnTensorDescriptor_t *tensorDesc);
+
+/**
+ * @param activationDesc SEND_RECV
+ */
+cudnnStatus_t cudnnCreateActivationDescriptor(
+        cudnnActivationDescriptor_t   *activationDesc);
+
+/**
+ * @param activationDesc SEND_ONLY
+ * @param mode SEND_ONLY
+ * @param reluNanOpt SEND_ONLY
+ * @param coef SEND_ONLY
+ */
+cudnnStatus_t cudnnSetActivationDescriptor(
+    cudnnActivationDescriptor_t         activationDesc,
+    cudnnActivationMode_t               mode,
+    cudnnNanPropagation_t               reluNanOpt,
+    double                              coef);
+
+/**
+ * @param handle SEND_ONLY
+ */
+cudnnStatus_t cudnnDestroy(cudnnHandle_t handle);

codegen/codegen.py

@@ -93,19 +93,22 @@ def client_rpc_write(self, f):
                 server_type=self.ptr.format(),
             )
         )
+
         f.write(
             "        ({param_name} != nullptr && rpc_write(0, {param_name}, sizeof({base_type})) < 0) ||\n".format(
                 param_name=self.parameter.name,
-                base_type=self.ptr.ptr_to.format(),
+                # void is treated differently from non void pointer types
+                base_type=(self.ptr.format() if self.ptr.ptr_to.format() == "const void" else self.ptr.ptr_to.format()),
             )
         )
 
     @property
     def server_declaration(self) -> str:
         c = self.ptr.ptr_to.const
         self.ptr.ptr_to.const = False
+        # void is treated differently from non void pointer types
         s = f"    {self.ptr.format()} {self.parameter.name}_null_check;\n" + \
-            f"    {self.ptr.ptr_to.format()} {self.parameter.name};\n"
+            f"    {self.ptr.format() if self.ptr.ptr_to.format() == "void" else self.ptr.ptr_to.format()} {self.parameter.name};\n"
         self.ptr.ptr_to.const = c
         return s
 
@@ -121,7 +124,8 @@ def server_rpc_read(self, f):
         f.write(
             "        ({param_name}_null_check && rpc_read(conn, &{param_name}, sizeof({base_type})) < 0) ||\n".format(
                 param_name=self.parameter.name,
-                base_type=self.ptr.ptr_to.format(),
+                # void is treated differently from non void pointer types
+                base_type=(self.ptr.format() if self.ptr.ptr_to.format() == "const void" else self.ptr.ptr_to.format()),
             )
         )
 
@@ -379,7 +383,11 @@ def client_rpc_write(self, f):
     def server_declaration(self) -> str:
         if isinstance(self.type_, Pointer) and self.recv:
             return f"    {self.type_.ptr_to.format()} {self.parameter.name};\n"
-        return f"    {self.type_.format()} {self.parameter.name};\n"
+        # ensure we don't have a const struct, otherwise we can't initialise it properly; ex: "const cudnnTensorDescriptor_t xDesc;" is invalid...
+        # but "const cudnnTensorDescriptor_t *xDesc" IS valid. This subtle change carries reprecussions.
+        elif "const " in self.type_.format() and not "void" in self.type_.format() and not "*" in self.type_.format():
+            return f"   {self.type_.format().replace("const", "")} {self.parameter.name};\n"
+        else: return f"    {self.type_.format()} {self.parameter.name};\n"
 
     def server_rpc_read(self, f):
         if not self.send:
@@ -605,6 +613,8 @@ def error_const(return_type: str) -> str:
         return "cudaErrorDevicesUnavailable"
     if return_type == "cublasStatus_t":
         return "CUBLAS_STATUS_NOT_INITIALIZED"
+    if return_type == "cudnnStatus_t":
+        return "CUDNN_STATUS_NOT_INITIALIZED"
     raise NotImplementedError("Unknown return type: %s" % return_type)
 
 
@@ -618,6 +628,8 @@ def main():
     options = ParserOptions(preprocessor=make_gcc_preprocessor(defines=["CUBLASAPI="]))
 
     nvml_ast: ParsedData = parse_file("/usr/include/nvml.h", options=options)
+    cudnn_graph_ast: ParsedData = parse_file("/usr/include/cudnn_graph.h", options=options)
+    cudnn_ops_ast: ParsedData = parse_file("/usr/include/cudnn_ops.h", options=options)
     cuda_ast: ParsedData = parse_file("/usr/include/cuda.h", options=options)
     cublas_ast: ParsedData = parse_file("/usr/include/cublas_api.h", options=options)
     cudart_ast: ParsedData = parse_file(
@@ -632,6 +644,8 @@ def main():
         + cuda_ast.namespace.functions
         + cudart_ast.namespace.functions
         + cublas_ast.namespace.functions
+        + cudnn_graph_ast.namespace.functions
+        + cudnn_ops_ast.namespace.functions
     )
 
     functions_with_annotations: list[tuple[Function, Function, list[Operation]]] = []
@@ -675,6 +689,7 @@ def main():
         f.write(
             "#include <nvml.h>\n"
             "#include <cuda.h>\n"
+            "#include <cudnn.h>\n"
             "#include <cublas_v2.h>\n"
             "#include <cuda_runtime_api.h>\n\n"
             "#include <cstring>\n"
@@ -815,7 +830,9 @@ def main():
     with open("gen_server.cpp", "w") as f:
         f.write(
             "#include <nvml.h>\n"
+            "#include <iostream>\n"
             "#include <cuda.h>\n"
+            "#include <cudnn.h>\n"
             "#include <cublas_v2.h>\n"
             "#include <cuda_runtime_api.h>\n\n"
             "#include <cstring>\n"

codegen/gen_api.h

@@ -888,3 +888,10 @@
 #define RPC_cublasCreate_v2 887
 #define RPC_cublasDestroy_v2 888
 #define RPC_cublasSgemm_v2 889
+#define RPC_cudnnCreate 890
+#define RPC_cudnnDestroy 891
+#define RPC_cudnnCreateTensorDescriptor 892
+#define RPC_cudnnSetTensor4dDescriptor 893
+#define RPC_cudnnCreateActivationDescriptor 894
+#define RPC_cudnnSetActivationDescriptor 895
+#define RPC_cudnnActivationForward 896

codegen/gen_client.cpp

@@ -1,5 +1,6 @@
 #include <nvml.h>
 #include <cuda.h>
+#include <cudnn.h>
 #include <cublas_v2.h>
 #include <cuda_runtime_api.h>
 
@@ -11160,6 +11161,103 @@ cublasStatus_t cublasSgemm_v2(cublasHandle_t handle, cublasOperation_t transa, c
     return return_value;
 }
 
+cudnnStatus_t cudnnCreate(cudnnHandle_t* handle)
+{
+    cudnnStatus_t return_value;
+    if (rpc_start_request(0, RPC_cudnnCreate) < 0 ||
+        rpc_wait_for_response(0) < 0 ||
+        rpc_read(0, handle, sizeof(cudnnHandle_t)) < 0 ||
+        rpc_end_response(0, &return_value) < 0)
+        return CUDNN_STATUS_NOT_INITIALIZED;
+    return return_value;
+}
+
+cudnnStatus_t cudnnDestroy(cudnnHandle_t handle)
+{
+    cudnnStatus_t return_value;
+    if (rpc_start_request(0, RPC_cudnnDestroy) < 0 ||
+        rpc_write(0, &handle, sizeof(cudnnHandle_t)) < 0 ||
+        rpc_wait_for_response(0) < 0 ||
+        rpc_end_response(0, &return_value) < 0)
+        return CUDNN_STATUS_NOT_INITIALIZED;
+    return return_value;
+}
+
+cudnnStatus_t cudnnCreateTensorDescriptor(cudnnTensorDescriptor_t* tensorDesc)
+{
+    cudnnStatus_t return_value;
+    if (rpc_start_request(0, RPC_cudnnCreateTensorDescriptor) < 0 ||
+        rpc_write(0, tensorDesc, sizeof(cudnnTensorDescriptor_t)) < 0 ||
+        rpc_wait_for_response(0) < 0 ||
+        rpc_read(0, tensorDesc, sizeof(cudnnTensorDescriptor_t)) < 0 ||
+        rpc_end_response(0, &return_value) < 0)
+        return CUDNN_STATUS_NOT_INITIALIZED;
+    return return_value;
+}
+
+cudnnStatus_t cudnnSetTensor4dDescriptor(cudnnTensorDescriptor_t tensorDesc, cudnnTensorFormat_t format, cudnnDataType_t dataType, int n, int c, int h, int w)
+{
+    cudnnStatus_t return_value;
+    if (rpc_start_request(0, RPC_cudnnSetTensor4dDescriptor) < 0 ||
+        rpc_write(0, &tensorDesc, sizeof(cudnnTensorDescriptor_t)) < 0 ||
+        rpc_write(0, &format, sizeof(cudnnTensorFormat_t)) < 0 ||
+        rpc_write(0, &dataType, sizeof(cudnnDataType_t)) < 0 ||
+        rpc_write(0, &n, sizeof(int)) < 0 ||
+        rpc_write(0, &c, sizeof(int)) < 0 ||
+        rpc_write(0, &h, sizeof(int)) < 0 ||
+        rpc_write(0, &w, sizeof(int)) < 0 ||
+        rpc_wait_for_response(0) < 0 ||
+        rpc_end_response(0, &return_value) < 0)
+        return CUDNN_STATUS_NOT_INITIALIZED;
+    return return_value;
+}
+
+cudnnStatus_t cudnnCreateActivationDescriptor(cudnnActivationDescriptor_t* activationDesc)
+{
+    cudnnStatus_t return_value;
+    if (rpc_start_request(0, RPC_cudnnCreateActivationDescriptor) < 0 ||
+        rpc_write(0, activationDesc, sizeof(cudnnActivationDescriptor_t)) < 0 ||
+        rpc_wait_for_response(0) < 0 ||
+        rpc_read(0, activationDesc, sizeof(cudnnActivationDescriptor_t)) < 0 ||
+        rpc_end_response(0, &return_value) < 0)
+        return CUDNN_STATUS_NOT_INITIALIZED;
+    return return_value;
+}
+
+cudnnStatus_t cudnnSetActivationDescriptor(cudnnActivationDescriptor_t activationDesc, cudnnActivationMode_t mode, cudnnNanPropagation_t reluNanOpt, double coef)
+{
+    cudnnStatus_t return_value;
+    if (rpc_start_request(0, RPC_cudnnSetActivationDescriptor) < 0 ||
+        rpc_write(0, &activationDesc, sizeof(cudnnActivationDescriptor_t)) < 0 ||
+        rpc_write(0, &mode, sizeof(cudnnActivationMode_t)) < 0 ||
+        rpc_write(0, &reluNanOpt, sizeof(cudnnNanPropagation_t)) < 0 ||
+        rpc_write(0, &coef, sizeof(double)) < 0 ||
+        rpc_wait_for_response(0) < 0 ||
+        rpc_end_response(0, &return_value) < 0)
+        return CUDNN_STATUS_NOT_INITIALIZED;
+    return return_value;
+}
+
+cudnnStatus_t cudnnActivationForward(cudnnHandle_t handle, cudnnActivationDescriptor_t activationDesc, const void* alpha, const cudnnTensorDescriptor_t xDesc, const void* x, const void* beta, const cudnnTensorDescriptor_t yDesc, void* y)
+{
+    cudnnStatus_t return_value;
+    if (rpc_start_request(0, RPC_cudnnActivationForward) < 0 ||
+        rpc_write(0, &handle, sizeof(cudnnHandle_t)) < 0 ||
+        rpc_write(0, &activationDesc, sizeof(cudnnActivationDescriptor_t)) < 0 ||
+        rpc_write(0, &alpha, sizeof(const void*)) < 0 ||
+        (alpha != nullptr && rpc_write(0, alpha, sizeof(const void*)) < 0) ||
+        rpc_write(0, &xDesc, sizeof(const cudnnTensorDescriptor_t)) < 0 ||
+        rpc_write(0, &x, sizeof(const void*)) < 0 ||
+        rpc_write(0, &beta, sizeof(const void*)) < 0 ||
+        (beta != nullptr && rpc_write(0, beta, sizeof(const void*)) < 0) ||
+        rpc_write(0, &yDesc, sizeof(const cudnnTensorDescriptor_t)) < 0 ||
+        rpc_write(0, &y, sizeof(void*)) < 0 ||
+        rpc_wait_for_response(0) < 0 ||
+        rpc_end_response(0, &return_value) < 0)
+        return CUDNN_STATUS_NOT_INITIALIZED;
+    return return_value;
+}
+
 std::unordered_map<std::string, void *> functionMap = {
     {"__cudaRegisterVar", (void *)__cudaRegisterVar},
     {"__cudaRegisterFunction", (void *)__cudaRegisterFunction},
@@ -12018,6 +12116,13 @@ std::unordered_map<std::string, void *> functionMap = {
     {"cublasCreate_v2", (void *)cublasCreate_v2},
     {"cublasDestroy_v2", (void *)cublasDestroy_v2},
     {"cublasSgemm_v2", (void *)cublasSgemm_v2},
+    {"cudnnCreate", (void *)cudnnCreate},
+    {"cudnnDestroy", (void *)cudnnDestroy},
+    {"cudnnCreateTensorDescriptor", (void *)cudnnCreateTensorDescriptor},
+    {"cudnnSetTensor4dDescriptor", (void *)cudnnSetTensor4dDescriptor},
+    {"cudnnCreateActivationDescriptor", (void *)cudnnCreateActivationDescriptor},
+    {"cudnnSetActivationDescriptor", (void *)cudnnSetActivationDescriptor},
+    {"cudnnActivationForward", (void *)cudnnActivationForward},
     {"cuMemcpy_ptds", (void *)cuMemcpy},
     {"cuMemcpyAsync_ptsz", (void *)cuMemcpyAsync},
     {"cuMemcpyPeer_ptds", (void *)cuMemcpyPeer},