Add gpu flag to initRDMA for GPU P2P RDMA support

Adds optional gpu and gpu_dev_id parameters to initRDMA(); when gpu=true,
the RDMA buffer is allocated from GPU memory instead of hugepages, enabling
direct FPGA-to-GPU data delivery without a CPU-staged copy.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: Maximilian

sw/include/coyote/cThread.hpp

@@ -344,8 +344,11 @@ class cThread {
 	 * @param buffer_size Size of the buffer to be allocated for RDMA operations
 	 * @param port Port number to be used for the out-of-band connection
 	 * @param server_address Optional server address to connect to; if not provided, this cThread acts as the server
+	 * @param gpu If true, allocates GPU memory as the RDMA buffer instead of hugepages (enables GPU P2P RDMA)
+	 * @param gpu_dev_id GPU device ID to use when gpu=true
 	 */
-	void* initRDMA(uint32_t buffer_size, uint16_t port, const char* server_address = nullptr);
+	void* initRDMA(uint32_t buffer_size, uint16_t port, const char* server_address = nullptr,
+	               bool gpu = false, uint32_t gpu_dev_id = 0);
 
 	/**
 	 * @brief Opposite of initRDMA; releases the the out-of-band connection which was used to exchange QP

sw/src/cThread.cpp

@@ -1026,7 +1026,8 @@ void cThread::connSync(bool client) {
     }
 }
 
-void* cThread::initRDMA(uint32_t buffer_size, uint16_t port, const char* server_address) {
+void* cThread::initRDMA(uint32_t buffer_size, uint16_t port, const char* server_address,
+                        bool gpu, uint32_t gpu_dev_id) {
     // Served address provided, so this node is the client
     if (server_address) {
         DBG3("cThread: initRDMA called from client side with server address " << server_address);
@@ -1067,7 +1068,9 @@ void* cThread::initRDMA(uint32_t buffer_size, uint16_t port, const char* server_
         }
 
         // Allocate memory for RDMA operations
-        void *mem = getMem({CoyoteAllocType::HPF, buffer_size, true});
+        void *mem = gpu
+            ? getMem({CoyoteAllocType::GPU, buffer_size, true, gpu_dev_id})
+            : getMem({CoyoteAllocType::HPF, buffer_size, true});
 
         // Send the memory address to the server
         if (write(connfd, &(qpair->local), sizeof(ibvQ)) != sizeof(ibvQ)) {