xref: /dpdk/doc/guides/prog_guide/gpudev.rst (revision 72c00ae9dba7fc3f8c892b0354b956e885a25770)
1.. SPDX-License-Identifier: BSD-3-Clause
2   Copyright (c) 2021 NVIDIA Corporation & Affiliates
3
4General-Purpose Graphics Processing Unit Library
5================================================
6
7When mixing networking activity with task processing on a GPU device,
8there may be the need to put in communication the CPU with the device
9in order to manage the memory, synchronize operations, exchange info, etc..
10
11By means of the generic GPU interface provided by this library,
12it is possible to allocate a chunk of GPU memory and use it
13to create a DPDK mempool with external mbufs having the payload
14on the GPU memory, enabling any network interface card
15(which support this feature like Mellanox NIC)
16to directly transmit and receive packets using GPU memory.
17
18Additionally, this library provides a number of functions
19to enhance the dialog between CPU and GPU.
20
21Out of scope of this library is to provide a wrapper for GPU specific libraries
22(e.g. CUDA Toolkit or OpenCL), thus it is not possible to launch workload
23on the device or create GPU specific objects
24(e.g. CUDA Driver context or CUDA Streams in case of NVIDIA GPUs).
25
26This library is optional in DPDK
27and can be disabled with ``-Ddisable_libs=gpudev``.
28
29
30Features
31--------
32
33This library provides a number of features:
34
35- Interoperability with device-specific library through generic handlers.
36- Allocate and free memory on the device.
37- Register CPU memory to make it visible from the device.
38- Communication between the CPU and the device.
39
40The whole CPU - GPU communication is implemented
41using CPU memory visible from the GPU.
42
43
44API Overview
45------------
46
47Child Device
48~~~~~~~~~~~~
49
50By default, DPDK PCIe module detects and registers physical GPU devices
51in the system.
52With the gpudev library is also possible to add additional non-physical devices
53through an ``uint64_t`` generic handler (e.g. CUDA Driver context)
54that will be registered internally by the driver as an additional device (child)
55connected to a physical device (parent).
56Each device (parent or child) is represented through a ID
57required to indicate which device a given operation should be executed on.
58
59Memory Allocation
60~~~~~~~~~~~~~~~~~
61
62gpudev can allocate on an input given GPU device a memory area
63returning the pointer to that memory.
64Later, it's also possible to free that memory with gpudev.
65GPU memory allocated outside of the gpudev library
66(e.g. with GPU-specific library) cannot be freed by the gpudev library.
67
68Memory Registration
69~~~~~~~~~~~~~~~~~~~
70
71gpudev can register a CPU memory area to make it visible from a GPU device.
72Later, it's also possible to unregister that memory with gpudev.
73CPU memory registered outside of the gpudev library
74(e.g. with GPU specific library) cannot be unregistered by the gpudev library.
75
76Memory Barrier
77~~~~~~~~~~~~~~
78
79Some GPU drivers may need, under certain conditions,
80to enforce the coherency of external devices writes (e.g. NIC receiving packets)
81into the GPU memory.
82gpudev abstracts and exposes this capability.
83
84Communication Flag
85~~~~~~~~~~~~~~~~~~
86
87Considering an application with some GPU task
88that's waiting to receive a signal from the CPU
89to move forward with the execution.
90The communication flag allocates a CPU memory GPU-visible ``uint32_t`` flag
91that can be used by the CPU to communicate with a GPU task.
92
93Communication list
94~~~~~~~~~~~~~~~~~~
95
96By default, DPDK pulls free mbufs from a mempool to receive packets.
97Best practice, especially in a multithreaded application,
98is to no make any assumption on which mbufs will be used
99to receive the next bursts of packets.
100Considering an application with a GPU memory mempool
101attached to a receive queue having some task waiting on the GPU
102to receive a new burst of packets to be processed,
103there is the need to communicate from the CPU
104the list of mbuf payload addresses where received packet have been stored.
105The ``rte_gpu_comm_*()`` functions are responsible to create a list of packets
106that can be populated with receive mbuf payload addresses
107and communicated to the task running on the GPU.
108
109
110CUDA Example
111------------
112
113In the example below, there is a pseudo-code to give an example
114about how to use functions in this library in case of a CUDA application.
115
116.. code-block:: c
117
118   //////////////////////////////////////////////////////////////////////////
119   ///// gpudev library + CUDA functions
120   //////////////////////////////////////////////////////////////////////////
121   #define GPU_PAGE_SHIFT 16
122   #define GPU_PAGE_SIZE (1UL << GPU_PAGE_SHIFT)
123
124   int main()
125   {
126       struct rte_gpu_flag quit_flag;
127       struct rte_gpu_comm_list *comm_list;
128       int nb_rx = 0;
129       int comm_list_entry = 0;
130       struct rte_mbuf *rx_mbufs[max_rx_mbufs];
131       cudaStream_t cstream;
132       struct rte_mempool *mpool_payload, *mpool_header;
133       struct rte_pktmbuf_extmem ext_mem;
134       int16_t dev_id;
135       int16_t port_id = 0;
136
137       /* Initialize CUDA objects (cstream, context, etc..). */
138       /* Use gpudev library to register a new CUDA context if any. */
139
140       /* Let's assume the application wants to use the default context of the GPU device 0. */
141       dev_id = 0;
142
143       /* Create an external memory mempool using memory allocated on the GPU. */
144       ext_mem.elt_size = mbufs_headroom_size;
145       ext_mem.buf_len = RTE_ALIGN_CEIL(mbufs_num * ext_mem.elt_size, GPU_PAGE_SIZE);
146       ext_mem.buf_iova = RTE_BAD_IOVA;
147       ext_mem.buf_ptr = rte_gpu_mem_alloc(dev_id, ext_mem.buf_len, 0);
148       rte_extmem_register(ext_mem.buf_ptr, ext_mem.buf_len, NULL, ext_mem.buf_iova, GPU_PAGE_SIZE);
149       rte_dev_dma_map(rte_eth_devices[port_id].device,
150               ext_mem.buf_ptr, ext_mem.buf_iova, ext_mem.buf_len);
151       mpool_payload = rte_pktmbuf_pool_create_extbuf("gpu_mempool", mbufs_num,
152                                                      0, 0, ext_mem.elt_size,
153                                                      rte_socket_id(), &ext_mem, 1);
154
155       /*
156        * Create CPU - device communication flag.
157        * With this flag, the CPU can tell to the CUDA kernel to exit from the main loop.
158        */
159       rte_gpu_comm_create_flag(dev_id, &quit_flag, RTE_GPU_COMM_FLAG_CPU);
160       rte_gpu_comm_set_flag(&quit_flag , 0);
161
162       /*
163        * Create CPU - device communication list.
164        * Each entry of this list will be populated by the CPU
165        * with a new set of received mbufs that the CUDA kernel has to process.
166        */
167       comm_list = rte_gpu_comm_create_list(dev_id, num_entries);
168
169       /* A very simple CUDA kernel with just 1 CUDA block and RTE_GPU_COMM_LIST_PKTS_MAX CUDA threads. */
170       cuda_kernel_packet_processing<<<1, RTE_GPU_COMM_LIST_PKTS_MAX, 0, cstream>>>(quit_flag->ptr, comm_list, num_entries, ...);
171
172       /*
173        * For simplicity, the CPU here receives only 2 bursts of mbufs.
174        * In a real application, network activity and device processing should overlap.
175        */
176       nb_rx = rte_eth_rx_burst(port_id, queue_id, &(rx_mbufs[0]), max_rx_mbufs);
177       rte_gpu_comm_populate_list_pkts(comm_list[0], rx_mbufs, nb_rx);
178       nb_rx = rte_eth_rx_burst(port_id, queue_id, &(rx_mbufs[0]), max_rx_mbufs);
179       rte_gpu_comm_populate_list_pkts(comm_list[1], rx_mbufs, nb_rx);
180
181       /*
182        * CPU waits for the completion of the packets' processing on the CUDA kernel
183        * and then it does a cleanup of the received mbufs.
184        */
185       while (rte_gpu_comm_cleanup_list(comm_list[0]));
186       while (rte_gpu_comm_cleanup_list(comm_list[1]));
187
188       /* CPU notifies the CUDA kernel that it has to terminate. */
189       rte_gpu_comm_set_flag(&quit_flag, 1);
190
191       /* gpudev objects cleanup/destruction */
192       rte_gpu_mem_free(dev_id, ext_mem.buf_len);
193
194       return 0;
195   }
196
197   //////////////////////////////////////////////////////////////////////////
198   ///// CUDA kernel
199   //////////////////////////////////////////////////////////////////////////
200
201   void cuda_kernel(uint32_t * quit_flag_ptr, struct rte_gpu_comm_list *comm_list, int comm_list_entries)
202   {
203       int comm_list_index = 0;
204       struct rte_gpu_comm_pkt *pkt_list = NULL;
205
206       /* Do some pre-processing operations. */
207
208       /* GPU kernel keeps checking this flag to know if it has to quit or wait for more packets. */
209       while (*quit_flag_ptr == 0) {
210           if (comm_list[comm_list_index]->status != RTE_GPU_COMM_LIST_READY)
211               continue;
212
213           if (threadIdx.x < comm_list[comm_list_index]->num_pkts)
214           {
215               /* Each CUDA thread processes a different packet. */
216               packet_processing(comm_list[comm_list_index]->addr, comm_list[comm_list_index]->size, ..);
217           }
218           __threadfence();
219           __syncthreads();
220
221           /* Wait for new packets on the next communication list entry. */
222           comm_list_index = (comm_list_index+1) % comm_list_entries;
223       }
224
225       /* Do some post-processing operations. */
226   }
227