printf

Product Support

Function

Provides the formatted output function for debugging in the CPU domain and NPU domain.

1
2
3

#include "kernel_operator.h"
AscendC::printf("fmt string %d\n", 0x123);
AscendC::PRINTF("fmt string %d\n", 0x123);

Prototype

1
2
3
4

template <class... Args>
__aicore__ inline void printf(__gm__ const char* fmt, Args&&... args)
template <class... Args>
__aicore__ inline void PRINTF(__gm__ const char* fmt, Args&&... args)

Parameters

Returns

None

Restrictions

• This API does not allow the printing of escape characters, except for newline characters.
• If developers need to include the standard library header files stdio.h and cstdio, include them before the kernel_operator.h header file to avoid printf symbol conflicts.
• The total amount of data printed by a single call to this API cannot exceed 1 MB (including a small amount of header and footer information required by the framework, which can usually be ignored). If this limit is exceeded, the data will not be printed. When a custom operator project is used for operator development, the total amount of data dumped by all dump APIs of an operator on each core cannot exceed 1 MB. You need to control the amount of data to be printed. If the limit is exceeded, no content will be printed.
• The way of outputting the printf result varies according to the operator execution mode. In the dynamic graph or single-operator scenario, the content to be output is parsed and printed on the screen. In the static graph scenario, all operators of the graph need to be offloaded to the NPU for execution, and the information about a single operator cannot be printed by directly calling the API. In this case, after the model is executed, the content to be output needs to be flushed to the dump file, which is further parsed into readable content by using a tool.
  • The dump file paths are sorted by priority as follows:
    • If the data dump function is enabled, the dump file is stored in dump_path configured by the developer. Dump function enabling depends on the network run mode. The following uses TensorFlow online inference as an example to describe how to configure parameters such as enable_dump, dump_path, and dump_mode. For details about how to configure them, see "API Reference > TF Adapter API (2.x) > npu.global_options > Configuration Options" in TensorFlow 2.6.5 Model Porting Guide.
    • If the data dump function is disabled but the ASCEND_WORK_PATH environment variable is configured, the dump file is flushed to the printf directory in ASCEND_WORK_PATH.
    • If the data dump function is disabled and the ASCEND_WORK_PATH environment variable is not configured, the dump file is flushed to the printf directory in the current execution directory.
  • The dump file needs to be parsed into readable content by using a tool.
    The show_kernel_debug_data tool can be used to parse a dump binary file into readable content. The command format is as follows.

    show_kernel_debug_data  bin_file  output_dir

• In the scenario of operator integration into a graph, if a dynamic shape model contains a part that can be offloaded, the framework splits the model into dynamic scheduling and offload scheduling (static subgraph). Operators in the static subgraph do not support the printf feature.

Example

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15

#include "kernel_operator.h"

// Integer printing:
AscendC::printf("fmt string %d\n", 0x123);
AscendC::PRINTF("fmt string %d\n", 0x123);

// Floating-point printing:
float a = 3.14;
AscendC::printf("fmt string %f\n", a);
AscendC::PRINTF("fmt string %f\n", a);

// Pointer printing:
int *b;
AscendC::printf("TEST %p\n", b);
AscendC::PRINTF("TEST %p\n", b);

In NPU mode, the runtime print effect of a program appears as follows (CANN Version and TimeStamp are printed only when a custom operator project is used):

1
2
3
4
5
6
7

CANN Version: XXX.XX, TimeStamp: 202408
fmt string 291
fmt string 291
fmt string 3.140000
fmt string 3.140000
TEST 0x12c08001a000
TEST 0x12c08001a000