OpenCLBuiltIn.md

OpenCL(TM) Built-In Intrinsics

Overview

This chapter describes OpenCL(TM) built-in intrinsics one may use inside the kernel code to retrieve some additional instruction-level information from Intel(R) Processor Graphics execution unit (EU).

Supported Runtimes:

• Intel(R) Graphics Compute Runtime for oneAPI Level Zero and OpenCL(TM) Driver

Supported OS:

• Linux
• Windows

Supported HW:

• Intel(R) Processor Graphics GEN9+

Needed Headers:

• OpenCL(TM) headers

Needed Libraries:

• OpenCL(TM) libraries

How To Use

The following instrinsics are currently available:

/* Returns the value from GPU EU timestamp register */
ulong __attribute__((overloadable)) intel_get_cycle_counter( void );

/* Returns active channel mask for current GPU hardware thread */
uint __attribute__((overloadable)) intel_get_active_channel_mask( void );

/* Returns the value for target control register */
uint __attribute__((overloadable)) intel_get_control_register( uint reg );

/* Returns global GPU hardware thread ID across the device */
uint __attribute__((overloadable)) intel_get_hw_thread_id( void );

/* Returns global GPU slice ID */
uint __attribute__((overloadable)) intel_get_slice_id( void );

/* Returns GPU subslice ID for current slice */
uint __attribute__((overloadable)) intel_get_subslice_id( void );

/* Returns GPU EU ID for current subslice */
uint __attribute__((overloadable)) intel_get_eu_id( void );

/* Returns GPU hardware thread ID for current EU */
uint __attribute__((overloadable)) intel_get_eu_thread_id( void );

/* Pauses current GPU hardware thread for (32 * value) cycles  */
void __attribute__((overloadable)) intel_eu_thread_pause( uint value );

To use them inside an OpenCL(TM) kernel, one need to declare their prototypes first and then use as normal functions:

ulong __attribute__((overloadable)) intel_get_cycle_counter( void );

__kernel kernel(/*...*/) {
  ulong thread_start = intel_get_cycle_counter();
  /* do some computations */
  ulong thread_end = intel_get_cycle_counter();
}

Note, that described instrinsics will work correctly only with Debug or ReleaseInternal builds of Intel(R) Graphics Compute Runtime for oneAPI Level Zero and OpenCL(TM) Driver and Intel(R) Processor Graphics Compiler.

Usage Details

• refer to Intel(R) Processor Graphics Compiler project to get more details on OpenCL(TM) built-in instrinsicts support (this file contains its implementations)
• refer to the paper Introduction to GEN Assembly to learn more on Intel(R) Processor Graphics instruction set
• refer to Compute Architecture Manuals to learn more on Intel(R) Processor Graphics Architecture

Samples

• OpenCL(TM) GEMM with Manual Code Instrumentation