File: tileid.cl

package info (click to toggle)
intel-graphics-compiler2 2.28.4-4
  • links: PTS, VCS
  • area: main
  • in suites: forky, sid
  • size: 792,744 kB
  • sloc: cpp: 5,761,745; ansic: 466,928; lisp: 312,143; python: 114,790; asm: 44,736; pascal: 10,930; sh: 8,033; perl: 7,914; ml: 3,625; awk: 3,523; yacc: 2,747; javascript: 2,667; lex: 1,898; f90: 1,028; cs: 573; xml: 474; makefile: 344; objc: 162
file content (34 lines) | stat: -rw-r--r-- 1,286 bytes parent folder | download
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
/*========================== begin_copyright_notice ============================

Copyright (C) 2021 Intel Corporation

SPDX-License-Identifier: MIT

============================= end_copyright_notice ===========================*/

extern int __SubDeviceID;

int __attribute__((overloadable)) __spirv_BuiltInSubDeviceIDINTEL(void) {
    if(BIF_FLAG_CTRL_GET(HasHWTileIDRegister)) {
        return __builtin_IB_hw_tile_id();
    }

    // When __SubDeviceID is declared as an extern int,
    // it is lowered to LLVM-IR like:
    // @__SubDeviceID = external addrspace(1) global i32, align 4
    // This global address is being then patched by the runtime
    // and can be set to null when implicit scaling is disabled.
    // One may wonder why __SubDeviceID is not declared as an extern int*
    // In this case this would end up as a pointer to pointer in LLVM-IR.
    // It would generate two loads and wouldn't be consistent with runtime behavior.
    volatile int* p = &__SubDeviceID;
    if (p == NULL) {
        return 0;
    }
    return *p;
}

int __attribute__((overloadable)) __spirv_GlobalHWThreadIDINTEL(void) {
    int subDeviceId = __spirv_BuiltInSubDeviceIDINTEL();
    return __builtin_IB_hw_thread_id() + subDeviceId * BIF_FLAG_CTRL_GET(MaxHWThreadIDPerSubDevice);
}