0
0
Fork 0
mirror of https://github.com/matrix-construct/construct synced 2024-11-05 13:28:54 +01:00
construct/include/ircd/simt/reduce_add.h

91 lines
2.4 KiB
C
Raw Normal View History

2021-04-02 21:51:59 +02:00
// Matrix Construct
//
// Copyright (C) Matrix Construct Developers, Authors & Contributors
// Copyright (C) 2016-2021 Jason Volk <jason@zemos.net>
//
// Permission to use, copy, modify, and/or distribute this software for any
// purpose with or without fee is hereby granted, provided that the above
// copyright notice and this permission notice is present in all copies. The
// full license for this software is available in the LICENSE file.
#pragma once
#define HAVE_IRCD_SIMT_REDUCE_ADD_H
#ifdef __OPENCL_VERSION__
2021-04-02 21:51:59 +02:00
/// Sum all elements in the buffer. All threads in the group participate;
/// result is placed in index [0], the rest of the buffer is trashed.
inline void
ircd_simt_reduce_add_f4lldr(__local float4 *const buf,
const uint ln,
const uint li)
2021-04-02 21:51:59 +02:00
{
for(uint stride = ln >> 1; stride > 0; stride >>= 1)
{
barrier(CLK_LOCAL_MEM_FENCE);
if(li < stride)
buf[li] += buf[li + stride];
}
if(!ircd_math_is_pow2(ln) && li == 0)
buf[li] += buf[li + 2];
2021-04-02 21:51:59 +02:00
}
#endif
2021-04-02 21:51:59 +02:00
#ifdef __OPENCL_VERSION__
/// Sum all elements in the buffer. All threads in the group participate;
/// result is placed in index [0], the rest of the buffer is trashed.
inline void
ircd_simt_reduce_add_flldr(__local float *const buf,
const uint ln,
const uint li)
#if defined(cl_khr_subgroups)
{
const float
ret = work_group_reduce_add(buf[li]);
if(li == 0)
buf[li] = ret;
}
#else
{
for(uint stride = ln >> 1; stride > 0; stride >>= 1)
{
barrier(CLK_LOCAL_MEM_FENCE);
if(li < stride)
buf[li] += buf[li + stride];
}
if(!ircd_math_is_pow2(ln) && li == 0)
buf[li] += buf[li + 2];
}
#endif
#endif
#ifdef __OPENCL_VERSION__
/// Sum all elements in the buffer. All threads in the group participate;
/// result is placed in index [0], the rest of the buffer is trashed.
inline void
ircd_simt_reduce_add_illdr(__local int *const buf,
const uint ln,
const uint li)
{
if(li > 0)
atomic_add(buf + 0, buf[li]);
}
#endif
#ifdef __OPENCL_VERSION__
/// Sum all elements in the buffer. All threads in the group participate;
/// result is placed in index [0], the rest of the buffer is trashed.
inline void
ircd_simt_reduce_add_ulldr(__local uint *const buf,
const uint ln,
const uint li)
{
if(li > 0)
atomic_add(buf + 0, buf[li]);
}
#endif