MESSAGE
| DATE | 2016-10-31 |
| FROM | Christopher League
|
| SUBJECT | Re: [Learn] cuda kernels
|
From learn-bounces-at-nylxs.com Mon Oct 31 17:33:52 2016
Return-Path:
X-Original-To: archive-at-mrbrklyn.com
Delivered-To: archive-at-mrbrklyn.com
Received: from www.mrbrklyn.com (www.mrbrklyn.com [96.57.23.82])
by mrbrklyn.com (Postfix) with ESMTP id 07F7E161312;
Mon, 31 Oct 2016 17:33:52 -0400 (EDT)
X-Original-To: learn-at-nylxs.com
Delivered-To: learn-at-nylxs.com
Received: from liucs.net (contrapunctus.net [174.136.110.10])
by mrbrklyn.com (Postfix) with ESMTP id DEA68160E77
for ; Mon, 31 Oct 2016 17:33:49 -0400 (EDT)
Received: from localhost (unknown [148.4.40.11])
by liucs.net (Postfix) with ESMTPSA id 8089FE08E;
Mon, 31 Oct 2016 17:33:48 -0400 (EDT)
From: Christopher League
To: Ruben Safir , Samir Iabbassen ,
learn-at-nylxs.com
In-Reply-To: <9570d66b-7c02-4e2d-3225-20e008158f0d-at-mrbrklyn.com>
References: <9570d66b-7c02-4e2d-3225-20e008158f0d-at-mrbrklyn.com>
User-Agent: Notmuch/0.21 (http://notmuchmail.org) Emacs/25.1.1
(x86_64-unknown-linux-gnu)
Date: Mon, 31 Oct 2016 17:33:47 -0400
Message-ID: <87lgx48apw.fsf-at-contrapunctus.net>
MIME-Version: 1.0
Subject: Re: [Learn] cuda kernels
X-BeenThere: learn-at-nylxs.com
X-Mailman-Version: 2.1.17
Precedence: list
List-Id:
List-Unsubscribe: ,
List-Archive:
List-Post:
List-Help:
List-Subscribe: ,
Content-Type: multipart/mixed; boundary="===============0787192310=="
Errors-To: learn-bounces-at-nylxs.com
Sender: "Learn"
--===============0787192310==
Content-Type: multipart/alternative; boundary="=-=-="
--=-=-=
Content-Type: text/plain
Ruben Safir writes:
> The video says that the below lines ask for 64 copys of the kernel on 64
> threads? I don't see that
>
> cudaMemcpy(d_in, h_in, ARRAY_BYTES, cudaMemcpyHostToDevice);
> //launch kernel
> square<<<1, ARRAY_SIZE>>>(d_in, d_out);
> cudaMemcpy(h_out, d_out, ARRAY_BYTES, cudaMemcpyDeviceToHost);
I don't really know anything about CUDA yet, but ARRAY_SIZE was 64. And
this syntax: `square<<<1, ARRAY_SIZE>>>` is NOT valid C/C++ syntax. So I
assume that the specialized GPU compiler interprets that in a way that
it distributes the computation of the `square` function across 64 GPU
threads...?
BTW, I'm pretty sure "kernel" here doesn't refer to "the Linux kernel."
CL
--=-=-=
Content-Type: text/html; charset=utf-8
Content-Transfer-Encoding: quoted-printable
1.0, user-scalable=3Dyes">
Ruben Safir ruben-at-mrbrklyn.com=
writes:
The video says that the below lines ask for 64 copys of the kernel on 64=
threads? I don=E2=80=99t see that
cudaMemcpy(d_in, h_in, ARRAY_BYTES, cudaMemcpyHostToDevice);
//launch kernel
square<<<1, ARRAY_SIZE>>>(d_in, d_out);
cudaMemcpy(h_out, d_out, ARRAY_BYTES, cudaMemcpyDeviceToHost);
I don=E2=80=99t really know anything about CUDA yet, but ARRAY_SIZE was =
64. And this syntax: square<<<1, ARRAY_SIZE>>>e> is NOT valid C/C++ syntax. So I assume that the specialized GPU compiler=
interprets that in a way that it distributes the computation of the =
square function across 64 GPU threads=E2=80=A6?
BTW, I=E2=80=99m pretty sure =E2=80=9Ckernel=E2=80=9D here doesn=E2=80=
=99t refer to =E2=80=9Cthe Linux kernel.=E2=80=9D
CL
--=-=-=--
--===============0787192310==
Content-Type: text/plain; charset="us-ascii"
MIME-Version: 1.0
Content-Transfer-Encoding: 7bit
Content-Disposition: inline
_______________________________________________
Learn mailing list
Learn-at-nylxs.com
http://lists.mrbrklyn.com/mailman/listinfo/learn
--===============0787192310==--
|
|
