From patchwork Fri Jan 20 20:46:50 2023 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 8bit X-Patchwork-Submitter: Thomas Schwinge X-Patchwork-Id: 63489 Return-Path: X-Original-To: patchwork@sourceware.org Delivered-To: patchwork@sourceware.org Received: from server2.sourceware.org (localhost [IPv6:::1]) by sourceware.org (Postfix) with ESMTP id C2657385B524 for ; Fri, 20 Jan 2023 20:47:15 +0000 (GMT) X-Original-To: gcc-patches@gcc.gnu.org Delivered-To: gcc-patches@gcc.gnu.org Received: from esa2.mentor.iphmx.com (esa2.mentor.iphmx.com [68.232.141.98]) by sourceware.org (Postfix) with ESMTPS id 9C8B73858D20 for ; Fri, 20 Jan 2023 20:46:57 +0000 (GMT) DMARC-Filter: OpenDMARC Filter v1.4.2 sourceware.org 9C8B73858D20 Authentication-Results: sourceware.org; dmarc=none (p=none dis=none) header.from=codesourcery.com Authentication-Results: sourceware.org; spf=pass smtp.mailfrom=mentor.com X-IronPort-AV: E=Sophos;i="5.97,233,1669104000"; d="scan'208,223";a="95184745" Received: from orw-gwy-01-in.mentorg.com ([192.94.38.165]) by esa2.mentor.iphmx.com with ESMTP; 20 Jan 2023 12:46:56 -0800 IronPort-SDR: mq2tCVm3Pjfs7n/DrPi7LYUP8AMVggvKvw6DgVpcZJg5h6cOnW/W0yELDNAkDH4hUsjHqbP8o4 RssFBPz3Mim2HicowZ/XPzvclEBTA7nRyzZlkXGvM2XYD2ntbQ7VG7aCykihIAxfg7mwath54o rEd2leVXEBRb3LEMDqsIPtU6TB2QIWJkrkI5y1YeVXJG0C8hqqSN609uv6kVZxsXU77SCHJypl N+lbcAJ0uSRkgnRu3X2L5q1YBh1FbNa/e2S5IW3Lxjuco7r6R8eIG+OYvJioS3U7C3vN5xhwJR /Y8= From: Thomas Schwinge To: CC: Tom de Vries Subject: [og12] nvptx: Support global constructors/destructors via 'collect2' for offloading (was: nvptx: Support global constructors/destructors via 'collect2') In-Reply-To: <87r0wqp7jf.fsf@euler.schwinge.homeip.net> References: <878rjqaku5.fsf@dem-tschwing-1.ger.mentorg.com> <87y1rq7wt4.fsf@dem-tschwing-1.ger.mentorg.com> <87r0wqp7jf.fsf@euler.schwinge.homeip.net> User-Agent: Notmuch/0.29.1+93~g67ed7df (https://notmuchmail.org) Emacs/26.3 (x86_64-pc-linux-gnu) Date: Fri, 20 Jan 2023 21:46:50 +0100 Message-ID: <87sfg50w91.fsf@dem-tschwing-1.ger.mentorg.com> MIME-Version: 1.0 X-Originating-IP: [137.202.0.90] X-ClientProxiedBy: svr-ies-mbx-14.mgc.mentorg.com (139.181.222.14) To svr-ies-mbx-10.mgc.mentorg.com (139.181.222.10) X-Spam-Status: No, score=-11.9 required=5.0 tests=BAYES_00, GIT_PATCH_0, HEADER_FROM_DIFFERENT_DOMAINS, KAM_DMARC_STATUS, RCVD_IN_MSPIKE_H2, SPF_HELO_PASS, SPF_PASS, TXREP autolearn=ham autolearn_force=no version=3.4.6 X-Spam-Checker-Version: SpamAssassin 3.4.6 (2021-04-09) on server2.sourceware.org X-BeenThere: gcc-patches@gcc.gnu.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Gcc-patches mailing list List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: gcc-patches-bounces+patchwork=sourceware.org@gcc.gnu.org Sender: "Gcc-patches" Hi! On 2022-12-23T14:35:16+0100, I wrote: > On 2022-12-02T14:35:35+0100, I wrote: >> On 2022-12-01T22:13:38+0100, I wrote: >>> I'm working on support for global constructors/destructors with >>> GCC/nvptx >> >> See "nvptx: Support global constructors/destructors via 'collect2'" >> [posted before] > > Building on that, attached is now the additional "for offloading" piece: > "nvptx: Support global constructors/destructors via 'collect2' for offloading". > OK to push? For now pushed to devel/omp/gcc-12 branch in commit 689a5340c7e4286b451f1bc600342550c7c94da2 "nvptx: Support global constructors/destructors via 'collect2' for offloading", see attached. > I did manually test this (by putting a few constructors/destructors into > 'libgomp/config/nvptx/oacc-parallel.c', and observing them be executed), > and also in my WIP development tree with standard libgfortran > constructors (with 'LIBGFOR_MINIMAL' disabled). Grüße Thomas ----------------- Siemens Electronic Design Automation GmbH; Anschrift: Arnulfstraße 201, 80634 München; Gesellschaft mit beschränkter Haftung; Geschäftsführer: Thomas Heurung, Frank Thürauf; Sitz der Gesellschaft: München; Registergericht München, HRB 106955 From 689a5340c7e4286b451f1bc600342550c7c94da2 Mon Sep 17 00:00:00 2001 From: Thomas Schwinge Date: Wed, 30 Nov 2022 22:09:35 +0100 Subject: [PATCH] nvptx: Support global constructors/destructors via 'collect2' for offloading This extends "nvptx: Support global constructors/destructors via 'collect2'" for offloading. libgcc/ * config/nvptx/crtstuff.c ["mgomp"] (__do_global_ctors__entry__mgomp) (__do_global_dtors__entry__mgomp): New. [!"mgomp"] (__do_global_ctors__entry, __do_global_dtors__entry): New. libgomp/ * plugin/plugin-nvptx.c (nvptx_do_global_cdtors): New. (nvptx_close_device, GOMP_OFFLOAD_load_image) (GOMP_OFFLOAD_unload_image): Call it. --- libgcc/ChangeLog.omp | 6 ++ libgcc/config/nvptx/crtstuff.c | 64 ++++++++++++++++++- libgomp/ChangeLog.omp | 4 ++ libgomp/plugin/plugin-nvptx.c | 113 ++++++++++++++++++++++++++++++++- 4 files changed, 185 insertions(+), 2 deletions(-) diff --git a/libgcc/ChangeLog.omp b/libgcc/ChangeLog.omp index 68a99cbe427..2e7bf5cc029 100644 --- a/libgcc/ChangeLog.omp +++ b/libgcc/ChangeLog.omp @@ -1,5 +1,11 @@ 2023-01-20 Thomas Schwinge + * config/nvptx/crtstuff.c ["mgomp"] + (__do_global_ctors__entry__mgomp) + (__do_global_dtors__entry__mgomp): New. + [!"mgomp"] (__do_global_ctors__entry, __do_global_dtors__entry): + New. + * config.host : Add 'crtbegin.o', 'crtend.o' to 'extra_parts'. * config/nvptx/crt0.c: Invoke '__do_global_ctors', diff --git a/libgcc/config/nvptx/crtstuff.c b/libgcc/config/nvptx/crtstuff.c index 0823fc49901..8dc80687e0a 100644 --- a/libgcc/config/nvptx/crtstuff.c +++ b/libgcc/config/nvptx/crtstuff.c @@ -29,6 +29,14 @@ files (via 'CRT_BEGIN' and 'CRT_END'): 'crtbegin.o' and 'crtend.o', but we do so anyway, for symmetry with other configurations. */ + +/* See 'crt0.c', 'mgomp.c'. */ +#if defined(__nvptx_softstack__) && defined(__nvptx_unisimt__) +extern void *__nvptx_stacks[32] __attribute__((shared,nocommon)); +extern unsigned __nvptx_uni[32] __attribute__((shared,nocommon)); +#endif + + #ifdef CRT_BEGIN void @@ -37,6 +45,33 @@ __do_global_ctors (void) DO_GLOBAL_CTORS_BODY; } +/* Need '.entry' wrapper for offloading. */ + +# if defined(__nvptx_softstack__) && defined(__nvptx_unisimt__) + +__attribute__((kernel)) void __do_global_ctors__entry__mgomp (void *); + +void +__do_global_ctors__entry__mgomp (void *nvptx_stacks_0) +{ + __nvptx_stacks[0] = nvptx_stacks_0; + __nvptx_uni[0] = 0; + + __do_global_ctors (); +} + +# else + +__attribute__((kernel)) void __do_global_ctors__entry (void); + +void +__do_global_ctors__entry (void) +{ + __do_global_ctors (); +} + +# endif + #elif defined(CRT_END) /* ! CRT_BEGIN */ void @@ -45,7 +80,7 @@ __do_global_dtors (void) /* In this configuration here, there's no way that "this routine is run more than once [...] when exit is called recursively": for nvptx target, the call to '__do_global_dtors' is registered via 'atexit', which doesn't - re-enter a function already run. + re-enter a function already run, and neither does nvptx offload target. Therefore, we do *not* "arrange to remember where in the list we left off processing". */ func_ptr *p; @@ -53,6 +88,33 @@ __do_global_dtors (void) (*p++) (); } +/* Need '.entry' wrapper for offloading. */ + +# if defined(__nvptx_softstack__) && defined(__nvptx_unisimt__) + +__attribute__((kernel)) void __do_global_dtors__entry__mgomp (void *); + +void +__do_global_dtors__entry__mgomp (void *nvptx_stacks_0) +{ + __nvptx_stacks[0] = nvptx_stacks_0; + __nvptx_uni[0] = 0; + + __do_global_dtors (); +} + +# else + +__attribute__((kernel)) void __do_global_dtors__entry (void); + +void +__do_global_dtors__entry (void) +{ + __do_global_dtors (); +} + +# endif + #else /* ! CRT_BEGIN && ! CRT_END */ #error "One of CRT_BEGIN or CRT_END must be defined." #endif diff --git a/libgomp/ChangeLog.omp b/libgomp/ChangeLog.omp index 4447b74a2ab..32aa9705296 100644 --- a/libgomp/ChangeLog.omp +++ b/libgomp/ChangeLog.omp @@ -1,5 +1,9 @@ 2023-01-20 Thomas Schwinge + * plugin/plugin-nvptx.c (nvptx_do_global_cdtors): New. + (nvptx_close_device, GOMP_OFFLOAD_load_image) + (GOMP_OFFLOAD_unload_image): Call it. + * plugin/plugin-nvptx.c (nvptx_exec): Assert what we know about 'blockDimX'. diff --git a/libgomp/plugin/plugin-nvptx.c b/libgomp/plugin/plugin-nvptx.c index b2fabc61cc8..8e7b63bd637 100644 --- a/libgomp/plugin/plugin-nvptx.c +++ b/libgomp/plugin/plugin-nvptx.c @@ -344,6 +344,11 @@ static struct ptx_device **ptx_devices; default is set here. */ static unsigned lowlat_pool_size = 8*1024; +static bool nvptx_do_global_cdtors (CUmodule, struct ptx_device *, + const char *); +static size_t nvptx_stacks_size (); +static void *nvptx_stacks_acquire (struct ptx_device *, size_t, int); + static inline struct nvptx_thread * nvptx_thread (void) { @@ -571,6 +576,17 @@ nvptx_close_device (struct ptx_device *ptx_dev) if (!ptx_dev) return true; + bool ret = true; + + for (struct ptx_image_data *image = ptx_dev->images; + image != NULL; + image = image->next) + { + if (!nvptx_do_global_cdtors (image->module, ptx_dev, + "__do_global_dtors__entry")) + ret = false; + } + for (struct ptx_free_block *b = ptx_dev->free_blocks; b;) { struct ptx_free_block *b_next = b->next; @@ -591,7 +607,8 @@ nvptx_close_device (struct ptx_device *ptx_dev) CUDA_CALL (cuCtxDestroy, ptx_dev->ctx); free (ptx_dev); - return true; + + return ret; } static int @@ -1313,6 +1330,93 @@ nvptx_set_clocktick (CUmodule module, struct ptx_device *dev) GOMP_PLUGIN_fatal ("cuMemcpyHtoD error: %s", cuda_error (r)); } +/* Invoke MODULE's global constructors/destructors. */ + +static bool +nvptx_do_global_cdtors (CUmodule module, struct ptx_device *ptx_dev, + const char *funcname) +{ + bool ret = true; + char *funcname_mgomp = NULL; + CUresult r; + CUfunction funcptr; + r = CUDA_CALL_NOCHECK (cuModuleGetFunction, + &funcptr, module, funcname); + GOMP_PLUGIN_debug (0, "cuModuleGetFunction (%s): %s\n", + funcname, cuda_error (r)); + if (r == CUDA_ERROR_NOT_FOUND) + { + /* Try '[funcname]__mgomp'. */ + + size_t funcname_len = strlen (funcname); + const char *mgomp_suffix = "__mgomp"; + size_t mgomp_suffix_len = strlen (mgomp_suffix); + funcname_mgomp + = GOMP_PLUGIN_malloc (funcname_len + mgomp_suffix_len + 1); + memcpy (funcname_mgomp, funcname, funcname_len); + memcpy (funcname_mgomp + funcname_len, + mgomp_suffix, mgomp_suffix_len + 1); + funcname = funcname_mgomp; + + r = CUDA_CALL_NOCHECK (cuModuleGetFunction, + &funcptr, module, funcname); + GOMP_PLUGIN_debug (0, "cuModuleGetFunction (%s): %s\n", + funcname, cuda_error (r)); + } + if (r == CUDA_ERROR_NOT_FOUND) + ; + else if (r != CUDA_SUCCESS) + { + GOMP_PLUGIN_error ("cuModuleGetFunction (%s) error: %s", + funcname, cuda_error (r)); + ret = false; + } + else + { + /* If necessary, set up soft stack. */ + void *nvptx_stacks_0; + void *kargs[1]; + if (funcname_mgomp) + { + size_t stack_size = nvptx_stacks_size (); + pthread_mutex_lock (&ptx_dev->omp_stacks.lock); + nvptx_stacks_0 = nvptx_stacks_acquire (ptx_dev, stack_size, 1); + nvptx_stacks_0 += stack_size; + kargs[0] = &nvptx_stacks_0; + } + r = CUDA_CALL_NOCHECK (cuLaunchKernel, + funcptr, + 1, 1, 1, 1, 1, 1, + /* sharedMemBytes */ 0, + /* hStream */ NULL, + /* kernelParams */ funcname_mgomp ? kargs : NULL, + /* extra */ NULL); + if (r != CUDA_SUCCESS) + { + GOMP_PLUGIN_error ("cuLaunchKernel (%s) error: %s", + funcname, cuda_error (r)); + ret = false; + } + + r = CUDA_CALL_NOCHECK (cuStreamSynchronize, + NULL); + if (r != CUDA_SUCCESS) + { + GOMP_PLUGIN_error ("cuStreamSynchronize (%s) error: %s", + funcname, cuda_error (r)); + ret = false; + } + + if (funcname_mgomp) + pthread_mutex_unlock (&ptx_dev->omp_stacks.lock); + } + + if (funcname_mgomp) + free (funcname_mgomp); + + return ret; +} + /* Load the (partial) program described by TARGET_DATA to device number ORD. Allocate and return TARGET_TABLE. If not NULL, REV_FN_TABLE will contain the on-device addresses of the functions for reverse offload. @@ -1485,6 +1589,9 @@ GOMP_OFFLOAD_load_image (int ord, unsigned version, const void *target_data, nvptx_set_clocktick (module, dev); + if (!nvptx_do_global_cdtors (module, dev, "__do_global_ctors__entry")) + return -1; + return fn_entries + var_entries + other_entries; } @@ -1510,6 +1617,10 @@ GOMP_OFFLOAD_unload_image (int ord, unsigned version, const void *target_data) for (prev_p = &dev->images; (image = *prev_p) != 0; prev_p = &image->next) if (image->target_data == target_data) { + if (!nvptx_do_global_cdtors (image->module, dev, + "__do_global_dtors__entry")) + ret = false; + *prev_p = image->next; if (CUDA_CALL_NOCHECK (cuModuleUnload, image->module) != CUDA_SUCCESS) ret = false; -- 2.25.1