[libgomp] Rewire OpenACC async

Message ID	56FA4F69.1060101@codesourcery.com
State	New
Headers	show Return-Path: <gcc-patches-return-423919-incoming=patchwork.ozlabs.org@gcc.gnu.org> DomainKey-Signature: a=rsa-sha1; c=nofws; d=gcc.gnu.org; h=list-id :list-unsubscribe:list-archive:list-post:list-help:sender :subject:to:references:cc:from:message-id:date:mime-version :in-reply-to:content-type; q=dns; s=default; b=v6jEFdmcy2GyBsQK8 0tS/s3Xfap0aOlyFohYh35Sy6Den1cMlP8IE0ghVq1fzEgkis81At4oPQaGzR3wi LHWq4N2H/JSp82U6JwGQ+sAyDY8vyA8ey3oTga3Dx1sEyvm5/a0D54dkuOW4QlwC jTg7Ygat/uAdiC4OKGCklMBiVw= Mailing-List: contact gcc-patches-help@gcc.gnu.org; run by ezmlm Precedence: bulk Sender: gcc-patches-owner@gcc.gnu.org Subject: Re: [PATCH, libgomp] Rewire OpenACC async To: gcc-patches <gcc-patches@gcc.gnu.org> References: <56543B8C.404@codesourcery.com> <567910D1.1060306@codesourcery.com> CC: Jakub Jelinek <jakub@redhat.com>, Thomas Schwinge <thomas_schwinge@mentor.com>, Cesar Philippidis <cesar_philippidis@mentor.com>, Julian Brown <julian@codesourcery.com> From: Chung-Lin Tang <cltang@codesourcery.com> Message-ID: <56FA4F69.1060101@codesourcery.com> Date: Tue, 29 Mar 2016 17:48:25 +0800 User-Agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10.11; rv:38.0) Gecko/20100101 Thunderbird/38.7.1 MIME-Version: 1.0 In-Reply-To: <567910D1.1060306@codesourcery.com> Content-Type: multipart/mixed; boundary="------------090105050600030702030606"

Index: oacc-host.c =================================================================== --- oacc-host.c (revision 234516) +++ oacc-host.c (working copy) @@ -144,7 +144,8 @@ host_openacc_exec (void (*fn) (void *), } static void -host_openacc_register_async_cleanup (void *targ_mem_desc __attribute__ ((unused))) +host_openacc_register_async_cleanup (void *targ_mem_desc __attribute__ ((unused)), + int async __attribute__ ((unused))) { } Index: oacc-mem.c =================================================================== --- oacc-mem.c (revision 234516) +++ oacc-mem.c (working copy) @@ -661,10 +661,7 @@ gomp_acc_remove_pointer (void *h, bool force_copyf if (async < acc_async_noval) gomp_unmap_vars (t, true); else - { - gomp_copy_from_async (t); - acc_dev->openacc.register_async_cleanup_func (t); - } + t->device_descr->openacc.register_async_cleanup_func (t, async); gomp_debug (0, " %s: mappings restored\n", __FUNCTION__); } Index: oacc-parallel.c =================================================================== --- oacc-parallel.c (revision 234516) +++ oacc-parallel.c (working copy) @@ -186,10 +186,7 @@ GOACC_parallel_keyed (int device, void (*fn) (void if (async < acc_async_noval) gomp_unmap_vars (tgt, true); else - { - gomp_copy_from_async (tgt); - acc_dev->openacc.register_async_cleanup_func (tgt); - } + tgt->device_descr->openacc.register_async_cleanup_func (tgt, async); acc_dev->openacc.async_set_async_func (acc_async_sync); } Index: target.c =================================================================== --- target.c (revision 234516) +++ target.c (working copy) @@ -663,7 +663,6 @@ gomp_map_vars (struct gomp_device_descr *devicep, tgt->list[i].offset = 0; tgt->list[i].length = k->host_end - k->host_start; k->refcount = 1; - k->async_refcount = 0; tgt->refcount++; array->left = NULL; array->right = NULL; @@ -813,40 +812,6 @@ gomp_unmap_tgt (struct target_mem_desc *tgt) free (tgt); } -/* Decrease the refcount for a set of mapped variables, and queue asychronous - copies from the device back to the host after any work that has been issued. - Because the regions are still "live", increment an asynchronous reference - count to indicate that they should not be unmapped from host-side data - structures until the asynchronous copy has completed. */ - -attribute_hidden void -gomp_copy_from_async (struct target_mem_desc *tgt) -{ - struct gomp_device_descr *devicep = tgt->device_descr; - size_t i; - - gomp_mutex_lock (&devicep->lock); - - for (i = 0; i < tgt->list_count; i++) - if (tgt->list[i].key == NULL) - ; - else if (tgt->list[i].key->refcount > 1) - { - tgt->list[i].key->refcount--; - tgt->list[i].key->async_refcount++; - } - else - { - splay_tree_key k = tgt->list[i].key; - if (tgt->list[i].copy_from) - devicep->dev2host_func (devicep->target_id, (void *) k->host_start, - (void *) (k->tgt->tgt_start + k->tgt_offset), - k->host_end - k->host_start); - } - - gomp_mutex_unlock (&devicep->lock); -} - /* Unmap variables described by TGT. If DO_COPYFROM is true, copy relevant variables back from device to host: if it is false, it is assumed that this has been done already, i.e. by gomp_copy_from_async above. */ @@ -883,13 +848,8 @@ gomp_unmap_vars (struct target_mem_desc *tgt, bool k->refcount--; else if (k->refcount == 1) { - if (k->async_refcount > 0) - k->async_refcount--; - else - { - k->refcount--; - do_unmap = true; - } + k->refcount--; + do_unmap = true; } if ((do_unmap && do_copyfrom && tgt->list[i].copy_from) @@ -1040,7 +1000,6 @@ gomp_load_image_to_device (struct gomp_device_desc k->tgt = tgt; k->tgt_offset = target_table[i].start; k->refcount = REFCOUNT_INFINITY; - k->async_refcount = 0; k->link_key = NULL; array->left = NULL; array->right = NULL; @@ -1073,7 +1032,6 @@ gomp_load_image_to_device (struct gomp_device_desc k->tgt = tgt; k->tgt_offset = target_var->start; k->refcount = target_size & link_bit ? REFCOUNT_LINK : REFCOUNT_INFINITY; - k->async_refcount = 0; k->link_key = NULL; array->left = NULL; array->right = NULL; @@ -2299,7 +2257,6 @@ omp_target_associate_ptr (void *host_ptr, void *de k->tgt = tgt; k->tgt_offset = (uintptr_t) device_ptr + device_offset; k->refcount = REFCOUNT_INFINITY; - k->async_refcount = 0; array->left = NULL; array->right = NULL; splay_tree_insert (&devicep->mem_map, array); Index: libgomp.h =================================================================== --- libgomp.h (revision 234516) +++ libgomp.h (working copy) @@ -837,8 +837,6 @@ struct splay_tree_key_s { uintptr_t tgt_offset; /* Reference count. */ uintptr_t refcount; - /* Asynchronous reference count. */ - uintptr_t async_refcount; /* Pointer to the original mapping of "omp declare target link" object. */ splay_tree_key link_key; }; @@ -874,7 +872,7 @@ typedef struct acc_dispatch_t unsigned *, void *); /* Async cleanup callback registration. */ - void (*register_async_cleanup_func) (void *); + void (*register_async_cleanup_func) (void *, int); /* Asynchronous routines. */ int (*async_test_func) (int); @@ -979,7 +977,6 @@ extern struct target_mem_desc *gomp_map_vars (stru size_t, void **, void **, size_t *, void *, bool, enum gomp_map_vars_kind); -extern void gomp_copy_from_async (struct target_mem_desc *); extern void gomp_unmap_vars (struct target_mem_desc *, bool); extern void gomp_init_device (struct gomp_device_descr *); extern void gomp_free_memmap (struct splay_tree_s *); Index: oacc-plugin.c =================================================================== --- oacc-plugin.c (revision 234516) +++ oacc-plugin.c (working copy) @@ -31,11 +31,14 @@ #include "oacc-int.h" void -GOMP_PLUGIN_async_unmap_vars (void *ptr) +GOMP_PLUGIN_async_unmap_vars (void *ptr, int async) { struct target_mem_desc *tgt = ptr; + struct gomp_device_descr *devicep = tgt->device_descr; - gomp_unmap_vars (tgt, false); + devicep->openacc.async_set_async_func (async); + gomp_unmap_vars (tgt, true); + devicep->openacc.async_set_async_func (acc_async_sync); } /* Return the target-specific part of the TLS data for the current thread. */ Index: oacc-plugin.h =================================================================== --- oacc-plugin.h (revision 234516) +++ oacc-plugin.h (working copy) @@ -27,7 +27,7 @@ #ifndef OACC_PLUGIN_H #define OACC_PLUGIN_H 1 -extern void GOMP_PLUGIN_async_unmap_vars (void *); +extern void GOMP_PLUGIN_async_unmap_vars (void *, int); extern void *GOMP_PLUGIN_acc_thread (void); #endif Index: plugin/plugin-nvptx.c =================================================================== --- plugin/plugin-nvptx.c (revision 234516) +++ plugin/plugin-nvptx.c (working copy) @@ -310,6 +310,7 @@ struct ptx_event int type; void *addr; int ord; + int val; struct ptx_event *next; }; @@ -783,6 +784,7 @@ static void event_gc (bool memmap_lockable) { struct ptx_event *ptx_event = ptx_events; + struct ptx_event *async_cleanups = NULL; struct nvptx_thread *nvthd = nvptx_thread (); pthread_mutex_lock (&ptx_event_lock); @@ -800,6 +802,7 @@ event_gc (bool memmap_lockable) r = cuEventQuery (*e->evt); if (r == CUDA_SUCCESS) { + bool append_async = false; CUevent *te; te = e->evt; @@ -824,7 +827,7 @@ event_gc (bool memmap_lockable) if (!memmap_lockable) continue; - GOMP_PLUGIN_async_unmap_vars (e->addr); + append_async = true; } break; } @@ -832,6 +835,7 @@ event_gc (bool memmap_lockable) cuEventDestroy (*te); free ((void *)te); + /* Unlink 'e' from ptx_events list. */ if (ptx_events == e) ptx_events = ptx_events->next; else @@ -842,15 +846,31 @@ event_gc (bool memmap_lockable) e_->next = e_->next->next; } - free (e); + if (append_async) + { + e->next = async_cleanups; + async_cleanups = e; + } + else + free (e); } } pthread_mutex_unlock (&ptx_event_lock); + + /* We have to do these here, after ptx_event_lock is released. */ + while (async_cleanups) + { + struct ptx_event *e = async_cleanups; + async_cleanups = async_cleanups->next; + + GOMP_PLUGIN_async_unmap_vars (e->addr, e->val); + free (e); + } } static void -event_add (enum ptx_event_type type, CUevent *e, void *h) +event_add (enum ptx_event_type type, CUevent *e, void *h, int val) { struct ptx_event *ptx_event; struct nvptx_thread *nvthd = nvptx_thread (); @@ -863,6 +883,7 @@ static void ptx_event->evt = e; ptx_event->addr = h; ptx_event->ord = nvthd->ptx_dev->ord; + ptx_event->val = val; pthread_mutex_lock (&ptx_event_lock); @@ -975,7 +996,7 @@ nvptx_exec (void (*fn), size_t mapnum, void **host if (r != CUDA_SUCCESS) GOMP_PLUGIN_fatal ("cuEventRecord error: %s", cuda_error (r)); - event_add (PTX_EVT_KNL, e, (void *)dev_str); + event_add (PTX_EVT_KNL, e, (void *)dev_str, 0); } #else r = cuCtxSynchronize (); @@ -1082,7 +1103,7 @@ nvptx_host2dev (void *d, const void *h, size_t s) if (r != CUDA_SUCCESS) GOMP_PLUGIN_fatal ("cuEventRecord error: %s", cuda_error (r)); - event_add (PTX_EVT_MEM, e, (void *)h); + event_add (PTX_EVT_MEM, e, (void *)h, 0); } else #endif @@ -1147,7 +1168,7 @@ nvptx_dev2host (void *h, const void *d, size_t s) if (r != CUDA_SUCCESS) GOMP_PLUGIN_fatal ("cuEventRecord error: %s", cuda_error (r)); - event_add (PTX_EVT_MEM, e, (void *)h); + event_add (PTX_EVT_MEM, e, (void *)h, 0); } else #endif @@ -1273,7 +1294,7 @@ nvptx_wait_async (int async1, int async2) if (r != CUDA_SUCCESS) GOMP_PLUGIN_fatal ("cuEventRecord error: %s", cuda_error (r)); - event_add (PTX_EVT_SYNC, e, NULL); + event_add (PTX_EVT_SYNC, e, NULL, 0); r = cuStreamWaitEvent (s2->stream, *e, 0); if (r != CUDA_SUCCESS) @@ -1355,7 +1376,7 @@ nvptx_wait_all_async (int async) if (r != CUDA_SUCCESS) GOMP_PLUGIN_fatal ("cuEventRecord error: %s", cuda_error (r)); - event_add (PTX_EVT_SYNC, e, NULL); + event_add (PTX_EVT_SYNC, e, NULL, 0); r = cuStreamWaitEvent (waiting_stream->stream, *e, 0); if (r != CUDA_SUCCESS) @@ -1667,7 +1688,7 @@ GOMP_OFFLOAD_openacc_parallel (void (*fn) (void *) } void -GOMP_OFFLOAD_openacc_register_async_cleanup (void *targ_mem_desc) +GOMP_OFFLOAD_openacc_register_async_cleanup (void *targ_mem_desc, int async) { CUevent *e; CUresult r; @@ -1683,7 +1704,7 @@ void if (r != CUDA_SUCCESS) GOMP_PLUGIN_fatal ("cuEventRecord error: %s", cuda_error (r)); - event_add (PTX_EVT_ASYNC_CLEANUP, e, targ_mem_desc); + event_add (PTX_EVT_ASYNC_CLEANUP, e, targ_mem_desc, async); } int

[libgomp] Rewire OpenACC async

Commit Message

Comments

Patch