stv0g/metalsvm
1
0
Fork 0
Code Issues Pull requests Releases Wiki Activity

Merge branch 'master' of git.lfbs.rwth-aachen.de:metalsvm into jacobi

Browse source
This commit is contained in:
Stefan Lankes 2011-04-19 17:52:26 +02:00
commit a62b4949ee
18 changed files with 1867 additions and 18 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

154
arch/x86/include/asm/iRCCE.h Normal file
View file

@ -0,0 +1,154 @@
//
// Copyright 2010 Intel Corporation
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
//
// [2010-10-25] added support for non-blocking send/recv operations
// - iRCCE_isend(), ..._test(), ..._wait(), ..._push()
// - iRCCE_irecv(), ..._test(), ..._wait(), ..._push()
// by Carsten Clauss, Chair for Operating Systems,
// RWTH Aachen University
//
// [2010-11-12] extracted non-blocking code into separate library
// by Carsten Scholtes, University of Bayreuth
//
// [2010-12-09] added functions for a convenient handling of multiple
// pending non-blocking requests
// by Jacek Galowicz, Chair for Operating Systems
// RWTH Aachen University
//
#ifndef IRCCE_H
#define IRCCE_H
#include <asm/RCCE.h>
#define iRCCE_SUCCESS RCCE_SUCCESS
#define iRCCE_PENDING -1
#define iRCCE_RESERVED -2
#define iRCCE_NOT_ENQUEUED -3
typedef struct _iRCCE_SEND_REQUEST {
char *privbuf; // source buffer in local private memory (send buffer)
t_vcharp combuf; // intermediate buffer in MPB
size_t chunk; // size of MPB available for this message (bytes)
RCCE_FLAG *ready; // flag indicating whether receiver is ready
RCCE_FLAG *sent; // flag indicating whether message has been sent by source
size_t size; // size of message (bytes)
int dest; // UE that will receive the message
size_t wsize; // offset within send buffer when putting in "chunk" bytes
size_t remainder; // bytes remaining to be sent
size_t nbytes; // number of bytes to be sent in single RCCE_put call
char *bufptr; // running pointer inside privbuf for current location
int label; // jump/goto label for the reentrance of the respective poll function
int finished; // flag that indicates whether the request has already been finished
struct _iRCCE_SEND_REQUEST *next;
} iRCCE_SEND_REQUEST;
typedef struct _iRCCE_RECV_REQUEST {
char *privbuf; // source buffer in local private memory (send buffer)
t_vcharp combuf; // intermediate buffer in MPB
size_t chunk; // size of MPB available for this message (bytes)
RCCE_FLAG *ready; // flag indicating whether receiver is ready
RCCE_FLAG *sent; // flag indicating whether message has been sent by source
size_t size; // size of message (bytes)
int source; // UE that will send the message
size_t wsize; // offset within send buffer when putting in "chunk" bytes
size_t remainder; // bytes remaining to be sent
size_t nbytes; // number of bytes to be sent in single RCCE_put call
char *bufptr; // running pointer inside privbuf for current location
int label; // jump/goto label for the reentrance of the respective poll function
int finished; // flag that indicates whether the request has already been finished
int started; // flag that indicates whether message parts have already been received
struct _iRCCE_RECV_REQUEST *next;
} iRCCE_RECV_REQUEST;
#define iRCCE_WAIT_LIST_RECV_TYPE 0
#define iRCCE_WAIT_LIST_SEND_TYPE 1
typedef struct _iRCCE_WAIT_LISTELEM {
int type;
struct _iRCCE_WAIT_LISTELEM * next;
void * req;
} iRCCE_WAIT_LISTELEM;
typedef struct _iRCCE_WAIT_LIST {
iRCCE_WAIT_LISTELEM * first;
iRCCE_WAIT_LISTELEM * last;
} iRCCE_WAIT_LIST;
///////////////////////////////////////////////////////////////
//
// THE iRCCE API:
//
// Initialize function:
int iRCCE_init(void);
//
// Non-blocking send/recv functions:
int iRCCE_isend(char *, size_t, int, iRCCE_SEND_REQUEST *);
int iRCCE_isend_test(iRCCE_SEND_REQUEST *, int *);
int iRCCE_isend_wait(iRCCE_SEND_REQUEST *);
int iRCCE_isend_push(void);
int iRCCE_irecv(char *, size_t, int, iRCCE_RECV_REQUEST *);
int iRCCE_irecv_test(iRCCE_RECV_REQUEST *, int *);
int iRCCE_irecv_wait(iRCCE_RECV_REQUEST *);
int iRCCE_irecv_push(void);
//
// Blocking but pipelined send/recv functions:
int iRCCE_send(char *, size_t, int);
int iRCCE_recv(char *, size_t, int);
//
// SCC-customized put/get and memcpy functions:
int iRCCE_put(t_vcharp, t_vcharp, int, int);
int iRCCE_get(t_vcharp, t_vcharp, int, int);
void* iRCCE_memcpy_put(void*, const void*, size_t);
void* iRCCE_memcpy_get(void*, const void*, size_t);
//
// Wait/test-all/any functions:
void iRCCE_init_wait_list(iRCCE_WAIT_LIST*);
void iRCCE_add_to_wait_list(iRCCE_WAIT_LIST*, iRCCE_SEND_REQUEST *, iRCCE_RECV_REQUEST *);
int iRCCE_test_all(iRCCE_WAIT_LIST*, int *);
int iRCCE_wait_all(iRCCE_WAIT_LIST*);
int iRCCE_test_any(iRCCE_WAIT_LIST*, iRCCE_SEND_REQUEST **, iRCCE_RECV_REQUEST **);
int iRCCE_wait_any(iRCCE_WAIT_LIST*, iRCCE_SEND_REQUEST **, iRCCE_RECV_REQUEST **);
//
// Cancel functions for yet not started non-blocking requests:
int iRCCE_isend_cancel(iRCCE_SEND_REQUEST *, int *);
int iRCCE_irecv_cancel(iRCCE_RECV_REQUEST *, int *);
//
///////////////////////////////////////////////////////////////
//
// Just for for convenience:
#if 1
#define RCCE_isend iRCCE_isend
#define RCCE_isend_test iRCCE_isend_test
#define RCCE_isend_wait iRCCE_isend_wait
#define RCCE_isend_push iRCCE_isend_push
#define RCCE_irecv iRCCE_irecv
#define RCCE_irecv_test iRCCE_irecv_test
#define RCCE_irecv_wait iRCCE_irecv_wait
#define RCCE_irecv_push iRCCE_irecv_push
#define RCCE_SEND_REQUEST iRCCE_SEND_REQUEST
#define RCCE_RECV_REQUEST iRCCE_RECV_REQUEST
#endif
///////////////////////////////////////////////////////////////
#endif

39
arch/x86/include/asm/iRCCE_lib.h Normal file
View file

@ -0,0 +1,39 @@
//
// Copyright 2010 Intel Corporation
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
//
// [2010-10-25] added support for non-blocking send/recv operations
// - iRCCE_isend(), ..._test(), ..._wait(), ..._push()
// - iRCCE_irecv(), ..._test(), ..._wait(), ..._push()
// by Carsten Clauss, Chair for Operating Systems,
// RWTH Aachen University
//
// [2010-11-12] extracted non-blocking code into separate library
// by Carsten Scholtes
//
#ifndef IRCCE_LIB_H
#define IRCCE_LIB_H
#include <asm/iRCCE.h>
#include <asm/RCCE_lib.h>
extern iRCCE_SEND_REQUEST* iRCCE_isend_queue;
extern iRCCE_RECV_REQUEST* iRCCE_irecv_queue[RCCE_MAXNP];
#ifdef _OPENMP
#pragma omp threadprivate (iRCCE_isend_queue, iRCCE_irecv_queue)
#endif
int iRCCE_test_flag(RCCE_FLAG, RCCE_FLAG_STATUS, int *);
#endif

2
arch/x86/scc/Makefile
View file

@ -1,4 +1,4 @@
C_source := scc_init.c SCC_API.c RCCE_malloc.c RCCE_shmalloc.c RCCE_debug.c RCCE_qsort.c RCCE_DCMflush.c RCCE_send.c RCCE_recv.c RCCE_flags.c RCCE_comm.c RCCE_put.c RCCE_get.c RCCE_synch.c RCCE_bcast.c RCCE_admin.c # RCCE_power_management.c
C_source := scc_init.c SCC_API.c iRCCE_admin.c iRCCE_send.c iRCCE_isend.c iRCCE_irecv.c iRCCE_recv.c iRCCE_get.c iRCCE_put.c iRCCE_synch.c RCCE_malloc.c RCCE_shmalloc.c RCCE_debug.c RCCE_qsort.c RCCE_DCMflush.c RCCE_send.c RCCE_recv.c RCCE_flags.c RCCE_comm.c RCCE_put.c RCCE_get.c RCCE_synch.c RCCE_bcast.c RCCE_admin.c # RCCE_power_management.c
ASM_source :=
MODULE := arch_x86_scc

2
arch/x86/scc/RCCE_admin.c
View file

@ -45,7 +45,7 @@
// #include <fcntl.h>
// En-/ or disable debug prints...
#define DEBUG 1
#define DEBUG 0
//......................................................................................
// GLOBAL VARIABLES USED BY THE LIBRARY

67
arch/x86/scc/iRCCE_admin.c Normal file
View file

@ -0,0 +1,67 @@
//***************************************************************************************
// Administrative routines.
//***************************************************************************************
//
// Author: Rob F. Van der Wijngaart
// Intel Corporation
// Date: 008/30/2010
//
//***************************************************************************************
//
//
// Copyright 2010 Intel Corporation
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
//
// [2010-10-25] added support for non-blocking send/recv operations
// - iRCCE_isend(), ..._test(), ..._wait(), ..._push()
// - iRCCE_irecv(), ..._test(), ..._wait(), ..._push()
// by Carsten Clauss, Chair for Operating Systems,
// RWTH Aachen University
//
// [2010-11-12] extracted non-blocking code into separate library
// by Carsten Scholtes
//
// [2011-02-21] added support for multiple incoming queues
// (one recv queue per remote rank)
//
#include <metalsvm/stddef.h>
#ifdef CONFIG_ROCKCREEK
#include <asm/iRCCE_lib.h>
// send request queue
iRCCE_SEND_REQUEST* iRCCE_isend_queue;
// recv request queue
iRCCE_RECV_REQUEST* iRCCE_irecv_queue[RCCE_MAXNP];
//--------------------------------------------------------------------------------------
// FUNCTION: iRCCE_init
//--------------------------------------------------------------------------------------
// initialize the library
//--------------------------------------------------------------------------------------
int iRCCE_init(void) {
int i;
for(i=0; i<RCCE_MAXNP; i++) {
iRCCE_irecv_queue[i] = NULL;
}
iRCCE_isend_queue = NULL;
return (iRCCE_SUCCESS);
}
#endif

85
arch/x86/scc/iRCCE_get.c Normal file
View file

@ -0,0 +1,85 @@
//***************************************************************************************
// Get data from communication buffer.
//***************************************************************************************
//
// Author: Rob F. Van der Wijngaart
// Intel Corporation
// Date: 008/30/2010
//
//***************************************************************************************
//
// Copyright 2010 Intel Corporation
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
//
// [2010-11-03] switched to SCC-optimized memcpy() functions in scc_memcpy.h:
// - memcpy_to_mpb()
// - memcpy_from_mpb()
// by Stefan Lankes, Carsten Clauss, Chair for Operating Systems,
// RWTH Aachen University
//
#include <metalsvm/stdlib.h>
#include <metalsvm/string.h>
#ifdef CONFIG_ROCKCREEK
#include <asm/iRCCE_lib.h>
#ifdef COPPERRIDGE
#include "scc_memcpy.h"
#endif
void* iRCCE_memcpy_get(void *dest, const void *src, size_t count)
{
#ifdef COPPERRIDGE
return memcpy_from_mpb(dest, src, count);
#else
return memcpy(dest, src, count);
#endif
}
//--------------------------------------------------------------------------------------
// FUNCTION: iRCCE_get
//--------------------------------------------------------------------------------------
// copy data from address "source" in the remote MPB to address "target" in either the
// local MPB, or in the calling UE's private memory. We do not test to see if a move
// into the calling UE's private memory stays within allocated memory *
//--------------------------------------------------------------------------------------
int iRCCE_get(
t_vcharp target, // target buffer, MPB or private memory
t_vcharp source, // source buffer, MPB
int num_bytes, // number of bytes to copy (must be multiple of cache line size
int ID // rank of source UE
) {
// in non-GORY mode we only need to retain the MPB source shift; we
// already know the source is in the MPB, not private memory
source = RCCE_comm_buffer[ID]+(source-RCCE_comm_buffer[RCCE_IAM]);
// do the actual copy, making sure we copy fresh data
#ifdef _OPENMP
#pragma omp flush
#endif
RC_cache_invalidate();
iRCCE_memcpy_get((void *)target, (void *)source, num_bytes);
// flush data to make sure it is visible to all threads; cannot use a flush list
// because it concerns malloced space
#ifdef _OPENMP
#pragma omp flush
#endif
return(iRCCE_SUCCESS);
}
#endif

393
arch/x86/scc/iRCCE_irecv.c Normal file
View file

@ -0,0 +1,393 @@
//***************************************************************************************
// Synchronized receive routines.
//***************************************************************************************
//
// Author: Rob F. Van der Wijngaart
// Intel Corporation
// Date: 008/30/2010
//
//***************************************************************************************
//
// Copyright 2010 Intel Corporation
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
//
// [2010-10-25] added support for non-blocking send/recv operations
// - iRCCE_isend(), ..._test(), ..._wait(), ..._push()
// - iRCCE_irecv(), ..._test(), ..._wait(), ..._push()
// by Carsten Clauss, Chair for Operating Systems,
// RWTH Aachen University
//
// [2010-11-12] extracted non-blocking code into separate library
// by Carsten Scholtes
//
// [2010-12-09] added cancel functions for non-blocking send/recv requests
// by Carsten Clauss
//
// [2011-02-21] added support for multiple incoming queues
// (one recv queue per remote rank)
//
#include <metalsvm/stdlib.h>
#include <metalsvm/string.h>
#ifdef CONFIG_ROCKCREEK
#include <asm/iRCCE_lib.h>
static int iRCCE_push_recv_request(iRCCE_RECV_REQUEST *request) {
char padline[RCCE_LINE_SIZE]; // copy buffer, used if message not multiple of line size
int test; // flag for calling iRCCE_test_flag()
if(request->finished) return(iRCCE_SUCCESS);
if(request->label == 1) goto label1;
if(request->label == 2) goto label2;
if(request->label == 3) goto label3;
// receive data in units of available chunk size of MPB
for (; request->wsize < (request->size / request->chunk) * request->chunk; request->wsize += request->chunk) {
request->bufptr = request->privbuf + request->wsize;
request->nbytes = request->chunk;
label1:
iRCCE_test_flag(*(request->sent), RCCE_FLAG_SET, &test);
if(!test) {
request->label = 1;
return(iRCCE_PENDING);
}
request->started = 1;
RCCE_flag_write(request->sent, RCCE_FLAG_UNSET, RCCE_IAM);
// copy data from source's MPB space to private memory
iRCCE_get((t_vcharp)request->bufptr, request->combuf, request->nbytes, request->source);
// tell the source I have moved data out of its comm buffer
RCCE_flag_write(request->ready, RCCE_FLAG_SET, request->source);
}
request->remainder = request->size % request->chunk;
// if nothing is left over, we are done
if (!request->remainder) {
request->finished = 1;
return(iRCCE_SUCCESS);
}
// receive remainder of data--whole cache lines
request->bufptr = request->privbuf + (request->size / request->chunk) * request->chunk;
request->nbytes = request->remainder - request->remainder % RCCE_LINE_SIZE;
if (request->nbytes) {
label2:
iRCCE_test_flag(*(request->sent), RCCE_FLAG_SET, &test);
if(!test) {
request->label = 2;
return(iRCCE_PENDING);
}
request->started = 1;
RCCE_flag_write(request->sent, RCCE_FLAG_UNSET, RCCE_IAM);
// copy data from source's MPB space to private memory
iRCCE_get((t_vcharp)request->bufptr, request->combuf, request->nbytes, request->source);
// tell the source I have moved data out of its comm buffer
RCCE_flag_write(request->ready, RCCE_FLAG_SET, request->source);
}
request->remainder = request->size % request->chunk;
request->remainder = request->remainder % RCCE_LINE_SIZE;
if (!request->remainder) {
request->finished = 1;
return(iRCCE_SUCCESS);
}
// remainder is less than cache line. This must be copied into appropriately sized
// intermediate space before exact number of bytes get copied to the final destination
request->bufptr = request->privbuf + (request->size / request->chunk) * request->chunk + request->nbytes;
request->nbytes = RCCE_LINE_SIZE;
label3:
iRCCE_test_flag(*(request->sent), RCCE_FLAG_SET, &test);
if(!test) {
request->label = 3;
return(iRCCE_PENDING);
}
request->started = 1;
RCCE_flag_write(request->sent, RCCE_FLAG_UNSET, RCCE_IAM);
// copy data from source's MPB space to private memory
iRCCE_get((t_vcharp)padline, request->combuf, request->nbytes, request->source);
memcpy(request->bufptr,padline,request->remainder);
// tell the source I have moved data out of its comm buffer
RCCE_flag_write(request->ready, RCCE_FLAG_SET, request->source);
request->finished = 1;
return(iRCCE_SUCCESS);
}
static void iRCCE_init_recv_request(
char *privbuf, // source buffer in local private memory (send buffer)
t_vcharp combuf, // intermediate buffer in MPB
size_t chunk, // size of MPB available for this message (bytes)
RCCE_FLAG *ready, // flag indicating whether receiver is ready
RCCE_FLAG *sent, // flag indicating whether message has been sent by source
size_t size, // size of message (bytes)
int source, // UE that will send the message
iRCCE_RECV_REQUEST *request
) {
request->privbuf = privbuf;
request->combuf = combuf;
request->chunk = chunk;
request->ready = ready;
request->sent = sent;
request->size = size;
request->source = source;
request->wsize = 0;
request->remainder = 0;
request->nbytes = 0;
request->bufptr = NULL;
request->label = 0;
request->finished = 0;
request->started = 0;
request->next = NULL;
return;
}
//--------------------------------------------------------------------------------------
// FUNCTION: iRCCE_irecv
//--------------------------------------------------------------------------------------
// non-blocking recv function; returns an handle of type iRCCE_RECV_REQUEST
//--------------------------------------------------------------------------------------
static iRCCE_RECV_REQUEST blocking_irecv_request;
int iRCCE_irecv(char *privbuf, size_t size, int source, iRCCE_RECV_REQUEST *request) {
if(request == NULL) request = &blocking_irecv_request;
if (source<0 || source >= RCCE_NP)
return(RCCE_error_return(RCCE_debug_comm,RCCE_ERROR_ID));
else {
iRCCE_init_recv_request(privbuf, RCCE_buff_ptr, RCCE_chunk,
&RCCE_ready_flag[RCCE_IAM], &RCCE_sent_flag[source],
size, source, request);
if(iRCCE_irecv_queue[source] == NULL) {
if(iRCCE_push_recv_request(request) == iRCCE_SUCCESS) {
return(iRCCE_SUCCESS);
}
else {
iRCCE_irecv_queue[source] = request;
if(request == &blocking_irecv_request) {
iRCCE_irecv_wait(request);
return(iRCCE_SUCCESS);
}
return(iRCCE_PENDING);
}
}
else {
if(iRCCE_irecv_queue[source]->next == NULL) {
iRCCE_irecv_queue[source]->next = request;
}
else {
iRCCE_RECV_REQUEST *run = iRCCE_irecv_queue[source];
while(run->next != NULL) run = run->next;
run->next = request;
}
if(request == &blocking_irecv_request) {
iRCCE_irecv_wait(request);
return(iRCCE_SUCCESS);
}
return(iRCCE_RESERVED);
}
}
}
//--------------------------------------------------------------------------------------
// FUNCTION: iRCCE_irecv_test
//--------------------------------------------------------------------------------------
// test function for completion of the requestes non-blocking recv operation
// Just provide NULL instead of the testvar if you don't need it
//--------------------------------------------------------------------------------------
int iRCCE_irecv_test(iRCCE_RECV_REQUEST *request, int *test) {
int source;
if(request == NULL) {
if(iRCCE_irecv_push() == iRCCE_SUCCESS) {
if (test) (*test) = 1;
return(iRCCE_SUCCESS);
}
else {
if (test) (*test) = 0;
return(iRCCE_PENDING);
}
}
source = request->source;
if(request->finished) {
if (test) (*test) = 1;
return(iRCCE_SUCCESS);
}
if(iRCCE_irecv_queue[source] != request) {
if (test) (*test) = 0;
return(iRCCE_RESERVED);
}
iRCCE_push_recv_request(request);
if(request->finished) {
iRCCE_irecv_queue[source] = request->next;
if (test) (*test) = 1;
return(iRCCE_SUCCESS);
}
if (test) (*test) = 0;
return(iRCCE_PENDING);
}
//--------------------------------------------------------------------------------------
// FUNCTION: iRCCE_irecv_push
//--------------------------------------------------------------------------------------
// progress function for pending requests in the irecv queue
//--------------------------------------------------------------------------------------
static int iRCCE_irecv_push_source(int source) {
iRCCE_RECV_REQUEST *request = iRCCE_irecv_queue[source];
if(request == NULL) {
return(iRCCE_SUCCESS);
}
if(request->finished) {
return(iRCCE_SUCCESS);
}
iRCCE_push_recv_request(request);
if(request->finished) {
iRCCE_irecv_queue[source] = request->next;
return(iRCCE_SUCCESS);
}
return(iRCCE_PENDING);
}
int iRCCE_irecv_push(void) {
int i, j;
int retval = iRCCE_SUCCESS;
for(i=0; i<RCCE_NP; i++) {
j = iRCCE_irecv_push_source(i);
if(j != iRCCE_SUCCESS) {
retval = j;
}
}
return retval;
}
//--------------------------------------------------------------------------------------
// FUNCTION: iRCCE_irecv_wait
//--------------------------------------------------------------------------------------
// just wait for completion of the requested non-blocking send operation
//--------------------------------------------------------------------------------------
int iRCCE_irecv_wait(iRCCE_RECV_REQUEST *request) {
if(request != NULL) {
while(!request->finished) {
iRCCE_irecv_push();
iRCCE_isend_push();
}
}
else {
do {
iRCCE_isend_push();
}
while( iRCCE_irecv_push() != iRCCE_SUCCESS );
}
return(iRCCE_SUCCESS);
}
//--------------------------------------------------------------------------------------
// FUNCTION: iRCCE_irecv_cancel
//--------------------------------------------------------------------------------------
// try to cancel a pending non-blocking recv request
//--------------------------------------------------------------------------------------
int iRCCE_irecv_cancel(iRCCE_RECV_REQUEST *request, int *test) {
int source;
iRCCE_RECV_REQUEST *run;
if( (request == NULL) || (request->finished) ) {
if (test) (*test) = 0;
return iRCCE_NOT_ENQUEUED;
}
source = request->source;
if(iRCCE_irecv_queue[source] == NULL) {
if (test) (*test) = 0;
return iRCCE_NOT_ENQUEUED;
}
if(iRCCE_irecv_queue[source] == request) {
// have parts of the message already been received?
if(request->started) {
if (test) (*test) = 0;
return iRCCE_PENDING;
}
else {
// no, thus request can be canceld just in time:
iRCCE_irecv_queue[source] = request->next;
if (test) (*test) = 1;
return iRCCE_SUCCESS;
}
}
for(run = iRCCE_irecv_queue[source]; run->next != NULL; run = run->next) {
// request found --> remove it from recv queue:
if(run->next == request) {
run->next = run->next->next;
if (test) (*test) = 1;
return iRCCE_SUCCESS;
}
}
if (test) (*test) = 0;
return iRCCE_NOT_ENQUEUED;
}
#endif

355
arch/x86/scc/iRCCE_isend.c Normal file
View file

@ -0,0 +1,355 @@
//***************************************************************************************
// Non-blocking send routines.
//***************************************************************************************
//
// Author: Rob F. Van der Wijngaart
// Intel Corporation
// Date: 008/30/2010
//
//***************************************************************************************
//
// Copyright 2010 Intel Corporation
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
//
// [2010-10-25] added support for non-blocking send/recv operations
// - iRCCE_isend(), ..._test(), ..._wait(), ..._push()
// - iRCCE_irecv(), ..._test(), ..._wait(), ..._push()
// by Carsten Clauss, Chair for Operating Systems,
// RWTH Aachen University
//
// [2010-11-12] extracted non-blocking code into separate library
// by Carsten Scholtes
//
// [2010-12-09] added cancel functions for non-blocking send/recv requests
// by Carsten Clauss
//
#include <metalsvm/stdlib.h>
#include <metalsvm/string.h>
#ifdef CONFIG_ROCKCREEK
#include <asm/iRCCE_lib.h>
static int iRCCE_push_send_request(iRCCE_SEND_REQUEST *request) {
char padline[RCCE_LINE_SIZE]; // copy buffer, used if message not multiple of line size
int test; // flag for calling iRCCE_test_flag()
if(request->finished) return(iRCCE_SUCCESS);
if(request->label == 1) goto label1;
if(request->label == 2) goto label2;
if(request->label == 3) goto label3;
// send data in units of available chunk size of comm buffer
for (; request->wsize< (request->size / request->chunk) * request->chunk; request->wsize += request->chunk) {
request->bufptr = request->privbuf + request->wsize;
request->nbytes = request->chunk;
// copy private data to own comm buffer
iRCCE_put(request->combuf, (t_vcharp) request->bufptr, request->nbytes, RCCE_IAM);
RCCE_flag_write(request->sent, RCCE_FLAG_SET, request->dest);
// wait for the destination to be ready to receive a message
label1:
iRCCE_test_flag(*(request->ready), RCCE_FLAG_SET, &test);
if(!test) {
request->label = 1;
return(iRCCE_PENDING);
}
RCCE_flag_write(request->ready, RCCE_FLAG_UNSET, RCCE_IAM);
}
request->remainder = request->size % request->chunk;
// if nothing is left over, we are done
if (!request->remainder) {
request->finished = 1;
return(iRCCE_SUCCESS);
}
// send remainder of data--whole cache lines
request->bufptr = request->privbuf + (request->size / request->chunk) * request->chunk;
request->nbytes = request->remainder - request->remainder % RCCE_LINE_SIZE;
if (request->nbytes) {
// copy private data to own comm buffer
iRCCE_put(request->combuf, (t_vcharp)request->bufptr, request->nbytes, RCCE_IAM);
RCCE_flag_write(request->sent, RCCE_FLAG_SET, request->dest);
// wait for the destination to be ready to receive a message
label2:
iRCCE_test_flag(*(request->ready), RCCE_FLAG_SET, &test);
if(!test) {
request->label = 2;
return(iRCCE_PENDING);
}
RCCE_flag_write(request->ready, RCCE_FLAG_UNSET, RCCE_IAM);
}
request->remainder = request->size % request->chunk;
request->remainder = request->remainder%RCCE_LINE_SIZE;
// if nothing is left over, we are done
if (!request->remainder)
{
request->finished = 1;
return(iRCCE_SUCCESS);
}
// remainder is less than a cache line. This must be copied into appropriately sized
// intermediate space before it can be sent to the receiver
request->bufptr = request->privbuf + (request->size / request->chunk) * request->chunk + request->nbytes;
request->nbytes = RCCE_LINE_SIZE;
// copy private data to own comm buffer
memcpy(padline,request->bufptr,request->remainder);
iRCCE_put(request->combuf, (t_vcharp)padline, request->nbytes, RCCE_IAM);
RCCE_flag_write(request->sent, RCCE_FLAG_SET, request->dest);
// wait for the destination to be ready to receive a message
label3:
iRCCE_test_flag(*(request->ready), RCCE_FLAG_SET, &test);
if(!test) {
request->label = 3;
return(iRCCE_PENDING);
}
RCCE_flag_write(request->ready, RCCE_FLAG_UNSET, RCCE_IAM);
request->finished = 1;
return(iRCCE_SUCCESS);
}
static void iRCCE_init_send_request(
char *privbuf, // source buffer in local private memory (send buffer)
t_vcharp combuf, // intermediate buffer in MPB
size_t chunk, // size of MPB available for this message (bytes)
RCCE_FLAG *ready, // flag indicating whether receiver is ready
RCCE_FLAG *sent, // flag indicating whether message has been sent by source
size_t size, // size of message (bytes)
int dest, // UE that will receive the message
iRCCE_SEND_REQUEST *request
) {
request->privbuf = privbuf;
request->combuf = combuf;
request->chunk = chunk;
request->ready = ready;
request->sent = sent;
request->size = size;
request->dest = dest;
request->wsize = 0;
request->remainder = 0;
request->nbytes = 0;
request->bufptr = NULL;
request->label = 0;
request->finished = 0;
request->next = NULL;
return;
}
//--------------------------------------------------------------------------------------
// FUNCTION: iRCCE_isend
//--------------------------------------------------------------------------------------
// non-blocking send function; returns a handle of type iRCCE_SEND_REQUEST
//--------------------------------------------------------------------------------------
static iRCCE_SEND_REQUEST blocking_isend_request;
int iRCCE_isend(char *privbuf, size_t size, int dest, iRCCE_SEND_REQUEST *request) {
if(request == NULL) request = &blocking_isend_request;
if (dest<0 || dest >= RCCE_NP)
return(RCCE_error_return(RCCE_debug_comm,RCCE_ERROR_ID));
else {
iRCCE_init_send_request(privbuf, RCCE_buff_ptr, RCCE_chunk,
&RCCE_ready_flag[dest], &RCCE_sent_flag[RCCE_IAM],
size, dest, request);
if(iRCCE_isend_queue == NULL) {
if(iRCCE_push_send_request(request) == iRCCE_SUCCESS) {
return(iRCCE_SUCCESS);
}
else {
iRCCE_isend_queue = request;
if(request == &blocking_isend_request) {
iRCCE_isend_wait(request);
return(iRCCE_SUCCESS);
}
return(iRCCE_PENDING);
}
}
else {
if(iRCCE_isend_queue->next == NULL) {
iRCCE_isend_queue->next = request;
}
else {
iRCCE_SEND_REQUEST *run = iRCCE_isend_queue;
while(run->next != NULL) run = run->next;
run->next = request;
}
if(request == &blocking_isend_request) {
iRCCE_isend_wait(request);
return(iRCCE_SUCCESS);
}
return(iRCCE_RESERVED);
}
}
}
//--------------------------------------------------------------------------------------
// FUNCTION: iRCCE_isend_test
//--------------------------------------------------------------------------------------
// test function for completion of the requestes non-blocking send operation
// Just provide NULL instead of testvar if you don't need it
//--------------------------------------------------------------------------------------
int iRCCE_isend_test(iRCCE_SEND_REQUEST *request, int *test) {
if(request == NULL) {
iRCCE_isend_push();
if(iRCCE_isend_queue == NULL) {
if (test) (*test) = 1;
return(iRCCE_SUCCESS);
}
else {
if (test) (*test) = 0;
return(iRCCE_PENDING);
}
}
if(request->finished) {
if (test) (*test) = 1;
return(iRCCE_SUCCESS);
}
if(iRCCE_isend_queue != request) {
if (test) (*test) = 0;
return(iRCCE_RESERVED);
}
iRCCE_push_send_request(request);
if(request->finished) {
iRCCE_isend_queue = request->next;
if (test) (*test) = 1;
return(iRCCE_SUCCESS);
}
if (test) (*test) = 0;
return(iRCCE_PENDING);
}
//--------------------------------------------------------------------------------------
// FUNCTION: iRCCE_isend_push
//--------------------------------------------------------------------------------------
// progress function for pending requests in the isend queue
//--------------------------------------------------------------------------------------
int iRCCE_isend_push(void) {
iRCCE_SEND_REQUEST *request = iRCCE_isend_queue;
if(request == NULL) {
return(iRCCE_SUCCESS);
}
if(request->finished) {
return(iRCCE_SUCCESS);
}
iRCCE_push_send_request(request);
if(request->finished) {
iRCCE_isend_queue = request->next;
return(iRCCE_SUCCESS);
}
return(iRCCE_PENDING);
}
//--------------------------------------------------------------------------------------
// FUNCTION: iRCCE_isend_wait
//--------------------------------------------------------------------------------------
// just wait for completion of the requestes non-blocking send operation
//--------------------------------------------------------------------------------------
int iRCCE_isend_wait(iRCCE_SEND_REQUEST *request) {
if(request != NULL) {
while(!request->finished) {
iRCCE_isend_push();
iRCCE_irecv_push();
}
}
else {
while(iRCCE_isend_queue != NULL) {
iRCCE_isend_push();
iRCCE_irecv_push();
}
}
return(iRCCE_SUCCESS);
}
//--------------------------------------------------------------------------------------
// FUNCTION: iRCCE_isend_cancel
//--------------------------------------------------------------------------------------
// try to cancel a pending non-blocking send request
//--------------------------------------------------------------------------------------
int iRCCE_isend_cancel(iRCCE_SEND_REQUEST *request, int *test) {
iRCCE_SEND_REQUEST *run;
if( (request == NULL) || (request->finished) ) {
if (test) (*test) = 0;
return iRCCE_NOT_ENQUEUED;
}
if(iRCCE_isend_queue == NULL) {
if (test) (*test) = 0;
return iRCCE_NOT_ENQUEUED;
}
if(iRCCE_isend_queue == request) {
if (test) (*test) = 0;
return iRCCE_PENDING;
}
for(run = iRCCE_isend_queue; run->next != NULL; run = run->next) {
// request found --> remove it from send queue:
if(run->next == request) {
run->next = run->next->next;
if (test) (*test) = 1;
return iRCCE_SUCCESS;
}
}
if (test) (*test) = 0;
return iRCCE_NOT_ENQUEUED;
}
#endif

87
arch/x86/scc/iRCCE_put.c Normal file
View file

@ -0,0 +1,87 @@
//***************************************************************************************
// Put data into communication buffer.
//***************************************************************************************
//
// Author: Rob F. Van der Wijngaart
// Intel Corporation
// Date: 008/30/2010
//
//***************************************************************************************
//
// Copyright 2010 Intel Corporation
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
//
// [2010-11-03] switched to SCC-optimized memcpy() functions in scc_memcpy.h:
// - memcpy_to_mpb()
// - memcpy_from_mpb()
// by Stefan Lankes, Carsten Clauss, Chair for Operating Systems,
// RWTH Aachen University
//
#include <metalsvm/stdlib.h>
#include <metalsvm/string.h>
#ifdef CONFIG_ROCKCREEK
#include <asm/iRCCE_lib.h>
#ifdef COPPERRIDGE
#include "scc_memcpy.h"
#endif
void* iRCCE_memcpy_put(void *dest, const void *src, size_t count)
{
#ifdef COPPERRIDGE
return memcpy_to_mpb(dest, src, count);
#else
return memcpy(dest, src, count);
#endif
}
//--------------------------------------------------------------------------------------
// FUNCTION: iRCCE_put
//--------------------------------------------------------------------------------------
// copy data from address "source" in the local MPB or the calling UE's private memory
// to address "target" in the remote MPB. We do not test to see if a move from the
// calling UE's private memory stays within allocated memory
//--------------------------------------------------------------------------------------
int iRCCE_put(
t_vcharp target, // target buffer, MPB
t_vcharp source, // source buffer, MPB or private memory
int num_bytes,
int ID
) {
// in non-GORY mode we only need to retain the MPB target shift; we
// already know the target is in the MPB, not private memory
target = RCCE_comm_buffer[ID]+(target-RCCE_comm_buffer[RCCE_IAM]);
// make sure that any data that has been put in our MPB by another UE is visible
#ifdef _OPENMP
#pragma omp flush
#endif
// do the actual copy
RC_cache_invalidate();
iRCCE_memcpy_put((void *)target, (void *)source, num_bytes);
// flush data to make it visible to all threads; cannot use flush list because it
// concerns malloced space
#ifdef _OPENMP
#pragma omp flush
#endif
return(iRCCE_SUCCESS);
}
#endif

190
arch/x86/scc/iRCCE_recv.c Normal file
View file

@ -0,0 +1,190 @@
//***************************************************************************************
// Non-blocking receive routines.
//***************************************************************************************
//
// Author: Rob F. Van der Wijngaart
// Intel Corporation
// Date: 008/30/2010
//
//***************************************************************************************
//
// Copyright 2010 Intel Corporation
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
//
// [2010-11-26] added a _pipelined_ version of blocking send/recv
// by Carsten Clauss, Chair for Operating Systems,
// RWTH Aachen University
//
#include <metalsvm/stdlib.h>
#include <metalsvm/string.h>
#ifdef CONFIG_ROCKCREEK
#include <asm/iRCCE_lib.h>
//--------------------------------------------------------------------------------------
// FUNCTION: iRCCE_recv_general
//--------------------------------------------------------------------------------------
// pipelined receive function
//--------------------------------------------------------------------------------------
static int iRCCE_recv_general(
char *privbuf, // destination buffer in local private memory (receive buffer)
t_vcharp combuf, // intermediate buffer in MPB
size_t chunk, // size of MPB available for this message (bytes)
RCCE_FLAG *ready, // flag indicating whether receiver is ready
RCCE_FLAG *sent, // flag indicating whether message has been sent by source
size_t size, // size of message (bytes)
int source, // UE that sent the message
int *test // if 1 upon entry, do nonblocking receive; if message available
// set to 1, otherwise to 0
) {
char padline[RCCE_LINE_SIZE]; // copy buffer, used if message not multiple of line size
size_t wsize, // offset within receive buffer when pulling in "chunk" bytes
remainder, // bytes remaining to be received
nbytes; // number of bytes to be received in single iRCCE_get call
int first_test; // only use first chunk to determine if message has been received yet
char *bufptr; // running pointer inside privbuf for current location
first_test = 1;
#if 0
// receive data in units of available chunk size of MPB
for (wsize=0; wsize< (size/chunk)*chunk; wsize+=chunk) {
bufptr = privbuf + wsize;
nbytes = chunk;
// if function is called in test mode, check if first chunk has been sent already.
// If so, proceed as usual. If not, exit immediately
if (*test && first_test) {
first_test = 0;
if (!(*test = RCCE_probe(*sent))) return(iRCCE_SUCCESS);
}
RCCE_wait_until(*sent, RCCE_FLAG_SET);
RCCE_flag_write(sent, RCCE_FLAG_UNSET, RCCE_IAM);
// copy data from local MPB space to private memory
iRCCE_get((t_vcharp)bufptr, combuf, nbytes, source);
// tell the source I have moved data out of its comm buffer
RCCE_flag_write(ready, RCCE_FLAG_SET, source);
}
#else
{ // pipelined version of send/recv:
size_t subchunk1 = chunk / 2;
size_t subchunk2 = chunk - subchunk1;
for (wsize=0; wsize < (size/chunk)*chunk; wsize+=chunk) {
if (*test && first_test) {
first_test = 0;
if (!(*test = RCCE_probe(*sent))) return(iRCCE_SUCCESS);
}
bufptr = privbuf + wsize;
nbytes = subchunk1;
RCCE_wait_until(*ready, RCCE_FLAG_SET);
RCCE_flag_write(ready, RCCE_FLAG_UNSET, RCCE_IAM);
iRCCE_get((t_vcharp)bufptr, combuf, nbytes, source);
RCCE_flag_write(ready, RCCE_FLAG_SET, source);
bufptr = privbuf + wsize + subchunk1;
nbytes = subchunk2;
RCCE_wait_until(*sent, RCCE_FLAG_SET);
RCCE_flag_write(sent, RCCE_FLAG_UNSET, RCCE_IAM);
iRCCE_get((t_vcharp)bufptr, combuf + subchunk1, nbytes, source);
RCCE_flag_write(sent, RCCE_FLAG_SET, source);
}
}
#endif
remainder = size%chunk;
// if nothing is left over, we are done
if (!remainder) return(iRCCE_SUCCESS);
// receive remainder of data--whole cache lines
bufptr = privbuf + (size/chunk)*chunk;
nbytes = remainder - remainder % RCCE_LINE_SIZE;
if (nbytes) {
// if function is called in test mode, check if first chunk has been sent already.
// If so, proceed as usual. If not, exit immediately
if (*test && first_test) {
first_test = 0;
if (!(*test = RCCE_probe(*sent))) return(iRCCE_SUCCESS);
}
RCCE_wait_until(*sent, RCCE_FLAG_SET);
RCCE_flag_write(sent, RCCE_FLAG_UNSET, RCCE_IAM);
// copy data from local MPB space to private memory
iRCCE_get((t_vcharp)bufptr, combuf, nbytes, source);
// tell the source I have moved data out of its comm buffer
RCCE_flag_write(ready, RCCE_FLAG_SET, source);
}
remainder = remainder % RCCE_LINE_SIZE;
if (!remainder) return(iRCCE_SUCCESS);
// remainder is less than cache line. This must be copied into appropriately sized
// intermediate space before exact number of bytes get copied to the final destination
bufptr = privbuf + (size/chunk)*chunk + nbytes;
nbytes = RCCE_LINE_SIZE;
// if function is called in test mode, check if first chunk has been sent already.
// If so, proceed as usual. If not, exit immediately
if (*test && first_test) {
first_test = 0;
if (!(*test = RCCE_probe(*sent))) return(iRCCE_SUCCESS);
}
RCCE_wait_until(*sent, RCCE_FLAG_SET);
RCCE_flag_write(sent, RCCE_FLAG_UNSET, RCCE_IAM);
// copy data from local MPB space to private memory
iRCCE_get((t_vcharp)padline, combuf, nbytes, source);
memcpy(bufptr,padline,remainder);
// tell the source I have moved data out of its comm buffer
RCCE_flag_write(ready, RCCE_FLAG_SET, source);
return(iRCCE_SUCCESS);
}
//--------------------------------------------------------------------------------------
// FUNCTION: iRCCE_recv
//--------------------------------------------------------------------------------------
// pipelined recv function (blocking!)
//--------------------------------------------------------------------------------------
int iRCCE_recv(char *privbuf, size_t size, int source) {
int ignore;
while(iRCCE_irecv_queue[source] != NULL) {
// wait for completion of pending non-blocking requests
iRCCE_irecv_push();
iRCCE_isend_push();
}
if (source<0 || source >= RCCE_NP)
return(RCCE_error_return(RCCE_debug_comm,RCCE_ERROR_ID));
else {
ignore = 0;
return(iRCCE_recv_general(privbuf, RCCE_buff_ptr, RCCE_chunk,
&RCCE_ready_flag[RCCE_IAM], &RCCE_sent_flag[source],
size, source, &ignore));
}
}
#endif

165
arch/x86/scc/iRCCE_send.c Normal file
View file

@ -0,0 +1,165 @@
//***************************************************************************************
// Synchronized receive routines.
//***************************************************************************************
//
// Author: Rob F. Van der Wijngaart
// Intel Corporation
// Date: 008/30/2010
//
//***************************************************************************************
//
// Copyright 2010 Intel Corporation
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
//
// [2010-11-26] added a _pipelined_ version of blocking send/recv
// by Carsten Clauss, Chair for Operating Systems,
// RWTH Aachen University
//
#include <metalsvm/stdlib.h>
#include <metalsvm/string.h>
#ifdef CONFIG_ROCKCREEK
#include <asm/iRCCE_lib.h>
//--------------------------------------------------------------------------------------
// FUNCTION: iRCCE_send_general
//--------------------------------------------------------------------------------------
// pipelined send function
//--------------------------------------------------------------------------------------
static int iRCCE_send_general(
char *privbuf, // source buffer in local private memory (send buffer)
t_vcharp combuf, // intermediate buffer in MPB
size_t chunk, // size of MPB available for this message (bytes)
RCCE_FLAG *ready, // flag indicating whether receiver is ready
RCCE_FLAG *sent, // flag indicating whether message has been sent by source
size_t size, // size of message (bytes)
int dest // UE that will receive the message
) {
char padline[RCCE_LINE_SIZE]; // copy buffer, used if message not multiple of line size
size_t wsize, // offset within send buffer when putting in "chunk" bytes
remainder, // bytes remaining to be sent
nbytes; // number of bytes to be sent in single iRCCE_put call
char *bufptr; // running pointer inside privbuf for current location
#if 0
// send data in units of available chunk size of comm buffer
for (wsize=0; wsize< (size/chunk)*chunk; wsize+=chunk) {
bufptr = privbuf + wsize;
nbytes = chunk;
// copy private data to own comm buffer
iRCCE_put(combuf, (t_vcharp) bufptr, nbytes, RCCE_IAM);
RCCE_flag_write(sent, RCCE_FLAG_SET, dest);
// wait for the destination to be ready to receive a message
RCCE_wait_until(*ready, RCCE_FLAG_SET);
RCCE_flag_write(ready, RCCE_FLAG_UNSET, RCCE_IAM);
}
#else
{ // pipelined version of send/recv:
size_t subchunk1 = chunk / 2;
size_t subchunk2 = chunk - subchunk1;
wsize = 0;
for (; wsize < (size/chunk)*chunk; wsize+=chunk) {
bufptr = privbuf + wsize;
nbytes = subchunk1;
iRCCE_put(combuf, (t_vcharp) bufptr, nbytes, RCCE_IAM);
RCCE_flag_write(ready, RCCE_FLAG_SET, dest);
if(wsize>0)
{
RCCE_wait_until(*sent, RCCE_FLAG_SET);
RCCE_flag_write(sent, RCCE_FLAG_UNSET, RCCE_IAM);
}
bufptr = privbuf + wsize + subchunk1;
nbytes = subchunk2;
iRCCE_put(combuf + subchunk1, (t_vcharp) bufptr, nbytes, RCCE_IAM);
RCCE_flag_write(sent, RCCE_FLAG_SET, dest);
RCCE_wait_until(*ready, RCCE_FLAG_SET);
RCCE_flag_write(ready, RCCE_FLAG_UNSET, RCCE_IAM);
}
if(wsize>0) {
RCCE_wait_until(*sent, RCCE_FLAG_SET);
RCCE_flag_write(sent, RCCE_FLAG_UNSET, RCCE_IAM);
}
}
#endif
remainder = size%chunk;
// if nothing is left over, we are done
if (!remainder) return(iRCCE_SUCCESS);
// send remainder of data--whole cache lines
bufptr = privbuf + (size/chunk)*chunk;
nbytes = remainder - remainder%RCCE_LINE_SIZE;
if (nbytes) {
// copy private data to own comm buffer
iRCCE_put(combuf, (t_vcharp)bufptr, nbytes, RCCE_IAM);
RCCE_flag_write(sent, RCCE_FLAG_SET, dest);
// wait for the destination to be ready to receive a message
RCCE_wait_until(*ready, RCCE_FLAG_SET);
RCCE_flag_write(ready, RCCE_FLAG_UNSET, RCCE_IAM);
}
remainder = remainder%RCCE_LINE_SIZE;
if (!remainder) return(iRCCE_SUCCESS);
// remainder is less than a cache line. This must be copied into appropriately sized
// intermediate space before it can be sent to the receiver
bufptr = privbuf + (size/chunk)*chunk + nbytes;
nbytes = RCCE_LINE_SIZE;
// copy private data to own comm buffer
memcpy(padline,bufptr,remainder);
iRCCE_put(combuf, (t_vcharp)padline, nbytes, RCCE_IAM);
RCCE_flag_write(sent, RCCE_FLAG_SET, dest);
// wait for the destination to be ready to receive a message
RCCE_wait_until(*ready, RCCE_FLAG_SET);
RCCE_flag_write(ready, RCCE_FLAG_UNSET, RCCE_IAM);
return(iRCCE_SUCCESS);
}
//--------------------------------------------------------------------------------------
// FUNCTION: iRCCE_send
//--------------------------------------------------------------------------------------
// pipelined send function (blocking!)
//--------------------------------------------------------------------------------------
int iRCCE_send(char *privbuf, size_t size, int dest) {
while(iRCCE_isend_queue != NULL) {
// wait for completion of pending non-blocking requests
iRCCE_isend_push();
iRCCE_irecv_push();
}
if (dest<0 || dest >= RCCE_NP)
return(RCCE_error_return(RCCE_debug_comm,RCCE_ERROR_ID));
else
return(iRCCE_send_general(privbuf, RCCE_buff_ptr, RCCE_chunk,
&RCCE_ready_flag[dest], &RCCE_sent_flag[RCCE_IAM],
size, dest));
}
#endif

127
arch/x86/scc/iRCCE_synch.c Normal file
View file

@ -0,0 +1,127 @@
///*************************************************************************************
// Synchronization functions.
// Single-bit and whole-cache-line flags are sufficiently different that we provide
// separate implementations of the synchronization routines for each case
//**************************************************************************************
//
// Author: Rob F. Van der Wijngaart
// Intel Corporation
// Date: 008/30/2010
//
//**************************************************************************************
//
// Copyright 2010 Intel Corporation
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
//
// [2010-10-25] added support for non-blocking send/recv operations
// - iRCCE_isend(), ..._test(), ..._wait(), ..._push()
// - iRCCE_irecv(), ..._test(), ..._wait(), ..._push()
// by Carsten Clauss, Chair for Operating Systems,
// RWTH Aachen University
//
// [2010-11-12] extracted non-blocking code into separate library
// by Carsten Scholtes
//
// [2011-01-21] updated the datatype of RCCE_FLAG according to the
// recent version of RCCE
//
// [2011-04-12] added marco test for rcce version
//
#include <metalsvm/stdlib.h>
#include <metalsvm/string.h>
#ifdef CONFIG_ROCKCREEK
#include <asm/iRCCE_lib.h>
#ifdef SINGLEBITFLAGS
int iRCCE_test_flag(RCCE_FLAG flag, RCCE_FLAG_STATUS val, int *result) {
t_vcharp cflag;
#ifdef RCCE_VERSION
// this is a newer version than V1.0.13
t_vcharp flaga;
#endif
cflag = flag.line_address;
#ifdef RCCE_VERSION
// this is a newer version than V1.0.13
flaga = flag.flag_addr;
#endif
// always flush/invalidate to ensure we read the most recent value of *flag
// keep reading it until it has the required value
#ifdef _OPENMP
#pragma omp flush
#endif
RC_cache_invalidate();
#ifdef RCCE_VERSION
// this is a newer version than V1.0.13
if(RCCE_bit_value(flaga, (flag.location)%RCCE_FLAGS_PER_BYTE) != val) {
#else
if(RCCE_bit_value(cflag, flag.location) != val) {
#endif
(*result) = 0;
}
else {
(*result) = 1;
}
return(iRCCE_SUCCESS);
}
#else
//////////////////////////////////////////////////////////////////
// LOCKLESS SYNCHRONIZATION USING ONE WHOLE CACHE LINE PER FLAG //
//////////////////////////////////////////////////////////////////
int iRCCE_test_flag(RCCE_FLAG flag, RCCE_FLAG_STATUS val, int *result) {
#ifdef RCCE_VERSION
// this is a newer version than V1.0.13
t_vcharp flaga = flag.flag_addr;
#endif
// always flush/invalidate to ensure we read the most recent value of *flag
// keep reading it until it has the required value. We only need to read the
// first int of the MPB cache line containing the flag
#ifdef _OPENMP
#pragma omp flush
#endif
RC_cache_invalidate();
#ifdef RCCE_VERSION
// this is a newer version than V1.0.13
if((RCCE_FLAG_STATUS)(*flaga) != val) {
#else
if((*flag) != val) {
#endif
(*result) = 0;
}
else {
(*result) = 1;
}
return(iRCCE_SUCCESS);
}
#endif
#endif

54
arch/x86/scc/scc_init.c
View file

@ -17,7 +17,8 @@
#include <metalsvm/processor.h>
#include <metalsvm/errno.h>
#include <asm/io.h>
#include <asm/RCCE_lib.h>
#include <asm/RCCE.h>
#include <asm/iRCCE.h>
#include <asm/SCC_API.h>
#ifdef CONFIG_ROCKCREEK
@ -35,6 +36,9 @@ bootinfo_t* bootinfo = (bootinfo_t*) SCC_BOOTINFO;
static int scc_clear(void)
{
int tmp, x, y, z, offset;
// Initialize API
InitAPI(0);
// Find out who I am...
tmp=ReadConfigReg(CRB_OWN+MYTILEID);
@ -57,6 +61,9 @@ static int scc_clear(void)
// Clear test&set register write. Next read-access will read "1" (lock granted).
SetConfigReg(CRB_ADDR(x,y)+((z)?LOCK1:LOCK0), 1);
// frees Message Passing Buffer
MPBunalloc(&MPB);
return 0;
}
@ -64,31 +71,60 @@ int scc_init(void)
{
int num_ranks;
int i, my_rank;
uint64_t start, end, ticks, freq = 533;
uint32_t cr4, msg = 0;
kputs("Initialize Rock Creek!\n");
/* Enable Messagepassing in CR4 */
cr4 = read_cr4();
cr4 = cr4 | _CR4_MPE;
write_cr4(cr4);
kprintf("Initialize Rock Creek!\n");
kprintf("address of the initrd: 0x%x\n", bootinfo->addr);
kprintf("size of the initrd: %d\n", bootinfo->size);
kprintf("rcce argc = %d\n", bootinfo->argc);
for(i=0; i<bootinfo->argc; i++)
kprintf("rcce argv[%d] = %s\n", i, bootinfo->argv[i]);
if (bootinfo->argc >= 3)
freq = atoi(bootinfo->argv[2]);
kputs("Reset SCC!\n");
scc_clear();
kputs("Wait some time...\n");
mb();
start = rdtsc();
do {
mb();
end = rdtsc();
ticks = end > start ? end - start : start - end;
} while(ticks*TIMER_FREQ < 300ULL*freq*1000000ULL);
if (RCCE_init(&bootinfo->argc, &bootinfo->argv) != RCCE_SUCCESS)
return -ENODEV;
if (iRCCE_init() != iRCCE_SUCCESS)
return -ENODEV;
// enable additional outputs
RCCE_debug_set(RCCE_DEBUG_ALL);
my_rank = RCCE_ue();
num_ranks = RCCE_num_ues();
kprintf("Got rank %d of %d ranks\n", my_rank, num_ranks);
/* Enable Messagepassing in CR4 */
uint32_t cr4 = read_cr4();
cr4 = cr4 | _CR4_MPE;
write_cr4(cr4);
i = ReadConfigReg(CRB_OWN+GLCFG0);
kprintf("glcfg0 0x%x\n", i);
/* synchronize before starting MetalSVM: */
//RCCE_barrier(&RCCE_COMM_WORLD);
RCCE_barrier(&RCCE_COMM_WORLD);
kputs("RCCE test...\t");
if (my_rank == 0)
msg = 0x4711;
if (RCCE_bcast((char*) &msg, sizeof(msg), 0, RCCE_COMM_WORLD) == RCCE_SUCCESS)
kprintf("successfull! (0x%x)\n", msg);
else
kprintf("failed! (0x%x)\n", msg);
kputs("Now, the SCC is initialized!\n");

2
include/metalsvm/config.h.example
View file

@ -58,7 +58,7 @@ extern "C" {
// RCCE specific flags
#define SCC
#define MS_BAREMETAL
#define GORY
//#define GORY
//#define SHMADD
//#define SHMADD_CACHEABLE
/* default values for 16 GB system */

6
libkern/string.c
View file

@ -8,9 +8,9 @@
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
*/
#include <metalsvm/string.h>
#include <metalsvm/stddef.h>
#ifndef HAVE_ARCH_MEMCPY
//#ifndef HAVE_ARCH_MEMCPY
void *memcpy(void *dest, const void *src, size_t count)
{
size_t i;
@ -23,7 +23,7 @@ void *memcpy(void *dest, const void *src, size_t count)
return dest;
}
#endif
//#endif
#ifndef HAVE_ARCH_MEMSET
void *memset(void *dest, int val, size_t count)

9
tools/Makefile
View file

@ -41,16 +41,19 @@ scc_bootinfo.asm: bootinfo.sh
scc_bootinfo.bin: scc_bootinfo.asm
$(NASM) $(NASMFLAGS) -o $@ $<
SCC: scc_bootinfo.bin scc_setup.bin reset_vector.bin initrd.img
bin2obj: bin2obj.c
$(CC) $(CFLAGS) -o $@ $<
SCC: scc_bootinfo.bin scc_setup.bin reset_vector.bin initrd.img bin2obj
cp ../metalsvm.elf .
$(CROSS_OBJCOPY) -j .mboot -j .text -j .data -j .rodata -j .bss -O binary metalsvm.elf metalsvm.bin
chmod a-x *.bin
. ./prepare.sh
/home/lankes/tools/bin2obj -m load.map -o metalsvm.obj
./bin2obj -m load.map -o metalsvm.obj
sccMerge -noimage -m 8 -n 12 -force ./metalsvm.mt
clean:
$(RM) -rf *.o *~ make_initrd initrd.img *.bin *.obj *.hex *.elf obj
$(RM) -rf *.o *~ bin2obj make_initrd initrd.img *.bin *.obj *.hex *.elf obj
depend:
$(CC) -MM $(CFLAGS) *.c > Makefile.dep

147
tools/bin2obj.c Normal file
View file

@ -0,0 +1,147 @@
#include <stdio.h>
#include <getopt.h>
#include <string.h>
const char BIN2OBJIDSTRING[] = "$Id: bin2obj.c 8016 2007-11-01 14:24:42Z tlehnig $";
long long convertToHex(char *fn, unsigned long origin, FILE *outfile) {
FILE *datafile;
unsigned char data1, data2, data3, data4;
int res = 0;
long long count = 0;
datafile = fopen(fn, "r");
if (!datafile) {
printf("Datafile >%s< could not be opened, not writing data for this file\n", fn);
return -1;
}
printf("Converting file >%s< to .32.obj format at origin 0x%08lx (0x%08lx) ... ",
fn, origin >> 2, origin);
fprintf(outfile, "/origin %08lx\n", origin >> 2);
do {
data1 = 0;
data2 = 0;
data3 = 0;
data4 = 0;
res = fscanf(datafile, "%c%c%c%c", &data1, &data2, &data3, &data4);
if (res > 0) {
count += res;
fprintf(outfile, "%02x%02x%02x%02x", data4, data3, data2, data1);
if ((count % 16) == 0)
fprintf(outfile, "\n");
else
fprintf(outfile, " ");
}
} while (res > 0);
if ((count % 16) != 0) fprintf(outfile, "\n");
printf("done with %lli Bytes.\n", count);
fclose(datafile);
return count;
}
void print_help() {
printf("Usage: bin2obj [FLAGS] [OPTIONS]\n");
printf("\nFLAGS: -h, -v\n");
printf("-h Print this help\n");
printf("-v Print Version ID\n");
printf("\nOPTIONS: -m, -o\n");
printf("-m <mapfile> Defines mapfile to use for bin2obj\n");
printf("-o <outfile> Defines output file to use for bin2obj\n");
printf("\nbin2obj converts the binary files defined in the mapfile to a hex based textfile\n");
printf("used by MCEMU\n");
}
int main(int argc, char **argv) {
FILE *mapfile = NULL, *outfile = NULL;
unsigned long origin;
char datafn[255];
char outfn[255] = "output.obj";
char mapfn[255] = "load.map";
int res = 0;
unsigned long long count = 0;
long long thiscount = 0;
int retval = 0;
int c, doOptLoop = 1;
while (doOptLoop) {
c = getopt(argc, argv, "m:o:hv");
if (c == -1) {
doOptLoop = 0;
break;
}
switch (c) {
case 'h':
print_help();
return 0;
break;
case 'v':
printf("%s %s\n", argv[0], BIN2OBJIDSTRING);
return 0;
break;
case 'm':
printf("Mapfile: >%s<\n", optarg);
strncpy(mapfn, optarg, 255);
break;
case 'o':
printf("Outfile: >%s<\n", optarg);
strncpy(outfn, optarg, 255);
break;
default:
print_help();
return 0;
}
}
mapfile = fopen(mapfn, "r");
if (!mapfile) {
printf("Mapfile >%s< not found, exiting.\n", mapfn);
return -1;
}
outfile = fopen(outfn, "w");
if (!outfile) {
printf("Outputfile >%s< could not be created, exiting\n", outfn);
return -1;
}
// res = fscanf(mapfile, "%lx %s\n", &origin, datafn);
while ((res = fscanf(mapfile, "%lx %s\n", &origin, datafn)) == 2) {
//printf("ReadMapFile origin: 0x%08lx, filename: >%s<\n", origin, datafn);
thiscount = convertToHex(datafn, origin, outfile);
if (thiscount < 0) {
retval = -1;
}
else count += thiscount;
}
fprintf(outfile, "/eof\n");
fclose(mapfile);
fclose (outfile);
printf("Total conversion: %lli Bytes\n", count);
return retval;
}

1
tools/metalsvm.mt
View file

@ -1,2 +1,3 @@
# pid mch-route mch-dest-id mch-offset-base testcase
0x00 0x00 6 0x00 metalsvm.obj
0x01 0x00 6 0x01 metalsvm.obj