2e83cf15ce
PSM2 enables support for GPU buffers and CUDA managed memory and it can directly recognize GPU buffers, handle copies between HFIs and GPUs. Therefore, it is not required for OMPI to handle GPU buffers for pt2pt cases. In this patch, we allow the PSM2 MTL to specify when it does not require CUDA convertor support. This allows us to skip CUDA convertor init phases and lets PSM2 handle the memory transfers. This translates to improvements in latency. The patch enables blocking collectives and workloads with GPU contiguous, GPU non-contiguous memory. Signed-off-by: Aravind Gopalakrishnan <Aravind.Gopalakrishnan@intel.com>
74 строки
2.8 KiB
C
74 строки
2.8 KiB
C
/*
|
|
* Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
|
|
* University Research and Technology
|
|
* Corporation. All rights reserved.
|
|
* Copyright (c) 2004-2016 The University of Tennessee and The University
|
|
* of Tennessee Research Foundation. All rights
|
|
* reserved.
|
|
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
|
|
* University of Stuttgart. All rights reserved.
|
|
* Copyright (c) 2004-2006 The Regents of the University of California.
|
|
* All rights reserved.
|
|
* Copyright (c) 2017 Intel, Inc. All rights reserved
|
|
* $COPYRIGHT$
|
|
*
|
|
* Additional copyrights may follow
|
|
*
|
|
* $HEADER$
|
|
*/
|
|
|
|
#ifndef PML_CM_REQUEST_H
|
|
#define PML_CM_REQUEST_H
|
|
|
|
#include "ompi/mca/pml/base/pml_base_sendreq.h"
|
|
#include "ompi/mca/pml/base/pml_base_bsend.h"
|
|
#include "ompi/mca/pml/pml.h"
|
|
#include "ompi/mca/mtl/mtl.h"
|
|
|
|
/**
|
|
* Type of request.
|
|
*/
|
|
typedef enum {
|
|
MCA_PML_CM_REQUEST_SEND_HEAVY,
|
|
MCA_PML_CM_REQUEST_SEND_THIN,
|
|
MCA_PML_CM_REQUEST_RECV_HEAVY,
|
|
MCA_PML_CM_REQUEST_RECV_THIN,
|
|
MCA_PML_CM_REQUEST_NULL
|
|
} mca_pml_cm_request_type_t;
|
|
|
|
/**
|
|
* Base type for PML CM P2P requests
|
|
*/
|
|
struct mca_pml_cm_request_t {
|
|
|
|
/* START: These fields have to match the definition of the mca_pml_base_request_t */
|
|
ompi_request_t req_ompi; /**< base request */
|
|
volatile int32_t req_pml_complete; /**< flag indicating if the pt-2-pt layer is done with this request */
|
|
volatile int32_t req_free_called; /**< flag indicating if the user has freed this request */
|
|
mca_pml_cm_request_type_t req_pml_type;
|
|
struct ompi_communicator_t *req_comm; /**< communicator pointer */
|
|
struct ompi_datatype_t *req_datatype; /**< pointer to data type */
|
|
opal_convertor_t req_convertor; /**< convertor that describes the memory layout */
|
|
/* END: These fields have to match the definition of the mca_pml_base_request_t */
|
|
};
|
|
typedef struct mca_pml_cm_request_t mca_pml_cm_request_t;
|
|
OBJ_CLASS_DECLARATION(mca_pml_cm_request_t);
|
|
|
|
/*
|
|
* Avoid CUDA convertor inits only for contiguous memory and if indicated by
|
|
* the MTL. For non-contiguous memory, do not skip CUDA convertor init phases.
|
|
*/
|
|
#if OPAL_CUDA_SUPPORT
|
|
#define MCA_PML_CM_SWITCH_CUDA_CONVERTOR_OFF(flags, datatype, count) \
|
|
{ \
|
|
if (opal_datatype_is_contiguous_memory_layout(&datatype->super, count) \
|
|
&& (ompi_mtl->mtl_flags & MCA_MTL_BASE_FLAG_CUDA_INIT_DISABLE)) { \
|
|
flags |= CONVERTOR_SKIP_CUDA_INIT; \
|
|
} \
|
|
}
|
|
#else
|
|
#define MCA_PML_CM_SWITCH_CUDA_CONVERTOR_OFF(flags, datatype, count)
|
|
#endif
|
|
|
|
#endif
|