What: I want to extend the BTL interface to add support for atomic operations. The initial cut presented here includes support for the following features:
- Support for atomic memory operations if supported by the hardware. The attached patch includes support for add, and, or, and xor. Adding additional operations should be trivial. - Support for both fetching (btl_fop) and non-fetching (btl_op) semantics. - Support for blocking or non-blocking operation. Admittedly, I have only implemented blocking versions for ugni so I am not sure if the interface is ideal. Feedback on the non-blocking support would be very appreciated. - Support for compare-and-swap operations. For simplicity this interface only supports 64-bit operations. I have not thought about how to extend the interface to support arbitrary sizes. Would it be useful to look at adding support for 32-bit or 128-bit operations? How about other datatypes (float)? Additionally, I added a new prepare function btl_prepare_rdma. The reasons for this new function are: 1) prepare a fragment with no indication of what endpoint will use it, 2) remove the need for a convertor where a convertor is not needed. The function is meant to do what is needed to prepare the region for arbitrary put, get, and atomic operations. Why: To provide optimal support for one-sided operations I need access to the following at a minimum: atomic add, atomic fetch-and-add, and atomic compare-and-swap. This interface was designed with these operations in mind. To give a little hint as to the performance improvements we can realize with atomics/rdma support in osc: Result obtained from osu_put_latency on two nodes of a Cray XE6 using the uGNI btl. # OSU MPI_Put latency Test # Window creation: MPI_Win_allocate # Synchronization: MPI_Win_lock/unlock # Size Latency (us) 0 7.07 1 7.33 2 7.44 4 7.26 8 7.29 16 7.37 32 7.08 64 7.10 128 7.22 256 7.48 512 7.70 Same benchmark, same nodes, atomic/rdma implementation: # OSU MPI_Put latency Test # Window creation: MPI_Win_allocate # Synchronization: MPI_Win_lock/unlock # Size Latency (us) 0 1.28 1 2.40 2 2.41 4 2.41 8 2.40 16 2.45 32 2.47 64 2.52 128 2.53 256 2.58 512 2.65 When: This RFC is intended to start a discussion on what the atomic interface for the BTLs should look like. I have no reservations with completely re-thinking the interface as long as it doesn't 1) add the convertor into my osc critical path, and 2) require allocation of btl fragments for atomic operations. Lets plan on discussing extending the BTLs at the June developer meeting. All I care about is getting the design finalized in time for the 1.9 branch. I attatched a patch with the proposed BTL extension. I am leaving out the ugni implementation of the interface for now. -Nathan Hjelm
diff --git a/ompi/mca/btl/btl.h b/ompi/mca/btl/btl.h
index ba09e4b..2b9321f 100644
--- a/ompi/mca/btl/btl.h
+++ b/ompi/mca/btl/btl.h
@@ -1,4 +1,4 @@
-/* -*- Mode: C; c-basic-offset:4 ; -*- */
+/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
/*
* Copyright (c) 2004-2007 The Trustees of Indiana University and Indiana
* University Research and Technology
@@ -10,7 +10,7 @@
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
- * Copyright (c) 2006-2013 Los Alamos National Security, LLC. All rights
+ * Copyright (c) 2006-2014 Los Alamos National Security, LLC. All rights
* reserved.
* Copyright (c) 2010 Oracle and/or its affiliates. All rights reserved.
* Copyright (c) 2012-2013 NVIDIA Corporation. All rights reserved.
@@ -211,6 +211,14 @@ typedef uint8_t mca_btl_base_tag_t;
*/
#define MCA_BTL_FLAGS_SIGNALED 0x4000
+/* btl can support atomic compare-and-swap */
+#define MCA_BTL_FLAGS_ATOMIC_CSWAP 0x8000
+
+/* btl can support atomic operations. the btl will indicate which operations
+ * are supported in the btl_atomic_flags field */
+#define MCA_BTL_FLAGS_ATOMIC_OPS 0x10000
+
+
/* Default exclusivity levels */
#define MCA_BTL_EXCLUSIVITY_HIGH (64*1024) /* internal loopback */
#define MCA_BTL_EXCLUSIVITY_DEFAULT 1024 /* GM/IB/etc. */
@@ -222,6 +230,34 @@ typedef uint8_t mca_btl_base_tag_t;
#define MCA_BTL_ERROR_FLAGS_ADD_CUDA_IPC 0x4
/**
+ * @short atomic operations that may be used with a btl that supports atomic
+ * operations.
+ */
+enum mca_btl_base_atomic_op_t {
+ /** atomic signed addition */
+ MCA_BTL_BASE_ATOMIC_ADD,
+ /** atomic bit-wise and */
+ MCA_BTL_BASE_ATOMIC_AND,
+ /** atomic bit-wise or */
+ MCA_BTL_BASE_ATOMIC_OR,
+ /** atomic bit-wise exlusive or */
+ MCA_BTL_BASE_ATOMIC_XOR,
+ /** not a valid atomic operation */
+ MCA_BTL_BASE_ATOMIC_MAX,
+};
+
+typedef enum mca_btl_base_atomic_op_t mca_btl_base_atomic_op_t;
+
+/* atomic operations supported by the btl */
+enum mca_btl_base_atomic_flag_t {
+ MCA_BTL_ATOMIC_FLAG_ADD = 0x01,
+ MCA_BTL_ATOMIC_FLAG_AND = 0x02,
+ MCA_BTL_ATOMIC_FLAG_OR = 0x04,
+ MCA_BTL_ATOMIC_FLAG_XOR = 0x08,
+ MCA_BTL_ATOMIC_FLAG_ALL = 0x0f,
+};
+
+/**
* Asynchronous callback function on completion of an operation.
* Completion Semantics: The descriptor can be reused or returned to the
* BTL via mca_btl_base_module_free_fn_t. The operation has been queued to
@@ -661,6 +697,39 @@ typedef struct mca_btl_base_descriptor_t*
(*mca_btl_base_module_prepare_fn_t)(
uint32_t flags
);
+
+/**
+ * Prepare a descriptor for put/get/atomics using the supplied
+ * pointer.
+ *
+ * The descriptor returned can be used in multiple concurrent operations
+ * (send/put/get) unless the BTL has the MCA_BTL_FLAGS_RDMA_MATCHED flag set
+ * in which case a corresponding prepare call must accompany the put/get call
+ * in addition, the address and length that is put/get must match the address
+ * and length which is prepared.
+ *
+ * The order tag value ensures that operations on the
+ * descriptor that is prepared will be ordered w.r.t. a previous
+ * operation on a particular descriptor. Ordering is only guaranteed if
+ * the previous descriptor had its local completion callback function
+ * called and the order tag of that descriptor is only valid upon the local
+ * completion callback function.
+ *
+ * @param btl (IN) BTL module
+ * @param base (IN) Base address of tegion to prepare
+ * @param len (IN) Length of region to prepare
+ * @param flags (IN) BTL descriptor flags
+ *
+ * Note that the base and length may be adjusted to fit the alignment
requirements
+ * of the underlying API. It is acceptable to modify the returned fragment to
+ * use a different base and length as long as the new base and length fall into
+ * the prepared region.
+ *
+ */
+
+typedef struct mca_btl_base_descriptor_t*
(*mca_btl_base_module_prepare_rdma_fn_t)(
+ struct mca_btl_base_module_t* btl, void *base, size_t len, uint32_t flags);
+
/**
* Initiate an asynchronous send.
* Completion Semantics: the descriptor has been queued for a send operation
@@ -796,6 +865,97 @@ typedef void (*mca_btl_base_module_dump_fn_t)(
);
/**
+ * @short perform a 64-bit atomic operation with put semantics
+ *
+ * @param[in] btl BTL module
+ * @param[in] endpoint BTL endpoint
+ * @param[in] rem_ptr Remote pointer value
+ * @param[in] rem_seg Remote segment
+ * @param[in] operand 64-bit operand to the atomic operation
+ * @param[in] op BTL atomic operation to execute
+ * @param[in] cbfunc Function to call when the operation completes (may be
NULL).
+ *
+ * When the operation completes the atomic operation will be visible in the
target
+ * memory. The implementation may be either blocking or non-blocking. If the
+ * operation is complete this functions will return 1. Otherwise the provided
callback
+ * will be called on completion.
+ */
+typedef int (*mca_btl_base_module_aop_fn_t) (struct mca_btl_base_module_t *btl,
+ struct mca_btl_base_endpoint_t
*endpoint,
+ ompi_ptr_t rem_ptr,
+ struct mca_btl_base_segment_t
*rem_seg,
+ int64_t operand,
+ mca_btl_base_atomic_op_t op,
+ mca_btl_base_completion_fn_t
cbfunc,
+ void *cbdata);
+
+/**
+ * @short perform a 64-bit atomic operation with get semantics
+ *
+ * @param[in] btl BTL module
+ * @param[in] endpoint BTL endpoint
+ * @param[in] rem_ptr Remote pointer value
+ * @param[in] rem_seg Remote segment
+ * @param[in] result_ptr Location to store the result in
+ * @param[in] result_seg Local segment for the result
+ * @param[in] operand 64-bit operand to the atomic operation
+ * @param[in] op BTL atomic operation to execute
+ * @param[in] cbfunc Function to call when the operation completes (may be
NULL).
+ *
+ * This function executes an atomic operation with get semantics. When the
operation
+ * is complete the result of the operation is visible in the target memory and
the
+ * previous value is visible in {result_ptr}. {result_ptr} must be a pointer
within
+ * the descriptor {result_des} and {rem_ptr} must be a pointer within
{rem_des}. The
+ * BTL may provide either blocking or non-blocking semantics. If the operation
is
+ * complete upon return this function returns 1. Otherwise the provided
callback
+ * will be called on completion.
+ */
+typedef int (*mca_btl_base_module_afop_fn_t) (struct mca_btl_base_module_t
*btl,
+ struct mca_btl_base_endpoint_t
*endpoint,
+ ompi_ptr_t rem_ptr,
+ struct mca_btl_base_segment_t
*rem_seg,
+ ompi_ptr_t result_ptr,
+ struct mca_btl_base_segment_t
*result_seg,
+ int64_t operand,
+ mca_btl_base_atomic_op_t op,
+ mca_btl_base_completion_fn_t
cbfunc,
+ void *cbdata);
+
+/**
+ * @short perform a 64-bit compare-and-swap operation
+ *
+ * @param[in] btl BTL module
+ * @param[in] endpoint BTL endpoint
+ * @param[in] rem_ptr Remote pointer value
+ * @param[in] rem_seg Remote segment
+ * @param[in] result_ptr Location to store the result in
+ * @param[in] result_seg Local segment for the result
+ * @param[in] compare_operand 64-bit operand to compare with
+ * @param[in] swap_operand 64-but operand to swap in
+ * @param[in] cbfunc Function to call when the operation completes
(may
+ * be NULL).
+ *
+ * This function executes an atomic operation with get semantics. When the
operation
+ * is complete the result of the operation is visible in the target memory and
the
+ * previous value is visible in {result_ptr}. {result_ptr} must be a pointer
within
+ * the descriptor {result_des} and {rem_ptr} must be a pointer within
{rem_des}. The
+ * BTL may provide either blocking or non-blocking versions of this function.
If the
+ * operation is complete upon return this functions returns 1. Otherwise the
function
+ * returns OMPI_SUCCESS and the provided callback will be called on completion.
+ */
+typedef int (*mca_btl_base_module_acswap_fn_t) (struct mca_btl_base_module_t
*btl,
+ struct mca_btl_base_endpoint_t
*endpoint,
+ ompi_ptr_t rem_ptr,
+ struct mca_btl_base_segment_t
*rem_seg,
+ ompi_ptr_t result_ptr,
+ struct mca_btl_base_segment_t
*result_seg,
+ int64_t compare_operand,
+ int64_t swap_operand,
+ mca_btl_base_completion_fn_t
cbfunc,
+ void *cbdata);
+
+
+/**
* Fault Tolerance Event Notification Function
* @param state Checkpoint Status
* @return OMPI_SUCCESS or failure status
@@ -815,10 +975,12 @@ struct mca_btl_base_module_t {
size_t btl_rdma_pipeline_send_length; /**< amount of bytes that
should be send by pipeline protocol */
size_t btl_rdma_pipeline_frag_size; /**< maximum rdma fragment size
supported by the BTL */
size_t btl_min_rdma_pipeline_size; /**< minimum packet size for
pipeline protocol */
+ size_t btl_max_inline_size; /**< maxium size recommended for the
btl_sendi function */
uint32_t btl_exclusivity; /**< indicates this BTL should be used
exclusively */
uint32_t btl_latency; /**< relative ranking of latency used to
prioritize btls */
uint32_t btl_bandwidth; /**< bandwidth (Mbytes/sec) supported by
each endpoint */
uint32_t btl_flags; /**< flags (put/get...) */
+ uint32_t btl_atomic_flags; /**< supported atomic operations */
size_t btl_seg_size; /**< size of a btl segment */
/* BTL function table */
@@ -831,11 +993,16 @@ struct mca_btl_base_module_t {
mca_btl_base_module_free_fn_t btl_free;
mca_btl_base_module_prepare_fn_t btl_prepare_src;
mca_btl_base_module_prepare_fn_t btl_prepare_dst;
+ mca_btl_base_module_prepare_rdma_fn_t btl_prepare_rdma;
mca_btl_base_module_send_fn_t btl_send;
mca_btl_base_module_sendi_fn_t btl_sendi;
mca_btl_base_module_put_fn_t btl_put;
mca_btl_base_module_get_fn_t btl_get;
- mca_btl_base_module_dump_fn_t btl_dump;
+ mca_btl_base_module_dump_fn_t btl_dump;
+
+ mca_btl_base_module_aop_fn_t btl_aop;
+ mca_btl_base_module_afop_fn_t btl_afop;
+ mca_btl_base_module_acswap_fn_t btl_acswap;
/** the mpool associated with this btl (optional) */
mca_mpool_base_module_t* btl_mpool;
@@ -853,9 +1020,12 @@ typedef struct mca_btl_base_module_t
mca_btl_base_module_t;
/*
* Macro for use in modules that are of type btl v2.0.1
*/
-#define MCA_BTL_BASE_VERSION_2_0_0 \
- MCA_BASE_VERSION_2_0_0, \
- "btl", 2, 0, 0
+#define MCA_BTL_BASE_VERSION_3_0_0 \
+ MCA_BASE_VERSION_2_0_0, \
+ .mca_type_name = "btl", \
+ .mca_type_major_version = 3, \
+ .mca_type_minor_version = 0, \
+ .mca_type_release_version = 0
END_C_DECLS
diff --git a/ompi/mca/btl/portals4/btl_portals4_component.c
b/ompi/mca/btl/portals4/btl_portals4_component.c
index 4249655..feee7bc 100644
--- a/ompi/mca/btl/portals4/btl_portals4_component.c
+++ b/ompi/mca/btl/portals4/btl_portals4_component.c
@@ -43,7 +43,7 @@ mca_btl_portals4_component_t mca_btl_portals4_component = {
/* First, the mca_base_module_t struct containing meta
information about the module itself */
{
- MCA_BTL_BASE_VERSION_2_0_0,
+ MCA_BTL_BASE_VERSION_3_0_0,
"portals4", /* MCA module name */
OMPI_MAJOR_VERSION, /* MCA module major version */
diff --git a/ompi/mca/btl/scif/btl_scif_component.c
b/ompi/mca/btl/scif/btl_scif_component.c
index 559f822..92e9e81 100644
--- a/ompi/mca/btl/scif/btl_scif_component.c
+++ b/ompi/mca/btl/scif/btl_scif_component.c
@@ -32,7 +32,7 @@ mca_btl_scif_component_t mca_btl_scif_component = {
about the component itself */
.btl_version = {
- MCA_BTL_BASE_VERSION_2_0_0,
+ MCA_BTL_BASE_VERSION_3_0_0,
.mca_component_name = "scif",
.mca_component_major_version = OMPI_MAJOR_VERSION,
diff --git a/ompi/mca/btl/self/btl_self_component.c
b/ompi/mca/btl/self/btl_self_component.c
index 3550184..d515f36 100644
--- a/ompi/mca/btl/self/btl_self_component.c
+++ b/ompi/mca/btl/self/btl_self_component.c
@@ -44,7 +44,7 @@ mca_btl_self_component_t mca_btl_self_component = {
/* First, the mca_base_component_t struct containing meta information
about the component itself */
{
- MCA_BTL_BASE_VERSION_2_0_0,
+ MCA_BTL_BASE_VERSION_3_0_0,
"self", /* MCA component name */
OMPI_MAJOR_VERSION, /* MCA component major version */
diff --git a/ompi/mca/btl/sm/btl_sm_component.c
b/ompi/mca/btl/sm/btl_sm_component.c
index 6c1652c..9733059 100644
--- a/ompi/mca/btl/sm/btl_sm_component.c
+++ b/ompi/mca/btl/sm/btl_sm_component.c
@@ -85,7 +85,7 @@ mca_btl_sm_component_t mca_btl_sm_component = {
/* First, the mca_base_component_t struct containing meta information
about the component itself */
{
- MCA_BTL_BASE_VERSION_2_0_0,
+ MCA_BTL_BASE_VERSION_3_0_0,
"sm", /* MCA component name */
OMPI_MAJOR_VERSION, /* MCA component major version */
diff --git a/ompi/mca/btl/smcuda/btl_smcuda_component.c
b/ompi/mca/btl/smcuda/btl_smcuda_component.c
index 9bf2508..155790d 100644
--- a/ompi/mca/btl/smcuda/btl_smcuda_component.c
+++ b/ompi/mca/btl/smcuda/btl_smcuda_component.c
@@ -88,7 +88,7 @@ mca_btl_smcuda_component_t mca_btl_smcuda_component = {
/* First, the mca_base_component_t struct containing meta information
about the component itself */
{
- MCA_BTL_BASE_VERSION_2_0_0,
+ MCA_BTL_BASE_VERSION_3_0_0,
"smcuda", /* MCA component name */
OMPI_MAJOR_VERSION, /* MCA component major version */
diff --git a/ompi/mca/btl/tcp/btl_tcp_component.c
b/ompi/mca/btl/tcp/btl_tcp_component.c
index 4934346..288acbb 100644
--- a/ompi/mca/btl/tcp/btl_tcp_component.c
+++ b/ompi/mca/btl/tcp/btl_tcp_component.c
@@ -90,7 +90,7 @@ mca_btl_tcp_component_t mca_btl_tcp_component = {
about the component itself */
{
- MCA_BTL_BASE_VERSION_2_0_0,
+ MCA_BTL_BASE_VERSION_3_0_0,
"tcp", /* MCA component name */
OMPI_MAJOR_VERSION, /* MCA component major version */
diff --git a/ompi/mca/btl/ugni/btl_ugni_component.c
b/ompi/mca/btl/ugni/btl_ugni_component.c
index 2464aeb..9e92c76 100644
--- a/ompi/mca/btl/ugni/btl_ugni_component.c
+++ b/ompi/mca/btl/ugni/btl_ugni_component.c
@@ -32,7 +32,7 @@ mca_btl_ugni_component_t mca_btl_ugni_component = {
/* First, the mca_base_component_t struct containing meta information
about the component itself */
.btl_version = {
- MCA_BTL_BASE_VERSION_2_0_0,
+ MCA_BTL_BASE_VERSION_3_0_0,
.mca_component_name = "ugni",
.mca_component_major_version = OMPI_MAJOR_VERSION,
diff --git a/ompi/mca/btl/usnic/btl_usnic_component.c
b/ompi/mca/btl/usnic/btl_usnic_component.c
index 6ffc8f2..1341ec0 100644
--- a/ompi/mca/btl/usnic/btl_usnic_component.c
+++ b/ompi/mca/btl/usnic/btl_usnic_component.c
@@ -133,7 +133,7 @@ ompi_btl_usnic_component_t mca_btl_usnic_component = {
/* First, the mca_base_component_t struct containing meta information
about the component itself */
{
- MCA_BTL_BASE_VERSION_2_0_0,
+ MCA_BTL_BASE_VERSION_3_0_0,
"usnic", /* MCA component name */
OMPI_MAJOR_VERSION, /* MCA component major version */
diff --git a/ompi/mca/btl/vader/Makefile.am b/ompi/mca/btl/vader/Makefile.am
index d47ca83..b799005 100644
--- a/ompi/mca/btl/vader/Makefile.am
+++ b/ompi/mca/btl/vader/Makefile.am
@@ -37,7 +37,8 @@ libmca_btl_vader_la_sources = \
btl_vader_get.c \
btl_vader_put.c \
btl_vader_xpmem.c \
- btl_vader_xpmem.h
+ btl_vader_xpmem.h \
+ btl_vader_atomic.c
# Make the output library in this directory, and name it either
# mca_<type>_<name>.la (for DSO builds) or libmca_<type>_<name>.la
diff --git a/ompi/mca/btl/vader/btl_vader.h b/ompi/mca/btl/vader/btl_vader.h
index 19a650e..f1df7ad 100644
--- a/ompi/mca/btl/vader/btl_vader.h
+++ b/ompi/mca/btl/vader/btl_vader.h
@@ -226,6 +226,36 @@ mca_btl_base_descriptor_t* mca_btl_vader_alloc (struct
mca_btl_base_module_t* bt
struct
mca_btl_base_endpoint_t* endpoint,
uint8_t order, size_t size,
uint32_t flags);
+
+/**
+ * Execute an atomic operation
+ *
+ * @param[in] btl BTL module
+ * @param[in] endpoint BTL endpoint
+ * @param[in] rem_ptr Remote pointer to perform the operation on
+ * @param[in] rem_seg Remote segment rem_ptr belongs to
+ * @param[in] operand Operation operand
+ * @param[in] op Operation to execute
+ * @param[in] cbfunc Callback function to call when complete (ignored)
+ * @param[in] cbdata Callback function data (ignored)
+ *
+ * This function will always return 1 if the operation was successful so
+ * will never need to call the callback.
+ */
+int mca_btl_vader_aop (struct mca_btl_base_module_t *btl, struct
mca_btl_base_endpoint_t *endpoint,
+ ompi_ptr_t rem_ptr, mca_btl_base_segment_t *rem_seg,
int64_t operand,
+ mca_btl_base_atomic_op_t op,
mca_btl_base_completion_fn_t cbfunc, void *cbdata);
+
+int mca_btl_vader_afop (struct mca_btl_base_module_t *btl, struct
mca_btl_base_endpoint_t *endpoint,
+ ompi_ptr_t rem_ptr, mca_btl_base_segment_t *rem_seg,
ompi_ptr_t result_ptr,
+ mca_btl_base_segment_t *result_seg, int64_t operand,
mca_btl_base_atomic_op_t op,
+ mca_btl_base_completion_fn_t cbfunc, void *cbdata);
+
+int mca_btl_vader_acswap (struct mca_btl_base_module_t *btl, struct
mca_btl_base_endpoint_t *endpoint,
+ ompi_ptr_t rem_ptr, mca_btl_base_segment_t *rem_seg,
ompi_ptr_t result_ptr,
+ mca_btl_base_segment_t *result_seg, int64_t
compare_operand, int64_t swap_operand,
+ mca_btl_base_completion_fn_t cbfunc, void *cbdata);
+
END_C_DECLS
#endif
diff --git a/ompi/mca/btl/vader/btl_vader_atomic.c
b/ompi/mca/btl/vader/btl_vader_atomic.c
new file mode 100644
index 0000000..af6f379
--- /dev/null
+++ b/ompi/mca/btl/vader/btl_vader_atomic.c
@@ -0,0 +1,105 @@
+/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
+/*
+ * Copyright (c) 2010-2014 Los Alamos National Security, LLC.
+ * All rights reserved.
+ * $COPYRIGHT$
+ *
+ * Additional copyrights may follow
+ *
+ * $HEADER$
+ */
+
+#include "btl_vader.h"
+#include "btl_vader_xpmem.h"
+
+#if OMPI_BTL_VADER_HAVE_XPMEM
+int mca_btl_vader_aop (struct mca_btl_base_module_t *btl, struct
mca_btl_base_endpoint_t *endpoint,
+ ompi_ptr_t rem_ptr, mca_btl_base_segment_t *rem_seg,
int64_t operand,
+ mca_btl_base_atomic_op_t op,
mca_btl_base_completion_fn_t cbfunc, void *cbdata)
+{
+ int64_t dummy;
+ ompi_ptr_t dummy_ptr = {.lval = (uintptr_t) &dummy};
+
+ return mca_btl_vader_afop (btl, endpoint, rem_ptr, rem_seg, dummy_ptr,
NULL, operand, op, NULL, NULL);
+}
+
+int mca_btl_vader_afop (struct mca_btl_base_module_t *btl, struct
mca_btl_base_endpoint_t *endpoint,
+ ompi_ptr_t rem_ptr, mca_btl_base_segment_t *rem_seg,
ompi_ptr_t result_ptr,
+ mca_btl_base_segment_t *result_seg, int64_t operand,
mca_btl_base_atomic_op_t op,
+ mca_btl_base_completion_fn_t cbfunc, void *cbdata)
+{
+ mca_mpool_base_registration_t *reg;
+ ptrdiff_t offset;
+ char *base_ptr;
+ int ret = 1;
+
+ offset = rem_ptr.lval - rem_seg->seg_addr.lval;
+
+ /* range check */
+ if (OPAL_UNLIKELY((size_t) offset > (rem_seg->seg_len - 8))) {
+ return OMPI_ERR_VALUE_OUT_OF_BOUNDS;
+ }
+
+ /* NTH: TODO -- implement more atomics if needed. Add is enough to
+ * correctly implement locks for RMA */
+ if (OPAL_UNLIKELY(MCA_BTL_BASE_ATOMIC_ADD != op)) {
+ return OMPI_ERR_NOT_SUPPORTED;
+ }
+
+ reg = vader_get_registation (endpoint, rem_seg->seg_addr.pval,
rem_seg->seg_len, 0, (void **) &base_ptr);
+ if (OPAL_UNLIKELY(NULL == reg)) {
+ return OMPI_ERROR;
+ }
+
+ base_ptr += offset;
+
+ opal_atomic_mb ();
+
+ ((int64_t *) result_ptr.pval)[0] = opal_atomic_add_64 ((int64_t *)
base_ptr, operand) - operand;
+
+ opal_atomic_mb ();
+
+ vader_return_registration (reg, endpoint);
+
+ /* ignore the callback as it is not needed for shared memory */
+
+ return ret;
+}
+
+int mca_btl_vader_acswap (struct mca_btl_base_module_t *btl, struct
mca_btl_base_endpoint_t *endpoint,
+ ompi_ptr_t rem_ptr, mca_btl_base_segment_t *rem_seg,
ompi_ptr_t result_ptr,
+ mca_btl_base_segment_t *result_seg, int64_t
compare_operand, int64_t swap_operand,
+ mca_btl_base_completion_fn_t cbfunc, void *cbdata)
+{
+ mca_mpool_base_registration_t *reg;
+ ptrdiff_t offset;
+ char *base_ptr;
+ int ret = 1;
+
+ offset = rem_ptr.lval - rem_seg->seg_addr.lval;
+
+ /* range check */
+ if (OPAL_UNLIKELY((size_t) offset > (rem_seg->seg_len - 8))) {
+ return OMPI_ERR_VALUE_OUT_OF_BOUNDS;
+ }
+
+ reg = vader_get_registation (endpoint, rem_seg->seg_addr.pval,
rem_seg->seg_len, 0, (void **) &base_ptr);
+ if (OPAL_UNLIKELY(NULL == reg)) {
+ return OMPI_ERROR;
+ }
+
+ base_ptr += offset;
+
+ opal_atomic_mb ();
+
+ ((int64_t *) result_ptr.pval)[0] = opal_atomic_cmpset_64 ((int64_t *)
base_ptr, compare_operand, swap_operand);
+
+ opal_atomic_mb ();
+
+ vader_return_registration (reg, endpoint);
+
+ /* ignore the callback as it is not needed for shared memory */
+
+ return ret;
+}
+#endif
diff --git a/ompi/mca/btl/vader/btl_vader_component.c
b/ompi/mca/btl/vader/btl_vader_component.c
index 08794e0..ac9e44a 100644
--- a/ompi/mca/btl/vader/btl_vader_component.c
+++ b/ompi/mca/btl/vader/btl_vader_component.c
@@ -51,7 +51,7 @@ mca_btl_vader_component_t mca_btl_vader_component = {
/* First, the mca_base_component_t struct containing meta information
about the component itself */
.btl_version = {
- MCA_BTL_BASE_VERSION_2_0_0,
+ MCA_BTL_BASE_VERSION_3_0_0,
.mca_component_name = "vader",
.mca_component_major_version = OMPI_MAJOR_VERSION,
@@ -166,6 +166,11 @@ static int mca_btl_vader_component_register (void)
mca_btl_vader.super.btl_flags = MCA_BTL_FLAGS_SEND_INPLACE;
#endif
+#if OMPI_BTL_VADER_HAVE_XPMEM
+ mca_btl_vader.super.btl_flags |= MCA_BTL_FLAGS_ATOMIC_CSWAP |
MCA_BTL_FLAGS_ATOMIC_OPS;
+ mca_btl_vader.super.btl_atomic_flags = MCA_BTL_ATOMIC_FLAG_ADD;
+#endif
+
mca_btl_vader.super.btl_seg_size = sizeof (mca_btl_base_segment_t);
#if OMPI_BTL_VADER_HAVE_XPMEM || OMPI_BTL_VADER_HAVE_CMA
mca_btl_vader.super.btl_bandwidth = 40000; /* Mbs */
diff --git a/ompi/mca/btl/vader/btl_vader_module.c
b/ompi/mca/btl/vader/btl_vader_module.c
index b52b188..b355be0 100644
--- a/ompi/mca/btl/vader/btl_vader_module.c
+++ b/ompi/mca/btl/vader/btl_vader_module.c
@@ -89,6 +89,9 @@ mca_btl_vader_t mca_btl_vader = {
#endif
.btl_dump = mca_btl_base_dump,
+ .btl_aop = mca_btl_vader_aop,
+ .btl_afop = mca_btl_vader_afop,
+ .btl_acswap = mca_btl_vader_acswap,
.btl_register_error = vader_register_error_cb,
.btl_ft_event = vader_ft_event
}
pgpTJICL1vxoB.pgp
Description: PGP signature
