mirror of
https://github.com/microsoft/mscclpp.git
synced 2026-05-11 17:00:22 +00:00
774 lines
17 KiB
JSON
774 lines
17 KiB
JSON
{
|
|
"name": "reduce_tbg",
|
|
"collective": "allreduce",
|
|
"protocol": "Simple",
|
|
"inplace": true,
|
|
"reuse_resources": false,
|
|
"gpus": [
|
|
{
|
|
"id": 0,
|
|
"input_chunks": 4,
|
|
"output_chunks": 4,
|
|
"scratch_chunks": 4,
|
|
"threadblocks": [
|
|
{
|
|
"id": 0,
|
|
"ops": [
|
|
{
|
|
"name": "copy",
|
|
"src_buff": [
|
|
{
|
|
"type": "i",
|
|
"index": 2,
|
|
"size": 2
|
|
}
|
|
],
|
|
"dst_buff": [
|
|
{
|
|
"type": "s",
|
|
"index": 2,
|
|
"size": 2
|
|
}
|
|
],
|
|
"tbg_info": {
|
|
"tb_id": 0,
|
|
"tbg_size": 2
|
|
}
|
|
},
|
|
{
|
|
"name": "nop"
|
|
},
|
|
{
|
|
"name": "signal",
|
|
"channel_ids": [
|
|
0
|
|
],
|
|
"channel_type": "memory"
|
|
},
|
|
{
|
|
"name": "nop"
|
|
},
|
|
{
|
|
"name": "wait",
|
|
"channel_ids": [
|
|
0
|
|
],
|
|
"channel_type": "memory"
|
|
},
|
|
{
|
|
"name": "nop"
|
|
},
|
|
{
|
|
"name": "rres",
|
|
"src_buff": [
|
|
{
|
|
"type": "i",
|
|
"index": 0,
|
|
"size": 1
|
|
},
|
|
{
|
|
"buffer_id": 0,
|
|
"index": 0,
|
|
"size": 1
|
|
}
|
|
],
|
|
"dst_buff": [
|
|
{
|
|
"type": "i",
|
|
"index": 0,
|
|
"size": 1
|
|
},
|
|
{
|
|
"buffer_id": 1,
|
|
"index": 0,
|
|
"size": 1
|
|
}
|
|
],
|
|
"channel_type": "memory",
|
|
"reduce_op": "sum",
|
|
"tbg_info": {
|
|
"tb_id": 0,
|
|
"tbg_size": 2
|
|
}
|
|
},
|
|
{
|
|
"name": "get",
|
|
"src_buff": [
|
|
{
|
|
"buffer_id": 0,
|
|
"index": 1,
|
|
"size": 1
|
|
}
|
|
],
|
|
"dst_buff": [
|
|
{
|
|
"type": "s",
|
|
"index": 1,
|
|
"size": 1
|
|
}
|
|
],
|
|
"channel_ids": [
|
|
0
|
|
],
|
|
"channel_type": "memory",
|
|
"tbg_info": {
|
|
"tb_id": 0,
|
|
"tbg_size": 2
|
|
}
|
|
},
|
|
{
|
|
"name": "nop"
|
|
},
|
|
{
|
|
"name": "res",
|
|
"src_buff": [
|
|
{
|
|
"type": "i",
|
|
"index": 1,
|
|
"size": 1
|
|
},
|
|
{
|
|
"type": "s",
|
|
"index": 1,
|
|
"size": 1
|
|
}
|
|
],
|
|
"dst_buff": [
|
|
{
|
|
"type": "i",
|
|
"index": 1,
|
|
"size": 1
|
|
},
|
|
{
|
|
"buffer_id": 1,
|
|
"index": 1,
|
|
"size": 1
|
|
}
|
|
],
|
|
"channel_type": "memory",
|
|
"reduce_op": "sum",
|
|
"tbg_info": {
|
|
"tb_id": 0,
|
|
"tbg_size": 2
|
|
}
|
|
},
|
|
{
|
|
"name": "nop"
|
|
},
|
|
{
|
|
"name": "signal",
|
|
"channel_ids": [
|
|
0
|
|
],
|
|
"channel_type": "memory"
|
|
},
|
|
{
|
|
"name": "nop"
|
|
},
|
|
{
|
|
"name": "wait",
|
|
"channel_ids": [
|
|
0
|
|
],
|
|
"channel_type": "memory"
|
|
},
|
|
{
|
|
"name": "nop"
|
|
}
|
|
],
|
|
"channels": [
|
|
{
|
|
"channel_type": "memory",
|
|
"channel_ids": [
|
|
0
|
|
]
|
|
}
|
|
],
|
|
"remote_buffer_refs": [
|
|
{
|
|
"access_channel_type": "memory",
|
|
"remote_buffer_ids": [
|
|
0,
|
|
1
|
|
]
|
|
}
|
|
]
|
|
},
|
|
{
|
|
"id": 1,
|
|
"ops": [
|
|
{
|
|
"name": "copy",
|
|
"src_buff": [
|
|
{
|
|
"type": "i",
|
|
"index": 2,
|
|
"size": 2
|
|
}
|
|
],
|
|
"dst_buff": [
|
|
{
|
|
"type": "s",
|
|
"index": 2,
|
|
"size": 2
|
|
}
|
|
],
|
|
"tbg_info": {
|
|
"tb_id": 1,
|
|
"tbg_size": 2
|
|
}
|
|
},
|
|
{
|
|
"name": "nop"
|
|
},
|
|
{
|
|
"name": "signal",
|
|
"channel_ids": [
|
|
0
|
|
],
|
|
"channel_type": "memory"
|
|
},
|
|
{
|
|
"name": "nop"
|
|
},
|
|
{
|
|
"name": "wait",
|
|
"channel_ids": [
|
|
0
|
|
],
|
|
"channel_type": "memory"
|
|
},
|
|
{
|
|
"name": "nop"
|
|
},
|
|
{
|
|
"name": "rres",
|
|
"src_buff": [
|
|
{
|
|
"type": "i",
|
|
"index": 0,
|
|
"size": 1
|
|
},
|
|
{
|
|
"buffer_id": 0,
|
|
"index": 0,
|
|
"size": 1
|
|
}
|
|
],
|
|
"dst_buff": [
|
|
{
|
|
"type": "i",
|
|
"index": 0,
|
|
"size": 1
|
|
},
|
|
{
|
|
"buffer_id": 1,
|
|
"index": 0,
|
|
"size": 1
|
|
}
|
|
],
|
|
"channel_type": "memory",
|
|
"reduce_op": "sum",
|
|
"tbg_info": {
|
|
"tb_id": 1,
|
|
"tbg_size": 2
|
|
}
|
|
},
|
|
{
|
|
"name": "get",
|
|
"src_buff": [
|
|
{
|
|
"buffer_id": 0,
|
|
"index": 1,
|
|
"size": 1
|
|
}
|
|
],
|
|
"dst_buff": [
|
|
{
|
|
"type": "s",
|
|
"index": 1,
|
|
"size": 1
|
|
}
|
|
],
|
|
"channel_ids": [
|
|
1
|
|
],
|
|
"channel_type": "memory",
|
|
"tbg_info": {
|
|
"tb_id": 1,
|
|
"tbg_size": 2
|
|
}
|
|
},
|
|
{
|
|
"name": "nop"
|
|
},
|
|
{
|
|
"name": "res",
|
|
"src_buff": [
|
|
{
|
|
"type": "i",
|
|
"index": 1,
|
|
"size": 1
|
|
},
|
|
{
|
|
"type": "s",
|
|
"index": 1,
|
|
"size": 1
|
|
}
|
|
],
|
|
"dst_buff": [
|
|
{
|
|
"type": "i",
|
|
"index": 1,
|
|
"size": 1
|
|
},
|
|
{
|
|
"buffer_id": 1,
|
|
"index": 1,
|
|
"size": 1
|
|
}
|
|
],
|
|
"channel_type": "memory",
|
|
"reduce_op": "sum",
|
|
"tbg_info": {
|
|
"tb_id": 1,
|
|
"tbg_size": 2
|
|
}
|
|
},
|
|
{
|
|
"name": "nop"
|
|
},
|
|
{
|
|
"name": "signal",
|
|
"channel_ids": [
|
|
0
|
|
],
|
|
"channel_type": "memory"
|
|
},
|
|
{
|
|
"name": "nop"
|
|
},
|
|
{
|
|
"name": "wait",
|
|
"channel_ids": [
|
|
0
|
|
],
|
|
"channel_type": "memory"
|
|
},
|
|
{
|
|
"name": "nop"
|
|
}
|
|
],
|
|
"channels": [
|
|
{
|
|
"channel_type": "memory",
|
|
"channel_ids": [
|
|
1,
|
|
0
|
|
]
|
|
}
|
|
],
|
|
"remote_buffer_refs": [
|
|
{
|
|
"access_channel_type": "memory",
|
|
"remote_buffer_ids": [
|
|
0,
|
|
1
|
|
]
|
|
}
|
|
]
|
|
}
|
|
],
|
|
"channels": [
|
|
{
|
|
"channel_type": "memory",
|
|
"connected_to": [
|
|
1,
|
|
1
|
|
]
|
|
}
|
|
],
|
|
"remote_buffers": [
|
|
{
|
|
"rank": 1,
|
|
"type": "s",
|
|
"access_channel_types": [
|
|
"memory"
|
|
]
|
|
},
|
|
{
|
|
"rank": 1,
|
|
"type": "i",
|
|
"access_channel_types": [
|
|
"memory"
|
|
]
|
|
}
|
|
],
|
|
"semaphores": []
|
|
},
|
|
{
|
|
"id": 1,
|
|
"input_chunks": 4,
|
|
"output_chunks": 4,
|
|
"scratch_chunks": 4,
|
|
"threadblocks": [
|
|
{
|
|
"id": 0,
|
|
"ops": [
|
|
{
|
|
"name": "copy",
|
|
"src_buff": [
|
|
{
|
|
"type": "i",
|
|
"index": 0,
|
|
"size": 2
|
|
}
|
|
],
|
|
"dst_buff": [
|
|
{
|
|
"type": "s",
|
|
"index": 0,
|
|
"size": 2
|
|
}
|
|
],
|
|
"tbg_info": {
|
|
"tb_id": 0,
|
|
"tbg_size": 2
|
|
}
|
|
},
|
|
{
|
|
"name": "nop"
|
|
},
|
|
{
|
|
"name": "signal",
|
|
"channel_ids": [
|
|
0
|
|
],
|
|
"channel_type": "memory"
|
|
},
|
|
{
|
|
"name": "nop"
|
|
},
|
|
{
|
|
"name": "wait",
|
|
"channel_ids": [
|
|
0
|
|
],
|
|
"channel_type": "memory"
|
|
},
|
|
{
|
|
"name": "nop"
|
|
},
|
|
{
|
|
"name": "rre",
|
|
"src_buff": [
|
|
{
|
|
"type": "i",
|
|
"index": 2,
|
|
"size": 2
|
|
},
|
|
{
|
|
"buffer_id": 0,
|
|
"index": 2,
|
|
"size": 2
|
|
}
|
|
],
|
|
"dst_buff": [
|
|
{
|
|
"type": "i",
|
|
"index": 2,
|
|
"size": 2
|
|
}
|
|
],
|
|
"channel_type": "memory",
|
|
"reduce_op": "sum",
|
|
"tbg_info": {
|
|
"tb_id": 0,
|
|
"tbg_size": 2
|
|
}
|
|
},
|
|
{
|
|
"name": "nop"
|
|
},
|
|
{
|
|
"name": "copy",
|
|
"src_buff": [
|
|
{
|
|
"type": "i",
|
|
"index": 2,
|
|
"size": 2
|
|
}
|
|
],
|
|
"dst_buff": [
|
|
{
|
|
"type": "s",
|
|
"index": 2,
|
|
"size": 2
|
|
}
|
|
],
|
|
"tbg_info": {
|
|
"tb_id": 0,
|
|
"tbg_size": 2
|
|
}
|
|
},
|
|
{
|
|
"name": "nop"
|
|
},
|
|
{
|
|
"name": "put",
|
|
"src_buff": [
|
|
{
|
|
"type": "s",
|
|
"index": 2,
|
|
"size": 2
|
|
}
|
|
],
|
|
"dst_buff": [
|
|
{
|
|
"buffer_id": 0,
|
|
"index": 2,
|
|
"size": 2
|
|
}
|
|
],
|
|
"channel_type": "memory",
|
|
"tbg_info": {
|
|
"tb_id": 0,
|
|
"tbg_size": 2
|
|
}
|
|
},
|
|
{
|
|
"name": "nop"
|
|
},
|
|
{
|
|
"name": "signal",
|
|
"channel_ids": [
|
|
0
|
|
],
|
|
"channel_type": "memory"
|
|
},
|
|
{
|
|
"name": "nop"
|
|
},
|
|
{
|
|
"name": "wait",
|
|
"channel_ids": [
|
|
0
|
|
],
|
|
"channel_type": "memory"
|
|
},
|
|
{
|
|
"name": "nop"
|
|
}
|
|
],
|
|
"channels": [
|
|
{
|
|
"channel_type": "memory",
|
|
"channel_ids": [
|
|
0
|
|
]
|
|
}
|
|
],
|
|
"remote_buffer_refs": [
|
|
{
|
|
"access_channel_type": "memory",
|
|
"remote_buffer_ids": [
|
|
0
|
|
]
|
|
}
|
|
]
|
|
},
|
|
{
|
|
"id": 1,
|
|
"ops": [
|
|
{
|
|
"name": "copy",
|
|
"src_buff": [
|
|
{
|
|
"type": "i",
|
|
"index": 0,
|
|
"size": 2
|
|
}
|
|
],
|
|
"dst_buff": [
|
|
{
|
|
"type": "s",
|
|
"index": 0,
|
|
"size": 2
|
|
}
|
|
],
|
|
"tbg_info": {
|
|
"tb_id": 1,
|
|
"tbg_size": 2
|
|
}
|
|
},
|
|
{
|
|
"name": "nop"
|
|
},
|
|
{
|
|
"name": "signal",
|
|
"channel_ids": [
|
|
0
|
|
],
|
|
"channel_type": "memory"
|
|
},
|
|
{
|
|
"name": "nop"
|
|
},
|
|
{
|
|
"name": "wait",
|
|
"channel_ids": [
|
|
0
|
|
],
|
|
"channel_type": "memory"
|
|
},
|
|
{
|
|
"name": "nop"
|
|
},
|
|
{
|
|
"name": "rre",
|
|
"src_buff": [
|
|
{
|
|
"type": "i",
|
|
"index": 2,
|
|
"size": 2
|
|
},
|
|
{
|
|
"buffer_id": 0,
|
|
"index": 2,
|
|
"size": 2
|
|
}
|
|
],
|
|
"dst_buff": [
|
|
{
|
|
"type": "i",
|
|
"index": 2,
|
|
"size": 2
|
|
}
|
|
],
|
|
"channel_type": "memory",
|
|
"reduce_op": "sum",
|
|
"tbg_info": {
|
|
"tb_id": 1,
|
|
"tbg_size": 2
|
|
}
|
|
},
|
|
{
|
|
"name": "nop"
|
|
},
|
|
{
|
|
"name": "copy",
|
|
"src_buff": [
|
|
{
|
|
"type": "i",
|
|
"index": 2,
|
|
"size": 2
|
|
}
|
|
],
|
|
"dst_buff": [
|
|
{
|
|
"type": "s",
|
|
"index": 2,
|
|
"size": 2
|
|
}
|
|
],
|
|
"tbg_info": {
|
|
"tb_id": 1,
|
|
"tbg_size": 2
|
|
}
|
|
},
|
|
{
|
|
"name": "nop"
|
|
},
|
|
{
|
|
"name": "put",
|
|
"src_buff": [
|
|
{
|
|
"type": "s",
|
|
"index": 2,
|
|
"size": 2
|
|
}
|
|
],
|
|
"dst_buff": [
|
|
{
|
|
"buffer_id": 0,
|
|
"index": 2,
|
|
"size": 2
|
|
}
|
|
],
|
|
"channel_type": "memory",
|
|
"tbg_info": {
|
|
"tb_id": 1,
|
|
"tbg_size": 2
|
|
}
|
|
},
|
|
{
|
|
"name": "nop"
|
|
},
|
|
{
|
|
"name": "signal",
|
|
"channel_ids": [
|
|
0
|
|
],
|
|
"channel_type": "memory"
|
|
},
|
|
{
|
|
"name": "nop"
|
|
},
|
|
{
|
|
"name": "wait",
|
|
"channel_ids": [
|
|
0
|
|
],
|
|
"channel_type": "memory"
|
|
},
|
|
{
|
|
"name": "nop"
|
|
}
|
|
],
|
|
"channels": [
|
|
{
|
|
"channel_type": "memory",
|
|
"channel_ids": [
|
|
1,
|
|
0
|
|
]
|
|
}
|
|
],
|
|
"remote_buffer_refs": [
|
|
{
|
|
"access_channel_type": "memory",
|
|
"remote_buffer_ids": [
|
|
0
|
|
]
|
|
}
|
|
]
|
|
}
|
|
],
|
|
"channels": [
|
|
{
|
|
"channel_type": "memory",
|
|
"connected_to": [
|
|
0,
|
|
0
|
|
]
|
|
}
|
|
],
|
|
"remote_buffers": [
|
|
{
|
|
"rank": 0,
|
|
"type": "i",
|
|
"access_channel_types": [
|
|
"memory"
|
|
]
|
|
}
|
|
],
|
|
"semaphores": []
|
|
}
|
|
],
|
|
"num_threads_per_block": 1024,
|
|
"use_double_scratch_buffer": false,
|
|
"buffer_alignment": 16,
|
|
"min_message_size": 0,
|
|
"max_message_size": 18446744073709551615
|
|
}
|