jaunatisblue/final-qibotn
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Remove eval_tn_mpi

Browse Source
This commit is contained in:
tankya2
2024-01-30 10:23:33 +08:00
parent 293af81b43
commit 1035629745
1 changed files with 0 additions and 75 deletions
Download Patch File Download Diff File Expand all files Collapse all files

75
src/qibotn/cutn.py
View File

@@ -433,80 +433,6 @@ def eval_tn_MPI_expectation(qibo_circ, datatype, n_samples=8):
return result, rank
def eval_tn_MPI(qibo_circ, datatype, n_samples=8):
"""Convert qibo circuit to tensornet (TN) format and perform contraction using multi node and multi GPU through MPI.
The conversion is performed by QiboCircuitToEinsum(), after which it goes through 2 steps: pathfinder and execution.
The pathfinder looks at user defined number of samples (n_samples) iteratively to select the least costly contraction path. This is sped up with multi thread.
After pathfinding the optimal path is used in the actual contraction to give a dense vector representation of the TN.
"""
from mpi4py import MPI # this line initializes MPI
import socket
# Get the hostname
# hostname = socket.gethostname()
ncpu_threads = multiprocessing.cpu_count() // 2
comm = MPI.COMM_WORLD
rank = comm.Get_rank()
size = comm.Get_size()
# mem_avail = cp.cuda.Device().mem_info[0]
# print("Mem avail: Start",mem_avail, "rank =",rank, "hostname =",hostname)
device_id = rank % getDeviceCount()
cp.cuda.Device(device_id).use()
handle = cutn.create()
network_opts = cutn.NetworkOptions(handle=handle, blocking="auto")
# mem_avail = cp.cuda.Device().mem_info[0]
# print("Mem avail: aft network opts",mem_avail, "rank =",rank)
cutn.distributed_reset_configuration(handle, *cutn.get_mpi_comm_pointer(comm))
# mem_avail = cp.cuda.Device().mem_info[0]
# print("Mem avail: aft distributed reset config",mem_avail, "rank =",rank)
# Perform circuit conversion
myconvertor = QiboCircuitToEinsum(qibo_circ, dtype=datatype)
# mem_avail = cp.cuda.Device().mem_info[0]
# print("Mem avail: aft convetor",mem_avail, "rank =",rank)
operands_interleave = myconvertor.state_vector_operands()
# mem_avail = cp.cuda.Device().mem_info[0]
# print("Mem avail: aft operand interleave",mem_avail, "rank =",rank)
# Pathfinder: To search for the optimal path. Optimal path are assigned to path and info attribute of the network object.
network = cutn.Network(*operands_interleave, options=network_opts)
# mem_avail = cp.cuda.Device().mem_info[0]
# print("Mem avail: aft cutn.Network(*operands_interleave,",mem_avail, "rank =",rank)
network.contract_path(
optimize={
"samples": n_samples,
"threads": ncpu_threads,
"slicing": {"min_slices": max(16, size)},
}
)
# mem_avail = cp.cuda.Device().mem_info[0]
# print("Mem avail: aft contract path",mem_avail, "rank =",rank)
# Execution: To execute the contraction using the optimal path found previously
# print("opt_cost",opt_info.opt_cost, "Process =",rank)
"""
path, opt_info = network.contract_path(optimize={"samples": n_samples, "threads": ncpu_threads, 'slicing': {'min_slices': max(16, size)}})
num_slices = opt_info.num_slices#Andy
chunk, extra = num_slices // size, num_slices % size#Andy
slice_begin = rank * chunk + min(rank, extra)#Andy
slice_end = num_slices if rank == size - 1 else (rank + 1) * chunk + min(rank + 1, extra)#Andy
slices = range(slice_begin, slice_end)#Andy
result = network.contract(slices=slices)
"""
result = network.contract()
# mem_avail = cp.cuda.Device().mem_info[0]
# print("Mem avail: aft contract",mem_avail, "rank =",rank)
cutn.destroy(handle)
return result, rank
def eval_mps(qibo_circ, gate_algo, datatype):
myconvertor = QiboCircuitToMPS(qibo_circ, gate_algo, dtype=datatype)
mps_helper = MPSContractionHelper(myconvertor.num_qubits)
@@ -515,7 +441,6 @@ def eval_mps(qibo_circ, gate_algo, datatype):
myconvertor.mps_tensors, {"handle": myconvertor.handle}
)
def PauliStringGen(nqubits):
if nqubits <= 0:
return "Invalid input. N should be a positive integer."