def matrix_multiply():
# Configure array dimensions. Force an unequal data distribution.
dims = [TOTALELEMS]*NDIM
chunk = [TOTALELEMS/nprocs-1]*NDIM
# Create a global array g_a and duplicate it to get g_b and g_c.
g_a = ga.create(ga.C_DBL, dims, "array A", chunk)
if not g_a: ga.error("create failed: A")
if not me: print "Created Array A"
g_b = ga.duplicate(g_a, "array B")
g_c = ga.duplicate(g_a, "array C")
if not g_b or not g_c: ga.eror("duplicate failed")
if not me: print "Created Arrays B and C"
# Initialize data in matrices a and b.
if not me: print "Initializing matrix A and B"
a = np.random.rand(*dims)*29
b = np.random.rand(*dims)*37
# Copy data to global arrays g_a and g_b.
if not me:
ga.put(g_a, a)
ga.put(g_b, b)
# Synchronize all processors to make sure everyone has data.
ga.sync()
# Determine which block of data is locally owned. Note that
# the same block is locally owned for all GAs.
lo,hi = ga.distribution(g_c)
# Get the blocks from g_a and g_b needed to compute this block in
# g_c and copy them into the local buffers a and b.
a = ga.get(g_a, (lo[0],0), (hi[0],dims[0]))
b = ga.get(g_b, (0,lo[1]), (dims[1],hi[1]))
# Do local matrix multiplication and store the result in local
# buffer c. Start by evaluating the transpose of b.
btrns = b.transpose()
# Multiply a and b to get c.
c = np.dot(a,b)
# Copy c back to g_c.
ga.put(g_c, c, lo, hi)
verify(g_a, g_b, g_c)
# Deallocate arrays.
ga.destroy(g_a)
ga.destroy(g_b)
ga.destroy(g_c)
def verify_using_ga(g_a, g_b, g_c):
g_v = ga.duplicate(g_c)
ga.gemm(False, False, N, N, N, 1, g_a, g_b, 0, g_v)
c = ga.access(g_c)
v = ga.access(g_v)
if c is not None:
val = int(np.abs(np.sum(c - v)) > 0.0001)
else:
val = 0
val = ga.gop_add(val)
ga.destroy(g_v)
return val == 0
def verify_using_ga(g_a, g_b, g_c):
g_v = ga.duplicate(g_c)
ga.gemm(False,False,N,N,N,1,g_a,g_b,0,g_v)
c = ga.access(g_c)
v = ga.access(g_v)
if c is not None:
val = int(np.abs(np.sum(c-v))>0.0001)
else:
val = 0
val = ga.gop_add(val)
ga.destroy(g_v)
return val == 0
def TRANSPOSE1D():
# Configure array dimensions. Force an unequal data distribution.
dims = [nprocs * TOTALELEMS + nprocs / 2]
chunk = [TOTALELEMS] # minimum data on each process
# create a global array g_a and duplicate it to get g_b
g_a = ga.create(ga.C_INT, dims, "array A", chunk)
if not g_a: ga.error("create failed: A")
if not me: print "Created Array A"
g_b = ga.duplicate(g_a, "array B")
if not g_b: ga.error("duplicate failed")
if not me: print "Created Array B"
# initialize data in g_a
if not me:
print "Initializing matrix A"
ga.put(g_a, np.arange(dims[0], dtype=np.int32))
# Synchronize all processors to guarantee that everyone has data
# before proceeding to the next step.
ga.sync()
# Start initial phase of inversion by inverting the data held locally on
# each processor. Start by finding out which data each processor owns.
lo, hi = ga.distribution(g_a)
# Get locally held data and copy it into local buffer a
a = ga.get(g_a, lo, hi)
# Invert data locally
b = a[::-1]
# Invert data globally by copying locally inverted blocks into
# their inverted positions in the GA
ga.put(g_b, b, dims[0] - hi[0], dims[0] - lo[0])
# Synchronize all processors to make sure inversion is complete
ga.sync()
# Check to see if inversion is correct
if not me: verify(g_a, g_b)
# Deallocate arrays
ga.destroy(g_a)
ga.destroy(g_b)
def TRANSPOSE1D():
# Configure array dimensions. Force an unequal data distribution.
dims = [nprocs*TOTALELEMS + nprocs/2]
chunk = [TOTALELEMS] # minimum data on each process
# create a global array g_a and duplicate it to get g_b
g_a = ga.create(ga.C_INT, dims, "array A", chunk)
if not g_a: ga.error("create failed: A")
if not me: print "Created Array A"
g_b = ga.duplicate(g_a, "array B")
if not g_b: ga.error("duplicate failed")
if not me: print "Created Array B"
# initialize data in g_a
if not me:
print "Initializing matrix A"
ga.put(g_a, np.arange(dims[0], dtype=np.int32))
# Synchronize all processors to guarantee that everyone has data
# before proceeding to the next step.
ga.sync()
# Start initial phase of inversion by inverting the data held locally on
# each processor. Start by finding out which data each processor owns.
lo,hi = ga.distribution(g_a)
# Get locally held data and copy it into local buffer a
a = ga.get(g_a, lo, hi)
# Invert data locally
b = a[::-1]
# Invert data globally by copying locally inverted blocks into
# their inverted positions in the GA
ga.put(g_b, b, dims[0]-hi[0], dims[0]-lo[0])
# Synchronize all processors to make sure inversion is complete
ga.sync()
# Check to see if inversion is correct
if not me: verify(g_a, g_b)
# Deallocate arrays
ga.destroy(g_a)
ga.destroy(g_b)
def verify(g_a, g_b, g_c):
g_chk = ga.duplicate(g_a, "array check")
if not g_chk: ga.error("duplicate failed")
ga.sync()
ga.gemm(False, False, TOTALELEMS, TOTALELEMS, TOTALELEMS, 1.0, g_a, g_b,
0.0, g_chk);
ga.sync()
ga.add(g_c, g_chk, g_chk, 1.0, -1.0)
rchk = ga.dot(g_chk, g_chk)
if not me:
print "Normed difference in matrices: %12.4f" % rchk
if not (-TOLERANCE < rchk < TOLERANCE):
ga.error("Matrix multiply verify failed")
else:
print "Matrix Multiply OK"
ga.destroy(g_chk)
return value < EPSILON
def convergence_test_L2(g_a, g_b):
# compute L2 norm of change
# subtract g_b from g_a, results stored in g_b
ga.add(g_a, g_b, g_b, beta=-1)
# compute elementwise dot product (i.e. treats N-d arrays as vectors)
value = ga.dot(g_b, g_b)
if DEBUG:
print_sync(value)
return value < EPSILON
# create GA, distribute entire rows
g_a = ga.create(ga.C_FLOAT, (dim,dim), chunk=(0,dim))
# create a duplicate GA for the convergence test
g_b = ga.duplicate(g_a)
# process 0 initializes global array
# Note: alternatively, each process could initialize its local data using
# ga.access() and ga.distribution()
a = np.zeros((dim,dim), dtype=np.float32)
if rank == 0:
a[0,:] = 100 #top row
a[:,0] = 75 #left column
a[:,a.shape[0] - 1] = 50 #right column
ga.put(g_a, a)
ga.sync()
# which piece of array do I own?
# note that rhi and chi follow python range conventions i.e. [lo,hi)
(rlo,clo),(rhi,chi) = ga.distribution(g_a)