def test_conv_layer():
dtype = np.float32
ng = NervanaGPU(stochastic_round=False, bench=True)
nc = NervanaCPU()
N, C, K = 64, 64, 64
D, H, W = 1, 5, 5
T, R, S = 1, 3, 3
padding_d, padding_h, padding_w = 0, 1, 1
strides_d, strides_h, strides_w = 1, 1, 1
conv_ng = ng.conv_layer(dtype, N, C, K, D, H, W, T, R, S, padding_d,
padding_h, padding_w, strides_d, strides_h,
strides_w)
conv_nc = nc.conv_layer(dtype, N, C, K, D, H, W, T, R, S, padding_d,
padding_h, padding_w, strides_d, strides_h,
strides_w)
assert conv_nc.dimI == conv_ng.dimI
assert conv_nc.dimF == conv_ng.dimF
assert conv_nc.dimO == conv_ng.dimO
assert conv_nc.M == conv_ng.M
dimI = conv_ng.dimI
dimF = conv_ng.dimF
dimO = conv_ng.dimO
# cpu input arrays
cpuI = np.random.uniform(-0.8, 0.8, slicable(dimI, 1)).astype(np.float32)
cpuF = np.random.uniform(0.0, 0.3, slicable(dimF)).astype(np.float32)
cpuE = np.random.uniform(-0.2, 0.2, dimO).astype(np.float32)
# zero pad the last row of cpu input for the sake of numpy
cpuI[-1, :] = 0.0
# =======GPU and CPU==========
beI = cpuI[:-1, :].reshape(dimI)
beF = cpuF.reshape(dimF)
beE = cpuE
start_gpu = default_timer()
ngO, ngB, ngU = run_backend_conv(ng, conv_ng, beI, beF, beE, dtype)
end_gpu = default_timer()
start_cpu = default_timer()
ncO, ncB, ncU = run_backend_conv(nc, conv_nc, beI, beF, beE, dtype)
end_cpu = default_timer()
print("gputime: %s, cputime %s" %
(end_gpu - start_gpu, end_cpu - start_cpu))
# ======numpy===========
# cpu output arrays
cpuO = np.zeros(dimO, dtype=dtype)
cpuB = np.zeros(slicable(dimI, 1), dtype=dtype)
cpuU = np.zeros(slicable(dimF), dtype=dtype)
D, H, W = conv_nc.DHW
T, R, S = conv_nc.TRS
M, P, Q = conv_nc.MPQ
pad_d, pad_h, pad_w = conv_nc.padding
str_d, str_h, str_w = conv_nc.strides
for m in range(M):
mt = m * str_d - pad_d
for p in range(P):
pr = p * str_h - pad_h
for q in range(Q):
qs = q * str_w - pad_w
idx = pixel_indices(conv_nc, mt, pr, qs)
cpuO[:, m, p, q, :] = np.dot(cpuF.T, cpuI[idx, :])
cpuB[idx, :] += np.dot(cpuF, cpuE[:, m, p, q, :])
cpuU += np.dot(cpuI[idx, :], cpuE[:, m, p, q, :].T)
for op, ngA, ncA, cpuA, w in (("fprop", ngO, ncO, cpuO,
Q), ("bprop", ngB, ncB.reshape(dimI),
cpuB[:-1, :].reshape(dimI), W),
("update", ngU, ncU.reshape(dimF),
cpuU.reshape(dimF), S)):
print op
assert np.allclose(ngA.get(), cpuA, rtol=0, atol=1e-4)
assert np.allclose(ncA.get(), cpuA, rtol=0, atol=1e-5)
ng.ctx.detach()
del ng
dict(alpha=2.0, beta=3.0),
dict(),
]
for config in configs:
kernelClass, N, C, K, determ, compound, override, convs = config
for conv in convs:
D, H, W, T, R, S, pad_d, pad_h, pad_w, str_d, str_h, str_w = conv
ng.deterministic = determ
layer = nc.conv_layer(np.float64,
N, C, K, D, H, W, T, R, S,
pad_d, pad_h, pad_w,
str_d, str_h, str_w)
(M, P, Q) = layer.MPQ
if kernelClass in (FpropCuda, BpropCuda, UpdateCuda):
dtypes = (np.float32,)
else:
dtypes = (np.float32, np.float16)
for dtype in (dtypes):
ng.scratch_buffer_reset()
if override is None:
kernel = kernelClass(ng, np.dtype(dtype),
dict(alpha=2.0, beta=3.0),
dict(),
]
for config in configs:
kernelClass, N, C, K, determ, compound, override, convs = config
for conv in convs:
D, H, W, T, R, S, pad_d, pad_h, pad_w, str_d, str_h, str_w = conv
ng.deterministic = determ
layer = nc.conv_layer(np.float64,
N, C, K, D, H, W, T, R, S,
pad_d, pad_h, pad_w,
str_d, str_h, str_w)
(M, P, Q) = layer.MPQ
if kernelClass in (FpropCuda, BpropCuda, UpdateCuda):
dtypes = (np.float32,)
else:
dtypes = (np.float32, np.float16)
for dtype in (dtypes):
ng.scratch_buffer_reset()
if override is None:
kernel = kernelClass(ng, np.dtype(dtype),
def test_conv_layer(fargs_tests):
dtype = np.float32
ng = NervanaGPU(stochastic_round=False, bench=True)
N, C, K = fargs_tests[0]
D, H, W = fargs_tests[1]
T, R, S = fargs_tests[2]
padding_d, padding_h, padding_w = 0, 1, 1
strides_d, strides_h, strides_w = 1, 1, 1
conv_ng = ng.conv_layer(
dtype,
N, C, K,
D, H, W,
T, R, S,
padding_d, padding_h, padding_w,
strides_d, strides_h, strides_w)
nc = NervanaCPU()
conv_nc = nc.conv_layer(
dtype,
N, C, K,
D, H, W,
T, R, S,
padding_d, padding_h, padding_w,
strides_d, strides_h, strides_w)
assert conv_nc.dimI == conv_ng.dimI
assert conv_nc.dimF == conv_ng.dimF
assert conv_nc.dimO == conv_ng.dimO
assert conv_nc.M == conv_ng.M
dimI = conv_ng.dimI
dimF = conv_ng.dimF
dimO = conv_ng.dimO
# cpu input arrays
cpuI = np.random.uniform(-0.8, 0.8, slicable(dimI, 1)).astype(np.float32)
cpuF = np.random.uniform(0.0, 0.3, slicable(dimF)).astype(np.float32)
cpuE = np.random.uniform(-0.2, 0.2, dimO).astype(np.float32)
# zero pad the last row of cpu input for the sake of numpy
cpuI[-1, :] = 0.0
# =======GPU and CPU==========
beI = cpuI[:-1, :].reshape(dimI)
beF = cpuF.reshape(dimF)
beE = cpuE
start_gpu = default_timer()
ngO, ngB, ngU = run_backend_conv(ng, conv_ng, beI, beF, beE, dtype)
end_gpu = default_timer()
start_cpu = default_timer()
ncO, ncB, ncU = run_backend_conv(nc, conv_nc, beI, beF, beE, dtype)
end_cpu = default_timer()
print("gputime: %s, cputime %s" %
(end_gpu - start_gpu, end_cpu - start_cpu))
# ======numpy===========
# cpu output arrays
cpuO = np.zeros(dimO, dtype=dtype)
cpuB = np.zeros(slicable(dimI, 1), dtype=dtype)
cpuU = np.zeros(slicable(dimF), dtype=dtype)
D, H, W = conv_nc.DHW
T, R, S = conv_nc.TRS
M, P, Q = conv_nc.MPQ
pad_d, pad_h, pad_w = conv_nc.padding
str_d, str_h, str_w = conv_nc.strides
for m in range(M):
mt = m * str_d - pad_d
for p in range(P):
pr = p * str_h - pad_h
for q in range(Q):
qs = q * str_w - pad_w
idx = pixel_indices(conv_nc, mt, pr, qs)
cpuO[:, m, p, q, :] = np.dot(cpuF.T, cpuI[idx, :])
cpuB[idx, :] += np.dot(cpuF, cpuE[:, m, p, q, :])
cpuU += np.dot(cpuI[idx, :], cpuE[:, m, p, q, :].T)
for op, ngA, ncA, cpuA, w in (
("fprop", ngO, ncO, cpuO, Q),
("bprop", ngB, ncB.reshape(dimI), cpuB[:-1, :].reshape(dimI), W),
("update", ngU, ncU.reshape(dimF), cpuU.reshape(dimF), S)):
print(op)
assert np.allclose(ngA.get(), cpuA, rtol=0, atol=1e-4)
assert np.allclose(ncA.get(), cpuA, rtol=0, atol=1e-4)
del ng
del nc
def test_conv_layer(fargs_tests, device_id):
dtype = np.float32
ng = NervanaGPU(stochastic_round=False, bench=True, device_id=device_id)
N, C, K = fargs_tests[0]
D, H, W = fargs_tests[1]
T, R, S = fargs_tests[2]
padding_d, padding_h, padding_w = fargs_tests[3]
strides_d, strides_h, strides_w = fargs_tests[4]
conv_ng = ng.conv_layer(dtype, N, C, K, D, H, W, T, R, S, padding_d,
padding_h, padding_w, strides_d, strides_h,
strides_w)
nc = NervanaCPU()
conv_nc = nc.conv_layer(dtype, N, C, K, D, H, W, T, R, S, padding_d,
padding_h, padding_w, strides_d, strides_h,
strides_w)
assert conv_nc.dimI == conv_ng.dimI
assert conv_nc.dimF == conv_ng.dimF
assert conv_nc.dimO == conv_ng.dimO
assert conv_nc.M == conv_ng.M
dimI = conv_ng.dimI
dimF = conv_ng.dimF
dimO = conv_ng.dimO
# cpu input arrays
cpuI = np.random.uniform(-0.8, 0.8, slicable(dimI, 1)).astype(np.float32)
cpuF = np.random.uniform(0.0, 0.3, slicable(dimF)).astype(np.float32)
cpuE = np.random.uniform(-0.2, 0.2, dimO).astype(np.float32)
# zero pad the last row of cpu input for the sake of numpy
cpuI[-1, :] = 0.0
# =======GPU and CPU==========
beI = cpuI[:-1, :].reshape(dimI)
beF = cpuF.reshape(dimF)
beE = cpuE
start_gpu = default_timer()
ngO, ngB, ngU = run_backend_conv(ng, conv_ng, beI, beF, beE, dtype)
end_gpu = default_timer()
start_cpu = default_timer()
ncO, ncB, ncU = run_backend_conv(nc, conv_nc, beI, beF, beE, dtype)
end_cpu = default_timer()
print("gputime: %s, cputime %s" %
(end_gpu - start_gpu, end_cpu - start_cpu))
# ======numpy===========
# cpu output arrays
cpuO = np.zeros(dimO, dtype=dtype)
cpuB = np.zeros(slicable(dimI, 1), dtype=dtype)
cpuU = np.zeros(slicable(dimF), dtype=dtype)
D, H, W = conv_nc.DHW
T, R, S = conv_nc.TRS
M, P, Q = conv_nc.MPQ
pad_d, pad_h, pad_w = conv_nc.padding
str_d, str_h, str_w = conv_nc.strides
for m in range(M):
mt = m * str_d - pad_d
for p in range(P):
pr = p * str_h - pad_h
for q in range(Q):
qs = q * str_w - pad_w
idx = pixel_indices(conv_nc, mt, pr, qs)
cpuO[:, m, p, q, :] = np.dot(cpuF.T, cpuI[idx, :])
cpuB[idx, :] += np.dot(cpuF, cpuE[:, m, p, q, :])
cpuU += np.dot(cpuI[idx, :], cpuE[:, m, p, q, :].T)
for op, ngA, ncA, cpuA, w in (("fprop", ngO, ncO, cpuO,
Q), ("bprop", ngB, ncB.reshape(dimI),
cpuB[:-1, :].reshape(dimI), W),
("update", ngU, ncU.reshape(dimF),
cpuU.reshape(dimF), S)):
print(op)
ncAnp = ncA.get().astype(np.float32)
ngAnp = ngA.get().astype(np.float32)
ncdif = cpuA - ncAnp
ngdif = cpuA - ngAnp
maxval = abs(cpuA).max()
ncmaxdif = abs(ncdif).max()
ngmaxdif = abs(ngdif).max()
ncRatio = ncmaxdif / maxval
ngRatio = ngmaxdif / maxval
assert ncRatio < 1e-5
assert ngRatio < 1e-5
assert allclose_with_out(ncA.get(), cpuA, rtol=0, atol=1e-4)
assert allclose_with_out(ngA.get(), cpuA, rtol=0, atol=1e-3)
del ng
del nc