def test():
"""
Execute the tests
Returns: (n_total, n_success)
"""
# TODO remove this line
return 0, 0
logger = TestLogger(TESTNAME)
# TODO change this loop to do more suitable tests
for size in [128, 1024, 4096]:
# generate the stimuli
vecA, vecB, result = gen_stimuli(size)
# prepare header file
# TODO generate the header file
header = HeaderFile("test_stimuli.h")
header.add(HeaderConstant("LENGTH", size))
header.add(HeaderArray("vecA", "int8_t", vecA))
header.add(HeaderArray("vecB", "int8_t", vecB))
header.write()
# compile and run
os.system("make clean all run > {}".format(RESULT_FILE))
# parse output
result = parse_output(RESULT_FILE)
# log the result
# TODO add meaningful name for the subcase
logger.show_subcase_result("size {:4}".format(size), result)
# return summary
return logger.summary()
def test():
"""
Execute the tests
Returns: (n_total, n_success)
"""
logger = TestLogger(TESTNAME, show_title=False)
for simd, parallel in [(False, False), (True, False), (True, True)]:
# generate makefile
mkf = Makefile()
mkf.add_fc_test_source("test.c")
mkf.add_cl_test_source("cluster.c")
mkf.add_cl_prog_source("net/layer3.c")
mkf.add_cl_prog_source("net/net.c")
mkf.add_cl_prog_source("func/transform.c")
mkf.add_cl_prog_source("func/conv.c")
if not simd:
mkf.add_define("NO_SIMD")
if parallel:
mkf.add_define("PARALLEL")
mkf.write()
random_input = False
# generate the stimuli
_, x_align, _, y_exp_align = gen_stimuli(random_input)
# prepare header file
header = HeaderFile("test_stimuli.h")
header.add(HeaderArray("x_vec", "int8_t", x_align.ravel()))
header.add(HeaderArray("y_exp_vec", "int8_t", y_exp_align.ravel()))
header.write()
# compile and run
os.system("make clean all run > {}".format(RESULT_FILE))
# parse output
result = parse_output(RESULT_FILE)
# log the result
options = []
if simd:
options.append("simd")
if parallel:
options.append("parallel")
subcase_name = "layer 3 "
if options:
subcase_name += " + ".join(options)
else:
subcase_name += "naive"
logger.show_subcase_result(subcase_name, result)
# return summary
return logger.summary()
def test():
"""
Execute the tests
Returns: (n_total, n_success)
"""
logger = TestLogger(TESTNAME, show_title=False)
for parallel in [False, True]:
# generate makefile
mkf = Makefile()
mkf.add_fc_test_source("test.c")
mkf.add_cl_test_source("cluster.c")
mkf.add_cl_prog_source("net/layer1.c")
mkf.add_cl_prog_source("net/net.c")
mkf.add_cl_prog_source("func/flip.c")
if parallel:
mkf.add_define("PARALLEL")
mkf.write()
# generate the stimuli
_, x_align, y_exp, y_exp_align = gen_stimuli()
# prepare header file
header = HeaderFile("test_stimuli.h")
header.add(HeaderArray("x_vec", "int8_t", x_align.ravel(),
const=False))
header.add(
HeaderArray("y_exp", "int8_t", y_exp_align.ravel(), const=False))
header.write()
# compile and run
os.system("make clean all run > {}".format(RESULT_FILE))
# parse output
result = parse_output(RESULT_FILE)
# log the result
subcase_name = "Layer 1 flip "
if parallel:
subcase_name += "parallel"
else:
subcase_name += "naive"
logger.show_subcase_result(subcase_name, result)
# return summary
return logger.summary()
def test():
"""
Execute the tests
Returns: (n_total, n_success)
"""
logger = TestLogger(TESTNAME, show_title=False)
for simd in [False, True]:
# generate makefile
mkf = Makefile()
mkf.add_fc_test_source("test.c")
mkf.add_cl_test_source("cluster.c")
mkf.add_cl_prog_source("net/layer5.c")
mkf.add_cl_prog_source("net/net.c")
mkf.add_cl_prog_source("func/transform.c")
mkf.add_cl_prog_source("func/dotp.c")
mkf.write()
if not simd:
mkf.add_define("NO_SIMD")
random_input = False
# generate the stimuli
_, x_align, _, y_exp_align = gen_stimuli(random_input)
# prepare header file
header = HeaderFile("test_stimuli.h")
header.add(HeaderArray("x_vec", "int8_t", x_align.ravel()))
header.add(HeaderArray("y_exp_vec", "int8_t", y_exp_align.ravel()))
header.write()
# compile and run
os.system("make clean all run > {}".format(RESULT_FILE))
# parse output
result = parse_output(RESULT_FILE)
# log the result
subcase_name = "Layer 5 "
if simd:
subcase_name += "simd"
else:
subcase_name += "naive"
logger.show_subcase_result(subcase_name, result)
# return summary
return logger.summary()
def gen_input_header(net, net_params, data, output_file):
# only allow nets with 255 levels
assert net_params["weightInqNumLevels"] == 255
assert net_params["actSTENumLevels"] == 255
# extract and prepare the input data
scale_factor = convert.ste_quant(net, "quant1")
input_quant = F.quantize_to_int(data, scale_factor)
input_quant_align = align_array(input_quant)
# also generate the padded input vector
_, C, T = input_quant.shape
T_pad = T + 63
assert T_pad % 4 == 0
input_pad = np.zeros((C, T_pad), dtype=np.int)
input_pad[:, 31:31 + T] = input_quant[0]
# generate the header file
header = HeaderFile(output_file, "__INPUT_H__", with_c=True)
header.add(HeaderArray("input_data", "int8_t", input_quant_align.ravel()))
header.add(HeaderArray("input_data_pad", "int8_t", input_pad.ravel()))
header.write()
def test():
"""
Execute the tests
Returns: (n_total, n_success)
"""
logger = TestLogger(TESTNAME)
for size_a, size_b in [(155, 16), (1021, 63), (1024, 63), (1188, 64),
(4096, 128)]:
for conv_version in [2, 3]:
# generate makefile
mkf = Makefile()
mkf.add_fc_test_source("test.c")
mkf.add_cl_test_source("cluster.c")
mkf.add_cl_prog_source("func/xcorr.c")
mkf.add_define("CONV_VERSION", conv_version)
mkf.write()
# generate the stimuli
vecA, vecB, vecExp = gen_stimuli(size_a, size_b)
# prepare header file
header = HeaderFile("test_stimuli.h")
header.add(HeaderConstant("LENGTH_A", size_a))
header.add(HeaderConstant("LENGTH_B", size_b))
header.add(HeaderConstant("LENGTH_RES", len(vecExp)))
header.add(HeaderArray("vecA", "int8_t", vecA))
header.add(HeaderArray("vecB", "int8_t", vecB))
header.add(HeaderArray("vecExp", "int32_t", vecExp))
header.write()
# compile and run
os.system("make clean all run > {}".format(RESULT_FILE))
# parse output
result = parse_output(RESULT_FILE)
casename = "V{}, {}x{}".format(conv_version, size_a, size_b)
# log the result
logger.show_subcase_result(casename, result)
# return summary
return logger.summary()
def test():
"""
Execute the tests
Returns: (n_total, n_success)
"""
logger = TestLogger(TESTNAME)
# generate makefile
mkf = Makefile()
mkf.add_fc_test_source("test.c")
mkf.add_cl_test_source("cluster.c")
mkf.add_cl_prog_source("func/dotp.c")
mkf.write()
for length in [22, 23, 1024, 1025]:
for a_stride, b_stride in [(1, 1), (4, 1), (8, 4)]:
# generate the stimuli
vec_a, vec_b, exp_result = gen_stimuli(length, a_stride, b_stride)
# prepare header file
header = HeaderFile("test_stimuli.h")
header.add(HeaderConstant("LENGTH", length))
header.add(HeaderConstant("EXP_RESULT", exp_result))
header.add(HeaderConstant("A_STRIDE", a_stride))
header.add(HeaderConstant("B_STRIDE", b_stride))
header.add(HeaderArray("vec_a", "int8_t", vec_a.ravel()))
header.add(HeaderArray("vec_b", "int8_t", vec_b.ravel()))
header.write()
# compile and run
os.system("make clean all run > {}".format(RESULT_FILE))
# parse output
result = parse_output(RESULT_FILE)
# log the result
subcase_name = "length: {}, stride: {}x{}".format(
length, a_stride, b_stride)
logger.show_subcase_result(subcase_name, result)
# return summary
return logger.summary()
def test():
"""
Execute the tests
Returns: (n_total, n_success)
"""
logger = TestLogger(TESTNAME)
for no_intermediate_scale, duplicate_featuremap in [(False, False),
(True, False),
(True, True)]:
# generate makefile
# mkf = Makefile(opt_level=2 if duplicate_featuremap else 3)
mkf = Makefile(opt_level=3)
mkf.add_fc_test_source("test.c")
mkf.add_cl_test_source("cluster.c")
mkf.add_cl_prog_source("net/fused_layer_1_2.c")
mkf.add_cl_prog_source("net/net.c")
mkf.add_cl_prog_source("func/conv.c")
mkf.add_cl_prog_source("func/xcorr.c")
mkf.add_cl_prog_source("func/dotp.c")
mkf.add_cl_prog_source("func/transform.c")
mkf.add_define("PARALLEL")
mkf.add_define("INTRINSIC_SCALE")
mkf.add_define("CROSS_CORRELATE")
mkf.add_define("FUSE_LAYERS")
mkf.add_define("DEFAULT_DIM")
if no_intermediate_scale:
mkf.add_define("NO_INTERMEDIATE_SCALE")
if duplicate_featuremap:
mkf.add_define("DUPLICATE_FEATUREMAP")
mkf.write()
random_input = False
# generate the stimuli
_, x_align, _, y_exp_align = gen_stimuli(random_input,
no_intermediate_scale,
duplicate_featuremap)
# prepare header file
header = HeaderFile("test_stimuli.h")
header.add(HeaderArray("x_vec", "int8_t", x_align.ravel()))
header.add(HeaderArray("y_exp_vec", "int8_t", y_exp_align.ravel()))
header.write()
# compile and run
os.system("make clean all run > {}".format(RESULT_FILE))
# parse output
result = parse_output(RESULT_FILE)
# log the result
options = []
if no_intermediate_scale:
options.append("no scale")
if duplicate_featuremap:
options.append("dup inp")
subcase_name = "Fused Layer 1+2 "
if options:
subcase_name += "; ".join(options)
logger.show_subcase_result(subcase_name, result)
# return summary
return logger.summary()
def gen_net_header(net_file, config_file, output_file):
# load network
net = np.load(net_file)
# load configuration file
with open(config_file, "r") as _f:
config = json.load(_f)
# we only need the network parameters
net_params = config["indiv"]["net"]["params"]
# only allow nets with 255 levels
assert net_params["weightInqNumLevels"] == 255
assert net_params["actSTENumLevels"] == 255
assert net_params["F2"] % 4 == 0
assert net_params["N"] == 4
# prepare params
if net_params["F2"] is None:
net_params["F2"] = net_params["F1"] * net_params["D"]
# only allow F2 = F1 * D
assert net_params["F2"] == net_params["F1"] * net_params["D"]
# start the header file
header = HeaderFile(output_file, "__NET_NET_H__", with_c=True)
# add network dimensions
header.add(HeaderComment("Network Dimensions", blank_line=False))
header.add(HeaderConstant("NET_F1", net_params["F1"], blank_line=False))
header.add(HeaderConstant("NET_F2", net_params["F2"], blank_line=False))
header.add(HeaderConstant("NET_D", net_params["D"], blank_line=False))
header.add(HeaderConstant("NET_C", net_params["C"], blank_line=False))
header.add(
HeaderConstant("NET_C_ALIGN",
align_array_size(net_params["C"]),
blank_line=False))
header.add(HeaderConstant("NET_T", net_params["T"], blank_line=False))
header.add(
HeaderConstant("NET_T_ALIGN",
align_array_size(net_params["T"]),
blank_line=False))
header.add(HeaderConstant("NET_T8", net_params["T"] // 8,
blank_line=False))
header.add(
HeaderConstant("NET_T8_ALIGN",
align_array_size(net_params["T"] // 8),
blank_line=False))
header.add(
HeaderConstant("NET_T64", (net_params["T"] // 8) // 8,
blank_line=False))
header.add(
HeaderConstant("NET_T64_ALIGN",
align_array_size((net_params["T"] // 8) // 8),
blank_line=False))
header.add(HeaderConstant("NET_N", net_params["N"], blank_line=True))
# Layer 1
input_scale = convert.ste_quant(net, "quant1")
weight, weight_scale = convert.inq_conv2d(net, "conv1")
weight = weight.reshape(net_params["F1"], 64)
weight_reverse, _ = convert.inq_conv2d(net, "conv1", store_reversed=True)
weight_reverse = weight_reverse.reshape(net_params["F1"], 64)
bn_scale, bn_offset = convert.batch_norm(net, "batch_norm1")
output_scale = convert.ste_quant(net, "quant2")
factor, offset = convert.div_factor_batch_norm(input_scale, weight_scale,
output_scale, bn_scale,
bn_offset)
# add padding to the weight vector of 4
if WEIGHT_L1_PAD > 0:
weight_reverse_pad = np.zeros((net_params["F1"], 64 + WEIGHT_L1_PAD))
weight_reverse_pad[:, :-WEIGHT_L1_PAD] = weight_reverse
else:
weight_reverse_pad = weight_reverse
header.add(
HeaderComment(
"Layer 1\n"
"=======\n"
"Convolution + BN\n\n"
"Input: [C, T]\n"
"Weight: [F1, 64]\n"
"Output: [F1, C, T]",
mode="/*"))
header.add(HeaderConstant("NET_L1_PAD_START", 31))
header.add(HeaderConstant("NET_L1_PAD_END", 32))
header.add(
HeaderConstant("NET_L1_PAD_INPUT_LEN", net_params["T"] + 31 + 32))
header.add(
HeaderConstant("NET_L1_PAD_INPUT_LEN_ALIGN",
align_array_size(net_params["T"] + 31 + 32)))
header.add(HeaderArray("net_l1_factor", "int32_t", factor.ravel()))
header.add(HeaderArray("net_l1_offset", "int32_t", offset.ravel()))
header.add(HeaderConstant("NET_L1_WEIGHT_LEN", weight.shape[-1]))
header.add(
HeaderConstant("NET_L1_WEIGHT_LEN_ALIGN",
weight_reverse_pad.shape[-1]))
header.add(HeaderArray("net_l1_weight", "int8_t", weight.ravel()))
header.add(
HeaderArray("net_l1_weight_reverse", "int8_t", weight_reverse.ravel()))
header.add(
HeaderArray("net_l1_weight_reverse_pad", "int8_t",
weight_reverse_pad.ravel()))
# layer2
input_scale = convert.ste_quant(net, "quant2")
weight, weight_scale = convert.inq_conv2d(net,
"conv2",
store_reversed=True)
bn_scale, bn_offset = convert.batch_norm(net, "batch_norm2")
output_scale = convert.ste_quant(net, "quant3")
factor, offset = convert.div_factor_batch_norm(input_scale,
weight_scale,
output_scale,
bn_scale,
bn_offset,
pool=8)
weight = weight.reshape(net_params["F2"], net_params["C"])
weight = align_array(weight)
header.add(
HeaderComment(
"Layer 2\n"
"=======\n"
"Convolution + BN + ReLU + Pooling\n\n"
"Input: [F1, C, T]\n"
"Weight: [F2, C] (aligned to [F2, 24]\n"
"Output: [F2, T // 8]",
mode="/*"))
header.add(HeaderArray("net_l2_factor", "int32_t", factor.ravel()))
header.add(HeaderArray("net_l2_offset", "int32_t", offset.ravel()))
header.add(HeaderConstant("NET_L2_WEIGHT_LEN", weight.shape[-1]))
header.add(HeaderArray("net_l2_weight", "int8_t", weight.ravel()))
header.add(HeaderArray("net_l2_weight_32", "int32_t", weight.ravel()))
# layer3
input_scale = convert.ste_quant(net, "quant3")
weight, weight_scale = convert.inq_conv2d(net, "sep_conv1")
output_scale = convert.ste_quant(net, "quant4")
factor = convert.div_factor(input_scale, weight_scale, output_scale)
weight = weight.reshape(net_params["F2"], 16)
header.add(
HeaderComment(
"Layer 3\n"
"=======\n"
"Convolution\n\n"
"Input: [F2, T // 8]\n"
"Weight: [F2, 16]\n"
"Output: [F2, T // 8]",
mode="/*",
blank_line=False))
header.add(HeaderConstant("NET_L3_PAD_START", 7))
header.add(HeaderConstant("NET_L3_PAD_END", 8))
header.add(
HeaderConstant("NET_L3_PAD_INPUT_LEN", net_params["T"] // 8 + 7 + 8))
header.add(
HeaderConstant("NET_L3_PAD_INPUT_LEN_ALIGN",
align_array_size(net_params["T"] // 8 + 7 + 8)))
header.add(HeaderConstant("NET_L3_FACTOR", factor))
header.add(HeaderConstant("NET_L3_WEIGHT_LEN", weight.shape[-1]))
header.add(HeaderArray("net_l3_weight", "int8_t", weight.ravel()))
# layer4
input_scale = convert.ste_quant(net, "quant4")
weight, weight_scale = convert.inq_conv2d(net, "sep_conv2")
output_scale = convert.ste_quant(net, "quant5")
bn_scale, bn_offset = convert.batch_norm(net, "batch_norm3")
factor, offset = convert.div_factor_batch_norm(input_scale,
weight_scale,
output_scale,
bn_scale,
bn_offset,
pool=8)
weight = weight.reshape(net_params["F2"], net_params["F2"])
header.add(
HeaderComment(
"Layer 4\n"
"=======\n"
"Convolution + BN + ReLU + Pooling\n\n"
"Input: [F2, T // 8]\n"
"Weight: [F2, F2]\n"
"Output: [F2, T // 64]",
mode="/*"))
header.add(HeaderArray("net_l4_factor", "int32_t", factor.ravel()))
header.add(HeaderArray("net_l4_offset", "int32_t", offset.ravel()))
header.add(HeaderConstant("NET_L4_WEIGHT_LEN", weight.shape[-1]))
header.add(HeaderArray("net_l4_weight", "int8_t", weight.ravel()))
# layer5
input_scale = convert.ste_quant(net, "quant5")
output_scale = convert.ste_quant(net, "quant6")
weight, bias, weight_scale = convert.inq_linear(net, "fc")
weight = weight.reshape(net_params["N"],
net_params["F2"] * (net_params["T"] // 64))
#weight = align_array(weight)
# we want to align, not for the product F2*T//64, but for T//64 itself.
t64 = net_params["T"] // 64
t64_align = align_array_size(t64)
weight_align = np.zeros((net_params["N"], net_params["F2"] * t64_align),
dtype=int)
for i in range(net_params["F2"]):
weight_align[:, i * t64_align:i * t64_align +
t64] = weight[:, i * t64:(i + 1) * t64]
factor = convert.div_factor(input_scale, weight_scale, output_scale)
header.add(
HeaderComment(
"Layer 5\n"
"=======\n"
"Linear Layer (without scaling in the end)\n\n"
"Input: [F2, T // 64]\n"
"Weight: [N, F2 * (T // 64)]\n"
"Bias: [N]\n"
"Output: [N]",
mode="/*"))
header.add(HeaderConstant("NET_L5_FACTOR", factor))
header.add(HeaderArray("net_l5_bias", "int8_t", bias.ravel()))
header.add(HeaderConstant("NET_L5_WEIGHT_LEN", weight_align.shape[-1]))
header.add(HeaderArray("net_l5_weight", "int8_t", weight_align.ravel()))
# store the header file
header.write()
def test():
"""
Execute the tests
Returns: (n_total, n_success)
"""
logger = TestLogger(TESTNAME)
for intrinsic, simd, flip_layers, parallel, stream, xcorr, fuse, no_div, reorder, dup_inp in [
(False, False, False, False, False, False, False, False, False, False),
(True, False, False, False, False, False, False, False, False, False),
(True, True, False, False, False, False, False, False, False, False),
(True, True, True, False, False, False, False, False, False, False),
(True, True, True, True, False, False, False, False, False, False),
(True, True, True, True, True, False, False, False, False, False),
(True, True, True, True, True, True, False, False, False, False),
(True, True, True, True, True, True, True, False, False, False),
(True, True, True, True, True, True, True, True, False, False),
(True, True, True, True, True, True, True, True, True, False),
(True, True, True, True, True, True, True, True, True, True)
]:
# generate makefile
mkf = Makefile()
mkf.add_fc_test_source("test.c")
mkf.add_cl_test_source("cluster.c")
mkf.add_cl_prog_source("net/model.c")
mkf.add_cl_prog_source("net/layer1.c")
mkf.add_cl_prog_source("net/layer2.c")
mkf.add_cl_prog_source("net/layer3.c")
mkf.add_cl_prog_source("net/layer4.c")
mkf.add_cl_prog_source("net/layer5.c")
mkf.add_cl_prog_source("net/fused_layer_1_2.c")
mkf.add_cl_prog_source("net/net.c")
mkf.add_cl_prog_source("func/transform.c")
mkf.add_cl_prog_source("func/dotp.c")
mkf.add_cl_prog_source("func/conv.c")
mkf.add_cl_prog_source("func/flip.c")
mkf.add_cl_prog_source("func/xcorr.c")
if not simd:
mkf.add_define("NO_SIMD")
if flip_layers:
mkf.add_define("FLIP_LAYERS")
if parallel:
mkf.add_define("PARALLEL")
if intrinsic:
mkf.add_define("INTRINSIC_SCALE")
if stream:
mkf.add_define("DMA_STREAM")
if xcorr:
mkf.add_define("CROSS_CORRELATE")
if fuse:
mkf.add_define("FUSE_LAYERS")
if no_div:
mkf.add_define("NO_INTERMEDIATE_SCALE")
if dup_inp:
mkf.add_define("DUPLICATE_FEATUREMAP")
if reorder:
mkf.add_define("REORDER_BN")
mkf.write()
# generate the stimuli
_, x_align, _, y_exp_align = gen_stimuli(no_div=no_div, pad_data=dup_inp, reorder_bn=reorder)
# prepare header file
header = HeaderFile("test_stimuli.h")
header.add(HeaderArray("x_vec", "int8_t", x_align.ravel()))
header.add(HeaderArray("y_exp_vec", "int8_t", y_exp_align.ravel()))
header.write()
# compile and run
os.system("make clean all run > {}".format(RESULT_FILE))
# parse output
result = parse_output(RESULT_FILE)
# skip the naive result
if not flip_layers:
result["1"]["result"] = None
# prepare the case name
subcase_name = "naive"
if intrinsic:
subcase_name = "+ intrinsic scale"
if simd:
subcase_name = "+ SIMD"
if flip_layers:
subcase_name = "+ flip"
if parallel:
subcase_name = "+ parallel"
if stream:
subcase_name = "+ double buffering"
if xcorr:
subcase_name = "+ cross correlations"
if fuse:
subcase_name = "+ fused layer 1+2"
if no_div:
subcase_name = "+ no division after layer 1"
if reorder:
subcase_name = "+ reorder BN"
if dup_inp:
subcase_name = "+ duplicate featuremap"
# log the result
logger.show_subcase_result(subcase_name, result)
# return summary
return logger.summary()
def test():
"""
Execute the tests
Returns: (n_total, n_success)
"""
logger = TestLogger(TESTNAME, show_title=False)
for simd in [False, True]:
for flip_layers in [False, True]:
for parallel in [False, True]:
for dma_stream in [False, True]:
for reorder in [False, True]:
if not simd and (flip_layers or parallel or dma_stream or reorder):
continue
if not flip_layers and (parallel or dma_stream):
# not implemented
continue
if not parallel and dma_stream:
# not implemented
continue
# generate makefile
mkf = Makefile()
mkf.add_fc_test_source("test.c")
mkf.add_cl_test_source("cluster.c")
mkf.add_cl_prog_source("net/layer2.c")
mkf.add_cl_prog_source("net/net.c")
mkf.add_cl_prog_source("func/transform.c")
mkf.add_cl_prog_source("func/dotp.c")
if not simd:
mkf.add_define("NO_SIMD")
if flip_layers:
mkf.add_define("FLIP_LAYERS")
if parallel:
mkf.add_define("PARALLEL")
if dma_stream:
mkf.add_define("DMA_STREAM")
if reorder:
mkf.add_define("REORDER_BN")
mkf.write()
random_input = False
# generate the stimuli
_, x_align, _, y_exp_align = gen_stimuli(random_input, flip_layers, reorder)
# prepare header file
header = HeaderFile("test_stimuli.h")
header.add(HeaderArray("x_vec", "int8_t", x_align.ravel()))
header.add(HeaderArray("y_exp_vec", "int8_t", y_exp_align.ravel()))
header.write()
# compile and run
os.system("make clean all run > {}".format(RESULT_FILE))
# parse output
result = parse_output(RESULT_FILE)
# log the result
subcase_name = "Layer 2 "
options = []
if simd:
options.append("simd")
if flip_layers:
options.append("flip")
if parallel:
options.append("par")
if dma_stream:
options.append("stream")
if reorder:
options.append("reorder")
if options:
subcase_name += "; ".join(options)
else:
subcase_name += "naive"
logger.show_subcase_result(subcase_name, result)
# return summary
return logger.summary()
def test():
"""
Execute the tests
Returns: (n_total, n_success)
"""
logger = TestLogger(TESTNAME)
for intrinsic_conv_scale in [False, True]:
for simd in [False, True]:
for parallel in [False, True]:
for cross_correlate in [False, True]:
if not simd and (parallel or cross_correlate):
continue
# parallel requires intrinsic conv scale
if parallel and not intrinsic_conv_scale:
continue
# not implemented
if cross_correlate and not parallel:
continue
# generate makefile
mkf = Makefile()
mkf.add_fc_test_source("test.c")
mkf.add_cl_test_source("cluster.c")
mkf.add_cl_prog_source("net/layer1.c")
mkf.add_cl_prog_source("net/net.c")
mkf.add_cl_prog_source("func/conv.c")
mkf.add_cl_prog_source("func/xcorr.c")
mkf.add_cl_prog_source("func/transform.c")
if parallel:
mkf.add_define("PARALLEL")
if intrinsic_conv_scale:
mkf.add_define("INTRINSIC_SCALE")
if cross_correlate:
mkf.add_define("CROSS_CORRELATE")
if not simd:
mkf.add_define("NO_SIMD")
mkf.write()
random_input = False
# generate the stimuli
x, y_exp = gen_stimuli(random_input)
x_align = align_array(x)
y_exp_align = align_array(y_exp)
# prepare header file
header = HeaderFile("test_stimuli.h")
header.add(HeaderArray("x_vec", "int8_t", x_align.ravel()))
header.add(
HeaderArray("y_exp_vec", "int8_t",
y_exp_align.ravel()))
header.write()
# compile and run
os.system("make clean all run > {}".format(RESULT_FILE))
# parse output
result = parse_output(RESULT_FILE)
# log the result
options = []
if simd:
options.append("simd")
if parallel:
options.append("par")
if intrinsic_conv_scale:
options.append("intr.s.")
if cross_correlate:
options.append("xcorr")
subcase_name = "Layer 1 "
if options:
subcase_name += "; ".join(options)
else:
subcase_name += "naive"
logger.show_subcase_result(subcase_name, result)
# return summary
return logger.summary()
def test():
"""
Execute the tests
Returns: (n_total, n_success)
"""
logger = TestLogger(TESTNAME)
# generate makefile
mkf = Makefile()
mkf.add_fc_test_source("test.c")
mkf.add_cl_test_source("cluster.c")
mkf.add_cl_prog_source("func/flip.c")
mkf.write()
for outer_len in [16, 17, 18, 19]:
for inner_len in [256, 257, 258, 259]:
# generate the stimuli
stim, exp = gen_stimuli(outer_len, inner_len)
# prepare header file
header = HeaderFile("test_stimuli.h")
header.add(HeaderConstant("OUTER_LEN", outer_len))
header.add(HeaderConstant("OUTER_LEN_ALIGN", align_array_size(outer_len)))
header.add(HeaderConstant("INNER_LEN", inner_len))
header.add(HeaderConstant("INNER_LEN_ALIGN", align_array_size(inner_len)))
header.add(HeaderArray("vec_x", "int8_t", stim))
header.add(HeaderArray("vec_exp", "int8_t", exp))
header.write()
# compile and run
os.system("make clean all run > {}".format(RESULT_FILE))
# parse output
result = parse_output(RESULT_FILE)
# log the result
subcase_name = "{}x{}".format(outer_len, inner_len)
logger.show_subcase_result(subcase_name, result)
# return summary
return logger.summary()
def test():
"""
Execute the tests
Returns: (n_total, n_success)
"""
logger = TestLogger(TESTNAME)
# generate makefile
mkf = Makefile()
mkf.add_fc_test_source("test.c")
mkf.add_cl_test_source("cluster.c")
mkf.add_cl_prog_source("func/transform.c")
mkf.write()
for size in [1024, 1025, 1026, 1027]:
# generate the stimuli
x, y, y_bias = gen_stimuli(size, scale_factor=SCALE_FACTOR, bias=BIAS, max_val=2560)
# prepare header file
header = HeaderFile("test_stimuli.h")
header.add(HeaderConstant("LENGTH", size))
header.add(HeaderScalar("div_factor", "int32_t", SCALE_FACTOR))
header.add(HeaderScalar("bias", "int32_t", BIAS))
header.add(HeaderArray("vec_x", "int32_t", x))
header.add(HeaderArray("vec_exp", "int8_t", y))
header.add(HeaderArray("vec_exp_bias", "int8_t", y_bias))
header.write()
# compile and run
os.system("make clean all run > {}".format(RESULT_FILE))
# parse output
result = parse_output(RESULT_FILE)
# log the result
subcase_name = "n={}".format(size)
logger.show_subcase_result(subcase_name, result)
# return summary
return logger.summary()