def eval_time(ind, module_file):
config = task.config_space.get(ind)
with autotvm.ApplyConfig(config):
with tvm.target.create("llvm -mcpu=core-avx2"):
s, arg_bufs = task.func(*task.args)
func = tvm.runtime.load_module(module_file)
a_np = np.random.uniform(size=(N, N))
b_np = np.random.uniform(size=(N, N, N))
c_np = np.zeros((N, N, N))
ctx = tvm.cpu()
a_tvm = tvm.nd.array(a_np.astype(np.float32), ctx=ctx)
b_tvm = tvm.nd.array(b_np.astype(np.float32), ctx=ctx)
c_tvm = tvm.nd.array(c_np.astype(np.float32), ctx=ctx)
evaluator = func.time_evaluator(
func.entry_name,
ctx,
repeat=10,
number=4,
)
variation = 1
while variation > 0.05:
res = np.array(sorted(evaluator(a_tvm, b_tvm, c_tvm).results)[:-5])
variation = res.std() / res.mean()
#if tuple(arg_bufs[1].shape) == b_tvm.shape:
# res = evaluator(c_tvm, b_tvm, a_tvm)
#else:
# res = evaluator(c_tvm, a_tvm, b_tvm)
return res.mean(), ind
def limited_test(ind):
tic = time.time()
lower_llvm_limit = 1
upper_llvm_limit = 2
lower_asm_limit = 0.5
upper_asm_limit = 2
results = []
config = task.config_space.get(ind)
with autotvm.ApplyConfig(config):
with tvm.target.create("llvm -mcpu=core-avx2"):
s, arg_bufs = task.func(*task.args)
op_func = tvm.build(s, arg_bufs)
build_time = time.time() - tic
ll_source = op_func.get_source()
funcs = ll_source.split('\n\n')
llvm_opint = 0
asm_opint = 0
length = 0
for func in funcs:
if 'fmuladd.v' in func and len(func) > length:
length = len(func)
longest = func
loads = 0
stores = 0
fmas = 0
if length > 0:
lines = longest.split('\n')
for line in lines:
if 'load <' in line:
loads += 1
elif 'store <' in line:
stores += 1
elif 'fmuladd.v8' in line:
fmas += 1
if loads + stores > 0:
llvm_opint = fmas / (loads + stores)
if llvm_opint >= lower_llvm_limit and llvm_opint <= upper_llvm_limit:
tic = time.time()
asm_source = op_func.get_source('asm')
asm_time = time.time() - tic
funcs = asm_source.split(':\n')
length = 0
for func in funcs:
if 'vfmadd' in func and len(func) > length:
length = len(func)
longest = func
moves = 0
fmas = 0
if length > 0:
lines = longest.split('\n')
for line in lines:
if 'vmov' in line and 'ymm' in line:
moves += 1
elif 'vfmadd' in line and 'ymm' in line:
fmas += 1
if '(%r' in line:
moves += 1
if moves > 0:
asm_opint = fmas / moves
if asm_opint >= lower_asm_limit and asm_opint <= upper_asm_limit:
module_file = os.path.join(
'/tmp/',
''.join(random.choice(letters) for i in range(10)) + '.o')
op_func.save(module_file)
return module_file, llvm_opint, asm_opint, ind, build_time, asm_time
return '', llvm_opint, asm_opint, ind, build_time, 0