Giovanni LB
66f517ab27
Change-Id: I036f45502e6ff18518dcdb954161d09ce3546fe0
[ROCm/rocprofiler commit: b8b870328b]
73 строки
2.6 KiB
Python
73 строки
2.6 KiB
Python
import numpy as np
|
|
import pandas
|
|
|
|
MAX_CU = 16
|
|
MAX_WAVE_SIZE = 64
|
|
MAXIMUM_ATT_HITS = 256*128//MAX_WAVE_SIZE//MAX_CU
|
|
|
|
kernel_name = "vectoradd_att"
|
|
csv_filename = "vadd_" + kernel_name + "_v0.csv"
|
|
output_folder = "/tmp/tests-v2/att"
|
|
|
|
|
|
def test_hitcount(csv):
|
|
hits = {m: True for m in csv['Hitcount'] if m != 0}
|
|
print('hits', hits)
|
|
assert(len(hits) > 0)
|
|
assert(np.max([k for k in hits.keys()]) <= MAXIMUM_ATT_HITS)
|
|
|
|
|
|
def test_addr(csv):
|
|
addrs = np.array([int(addr, 16) for addr in csv['Addr'] if addr != '0x0'])
|
|
print('addrs', addrs)
|
|
assert(addrs.max() - addrs.min() > 32) # 32 bytes is a safe minimum value
|
|
assert(addrs.max() - addrs.min() < 2**24) # Kernels are not anywhere near that large
|
|
|
|
|
|
def test_memory_list(csv):
|
|
inst_list = ' '.join(csv['Instruction'])
|
|
assert('vectoradd_' in inst_list)
|
|
assert('s_load_' in inst_list)
|
|
assert('_store_' in inst_list)
|
|
assert('s_waitcnt' in inst_list)
|
|
assert('v_add' in inst_list)
|
|
assert('global_load' in inst_list or 'buffer_load' in inst_list or 'flat_load' in inst_list)
|
|
|
|
|
|
def test_mean_cycles(csv):
|
|
cycles = np.array([c/float(h) for c, h in zip(csv['Cycles'], csv['Hitcount']) if c != 0])
|
|
print('cycles', cycles)
|
|
assert(cycles.min() < 5) # Waves should have some instructions with very few cycles
|
|
assert(cycles.max() > 100) # s_waitcnt should have a large cost
|
|
assert(cycles.mean() > 1) # Minimum cost per inst is 1
|
|
assert(np.median(cycles) <= 16) # Majority of instructions are not that expensive
|
|
|
|
maxv = int(4*cycles.max()+5)//4
|
|
histogram = np.histogram(cycles, range=[0,maxv], bins=max(maxv//8, 1))[0]
|
|
assert(histogram[0] == np.max(histogram)) # 1~8 cycles should be most common cost
|
|
|
|
|
|
def test_memory_cycles(csv):
|
|
is_memory_op = lambda s: ('waitcnt' in s) or ('_load_' in s) or ('_store_' in s)
|
|
|
|
max_cycles = np.max(csv['Cycles'])
|
|
most_exp_inst = [f for f in csv[csv['Cycles'] == max_cycles]['Instruction']][0]
|
|
print('most_exp_inst', most_exp_inst)
|
|
assert(is_memory_op(most_exp_inst)) # Memory ops should be the most expensive insts
|
|
|
|
memory_ops = [c for s, c in zip(csv['Instruction'],csv['Cycles']) if is_memory_op(s)]
|
|
print('memory_ops', memory_ops) # Memory ops should be more than half the total cycles
|
|
assert(np.sum(memory_ops) > np.sum(csv['Cycles'])*0.5)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
csv = pandas.read_csv(f"{output_folder}/{csv_filename}")
|
|
|
|
test_hitcount(csv)
|
|
test_addr(csv)
|
|
test_memory_list(csv)
|
|
test_mean_cycles(csv)
|
|
test_memory_cycles(csv)
|
|
|
|
print("Test Passed: All ATT correctness tests passed.")
|