def pack_smimms_int(x):
smimms_int = {}
for i in range(16):
smimms_int[i] = i
smimms_int[i - 16] = i + 16
smimms_int[pack_unpack('i', 'I', i - 16)] = i + 16
smimms_int[pack_unpack('f', 'I', 2 ** (i - 8))] = i + 32
return smimms_int[x]
def pack(self):
if self.addr_label is not None:
addr = pack_unpack('i', 'I',
(int(self.addr_label) - self.serial - 4) * 8)
elif self.addr is not None:
addr = self.addr
else:
addr = 0
set_link = 1 if self.set_link else 0
msfign = 0b00
return 0 \
| (0b10 << 56) \
| (((addr & ((1 << 24) - 1)) >> 3) << 35) \
| (self.cond_br << 32) \
| ((addr >> 24) << 24) \
| (set_link << 23) \
| (msfign << 21) \
| (self.bdu << 15) \
| (self.ub << 14) \
| (self.bdi << 12) \
| ((self.raddr_a if self.raddr_a is not None else 0) << 6)
def pack_smimms_float(x):
smimms_float = {}
for i in range(16):
# Denormal numbers
smimms_float[pack_unpack('I', 'f', i)] = i
smimms_float[2 ** (i - 8)] = i + 32
return smimms_float[x]
def run(drv,
unif,
src,
dst,
num_qpus,
rows,
cols,
tile_rows,
tile_cols,
subtile_rows,
subtile_cols,
code_offset=0):
code = drv.program(qpu_comatcopy_t,
num_qpus=num_qpus,
tile_rows=tile_rows,
tile_cols=tile_cols,
subtile_rows=subtile_rows,
subtile_cols=subtile_cols,
code_offset=code_offset)
src[:, :] = np.arange(src.size, dtype=src.dtype).reshape(src.shape)
dst[:, :] = np.arange(dst.size, dtype=dst.dtype).reshape(dst.shape)
unif[0] = rows
unif[1] = cols
unif[2] = pack_unpack('f', 'I', 1.)
unif[3] = pack_unpack('f', 'I', 0.)
unif[4] = src.addresses()[0, 0]
unif[5] = cols * 8
unif[6] = dst.addresses()[0, 0]
unif[7] = rows * 8
start = monotonic()
drv.execute(code, unif.addresses()[0], thread=num_qpus)
end = monotonic()
print(f'{num_qpus} QPUs,', f'{rows} x {cols} matrix,',
f'{tile_rows:2} x {tile_cols:2} tile,',
f'{subtile_rows:2} x {subtile_cols:2} subtile:',
f'{end - start} seconds,',
f'{rows * cols * 8 / (end - start) * 1e-6} MB/s')
def block_2x4_params(i, j):
tile_P = P // 2
tile_R = R // 4
return [
tile_P,
Q,
tile_R,
A.addresses()[tile_P * i, 0],
A.strides[0],
B.addresses()[0, tile_R * j],
B.strides[0],
C.addresses()[tile_P * i, tile_R * j],
C.strides[0],
*pack_unpack('f', 'I', [alpha, beta]),
]