Skip to content

Commit

Permalink
Merge pull request #1965 from StanfordAHA/hetero-cgra-lopsided-layout
Browse files Browse the repository at this point in the history 
Hetero cgra lopsided layout
  • Loading branch information
@mcoduoza
mcoduoza authored Dec 11, 2024
2 parents b7a3cbd + a07df0a commit b8600d1
Show file tree
Hide file tree
Showing 3 changed files with 67 additions and 27 deletions.
90 changes: 65 additions & 25 deletions aha/util/regress.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -26,6 +26,7 @@ def add_subparser(subparser):
parser.add_argument("--unroll", default=1, type=int)
parser.add_argument("--using-matrix-unit", action="store_true")
parser.add_argument("--mu-datawidth", default=16, type=int)
parser.add_argument("--num-fabric-cols-removed", default=0, type=int)
parser.set_defaults(dispatch=dispatch)


Expand Down Expand Up @@ -68,7 +69,7 @@ def buildkite_call(command, env={}, return_output=False, out_file=None):
else:
raise

def gen_garnet(width, height, dense_only=False, using_matrix_unit=False, mu_datawidth=16):
def gen_garnet(width, height, dense_only=False, using_matrix_unit=False, mu_datawidth=16, num_fabric_cols_removed=0):
print("--- Generating Garnet", flush=True)
start = time.time()
if not os.path.exists("/aha/garnet/garnet.v"):
Expand All @@ -93,13 +94,16 @@ def gen_garnet(width, height, dense_only=False, using_matrix_unit=False, mu_data
buildkite_args.append("--using-matrix-unit")
buildkite_args.append("--mu-datawidth")
buildkite_args.append(str(mu_datawidth))
buildkite_args.append("--give-north-io-sbs")
buildkite_args.append("--num-fabric-cols-removed")
buildkite_args.append(str(num_fabric_cols_removed))

buildkite_call(buildkite_args)

return time.time() - start


def generate_sparse_bitstreams(sparse_tests, width, height, seed_flow, data_tile_pairs, kernel_name, opal_workaround=False, unroll=1, using_matrix_unit=False):
def generate_sparse_bitstreams(sparse_tests, width, height, seed_flow, data_tile_pairs, kernel_name, opal_workaround=False, unroll=1, using_matrix_unit=False, num_fabric_cols_removed=0):
if len(sparse_tests) == 0:
return 0

Expand Down Expand Up @@ -133,6 +137,9 @@ def generate_sparse_bitstreams(sparse_tests, width, height, seed_flow, data_tile
build_tb_cmd.append("--opal-workaround")
if using_matrix_unit:
build_tb_cmd.append("--using-matrix-unit")
build_tb_cmd.append("--give-north-io-sbs")
build_tb_cmd.append("--num-fabric-cols-removed")
build_tb_cmd.append(str(num_fabric_cols_removed))
buildkite_call(
build_tb_cmd,
env=env_vars,
Expand Down Expand Up @@ -166,6 +173,9 @@ def generate_sparse_bitstreams(sparse_tests, width, height, seed_flow, data_tile
build_tb_cmd.append("--opal-workaround")
if using_matrix_unit:
build_tb_cmd.append("--using-matrix-unit")
build_tb_cmd.append("--give-north-io-sbs")
build_tb_cmd.append("--num-fabric-cols-removed")
build_tb_cmd.append(str(num_fabric_cols_removed))
buildkite_call(
build_tb_cmd,
env=env_vars,
Expand Down Expand Up @@ -212,15 +222,16 @@ def format_concat_tiles(test, data_tile_pairs, kernel_name, pipeline_num=32, unr
return all_tiles, num_list


def test_sparse_app(testname, seed_flow, data_tile_pairs, pipeline_num_l=None, opal_workaround=False, test="", test_dataset_runtime_dict=None, using_matrix_unit=False, cgra_height=32, mu_datawidth=16):
def test_sparse_app(testname, seed_flow, data_tile_pairs, pipeline_num_l=None, opal_workaround=False, test="", test_dataset_runtime_dict=None, using_matrix_unit=False, cgra_height=32, mu_datawidth=16, num_fabric_cols_removed=0):
if test == "":
test = testname

print(f"--- {test}")

env_vars = {"PYTHONPATH": "/aha/garnet/"}
if using_matrix_unit:
env_vars["WEST_IN_IO_SIDES"] = "1"
if num_fabric_cols_removed == 0:
env_vars["WEST_IN_IO_SIDES"] = "1"
env_vars["USING_MATRIX_UNIT"] = "1"
env_vars["OC_0"] = str(2*cgra_height)
env_vars["MU_DATAWIDTH"] = str(mu_datawidth)
Expand Down Expand Up @@ -294,7 +305,7 @@ def test_sparse_app(testname, seed_flow, data_tile_pairs, pipeline_num_l=None, o
return 0, 0, time_test


def test_dense_app(test, width, height, env_parameters, extra_args, layer=None, dense_only=False, use_fp=False, using_matrix_unit=False, cgra_height=32, mu_datawidth=16):
def test_dense_app(test, width, height, env_parameters, extra_args, layer=None, dense_only=False, use_fp=False, using_matrix_unit=False, cgra_height=32, mu_datawidth=16, num_fabric_cols_removed=0):
env_parameters = str(env_parameters)
testname = layer if layer is not None else test
print(f"--- {testname}")
Expand Down Expand Up @@ -342,7 +353,13 @@ def test_dense_app(test, width, height, env_parameters, extra_args, layer=None,

if using_matrix_unit:
buildkite_args.append("--using-matrix-unit")
env_vars["WEST_IN_IO_SIDES"] = "1"
buildkite_args.append("--give-north-io-sbs")
buildkite_args.append("--num-fabric-cols-removed")
buildkite_args.append(str(num_fabric_cols_removed))

if num_fabric_cols_removed == 0:
env_vars["WEST_IN_IO_SIDES"] = "1"

env_vars["USING_MATRIX_UNIT"] = "1"
env_vars["OC_0"] = str(2*cgra_height)
env_vars["MU_DATAWIDTH"] = str(mu_datawidth)
Expand All @@ -362,7 +379,7 @@ def test_dense_app(test, width, height, env_parameters, extra_args, layer=None,
return time_compile, time_map, time_test


def test_hardcoded_dense_app(test, width, height, env_parameters, extra_args, layer=None, dense_only=False, using_matrix_unit=False, cgra_height=32, mu_datawidth=16):
def test_hardcoded_dense_app(test, width, height, env_parameters, extra_args, layer=None, dense_only=False, using_matrix_unit=False, cgra_height=32, mu_datawidth=16, num_fabric_cols_removed=0):
env_parameters = str(env_parameters)
testname = layer if layer is not None else test
print(f"--- {testname}")
Expand Down Expand Up @@ -431,7 +448,11 @@ def test_hardcoded_dense_app(test, width, height, env_parameters, extra_args, la

if using_matrix_unit:
buildkite_args.append("--using-matrix-unit")
env_vars["WEST_IN_IO_SIDES"] = "1"
buildkite_args.append("--give-north-io-sbs")

if num_fabric_cols_removed == 0:
env_vars["WEST_IN_IO_SIDES"] = "1"

env_vars["USING_MATRIX_UNIT"] = "1"
env_vars["OC_0"] = str(2*cgra_height)
env_vars["MU_DATAWIDTH"] = str(mu_datawidth)
Expand All @@ -454,8 +475,13 @@ def dispatch(args, extra_args=None):
pipeline_num = args.pipeline_num
using_matrix_unit = args.using_matrix_unit
mu_datawidth = args.mu_datawidth
num_fabric_cols_removed = args.num_fabric_cols_removed
unroll = args.unroll

# Can only remove col that is multiple of 4
assert num_fabric_cols_removed % 4 == 0, "ERROR: Number of cols removed must be a multiple of 4"
assert num_fabric_cols_removed <= 8, "ERROR: Removing more than 8 columns is not supported yet. Hardware modifications may be necessary to proceed."

# Preserve backward compatibility
if args.config == "daily": args.config = "pr_aha" # noqa
if args.config == "pr": args.config = "pr_submod" # noqa
Expand Down Expand Up @@ -499,7 +525,7 @@ def dispatch(args, extra_args=None):

print(f"--- Running regression: {args.config}", flush=True)
info = []
t = gen_garnet(width, height, dense_only=False, using_matrix_unit=using_matrix_unit, mu_datawidth=mu_datawidth)
t = gen_garnet(width, height, dense_only=False, using_matrix_unit=using_matrix_unit, mu_datawidth=mu_datawidth, num_fabric_cols_removed=num_fabric_cols_removed)
info.append(["garnet with sparse and dense", t])

data_tile_pairs = []
Expand All @@ -526,16 +552,19 @@ def dispatch(args, extra_args=None):
print("HERE ARE THE DATA TILE PAIRS!")
print(data_tile_pairs)

generate_sparse_bitstreams(sparse_tests, width, height, seed_flow, data_tile_pairs, kernel_name, opal_workaround=args.opal_workaround, unroll=unroll, using_matrix_unit=using_matrix_unit)
generate_sparse_bitstreams(sparse_tests, width, height, seed_flow, data_tile_pairs, kernel_name,
opal_workaround=args.opal_workaround, unroll=unroll, using_matrix_unit=using_matrix_unit, num_fabric_cols_removed=num_fabric_cols_removed)

for test in sparse_tests:
if use_pipeline:
assert (not seed_flow), "Pipeline mode is not supported with seed flow"
tile_pairs, pipeline_num_l = format_concat_tiles(test, data_tile_pairs, kernel_name, pipeline_num, unroll)
t0, t1, t2 = test_sparse_app(test, seed_flow, tile_pairs, pipeline_num_l, opal_workaround=args.opal_workaround, test_dataset_runtime_dict=test_dataset_runtime_dict, using_matrix_unit=using_matrix_unit, cgra_height=height, mu_datawidth=mu_datawidth)
t0, t1, t2 = test_sparse_app(test, seed_flow, tile_pairs, pipeline_num_l, opal_workaround=args.opal_workaround, test_dataset_runtime_dict=test_dataset_runtime_dict,
using_matrix_unit=using_matrix_unit, cgra_height=height, mu_datawidth=mu_datawidth, num_fabric_cols_removed=num_fabric_cols_removed)
info.append([test + "_glb", t0 + t1 + t2, t0, t1, t2])
else:
t0, t1, t2 = test_sparse_app(test, seed_flow, data_tile_pairs, opal_workaround=args.opal_workaround, test_dataset_runtime_dict=test_dataset_runtime_dict, using_matrix_unit=using_matrix_unit, cgra_height=height, mu_datawidth=mu_datawidth)
t0, t1, t2 = test_sparse_app(test, seed_flow, data_tile_pairs, opal_workaround=args.opal_workaround, test_dataset_runtime_dict=test_dataset_runtime_dict,
using_matrix_unit=using_matrix_unit, cgra_height=height, mu_datawidth=mu_datawidth, num_fabric_cols_removed=num_fabric_cols_removed)
info.append([test + "_glb", t0 + t1 + t2, t0, t1, t2])

# remove the generated collateral for tiles that passed to avoid overrunning the disk
Expand All @@ -547,46 +576,55 @@ def dispatch(args, extra_args=None):
for dataset, time_value in dataset_runtime_dict.items():
perf_out_file.write(f"{testname} {dataset} {time_value}\n")
else:
generate_sparse_bitstreams(sparse_tests, width, height, seed_flow, data_tile_pairs, kernel_name, opal_workaround=args.opal_workaround, unroll=unroll, using_matrix_unit=using_matrix_unit)
generate_sparse_bitstreams(sparse_tests, width, height, seed_flow, data_tile_pairs, kernel_name,
opal_workaround=args.opal_workaround, unroll=unroll, using_matrix_unit=using_matrix_unit, num_fabric_cols_removed=num_fabric_cols_removed)

for test in sparse_tests:
assert(not use_pipeline), "Pipeline mode is not supported with seed flow"
t0, t1, t2 = test_sparse_app(test, seed_flow, data_tile_pairs, opal_workaround=args.opal_workaround, using_matrix_unit=using_matrix_unit, cgra_height=height, mu_datawidth=mu_datawidth)
t0, t1, t2 = test_sparse_app(test, seed_flow, data_tile_pairs, opal_workaround=args.opal_workaround,
using_matrix_unit=using_matrix_unit, cgra_height=height, mu_datawidth=mu_datawidth, num_fabric_cols_removed=num_fabric_cols_removed)
info.append([test + "_glb", t0 + t1 + t2, t0, t1, t2])

for test in glb_tests:
t0, t1, t2 = test_dense_app(test,
width, height, args.env_parameters, extra_args, using_matrix_unit=using_matrix_unit, cgra_height=height, mu_datawidth=mu_datawidth)
width, height, args.env_parameters, extra_args,
using_matrix_unit=using_matrix_unit, cgra_height=height, mu_datawidth=mu_datawidth, num_fabric_cols_removed=num_fabric_cols_removed)
info.append([test + "_glb", t0 + t1 + t2, t0, t1, t2])

for test in glb_tests_fp:
t0, t1, t2 = test_dense_app(test,
width, height, args.env_parameters, extra_args, use_fp=True, using_matrix_unit=using_matrix_unit, cgra_height=height, mu_datawidth=mu_datawidth)
width, height, args.env_parameters, extra_args, use_fp=True,
using_matrix_unit=using_matrix_unit, cgra_height=height, mu_datawidth=mu_datawidth, num_fabric_cols_removed=num_fabric_cols_removed)
info.append([test + "_glb", t0 + t1 + t2, t0, t1, t2])

for test in resnet_tests:
if "residual" in test:
t0, t1, t2 = test_dense_app("apps/resnet_residual",
width, height, args.env_parameters, extra_args, layer=test, using_matrix_unit=using_matrix_unit, cgra_height=height, mu_datawidth=mu_datawidth)
width, height, args.env_parameters, extra_args, layer=test,
using_matrix_unit=using_matrix_unit, cgra_height=height, mu_datawidth=mu_datawidth, num_fabric_cols_removed=num_fabric_cols_removed)
info.append([test + "_glb", t0 + t1 + t2, t0, t1, t2])
else:
t0, t1, t2 = test_dense_app("apps/resnet_output_stationary",
width, height, args.env_parameters, extra_args, layer=test, using_matrix_unit=using_matrix_unit, cgra_height=height, mu_datawidth=mu_datawidth)
width, height, args.env_parameters, extra_args, layer=test,
using_matrix_unit=using_matrix_unit, cgra_height=height, mu_datawidth=mu_datawidth, num_fabric_cols_removed=num_fabric_cols_removed)
info.append([test + "_glb", t0 + t1 + t2, t0, t1, t2])

for test in resnet_tests_fp:
if "residual" in test:
t0, t1, t2 = test_dense_app("apps/conv2D_residual_fp",
width, height, args.env_parameters, extra_args, layer=test, use_fp=True, using_matrix_unit=using_matrix_unit, cgra_height=height, mu_datawidth=mu_datawidth)
width, height, args.env_parameters, extra_args, layer=test, use_fp=True,
using_matrix_unit=using_matrix_unit, cgra_height=height, mu_datawidth=mu_datawidth, num_fabric_cols_removed=num_fabric_cols_removed)
info.append([test + "_glb", t0 + t1 + t2, t0, t1, t2])
else:
t0, t1, t2 = test_dense_app("apps/conv2D_fp",
width, height, args.env_parameters, extra_args, layer=test, use_fp=True, using_matrix_unit=using_matrix_unit, cgra_height=height, mu_datawidth=mu_datawidth)
width, height, args.env_parameters, extra_args, layer=test, use_fp=True,
using_matrix_unit=using_matrix_unit, cgra_height=height, mu_datawidth=mu_datawidth, num_fabric_cols_removed=num_fabric_cols_removed)
info.append([test + "_glb", t0 + t1 + t2, t0, t1, t2])

for test in hardcoded_dense_tests:
t0, t1, t2 = test_hardcoded_dense_app(test,
width, height, args.env_parameters, extra_args, using_matrix_unit=using_matrix_unit, cgra_height=height, mu_datawidth=mu_datawidth)
width, height, args.env_parameters, extra_args,
using_matrix_unit=using_matrix_unit, cgra_height=height, mu_datawidth=mu_datawidth, num_fabric_cols_removed=num_fabric_cols_removed)
info.append([test + "_glb", t0 + t1 + t2, t0, t1, t2])

if args.include_dense_only_tests:
Expand All @@ -596,27 +634,29 @@ def dispatch(args, extra_args=None):
if os.WEXITSTATUS(exit_status) != 0:
raise RuntimeError(f"Command 'rm /aha/garnet/garnet.v' returned non-zero exit status {os.WEXITSTATUS(exit_status)}.")

t = gen_garnet(width, height, dense_only=True, using_matrix_unit=using_matrix_unit, mu_datawidth=mu_datawidth)
t = gen_garnet(width, height, dense_only=True, using_matrix_unit=using_matrix_unit, mu_datawidth=mu_datawidth, num_fabric_cols_removed=num_fabric_cols_removed)
info.append(["garnet with dense only", t])

num_dense_only_glb_tests = 5
for test_index, test in enumerate(glb_tests):
if test_index == num_dense_only_glb_tests:
break
t0, t1, t2 = test_dense_app(test,
width, height, args.env_parameters, extra_args, dense_only=True, using_matrix_unit=using_matrix_unit, cgra_height=height, mu_datawidth=mu_datawidth)
width, height, args.env_parameters, extra_args, dense_only=True,
using_matrix_unit=using_matrix_unit, cgra_height=height, mu_datawidth=mu_datawidth, num_fabric_cols_removed=num_fabric_cols_removed)
info.append([test + "_glb dense only", t0 + t1 + t2, t0, t1, t2])

for test in resnet_tests:
# residual resnet test is not working with dense only mode
if "residual" not in test:
t0, t1, t2 = test_dense_app("apps/resnet_output_stationary",
width, height, args.env_parameters, extra_args, layer=test, using_matrix_unit=using_matrix_unit, cgra_height=height, mu_datawidth=mu_datawidth)
width, height, args.env_parameters, extra_args, layer=test,
using_matrix_unit=using_matrix_unit, cgra_height=height, mu_datawidth=mu_datawidth, num_fabric_cols_removed=num_fabric_cols_removed)
info.append([test + "_glb dense only", t0 + t1 + t2, t0, t1, t2])

print(f"+++ TIMING INFO", flush=True)
print(tabulate(info, headers=["step", "total", "compile", "map", "test"], floatfmt=".0f"), flush=True)


def gather_tests(tags):
pass
pass
2 changes: 1 addition & 1 deletion canal
Submodule canal updated 3 files
+40 −1 canal/circuit.py
+24 −7 canal/interconnect.py
+83 −15 canal/util.py
2 changes: 1 addition & 1 deletion garnet
Submodule garnet updated 3 files
+46 −7 cgra/util_onyx.py
+17 −7 garnet.py
+11 −2 tests/test_memory_core/build_tb.py

0 comments on commit b8600d1

Please sign in to comment.