Hetero cgra lopsided layout

aha/util/regress.py

@@ -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)
 
 
@@ -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"):
@@ -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
 
@@ -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,
@@ -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,
@@ -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)
@@ -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}")
@@ -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)
@@ -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}")
@@ -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)
@@ -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
@@ -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 = []
@@ -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
@@ -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:
@@ -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