Add baseline comparisons

Author: eagarvey-amd

models/turbine_models/custom_models/torchbench/export.py

@@ -36,19 +36,19 @@
 torchbench_models_dict = {
     # "BERT_pytorch": {
     #     "dim": 128,
-    # },
+    # }, # Dynamo Export Issue
     # "Background_Matting": {
     #     "dim": 16,
-    # },
-    # "LearningToPaint": {
-    #     "dim": 1024,
-    # },
+    # }, # Transpose Bubbling Pattern Failed
+    "LearningToPaint": {
+        "dim": 1024,
+    },
     "alexnet": {
         "dim": 1024,
     },
-    # "densenet121": {
-    #     "dim": 64,
-    # },
+    "densenet121": {
+        "dim": 64,
+    },
     # "hf_Albert": {"dim": 32, "buffer_prefix": "albert"},
     # "hf_Bart": {
     #     "dim": 16,
@@ -131,17 +131,28 @@ def get_runner(tb_dir, tb_args):
     return runner
 
 
-def get_model_and_inputs(model_id, batch_size, tb_dir, tb_args):
+def get_model_and_inputs(model_id, batch_size, tb_dir, tb_args, get_baseline=False):
     runner = get_runner(tb_dir, tb_args)
-    return runner.load_model(
+    _, model_name, model, forward_args, _ = runner.load_model(
         "cuda:0",
         model_id,
         batch_size=batch_size,
     )
-
-
+    match get_baseline:
+        case True:
+            start_t = time.time()
+            res = runner.forward_pass(model, forward_args, collect_outputs=True)
+            baseline = time.time() - start_t
+            return model_name, model, forward_args, res, baseline
+        case False:
+            return model_name, model, forward_args
+
+
+'''
+Imports models from torchbench model tooling, exports them with turbine AOT, and does simple benchmarking.
+'''
 @torch.no_grad()
-def export_torchbench_model(
+def benchmark_torchbench_model(
     model_id,
     tb_dir,
     tb_args,
@@ -159,6 +170,7 @@ def export_torchbench_model(
     input_mlir=None,
     weights_only=False,
     upload_ir=False,
+    compare_vs_eager=False,
 ):
     static_dim = torchbench_models_dict[model_id]["dim"]
     dtype = torch.float16 if precision == "fp16" else torch.float32
@@ -187,9 +199,16 @@ def export_torchbench_model(
         )
         return vmfb_path
 
-    _, model_name, model, forward_args, _ = get_model_and_inputs(
-        model_id, batch_size, tb_dir, tb_args
-    )
+    if compare_vs_eager:   
+        model_name, model, forward_args, golden, baseline = get_model_and_inputs(
+            model_id, batch_size, tb_dir, tb_args, get_baseline=True
+        )
+    else:
+        model_name, model, forward_args = get_model_and_inputs(
+            model_id, batch_size, tb_dir, tb_args
+        )
+        golden = None
+        baseline = None
 
     if dtype == torch.float16:
         model = model.half()
@@ -275,7 +294,8 @@ class CompiledTorchbenchModel(CompiledModule):
         inst = CompiledTorchbenchModel(context=Context(), import_to="IMPORT")
 
         module = CompiledModule.get_mlir_module(inst)
-
+    model.to("cpu")
+    del model
     if compile_to != "vmfb":
         return str(module)
     else:
@@ -288,17 +308,21 @@ class CompiledTorchbenchModel(CompiledModule):
             return_path=not exit_on_vmfb,
             attn_spec=attn_spec,
         )
-        return vmfb_path, external_weight_path, forward_args
+        return vmfb_path, external_weight_path, forward_args, golden, baseline
 
 
 def _run_iter(runner, inputs):
     start = time.time()
     res = runner.ctx.modules.compiled_torchbench_model["main"](*inputs)
     return res, time.time() - start
 
+def do_compare(shark_results, shark_latency, golden_results, golden_latency):
+    numerics_pass_fail = np.allclose(shark_results.to_host(), golden_results.clone().cpu().numpy(), rtol=1e-4, atol=1e-4)
+    speedup = golden_latency / shark_latency
+    return speedup, numerics_pass_fail
 
 def run_benchmark(
-    device, vmfb_path, weights_path, example_args, model_id, csv_path, iters
+    device, vmfb_path, weights_path, example_args, model_id, csv_path, iters, golden=None, baseline=None,
 ):
     if "rocm" in device:
         device = "hip" + device.split("rocm")[-1]
@@ -311,16 +335,23 @@ def run_benchmark(
     avg_latency = sum(iter_latencies) / len(iter_latencies)
     it_per_sec = 1 / avg_latency
 
+    if golden is not None and baseline is not None:
+        speedup, numerics_pass_fail = do_compare(results, avg_latency, golden, baseline)
+    else:
+        speedup, numerics_pass_fail = ("N/A", "N/A")
+
     needs_header = True
     if os.path.exists(csv_path):
         needs_header = False
     with open(csv_path, "a") as csvfile:
-        fieldnames = ["model", "avg_latency", "avg_iter_per_sec"]
+        fieldnames = ["model", "avg_latency", "avg_iter_per_sec", "speedup_over_eager", "numerics"]
         data = [
             {
                 "model": model_id,
                 "avg_latency": avg_latency,
                 "avg_iter_per_sec": it_per_sec,
+                "speedup_over_eager": speedup,
+                "numerics": numerics_pass_fail,
             }
         ]
         writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
@@ -346,7 +377,7 @@ def torch_to_iree(iree_runner, example_args):
 
 def run_main(model_id, args, tb_dir, tb_args):
     print(f"exporting {model_id}")
-    mod_str, weights_path, example_args = export_torchbench_model(
+    mod_str, weights_path, example_args, golden, baseline = benchmark_torchbench_model(
         model_id,
         tb_dir,
         tb_args,
@@ -361,6 +392,7 @@ def run_main(model_id, args, tb_dir, tb_args):
         decomp_attn=args.decomp_attn,
         attn_spec=args.attn_spec,
         input_mlir=args.input_mlir,
+        compare_vs_eager=args.compare_vs_torch,
     )
     if args.compile_to in ["torch", "mlir"]:
         safe_name = utils.create_safe_name(
@@ -379,6 +411,8 @@ def run_main(model_id, args, tb_dir, tb_args):
             model_id,
             args.output_csv,
             args.num_iters,
+            golden,
+            baseline,
         )
 
     gc.collect()

models/turbine_models/custom_models/torchbench/utils.py

@@ -12,22 +12,20 @@
 MI_flags = {
     "all": [
         "--iree-global-opt-propagate-transposes=true",
-        "--iree-opt-const-eval=false",
-        "--iree-llvmgpu-enable-prefetch=true",
-        "--iree-execution-model=async-external",
-        "--iree-dispatch-creation-enable-aggressive-fusion",
         "--iree-dispatch-creation-enable-fuse-horizontal-contractions=true",
+        "--iree-dispatch-creation-enable-aggressive-fusion=true",
         "--iree-opt-aggressively-propagate-transposes=true",
         "--iree-opt-outer-dim-concat=true",
-        "--iree-codegen-llvmgpu-use-vector-distribution=true",
+        "--iree-vm-target-truncate-unsupported-floats",
+        "--iree-llvmgpu-enable-prefetch=true",
+        "--iree-opt-data-tiling=false",
+        "--iree-codegen-gpu-native-math-precision=true",
+        "--iree-codegen-llvmgpu-use-vector-distribution",
+        "--iree-hip-waves-per-eu=2",
+        "--iree-execution-model=async-external",
     ],
     "preprocess_default": [
-        "--iree-preprocessing-pass-pipeline=builtin.module(iree-preprocessing-pad-to-intrinsics)",
-    ],
-    "preprocess_transpose": [
-        "--iree-preprocessing-pass-pipeline=builtin.module(iree-preprocessing-transpose-convolution-pipeline, iree-global-opt-raise-special-ops, iree-preprocessing-pad-to-intrinsics)",
-        "--iree-codegen-llvmgpu-enable-transform-dialect-jit=false",
-    ],
+    ]
 }
 GFX11_flags = {
     "all": [