Merge d1f46da into 0d542c3

Author: andrej

aie_kernels/aie2p/softmax.cc

@@ -3,6 +3,7 @@
 
 #include <aie_api/aie.hpp>
 #include <stdint.h>
+#include <math.h>
 
 #define SM_VEC_LEN 64   // 32
 #define log2e 1.4453125 // 1.44269504089
@@ -30,7 +31,7 @@ void softmax_simple_bf16(bfloat16 *restrict input_vector, bfloat16 *restrict out
     aie::vector<bfloat16, SM_VEC_LEN> in_elems, exp_val, input_bf16, log2e_vec, max_val_vec;
     aie::accum<accfloat, SM_VEC_LEN> out_vals, exp_val_accum, scaled_accum, exp_in_accum;
 
-    float max_val = 0;
+    float max_val = -INFINITY;
     float accum_exp_val = 0;
     float running_max = 0;
     bfloat16 col_sum_inv;

conftest.py

@@ -59,6 +59,14 @@ def __init__(self, csv_path):
         self.commit = get_git_commit()
         self.date = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
         self.test_metrics = {}  # test_name -> {metric_name -> [values]}
+        self._initialize_csv()
+
+    def _initialize_csv(self):
+        """Initialize CSV file - will be written incrementally as tests complete"""
+        self.csv_path.parent.mkdir(parents=True, exist_ok=True)
+        # Clear the file at the start of a new test run
+        with open(self.csv_path, "w", newline="") as f:
+            pass  # Create empty file, header will be written with first result
 
     def add_result(self, test_name, passed, captured_output, metric_patterns):
         self.test_metrics.setdefault(test_name, {}).setdefault("passed", []).append(
@@ -72,40 +80,70 @@ def add_result(self, test_name, passed, captured_output, metric_patterns):
             value = float(match.group("value"))
             self.test_metrics[test_name].setdefault(metric_name, []).append(value)
 
-    def finalize_results(self):
-        """Compute statistics for all collected metrics"""
-        for test_name, data in self.test_metrics.items():
-            row = {
-                "Commit": self.commit,
-                "Date": self.date,
-                "Test": test_name,
-                "Checks": f"{sum(data['passed'])}/{len(data['passed'])}",
-            }
-            for metric_name, values in data.items():
-                if metric_name == "passed":
-                    continue
-                if values:
-                    row[f"{metric_name} (mean)"] = statistics.mean(values)
-                    row[f"{metric_name} (median)"] = statistics.median(values)
-                    row[f"{metric_name} (min)"] = min(values)
-                    row[f"{metric_name} (max)"] = max(values)
-                    row[f"{metric_name} (stddev)"] = (
-                        statistics.stdev(values) if len(values) > 1 else 0.0
-                    )
+    def _compute_row(self, test_name, data):
+        """Compute statistics row for a single test"""
+        row = {
+            "Commit": self.commit,
+            "Date": self.date,
+            "Test": test_name,
+            "Checks": f"{sum(data['passed'])}/{len(data['passed'])}",
+        }
+        for metric_name, values in data.items():
+            if metric_name == "passed":
+                continue
+            if values:
+                row[f"{metric_name} (mean)"] = statistics.mean(values)
+                row[f"{metric_name} (median)"] = statistics.median(values)
+                row[f"{metric_name} (min)"] = min(values)
+                row[f"{metric_name} (max)"] = max(values)
+                row[f"{metric_name} (stddev)"] = (
+                    statistics.stdev(values) if len(values) > 1 else 0.0
+                )
+        return row
+
+    def write_test_result(self, test_name):
+        """Write or update result for a specific test incrementally"""
+        if test_name not in self.test_metrics:
+            return
+
+        data = self.test_metrics[test_name]
+        row = self._compute_row(test_name, data)
+
+        # Update or add to results
+        existing_idx = None
+        for idx, existing_row in enumerate(self.results):
+            if existing_row["Test"] == test_name:
+                existing_idx = idx
+                break
+
+        if existing_idx is not None:
+            self.results[existing_idx] = row
+        else:
             self.results.append(row)
 
-    def write_csv(self):
-        self.results.sort(key=lambda x: (x["Test"], x["Date"]))
+        # Rewrite entire CSV to handle column changes and updates
+        self._write_csv_internal()
+
+    def _write_csv_internal(self):
+        """Internal method to write CSV file"""
+        if not self.results:
+            return
+
+        sorted_results = sorted(self.results, key=lambda x: (x["Test"], x["Date"]))
 
         cols = {}
-        for row in self.results:
+        for row in sorted_results:
             cols.update({k: None for k in row.keys()})
 
         self.csv_path.parent.mkdir(parents=True, exist_ok=True)
         with open(self.csv_path, "w", newline="") as f:
             writer = csv.DictWriter(f, cols.keys())
             writer.writeheader()
-            writer.writerows(self.results)
+            writer.writerows(sorted_results)
+
+    def write_csv(self):
+        """Final write at session end - ensures all results are flushed"""
+        self._write_csv_internal()
 
 
 # Initialize the CSV writer once at test session setup
@@ -147,6 +185,9 @@ def pytest_runtest_makereport(item, call):
 
             csv_reporter.add_result(test_name, passed, captured, metric_patterns)
 
+            # Write results incrementally after each test
+            csv_reporter.write_test_result(test_name)
+
 
 def pytest_configure(config):
     csv_path = config.getoption("--csv-output")
@@ -170,7 +211,6 @@ def pytest_collection_modifyitems(config, items):
 
 def pytest_sessionfinish(session, exitstatus):
     if hasattr(session.config, "_csv_reporter"):
-        session.config._csv_reporter.finalize_results()
         session.config._csv_reporter.write_csv()
 
 

iron/common/compilation/base.py

@@ -502,6 +502,7 @@ def compile(self, graph):
                 str(self.aiecc_path),
                 "-v",
                 "-j1",
+                "--dynamic-objFifos",
                 "--no-compile-host",
                 "--no-xchesscc",
                 "--no-xbridge",

iron/common/fusion.py

@@ -5,6 +5,7 @@
 import ml_dtypes
 import pyxrt
 import ctypes
+import time
 from . import compilation as comp
 from .base import AIEOperatorBase, MLIROperator
 from .utils import XRTSubBuffer
@@ -42,8 +43,7 @@ def get_kernel_artifacts(self):
         """Collect all kernel artifacts from child operators.
 
         Returns:
-            List of KernelObjectArtifact instances from all unique child operators,
-            with filenames and symbol prefixes disambiguated per operator index.
+            List of KernelObjectArtifact instances from all unique child operators.
         """
         kernel_artifacts = []
         seen: dict[int, object] = {}
@@ -52,9 +52,6 @@ def get_kernel_artifacts(self):
         ]
         for idx, op in enumerate(unique_operators):
             objs = op.get_kernel_artifacts()
-            for obj in objs:
-                obj.filename = f"op{idx}_{obj.filename}"
-                obj.prefix_symbols = f"op{idx}_"
             kernel_artifacts.extend(objs)
         return kernel_artifacts
 
@@ -82,8 +79,6 @@ def get_mlir_artifact(self):
         ]
         for idx, op in enumerate(unique_operators):
             mlir_artifact = op.get_mlir_artifact()
-            if len(op.get_kernel_artifacts()) > 0:
-                mlir_artifact.generator.kwargs["func_prefix"] = f"op{idx}_"
             op_name = f"op{idx}_{op.__class__.__name__}"
             op_names[id(op)] = op_name
             operator_mlir_map[op_name] = mlir_artifact
@@ -290,8 +285,10 @@ def __call__(self, *args):
         for i, arg in enumerate(args):
             assert isinstance(arg, pyxrt.bo), f"Argument {i} is not a pyxrt.bo"
             run.set_arg(i, arg)
+        t0 = time.perf_counter()
         run.start()
         ret_code = run.wait()
+        self.last_elapsed = time.perf_counter() - t0
         if ret_code != pyxrt.ert_cmd_state.ERT_CMD_STATE_COMPLETED:
             raise RuntimeError(f"Kernel execution failed with return code {ret_code}")
 
@@ -371,10 +368,10 @@ def get_buffer(self, buffer_name):
         return sub_buffer
 
     def __call__(self):
-        self.input_buffer.to("npu")
+        self.input_buffer._sync_to_device()
         super().__call__(
             self.input_buffer.buffer_object(),
             self.output_buffer.buffer_object(),
             self.scratch_buffer.buffer_object(),
         )
-        self.output_buffer.to("cpu")
+        self.output_buffer._sync_from_device()

iron/operators/gemm/design.py

@@ -299,7 +299,7 @@ def my_matmul(
             gemm_object,
             [C_l1_ty_internal],
         )
-        matmul_func_name = f"matmul{scalar_suffix}_{dtype_in_str}_f32"
+        matmul_func_name = f"{func_prefix}matmul{scalar_suffix}_{dtype_in_str}_f32"
         matmul_kernel = Kernel(
             matmul_func_name,
             gemm_object,
@@ -314,7 +314,9 @@ def my_matmul(
             gemm_object,
             [C_l1_ty],
         )
-        matmul_func_name = f"matmul{scalar_suffix}_{dtype_in_str}_{dtype_out_str}"
+        matmul_func_name = (
+            f"{func_prefix}matmul{scalar_suffix}_{dtype_in_str}_{dtype_out_str}"
+        )
         matmul_kernel = Kernel(
             matmul_func_name,
             gemm_object,

iron/operators/mha_prefill_lxl_sd/__init__.py

@@ -0,0 +1,2 @@
+# SPDX-FileCopyrightText: Copyright (C) 2026 Advanced Micro Devices, Inc. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0