Windows support (#136)

* Windows support

- Switched to using ``time.perf_counter_ns`` to prevent divide by 0 errors.
- Hoisted object FIFO acquires out of conditional structures to satisfy cyclostatic compiler analysis constraints.
- Swapped shell/awk commands for platform-independent Python execution blocks to handle symbol mapping on Windows.
- Added test skip validation for missing gated weights to stabilize local test executions.

Co-authored-by: Gemini AI <gemini-ai@google.com>

* Potential fix for pull request finding

Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>

* fix formatting issue and revert mha changes

---------

Co-authored-by: Gemini AI <gemini-ai@google.com>
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
Co-authored-by: André Rösti <an.roesti@gmail.com>

Author: 4 people

iron/applications/llama_3.2_1b/test.py

@@ -5,6 +5,7 @@
 import subprocess
 import pytest
 import os
+import sys
 from pathlib import Path
 
 test_dir = Path(__file__).parent
@@ -27,14 +28,21 @@ def generate_test_params():
 params, names = generate_test_params()
 
 
+@pytest.mark.skipif(
+    not (
+        (weights_dir / "llama3.2-1b" / "model.safetensors").exists()
+        and (weights_dir / "llama3.2-1b" / "tokenizer.model").exists()
+    ),
+    reason="llama3.2-1b weights not found",
+)
 @pytest.mark.supported_devices("npu2")
 @pytest.mark.metrics(
     TTFT=r"\[Prefill\]\s*Time to first token:\s*(?P<value>[\d\.e\+-]+) s",
     TPS=r"\[Decode\]\s*Tokens per second:\s*(?P<value>[\d\.e\+-]+)",
 )
 @pytest.mark.parametrize("prompt_len,num_tokens", params, ids=names)
 def test_llama_3_2_1b(prompt_len, num_tokens):
-    command = f"python3 {test_dir}/llama_npu.py {weights_dir}/llama3.2-1b/model.safetensors {weights_dir}/llama3.2-1b/tokenizer.model --num-tokens {num_tokens} --prompt-len {prompt_len}"
+    command = f"{sys.executable} {test_dir}/llama_npu.py {weights_dir}/llama3.2-1b/model.safetensors {weights_dir}/llama3.2-1b/tokenizer.model --num-tokens {num_tokens} --prompt-len {prompt_len}"
 
     result = subprocess.run(
         command,

iron/common/compilation/base.py

@@ -795,20 +795,37 @@ def _prefix_symbols(self, artifact, prefix):
         nm_path = self._find_working_tool("llvm-nm")
         symbol_map_file = artifact.filename + ".symbol_map"
 
-        # Extract defined symbols and build the redefine-syms map. Run nm to a
-        # file, THEN awk (joined by `&&`) rather than `nm | awk`: a pipe reports
-        # only awk's exit status, so a failing nm silently produces an EMPTY map
-        # and the prefix rename is skipped, surfacing much later as
-        # `undefined symbol: {prefix}<sym>` at the per-core link. With `&&` a
-        # failed nm aborts here loudly instead.
-        nm_cmd = [
-            "sh",
-            "-c",
-            f"{nm_path} --defined-only --extern-only {artifact.filename} "
-            f"> {symbol_map_file}.syms && "
-            f"awk '{{print $3 \" {prefix}\" $3}}' {symbol_map_file}.syms "
-            f"> {symbol_map_file}",
-        ]
+        if os.name == "nt":
+            # Pure python code execution block wrapped cleanly for Windows
+            python_script = f"""
+import subprocess
+nm_cmd = [{repr(nm_path)}, '--defined-only', '--extern-only', {repr(artifact.filename)}]
+res = subprocess.run(nm_cmd, capture_output=True, text=True, check=True)
+lines = []
+for line in res.stdout.splitlines():
+    parts = line.strip().split()
+    if len(parts) >= 3:
+        sym = parts[-1]
+        lines.append(f"{{sym}} {prefix}{{sym}}\\n")
+with open({repr(symbol_map_file)}, 'w') as f:
+    f.writelines(lines)
+"""
+            nm_cmd = [sys.executable, "-c", python_script.strip()]
+        else:
+            # Extract defined symbols and build the redefine-syms map. Run nm to a
+            # file, THEN awk (joined by `&&`) rather than `nm | awk`: a pipe reports
+            # only awk's exit status, so a failing nm silently produces an EMPTY map
+            # and the prefix rename is skipped, surfacing much later as
+            # `undefined symbol: {prefix}<sym>` at the per-core link. With `&&` a
+            # failed nm aborts here loudly instead.
+            nm_cmd = [
+                "sh",
+                "-c",
+                f"{nm_path} --defined-only --extern-only {artifact.filename} "
+                f"> {symbol_map_file}.syms && "
+                f"awk '{{print $3 \" {prefix}\" $3}}' {symbol_map_file}.syms "
+                f"> {symbol_map_file}",
+            ]
 
         # Apply the renaming using the symbol map
         objcopy_cmd = [