emx-onnx-cgen 0.2.0__py3-none-any.whl → 0.3.0__py3-none-any.whl

emx_onnx_cgen/testbench.py

@@ -0,0 +1,23 @@
+from __future__ import annotations
+
+from typing import Any
+
+import numpy as np
+
+
+def _convert_hex_floats(value: Any) -> Any:
+    if isinstance(value, list):
+        return [_convert_hex_floats(item) for item in value]
+    if isinstance(value, str):
+        return float.fromhex(value)
+    return value
+
+
+def decode_testbench_array(data: object, dtype: np.dtype) -> np.ndarray:
+    """Decode testbench JSON data into a numpy array.
+
+    Floating-point values are expected to be hex strings (C99 %a formatting).
+    """
+    if np.issubdtype(dtype, np.floating):
+        data = _convert_hex_floats(data)
+    return np.array(data, dtype=dtype)

emx_onnx_cgen/verification.py

@@ -0,0 +1,61 @@
+from __future__ import annotations
+
+import numpy as np
+
+
+def _float_uint_dtype(values: np.ndarray) -> type[np.unsignedinteger]:
+    if values.dtype == np.float16:
+        return np.uint16
+    if values.dtype == np.float32:
+        return np.uint32
+    if values.dtype == np.float64:
+        return np.uint64
+    raise ValueError(f"Unsupported floating dtype for ULP calculation: {values.dtype}")
+
+
+def _float_to_ordered_int(values: np.ndarray) -> np.ndarray:
+    uint_dtype = _float_uint_dtype(values)
+    bits = np.dtype(uint_dtype).itemsize * 8
+    sign_mask = np.array(1 << (bits - 1), dtype=uint_dtype)
+    as_uint = values.view(uint_dtype)
+    ordered = np.where(as_uint & sign_mask, ~as_uint, as_uint | sign_mask)
+    return ordered.astype(np.uint64, copy=False)
+
+
+def max_ulp_diff(actual: np.ndarray, expected: np.ndarray) -> int:
+    if actual.shape != expected.shape:
+        raise ValueError(
+            f"Shape mismatch for ULP calculation: {actual.shape} vs {expected.shape}"
+        )
+    if not np.issubdtype(expected.dtype, np.floating):
+        return 0
+    dtype = expected.dtype
+    if dtype not in (np.float16, np.float32, np.float64):
+        raise ValueError(f"Unsupported floating dtype for ULP calculation: {dtype}")
+    actual_cast = actual.astype(dtype, copy=False)
+    expected_cast = expected.astype(dtype, copy=False)
+    nan_mask = np.isnan(actual_cast) | np.isnan(expected_cast)
+    if nan_mask.any():
+        both_nan = np.isnan(actual_cast) & np.isnan(expected_cast)
+        if not np.all(both_nan):
+            uint_dtype = _float_uint_dtype(expected_cast)
+            return int(np.iinfo(uint_dtype).max)
+        actual_cast = actual_cast[~nan_mask]
+        expected_cast = expected_cast[~nan_mask]
+        if actual_cast.size == 0:
+            return 0
+    eps = np.finfo(dtype).eps
+    near_zero = (np.abs(actual_cast) < eps) & (np.abs(expected_cast) < eps)
+    if np.any(near_zero):
+        actual_cast = actual_cast.copy()
+        expected_cast = expected_cast.copy()
+        actual_cast[near_zero] = 0
+        expected_cast[near_zero] = 0
+    ordered_actual = _float_to_ordered_int(actual_cast)
+    ordered_expected = _float_to_ordered_int(expected_cast)
+    deltas = ordered_actual.astype(np.int64) - ordered_expected.astype(np.int64)
+    return int(np.max(np.abs(deltas)))
+
+
+def format_success_message(max_ulp: int) -> str:
+    return f"OK (max ULP {max_ulp})"

{emx_onnx_cgen-0.2.0.dist-info → emx_onnx_cgen-0.3.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: emx-onnx-cgen
-Version: 0.2.0
+Version: 0.3.0
 Summary: emmtrix ONNX-to-C Code Generator
 Requires-Python: >=3.10
 Description-Content-Type: text/markdown
@@ -9,14 +9,14 @@ Description-Content-Type: text/markdown
 
 [![PyPI - Version](https://img.shields.io/pypi/v/emx-onnx-cgen.svg)](https://pypi.org/project/emx-onnx-cgen)
 
-`emx-onnx-cgen` compiles ONNX models to portable, deterministic C code for deeply embedded systems. The generated code is designed to run without dynamic memory allocation, operating systems, or external runtimes, making it suitable for safety-critical and resource-constrained targets.
+`emx-onnx-cgen` compiles ONNX models to portable, deterministic C code for deeply embedded systems. The generated code is designed to run without dynamic memory allocation, operating-system services, or external runtimes, making it suitable for safety-critical and resource-constrained targets.
 
 Key characteristics:
 
 - **No dynamic memory allocation** (`malloc`, `free`, heap usage)
 - **Static, compile-time known memory layout** for parameters, activations, and temporaries
 - **Deterministic control flow** (explicit loops, no hidden dispatch or callbacks)
-- **No OS or libc dependencies** beyond basic C
+- **No OS dependencies**, using only standard C headers (for example, `stdint.h` and `stddef.h`)
 - **Single-threaded execution model**
 - **Bitwise-stable code generation** for reproducible builds
 - **Readable, auditable C code** suitable for certification and code reviews
@@ -47,7 +47,7 @@ Key characteristics:
   - `float`, `double`, `float16`
   - `int8_t`, `uint8_t`, `int16_t`, `uint16_t`, `int32_t`, `uint32_t`, `int64_t`, `uint64_t`
   - `bool`
-- Supporting dynamic dimensions by utilizing C99 variable-length arrays (VLAs).
+- Optional support for dynamic dimensions using C99 variable-length arrays (VLAs), when the target compiler supports them.
 
 ## Installation
 
@@ -93,6 +93,8 @@ Options:
 - `--model-name`: Override the generated model name (default: output file stem).
 - `--emit-testbench`: Emit a JSON-producing `main()` testbench for validation.
 - `--emit-data-file`: Emit constant data arrays into a companion `_data` C file.
+- `--large-weight-threshold`: Store weights larger than this element count in a binary file (default: `1024`).
+- `--large-temp-threshold-bytes`: Mark temporary buffers larger than this threshold as static (default: `1024`).
 - `--no-restrict-arrays`: Disable `restrict` qualifiers on generated array parameters.
 
 ### `verify`
@@ -106,6 +108,25 @@ Options:
 - `--template-dir`: Directory containing the C templates (default: `templates`).
 - `--model-name`: Override the generated model name (default: model file stem).
 - `--cc`: Explicit C compiler command for building the testbench binary.
+- `--large-weight-threshold`: Store weights larger than this element count in a binary file (default: `1024`).
+- `--large-temp-threshold-bytes`: Mark temporary buffers larger than this threshold as static (default: `1024`).
+- `--max-ulp`: Maximum allowed ULP distance for floating outputs (default: `100`).
+
+How verification works:
+
+1. **Compile with a testbench**: the compiler is invoked with `--emit-testbench`,
+   generating a C program that runs the model and prints inputs/outputs as JSON.
+2. **Build and execute**: the testbench is compiled with the selected C compiler
+   (`--cc`, `CC`, or a detected `cc/gcc/clang`) and executed in a temporary
+   directory.
+3. **Run ONNX Runtime**: the JSON inputs from the testbench are fed to ORT using
+   the same model.
+4. **Compare outputs**: floating outputs are compared by maximum ULP distance
+   (see https://www.emmtrix.com/wiki/ULP_Difference_of_Float_Numbers for the
+   ULP definition and algorithm); non-floating outputs must match exactly.
+   Missing outputs or mismatches are treated as failures.
+5. **ORT unsupported models**: if ORT reports `NOT_IMPLEMENTED`, verification is
+   skipped with a warning (exit code 0).
 
 ## Output
 
@@ -113,15 +134,20 @@ By default, the compiler emits a single C source file that includes:
 
 - A generated entry point that mirrors the ONNX graph inputs/outputs.
 - Tensor buffers for constants and temporaries.
-- A lightweight runtime implemented via templates in `templates/`.
 
 When `--emit-data-file` is enabled, the main C source declares constant arrays
 as `extern`, and a second file named like the output with a `_data` suffix
 contains the constant definitions.
 
+When `--large-weight-threshold` is set and a weight exceeds the threshold, the
+compiler emits a `<model>.bin` file with weights packed contiguously and
+generates a `<model>_load` helper that loads weights from the binary file at
+runtime.
+
 ## Official ONNX test coverage
 
 See [`OFFICIAL_ONNX_FILE_SUPPORT.md`](OFFICIAL_ONNX_FILE_SUPPORT.md) for the generated support matrix.
+See [`SUPPORT_OPS.md`](SUPPORT_OPS.md) for operator-level support derived from the expectation JSON files.
 
 ## Maintained by
 

{emx_onnx_cgen-0.2.0.dist-info → emx_onnx_cgen-0.3.0.dist-info}/RECORD

@@ -1,76 +1,93 @@
 emx_onnx_cgen/__init__.py,sha256=jUSbu1kJ0krzVTYEcph3jCprBhD7tWNtiSdL6r29KrM,221
 emx_onnx_cgen/__main__.py,sha256=iC1lLVtR6-TmpL6OxXcy3oIntExUtajn9-q627R1XyI,140
-emx_onnx_cgen/_build_info.py,sha256=tp1Kzo4PNvj3ydRSCdr84YUd2wEUVUopl7AoaJfOaHk,112
-emx_onnx_cgen/cli.py,sha256=2eQ6JxfdamiLDqo-ZfZNnnb1a1I7NTEvecfUJaAA_3M,11024
-emx_onnx_cgen/compiler.py,sha256=foB9JK1Z2NRcRg_Fn-kL9K7L7tKGWxI9-kaMbYg3dnM,20877
+emx_onnx_cgen/_build_info.py,sha256=zS8xdzMihYIqmbd58Pfku76TrBsTlBdkLVrWqTCiUs4,112
+emx_onnx_cgen/_version.py,sha256=5zTqm8rgXsWYBpB2M3Zw_K1D-aV8wP7NsBLrmMKkrAQ,704
+emx_onnx_cgen/cli.py,sha256=hRF2xG6t2YUkkTYrAUVsOyz1lTAdjokE-1pxFffsG2c,20643
+emx_onnx_cgen/compiler.py,sha256=Q4a4_a1DkGmbrRJaRgpk0uyOqrJQExqDQD_BNw3AUcw,28585
 emx_onnx_cgen/dtypes.py,sha256=jRx3BBvk0qFW14bngoL1B7L_IRasyNJ4jqhpM5YhcOM,1335
 emx_onnx_cgen/errors.py,sha256=HpOv95mTgr9ZX2gYe1RtwVMbPskh7zkqjU_FgAD-uIM,363
-emx_onnx_cgen/onnx_import.py,sha256=aMvSxT3ycg4UmnutWYvsQpzGt2m_KpNnDIiddlg-vDA,9028
-emx_onnx_cgen/ops.py,sha256=pW3ks2EJITiJxvThfU58KjQZE7AFUiPmMwKZRCNb1mY,16586
+emx_onnx_cgen/onnx_import.py,sha256=IF7KZGfEP9H4H1fHYjobGbB_381fqD_67KtqZYs9AZ4,9168
+emx_onnx_cgen/onnxruntime_utils.py,sha256=mEsC1x00M1jyBgVBKqnKoqx6H1tdgsFFUy7rbITs3bs,308
+emx_onnx_cgen/ops.py,sha256=qpPOaqsYprlJrhCNLVBZ3XnREBRDdmkXbd1zaAkywOI,16732
+emx_onnx_cgen/testbench.py,sha256=-NbqD1aC7OXvFMLiLzd2IPObenQdHFH85cNxNSB1GeY,640
 emx_onnx_cgen/validation.py,sha256=KFdUdGjQbzTj1szCJcjxnTi8f5l6ywNgCB9abbBpTbM,2360
+emx_onnx_cgen/verification.py,sha256=eTnhl_9YObyvs0fqJAw8796TlRzp-IoFM4JuMkQ8XOc,2403
 emx_onnx_cgen/codegen/__init__.py,sha256=-_sxL87uyAIunaetjUvIUo2bc46ugVlaNtSsidegMRM,362
-emx_onnx_cgen/codegen/c_emitter.py,sha256=gCgbqTgDWGAmp7H8TCFPBKp7DCNb_nskkMCAzKwDG0Y,337438
+emx_onnx_cgen/codegen/c_emitter.py,sha256=vYrRx3UQvve_s4ElLiuh25lsbEt7mDTCXjvr-kdkggM,422455
 emx_onnx_cgen/ir/__init__.py,sha256=fD2D8qxlGoCFJb0m9v6u3XTgzSxDOhB4cfLBiCLovzg,102
-emx_onnx_cgen/ir/model.py,sha256=e8vRA0RNDU8Ioz3TXQKpdUhDtUK6Hm71KouUPMhCcpg,1213
+emx_onnx_cgen/ir/model.py,sha256=SZ3K8t4dKUqWuXWe5ozApofXx4bdcf4p0WYCdeU-mFA,1265
 emx_onnx_cgen/lowering/__init__.py,sha256=wrxLMWcPUH1RbPJOs0Tsdb12FhXjAAeZVDYwKqcIuzw,103
 emx_onnx_cgen/lowering/arg_reduce.py,sha256=2AowDRCJRkIvrVBphbA0rM18oCWEpCDEV5Y4K9wSDII,3388
 emx_onnx_cgen/lowering/attention.py,sha256=19Jq_k0DXwH71a3pmLTWCNMttmw5uuiNK6Jhln5HC4A,16488
 emx_onnx_cgen/lowering/average_pool.py,sha256=9kg3pYHG7QLid_M2dbleC1VoNlVlGsKdOrsWp3pt7sc,8085
 emx_onnx_cgen/lowering/batch_normalization.py,sha256=_aFCm4QaC5jH-JNEvqDFYOyAMdzgUFS_3Gmo1vdPyKE,3987
 emx_onnx_cgen/lowering/cast.py,sha256=zKiE4wI7oWP_TjxBV4fY3-FXvZxK2zy58O6tWJ2dODQ,2852
-emx_onnx_cgen/lowering/common.py,sha256=4w9kjKW3_LilOGgmXYcUGg5lohbYsaLudoL4ALoDUkk,2356
+emx_onnx_cgen/lowering/common.py,sha256=OF5UTin4teEFSp-rbiUArYCJogZ636Rujhkgrm2vj_w,16083
 emx_onnx_cgen/lowering/concat.py,sha256=TefckPfuaIHVHxGExJiO9wlkjyRO1TGg-QAMeoW8hW0,1097
 emx_onnx_cgen/lowering/constant_of_shape.py,sha256=btQflQFMP_y22sK7RrhkbGdaeSSLPC_DWhLjxY7CAgk,3208
 emx_onnx_cgen/lowering/conv.py,sha256=I1_tssw_ySf4beKV0sCVe8DRhNxL58PqC0wxtWjD79s,7309
+emx_onnx_cgen/lowering/conv_transpose.py,sha256=vMbH7g3V9o68BjsW-FurNp1G8Dgr3NrV7JPLLfopHG0,11164
 emx_onnx_cgen/lowering/cumsum.py,sha256=eX0bDtwY-qevz0KXNHtJaDiKUUHIOhDX0uDiSxcC0ZU,4125
 emx_onnx_cgen/lowering/depth_space.py,sha256=M4md379jiumGWmg7EgR-CinoPzwof2RdfOiNqOzxd9o,4217
 emx_onnx_cgen/lowering/dropout.py,sha256=oBKPMN-J9Gnw8dRXvf-bN15L1-5W7-qKhR72Z6AgLXQ,1775
+emx_onnx_cgen/lowering/einsum.py,sha256=g0KEZNJb87SzH-TqDDcfNPTcAaRioq455eN6HHLZNNo,6128
 emx_onnx_cgen/lowering/elementwise.py,sha256=HN6vEW58lceYECp-7QWLCWOBo1ImyY66aZIg06nA5g8,6231
 emx_onnx_cgen/lowering/expand.py,sha256=4msnYM-6RnzMplQqde2ovOLsjmWQ4bnXEoUiEM6CT6k,5529
 emx_onnx_cgen/lowering/eye_like.py,sha256=76HEdT-EofDCCy7DewjIpILJdIJyJ-YVCbLXO54SX5E,1734
 emx_onnx_cgen/lowering/flatten.py,sha256=sGol05FDN0xoNgSl_DlVbjYvBHCHWjQC2KB15ytYfPs,2142
 emx_onnx_cgen/lowering/gather.py,sha256=9zMB9fcdJi1fkTmDs_-L6FvQi1fnhdk0h7RmeN5MP2M,1814
-emx_onnx_cgen/lowering/gather_elements.py,sha256=0E-WAge15HhGeWPRC_ZE94fb9C6LIoef8p5D1usWvBQ,2358
+emx_onnx_cgen/lowering/gather_elements.py,sha256=K-3w__F_I_gq3Kykk7LydTR5syH_Zpi-0-rdShLumbo,2329
+emx_onnx_cgen/lowering/gather_nd.py,sha256=_0IW93RMRa9VtXSu4KMpBBA18ovLBGTmH90Y-ANOk1M,3101
 emx_onnx_cgen/lowering/gemm.py,sha256=Ps2T4tZgXr5FObz5figwbLZq-Njzg44iBQ9cFmvH78k,4590
+emx_onnx_cgen/lowering/global_max_pool.py,sha256=xyoqQyRFpDKCXBO8bqp7JstVxVfbj9pMd06-848ix5o,2223
 emx_onnx_cgen/lowering/grid_sample.py,sha256=Ne-97ljxSdqfjBJtVHp2AQnEeXGQ5HE-HegCoxcNCm0,5228
 emx_onnx_cgen/lowering/group_normalization.py,sha256=RqXud5_xNxMu8DP5EsPr4L2b6CZghQWCcG6Bh1x2gMA,2664
-emx_onnx_cgen/lowering/identity.py,sha256=fn1Tg56xACwAjhesy0wyr9TJjCmmddnd8QrQ4-uCdO0,1843
+emx_onnx_cgen/lowering/hardmax.py,sha256=Lqnt9g48-Kpcklq_uJpA93FYdhTI8SjUu4Cnh9_BhYY,1961
+emx_onnx_cgen/lowering/identity.py,sha256=h8cy6T9n10IWEnOKgIGOca9sNVaw5vU1fdt_p8AAqq8,1859
 emx_onnx_cgen/lowering/instance_normalization.py,sha256=1Yx2KPKq_BHberCBTrGQXQswAS0FfDle9NpyeD41ypU,1950
 emx_onnx_cgen/lowering/layer_normalization.py,sha256=ZvqGZOhuoYh8ZPyzb-PV0kIc2bbunWTYj12wmrGu9YY,4529
-emx_onnx_cgen/lowering/logsoftmax.py,sha256=1FEaX45GdDr6jIdS_sOwXOy_DdVDruZem4yZ9XA4a38,1669
+emx_onnx_cgen/lowering/logsoftmax.py,sha256=gdPYJdRNjwRDRXozzKOkKHV7HeOw48Zl4guiAN5pgKs,1895
 emx_onnx_cgen/lowering/lp_normalization.py,sha256=61CGS-2yN0bf5dby5b7Ug1PH3CStZN1xZmYWa5TysTI,1712
+emx_onnx_cgen/lowering/lp_pool.py,sha256=96M-CeIqOkPstVr2BEoASImG6-Z4_S7ngB8bmPQlo7M,4873
 emx_onnx_cgen/lowering/lrn.py,sha256=zGw1Jk7iBk1jHdjdDqfAREsV5VcSdOG3LcAmEllIB08,3370
 emx_onnx_cgen/lowering/lstm.py,sha256=JhGxiF3bTSY3flkw_u9mil2esRxvIjr5Tc4vSPULDr4,12305
-emx_onnx_cgen/lowering/matmul.py,sha256=NEfBa140ofpgm9xnqUBulMSA-yQlb29F2NqhCJpmKSY,4262
+emx_onnx_cgen/lowering/matmul.py,sha256=QJ2DfMk6g5hNNpfUUfHH732cfSLL0LFzZxw-4GOuRYQ,4259
 emx_onnx_cgen/lowering/maxpool.py,sha256=MRLeoCEdIwO8JNWOi7iKoeIsJvukqpx_w6GCHaDaYHU,7494
 emx_onnx_cgen/lowering/mean_variance_normalization.py,sha256=L_6ECH9wPEnNX2mL6yroZRexZM8JV5ZnJvoPQS6IAuc,1875
-emx_onnx_cgen/lowering/negative_log_likelihood_loss.py,sha256=wMWvJ9ymkA-ptFgqx0VmHAJHS5NVvDeo2GAlrECztJ8,9307
+emx_onnx_cgen/lowering/negative_log_likelihood_loss.py,sha256=nqNETCqZP9MZhBU_Wcgaxu1w0uNKHa6VR5KTw--CWP0,9344
+emx_onnx_cgen/lowering/nonzero.py,sha256=d_81vdF2BkYi_Z7jbypQ3qwSe-ctNoSxaJRYO8qDMjc,1637
+emx_onnx_cgen/lowering/one_hot.py,sha256=twxMlNosGfm3iAeGv5LDGPliMCJ36o7Dxusn4gVJE00,4352
 emx_onnx_cgen/lowering/pad.py,sha256=cNlh-rA3CRPfO-u8gvJ1MeF1j-vdBMXLuJBpp2DkFTc,10416
+emx_onnx_cgen/lowering/quantize_linear.py,sha256=3IvZTQbWAYh61nVNAbd-CeHTpYaULUSEjSdnu-nIOI8,4771
 emx_onnx_cgen/lowering/range.py,sha256=xLbG3SGvQiboPqSIh5qZyw3Krbxjk0nd3YvRlUI_q64,3463
-emx_onnx_cgen/lowering/reduce.py,sha256=XRxPopJCU9FGd1XmdxCZickmSTEvmkpAgLUPDFfjRm0,18431
+emx_onnx_cgen/lowering/reduce.py,sha256=xetOc7mfZgKsD57O00K8NKmb0sEcHbMTpdDlqi4U6b4,18413
 emx_onnx_cgen/lowering/registry.py,sha256=rKyWnLDBFHJbHw-iyOtXv2Qc5LEBCwgopUXewvQpEpg,1392
-emx_onnx_cgen/lowering/reshape.py,sha256=SF46eP95Z19PT67ayJDKW1fwllBWEZmGfXAL5q9wy-I,6881
+emx_onnx_cgen/lowering/reshape.py,sha256=dXZx3tTklPHSHAi-ZroiS6niCmHfHb7yGioJUYUDkLY,13452
 emx_onnx_cgen/lowering/resize.py,sha256=J_x53hVHlfJemLwEhq5n_11Pe1TlF9nRMEpkw6IpzN8,14644
 emx_onnx_cgen/lowering/rms_normalization.py,sha256=_H56Pf9T80FYbmy1m3oc7_D5TbNxRrVeJScD5VmLZRo,2536
+emx_onnx_cgen/lowering/scatter_nd.py,sha256=q7rBQ0AdPrwsHnsN5qpCTcnwCWyKzSaG86wpAhNiDzE,3221
 emx_onnx_cgen/lowering/shape.py,sha256=Vvd2zQB06wZcEe4mW5WBRrQuVF8f_tXSM9fpGxe9PEo,2913
 emx_onnx_cgen/lowering/size.py,sha256=Z_DTevdpx2W_3k0GoyQ2uWE3ms_PN1d_Ti7hh6HhB1Q,1261
 emx_onnx_cgen/lowering/slice.py,sha256=yHm_mXeHcLufDmVNvj_kv08zMdbvI39ViHcE-tVPKa0,14816
-emx_onnx_cgen/lowering/softmax.py,sha256=ZaOZf00f5PNHRjSki08Fv-iod6UgqL7cmblfpE_OQRU,1648
+emx_onnx_cgen/lowering/softmax.py,sha256=qmg9AcxFgYZcz98VyxxgyaFmSEcUCxhGtfSX6zLTRgE,1874
 emx_onnx_cgen/lowering/softmax_cross_entropy_loss.py,sha256=I0pbWyJdnf-9vAuX8-xsnovDKDGxynlBhfj5k_IVIa4,5230
 emx_onnx_cgen/lowering/split.py,sha256=ImGnsqrl7IdWbPTPazfXYjcsoRoziqsqtJBum12xTXY,5894
-emx_onnx_cgen/lowering/squeeze.py,sha256=rgICFprcWhC03h-GXZNaIQsdFjsqyxybJYtPklTfaYM,6086
+emx_onnx_cgen/lowering/squeeze.py,sha256=ihLbOmUW3PIrV9_do78Qal-J-Ten7sxTnd4a2tm5aEo,6095
 emx_onnx_cgen/lowering/tile.py,sha256=fT-ybiBZfb4bqBAPrCORZCNm3KWeu4rRW2BJ_UVIVZU,3041
+emx_onnx_cgen/lowering/topk.py,sha256=u2FGoCt1RLI8GDCP-833Raxw5yMlXQY3ElJ-vJj-FPY,4113
 emx_onnx_cgen/lowering/transpose.py,sha256=TrRXUt-4UFNHZWaOpS3N5zEz5-OCK6-twZdlrnw7Pqg,1762
-emx_onnx_cgen/lowering/unsqueeze.py,sha256=sE3vribz8EyHqDG8lEcreKII7rQDElnHf1OpoM5HiAo,5987
+emx_onnx_cgen/lowering/trilu.py,sha256=irA0fZV_OzKRYMhbJGuAZQPcDgxztNE0a-fMw_seU6E,3277
+emx_onnx_cgen/lowering/unsqueeze.py,sha256=tlmdF8OMS9u-aT3jSwmuBo_VHzgMu3QiCnf45UNMNuY,5996
 emx_onnx_cgen/lowering/variadic.py,sha256=hmPzRIj0kcZriGRTR2ma1YMH9g21K_4f-3FXw6qO3jE,3298
 emx_onnx_cgen/lowering/where.py,sha256=uiaWU9RM6o-n38N0AEINIkXS33yVK3-ohkfKIApJOoA,2655
 emx_onnx_cgen/runtime/__init__.py,sha256=88xGpAs1IEBlzlWL_e9tnKUlaSRdc7pQUeVCu5LC4DY,50
-emx_onnx_cgen/runtime/evaluator.py,sha256=GFxrBXcKuQkZ0HY46twOTrNc955UqW3cRKAu5AYVJzQ,84910
+emx_onnx_cgen/runtime/evaluator.py,sha256=Zd0RRwn0c7Lr3eW4OSvJEX9oBEX9p4cr_fmfvJK4LHY,102372
 shared/__init__.py,sha256=bmP79AVZdY_1aNULJap9pm76Q41Rabrza6X-0A8lDzw,45
-shared/scalar_functions.py,sha256=OAFO6kT6Gtcv5jp7UBLRhifhGmAbWhDKAmapTvqQruc,89911
+shared/scalar_functions.py,sha256=KawY6sleIcVf5FdffFABQWeh4P_I8Oz7IMPcXjMyRfw,90843
 shared/scalar_types.py,sha256=kEpsl5T-NVFxCcTzXqPJbtpvDiCgKHfz91dphLLZxZA,4912
-emx_onnx_cgen-0.2.0.dist-info/METADATA,sha256=xTbPSAdUMfyXwvkxGGVYhsGhKjpxgPSzV1rimuB8zn0,4256
-emx_onnx_cgen-0.2.0.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-emx_onnx_cgen-0.2.0.dist-info/entry_points.txt,sha256=b7Rvmz_Bi9kWyn7QayQC_FEXiRpt4cS1RnluKh49yoo,57
-emx_onnx_cgen-0.2.0.dist-info/top_level.txt,sha256=g39fo-blEbgiVcC_GRqAnBzN234w3LXbcVdLUoItSLk,21
-emx_onnx_cgen-0.2.0.dist-info/RECORD,,
+shared/ulp.py,sha256=o_JQ0pyeu1BD5Jx5tWuKnFQYWm1Q8zA8i8EHecf6Ys8,1371
+emx_onnx_cgen-0.3.0.dist-info/METADATA,sha256=Q4l5q1s5a6pLWYmAujI9CzU-kMm40d4JffGtQQwVCjw,6036
+emx_onnx_cgen-0.3.0.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+emx_onnx_cgen-0.3.0.dist-info/entry_points.txt,sha256=b7Rvmz_Bi9kWyn7QayQC_FEXiRpt4cS1RnluKh49yoo,57
+emx_onnx_cgen-0.3.0.dist-info/top_level.txt,sha256=g39fo-blEbgiVcC_GRqAnBzN234w3LXbcVdLUoItSLk,21
+emx_onnx_cgen-0.3.0.dist-info/RECORD,,

shared/scalar_functions.py

@@ -1,6 +1,6 @@
 from __future__ import annotations
 
-from dataclasses import dataclass
+from dataclasses import dataclass, field
 from enum import Enum
 import math
 from typing import Callable, Dict, List, Mapping, Set
@@ -26,6 +26,7 @@ class _GeneratedScalar:
     lines: List[str]
     deps: Set[ScalarFunctionKey]
     includes: Set[str]
+    constants: Set[str] = field(default_factory=set)
 
 
 def _scalar_function_spec(
@@ -396,6 +397,7 @@ _ONNX_OP_TO_SCALAR_FUNCTION = {
     "Max": ScalarFunction.MAXIMUM,
     "Mean": ScalarFunction.MEAN,
     "Min": ScalarFunction.MINIMUM,
+    "Mish": ScalarFunction.MISH,
     "Mod": ScalarFunction.FMOD,
     "Mul": ScalarFunction.MUL,
     "Neg": ScalarFunction.NEG,
@@ -1071,7 +1073,7 @@ def _float_sign(dtype_info: _ScalarTypeInfo) -> _GeneratedScalar:
 
 
 def _float_round(dtype_info: _ScalarTypeInfo) -> _GeneratedScalar:
-    return _float_unary_math(dtype_info, "round", "round")
+    return _float_unary_math(dtype_info, "round", "rint")
 
 
 def _float_trunc(dtype_info: _ScalarTypeInfo) -> _GeneratedScalar:
@@ -1089,7 +1091,7 @@ def _float_angle(dtype_info: _ScalarTypeInfo) -> _GeneratedScalar:
         f"    return a < {zero} ? {pi} : {zero};",
         "}",
     ]
-    return _GeneratedScalar(lines=lines, deps=set(), includes=set())
+    return _GeneratedScalar(lines=lines, deps=set(), includes=set(), constants={pi})
 
 
 def _float_conj(dtype_info: _ScalarTypeInfo, name: str) -> _GeneratedScalar:
@@ -1099,13 +1101,25 @@ def _float_conj(dtype_info: _ScalarTypeInfo, name: str) -> _GeneratedScalar:
 def _float_deg2rad(dtype_info: _ScalarTypeInfo) -> _GeneratedScalar:
     pi = "REF_PI_F" if dtype_info.suffix == "f32" else "REF_PI_D"
     one_eighty = _float_literal(180.0, dtype_info)
-    return _simple_unary(dtype_info, "deg2rad", f"a * ({pi} / {one_eighty})")
+    generated = _simple_unary(dtype_info, "deg2rad", f"a * ({pi} / {one_eighty})")
+    return _GeneratedScalar(
+        lines=generated.lines,
+        deps=generated.deps,
+        includes=generated.includes,
+        constants={pi},
+    )
 
 
 def _float_rad2deg(dtype_info: _ScalarTypeInfo) -> _GeneratedScalar:
     pi = "REF_PI_F" if dtype_info.suffix == "f32" else "REF_PI_D"
     one_eighty = _float_literal(180.0, dtype_info)
-    return _simple_unary(dtype_info, "rad2deg", f"a * ({one_eighty} / {pi})")
+    generated = _simple_unary(dtype_info, "rad2deg", f"a * ({one_eighty} / {pi})")
+    return _GeneratedScalar(
+        lines=generated.lines,
+        deps=generated.deps,
+        includes=generated.includes,
+        constants={pi},
+    )
 
 
 def _float_digamma_f64() -> _GeneratedScalar:
@@ -1135,7 +1149,9 @@ def _float_digamma_f64() -> _GeneratedScalar:
         "    return result;",
         "}",
     ]
-    return _GeneratedScalar(lines=lines, deps=set(), includes=set())
+    return _GeneratedScalar(
+        lines=lines, deps=set(), includes=set(), constants={"REF_PI_D"}
+    )
 
 
 def _float_digamma(dtype_info: _ScalarTypeInfo) -> _GeneratedScalar:
@@ -1186,7 +1202,7 @@ def _float_erfinv(dtype_info: _ScalarTypeInfo) -> _GeneratedScalar:
         "    return approx;",
         "}",
     ]
-    return _GeneratedScalar(lines=lines, deps=set(), includes=set())
+    return _GeneratedScalar(lines=lines, deps=set(), includes=set(), constants={pi})
 
 
 def _float_frac(dtype_info: _ScalarTypeInfo) -> _GeneratedScalar:
@@ -1288,7 +1304,7 @@ def _float_sinc(dtype_info: _ScalarTypeInfo) -> _GeneratedScalar:
         f"    return {_math_fn('sin', dtype_info)}(x) / x;",
         "}",
     ]
-    return _GeneratedScalar(lines=lines, deps=set(), includes=set())
+    return _GeneratedScalar(lines=lines, deps=set(), includes=set(), constants={pi})
 
 
 def _float_square(dtype_info: _ScalarTypeInfo) -> _GeneratedScalar:
@@ -2285,7 +2301,12 @@ def _generate_scalar(key: ScalarFunctionKey) -> _GeneratedScalar:
             includes.add("#include <limits.h>")
     if dtype_info.is_bool:
         includes.add("#include <stdbool.h>")
-    return _GeneratedScalar(lines=generated.lines, deps=generated.deps, includes=includes)
+    return _GeneratedScalar(
+        lines=generated.lines,
+        deps=generated.deps,
+        includes=includes,
+        constants=generated.constants,
+    )
 
 
 def _function_name_for_key(key: ScalarFunctionKey) -> str:
@@ -2352,6 +2373,7 @@ class ScalarFunctionRegistry:
     def include_lines(self) -> List[str]:
         includes: Set[str] = set()
         visited: Set[ScalarFunctionKey] = set()
+        constants: Set[str] = set()
 
         def collect(key: ScalarFunctionKey) -> None:
             if key in visited:
@@ -2362,18 +2384,28 @@ class ScalarFunctionRegistry:
             for dep in entry.deps:
                 collect(dep)
             includes.update(entry.includes)
+            constants.update(entry.constants)
 
         for key in self._requested:
             collect(key)
         ordered = sorted(includes)
-        preamble = [
-            "#ifndef REF_PI_F",
-            "#define REF_PI_F 3.14159265358979323846f",
-            "#endif",
-            "#ifndef REF_PI_D",
-            "#define REF_PI_D 3.14159265358979323846",
-            "#endif",
-        ]
+        preamble: List[str] = []
+        if "REF_PI_F" in constants:
+            preamble.extend(
+                [
+                    "#ifndef REF_PI_F",
+                    "#define REF_PI_F 3.14159265358979323846f",
+                    "#endif",
+                ]
+            )
+        if "REF_PI_D" in constants:
+            preamble.extend(
+                [
+                    "#ifndef REF_PI_D",
+                    "#define REF_PI_D 3.14159265358979323846",
+                    "#endif",
+                ]
+            )
         return ordered + preamble
 
     def render(self) -> List[str]:

shared/ulp.py

@@ -0,0 +1,48 @@
+from __future__ import annotations
+
+from typing import Dict
+
+import numpy as np
+
+from shared.scalar_types import ScalarFunctionError
+
+_FLOAT_TO_UINT: Dict[np.dtype, np.dtype] = {
+    np.dtype("float16"): np.dtype("uint16"),
+    np.dtype("float32"): np.dtype("uint32"),
+    np.dtype("float64"): np.dtype("uint64"),
+}
+
+
+def _coerce_float_scalar(value: object, dtype: np.dtype) -> np.ndarray:
+    return np.asarray(value, dtype=dtype).reshape(())
+
+
+def _ulp_intdiff_same_sign(
+    f1: np.ndarray, f2: np.ndarray, uint_dtype: np.dtype
+) -> int:
+    i1 = f1.view(uint_dtype).item()
+    i2 = f2.view(uint_dtype).item()
+    return int(i1 - i2) if i1 > i2 else int(i2 - i1)
+
+
+def ulp_intdiff_float(f1: object, f2: object) -> int:
+    dtype = np.result_type(f1, f2)
+    try:
+        uint_dtype = _FLOAT_TO_UINT[dtype]
+    except KeyError as exc:
+        raise ScalarFunctionError(
+            f"unsupported dtype for ULP diff: {dtype}"
+        ) from exc
+
+    f1_scalar = _coerce_float_scalar(f1, dtype)
+    f2_scalar = _coerce_float_scalar(f2, dtype)
+
+    if np.signbit(f1_scalar) != np.signbit(f2_scalar):
+        zero = _coerce_float_scalar(0.0, dtype)
+        return (
+            ulp_intdiff_float(zero, np.abs(f1_scalar))
+            + ulp_intdiff_float(zero, np.abs(f2_scalar))
+            + 1
+        )
+
+    return _ulp_intdiff_same_sign(f1_scalar, f2_scalar, uint_dtype)