Add support to hash via GPUs

Signed-off-by: Ashutosh Gupta <[email protected]>

Author: Ashutosh Gupta

pyproject.toml

@@ -59,6 +59,10 @@ otel = [
   "opentelemetry-exporter-otlp",
 ]
 
+gpu = [
+  "torch",
+]
+
 [project.scripts]
 model_signing = "model_signing._cli:main"
 

src/model_signing/_hashing/gpu.py

@@ -0,0 +1,234 @@
+# Copyright 2024 The Sigstore Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""GPU-enabled hashing engines.
+
+This module contains hashing engines backed by PyTorch. When a CUDA capable
+GPU is available, the hashing computation runs on the GPU. Otherwise, the
+implementation gracefully falls back to CPU execution while exposing the same
+API.
+
+The hashing algorithm implemented is SHA256 and mirrors the one from
+:mod:`hashlib`. The implementation is self-contained and relies only on basic
+PyTorch tensor operations, making it suitable for execution on both CPU and
+GPU devices.
+"""
+
+from __future__ import annotations
+
+import importlib
+from typing import Any
+
+from typing_extensions import override
+
+from model_signing._hashing import hashing
+
+
+# PyTorch lacks bitwise shifts for unsigned tensors, so we operate on int64 and
+# mask values to 32 bits explicitly.
+# NOTE: `_MASK32` is a Python integer so that it can be broadcast on any device
+# without an explicit copy.
+_MASK32 = 0xFFFFFFFF
+
+
+_K_VALUES = [
+    0x428A2F98,
+    0x71374491,
+    0xB5C0FBCF,
+    0xE9B5DBA5,
+    0x3956C25B,
+    0x59F111F1,
+    0x923F82A4,
+    0xAB1C5ED5,
+    0xD807AA98,
+    0x12835B01,
+    0x243185BE,
+    0x550C7DC3,
+    0x72BE5D74,
+    0x80DEB1FE,
+    0x9BDC06A7,
+    0xC19BF174,
+    0xE49B69C1,
+    0xEFBE4786,
+    0x0FC19DC6,
+    0x240CA1CC,
+    0x2DE92C6F,
+    0x4A7484AA,
+    0x5CB0A9DC,
+    0x76F988DA,
+    0x983E5152,
+    0xA831C66D,
+    0xB00327C8,
+    0xBF597FC7,
+    0xC6E00BF3,
+    0xD5A79147,
+    0x06CA6351,
+    0x14292967,
+    0x27B70A85,
+    0x2E1B2138,
+    0x4D2C6DFC,
+    0x53380D13,
+    0x650A7354,
+    0x766A0ABB,
+    0x81C2C92E,
+    0x92722C85,
+    0xA2BFE8A1,
+    0xA81A664B,
+    0xC24B8B70,
+    0xC76C51A3,
+    0xD192E819,
+    0xD6990624,
+    0xF40E3585,
+    0x106AA070,
+    0x19A4C116,
+    0x1E376C08,
+    0x2748774C,
+    0x34B0BCB5,
+    0x391C0CB3,
+    0x4ED8AA4A,
+    0x5B9CCA4F,
+    0x682E6FF3,
+    0x748F82EE,
+    0x78A5636F,
+    0x84C87814,
+    0x8CC70208,
+    0x90BEFFFA,
+    0xA4506CEB,
+    0xBEF9A3F7,
+    0xC67178F2,
+]
+
+
+_H0_VALUES = [
+    0x6A09E667,
+    0xBB67AE85,
+    0x3C6EF372,
+    0xA54FF53A,
+    0x510E527F,
+    0x9B05688C,
+    0x1F83D9AB,
+    0x5BE0CD19,
+]
+
+
+_TORCH: Any | None = None
+
+
+def _ensure_torch() -> Any:
+    """Import :mod:`torch` on demand."""
+    global _TORCH
+    if _TORCH is None:
+        try:
+            _TORCH = importlib.import_module("torch")
+        except ModuleNotFoundError as exc:
+            raise ModuleNotFoundError(
+                "TorchSHA256 requires the optional 'torch' dependency; install "
+                "with `pip install model-signing[gpu]`"
+            ) from exc
+    return _TORCH
+
+
+def _rotr(x: Any, n: int, torch: Any) -> Any:
+    """Right rotation for 32-bit tensors."""
+    return ((x >> n) | (x << (32 - n))) & _MASK32
+
+
+def _sha256_torch(data: bytes, device: Any) -> bytes:
+    """Pure PyTorch SHA256 implementation."""
+    torch = _ensure_torch()
+    msg = bytearray(data)
+    bit_len = (len(msg) * 8) & 0xFFFFFFFFFFFFFFFF
+    msg.append(0x80)
+    while (len(msg) * 8) % 512 != 448:
+        msg.append(0)
+    msg.extend(bit_len.to_bytes(8, "big"))
+
+    words = torch.tensor(
+        [int.from_bytes(msg[i : i + 4], "big") for i in range(0, len(msg), 4)],
+        dtype=torch.int64,
+        device=device,
+    )
+
+    k = torch.tensor(_K_VALUES, dtype=torch.int64, device=device)
+    h = torch.tensor(_H0_VALUES, dtype=torch.int64, device=device)
+    for chunk_start in range(0, words.shape[0], 16):
+        w = torch.zeros(64, dtype=torch.int64, device=device)
+        w[:16] = words[chunk_start : chunk_start + 16]
+        for i in range(16, 64):
+            s0 = (
+                _rotr(w[i - 15], 7, torch)
+                ^ _rotr(w[i - 15], 18, torch)
+                ^ (w[i - 15] >> 3)
+            )
+            s1 = (
+                _rotr(w[i - 2], 17, torch)
+                ^ _rotr(w[i - 2], 19, torch)
+                ^ (w[i - 2] >> 10)
+            )
+            w[i] = (w[i - 16] + s0 + w[i - 7] + s1) & _MASK32
+
+        a, b, c, d, e, f, g, hv = h
+        for i in range(64):
+            s1 = _rotr(e, 6, torch) ^ _rotr(e, 11, torch) ^ _rotr(e, 25, torch)
+            ch = (e & f) ^ (((~e) & _MASK32) & g)
+            temp1 = (hv + s1 + ch + k[i] + w[i]) & _MASK32
+            s0 = _rotr(a, 2, torch) ^ _rotr(a, 13, torch) ^ _rotr(a, 22, torch)
+            maj = (a & b) ^ (a & c) ^ (b & c)
+            temp2 = (s0 + maj) & _MASK32
+
+            hv = g
+            g = f
+            f = e
+            e = (d + temp1) & _MASK32
+            d = c
+            c = b
+            b = a
+            a = (temp1 + temp2) & _MASK32
+
+        h = (h + torch.stack([a, b, c, d, e, f, g, hv])) & _MASK32
+
+    return b"".join(int(x.item()).to_bytes(4, "big") for x in h)
+
+
+class TorchSHA256(hashing.StreamingHashEngine):
+    """SHA256 hashing engine powered by PyTorch."""
+
+    def __init__(self, initial_data: bytes = b"", device: Any | None = None):
+        torch = _ensure_torch()
+        self._buffer = bytearray(initial_data)
+        if device is None:
+            device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+        self._device = torch.device(device)
+
+    @override
+    def update(self, data: bytes) -> None:
+        self._buffer.extend(data)
+
+    @override
+    def reset(self, data: bytes = b"") -> None:
+        self._buffer = bytearray(data)
+
+    @override
+    def compute(self) -> hashing.Digest:
+        digest_bytes = _sha256_torch(bytes(self._buffer), self._device)
+        return hashing.Digest(self.digest_name, digest_bytes)
+
+    @property
+    @override
+    def digest_name(self) -> str:
+        return "sha256"
+
+    @property
+    @override
+    def digest_size(self) -> int:
+        return 32

tests/_hashing/gpu_test.py

@@ -0,0 +1,33 @@
+"""Tests for GPU backed hashing engines."""
+
+# Copyright 2024 The Sigstore Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import pytest
+
+from model_signing._hashing.gpu import TorchSHA256
+from model_signing._hashing.memory import SHA256
+
+
+pytest.importorskip("torch")
+
+
+def test_torch_sha256_matches_hashlib():
+    data = b"sigstore"
+    gpu_hasher = TorchSHA256(data)
+    cpu_hasher = SHA256(data)
+    gpu_digest = gpu_hasher.compute()
+    cpu_digest = cpu_hasher.compute()
+    assert gpu_digest.digest_value == cpu_digest.digest_value
+    assert gpu_digest.algorithm == cpu_digest.algorithm == "sha256"