MOVOU	X15, 240(DI)
	SUBQ	$256, BX
	ADDQ	$256, DI
	CMPQ	BX, $256
	JAE	loop
	JMP	tail
#endif

loop_preheader_avx2:
	VPXOR X0, X0, X0
	// For smaller sizes MOVNTDQ may be faster or slower depending on hardware.
	// For larger sizes it is always faster, even on dual Xeons with 30M cache.
	// TODO take into account actual LLC size. E. g. glibc uses LLC size/2.
	CMPQ    BX, $0x2000000
	JAE	loop_preheader_avx2_huge

loop_avx2:

TensorFlow in a multitenant design mixes the risks described above with the
inherent ones from multitenant configurations. The primary areas of concern are
tenant isolation, resource allocation, model sharing and hardware attacks.

### Tenant isolation

Since any tenants or users providing models, graphs or checkpoints can execute
code in context of the TensorFlow service, it is important to design isolation

// license that can be found in the LICENSE file.

package windows

import (
	"errors"
	"sync"
	"syscall"
	"unsafe"
)

// https://learn.microsoft.com/en-us/windows-hardware/drivers/ddi/wdm/ns-wdm-_osversioninfow
type _OSVERSIONINFOW struct {
	osVersionInfoSize uint32
	majorVersion      uint32
	minorVersion      uint32
	buildNumber       uint32
	platformId        uint32

// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.

// This file contains CRC32 algorithms that are not specific to any architecture
// and don't use hardware acceleration.
//
// The simple (and slow) CRC32 implementation only uses a 256*4 bytes table.
//
// The slicing-by-8 algorithm is a faster implementation that uses a bigger
// table (8*256*4 bytes).

package crc32

// exported as a TF SavedModel.
void AddCallModuleSerializationPasses(OpPassManager& pm);

// Passes for unpacking quantized ops to int valued StableHLO ops. This is
// useful when uniform quantized types are suboptimal for the hardware. It goes
// through a StableHLO <-> MHLO roundtrip to utilize the MHLOQuantToInt pass.
void AddStablehloQuantToIntPasses(OpPassManager& pm);

// Processes tensors with NCHW format (== (batch, channel, height, weight)) by

// can be used for increased type safety within applications.
type PrivateKey any

// Signer is an interface for an opaque private key that can be used for
// signing operations. For example, an RSA key kept in a hardware module.
type Signer interface {
	// Public returns the public key corresponding to the opaque,
	// private key.
	Public() PublicKey

	// Sign signs digest with the private key, possibly using entropy from

The actual name is semantically irrelevant but should be used to document
the argument's name.
It is worth stressing that <code>FP</code> is always a
pseudo-register, not a hardware
register, even on architectures with a hardware frame pointer.
</p>

<p>
For assembly functions with Go prototypes, <code>go</code> <code>vet</code> will check that the argument names
and offsets match.

    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "FacwMD9MPUew"
      },
      "source": [
        "StableHLO Quantizer is a quantization API to enable ML framework optionality and hardware retargetability."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "RXZUHZQoQZOo"
      },
      "outputs": [],
      "source": [

//
//     When AES hardware is available, AES-128-GCM and AES-256-GCM are faster
//     than ChaCha20Poly1305.
//
//     When AES hardware is not available, AES-128-GCM is one or more of: much
//     slower, way more complex, and less safe (because not constant time)
//     than ChaCha20Poly1305.
//
//     We use this list if we think both peers have AES hardware, and

        // The hardware is big-endian, so we need to reverse the order of the bytes.
        return Long.reverseBytes(bigEndian);
      }

      @Override
      public void putLongLittleEndian(byte[] array, int offset, long value) {
        // Reverse the order of the bytes before storing, since we're on big-endian hardware.
        long littleEndianValue = Long.reverseBytes(value);