MOVD	iv+32(FP), IVP
	MOVD	enc+40(FP), ENC
	MOVD	nr+48(FP), ROUNDS

#ifdef NEEDS_ESPERM
	MOVD	$·rcon(SB), R11
	LVX	(R11), ESPERM   // Permute value for P8_ macros.
#endif

	// Assume len > 0 && len % blockSize == 0.
	CMPW	ENC, $0
	P8_LXVB16X(IVP, R0, IVEC)
	CMPU	ROUNDS, $10, CR1
	CMPU	ROUNDS, $12, CR2 // Only sizes 10/12/14 are supported.

	// Setup key in VSRs, and set loop count in CTR.
	LOAD_KEY(KEYP)

	var erasure Erasure
	if !latestMeta.Deleted && !latestMeta.IsRemote() {
		// Initialize erasure coding
		erasure, err = NewErasure(ctx, latestMeta.Erasure.DataBlocks,
			latestMeta.Erasure.ParityBlocks, latestMeta.Erasure.BlockSize)
		if err != nil {
			return result, err
		}
	}

	result.ObjectSize, err = latestMeta.ToObjectInfo(bucket, object, true).GetActualSize()
	if err != nil {
		return result, err
	}

  int64_t out_channel = filter_shape[3];
  // Value/Op before reshape op.
  Value before_reshape_value = filter;
  if (height % block_size != 0 || width % block_size != 0) {
    // Calculate paddings for height and width.
    int32_t pad_h = block_size - height % block_size;
    int32_t pad_w = block_size - width % block_size;
    auto pad_op =
        GetPadOpForConv2DFilter(filter_shape, filter, &builder, pad_h, pad_w);

  float curr_sparsity = 0;
  std::vector<int> selected_block_size;
  result.needs_densify = true;
  for (const auto& block_size : supported_block_size) {
    curr_sparsity = CalculateBlockSparsity(attr, type, block_size);
    if (curr_sparsity / random_sparsity > ratio_threshold) {
      selected_block_size = block_size;
      result.can_compress = true;
      result.needs_densify = false;

def LegalizeSpaceToDepth : Pat<
  (TF_SpaceToDepthOp $input, $block_size, IsDataFormatNHWC:$data_format),
  (TFL_SpaceToDepthOp $input, (convertIntAttrTo32Bit $block_size))>;

def LegalizeDepthToSpace : Pat<
  (TF_DepthToSpaceOp $input, $block_size, IsDataFormatNHWC:$data_format),
  (TFL_DepthToSpaceOp $input, (convertIntAttrTo32Bit $block_size))>;

def LegalizeResizeBilinear : Pat<

// CHECK: return %[[SUB_0]] : tensor<15x28x28x1xf32>
}

// CHECK-LABEL: depth_to_space
func.func @depth_to_space(%arg0: tensor<1x1x1x4xf32>) -> tensor<1x2x2x1xf32> {
  %0 = "tf.DepthToSpace"(%arg0) {block_size = 2: i64,  data_format = "NHWC"}: (tensor<1x1x1x4xf32>) -> tensor<1x2x2x1xf32>
  func.return %0 : tensor<1x2x2x1xf32>

table BatchToSpaceNDOptions {
}

table SkipGramOptions {
  ngram_size: int;
  max_skip_size: int;
  include_all_ngrams: bool;
}

table SpaceToDepthOptions {
  block_size: int;
}

table DepthToSpaceOptions {
  block_size: int;
}

table SubOptions {
  fused_activation_function:ActivationFunctionType;
  // Parameters supported by version 5
  pot_scale_int16:bool = true;
}

}

table SkipGramOptions {
  ngram_size: int;
  max_skip_size: int;
  include_all_ngrams: bool;
}

table SpaceToDepthOptions {
  block_size: int;
}

table DepthToSpaceOptions {
  block_size: int;
}

table SubOptions {
  fused_activation_function:ActivationFunctionType;
  // Parameters supported by version 5
  pot_scale_int16:bool = true;
}

	// on systems with coarse timer granularity.
	const blockTime = 100 * time.Millisecond

	// Make sure the goroutine spawned above actually blocks on the lock.
	for {
		if runtime.GIsWaitingOnMutex(gp) {
			break
		}
		runtime.Gosched()
	}

	// Let some amount of time pass.
	time.Sleep(blockTime)

	// Let the other goroutine acquire the lock.
	mu.Unlock1()
	done <- true