// storage type. The available values use the full range of the storage value,
// i.e. [-128, 127]. Assumes asymmetric quantization, meaning the zero point
// value can be a non-zero value.
// If `narrow_range` is set true (ex: for weights), a restricted range of
// integers will be used for symmetric mapping, i.e. [-127, 127].
UniformQuantizedType CreateI8F32UniformQuantizedType(Location loc,
                                                     MLIRContext& context,

// RUN: tf-quant-opt %s -split-input-file -quant-lift-quantizable-spots-as-functions-drq="min-num-elements-for-weights=2500000" | FileCheck %s

// CHECK-LABEL: lift_float_matmul
func.func @lift_float_matmul(%arg0: tensor<1x12x12x512xf32>) -> (tensor<*xf32>, tensor<*xf32>) {
  %cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<512x512xf32>} : () -> tensor<512x512xf32>
  %out_1 = "tf.MatMul"(%arg0, %cst) {
    device = "", transpose_a = false, transpose_b = false

				weights := make([]int, 0)
				for _, localityEndpoint := range cluster.LoadAssignment.Endpoints {
					weights = append(weights, int(localityEndpoint.LoadBalancingWeight.GetValue()))
				}
				if !reflect.DeepEqual(weights, tt.expected) {
					t.Errorf("Got weights %v expected %v", weights, tt.expected)
				}
			})
		}
	})

// point constant.
def : Pat<(TFL_DequantizeOp
             (TFL_QuantizeOp (Arith_ConstantOp F32ElementsAttr:$cst), $qt)),
          (TFL_ConstOp $cst)>;

// Transpose conv supports hybrid computation with quantized weights.
def FoldQuantWeightsIntoTposeConv : Pat<
  (TFL_TransposeConvOp
    $output_shape,
    (TFL_DequantizeOp $quant_weights),
    $quant_input,
    $bias, $padding, $stride_h, $stride_w, $faf),

    (binaryOp (TFL_TransposeConvOp:$output $output_shape, $weights, $input,
                (Arith_ConstantOp FloatElementsAttr:$bias), $padding,
                $stride_h, $stride_w, TFL_AF_None),
              (Arith_ConstantOp FloatElementsAttr:$value), $act_fn),
    (TFL_TransposeConvOp $output_shape, $weights, $input,
      (binaryOp (Arith_ConstantOp $bias),
         (Arith_ConstantOp $value), TFL_AF_None),

		return scaleFactor
	}
	weight := uint32(math.MaxUint32)
	if ep.GetLoadBalancingWeight().Value < math.MaxUint32/scaleFactor {
		weight = ep.GetLoadBalancingWeight().Value * scaleFactor
	}
	return weight
}

// Apply the weight for this endpoint to the network gateways.

  // Whether to allow weight-only quantization. This scheme quantizes
  // weights but will dequantize them back at runtime which is useful for
  // memory bound case without kernel support available in lower precisions.
  // Used in MLIR dynamic range quantizer.
  bool weight_only_quantization = false;

  // The minimum number of elements in a weights array required to apply

    TFL_TensorOf<[F32, I8]>:$fw_recurrent_to_output_weights,

    // Forward Cell weights
    TFL_TensorOfOrNone<[F32, I8]>:$fw_cell_to_input_weights,
    // Optional Forward cell weights
    TFL_TensorOfOrNone<[F32, I8]>:$fw_cell_to_forget_weights,
    // Optional Forward cell weights
    TFL_TensorOfOrNone<[F32, I8]>:$fw_cell_to_output_weights,

    // Forward Bias

	}
	return postProcessNamedEdgeMap(weight, totalWeight)
}

func sortByWeight(edges []NamedCallEdge, weight map[NamedCallEdge]int64) {
	sort.Slice(edges, func(i, j int) bool {
		ei, ej := edges[i], edges[j]
		if wi, wj := weight[ei], weight[ej]; wi != wj {
			return wi > wj // want larger weight first
		}
		// same weight, order by name/line number
		if ei.CallerName != ej.CallerName {

    // determined. The activation and weight are quantized to INT8 while bias is
    // quantized to INT32.
    METHOD_STATIC_RANGE_INT8 = 2;

    // Dynamic range quantization. Quantized tensor values' ranges are
    // determined in the graph executions. The weights are quantized during
    // conversion.
    METHOD_DYNAMIC_RANGE_INT8 = 3;

    // Weight-only quantization. Only weights are quantized during conversion.