// range.
    CALIBRATION_METHOD_HISTOGRAM_MSE_SYMMETRIC = 6;
  }

  // Parameters required for calibration.
  // Next ID: 4
  message CalibrationParameters {
    // The number of histogram bins. Default to 512.
    int32 num_bins = 1;
    // min_percentile is only used in HISTOGRAM_PERCENTILE.
    // min_percentile is 0.001 by default.
    float min_percentile = 2;

   * stripes, the intended concurrency level, and the typical number of keys used in a {@code
   * bulkGet(keys)} operation. See <a href="http://www.mathpages.com/home/kmath199.htm">Balls in
   * Bins model</a> for mathematical formulas that can be used to estimate the probability of
   * collisions.
   *
   * @param keys arbitrary non-null keys

		count[i] += count[i-1]
	}
	if count[len(count)-1] != int32(len(values)) {
		f.Fatalf("storeOrder: value is missing, total count = %d, values = %v", count[len(count)-1], values)
	}

	// place values in count-indexed bins, which are in the desired store order
	order := make([]*Value, len(values))
	for _, v := range values {
		s := storeNumber[v.ID]
		order[count[s-1]] = v
		count[s-1]++
	}

   * stripes, the intended concurrency level, and the typical number of keys used in a {@code
   * bulkGet(keys)} operation. See <a href="http://www.mathpages.com/home/kmath199.htm">Balls in
   * Bins model</a> for mathematical formulas that can be used to estimate the probability of
   * collisions.
   *
   * @param keys arbitrary non-null keys

			var a [16]byte
			m := make(map[uint16]struct{}, 1<<16)
			for n := 0; n < 1<<16; n++ {
				a[i] = byte(n)
				a[j] = byte(n >> 8)
				m[uint16(BytesHash(a[:], 0))] = struct{}{}
			}
			// N balls in N bins, for N=65536
			avg := 41427
			stdDev := 123
			if len(m) < avg-40*stdDev || len(m) > avg+40*stdDev {
				t.Errorf("bad number of collisions i=%d j=%d outputs=%d out of 65536\n", i, j, len(m))
			}
		}
	}

// CHECK: return %[[v0]]
}

func.func @batch_multilhs_einsum(%arg0: tensor<2x1x1x11xf32>, %arg1: tensor<2x11x2xf32>) -> tensor<2x1x1x2xf32> {
  %0 = "tf.Einsum"(%arg0, %arg1) {T = "tfdtype$DT_FLOAT", equation = "BiNj,BjS->BiNS"} : (tensor<2x1x1x11xf32>, tensor<2x11x2xf32>) -> tensor<2x1x1x2xf32>
  func.return %0 : tensor<2x1x1x2xf32>
// CHECK-LABEL: batch_multilhs_einsum
// CHECK-DAG: %[[cst:.*]] = arith.constant dense<[2, 1, 11]> : tensor<3xi64>

if(p1===x.bins.length)throw new Error('Cannot subtract');if(x.bins[p1].max!==y.bins[p2].max){throw new Error('Incompatible bins');}
if(x.bins[p1].count<y.bins[p2].count){throw new Error('Cannot subtract');}
x.bins[p1].count-=y.bins[p2].count;for(const event of y.bins[p2].events){x.bins[p1].events.add(event);}

            `num_tokens`.
        *   `Discretization` added `num_bins` argument for learning bins
            boundaries through calling `adapt` on a dataset. Renamed `bins`
            argument to `bin_boundaries` for specifying bins without `adapt`.
    *   Improvements to model saving/loading:
        *   `model.load_weights` now accepts paths to saved models.

  );

  let results = (outs
    Res<TensorOf<[TF_Float32, TF_Float64, TF_Int32, TF_Int64]>, [{1D `Tensor` with length equal to `size`. The counts or summed weights for
each value in the range [0, size).}]>:$bins
  );

  TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<2>;
}

def TF_BitcastOp : TF_Op<"Bitcast", [Pure]> {
  let summary = [{