// Sqrt op patterns.
//===----------------------------------------------------------------------===//

// SqrtGrad(y, dy) = dy * 0.5 / y
def LowerSqrtGradOp : Pat<
  (TF_SqrtGradOp $y, $dy),
  (TF_DivOp
    (TF_MulOp $dy, (TF_ConstOp (GetScalarOfFloatType<"0.5"> $dy))),
    $y
  )>;

//===----------------------------------------------------------------------===//
// TanhGrad op patterns.

	e.buf[0] = sImageDescriptor
	byteorder.LePutUint16(e.buf[1:3], uint16(b.Min.X))
	byteorder.LePutUint16(e.buf[3:5], uint16(b.Min.Y))
	byteorder.LePutUint16(e.buf[5:7], uint16(b.Dx()))
	byteorder.LePutUint16(e.buf[7:9], uint16(b.Dy()))
	e.write(e.buf[:9])

	// To determine whether or not this frame's palette is the same as the
	// global palette, we can check a couple things. First, do they actually

		LoopCount: 5,
	}
	for i, f := range frames {
		g, err := readGIF(f)
		if err != nil {
			t.Fatal(f, err)
		}
		m := g.Image[0]
		if m.Bounds().Dx() != width || m.Bounds().Dy() != height {
			t.Fatalf("frame %d had unexpected bounds: got %v, want width/height = %d/%d",
				i, m.Bounds(), width, height)
		}
		g0.Image[i] = m
	}

// Training OPs
// =============================================================================

// `grad = dy * y * (1 - y)`, where `y = sigmoid(x)`
def LegalizeSigmoidGrad : Pat<(TF_SigmoidGradOp $y, $dy),
  (TFL_MulOp $dy, (TFL_MulOp $y, (TFL_SubOp
        (Arith_ConstantOp ConstantAttr<RankedF32ElementsAttr<[]>, "1.0f">),
        $y, TFL_AF_None),

// defer list means that the main goroutine's defers can still
// be handled non-atomically.
//
// In the diagram, dY and dX are meant to be processed when
// drangefunc would be processed, which is to say the defer order
// should be d4, d3, dY, dX, d2, d1. To make that happen,
// when defer processing reaches a d with rangefunc=true,
// it calls deferconvert to atomically take the extras

000001c0  b1 f3 66 30 f8 aa 90 f7  90 49 19 48 4c 25 4a 1f  |..f0.....I.HL%J.|
000001d0  12 9d 67 32 79 bb 53 d8  c5 d1 b4 6e 89 75 49 c0  |..g2y.S....n.uI.|
000001e0  65 86 ac 72 23 2f 97 d3  ae e2 64 79 5e e2 10 4e  |e..r#/....dy^..N|
000001f0  55 0c c6 70 d3 2e 4a 6c  b0 73 0a 11 eb ae f7 a1  |U..p..Jl.s......|
00000200  a1 f0 5f 67 45 46 d3 8c  11 ff 21 62 7d ed f9 0e  |.._gEF....!b}...|

000001e0  6b df 54 88 a8 f3 af 0d  4a 1b 29 91 3a bb 90 27  |k.T.....J.).:..'|
000001f0  29 3a 97 17 03 03 00 17  09 78 1d f3 95 2e 92 af  |):.......x......|
00000200  d8 af 64 59 fb 77 e3 85  de f0 37 be 11 91 2f 17  |..dY.w....7.../.|
00000210  03 03 00 13 4e a2 06 c3  90 6f d3 b0 04 a8 25 32  |....N....o....%2|

example, suppose y = f(x) and we wish to apply a custom function g for backprop
such that dx = g(dy). In Python,

```python
with tf.get_default_graph().gradient_override_map(
    {'IdentityN': 'OverrideGradientWithG'}):
  y, _ = identity_n([f(x), x])

@tf.RegisterGradient('OverrideGradientWithG')
def ApplyG(op, dy, _):
  return [None, g(dy)]  # Do not backprop to f(x).
```
  }];

  let arguments = (ins

github.com/prometheus/client_model v0.6.0 h1:k1v3CzpSRUTrKMppY35TLwPvxHqBu0bYgxZzqGIgaos=
github.com/prometheus/client_model v0.6.0/go.mod h1:NTQHnmxFpouOD0DpvP4XujX3CdOAGQPoaGhyTchlyt8=
github.com/prometheus/common v0.48.0 h1:QO8U2CdOzSn1BBsmXJXduaaW+dY/5QLjfB8svtSzKKE=
github.com/prometheus/common v0.48.0/go.mod h1:0/KsvlIEfPQCQ5I2iNSAWKPZziNCvRs5EC6ILDTlAPc=
github.com/prometheus/procfs v0.12.0 h1:jluTpSng7V9hY0O2R9DzzJHYb2xULk9VTR1V1R/k6Bo=

github.com/prometheus/client_model v0.6.0 h1:k1v3CzpSRUTrKMppY35TLwPvxHqBu0bYgxZzqGIgaos=
github.com/prometheus/client_model v0.6.0/go.mod h1:NTQHnmxFpouOD0DpvP4XujX3CdOAGQPoaGhyTchlyt8=
github.com/prometheus/common v0.48.0 h1:QO8U2CdOzSn1BBsmXJXduaaW+dY/5QLjfB8svtSzKKE=
github.com/prometheus/common v0.48.0/go.mod h1:0/KsvlIEfPQCQ5I2iNSAWKPZziNCvRs5EC6ILDTlAPc=
github.com/prometheus/procfs v0.12.0 h1:jluTpSng7V9hY0O2R9DzzJHYb2xULk9VTR1V1R/k6Bo=