# ADR-0001 - Use Architectural Decision Records

## Date

2023-12-01

## Context

In a distributed team with many subteams, the best solution to communicate decisions is to use a format accessible by everyone in charge of development.

We use *Specification* and *Discovery* documents stored in Google Drive, but they present some downsides:

  func.return %0#0, %0#1 : tensor<2xf32>, tensor<2xf32>

// CHECK-NEXT: %[[split:.*]]:2 = "tfl.split"
// CHECK-NEXT: %[[stats1:.*]] = "quantfork.stats"(%[[split]]#0) <{layerStats = dense<[-1.000000e+00, 1.000000e+00]>
// CHECK-NEXT: %[[stats2:.*]] = "quantfork.stats"(%[[split]]#1) <{layerStats = dense<[-1.000000e+00, 1.000000e+00]>

// CHECK-NEXT: %[[Y:[a-z0-9]*]], {{.*}}_FusedBatchNormEx
// CHECK-NEXT: return %[[Y]]
  %y:6 = "tf.FusedBatchNormV3"(%arg0, %arg1, %arg2, %arg3, %arg4) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", epsilon = 0.001 : f32, is_training = false} : (tensor<8x8x8x8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>) -> (tensor<8x8x8x8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>)

//   %0:6 = "tf.FusedBatchNormV3"(%cst_0, %cst_1, %cst_2, %cst_1, %cst_2) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", epsilon = 0.001 : f32, is_training = false} : (tensor<1x1x2x8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>) -> (tensor<1x1x2x8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>)

    tf_executor.fetch %ro#0, %ro#1 : tensor<*x!tf_type.string>, tensor<*x!tf_type.string>
  }
  return %0#0, %0#1 : tensor<*x!tf_type.string>, tensor<*x!tf_type.string>
}

// -----

//  Serialized string:
//   "\08\01\1A\01\01\22\01\00"
//  Proto debug string:
//   type: MAXIMAL
//   tile_assignment_dimensions: 1
//   tile_assignment_devices: 0

    "tf.Yield"(%3, %4) : (tensor<i32>, tensor<i32>) -> ()
  }) {is_stateless = true, parallel_iterations = 10 : i64} : (tensor<i32>, tensor<i32>) -> (tensor<i32>, tensor<i32>)
  return %0#0, %0#1 : tensor<i32>, tensor<i32>
}

// -----

// Check that `tf.ReadVariableOp` is hoisted because the variable is readonly.
// The following ops that depend on `tf.ReadVariableOp` and satisfy hoisting

0.675 responseHeadersEnd
0.676 responseBodyStart
0.679 responseBodyEnd
0.679 connectionReleased
0.680 callEnd
REQUEST 2 (pooled connection)
0.000 callStart
0.001 connectionAcquired
0.001 requestHeadersStart
0.001 requestHeadersEnd
0.002 responseHeadersStart
0.082 responseHeadersEnd
0.082 responseBodyStart
0.082 responseBodyEnd
0.083 connectionReleased
0.083 callEnd
```

(0x0,0x0)
(0x0,0x0)
0x0
[0/0]0x0
-7
7
7
7
7
7
7
+8.000000e+000
(+9.000000e+000+1.000000e+001i)
true
false
hello
one two
one two
hello
false
true
(+1.400000e+001+1.500000e+001i)
+1.300000e+001
12
12
12
12
12
12
-11
[0/0]0x0
0x0
(0x0,0x0)

	  tf_device.return %3#0, %3#1 : tensor<2048xi32>, tensor<2048xi32>
	}) {device = "TPU_REPLICATED_HOST_0"} : () -> (tensor<2048xi32>, tensor<2048xi32>)
    return %0#0, %0#1: tensor<2048xi32>, tensor<2048xi32>
  }

  // CHECK-LABEL: func @valid_copy_op_in_non_replicated_host

  // CHECK: "tf_device.launch"
  // CHECK-SAME: "/job:localhost/replica:0/task:0/device:CPU:0"

    tf_device.return %arg0, %arg1 : tensor<!tf_type.string>, tensor<!tf_type.string>
  }) : () -> (tensor<!tf_type.string>, tensor<!tf_type.string>)
  // CHECK: return %arg0, %arg1
  func.return %0#0, %0#1 : tensor<!tf_type.string>, tensor<!tf_type.string>
}

// Verifies handling op a cluster op requiring no rewrites.
// CHECK-LABEL: func @canonical_cluster