let regions = (region SizedRegion<1>:$body);
  let hasVerifier = 1;
}

def Quantization_ReturnOp : Quantization_Op<"return", [Terminator]> {
  let summary = [{
    The `return` operation terminates a quantize region and returns values.
  }];

  let arguments = (ins Variadic<AnyTensor>:$results);
}

//===----------------------------------------------------------------------===//

#include "absl/base/call_once.h"
#include "llvm-c/Target.h"
#include "llvm/Support/ManagedStatic.h"
#include "tensorflow/compiler/aot/codegen.h"
#include "tensorflow/compiler/aot/flags.h"
#include "tensorflow/compiler/aot/quantize.h"
#include "tensorflow/compiler/tf2xla/tf2xla.h"
#include "tensorflow/compiler/tf2xla/tf2xla_util.h"
#include "xla/client/client_library.h"
#include "xla/client/compile_only_client.h"
#include "xla/client/xla_computation.h"

      func.return %add : tensor<3x2xf32>
    }
  }
)mlir";

// Quantizable ops: XlaCallModule op with "fully_quantizable" attribute and
// same-scale StableHLO ops
// Non-quantizable ops: quantize/dequantize ops
constexpr absl::string_view kModuleCompositeSameScale = R"mlir(
  module {
    func.func @same_scale_after_composite() -> tensor<3x1xf32> {

// RUN: tf-quant-opt %s -split-input-file -quant-lift-quantizable-spots-as-functions -quant-quantize='target-opset=XLA' -verify-each=false | FileCheck %s

func.func private @conv(%input: tensor<1x3x4x3xf32> {tf._user_specified_name = "input_tensor"}) -> tensor<*xf32> attributes {tf._construction_context = "kEagerRuntime", tf._input_shapes = [#tf_type.shape<1x3x4x3>]} {
  %weight = arith.constant dense_resource<__elided__> : tensor<2x3x3x2xf32>

  let regions = (region SizedRegion<1>:$body);
  let hasVerifier = 1;
}

def quantfork_ReturnOp : quantfork_Op<"return", [Terminator]> {
  let summary = [{
    The `return` operation terminates a quantize region and returns values.
  }];

  let arguments = (ins Variadic<AnyTensor>:$results);
}

//===----------------------------------------------------------------------===//

  // CHECK-DAG: %[[VAL5:.+]] = "tfl.quantize"(%4) <{qtype = tensor<1x32x32x16x!quant.uniform<i8:f32, 1.000000e+00>>}>
  // CHECK: return %[[VAL5]]

        "passes/prepare_quantize.inc",
        "passes/prepare_quantize_drq.cc",
        "passes/preprocess_op.cc",
        "passes/preprocess_op.inc",
        "passes/propagate_quantize_type.cc",
        "passes/quantize.cc",
        "passes/quantize_composite_functions.cc",
        "passes/quantize_composite_functions.inc",
        "passes/quantize_weights.cc",
        "passes/quantized_function_library.h",

    bool enable_legacy_weight_only = false,
    std::optional<const absl::string_view> mlir_dump_file_prefix =
        std::nullopt);

// Converts dequantize-(quantizable) call-quantize pattern to a single call op
// that has quantized input and output types. It is expected for this pass to
// emit illegal IR with unsupported quantized input and output types. The

	return func() float64 {
		sum := f()
		for _, s := range fs {
			sum += s()
		}
		return sum
	}
}

// quantize returns a new stream that rounds f to a multiple
// of mult at each step.
func (f float64Stream) quantize(mult float64) float64Stream {
	return func() float64 {
		r := f() / mult
		if r < 0 {
			return math.Ceil(r) * mult
		}
		return math.Floor(r) * mult

Cela prendrait autant de temps pour finir avec ou sans sections (concurrence) et vous auriez effectué la même quantité de travail.