// RUN: stablehlo-quant-opt %s -stablehlo-merge-fusion-with-dequantize -split-input-file -verify-diagnostics | FileCheck %s

// Merge fusion with dequantize for relu case.

module attributes {tf_saved_model.semantics} {
  // CHECK-LABEL: func.func private @merge_relu_fusion
  func.func private @merge_relu_fusion(%arg0: tensor<1x4xf32>) -> tensor<1x3xf32> {

// CHECK-DAG: %[[input_11:.*]] = "tfl.dequantize"({{.*}}) : (tensor<3x!quant.uniform<i32:f32, 4.8060645560249487E-8>>) -> tensor<3xf32>
// CHECK-DAG: %[[input_12:.*]] = "tfl.dequantize"({{.*}}) : (tensor<3x!quant.uniform<i32:f32, 7.2090970130772759E-8>>) -> tensor<3xf32>

// RUN: tf-opt %s -tfl-prepare-quantize="post-training-quantize=true" -tfl-quantize="numeric-verify=true log-if-failed=true" | FileCheck --check-prefix=DEBUG %s
// RUN: tf-opt %s -tfl-prepare-quantize="post-training-quantize=true" -tfl-quantize="numeric-verify=true log-if-failed=true whole-model-verify=true" | FileCheck --check-prefix=MODEL-DEBUG %s

// DEBUG-LABEL: QuantizeConv2D

// CHECK: %[[QUANTIZE:.*]] = "tfl.quantize"(%[[CST]]) <{qtype = tensor<1x32x42x128x!quant.uniform<i8<-127:127>:f32:0, {1.000000e+00}>>}> {volatile}
// CHECK: %[[DEQUANTIZE:.*]] = "tfl.dequantize"(%[[QUANTIZE]]) : (tensor<1x32x42x128x!quant.uniform<i8<-127:127>:f32:0, {1.000000e+00}>>) -> tensor<1x32x42x128xf32>
// CHECK: "tfl.transpose_conv"(%arg1, %[[DEQUANTIZE]], %arg0,

// RUN: tf-opt %s -tfl-prepare-quantize -tfl-quantize  | FileCheck %s
// RUN: tf-opt %s -tfl-quantize="legacy-quantize=true" | FileCheck --check-prefix=LEGACY %s
// RUN: tf-opt %s -tfl-prepare-quantize -tfl-quantize="ops-blocklist=tfl.fully_connected,tfl.softmax locs-blocklist=Block,NullBlock" | FileCheck --check-prefix=BLOCK %s

// CHECK-LABEL: QuantizeFloatConst
func.func @QuantizeFloatConst() -> tensor<2x2x!quant.uniform<u8:f32, 7.8431372549019615E-4:128>> {

// RUN: tf-opt %s -tfl-quantize-variables | FileCheck %s
// RUN: tf-opt %s -tfl-prepare-quantize -tfl-quantize -tfl-post-quantize -tfl-quantize-variables -tfl-quantize -tfl-post-quantize | FileCheck --check-prefix=WHOLE-PASSES %s

// CHECK-LABEL: QuantizeReadVariable
func.func @QuantizeReadVariable() -> (tensor<1x2x1x3x!quant.uniform<i8:f32, 1.0>>) {
  %1 = "tfl.var_handle"() : () -> tensor<!tf_type.resource>

// CHECK-DAG: %[[q_w1:.*]] = "tfl.quantize"(%[[w]]) <{qtype = tensor<64x3x3x3x!quant.uniform<i8<-127:127>:f32:3, {1.000000e+00,1.000000e+00,1.000000e+00}>
// CHECK-DAG: %[[q_w2:.*]] = "tfl.quantize"(%[[w]]) <{qtype = tensor<64x3x3x3x!quant.uniform<i8<-127:127>:f32:0, {1.000000e+00,1.000000e+00,1.000000e+00
// CHECK-DAG: %[[dq_w1:.*]] = "tfl.dequantize"(%[[q_w1]])
// CHECK-DAG: %[[dq_w2:.*]] = "tfl.dequantize"(%[[q_w2]])

// RUN: tf-opt %s -tfl-prepare-quantize-dynamic-range -tfl-quantize="enable-dynamic-range-quantization=true" | FileCheck %s
// RUN: tf-opt %s -tfl-prepare-quantize-dynamic-range -tfl-quantize="enable-dynamic-range-quantization=true enable-weight-only-quantization=true" | FileCheck --check-prefix=PerChannelWeightOnly %s

  func.return %rst : tensor<8xf32>

// CHECK: %[[CONSTANT:.*]] = arith.constant dense<0.000000e+00> : tensor<8xf32>
// CHECK: %[[QUANTIZE:.*]] = "tfl.quantize"(%[[CONSTANT]]) <{qtype = tensor<8x!quant.uniform<u8:f32, 1.000000e+00>>}>
// CHECK: %[[DEQUANTIZE:.*]] = "tfl.dequantize"(%[[QUANTIZE]])
// CHECK: return %[[DEQUANTIZE]] : tensor<8xf32>
}

// CHECK-LABEL: fakeQuantFolded
func.func @fakeQuantFolded() -> (tensor<8xf32>) {

  func.return %rst : tensor<8xf32>

// CHECK: %[[CONSTANT:.*]] = arith.constant dense<0.000000e+00> : tensor<8xf32>
// CHECK: %[[QUANTIZE:.*]] = "tfl.quantize"(%[[CONSTANT]]) <{qtype = tensor<8x!quant.uniform<u4:f32, 1.000000e+00>>}>
// CHECK: %[[DEQUANTIZE:.*]] = "tfl.dequantize"(%[[QUANTIZE]])
// CHECK: return %[[DEQUANTIZE]] : tensor<8xf32>
}

// CHECK-LABEL: fakeQuantFolded
func.func @fakeQuantFolded() -> (tensor<8xf32>) {