experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Set the calibrated quantization parameters to be volatile

Browse Source 
The QDQ will be removed if there are not fused after quantization.

PiperOrigin-RevId: 354638972
Change-Id: I1a0cd1514f302ec47d1f64fbba4232a354fc9fb6
This commit is contained in:
Feng Liu 2021-01-29 17:02:48 -08:00 committed by TensorFlower Gardener
parent 23d9a2b49d
commit c6de267c22
5 changed files with 19 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
tensorflow/compiler/mlir/lite/quantization/quantization_utils.h
View File

@ -152,6 +152,8 @@ struct ConvertStatsToQDQs : public OpRewritePattern<quant::StatisticsOp> {
rewriter.setInsertionPointAfter(op.getOperation()); rewriter.setInsertionPointAfter(op.getOperation());
Type result_type = quant_type.castFromExpressedType(op.getType()); Type result_type = quant_type.castFromExpressedType(op.getType());
auto q = rewriter.create<Q>(op.getLoc(), result_type, op.arg()); auto q = rewriter.create<Q>(op.getLoc(), result_type, op.arg());
q->setAttr(kVolatileOpAttrName, rewriter.getUnitAttr());
auto dq = rewriter.create<DQ>(op.getLoc(), op.getType(), q); auto dq = rewriter.create<DQ>(op.getLoc(), op.getType(), q);
op.getResult().replaceAllUsesWith(dq); op.getResult().replaceAllUsesWith(dq);
q.getOperation()->replaceUsesOfWith(dq, op.arg()); q.getOperation()->replaceUsesOfWith(dq, op.arg());

10
tensorflow/compiler/mlir/lite/tests/prepare-quantize-post-training.mlir
View File

@ -112,7 +112,7 @@ func @QuantizeWithoutNorm(%arg0: tensor<1x5xf32>) -> tensor<*xf32> attributes {t
// CHECK-SAME: %[[input_9]], %[[input_10]], %[[input_11]], %[[input_12]], %[[input_13]], %[[input_14]], %[[input_15]], %[[input_16]], %[[input_17]], %[[input_18]], %[[input_19]] // CHECK-SAME: %[[input_9]], %[[input_10]], %[[input_11]], %[[input_12]], %[[input_13]], %[[input_14]], %[[input_15]], %[[input_16]], %[[input_17]], %[[input_18]], %[[input_19]]
// CHECK-SAME: effective_hidden_scale_intermediate = tensor<!quant.uniform<i8:f32, 0.0039215686274509803:-1>> // CHECK-SAME: effective_hidden_scale_intermediate = tensor<!quant.uniform<i8:f32, 0.0039215686274509803:-1>>
// CHECK: "tfl.quantize"(%[[lstm]]) {qtype = tensor<*x!quant.uniform<i8:f32, 0.015686274509803921:-1>>} // CHECK: "tfl.quantize"(%[[lstm]]) {qtype = tensor<*x!quant.uniform<i8:f32, 0.015686274509803921:-1>>, volatile}
} }
// CHECK-LABEL: QuantizeLstmCifg // CHECK-LABEL: QuantizeLstmCifg
@ -194,7 +194,7 @@ func @QuantizeLstmCifg(%arg0: tensor<1x5xf32>) -> tensor<*xf32> attributes {tf.e
// CHECK-SAME: input_to_forget_intermediate = tensor<!quant.uniform<i16<-32767:32767>:f32, 4.8829615161595508E-4>> // CHECK-SAME: input_to_forget_intermediate = tensor<!quant.uniform<i16<-32767:32767>:f32, 4.8829615161595508E-4>>
// CHECK-SAME: input_to_output_intermediate = tensor<!quant.uniform<i16<-32767:32767>:f32, 3.0518509475997192E-5>> // CHECK-SAME: input_to_output_intermediate = tensor<!quant.uniform<i16<-32767:32767>:f32, 3.0518509475997192E-5>>
// CHECK: "tfl.quantize"(%[[lstm]]) {qtype = tensor<*x!quant.uniform<i8:f32, 0.015686274509803921:-1>>} // CHECK: "tfl.quantize"(%[[lstm]]) {qtype = tensor<*x!quant.uniform<i8:f32, 0.015686274509803921:-1>>, volatile}
} }
// CHECK-LABEL: QuantizeUnidirectionalLstmFull // CHECK-LABEL: QuantizeUnidirectionalLstmFull
@ -285,7 +285,7 @@ func @QuantizeUnidirectionalLstmFull(%arg0: tensor<1x5xf32>) -> tensor<*xf32> at
// CHECK-SAME: input_to_input_intermediate = tensor<!quant.uniform<i16<-32767:32767>:f32, 9.7659230323191015E-4>> // CHECK-SAME: input_to_input_intermediate = tensor<!quant.uniform<i16<-32767:32767>:f32, 9.7659230323191015E-4>>
// CHECK-SAME: input_to_output_intermediate = tensor<!quant.uniform<i16<-32767:32767>:f32, 3.0518509475997192E-5>> // CHECK-SAME: input_to_output_intermediate = tensor<!quant.uniform<i16<-32767:32767>:f32, 3.0518509475997192E-5>>
// CHECK: "tfl.quantize"(%[[lstm]]) {qtype = tensor<*x!quant.uniform<i8:f32, 0.015686274509803921:-1>>} // CHECK: "tfl.quantize"(%[[lstm]]) {qtype = tensor<*x!quant.uniform<i8:f32, 0.015686274509803921:-1>>, volatile}
} }
// CHECK-LABEL: QuantizeLstmFull // CHECK-LABEL: QuantizeLstmFull
@ -377,7 +377,7 @@ func @QuantizeLstmFull(%arg0: tensor<1x5xf32>) -> tensor<*xf32> attributes {tf.e
// CHECK-SAME: input_to_input_intermediate = tensor<!quant.uniform<i16<-32767:32767>:f32, 9.7659230323191015E-4>> // CHECK-SAME: input_to_input_intermediate = tensor<!quant.uniform<i16<-32767:32767>:f32, 9.7659230323191015E-4>>
// CHECK-SAME: input_to_output_intermediate = tensor<!quant.uniform<i16<-32767:32767>:f32, 3.0518509475997192E-5>> // CHECK-SAME: input_to_output_intermediate = tensor<!quant.uniform<i16<-32767:32767>:f32, 3.0518509475997192E-5>>
// CHECK: "tfl.quantize"(%[[lstm]]) {qtype = tensor<*x!quant.uniform<i8:f32, 0.015686274509803921:-1>>} // CHECK: "tfl.quantize"(%[[lstm]]) {qtype = tensor<*x!quant.uniform<i8:f32, 0.015686274509803921:-1>>, volatile}
} }
// CHECK-LABEL: QuantizeSVDF // CHECK-LABEL: QuantizeSVDF
@ -398,7 +398,7 @@ func @QuantizeSVDF(%arg0: tensor<1x3xf32>) -> tensor<1x2xf32> {
// CHECK-DAG: %[[input_3:.*]] = "tfl.dequantize"({{.*}}) : (tensor<2x!quant.uniform<i32:f32, 1.3900876031311922E-5>>) // CHECK-DAG: %[[input_3:.*]] = "tfl.dequantize"({{.*}}) : (tensor<2x!quant.uniform<i32:f32, 1.3900876031311922E-5>>)
// CHECK-DAG: %[[input_4:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x4x!quant.uniform<i16<-32767:32767>:f32, 0.0037514108011770368>>) // CHECK-DAG: %[[input_4:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x4x!quant.uniform<i16<-32767:32767>:f32, 0.0037514108011770368>>)
// CHECK: %[[svdf:.*]] = "tfl.svdf"(%[[input_0]], %[[input_1]], %[[input_2]], %[[input_3]], %[[input_4]]) // CHECK: %[[svdf:.*]] = "tfl.svdf"(%[[input_0]], %[[input_1]], %[[input_2]], %[[input_3]], %[[input_4]])
// CHECK: %[[q:.*]] = "tfl.quantize"(%[[svdf]]) {qtype = tensor<1x2x!quant.uniform<i8:f32, 0.12954867493872549:-128>>} // CHECK: %[[q:.*]] = "tfl.quantize"(%[[svdf]]) {qtype = tensor<1x2x!quant.uniform<i8:f32, 0.12954867493872549:-128>>, volatile}
// CHECK: %[[dq:.*]] = "tfl.dequantize"(%11) // CHECK: %[[dq:.*]] = "tfl.dequantize"(%11)
// CHECK: return %[[dq]] // CHECK: return %[[dq]]
} }

8
tensorflow/compiler/mlir/lite/tests/prepare-quantize-signed.mlir
View File

@ -49,9 +49,9 @@ func @prepareStatistics(%arg0: tensor<8x4x3xf32>) -> tensor<8x4x3xf32> {
} : (tensor<8x4x3xf32>) -> tensor<8x4x3xf32> } : (tensor<8x4x3xf32>) -> tensor<8x4x3xf32>
return %1 : tensor<8x4x3xf32> return %1 : tensor<8x4x3xf32>
// CHECK: %[[q1:.*]] = "tfl.quantize"(%arg0) {qtype = tensor<8x4x3x!quant.uniform<i8:f32, 0.0078431372549019607:-1>>} // CHECK: %[[q1:.*]] = "tfl.quantize"(%arg0) {qtype = tensor<8x4x3x!quant.uniform<i8:f32, 0.0078431372549019607:-1>>, volatile}
// CHECK: %[[dq1:.*]] = "tfl.dequantize"(%[[q1]]) // CHECK: %[[dq1:.*]] = "tfl.dequantize"(%[[q1]])
// CHECK: %[[q2:.*]] = "tfl.quantize"(%[[dq1]]) {qtype = tensor<8x4x3x!quant.uniform<i8:f32:2, {0.0078431372549019607:-1,0.062745098039215685:-1,0.0039215686274509803:-1}>>} // CHECK: %[[q2:.*]] = "tfl.quantize"(%[[dq1]]) {qtype = tensor<8x4x3x!quant.uniform<i8:f32:2, {0.0078431372549019607:-1,0.062745098039215685:-1,0.0039215686274509803:-1}>>, volatile}
// CHECK: %[[dq2:.*]] = "tfl.dequantize"(%[[q2]]) // CHECK: %[[dq2:.*]] = "tfl.dequantize"(%[[q2]])
// CHECK: return %[[dq2]] // CHECK: return %[[dq2]]
} }
@ -71,9 +71,9 @@ func @prepareStatisticsNudge(%arg0: tensor<8x4x3xf32>) -> tensor<8x4x3xf32> {
} : (tensor<8x4x3xf32>) -> tensor<8x4x3xf32> } : (tensor<8x4x3xf32>) -> tensor<8x4x3xf32>
return %1 : tensor<8x4x3xf32> return %1 : tensor<8x4x3xf32>
// CHECK: %[[q1:.*]] = "tfl.quantize"(%arg0) {qtype = tensor<8x4x3x!quant.uniform<i8:f32, 0.0039215686274509803:-128>>} // CHECK: %[[q1:.*]] = "tfl.quantize"(%arg0) {qtype = tensor<8x4x3x!quant.uniform<i8:f32, 0.0039215686274509803:-128>>, volatile}
// CHECK: %[[dq1:.*]] = "tfl.dequantize"(%[[q1]]) // CHECK: %[[dq1:.*]] = "tfl.dequantize"(%[[q1]])
// CHECK: %[[q2:.*]] = "tfl.quantize"(%[[dq1]]) {qtype = tensor<8x4x3x!quant.uniform<i8:f32:2, {0.0078431372549019607:-1,0.031372549019607843:127,0.0039215686274509803:-1}>>} // CHECK: %[[q2:.*]] = "tfl.quantize"(%[[dq1]]) {qtype = tensor<8x4x3x!quant.uniform<i8:f32:2, {0.0078431372549019607:-1,0.031372549019607843:127,0.0039215686274509803:-1}>>, volatile}
// CHECK: %[[dq2:.*]] = "tfl.dequantize"(%[[q2]]) // CHECK: %[[dq2:.*]] = "tfl.dequantize"(%[[q2]])
// CHECK: return %[[dq2]] // CHECK: return %[[dq2]]
} }

4
tensorflow/compiler/mlir/lite/tests/prepare-quantize.mlir
View File

@ -48,9 +48,9 @@ func @prepareStatistics(%arg0: tensor<8x4x3xf32>) -> tensor<8x4x3xf32> {
} : (tensor<8x4x3xf32>) -> tensor<8x4x3xf32> } : (tensor<8x4x3xf32>) -> tensor<8x4x3xf32>
return %1 : tensor<8x4x3xf32> return %1 : tensor<8x4x3xf32>
// CHECK: %[[q1:.*]] = "tfl.quantize"(%arg0) {qtype = tensor<8x4x3x!quant.uniform<u8:f32, 0.0078431372549019607:128>>} // CHECK: %[[q1:.*]] = "tfl.quantize"(%arg0) {qtype = tensor<8x4x3x!quant.uniform<u8:f32, 0.0078431372549019607:128>>, volatile}
// CHECK: %[[dq1:.*]] = "tfl.dequantize"(%[[q1]]) // CHECK: %[[dq1:.*]] = "tfl.dequantize"(%[[q1]])
// CHECK: %[[q2:.*]] = "tfl.quantize"(%[[dq1]]) {qtype = tensor<8x4x3x!quant.uniform<u8:f32:2, {0.0078431372549019607:128,0.062745098039215685:128,0.0039215686274509803:128}>>} // CHECK: %[[q2:.*]] = "tfl.quantize"(%[[dq1]]) {qtype = tensor<8x4x3x!quant.uniform<u8:f32:2, {0.0078431372549019607:128,0.062745098039215685:128,0.0039215686274509803:128}>>, volatile}
// CHECK: %[[dq2:.*]] = "tfl.dequantize"(%[[q2]]) // CHECK: %[[dq2:.*]] = "tfl.dequantize"(%[[q2]])
// CHECK: return %[[dq2]] // CHECK: return %[[dq2]]
} }

6
tensorflow/compiler/mlir/lite/transforms/post_quantize.cc
View File

@ -132,6 +132,12 @@ struct RemoveVolatileOps : public OpRewritePattern<DequantizeOp> {
if (auto q = llvm::dyn_cast_or_null<QuantizeOp>(input_op)) { if (auto q = llvm::dyn_cast_or_null<QuantizeOp>(input_op)) {
if (!q->getAttr(mlir::quant::kVolatileOpAttrName)) return failure(); if (!q->getAttr(mlir::quant::kVolatileOpAttrName)) return failure();
// Don't remove leading and tailing QDQ for PQT workflow, so the io
// modifying lib can work correctly.
if (!q.input().getDefiningOp()) return failure();
if (op->hasOneUse() && op->user_begin()->isKnownTerminator())
return failure();
op.replaceAllUsesWith(q.input()); op.replaceAllUsesWith(q.input());
return success(); return success();
} }