Add TENSOR_QUANT8_ASYMM_SIGNED support to ADD, MUL and SUB.
* Update legacy TF Lite functions calls to the current ones.
* Add QuantizeMultiplierSmallerThanOneExp function (same as TF Lite
function, does the same thing as QuantizeMultiplierSmallerThanOne but
returns left shift instead of right shift so that all the
QuantizeMultiplier* functions have the same interface.
Bug: 143934463
Bug: 143934627
Bug: 143935040
Test: NNTest_static and 1.3 VTS QuantizationCouplingTest
Change-Id: I897c07bc5d93dfa7c03a0c66eeb77fcbdec6aea0
Merged-In: I897c07bc5d93dfa7c03a0c66eeb77fcbdec6aea0
(cherry picked from commit c7b32bf9555a556cd1b9e5ffe9372a7f6d3fadb8)
diff --git a/common/OperationsUtils.cpp b/common/OperationsUtils.cpp
index 5c251a2..89ed5dc 100644
--- a/common/OperationsUtils.cpp
+++ b/common/OperationsUtils.cpp
@@ -27,6 +27,9 @@
#include "Operations.h"
#include "Utils.h"
+#include "Operations.h"
+#include "Utils.h"
+
namespace android {
namespace nn {
@@ -47,6 +50,32 @@
return true;
}
+void CalculateActivationRangeImpl(int32_t activation, const Shape& outputShape, int32_t qmin,
+ int32_t qmax, int32_t* act_min, int32_t* act_max) {
+ const auto scale = outputShape.scale;
+ const auto zero_point = outputShape.offset;
+
+ auto quantize = [scale, zero_point](float f) {
+ return zero_point + static_cast<int32_t>(std::round(f / scale));
+ };
+
+ if (activation == kActivationRelu) {
+ *act_min = std::max(qmin, quantize(0.0));
+ *act_max = qmax;
+ } else if (activation == kActivationRelu6) {
+ *act_min = std::max(qmin, quantize(0.0));
+ *act_max = std::min(qmax, quantize(6.0));
+ } else if (activation == kActivationRelu1) {
+ *act_min = std::max(qmin, quantize(-1.0));
+ *act_max = std::min(qmax, quantize(1.0));
+ } else if (activation == kActivationNone) {
+ *act_min = qmin;
+ *act_max = qmax;
+ } else {
+ LOG(ERROR) << "Unsupported fused activation function.";
+ }
+}
+
} // namespace
bool validateInputTypes(const IOperationValidationContext* context,
@@ -171,7 +200,7 @@
return true;
}
-bool QuantizeMultiplier(double double_multiplier, int32_t* quantized_multiplier, int* shift) {
+bool QuantizeMultiplier(double double_multiplier, int32_t* quantized_multiplier, int32_t* shift) {
if (double_multiplier == 0.) {
*quantized_multiplier = 0;
*shift = 0;
@@ -189,6 +218,15 @@
return true;
}
+bool QuantizeMultiplierSmallerThanOneExp(double double_multiplier, int32_t* quantized_multiplier,
+ int32_t* left_shift) {
+ NN_RET_CHECK(double_multiplier > 0.);
+ NN_RET_CHECK(double_multiplier < 1.);
+ NN_RET_CHECK(QuantizeMultiplier(double_multiplier, quantized_multiplier, left_shift));
+ NN_RET_CHECK(*left_shift <= 0);
+ return true;
+}
+
bool QuantizeMultiplierSmallerThanOne(double double_multiplier, int32_t* quantized_multiplier,
int32_t* right_shift) {
NN_OPS_CHECK(double_multiplier >= 0.);
@@ -249,28 +287,15 @@
const int32_t qmin = std::numeric_limits<uint8_t>::min();
const int32_t qmax = std::numeric_limits<uint8_t>::max();
- const auto scale = outputShape.scale;
- const auto zero_point = outputShape.offset;
+ CalculateActivationRangeImpl(activation, outputShape, qmin, qmax, act_min, act_max);
+}
- auto quantize = [scale, zero_point](float f) {
- return zero_point + static_cast<int32_t>(std::round(f / scale));
- };
+void CalculateActivationRangeInt8(int32_t activation, const Shape& outputShape, int32_t* act_min,
+ int32_t* act_max) {
+ const int32_t qmin = std::numeric_limits<int8_t>::min();
+ const int32_t qmax = std::numeric_limits<int8_t>::max();
- if (activation == kActivationRelu) {
- *act_min = std::max(qmin, quantize(0.0));
- *act_max = qmax;
- } else if (activation == kActivationRelu6) {
- *act_min = std::max(qmin, quantize(0.0));
- *act_max = std::min(qmax, quantize(6.0));
- } else if (activation == kActivationRelu1) {
- *act_min = std::max(qmin, quantize(-1.0));
- *act_max = std::min(qmax, quantize(1.0));
- } else if (activation == kActivationNone) {
- *act_min = qmin;
- *act_max = qmax;
- } else {
- LOG(ERROR) << "Unsupported fused activation function.";
- }
+ CalculateActivationRangeImpl(activation, outputShape, qmin, qmax, act_min, act_max);
}
void CalculateActivationRangeFloat(int32_t activation, float* activation_min,
