commit 62cf7fdcc03da02682d47f731094c2032b4fc14a
author Michael Butler <butlermichael@google.com> Thu Dec 21 18:51:27 2017 -0800
committer Michael Butler <butlermichael@google.com> Thu Dec 21 18:54:42 2017 -0800
tree e84efdd10d1b2d1e814a9ce26d825c3c5962a2c3
parent dd779b6e6891811fb7dc317e72c50ff97e672f8c

Adjust Hexagon NN library's output quantized range

The Hexagon NN library chooses its own quantized ranges. To restore the
quantized range provided by the NN API, new operations are added to the
end of a hexagon nn graph to dequantize to float and quantize to set the
NN API's range.

Bug: 67478959
Test: mm
Test: runs on cts and vts binaries
Change-Id: I9cabf4c7c92335aaa0f04bd5e8d2133a6a53f001

1.0/HexagonModel.cpp

diff --git a/1.0/HexagonModel.cpp b/1.0/HexagonModel.cpp
index e165057..136818d 100644
--- a/1.0/HexagonModel.cpp
+++ b/1.0/HexagonModel.cpp
@@ -191,7 +191,7 @@
     return hexagon::getPadding(padding);
 }
 
-hexagon_nn_input Model::createQuantizationValue(uint32_t operand, uint32_t quant_value) {
+hexagon_nn_input Model::createQuantizationValue(uint32_t operand, int32_t quant_value) {
     OperandInfo& operandInfo = mOperands[operand];
     float real_value = (quant_value - operandInfo.zeroPoint) * operandInfo.scale;
     return createValues<float>({real_value});
@@ -526,12 +526,30 @@
 
 bool Model::addOutputs() {
     // prepare OP_OUTPUT's inputs
-    std::vector<hexagon_nn_input> ins(mOutputs.size());
-    for (size_t i = 0; i < mOutputs.size(); ++i) {
-        OperandInfo& operand = mOperands[mOutputs[i]];
+    std::vector<hexagon_nn_input> ins;
+    for (size_t out : mOutputs) {
+        OperandInfo& operand = mOperands[out];
         HEXAGON_SOFT_ASSERT_NE(operand.hexagon_input, hexagon_nn_input{},
                                "output operand has not been registered");
-        ins[i] = operand.hexagon_input;
+
+        if (operand.type == OperandType::TENSOR_QUANT8_ASYMM) {
+            // Adjust quantized range of outputs
+            uint32_t dequant = addOperationInternal(
+                OP_Dequantize, NN_PAD_NA,
+                {operand.hexagon_input, operand.hexagon_input_min, operand.hexagon_input_max},
+                {make_hexagon_nn_output(operand.dimensions, sizeof(float))});
+            uint32_t quant =
+                addOperationInternal(OP_Quantize, NN_PAD_NA,
+                                     {{.src_id = dequant, .output_idx = 0},
+                                      createQuantizationValue(out, 0),
+                                      createQuantizationValue(out, 255)},
+                                     {make_hexagon_nn_output(operand.dimensions, sizeof(uint8_t)),
+                                      make_hexagon_nn_output({1, 1, 1, 1}, sizeof(float)),
+                                      make_hexagon_nn_output({1, 1, 1, 1}, sizeof(float))});
+            ins.push_back({.src_id = quant, .output_idx = 0});
+        } else {
+            ins.push_back(operand.hexagon_input);
+        }
     }
 
     // add single output node for entire graph

1.0/HexagonModel.h

diff --git a/1.0/HexagonModel.h b/1.0/HexagonModel.h
index 4661ad9..797620b 100644
--- a/1.0/HexagonModel.h
+++ b/1.0/HexagonModel.h
@@ -90,7 +90,7 @@
     const hexagon_nn_input& getTensor(uint32_t operand);
     const hexagon_nn_input& getQuantizationMin(uint32_t operand);
     const hexagon_nn_input& getQuantizationMax(uint32_t operand);
-    hexagon_nn_input createQuantizationValue(uint32_t operand, uint32_t quant_value);
+    hexagon_nn_input createQuantizationValue(uint32_t operand, int32_t quant_value);
     hexagon_nn_input createConvFilterTensor(uint32_t operand);
     hexagon_nn_input createDepthwiseFilterTensor(uint32_t operand, int32_t depth_multiplier);
     hexagon_nn_input createFullyConnectedWeightTensor(uint32_t operand);