--- /dev/null
+#ifndef __NN_H__
+#define __NN_H__
+
+#include <vector>
+
+#include <cstdint>
+
+namespace vector
+{
+
+template <typename T> struct View
+{
+ virtual ~View() = default;
+
+ virtual int32_t size(void) const = 0;
+ virtual T at(uint32_t off) const = 0;
+};
+
+}
+
+namespace feature
+{
+
+struct Shape
+{
+ int32_t C;
+ int32_t H;
+ int32_t W;
+
+ Shape() = default;
+ Shape(int32_t depth, int32_t height, int32_t width)
+ : C{depth}, H{height}, W{width}
+ {
+ // DO NOTHING
+ }
+};
+
+template <typename T> struct View
+{
+ virtual ~View() = default;
+
+ virtual const Shape &shape(void) const = 0;
+ virtual T at(uint32_t ch, uint32_t row, uint32_t col) const = 0;
+};
+
+}
+
+namespace kernel
+{
+
+struct Shape
+{
+ int32_t N;
+ int32_t C;
+ int32_t H;
+ int32_t W;
+
+ Shape() = default;
+ Shape(int32_t count, int32_t depth, int32_t height, int32_t width)
+ : N{count}, C{depth}, H{height}, W{width}
+ {
+ // DO NOTHING
+ }
+};
+
+template <typename T> struct View
+{
+ virtual ~View() = default;
+
+ virtual const Shape &shape(void) const = 0;
+ virtual T at(uint32_t nth, uint32_t ch, uint32_t row, uint32_t col) const = 0;
+};
+
+}
+
+class SampleBiasObject final : public vector::View<float>
+{
+public:
+ SampleBiasObject(int32_t size);
+
+public:
+ int32_t size(void) const override;
+
+public:
+ float at(uint32_t off) const override;
+
+private:
+ int32_t _size;
+};
+
+class SampleFeatureObject final : public feature::View<float>
+{
+public:
+ SampleFeatureObject(const feature::Shape &shape);
+
+public:
+ const feature::Shape &shape(void) const override;
+
+ float at(uint32_t ch, uint32_t row, uint32_t col) const override;
+
+public:
+ float &at(uint32_t ch, uint32_t row, uint32_t col);
+
+private:
+ feature::Shape _shape;
+ std::vector<float> _value;
+};
+
+class SampleKernelObject final : public kernel::View<float>
+{
+public:
+ SampleKernelObject(const kernel::Shape &shape);
+
+public:
+ const kernel::Shape &shape(void) const override;
+
+ float at(uint32_t nth, uint32_t ch, uint32_t row, uint32_t col) const override;
+
+private:
+ kernel::Shape _shape;
+ std::vector<float> _value;
+};
+
+#endif
--- /dev/null
+#include <nn.h>
+
+#include <cassert>
+
+//
+// SampleBiasObject
+//
+SampleBiasObject::SampleBiasObject(int32_t size) : _size(size)
+{
+ // DO NOTHING
+}
+
+int32_t SampleBiasObject::size(void) const { return _size; }
+
+float SampleBiasObject::at(uint32_t off) const { return 0.0f; }
+
+//
+// SampleFeatureObject
+//
+SampleFeatureObject::SampleFeatureObject(const feature::Shape &shape) : _shape{shape}
+{
+ const uint32_t size = _shape.C * _shape.H * _shape.W;
+
+ for (uint32_t off = 0; off < size; ++off)
+ {
+ _value.emplace_back(off);
+ }
+
+ assert(_value.size() == size);
+}
+
+const feature::Shape &SampleFeatureObject::shape(void) const
+{
+ return _shape;
+}
+
+float SampleFeatureObject::at(uint32_t ch, uint32_t row, uint32_t col) const
+{
+ return _value.at(ch * _shape.H * _shape.W + row * _shape.W + col);
+}
+
+float &SampleFeatureObject::at(uint32_t ch, uint32_t row, uint32_t col)
+{
+ return _value.at(ch * _shape.H * _shape.W + row * _shape.W + col);
+}
+
+//
+// SampleKernelObject
+//
+SampleKernelObject::SampleKernelObject(const kernel::Shape &shape) : _shape{shape}
+{
+ const uint32_t size = _shape.N * _shape.C * _shape.H * _shape.W;
+
+ for (uint32_t off = 0; off < size; ++off)
+ {
+ _value.emplace_back(static_cast<float>(off));
+ }
+
+ assert(_value.size() == size);
+}
+
+const kernel::Shape &SampleKernelObject::shape(void) const { return _shape; };
+
+float SampleKernelObject::at(uint32_t nth, uint32_t ch, uint32_t row, uint32_t col) const
+{
+ return _value.at(nth * _shape.C * _shape.H * _shape.W + ch * _shape.H * _shape.W + row * _shape.W + col);
+}
--- /dev/null
+#include "tensorflow/contrib/lite/kernels/register.h"
+#include "tensorflow/contrib/lite/model.h"
+#include "tensorflow/contrib/lite/builtin_op_data.h"
+
+#include "nn.h"
+#include "env.h"
+
+#include <iostream>
+
+using namespace tflite;
+using namespace tflite::ops::builtin;
+
+int main(int argc, char **argv)
+{
+#define INT_VALUE(NAME, VALUE) IntVar NAME##_Value(#NAME, VALUE);
+#include "conv_1.lst"
+#undef INT_VALUE
+
+ const int32_t IFM_C = IFM_C_Value();
+ const int32_t IFM_H = IFM_H_Value();
+ const int32_t IFM_W = IFM_W_Value();
+
+ const int32_t KER_N = KER_N_Value();
+ const int32_t KER_C = IFM_C_Value();
+ const int32_t KER_H = KER_H_Value();
+ const int32_t KER_W = KER_W_Value();
+
+ const int32_t OFM_C = KER_N;
+ const int32_t OFM_H = (IFM_H - KER_H) + 1;
+ const int32_t OFM_W = (IFM_W - KER_W) + 1;
+
+ const SampleFeatureObject ifm{feature::Shape{IFM_C, IFM_H, IFM_W}};
+ const SampleKernelObject kernel{kernel::Shape{KER_N, KER_C, KER_H, KER_W}};
+ const SampleBiasObject bias{KER_N};
+
+ // Configure Kernel Data
+ const uint32_t kernel_size = KER_N * KER_C * KER_H * KER_W;
+ float kernel_data[kernel_size] = { 0.0f, };
+
+ // Fill kernel data in NHWC order
+ {
+ uint32_t off = 0;
+
+ for (uint32_t nth = 0; nth < KER_N; ++nth)
+ {
+ for (uint32_t row = 0; row < KER_H; ++row)
+ {
+ for (uint32_t col = 0; col < KER_W; ++col)
+ {
+ for (uint32_t ch = 0; ch < KER_C; ++ch)
+ {
+ const auto value = kernel.at(nth, ch, row, col);
+ kernel_data[off++] = value;
+ }
+ }
+ }
+ }
+
+ assert(kernel_size == off);
+ }
+
+ // Assumption on this example
+ assert(IFM_C == KER_C);
+ assert(KER_N == bias.size());
+
+ auto setup = [&](Interpreter &interp)
+ {
+ // Comment from 'context.h'
+ //
+ // Parameters for asymmetric quantization. Quantized values can be converted
+ // back to float using:
+ // real_value = scale * (quantized_value - zero_point);
+ //
+ // Q: Is this necessary?
+ TfLiteQuantizationParams quantization;
+
+ quantization.scale = 1;
+ quantization.zero_point = 0;
+
+ // On AddTensors(N) call, T/F Lite interpreter creates N tensors whose index is [0 ~ N)
+ interp.AddTensors(5);
+
+ // Configure OFM
+ interp.SetTensorParametersReadWrite(0,
+ kTfLiteFloat32 /* type */,
+ "output" /* name */,
+ {1 /*N*/, OFM_H, OFM_W, OFM_C} /* dims */,
+ quantization);
+
+
+ // Configure IFM
+ interp.SetTensorParametersReadWrite(1,
+ kTfLiteFloat32 /* type */,
+ "input" /* name */,
+ {1 /*N*/, IFM_H, IFM_W, IFM_C} /* dims */,
+ quantization);
+
+ // NOTE kernel_data should live longer than interpreter!
+ interp.SetTensorParametersReadOnly(2,
+ kTfLiteFloat32 /* type */,
+ "filter" /* name */,
+ {KER_N, KER_H, KER_W, KER_C} /* dims */,
+ quantization,
+ reinterpret_cast<const char *>(kernel_data), kernel_size * sizeof(float));
+
+ // Configure Bias
+ const uint32_t bias_size = bias.size();
+ float bias_data[bias_size] = { 0.0f, };
+
+ // Fill bias data
+ for (uint32_t off = 0; off < bias.size(); ++off)
+ {
+ bias_data[off] = bias.at(off);
+ }
+
+ interp.SetTensorParametersReadOnly(3,
+ kTfLiteFloat32 /* type */,
+ "bias" /* name */,
+ { bias.size() } /* dims */,
+ quantization,
+ reinterpret_cast<const char *>(bias_data), sizeof(bias_data));
+
+ // Add Convolution Node
+ //
+ // NOTE AddNodeWithParameters take the ownership of param, and deallocate it with free
+ // So, param should be allocated with malloc
+ TfLiteConvParams *param = reinterpret_cast<TfLiteConvParams *>(malloc(sizeof(TfLiteConvParams)));
+
+ param->padding = kTfLitePaddingValid;
+ param->stride_width = 1;
+ param->stride_height = 1;
+ param->activation = kTfLiteActRelu;
+
+ // Run Convolution and store its result into Tensor #0
+ // - Read IFM from Tensor #1
+ // - Read Filter from Tensor #2,
+ // - Read Bias from Tensor #3
+ interp.AddNodeWithParameters({1, 2, 3}, {0},
+ nullptr, 0,
+ reinterpret_cast<void *>(param),
+ BuiltinOpResolver().FindOp(BuiltinOperator_CONV_2D));
+
+ // Set Tensor #1 as Input #0, and Tensor #0 as Output #0
+ interp.SetInputs({1});
+ interp.SetOutputs({0});
+
+ // Allocate Tensor
+ interp.AllocateTensors();
+
+ // Fill IFM data in HWC order
+ {
+ uint32_t off = 0;
+
+ for (uint32_t row = 0; row < ifm.shape().H; ++row)
+ {
+ for (uint32_t col = 0; col < ifm.shape().W; ++col)
+ {
+ for (uint32_t ch = 0; ch < ifm.shape().C; ++ch)
+ {
+ const auto value = ifm.at(ch, row, col);
+ interp.typed_input_tensor<float>(0)[off++] = value;
+ }
+ }
+ }
+ }
+ };
+
+ Interpreter pure;
+ {
+ pure.UseNNAPI(false);
+
+ setup(pure);
+
+ // Let's Rock-n-Roll!
+ pure.Invoke();
+ }
+
+ Interpreter delegated;
+ {
+ // Let's use NNAPI (if possible)
+ delegated.UseNNAPI(true);
+
+ setup(delegated);
+
+ delegated.Invoke();
+ }
+
+ // Compare OFM
+ {
+ uint32_t off = 0;
+
+ for (uint32_t row = 0; row < OFM_H; ++row)
+ {
+ for (uint32_t col = 0; col < OFM_W; ++col)
+ {
+ for (uint32_t ch = 0; ch < kernel.shape().N; ++ch)
+ {
+ const auto value_from_interp = pure.typed_output_tensor<float>(0)[off];
+ const auto value_from_NNAPI = delegated.typed_output_tensor<float>(0)[off];
+
+ if (value_from_interp != value_from_NNAPI)
+ {
+ std::cerr << "Diff at (ch: " << ch << ", row: " << row << ", col: " << col << ")" << std::endl;
+ std::cerr << " Value from interpreter: " << value_from_interp << std::endl;
+ std::cerr << " Value from NNAPI: " << value_from_NNAPI << std::endl;
+ }
+
+ off += 1;
+ }
+ }
+ }
+ }
+
+ return 0;
+}
projects / platform / core / ml / nnfw.git / commitdiff