!4349 [MS][LITE][Develop]compare ops support quant

Merge pull request !4349 from chenjianping/lite_dev2

mindspore/lite/src/runtime/kernel/arm/int8/arithmetic_int8.cc

@@ -15,7 +15,6 @@
  */
 
 #include "src/runtime/kernel/arm/int8/arithmetic_int8.h"
-#include "src/runtime/kernel/arm/nnacl/int8/arithmetic_int8.h"
 #include "src/runtime/kernel/arm/nnacl/arithmetic_common.h"
 #include "schema/model_generated.h"
 #include "src/kernel_registry.h"
@@ -42,7 +41,7 @@ int ArithmeticsInt8Launch(int thread_id, LiteParallelGroupEnv *penv, void *cdata
   auto error_code = arithmetic_kernel->DoArithmetic(thread_id);
   if (error_code != RET_OK) {
     MS_LOG(ERROR) << "ArithmeticsRun error thread_id[" << thread_id << "] error_code[" << error_code << "]";
-    return RET_ERROR;
+    return error_code;
   }
   return RET_OK;
 }
@@ -79,28 +78,43 @@ ArithmeticInt8CPUKernel::~ArithmeticInt8CPUKernel() {
 int ArithmeticInt8CPUKernel::Init() {
   switch (op_parameter_->type_) {
     case PrimitiveType_Equal:
-      arithmetic_run_ = ElementEqual;
+      arithmetic_run_ = ElementEqualInt8;
       break;
     case PrimitiveType_NotEqual:
-      arithmetic_run_ = ElementNotEqual;
+      arithmetic_run_ = ElementNotEqualInt8;
       break;
     case PrimitiveType_Less:
-      arithmetic_run_ = ElementLess;
+      arithmetic_run_ = ElementLessInt8;
       break;
     case PrimitiveType_LessEqual:
-      arithmetic_run_ = ElementLessEqual;
+      arithmetic_run_ = ElementLessEqualInt8;
       break;
     case PrimitiveType_Greater:
-      arithmetic_run_ = ElementGreater;
+      arithmetic_run_ = ElementGreaterInt8;
       break;
     case PrimitiveType_GreaterEqual:
-      arithmetic_run_ = ElementGreaterEqual;
+      arithmetic_run_ = ElementGreaterEqualInt8;
       break;
     default:
       MS_LOG(ERROR) << "Error Operator type " << op_parameter_->type_;
       arithmetic_run_ = nullptr;
       return RET_PARAM_INVALID;
   }
+
+  auto *input0_tensor = in_tensors_.at(0);
+  auto in0_quant_args = input0_tensor->GetQuantParams();
+  quant_args_.in0_args_.scale_ = in0_quant_args.front().scale;
+  quant_args_.in0_args_.zp_ = in0_quant_args.front().zeroPoint;
+
+  auto *input1_tensor = in_tensors_.at(1);
+  auto in1_quant_args = input1_tensor->GetQuantParams();
+  quant_args_.in1_args_.scale_ = in1_quant_args.front().scale;
+  quant_args_.in1_args_.zp_ = in1_quant_args.front().zeroPoint;
+
+  auto *out_tensor = out_tensors_.at(kOutputIndex);
+  auto out_quant_args = out_tensor->GetQuantParams();
+  quant_args_.out_args_.scale_ = out_quant_args.front().scale;
+  quant_args_.out_args_.zp_ = out_quant_args.front().zeroPoint;
   if (!InferShapeDone()) {
     return RET_OK;
   }
@@ -142,16 +156,16 @@ int ArithmeticInt8CPUKernel::DoArithmetic(int thread_id) {
     }
 
     int error_code = arithmetic_run_(tile_data0_ + stride * thread_id, tile_data1_ + stride * thread_id,
-                                     output_data + stride * thread_id, count);
+                                     output_data + stride * thread_id, count, &quant_args_);
     if (error_code != RET_OK) {
       MS_LOG(ERROR) << "Arithmetic run fail! ret: " << error_code;
-      return RET_ERROR;
+      return error_code;
     }
   } else if (arithmetic_run_ != nullptr) {
-    int error_code = arithmetic_run_(input0_data, input1_data1, output_data, element_num);
+    int error_code = arithmetic_run_(input0_data, input1_data1, output_data, element_num, &quant_args_);
     if (error_code != RET_OK) {
       MS_LOG(ERROR) << "Arithmetic run fail!ret: " << error_code;
-      return RET_ERROR;
+      return error_code;
     }
   } else {
     MS_LOG(ERROR) << "arithmetic_run function is nullptr!";

mindspore/lite/src/runtime/kernel/arm/int8/arithmetic_int8.h

@@ -20,10 +20,12 @@
 #include <vector>
 #include "src/lite_kernel.h"
 #include "schema/model_generated.h"
+#include "src/runtime/kernel/arm/nnacl/int8/arithmetic_int8.h"
 
 namespace mindspore::kernel {
 class ArithmeticInt8CPUKernel : public LiteKernel {
-  typedef int (*ArithmeticRunInt8)(int8_t *input0, int8_t *input1, int8_t *output, int element_size);
+  typedef int (*ArithmeticRunInt8)(int8_t *input0, int8_t *input1, int8_t *output, int element_size,
+                                   ArithmeticQuantArg *quant_arg);
 
  public:
   ArithmeticInt8CPUKernel(OpParameter *parameter, const std::vector<lite::tensor::Tensor *> &inputs,
@@ -39,10 +41,10 @@ class ArithmeticInt8CPUKernel : public LiteKernel {
 
  private:
   void FreeTileData();
-  int thread_count_;
   int8_t *tile_data0_;
   int8_t *tile_data1_;
   ArithmeticRunInt8 arithmetic_run_;
+  ArithmeticQuantArg quant_args_;
 };
 }  // namespace mindspore::kernel
 #endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_INT8_ARITHMETIC_INT8_H_

mindspore/lite/src/runtime/kernel/arm/nnacl/fp32/arithmetic.cc

@@ -17,6 +17,8 @@
 #include "nnacl/fp32/arithmetic.h"
 #include <math.h>
 
+#define ACCURACY_DATA 0.00000001
+
 int ElementMul(float *input0, float *input1, float *output, int element_size) {
   int block_mod = element_size % C4NUM;
   int block_c4 = element_size - block_mod;
@@ -549,6 +551,14 @@ int BroadcastMinimum(float *input0, float *input1, float *tile_input0, float *ti
   return ElementMinimum(tile_input0, tile_input1, output, element_size);
 }
 
+float FloatNotEqualCheck(float in0, float in1) {
+  float minus = in0 - in1;
+  if (minus <= ACCURACY_DATA && minus >= -ACCURACY_DATA) {
+    return (float)false;
+  }
+  return (float)true;
+}
+
 int ElementNotEqual(float *input0, float *input1, float *output, int element_size) {
   int block_mod = element_size % C4NUM;
   int block_c4 = element_size - block_mod;
@@ -563,10 +573,10 @@ int ElementNotEqual(float *input0, float *input1, float *output, int element_siz
     float32x4_t vout = vbslq_f32(vceqq_f32(vin0, vin1), vfalse, vtrue);
     vst1q_f32(output, vout);
 #else
-    output[0] = (float)(input0[0] != input1[0]);
-    output[1] = (float)(input0[1] != input1[1]);
-    output[2] = (float)(input0[2] != input1[2]);
-    output[3] = (float)(input0[3] != input1[3]);
+    output[0] = FloatNotEqualCheck(input0[0], input1[0]);
+    output[1] = FloatNotEqualCheck(input0[1], input1[1]);
+    output[2] = FloatNotEqualCheck(input0[2], input1[2]);
+    output[3] = FloatNotEqualCheck(input0[3], input1[3]);
 #endif
     input0 += C4NUM;
     input1 += C4NUM;
@@ -584,6 +594,14 @@ int BroadcastNotEqual(float *input0, float *input1, float *tile_input0, float *t
   return ElementNotEqual(tile_input0, tile_input1, output, element_size);
 }
 
+float FloatEqualCheck(float in0, float in1) {
+  float minus = in0 - in1;
+  if (minus <= ACCURACY_DATA && minus >= -ACCURACY_DATA) {
+    return (float)true;
+  }
+  return (float)false;
+}
+
 int ElementEqual(float *input0, float *input1, float *output, int element_size) {
   int block_mod = element_size % C4NUM;
   int block_c4 = element_size - block_mod;
@@ -598,10 +616,10 @@ int ElementEqual(float *input0, float *input1, float *output, int element_size)
     float32x4_t vout = vbslq_f32(vceqq_f32(vin0, vin1), vtrue, vfalse);
     vst1q_f32(output, vout);
 #else
-    output[0] = (float)(input0[0] == input1[0]);
-    output[1] = (float)(input0[1] == input1[1]);
-    output[2] = (float)(input0[2] == input1[2]);
-    output[3] = (float)(input0[3] == input1[3]);
+    output[0] = FloatEqualCheck(input0[0], input1[0]);
+    output[1] = FloatEqualCheck(input0[1], input1[1]);
+    output[2] = FloatEqualCheck(input0[2], input1[2]);
+    output[3] = FloatEqualCheck(input0[3], input1[3]);
 #endif
     input0 += C4NUM;
     input1 += C4NUM;
@@ -758,3 +776,5 @@ int BroadcastGreaterEqual(float *input0, float *input1, float *tile_input0, floa
   TileDimensions(input0, input1, tile_input0, tile_input1, param);
   return ElementGreaterEqual(tile_input0, tile_input1, output, element_size);
 }
+
+#undef ACCURACY_DATA

mindspore/lite/src/runtime/kernel/arm/nnacl/int8/arithmetic_int8.cc

@@ -20,44 +20,102 @@
 #endif
 #include "nnacl/errorcode.h"
 
-int ElementNotEqual(int8_t *input0, int8_t *input1, int8_t *output, int element_size) {
+#define ACCURACY_DATA 0.00000001
+
+int ElementNotEqualInt8(int8_t *input0, int8_t *input1, int8_t *output, int element_size,
+                        ArithmeticQuantArg *quant_arg) {
+  float in0_bias = -quant_arg->in0_args_.zp_ * quant_arg->in0_args_.scale_;
+  float in1_bias = -quant_arg->in1_args_.zp_ * quant_arg->in1_args_.scale_;
+  float output_inverse_scale = 1.f / quant_arg->out_args_.scale_;
+  float out_zp = quant_arg->out_args_.zp_;
   for (int index = 0; index < element_size; ++index) {
-    output[index] = (int8_t)(input0[index] != input1[index]);
+    float in0_real = input0[index] * quant_arg->in0_args_.scale_ + in0_bias;
+    float in1_real = input1[index] * quant_arg->in1_args_.scale_ + in1_bias;
+    float minus_inputs = in0_real - in1_real;
+    float out_real = (float)true;
+    if (minus_inputs >= -ACCURACY_DATA && minus_inputs <= ACCURACY_DATA) {
+      out_real = (float)false;
+    }
+    output[index] = (int8_t)(out_real * output_inverse_scale + out_zp);
   }
   return NNACL_OK;
 }
 
-int ElementEqual(int8_t *input0, int8_t *input1, int8_t *output, int element_size) {
+int ElementEqualInt8(int8_t *input0, int8_t *input1, int8_t *output, int element_size, ArithmeticQuantArg *quant_arg) {
+  float in0_bias = -quant_arg->in0_args_.zp_ * quant_arg->in0_args_.scale_;
+  float in1_bias = -quant_arg->in1_args_.zp_ * quant_arg->in1_args_.scale_;
+  float output_inverse_scale = 1.f / quant_arg->out_args_.scale_;
+  float out_zp = quant_arg->out_args_.zp_;
   for (int index = 0; index < element_size; ++index) {
-    output[index] = (int8_t)(input0[index] == input1[index]);
+    float in0_real = input0[index] * quant_arg->in0_args_.scale_ + in0_bias;
+    float in1_real = input1[index] * quant_arg->in1_args_.scale_ + in1_bias;
+    float minus_inputs = in0_real - in1_real;
+    float out_real = (float)false;
+    if (minus_inputs >= -ACCURACY_DATA && minus_inputs <= ACCURACY_DATA) {
+      out_real = (float)true;
+    }
+    output[index] = (int8_t)(out_real * output_inverse_scale + out_zp);
   }
   return NNACL_OK;
 }
 
-int ElementLess(int8_t *input0, int8_t *input1, int8_t *output, int element_size) {
+int ElementLessInt8(int8_t *input0, int8_t *input1, int8_t *output, int element_size, ArithmeticQuantArg *quant_arg) {
+  float in0_bias = -quant_arg->in0_args_.zp_ * quant_arg->in0_args_.scale_;
+  float in1_bias = -quant_arg->in1_args_.zp_ * quant_arg->in1_args_.scale_;
+  float output_inverse_scale = 1.f / quant_arg->out_args_.scale_;
+  float out_zp = quant_arg->out_args_.zp_;
   for (int index = 0; index < element_size; ++index) {
-    output[index] = (int8_t)(input0[index] < input1[index]);
+    float in0_real = input0[index] * quant_arg->in0_args_.scale_ + in0_bias;
+    float in1_real = input1[index] * quant_arg->in1_args_.scale_ + in1_bias;
+    float out_real = (float)(in0_real < in1_real);
+    output[index] = (int8_t)(out_real * output_inverse_scale + out_zp);
   }
   return NNACL_OK;
 }
 
-int ElementLessEqual(int8_t *input0, int8_t *input1, int8_t *output, int element_size) {
+int ElementLessEqualInt8(int8_t *input0, int8_t *input1, int8_t *output, int element_size,
+                         ArithmeticQuantArg *quant_arg) {
+  float in0_bias = -quant_arg->in0_args_.zp_ * quant_arg->in0_args_.scale_;
+  float in1_bias = -quant_arg->in1_args_.zp_ * quant_arg->in1_args_.scale_;
+  float output_inverse_scale = 1.f / quant_arg->out_args_.scale_;
+  float out_zp = quant_arg->out_args_.zp_;
   for (int index = 0; index < element_size; ++index) {
-    output[index] = (int8_t)(input0[index] <= input1[index]);
+    float in0_real = input0[index] * quant_arg->in0_args_.scale_ + in0_bias;
+    float in1_real = input1[index] * quant_arg->in1_args_.scale_ + in1_bias;
+    float out_real = (float)(in0_real <= in1_real);
+    output[index] = (int8_t)(out_real * output_inverse_scale + out_zp);
   }
   return NNACL_OK;
 }
 
-int ElementGreater(int8_t *input0, int8_t *input1, int8_t *output, int element_size) {
+int ElementGreaterInt8(int8_t *input0, int8_t *input1, int8_t *output, int element_size,
+                       ArithmeticQuantArg *quant_arg) {
+  float in0_bias = -quant_arg->in0_args_.zp_ * quant_arg->in0_args_.scale_;
+  float in1_bias = -quant_arg->in1_args_.zp_ * quant_arg->in1_args_.scale_;
+  float output_inverse_scale = 1.f / quant_arg->out_args_.scale_;
+  float out_zp = quant_arg->out_args_.zp_;
   for (int index = 0; index < element_size; ++index) {
-    output[index] = (int8_t)(input0[index] > input1[index]);
+    float in0_real = input0[index] * quant_arg->in0_args_.scale_ + in0_bias;
+    float in1_real = input1[index] * quant_arg->in1_args_.scale_ + in1_bias;
+    float out_real = (float)(in0_real > in1_real);
+    output[index] = (int8_t)(out_real * output_inverse_scale + out_zp);
   }
   return NNACL_OK;
 }
 
-int ElementGreaterEqual(int8_t *input0, int8_t *input1, int8_t *output, int element_size) {
+int ElementGreaterEqualInt8(int8_t *input0, int8_t *input1, int8_t *output, int element_size,
+                            ArithmeticQuantArg *quant_arg) {
+  float in0_bias = -quant_arg->in0_args_.zp_ * quant_arg->in0_args_.scale_;
+  float in1_bias = -quant_arg->in1_args_.zp_ * quant_arg->in1_args_.scale_;
+  float output_inverse_scale = 1.f / quant_arg->out_args_.scale_;
+  float out_zp = quant_arg->out_args_.zp_;
   for (int index = 0; index < element_size; ++index) {
-    output[index] = (int8_t)(input0[index] >= input1[index]);
+    float in0_real = input0[index] * quant_arg->in0_args_.scale_ + in0_bias;
+    float in1_real = input1[index] * quant_arg->in1_args_.scale_ + in1_bias;
+    float out_real = (float)(in0_real >= in1_real);
+    output[index] = (int8_t)(out_real * output_inverse_scale + out_zp);
   }
   return NNACL_OK;
 }
+
+#undef ACCURACY_DATA

mindspore/lite/src/runtime/kernel/arm/nnacl/int8/arithmetic_int8.h

@@ -17,16 +17,21 @@
 #define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_NNACL_INT8_ARITHMETIC_INT8_H_
 
 #include "nnacl/op_base.h"
+#include "nnacl/quantization/quantize.h"
 
-int ElementNotEqual(int8_t *input0, int8_t *input1, int8_t *output, int element_size);
+int ElementNotEqualInt8(int8_t *input0, int8_t *input1, int8_t *output, int element_size,
+                        ArithmeticQuantArg *quant_arg);
 
-int ElementEqual(int8_t *input0, int8_t *input1, int8_t *output, int element_size);
+int ElementEqualInt8(int8_t *input0, int8_t *input1, int8_t *output, int element_size, ArithmeticQuantArg *quant_arg);
 
-int ElementLess(int8_t *input0, int8_t *input1, int8_t *output, int element_size);
+int ElementLessInt8(int8_t *input0, int8_t *input1, int8_t *output, int element_size, ArithmeticQuantArg *quant_arg);
 
-int ElementLessEqual(int8_t *input0, int8_t *input1, int8_t *output, int element_size);
+int ElementLessEqualInt8(int8_t *input0, int8_t *input1, int8_t *output, int element_size,
+                         ArithmeticQuantArg *quant_arg);
 
-int ElementGreater(int8_t *input0, int8_t *input1, int8_t *output, int element_size);
+int ElementGreaterInt8(int8_t *input0, int8_t *input1, int8_t *output, int element_size,
+                       ArithmeticQuantArg *quant_arg);
 
-int ElementGreaterEqual(int8_t *input0, int8_t *input1, int8_t *output, int element_size);
+int ElementGreaterEqualInt8(int8_t *input0, int8_t *input1, int8_t *output, int element_size,
+                            ArithmeticQuantArg *quant_arg);
 #endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_NNACL_INT8_ARITHMETIC_INT8_H_

mindspore/lite/src/runtime/kernel/arm/nnacl/quantization/quantize.h

@@ -193,6 +193,12 @@ typedef struct SubQuantArg {
   int right_shift_out_;
 } SubQuantArg;
 
+typedef struct ArithmeticQuantArg {
+  QuantArg in0_args_;
+  QuantArg in1_args_;
+  QuantArg out_args_;
+} ArithmeticQuantArg;
+
 void QuantizeMultiplier(double double_multiplier, int32_t *quantized_multiplier, int *shift);
 
 inline void QuantizeMultiplierSmallerThanOne(double double_multiplier, int32_t *quantized_multiplier,