COMPMID-2805: Add QASYMM8_SIGNED support in NEGEMMLowpOutputStage
Add support from requantizing down from S32 to Int8 with fixed point
requantization. This involves the following:
- Compute fixed point multiplication between each entry of input by
result_fixedpoint_multiplier
- Add bias to final result if bias tensor is not a nullptr
- Round to nearest division by a power-of-two using result_shift
- Add offset to each result
- Clamp the value between the specified min and max bounds
- Cast to int8 data type
Change-Id: I641b3fac0833c568d8565ccb859bbc561a24c17d
Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2340
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
diff --git a/tests/validation/fixtures/GEMMLowpFixture.h b/tests/validation/fixtures/GEMMLowpFixture.h
index 5d092ec..c17105e 100644
--- a/tests/validation/fixtures/GEMMLowpFixture.h
+++ b/tests/validation/fixtures/GEMMLowpFixture.h
@@ -254,8 +254,8 @@
output_stage.gemmlowp_offset, output_stage.gemmlowp_multipliers, output_stage.gemmlowp_shifts, output_stage.gemmlowp_min_bound, output_stage.gemmlowp_max_bound);
break;
case GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT:
- return reference::gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint<int32_t>(output, bias,
- output_stage.gemmlowp_multipliers, output_stage.gemmlowp_shifts, output_stage.gemmlowp_offset, output_stage.gemmlowp_min_bound, output_stage.gemmlowp_max_bound);
+ return reference::gemmlowp_quantize_down_scale_by_fixedpoint<int32_t, uint8_t>(output, bias,
+ output_stage.gemmlowp_multipliers, output_stage.gemmlowp_shifts, output_stage.gemmlowp_offset, output_stage.gemmlowp_min_bound, output_stage.gemmlowp_max_bound);
break;
default:
ARM_COMPUTE_ERROR("Not Supported!");
@@ -361,6 +361,101 @@
};
template <typename TensorType, typename AccessorType, typename FunctionType>
+class GEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointValidationFixture : public framework::Fixture
+{
+public:
+ template <typename...>
+ void setup(TensorShape shape, int32_t result_fixedpoint_multiplier, int32_t result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max, bool add_bias)
+ {
+ _target = compute_target(shape, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max, add_bias);
+ _reference = compute_reference(shape, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max, add_bias);
+ }
+
+protected:
+ template <typename U>
+ void fill(U &&tensor, int i)
+ {
+ std::uniform_int_distribution<> distribution(-6000, 6000);
+ library->fill(tensor, distribution, i);
+ }
+
+ TensorType compute_target(const TensorShape &shape, int32_t result_fixedpoint_multiplier, int32_t result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max, bool add_bias)
+ {
+ TensorShape shape_bias(shape[0]);
+
+ // Create tensors
+ TensorType a = create_tensor<TensorType>(shape, DataType::S32, 1);
+ TensorType b = create_tensor<TensorType>(shape_bias, DataType::S32, 1);
+ TensorType c = create_tensor<TensorType>(shape, DataType::QASYMM8_SIGNED, 1);
+
+ // Create and configure function
+ FunctionType output_stage;
+ output_stage.configure(&a, add_bias ? &b : nullptr, &c, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max);
+
+ ARM_COMPUTE_EXPECT(a.info()->is_resizable(), framework::LogLevel::ERRORS);
+ ARM_COMPUTE_EXPECT(c.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+ // Allocate tensors
+ a.allocator()->allocate();
+ c.allocator()->allocate();
+
+ ARM_COMPUTE_EXPECT(!a.info()->is_resizable(), framework::LogLevel::ERRORS);
+ ARM_COMPUTE_EXPECT(!c.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+ // Fill tensor
+ fill(AccessorType(a), 0);
+
+ if(add_bias)
+ {
+ ARM_COMPUTE_EXPECT(b.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+ // Allocate bias tensor
+ b.allocator()->allocate();
+
+ ARM_COMPUTE_EXPECT(!b.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+ // Fill tensor
+ fill(AccessorType(b), 1);
+ }
+
+ // Compute GEMM function
+ output_stage.run();
+ return c;
+ }
+
+ SimpleTensor<int8_t> compute_reference(const TensorShape &shape, int32_t result_fixed_point_multiplier, int32_t result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max,
+ bool add_bias)
+ {
+ // Create reference
+ TensorShape shape_bias(shape[0]);
+
+ SimpleTensor<int32_t> a{ shape, DataType::S32, 1 };
+ SimpleTensor<int32_t> b{ shape_bias, DataType::S32, 1 };
+
+ // Fill reference
+ fill(a, 0);
+
+ const std::vector<int32_t> result_fixed_point_multiplier_vec = { result_fixed_point_multiplier };
+ const std::vector<int32_t> result_shift_vec = { result_shift };
+
+ if(add_bias)
+ {
+ // Fill bias
+ fill(b, 1);
+
+ return reference::gemmlowp_quantize_down_scale_by_fixedpoint<int32_t, int8_t>(a, b, result_fixed_point_multiplier_vec, result_shift_vec, result_offset_after_shift, min, max);
+ }
+ else
+ {
+ return reference::gemmlowp_quantize_down_scale_by_fixedpoint<int32_t, int8_t>(a, result_fixed_point_multiplier_vec, result_shift_vec, result_offset_after_shift, min, max);
+ }
+ }
+
+ TensorType _target{};
+ SimpleTensor<int8_t> _reference{};
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType>
class GEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointValidationFixture : public framework::Fixture
{
public:
@@ -443,11 +538,11 @@
// Fill bias
fill(b, 1);
- return reference::gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint<int32_t>(a, b, result_fixed_point_multiplier_vec, result_shift_vec, result_offset_after_shift, min, max);
+ return reference::gemmlowp_quantize_down_scale_by_fixedpoint<int32_t, uint8_t>(a, b, result_fixed_point_multiplier_vec, result_shift_vec, result_offset_after_shift, min, max);
}
else
{
- return reference::gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint<int32_t>(a, result_fixed_point_multiplier_vec, result_shift_vec, result_offset_after_shift, min, max);
+ return reference::gemmlowp_quantize_down_scale_by_fixedpoint<int32_t, uint8_t>(a, result_fixed_point_multiplier_vec, result_shift_vec, result_offset_after_shift, min, max);
}
}
@@ -530,16 +625,19 @@
// Fill reference
fill(a, 0);
+ const std::vector<int32_t> result_fixed_point_multiplier_vec = { result_fixed_point_multiplier };
+ const std::vector<int32_t> result_shift_vec = { result_shift };
+
if(add_bias)
{
// Fill bias
fill(b, 1);
- return reference::gemmlowp_quantize_down_int32_to_int16_scale_by_fixedpoint<int32_t>(a, b, result_fixed_point_multiplier, result_shift, min, max);
+ return reference::gemmlowp_quantize_down_scale_by_fixedpoint<int32_t, int16_t>(a, b, result_fixed_point_multiplier_vec, result_shift_vec, 0, min, max);
}
else
{
- return reference::gemmlowp_quantize_down_int32_to_int16_scale_by_fixedpoint<int32_t>(a, result_fixed_point_multiplier, result_shift, min, max);
+ return reference::gemmlowp_quantize_down_scale_by_fixedpoint<int32_t, int16_t>(a, result_fixed_point_multiplier_vec, result_shift_vec, 0, min, max);
}
}