COMPMID-2306: CLDepthwiseConvolution: support for QUANT8_PER_CHANNEL_SYMM - Reference
This patch modifies the reference implementation and the fixtures of
depthwise convolution layer to support QSYMM8_PER_CHANNEL quantization.
Change-Id: I28adb5c110308b1024a213bec2d35a89180a46dc
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2063
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Giuseppe Rossini <giuseppe.rossini@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
diff --git a/tests/validation/reference/DepthwiseConvolutionLayer.cpp b/tests/validation/reference/DepthwiseConvolutionLayer.cpp
index b1d2b92..7458f81 100644
--- a/tests/validation/reference/DepthwiseConvolutionLayer.cpp
+++ b/tests/validation/reference/DepthwiseConvolutionLayer.cpp
@@ -40,7 +40,9 @@
{
namespace reference
{
-/** Perform a depthwise convolution
+namespace
+{
+/** Perform a depthwise convolution for floating-point types
*
* - Three dimensions tensors
* - Third dimention is number of channels
@@ -48,9 +50,9 @@
* - Padding, stride and output shape "match"
*
*/
-template <typename T, typename TB>
-SimpleTensor<T> depthwise_convolution(const SimpleTensor<T> &src, const SimpleTensor<T> &weights, const SimpleTensor<TB> &biases, const TensorShape &dst_shape, const PadStrideInfo &conv_info,
- unsigned int depth_multiplier, const Size2D &dilation, const QuantizationInfo &out_quant_info)
+template <typename T>
+SimpleTensor<T> depthwise_convolution_fp(const SimpleTensor<T> &src, const SimpleTensor<T> &weights, const SimpleTensor<T> &biases, const TensorShape &dst_shape, const PadStrideInfo &conv_info,
+ unsigned int depth_multiplier, const Size2D &dilation, const QuantizationInfo &out_quant_info)
{
ARM_COMPUTE_UNUSED(out_quant_info);
@@ -114,7 +116,7 @@
}
}
- dst[out_pos++] = saturate_cast<T>(val + *static_cast<const TB *>(biases(Coordinates(out_z))));
+ dst[out_pos++] = saturate_cast<T>(val + *static_cast<const T *>(biases(Coordinates(out_z))));
}
}
}
@@ -124,26 +126,32 @@
return dst;
}
-template <>
-SimpleTensor<uint8_t> depthwise_convolution(const SimpleTensor<uint8_t> &src, const SimpleTensor<uint8_t> &weights, const SimpleTensor<int32_t> &biases, const TensorShape &dst_shape,
- const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation, const QuantizationInfo &out_quant_info)
+/** Perform a quantized depthwise convolution
+ *
+ * - Three dimensions tensors
+ * - Third dimention is number of channels
+ * - Depths of input tensor and filter are equals
+ * - Padding, stride and output shape "match"
+ * - QASYMM8 input, output
+ * - QASYMM8 or QSYMM8_PER_CHANNEL filter
+ *
+ */
+template <typename T, typename TW, typename TB>
+SimpleTensor<T> depthwise_convolution_quantized(const SimpleTensor<T> &src, const SimpleTensor<TW> &weights, const SimpleTensor<int32_t> &biases, const TensorShape &dst_shape,
+ const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation, const QuantizationInfo &out_quant_info)
{
// if no explicit quantization has been set you the same as src
const QuantizationInfo &dst_qinfo = out_quant_info.uniform().empty() ? src.quantization_info() : out_quant_info;
- SimpleTensor<uint8_t> dst{ dst_shape, src.data_type(), 1, dst_qinfo };
+ SimpleTensor<T> dst{ dst_shape, src.data_type(), 1, dst_qinfo };
// Create reference
const int input_offset = -src.quantization_info().uniform().offset;
const float input_scale = src.quantization_info().uniform().scale;
const int weights_offset = -weights.quantization_info().uniform().offset;
- const float weights_scale = weights.quantization_info().uniform().scale;
const int output_offset = dst_qinfo.uniform().offset;
const float output_scale = dst_qinfo.uniform().scale;
- int output_multiplier = 0;
- int output_shift = 0;
- const float multiplier = input_scale * weights_scale / output_scale;
- arm_compute::quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
+ const std::vector<float> weights_scale_vec = weights.quantization_info().scale();
// Compute reference
const int filter_width = weights.shape().x();
@@ -173,11 +181,19 @@
const int maximum_x = input_width + pad_left + pad_right - static_cast<int>(patch_width);
const int maximum_y = input_height + pad_top + pad_bottom - static_cast<int>(patch_height);
+ const bool is_quantized_per_channel = is_data_type_quantized_per_channel(weights.data_type());
+
int out_pos = 0;
for(int r = 0; r < num_batches; ++r)
{
for(int z = 0; z < input_depth; ++z)
{
+ int output_multiplier = 0;
+ int output_shift = 0;
+ const float weights_scale = (is_quantized_per_channel) ? weights_scale_vec[z] : weights_scale_vec[0];
+ const float multiplier = input_scale * weights_scale / output_scale;
+ arm_compute::quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
+
for(unsigned int m = 0; m < depth_multiplier; ++m)
{
const int out_z = z * depth_multiplier + m;
@@ -197,8 +213,8 @@
{
coords.set(0, i);
coords.set(1, j);
- const auto in_val = tensor_elem_at<uint8_t>(src, coords, BorderMode::CONSTANT, -input_offset);
- const uint8_t w_val = *(weights.data() + filter_offset);
+ const auto in_val = tensor_elem_at<T>(src, coords, BorderMode::CONSTANT, -input_offset);
+ const TW w_val = *(weights.data() + filter_offset);
val += (in_val + input_offset) * (w_val + weights_offset);
++filter_offset;
}
@@ -206,8 +222,7 @@
val += bias_val;
val = asymm_rounding_divide_by_pow2(asymm_int_mult(val, output_multiplier), output_shift);
val += output_offset;
- val = std::max<int32_t>(val, 0);
- val = std::min<int32_t>(val, 255);
+ val = utility::clamp<int32_t>(val, 0, 255);
// Store the result
dst[out_pos++] = val;
@@ -219,12 +234,35 @@
return dst;
}
+} // namespace
-template SimpleTensor<float> depthwise_convolution(const SimpleTensor<float> &src, const SimpleTensor<float> &weights, const SimpleTensor<float> &biases, const TensorShape &dst_shape,
- const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation, const QuantizationInfo &out_quant_info);
+template <>
+SimpleTensor<float> depthwise_convolution(const SimpleTensor<float> &src, const SimpleTensor<float> &weights, const SimpleTensor<float> &biases, const TensorShape &dst_shape,
+ const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation, const QuantizationInfo &out_quant_info)
+{
+ return depthwise_convolution_fp(src, weights, biases, dst_shape, conv_info, depth_multiplier, dilation, out_quant_info);
+}
-template SimpleTensor<half> depthwise_convolution(const SimpleTensor<half> &src, const SimpleTensor<half> &weights, const SimpleTensor<half> &biases, const TensorShape &dst_shape,
- const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation, const QuantizationInfo &out_quant_info);
+template <>
+SimpleTensor<half> depthwise_convolution(const SimpleTensor<half> &src, const SimpleTensor<half> &weights, const SimpleTensor<half> &biases, const TensorShape &dst_shape,
+ const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation, const QuantizationInfo &out_quant_info)
+{
+ return depthwise_convolution_fp(src, weights, biases, dst_shape, conv_info, depth_multiplier, dilation, out_quant_info);
+}
+
+template <>
+SimpleTensor<uint8_t> depthwise_convolution(const SimpleTensor<uint8_t> &src, const SimpleTensor<uint8_t> &weights, const SimpleTensor<int32_t> &biases, const TensorShape &dst_shape,
+ const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation, const QuantizationInfo &out_quant_info)
+{
+ return depthwise_convolution_quantized<uint8_t, uint8_t, int32_t>(src, weights, biases, dst_shape, conv_info, depth_multiplier, dilation, out_quant_info);
+}
+
+template <>
+SimpleTensor<uint8_t> depthwise_convolution(const SimpleTensor<uint8_t> &src, const SimpleTensor<int8_t> &weights, const SimpleTensor<int32_t> &biases, const TensorShape &dst_shape,
+ const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation, const QuantizationInfo &out_quant_info)
+{
+ return depthwise_convolution_quantized<uint8_t, int8_t, int32_t>(src, weights, biases, dst_shape, conv_info, depth_multiplier, dilation, out_quant_info);
+}
} // namespace reference
} // namespace validation
} // namespace test