COMPMID-2601 [CL] add mixed precision support to PoolingLayer
* PoolingLayerInfo is updated with a new flag.
* CL Kernel is updated to use FP32 accumulation.
* CL pooling layer testscases are added for mixed precision.
* Reference pooling layer is updated to use FP32 accumulation.
Change-Id: I4ab2167cc7f86c86293cf50a0ca5119c04dc9c7e
Signed-off-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1973
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: VidhyaSudhan Loganathan <vidhyasudhan.loganathan@arm.com>
diff --git a/tests/validation/reference/PoolingLayer.cpp b/tests/validation/reference/PoolingLayer.cpp
index 34b19ff..010412c 100644
--- a/tests/validation/reference/PoolingLayer.cpp
+++ b/tests/validation/reference/PoolingLayer.cpp
@@ -37,8 +37,8 @@
{
using namespace arm_compute::misc::shape_calculator;
-template <typename T>
-SimpleTensor<T> pooling_layer(const SimpleTensor<T> &src, const PoolingLayerInfo &info, const QuantizationInfo &output_qinfo)
+template <typename T, typename ACC_T, typename std::enable_if<is_floating_point<T>::value, int>::type>
+SimpleTensor<T> pooling_layer_internal(const SimpleTensor<T> &src, const PoolingLayerInfo &info, const QuantizationInfo &output_qinfo)
{
ARM_COMPUTE_UNUSED(output_qinfo); // requantization occurs in pooling_layer<uint8_t>
ARM_COMPUTE_ERROR_ON(info.is_global_pooling() && (src.shape().x() != src.shape().y()));
@@ -79,12 +79,12 @@
wstart = std::max(wstart, 0);
hstart = std::max(hstart, 0);
- T max_val = std::numeric_limits<T>::lowest();
+ auto max_val = std::numeric_limits<ACC_T>::lowest();
for(int y = hstart; y < hend; ++y)
{
for(int x = wstart; x < wend; ++x)
{
- const T val = src[r * h_src * w_src + y * w_src + x];
+ const auto val = static_cast<ACC_T>(src[r * h_src * w_src + y * w_src + x]);
if(val > max_val)
{
max_val = val;
@@ -92,7 +92,7 @@
}
}
- dst[r * h_dst * w_dst + h * w_dst + w] = max_val;
+ dst[r * h_dst * w_dst + h * w_dst + w] = static_cast<T>(max_val);
}
}
}
@@ -105,16 +105,16 @@
{
for(int w = 0; w < w_dst; ++w)
{
- T avg_val(0);
- int wstart = w * pool_stride_x - pad_left;
- int hstart = h * pool_stride_y - pad_top;
- int wend = std::min(wstart + pool_size_x, w_src + pad_right);
- int hend = std::min(hstart + pool_size_y, h_src + pad_bottom);
- int pool = (hend - hstart) * (wend - wstart);
- wstart = std::max(wstart, 0);
- hstart = std::max(hstart, 0);
- wend = std::min(wend, w_src);
- hend = std::min(hend, h_src);
+ ACC_T avg_val(0);
+ int wstart = w * pool_stride_x - pad_left;
+ int hstart = h * pool_stride_y - pad_top;
+ int wend = std::min(wstart + pool_size_x, w_src + pad_right);
+ int hend = std::min(hstart + pool_size_y, h_src + pad_bottom);
+ int pool = (hend - hstart) * (wend - wstart);
+ wstart = std::max(wstart, 0);
+ hstart = std::max(hstart, 0);
+ wend = std::min(wend, w_src);
+ hend = std::min(hend, h_src);
// Exclude padding pixels from the average
if(exclude_padding)
{
@@ -127,7 +127,7 @@
{
for(int x = wstart; x < wend; ++x)
{
- avg_val += src[r * h_src * w_src + y * w_src + x];
+ avg_val += static_cast<ACC_T>(src[r * h_src * w_src + y * w_src + x]);
}
}
dst[r * h_dst * w_dst + h * w_dst + w] = avg_val / pool;
@@ -138,11 +138,11 @@
{
for(int x = wstart; x < wend; ++x)
{
- const T val = src[r * h_src * w_src + y * w_src + x];
+ const auto val = static_cast<ACC_T>(src[r * h_src * w_src + y * w_src + x]);
avg_val += val * val;
}
}
- dst[r * h_dst * w_dst + h * w_dst + w] = std::sqrt(avg_val / pool);
+ dst[r * h_dst * w_dst + h * w_dst + w] = static_cast<T>(std::sqrt(avg_val / pool));
}
}
}
@@ -152,17 +152,37 @@
return dst;
}
+template SimpleTensor<float> pooling_layer_internal<float>(const SimpleTensor<float> &src, const PoolingLayerInfo &info, const QuantizationInfo &output_qinfo);
+template SimpleTensor<half> pooling_layer_internal<half>(const SimpleTensor<half> &src, const PoolingLayerInfo &info, const QuantizationInfo &output_qinfo);
+template SimpleTensor<half> pooling_layer_internal<half, float>(const SimpleTensor<half> &src, const PoolingLayerInfo &info, const QuantizationInfo &output_qinfo);
+
+template <typename T>
+SimpleTensor<T> pooling_layer(const SimpleTensor<T> &src, const PoolingLayerInfo &info, const QuantizationInfo &output_qinfo)
+{
+ return pooling_layer_internal<T, T>(src, info, output_qinfo);
+}
+
template <>
SimpleTensor<uint8_t> pooling_layer<uint8_t>(const SimpleTensor<uint8_t> &src, const PoolingLayerInfo &info, const QuantizationInfo &output_qinfo)
{
SimpleTensor<float> src_tmp = convert_from_asymmetric(src);
- SimpleTensor<float> dst_tmp = pooling_layer<float>(src_tmp, info, output_qinfo);
+ SimpleTensor<float> dst_tmp = pooling_layer_internal<float>(src_tmp, info, output_qinfo);
SimpleTensor<uint8_t> dst = convert_to_asymmetric<uint8_t>(dst_tmp, output_qinfo);
return dst;
}
+template <>
+SimpleTensor<half> pooling_layer(const SimpleTensor<half> &src, const PoolingLayerInfo &info, const QuantizationInfo &output_qinfo)
+{
+ if(src.data_type() == DataType::F16 && info.fp_mixed_precision())
+ {
+ return pooling_layer_internal<half, float>(src, info, output_qinfo);
+ }
+
+ return pooling_layer_internal<half>(src, info, output_qinfo);
+}
+
template SimpleTensor<float> pooling_layer(const SimpleTensor<float> &src, const PoolingLayerInfo &info, const QuantizationInfo &output_qinfo);
-template SimpleTensor<half> pooling_layer(const SimpleTensor<half> &src, const PoolingLayerInfo &info, const QuantizationInfo &output_qinfo);
} // namespace reference
} // namespace validation
} // namespace test
diff --git a/tests/validation/reference/PoolingLayer.h b/tests/validation/reference/PoolingLayer.h
index 1c0b7ff..fc36d51 100644
--- a/tests/validation/reference/PoolingLayer.h
+++ b/tests/validation/reference/PoolingLayer.h
@@ -35,6 +35,8 @@
{
namespace reference
{
+template <typename T, typename ACC_T = T, typename std::enable_if<is_floating_point<T>::value, int>::type = 0>
+SimpleTensor<T> pooling_layer_internal(const SimpleTensor<T> &src, const PoolingLayerInfo &info, const QuantizationInfo &output_qinfo);
template <typename T>
SimpleTensor<T> pooling_layer(const SimpleTensor<T> &src, const PoolingLayerInfo &info, const QuantizationInfo &output_qinfo);
} // namespace reference