Add CLPool3d Int8 Support
- Adds Qasymm8 and Qasymm8_signed support to the 3d pool operator
Resolves: COMPMID-4669
Signed-off-by: Mohammed Suhail Munshi <MohammedSuhail.Munshi@arm.com>
Change-Id: I36038c2b7c4f36baf67f7aae801356890e104538
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/c/VisualCompute/ComputeLibrary/+/410496
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Sheri Zhang <sheri.zhang@arm.com>
Comments-Addressed: bsgcomp <bsgcomp@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7391
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/core/CL/cl_kernels/nhwc/pooling_3d_layer.cl b/src/core/CL/cl_kernels/nhwc/pooling_3d_layer.cl
index 7c64143..4e5481d 100644
--- a/src/core/CL/cl_kernels/nhwc/pooling_3d_layer.cl
+++ b/src/core/CL/cl_kernels/nhwc/pooling_3d_layer.cl
@@ -177,8 +177,15 @@
res0 = SQRT_OP(res0);
#endif // defined(POOL_L2)
- // Store result
-#if defined(FP_MIXED_PRECISION)
+ VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+ out_q0 = CONVERT(res0, VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE));
+
+
+
+ // Store result
+#if defined(QUANTIZED)
+ STORE_VECTOR_SELECT(out_q, DATA_TYPE, out_base_ptr, VEC_SIZE, VEC_SIZE_LEFTOVER, (VEC_SIZE_LEFTOVER != 0) && get_global_id(0) == 0);
+#elif defined(FP_MIXED_PRECISION)
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
res_converted0 = CONVERT(res0, VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE));
STORE_VECTOR_SELECT(res_converted, DATA_TYPE, out_base_ptr, VEC_SIZE, VEC_SIZE_LEFTOVER, (VEC_SIZE_LEFTOVER != 0) && get_global_id(0) == 0);
diff --git a/src/core/CL/cl_kernels/nhwc/pooling_3d_layer_quantized.cl b/src/core/CL/cl_kernels/nhwc/pooling_3d_layer_quantized.cl
new file mode 100644
index 0000000..abf0db9
--- /dev/null
+++ b/src/core/CL/cl_kernels/nhwc/pooling_3d_layer_quantized.cl
@@ -0,0 +1,185 @@
+/*
+ * Copyright (c) 2022 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "helpers.h"
+#include "tile_helpers.h" // Needed for GET_SPATIAL_IDX()
+
+#if defined(POOL_AVG)
+#define POOL_OP(x, y) ((x) + (y))
+#else /* defined(POOL_AVG) */
+#define POOL_OP(x, y) (max((x), (y)))
+#endif /* defined(POOL_AVG) */
+
+#define SQRT_OP(x) sqrt((x))
+
+#if defined(VEC_SIZE) && defined(VEC_SIZE_LEFTOVER) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(SRC_DEPTH) && defined(DST_CHANNELS) && defined(DST_HEIGHT) && defined(DST_DEPTH) && defined(DST_BATCH_SIZE) && defined(ACC_DATA_TYPE)
+
+#if defined(POOL_SIZE_X) && defined(POOL_SIZE_Y) && defined(POOL_SIZE_Z)
+
+#if defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT)
+#define VEC_FLOAT(VEC_SIZE) VEC_DATA_TYPE(float, VEC_SIZE)
+#define VEC_INT(VEC_SIZE) VEC_DATA_TYPE(int, VEC_SIZE)
+#define CONVERT_RTE(x, type) (convert_##type##_rte((x)))
+#define CONVERT_DOWN(x, type) CONVERT_RTE(x, type)
+#define REQUANTIZE(VEC_SIZE, input, in_offset, out_offset, in_scale, out_scale, res) \
+ { \
+ const VEC_FLOAT(VEC_SIZE) in_f32 = (CONVERT(input, VEC_FLOAT(VEC_SIZE)) - (VEC_FLOAT(VEC_SIZE))((float)in_offset)) * (VEC_FLOAT(VEC_SIZE))((float)in_scale); \
+ const VEC_FLOAT(VEC_SIZE) out_f32 = in_f32 / ((VEC_FLOAT(VEC_SIZE))(float)out_scale) + ((VEC_FLOAT(VEC_SIZE))((float)out_offset)); \
+ res = CONVERT_SAT(CONVERT_DOWN(out_f32, VEC_INT(VEC_SIZE)), VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)); \
+ }
+#endif /* defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) */
+
+#if defined(POOL_L2)
+#error "L2 pooling is not supported"
+#endif /* defined(POOL_L2) */
+
+/** Performs 3d pooling layer of size equal to MxNXD. This OpenCL kernel can perform the following pooling types:
+ * -# max, -DPOOL_MAX must be passed at compile time
+ * -# average, -DPOOL_AVG must be passed at compile time. If padding has to be excluded, -DEXCLUDE_PADDING should be passed at compile time
+ *
+ * @note Datatype must be passed at compile type using -DDATA_TYPE e.g. -DDATA_TYPE=half. Supported data types are QASYMM8_SIGNED, QASYMM8
+ * @note Accumulation data type must be passed at compile time using -DACC_DATA_TYPE e.g. -DACC_DATA_TYPE=float
+ * @note If -DFP_MIXED_PRECISION is passed at compile time, the kernel will use F32 for the partial result
+ * @note Pool size must be passed at compile time using -DPOOL_SIZE_X, -DPOOL_SIZE_Y, and -DPOOL_SIZE_Z. e.g. -DPOOL_SIZE_X=4, -DPOOL_SIZE_Y=4, -DPOOL_SIZE_Z=2
+ * @note Input tensor width, height and depth must be passed at compile time using -DSRC_WIDTH, -DSRC_HEIGHT, and -DSRC_DEPTH
+ * @note Output tensor height, channels, depth, and batch size must be passed at compile time using -DDST_HEIGHT, -DDST_CHANNELS, -DDST_DEPTH, and -DDST_BATCH_SIZE
+ * @note Pool strides must be passed at compile time using -DSTRIDE_X, -DSTRIDE_Y and -DSTRIDE_Z which are the steps of the window along the x, y and z directions
+ * @note Pool pads must be passed at compile time using -DPAD_X, -DPAD_Y, -DPAD_Z
+ * @note Vector size must be passed at compile time using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16
+ * @note Leftover vector size must be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VEC_SIZE
+ * @note The initial value for the pooling operation must be passed at compile time using -DINITIAL_VALUE e.g. -DINITIAL_VALUE=0
+ *
+ * @param[in] input_ptr Pointer to the source tensor. Supported data types: QASYMM8_SIGNED, QASYMM8
+ * @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes)
+ * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] input_stride_y Stride of the source tensor in Y dimension (in bytes)
+ * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes)
+ * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in] input_stride_w Stride of the source tensor in W dimension (in bytes)
+ * @param[in] input_step_w input_stride_w * number of elements along W processed per workitem(in bytes)
+ * @param[in] input_stride_v Stride of the source tensor in V dimension (in bytes)
+ * @param[in] input_step_v input_stride_v * number of elements along V processed per workitem(in bytes)
+ * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source tensor
+ * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr
+ * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes)
+ * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] output_stride_z Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in] output_stride_w Stride of the destination tensor in W dimension (in bytes)
+ * @param[in] output_step_w output_stride_w * number of elements along W processed per workitem(in bytes)
+ * @param[in] output_stride_v Stride of the destination tensor in V dimension (in bytes)
+ * @param[in] output_step_v output_stride_v * number of elements along V processed per workitem(in bytes)
+ * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ */
+__kernel void pooling_3d_layer_MxN_ndhwc_quantized(
+ TENSOR5D_DECLARATION(input),
+ TENSOR5D_DECLARATION(output))
+{
+ // Note: If C is not multiple of VEC_SIZE, we shift back of VEC_SIZE_LEFTOVER elements to compute the leftover elements for get_global_id(0) == 0
+ // Note: If C is less than VEC_SIZE, VEC_SIZE should be shrunk to the closest smaller VEC_SIZE. This operation is performed on the host side
+ int idx_out_c = GET_SPATIAL_IDX(0, VEC_SIZE, VEC_SIZE_LEFTOVER);
+ int idx_out_w = GET_SPATIAL_IDX(1, 1, 0);
+
+ // The depth size dimension and the batch size dimension are collapsed over the height dimension
+ int idx_out_h = GET_SPATIAL_IDX(2, 1, 0) % DST_HEIGHT;
+ int idx_out_d = (GET_SPATIAL_IDX(2, 1, 0) / DST_HEIGHT) % DST_DEPTH;
+ int idx_out_n = (GET_SPATIAL_IDX(2, 1, 0) / DST_HEIGHT) / DST_DEPTH;
+
+ __global unsigned char *in_base_ptr = input_ptr + input_offset_first_element_in_bytes + idx_out_c * sizeof(DATA_TYPE) + idx_out_n * input_stride_v;
+
+ __global unsigned char *out_base_ptr = output_ptr + output_offset_first_element_in_bytes + idx_out_c * sizeof(DATA_TYPE) + idx_out_w * output_stride_y + idx_out_h * output_stride_z + idx_out_d *
+ output_stride_w + idx_out_n * output_stride_v;
+
+ VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE)
+ res0 = INITIAL_VALUE;
+
+ int idx_in_w = idx_out_w * STRIDE_X - (int)PAD_X;
+ int idx_in_h = idx_out_h * STRIDE_Y - (int)PAD_Y;
+ int idx_in_d = idx_out_d * STRIDE_Z - (int)PAD_Z;
+
+ // The start of width to consider in calculation should exclude padding
+ int pool_x_s = max((int)0, -idx_in_w);
+ // Assumed Symmetric Padding (left padding = right padding = PAD_X), the filter end should be either the pool width or what is remaining from current pos to the (src width + pad right)
+ int pool_x_e = min((int)POOL_SIZE_X, (int)SRC_WIDTH + PAD_X - idx_in_w);
+ int pool_y_s = max((int)0, -idx_in_h);
+ int pool_y_e = min((int)POOL_SIZE_Y, (int)SRC_HEIGHT + PAD_Y - idx_in_h);
+ int pool_z_s = max((int)0, -idx_in_d);
+ int pool_z_e = min((int)POOL_SIZE_Z, (int)SRC_DEPTH + PAD_Z - idx_in_d);
+
+#if defined(POOL_AVG) && defined(EXCLUDE_PADDING)
+ int filter_size = 0;
+#elif defined(POOL_AVG) && !defined(EXCLUDE_PADDING) // defined(POOL_AVG) && defined(EXCLUDE_PADDING)
+ int filter_size = pool_z_e * pool_y_e * pool_x_e;
+#endif // defined(POOL_AVG) && !defined(EXCLUDE_PADDING)
+
+ // The end of width to consider in calculation should exclude PAD_X
+ pool_x_e = min(pool_x_e, SRC_WIDTH - idx_in_w);
+ pool_y_e = min(pool_y_e, SRC_HEIGHT - idx_in_h);
+ pool_z_e = min(pool_z_e, SRC_DEPTH - idx_in_d);
+
+ for(int z = pool_z_s; z < pool_z_e; ++z)
+ {
+ int depth_offset_src = (z + idx_in_d) * input_stride_w;
+ for(int y = pool_y_s; y < pool_y_e; ++y)
+ {
+ int height_offset_src = (y + idx_in_h) * input_stride_z;
+#pragma unroll 8
+ for(int x = pool_x_s; x < pool_x_e; ++x)
+ {
+ int width_offset_src = (x + idx_in_w) * input_stride_y;
+
+ VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+ data;
+ VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE)
+ data0;
+
+ data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(in_base_ptr + width_offset_src + height_offset_src + depth_offset_src));
+ data0 = CONVERT(data, VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE));
+
+ res0 = POOL_OP(res0, data0);
+
+#if defined(POOL_AVG) && defined(EXCLUDE_PADDING)
+ filter_size++;
+#endif // defined(POOL_AVG) && defined(EXCLUDE_PADDING)
+ }
+ }
+ }
+
+#if defined(POOL_AVG)
+ res0 = (res0 + (VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE))(filter_size >> 1)) / filter_size;
+#endif // defined(POOL_AVG)
+
+ VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+ out_q0 = CONVERT(res0, VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE));
+
+#if defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT)
+ REQUANTIZE(VEC_SIZE, out_q0, OFFSET_IN1, OFFSET_OUT, SCALE_IN1, SCALE_OUT, out_q0);
+#endif /* defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) */
+
+ STORE_VECTOR_SELECT(out_q, DATA_TYPE, out_base_ptr, VEC_SIZE, VEC_SIZE_LEFTOVER, (VEC_SIZE_LEFTOVER != 0) && get_global_id(0) == 0);
+}
+#endif // defined(POOL_SIZE_X) && defined(POOL_SIZE_Y) && defined(POOL_SIZE_Z)
+#endif // defined(VEC_SIZE) && defined(VEC_SIZE_LEFTOVER) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(SRC_DEPTH) && defined(DST_CHANNELS) && defined(DST_HEIGHT) && defined(DST_DEPTH) && defined(DST_BATCH_SIZE) && defined(ACC_DATA_TYPE)
diff --git a/src/gpu/cl/ClKernelLibrary.cpp b/src/gpu/cl/ClKernelLibrary.cpp
index f675dbe..1bf7f2b 100644
--- a/src/gpu/cl/ClKernelLibrary.cpp
+++ b/src/gpu/cl/ClKernelLibrary.cpp
@@ -438,6 +438,7 @@
{ "pooling_layer_2x2_nhwc", "nhwc/pooling_layer.cl" },
{ "pooling_layer_MxN_quantized_nhwc", "nhwc/pooling_layer_quantized.cl" },
{ "pooling_3d_layer_MxN_ndhwc", "nhwc/pooling_3d_layer.cl" },
+ { "pooling_3d_layer_MxN_ndhwc_quantized", "nhwc/pooling_3d_layer_quantized.cl" },
{ "reorg_layer_nhwc", "nhwc/reorg_layer.cl" },
{ "scale_nearest_neighbour_nhwc", "nhwc/scale.cl" },
{ "scale_bilinear_nhwc", "nhwc/scale.cl" },
@@ -907,6 +908,10 @@
#include "./cl_kernels/nhwc/pooling_3d_layer.clembed"
},
{
+ "nhwc/pooling_3d_layer_quantized.cl",
+#include "./cl_kernels/nhwc/pooling_3d_layer_quantized.clembed"
+ },
+ {
"nhwc/pooling_layer_quantized.cl",
#include "./cl_kernels/nhwc/pooling_layer_quantized.clembed"
},
diff --git a/src/gpu/cl/kernels/ClPool3dKernel.cpp b/src/gpu/cl/kernels/ClPool3dKernel.cpp
index 929ccf7..a090ac5 100644
--- a/src/gpu/cl/kernels/ClPool3dKernel.cpp
+++ b/src/gpu/cl/kernels/ClPool3dKernel.cpp
@@ -49,19 +49,22 @@
ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src);
ARM_COMPUTE_RETURN_ERROR_ON_MSG((pool_info.stride.x() == 0 || pool_info.stride.y() == 0 || pool_info.stride.z() == 0), "Strides cannot be zero.");
- ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F16, DataType::F32);
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F16, DataType::F32, DataType::QASYMM8_SIGNED, DataType::QASYMM8);
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG((!is_data_type_float(src->data_type())) && (!pool_info.exclude_padding
+ && (pool_info.pool_type == PoolingType::AVG)),
+ "Exclude padding is unsupported for non-float types for Avg op");
- const auto data_layout = src->data_layout();
- const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
- const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
- const int idx_depth = get_data_layout_dimension_index(data_layout, DataLayoutDimension::DEPTH);
- const bool is_global_pooling = pool_info.is_global_pooling;
+ const auto data_layout = src->data_layout();
+ const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
+ const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
+ const int idx_depth = get_data_layout_dimension_index(data_layout, DataLayoutDimension::DEPTH);
+ const bool is_global_pooling = pool_info.is_global_pooling;
const unsigned int pool_size_x = is_global_pooling ? src->dimension(idx_width) : pool_info.pool_size.width;
const unsigned int pool_size_y = is_global_pooling ? src->dimension(idx_height) : pool_info.pool_size.height;
const unsigned int pool_size_z = is_global_pooling ? src->dimension(idx_depth) : pool_info.pool_size.depth;
- int output_width = 0;
- int output_height = 0;
- int output_depth = 0;
+ int output_width = 0;
+ int output_height = 0;
+ int output_depth = 0;
bool round_type_ceil_with_asymm_padding = (pool_info.round_type == DimensionRoundingType::CEIL) && (!is_symmetric(pool_info.padding));
ARM_COMPUTE_RETURN_ERROR_ON_MSG(round_type_ceil_with_asymm_padding, "Cannot use dimension round type CEIL when padding is asymmetric.");
@@ -143,10 +146,31 @@
build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(src->dimension(idx_height)));
build_opts.add_option("-DSRC_DEPTH=" + support::cpp11::to_string(src->dimension(idx_depth)));
+ // If datatype is quantized add relevant parameters
+ if(is_data_type_quantized_asymmetric(data_type) && src->quantization_info() != dst->quantization_info())
+ {
+ const UniformQuantizationInfo iq_info = src->quantization_info().uniform();
+ const UniformQuantizationInfo oq_info = dst->quantization_info().uniform();
+
+ build_opts.add_option("-DOFFSET_IN1=" + float_to_string_with_full_precision(iq_info.offset));
+ build_opts.add_option("-DOFFSET_OUT=" + float_to_string_with_full_precision(oq_info.offset));
+ build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(iq_info.scale));
+ build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oq_info.scale));
+ }
+
// Set the initial value for the pooling operation accordingly with the data type
if(pool_type == PoolingType::MAX)
{
- build_opts.add_option("-DINITIAL_VALUE=" + float_to_string_with_full_precision(std::numeric_limits<float>::lowest()));
+ if(is_data_type_quantized(data_type))
+ {
+ PixelValue type_min{};
+ std::tie(type_min, std::ignore) = get_min_max(data_type);
+ build_opts.add_option("-DINITIAL_VALUE=" + support::cpp11::to_string(type_min.get<int32_t>()));
+ }
+ else
+ {
+ build_opts.add_option("-DINITIAL_VALUE=" + float_to_string_with_full_precision(std::numeric_limits<float>::lowest()));
+ }
}
else
{
@@ -164,6 +188,11 @@
{
acc_data_type = DataType::F32;
}
+ else if(is_data_type_quantized(data_type) && pool_type != PoolingType::MAX) // Use S32 for avg pooling to allow for integer division
+ {
+ acc_data_type = DataType::S32;
+ }
+
build_opts.add_option("-DACC_DATA_TYPE=" + get_cl_type_from_data_type(acc_data_type));
build_opts.add_option_if(use_fp_mixed_precision, "-DFP_MIXED_PRECISION");
build_opts.add_option_if(exclude_padding, "-DEXCLUDE_PADDING");
@@ -172,8 +201,10 @@
build_opts.add_option("-DDST_CHANNELS=" + support::cpp11::to_string(dst->dimension(idx_channel)));
build_opts.add_option("-DDST_BATCH_SIZE=" + support::cpp11::to_string(dst->dimension(idx_batch_size)));
build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(src->dimension(0) % _num_elems_processed_per_iteration));
- std::string kernel_name = "pooling_3d_layer_MxN_ndhwc";
- _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
+
+ // if datatype is quantized use quantized kernel function
+ std::string kernel_name = (is_data_type_quantized_asymmetric(data_type) ? "pooling_3d_layer_MxN_ndhwc_quantized" : "pooling_3d_layer_MxN_ndhwc");
+ _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
// Configure kernel window
Window win = calculate_max_window(*dst, Steps(_num_elems_processed_per_iteration));
diff --git a/src/gpu/cl/kernels/ClPool3dKernel.h b/src/gpu/cl/kernels/ClPool3dKernel.h
index 30c76ed..0085234 100644
--- a/src/gpu/cl/kernels/ClPool3dKernel.h
+++ b/src/gpu/cl/kernels/ClPool3dKernel.h
@@ -46,7 +46,7 @@
* @note Asymmetric padding is not supported when dimension rounding type == CEIL.
*
* @param[in] compile_context The compile context to be used.
- * @param[in] src Source tensor info. Data types supported: F16/F32.
+ * @param[in] src Source tensor info. Data types supported: F16/F32/QASYMM8/QASYMM8_SIGNED
* @param[out] dst Destination tensor info. Data types supported: same as @p src.
* @param[in] pool_info Contains pooling operation information described in @ref Pooling3dLayerInfo.
*/