COMPMID-2060: Support different qinfo in PoolingLayer
CL and Neon back ends now support different qinfos
Change-Id: I638d5f258ab2f99b40659601b4c5398d2c34c43b
Signed-off-by: Pablo Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/927
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
diff --git a/src/core/CL/cl_kernels/pooling_layer_quantized.cl b/src/core/CL/cl_kernels/pooling_layer_quantized.cl
index 198250b..919b76e 100644
--- a/src/core/CL/cl_kernels/pooling_layer_quantized.cl
+++ b/src/core/CL/cl_kernels/pooling_layer_quantized.cl
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -23,6 +23,19 @@
*/
#include "helpers.h"
+#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 VEC_UCHAR(VEC_SIZE) VEC_DATA_TYPE(uchar, 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_UCHAR(VEC_SIZE)); \
+ }
+#endif /* defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) */
+
#if defined(POOL_AVG)
#define POOL_OP(x, y) ((x) + (y))
#else /* defined(POOL_AVG) */
@@ -118,8 +131,22 @@
res = round(DIV_OP(res, calculate_avg_scale(POOL_SIZE_X, POOL_SIZE_Y, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y)));
#endif /* defined(POOL_AVG) */
- // Store result
- *(__global uchar *)output.ptr = convert_uchar(res);
+ uchar result_u8 = convert_uchar(res);
+
+#if defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT)
+
+ const float result_f32 = convert_float(result_u8);
+ const float input_offset = (float)OFFSET_IN1;
+ const float input_scale = (float)SCALE_IN1;
+ const float scale_out = (float)SCALE_OUT;
+ const float offset_out = (float)OFFSET_OUT;
+ const float in_f32 = (result_f32 - input_offset) * input_scale;
+ const float out_f32 = in_f32 / scale_out + offset_out;
+ result_u8 = convert_uchar_sat(convert_int_rte(out_f32));
+
+#endif /* defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) */
+
+ *(__global uchar *)output.ptr = result_u8;
}
int calculate_avg_scale_nhwc(const int pool_size_x, const int pool_size_y, int upper_bound_w, int upper_bound_h,
@@ -217,6 +244,11 @@
vdata = convert_int8(round(DIV_OP_NHWC(vdata, calculate_avg_scale_nhwc(POOL_SIZE_X, POOL_SIZE_Y, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y))));
#endif /* defined(POOL_AVG) */
+ uchar8 out_u8 = convert_uchar8(vdata);
+#if defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT)
+ REQUANTIZE(8, out_u8, OFFSET_IN1, OFFSET_OUT, SCALE_IN1, SCALE_OUT, out_u8);
+#endif /* defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) */
+
// Store result
- vstore8(convert_uchar8(vdata), 0, (__global uchar *)output.ptr);
-}
\ No newline at end of file
+ vstore8(out_u8, 0, (__global uchar *)output.ptr);
+}
diff --git a/src/core/CL/kernels/CLPoolingLayerKernel.cpp b/src/core/CL/kernels/CLPoolingLayerKernel.cpp
index 7081688..7ccbda9 100644
--- a/src/core/CL/kernels/CLPoolingLayerKernel.cpp
+++ b/src/core/CL/kernels/CLPoolingLayerKernel.cpp
@@ -78,7 +78,6 @@
{
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output);
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
TensorInfo out_info(TensorInfo(compute_pool_shape(*input, pool_info), 1, output->data_type()));
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &out_info);
}
@@ -201,6 +200,17 @@
const int pool_pad_top = pad_stride_info.pad_top();
const int pool_pad_left = pad_stride_info.pad_left();
+ // Set build options
+ CLBuildOptions build_opts;
+
+ if(is_data_type_quantized_asymmetric(input->info()->data_type()) && input->info()->quantization_info() != output->info()->quantization_info())
+ {
+ build_opts.add_option("-DOFFSET_IN1=" + float_to_string_with_full_precision(input->info()->quantization_info().offset));
+ build_opts.add_option("-DOFFSET_OUT=" + float_to_string_with_full_precision(output->info()->quantization_info().offset));
+ build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(input->info()->quantization_info().scale));
+ build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(output->info()->quantization_info().scale));
+ }
+
// Check output dimensions
auto_init(input->info(), output->info(), pool_info);
ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), pool_info));
@@ -212,8 +222,6 @@
const DataType data_type = input->info()->data_type();
- // Set build options
- CLBuildOptions build_opts;
build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type));
build_opts.add_option("-DPOOL_" + string_from_pooling_type(pool_type));
build_opts.add_option("-DSTRIDE_X=" + support::cpp11::to_string(pool_stride_x));
@@ -222,6 +230,7 @@
build_opts.add_option("-DPAD_Y=" + support::cpp11::to_string(pool_pad_top));
build_opts.add_option("-DPOOL_SIZE_X=" + support::cpp11::to_string(pool_size_x));
build_opts.add_option("-DPOOL_SIZE_Y=" + support::cpp11::to_string(pool_size_y));
+
build_opts.add_option_if(data_type == DataType::F16, "-DFP16");
// Create kernel
diff --git a/src/core/NEON/kernels/NEPoolingLayerKernel.cpp b/src/core/NEON/kernels/NEPoolingLayerKernel.cpp
index d00a4af..308fad5 100644
--- a/src/core/NEON/kernels/NEPoolingLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEPoolingLayerKernel.cpp
@@ -138,7 +138,6 @@
{
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output);
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
ARM_COMPUTE_RETURN_ERROR_ON((output->dimension(get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH)) != pooled_w)
|| (output->dimension(get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT)) != pooled_h));
}
@@ -640,6 +639,15 @@
}
}
+ const QuantizationInfo &input_qinfo = _input->info()->quantization_info();
+ const QuantizationInfo &output_qinfo = _output->info()->quantization_info();
+ if(input_qinfo != output_qinfo)
+ {
+ const auto requantized_output = vquantize(vdequantize(vcombine_u8(lower_res, upper_res), input_qinfo), output_qinfo);
+ lower_res = vget_low_u8(requantized_output);
+ upper_res = vget_high_u8(requantized_output);
+ }
+
// Store result
if(pool_stride_x == 1)
{
@@ -1641,6 +1649,11 @@
}
// Store result
+ const QuantizationInfo &input_qinfo = _input->info()->quantization_info();
+ const QuantizationInfo &output_qinfo = _output->info()->quantization_info();
+ res = (input_qinfo != output_qinfo) ? sqcvt_qasymm8_f32(scvt_f32_qasymm8(res, input_qinfo.scale, input_qinfo.offset), output_qinfo.scale,
+ output_qinfo.offset) :
+ res;
*(reinterpret_cast<uint8_t *>(output.ptr())) = res;
},
input, output);
@@ -1663,7 +1676,9 @@
const int upper_bound_w = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_right);
const int upper_bound_h = _input->info()->dimension(2) + (exclude_padding ? 0 : pool_pad_bottom);
- const float32x4_t half_scale_v = vdupq_n_f32(0.5f);
+ const float32x4_t half_scale_v = vdupq_n_f32(0.5f);
+ const QuantizationInfo &input_qinfo = _input->info()->quantization_info();
+ const QuantizationInfo &output_qinfo = _output->info()->quantization_info();
execute_window_loop(window, [&](const Coordinates & id)
{
@@ -1713,6 +1728,12 @@
uint8x8_t res1 = vmovn_u16(vcombine_u16(vmovn_u32(vres1), vmovn_u32(vres2)));
uint8x8_t res2 = vmovn_u16(vcombine_u16(vmovn_u32(vres3), vmovn_u32(vres4)));
+ if(input_qinfo != output_qinfo)
+ {
+ const auto requantized_output = vquantize(vdequantize(vcombine_u8(res1, res2), input_qinfo), output_qinfo);
+ res1 = vget_low_u8(requantized_output);
+ res2 = vget_high_u8(requantized_output);
+ }
// Store result
vst1_u8(output.ptr(), res1);
@@ -1733,7 +1754,7 @@
}
// Store result
- vst1q_u8(output.ptr(), vres);
+ vst1q_u8(output.ptr(), (input_qinfo != output_qinfo) ? vquantize(vdequantize(vres, input_qinfo), output_qinfo) : vres);
}
},
input, output);