COMPMID-665 - NEON: Add QASYMM8 in place Activation layer
- Added min and max arguments for QuantizeDownInt32ToUint8Scale in order
to apply bounded relu
- Added support for int32_t biases
- Extended tests
Change-Id: I015dae17faa7284766b5435ca33bcf593c1b2b69
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/96512
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
diff --git a/arm_compute/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.h b/arm_compute/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.h
index 04b8433..8c1bae9 100644
--- a/arm_compute/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.h
+++ b/arm_compute/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.h
@@ -35,6 +35,13 @@
* This kernel takes a final int32 accumulator value (the output of @NEGEMMLowpMatrixMultiplyKernel),
* and adds to it the offset contribution of matrix A and matrix B in-place.
*
+ * The final result is:
+ *
+ * mm_result[i][k] = mm_result[i][k] +
+ * (vector_sum_col[k] * a_offset) +
+ * (vector_sum_row[i] * b_offset) +
+ * (a_offset * b_offset * k)
+ *
*/
class NEGEMMLowpOffsetContributionKernel : public INEKernel
{
diff --git a/arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.h b/arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.h
index 65f1042..4ec0e9d 100644
--- a/arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.h
+++ b/arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.h
@@ -36,7 +36,10 @@
* The following computations will be performed by the kernel:
*
* -# Add offset terms to final result
- * -# Multiply each entry of result and round to nearest integer
+ * -# Multiply each entry of result by result_mult_int
+ * -# Add bias to final result if bias tensor is not a nullptr
+ * -# Shift the int32 accumulator by result_shift
+ * -# Clamp the value between the specified min and max bounds
* -# Clamp the resulting int32 values to the [0..255] range and cast to QASYMM8.
*
*/
@@ -56,22 +59,44 @@
/** Initialise the kernel's input and output.
*
* @param[in] input Input tensor. Data type supported: S32
+ * @param[in] bias Biases tensor. Only shared biases supported and it can be a nullptr if the biases addition is not required.
+ * Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
* @param[out] output Output tensor. Data type supported: Data type supported: QASYMM8
* @param[in] result_offset Offset to be added to each element of the input matrix
* @param[in] result_mult_int Value to be multiplied to each element of the input matrix when once the result_offset has been add
* @param[in] result_shift Number of bits to shift right the result before converting back to QASYMM8
+ * @param[in] min (Optional) Min value used to saturate down the output result before converting back to QASYMM8
+ * @param[in] max (Optional) Max value used to saturate up the output result before converting back to QASYMM8,
+ * Along with @p min, this value can be used to implement "rectified linear unit" activation functions
*/
- void configure(const ITensor *input, ITensor *output, int result_offset, int result_mult_int, int result_shift);
+ void configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_offset, int result_mult_int, int result_shift, int min = 0, int max = 0);
// Inherited methods overridden:
void run(const Window &window, const ThreadInfo &info) override;
private:
- const ITensor *_input;
- ITensor *_output;
- int32_t _result_offset;
- int32_t _result_mult_int;
- int32_t _result_shift;
+ /** Template function to run the NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel
+ *
+ * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()).
+ */
+ template <bool is_bounded_relu>
+ void run(const Window &window);
+
+ /** Common signature for all the specialised NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel functions
+ *
+ * @param[in] window Region on which to execute the kernel.
+ */
+ using QuantizeDownFunctionPtr = void (NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::*)(const Window &window);
+
+ QuantizeDownFunctionPtr _func;
+ const ITensor *_input;
+ const ITensor *_bias;
+ ITensor *_output;
+ int _result_offset;
+ int _result_mult_int;
+ int _result_shift;
+ int _min;
+ int _max;
};
} // namespace arm_compute
diff --git a/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h b/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h
index 8557ef4..a3db23a 100644
--- a/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h
+++ b/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h
@@ -43,14 +43,18 @@
* NEGEMMLowpQuantizeDownInt32ToUint8Scale depends on 3 parameters: result_offset, result_mult_int, result_shift
* The final result is:
*
- * ((input[i][k] + result_offset) * result_mult_int + rounding) >> result_shift
+ * ((input[i][k] + result_offset) * result_mult_int) >> result_shift
*
- * where rounding = (result_shift < 1) ? 0 : (1 << (result_shift - 1))
+ * In case the bias tensor is provided, the final result is:
+ *
+ * ((input[i][k] + result_offset) * result_mult_int + bias[k]) >> result_shift
*
* This function calls the following NEON kernels:
*
* -# @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel
*
+ * @note The function accepts also 2 optional input arguments (min and max) which can be used to implement "rectified linear unit" activation functions
+ * before the result is shifted right by result_shift
*/
class NEGEMMLowpQuantizeDownInt32ToUint8Scale : public INESimpleFunction
{
@@ -58,12 +62,17 @@
/** Initialise the kernel's inputs, output
*
* @param[in] input Input tensor. It is the output of @ref NEGEMMLowpMatrixMultiplyCore function. Data type supported: S32
+ * @param[in] bias Biases tensor. Only shared biases supported and it can be a nullptr if the addition of biases is not required.
+ * Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
* @param[out] output Output tensor. Data type supported: Data type supported: QASYMM8
* @param[in] result_offset Offset to be added to each element of the input matrix
* @param[in] result_mult_int Value to be multiplied to each element of the input matrix when once the result_offset has been add
* @param[in] result_shift Number of bits to shift right the result before converting back to QASYMM8
+ * @param[in] min (Optional) Min value used to saturate down the output result before converting back to QASYMM8
+ * @param[in] max (Optional) Max value used to saturate up the output result before converting back to QASYMM8,
+ * Along with @p min, this value can be used to implement "rectified linear unit" activation functions
*/
- void configure(const ITensor *input, ITensor *output, int result_offset, int result_mult_int, int result_shift);
+ void configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_offset, int result_mult_int, int result_shift, int min = 0, int max = 0);
};
}
#endif /*__ARM_COMPUTE_NEGEMMLOWPOUTPUTSTAGE_H__ */
\ No newline at end of file