COMPMID-661: Convolution quantized (#32)
Change-Id: Id69df4ce98d1d89bdf9c9aa5c4d909659909b30f
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/110456
Tested-by: BSG Visual Compute Jenkins server to access repositories on http://mpd-gerrit.cambridge.arm.com <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
diff --git a/arm_compute/core/CL/kernels/CLWeightsReshapeKernel.h b/arm_compute/core/CL/kernels/CLWeightsReshapeKernel.h
index ec8940e..6c84ded 100644
--- a/arm_compute/core/CL/kernels/CLWeightsReshapeKernel.h
+++ b/arm_compute/core/CL/kernels/CLWeightsReshapeKernel.h
@@ -50,6 +50,7 @@
* and 5D tensor with dimensions [kernel_x, kernel_y, IFM, OFM, num_patches] if unshared. Data types supported: QS8/QS16/QASYMM8/F16/F32
* @param[in] biases The shared biases tensor to append. Bias is 1D tensor with dimensions [OFM] if shared and 2D tensor with
* dimensions [OFM, num_patches] if unshared. Data types supported: Same as @p input
+ * @warning Appending biases to weights reshaped matrix is not supported for quantized asymmetric types.
* @param[out] output The output tensor. Should be a 2D Tensor. Data types supported: Same as @p input
*/
void configure(const ICLTensor *input, const ICLTensor *biases, ICLTensor *output);
diff --git a/arm_compute/core/Types.h b/arm_compute/core/Types.h
index c77f1d4..beaec14 100644
--- a/arm_compute/core/Types.h
+++ b/arm_compute/core/Types.h
@@ -827,6 +827,58 @@
const unsigned int _num_kernels;
};
+/** GEMM Information class. This class stores the necessary information to compute GEMM functions */
+class GEMMInfo
+{
+public:
+ /** Default constructor */
+ GEMMInfo()
+ : _is_a_reshaped(false), _is_b_reshaped(false), _reshape_b_only_on_first_run(false)
+ {
+ }
+ /** Constructor
+ *
+ * @param[in] is_a_reshaped True if the matrix A has been reshaped
+ * @param[in] is_b_reshaped True if the matrix B has been reshaped
+ * @param[in] reshape_b_only_on_first_run Reshape matrix B only for the first run
+ */
+ GEMMInfo(bool is_a_reshaped, bool is_b_reshaped, bool reshape_b_only_on_first_run)
+ : _is_a_reshaped(is_a_reshaped), _is_b_reshaped(is_b_reshaped), _reshape_b_only_on_first_run(reshape_b_only_on_first_run)
+ {
+ }
+ /** Flag which specifies if the matrix A has been reshaped
+ *
+ * @return True if the matrix A has been reshaped
+ */
+ bool is_a_reshaped() const
+ {
+ return _is_a_reshaped;
+ };
+ /** Flag which specifies if the matrix B has been reshaped
+ *
+ * @return True if the matrix B has been reshaped
+ */
+ bool is_b_reshaped() const
+ {
+ return _is_b_reshaped;
+ };
+ /** Flag which specifies if the reshape of matrix B should executed only for the first
+ *
+ * @note This flag could be set to TRUE when GEMM is used to accelerate convolution layer
+ *
+ * @return True if the reshaped of matrix B happens only for the first run
+ */
+ bool reshape_b_only_on_first_run() const
+ {
+ return _reshape_b_only_on_first_run;
+ };
+
+private:
+ const bool _is_a_reshaped;
+ const bool _is_b_reshaped;
+ const bool _reshape_b_only_on_first_run;
+};
+
/** IO formatting information class*/
struct IOFormatInfo
{
diff --git a/arm_compute/runtime/CL/functions/CLConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLConvolutionLayer.h
index cd1ea70..74c1244 100644
--- a/arm_compute/runtime/CL/functions/CLConvolutionLayer.h
+++ b/arm_compute/runtime/CL/functions/CLConvolutionLayer.h
@@ -36,6 +36,8 @@
#include "arm_compute/core/Types.h"
#include "arm_compute/runtime/CL/CLMemoryGroup.h"
#include "arm_compute/runtime/CL/CLTensor.h"
+#include "arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h"
+#include "arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h"
#include "arm_compute/runtime/IMemoryManager.h"
#include <memory>
@@ -55,7 +57,8 @@
CLConvolutionLayerReshapeWeights(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
/** Set the input and output tensors.
*
- * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: QS8/QS16/F16/F32.
+ * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
+ * Data type supported: QS8/QASYMM8/QS16/F16/F32.
* @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p weights.
* @param[out] output Destination tensor. Data types supported: Same as @p weights.
* @param[in] transpose1xW True if the weights are to undergo a 1xW transposition after reshaping (in case of GEMM operation), false otherwise.
@@ -79,7 +82,8 @@
* -# @ref CLGEMMTranspose1xWKernel (executed only once for each configuration)
* -# @ref CLIm2ColKernel
* -# @ref CLGEMMInterleave4x4Kernel
- * -# @ref CLGEMMMatrixMultiplyKernel
+ * -# @ref CLGEMMMatrixMultiplyKernel or @ref CLGEMMLowpMatrixMultiplyCore (if quantized asymmetric)
+ * -# @ref CLGEMMLowpQuantizeDownInt32ToUint8Scale (if quantized asymmetric)
* -# @ref CLCol2ImKernel
*/
class CLConvolutionLayer : public IFunction
@@ -91,9 +95,10 @@
*
* @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
* while every optional dimension from 4 and above represent a batch of inputs.
- * Data types supported: QS8/QS16/F16/F32.
- * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input.
- * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported:Same as @p input.
+ * Data types supported: QS8/QASYMM8/QS16/F16/F32.
+ * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: Same as @p input.
+ * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
+ * Data type supported: Should match @p input data type, except for input of QASYMM8 type where biases should be of S32 type.
* @param[out] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
* Data types supported: Same as @p input.
* @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo.
@@ -106,20 +111,37 @@
void run() override;
private:
- CLMemoryGroup _memory_group;
- CLConvolutionLayerReshapeWeights _reshape_weights;
- CLIm2ColKernel _input_im2col_kernel;
- CLGEMMInterleave4x4Kernel _input_interleave_kernel;
- CLGEMMMatrixMultiplyKernel _mm_kernel;
- CLCol2ImKernel _output_col2im_kernel;
- CLTensor _input_im2col_reshaped;
- CLTensor _input_interleaved_reshaped;
- CLTensor _weights_reshaped;
- CLTensor _weights_transposed;
- CLTensor _gemm_output;
- bool _has_bias;
- bool _is_fully_connected_convolution;
- bool _are_weights_reshaped;
+ /** Configures the appropriate matrix multiply routine
+ *
+ * @param input Input tensor. Data types supported: QS8/QASYMM8/QS16/F16/F32.
+ * @param weights Weights tensor. Data type supported: Same as @p input.
+ * @param output Output tensor. Data types supported: Same as @p input,
+ * except for input of QASYMM8 type where output should be of S32 type.
+ * @param is_interleaved_transposed Flag that signals if matrix is interleaved transposed
+ */
+ void configure_mm(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output, bool is_interleaved_transposed = true);
+
+private:
+ CLMemoryGroup _memory_group;
+ CLConvolutionLayerReshapeWeights _reshape_weights;
+ CLIm2ColKernel _input_im2col_kernel;
+ CLGEMMInterleave4x4Kernel _input_interleave_kernel;
+ CLGEMMMatrixMultiplyKernel _mm_kernel;
+ CLGEMMLowpMatrixMultiplyCore _mm_gemmlowp;
+ CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint _gemmlowp_output_stage;
+ CLCol2ImKernel _output_col2im_kernel;
+
+ CLTensor _input_im2col_reshaped;
+ CLTensor _input_interleaved_reshaped;
+ CLTensor _weights_reshaped;
+ CLTensor _weights_transposed;
+ CLTensor _gemm_output;
+ CLTensor _tmp_output;
+
+ bool _append_bias;
+ bool _is_fully_connected_convolution;
+ bool _are_weights_reshaped;
+ bool _is_quantized;
};
}
#endif /* __ARM_COMPUTE_CLCONVOLUTIONLAYER_H__ */
diff --git a/arm_compute/runtime/CL/functions/CLGEMMInterleave4x4.h b/arm_compute/runtime/CL/functions/CLGEMMInterleave4x4.h
index 04f55c1..ae05b0f 100644
--- a/arm_compute/runtime/CL/functions/CLGEMMInterleave4x4.h
+++ b/arm_compute/runtime/CL/functions/CLGEMMInterleave4x4.h
@@ -40,11 +40,11 @@
public:
/** Initialise the kernel's inputs, output
*
- * @param[in] input First input tensor. Data types supported: U8/S8/QS8/QASYMM8/U16/S16/F16/U32/S32/F32
+ * @param[in] input First input tensor. Data types supported: U8/S8/QS8/QASYMM8/QS16/U16/S16/F16/U32/S32/F32
* @param[out] output Output tensor. Data type supported: same as @p input
*/
void configure(const ICLTensor *input, ICLTensor *output);
};
}
-#endif /* __ARM_COMPUTE_CLGEMMINTERLEAVE4X4_H__ */
\ No newline at end of file
+#endif /* __ARM_COMPUTE_CLGEMMINTERLEAVE4X4_H__ */
diff --git a/arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h b/arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h
index 9944afe..e316144 100644
--- a/arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h
+++ b/arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h
@@ -62,11 +62,13 @@
* -# Convert b values from QASYMM8 to int32 add b_offset to each of them.
* -# Compute the matrix product of the resulting a * b in int32.
*
- * @param[in] a First input tensor (Matrix A). Data type supported: QASYMM8.
- * @param[in] b Second input tensor (Matrix B). Data type supported: same as @p a
- * @param[out] output Output tensor. Data type supported: Data type supported: S32
+ * @param[in] a First input tensor (Matrix A). Data type supported: QASYMM8.
+ * @param[in] b Second input tensor (Matrix B). Data type supported: same as @p a
+ * @param[out] output Output tensor. Data type supported: Data type supported: S32
+ * @param[in] gemm_info (Optional) Specifies if the matrix A and/or matrix B have been reshaped and
+ * if the reshape of matrix B should be executed only for the first run
*/
- void configure(const ICLTensor *a, const ICLTensor *b, ICLTensor *output);
+ void configure(const ICLTensor *a, const ICLTensor *b, ICLTensor *output, const GEMMInfo &gemm_info = GEMMInfo());
// Inherited methods overridden:
void run() override;
@@ -86,6 +88,8 @@
int32_t _a_offset;
int32_t _b_offset;
bool _is_interleaved_transposed;
+ bool _is_first_run;
+ bool _reshape_b_only_on_first_run;
};
}
#endif /*__ARM_COMPUTE_CLGEMMLOWPMATRIXMULTIPLYCORE_H__ */
diff --git a/arm_compute/runtime/CL/functions/CLGEMMTranspose1xW.h b/arm_compute/runtime/CL/functions/CLGEMMTranspose1xW.h
index 3d02aa9..ae56548 100644
--- a/arm_compute/runtime/CL/functions/CLGEMMTranspose1xW.h
+++ b/arm_compute/runtime/CL/functions/CLGEMMTranspose1xW.h
@@ -38,7 +38,7 @@
public:
/** Initialise the kernel's inputs, output
*
- * @param[in] input First input tensor. Data type supported: U8/S8/QS8/QASYMM8/U16/S16/F16/U32/S32/F32/
+ * @param[in] input First input tensor. Data type supported: U8/S8/QS8/QASYMM8/QS16/U16/S16/F16/U32/S32/F32
* @param[out] output Output tensor. Data type supported: same as @p input
*/
void configure(const ICLTensor *input, ICLTensor *output);
diff --git a/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h b/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h
index cc513ad..a4f3511 100644
--- a/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h
+++ b/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h
@@ -66,20 +66,24 @@
* -# Convert b values from QASYMM8 to int32 add b_offset to each of them.
* -# Compute the matrix product of the resulting a * b in int32.
*
- * @param[in] a First input tensor (Matrix A). Data type supported: QASYMM8.
- * @param[in] b Second input tensor (Matrix B). Data type supported: same as @p a
- * @param[out] output Output tensor. Data type supported: Data type supported: S32
+ * @param[in] a First input tensor (Matrix A). Data type supported: QASYMM8.
+ * @param[in] b Second input tensor (Matrix B). Data type supported: same as @p a
+ * @param[out] output Output tensor. Data type supported: Data type supported: S32
+ * @param[in] gemm_info (Optional) Specifies if the matrix A and/or matrix B have been reshaped and
+ * if the reshape of matrix B should be executed only for the first run
*/
- void configure(const ITensor *a, const ITensor *b, ITensor *output);
+ void configure(const ITensor *a, const ITensor *b, ITensor *output, const GEMMInfo &gemm_info = GEMMInfo());
/** Static function to check if given info will lead to a valid configuration of @ref NEGEMMLowpMatrixMultiplyCore
*
- * @param[in] a First input tensor (Matrix A). Data type supported: QASYMM8.
- * @param[in] b Second input tensor (Matrix B). Data type supported: same as @p a
- * @param[out] output Output tensor. Data type supported: Data type supported: S32
+ * @param[in] a First input tensor (Matrix A). Data type supported: QASYMM8.
+ * @param[in] b Second input tensor (Matrix B). Data type supported: same as @p a
+ * @param[out] output Output tensor. Data type supported: Data type supported: S32
+ * @param[in] gemm_info (Optional) Specifies if the matrix A and/or matrix B have been reshaped and
+ * if the reshape of matrix B should be executed only for the first run
*
* @return an error status
*/
- static Error validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *output);
+ static Error validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *output, const GEMMInfo &gemm_info = GEMMInfo());
// Inherited methods overridden:
void run() override;