COMPMID-1266 : Add support for FP16 in CLWinogradConvolutionLayer: 5x5 kernels

Introduced F32 accumulation for F16 winograd gemm and output transform
WinogradConvolution will be available for F16 only if fast math flag is enabled

Change-Id: I215593c205236a0f9669218437bb40b184ec6a4f

2 files changed, 19 insertions, 5 deletions

diff --git a/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h b/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h
index e030fa2d2a..f61c330de6 100644
--- a/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h
+++ b/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h
@@ -59,9 +59,11 @@ public:
      * @param[in]  alpha                     Weight of the matrix product
      * @param[in]  is_interleaved_transposed (Optional) True if input0 and input1 have been reshaped respectively using @ref CLGEMMInterleave4x4Kernel and @ref CLGEMMTranspose1xWKernel
      * @param[in]  reshape_info              (Optional) GEMM reshape info. If is_interleaved_transposed = true, this object must contain the information to understand how the matrix A and matrix B have been reshaped
+     * @param[in]  fp_mixed_precision        (Optional) Use wider accumulators (32 bit instead of 16 for FP16) to improve accuracy
      *
      */
-    void configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, float alpha, bool is_interleaved_transposed = true, const GEMMReshapeInfo &reshape_info = GEMMReshapeInfo());
+    void configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, float alpha, bool is_interleaved_transposed = true, const GEMMReshapeInfo &reshape_info = GEMMReshapeInfo(),
+                   bool fp_mixed_precision = false);
     /** Static function to check if given info will lead to a valid configuration of @ref CLGEMMMatrixMultiplyKernel
      *
      * @param[in] input0                    Input tensor containing the Matrix A. Data types supported: F16/F32
@@ -71,11 +73,12 @@ public:
      * @param[in] is_interleaved_transposed True if input0 and input1 have been reshaped respectively using @ref CLGEMMInterleave4x4Kernel and @ref CLGEMMTranspose1xWKernel
      * @param[in] reshape_info              GEMM reshape info. If is_interleaved_transposed = true, this object must contain the information to understand how the matrix A and matrix B have been reshaped
      * @param[in] gpu_target                GPU Target
+     * @param[in] fp_mixed_precision        (Optional) Use wider accumulators (32 bit instead of 16 for FP16) to improve accuracy
      *
      * @return a status
      */
     static Status validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, float alpha, bool is_interleaved_transposed, const GEMMReshapeInfo &reshape_info,
-                           GPUTarget gpu_target);
+                           GPUTarget gpu_target, bool fp_mixed_precision = false);
 
     // Inherited methods overridden:
     void run(const Window &window, cl::CommandQueue &queue) override;
diff --git a/arm_compute/core/Types.h b/arm_compute/core/Types.h
index fb277584fd..4eb8129b62 100644
--- a/arm_compute/core/Types.h
+++ b/arm_compute/core/Types.h
@@ -1593,7 +1593,8 @@ class GEMMInfo
 public:
     /** Default constructor */
     GEMMInfo()
-        : _is_a_reshaped(false), _is_b_reshaped(false), _reshape_b_only_on_first_run(false), _depth_output_gemm3d(0), _reinterpret_input_as_3d(false), _retain_internal_weights(false), _gemmlowp_output_stage()
+        : _is_a_reshaped(false), _is_b_reshaped(false), _reshape_b_only_on_first_run(false), _depth_output_gemm3d(0), _reinterpret_input_as_3d(false), _retain_internal_weights(false),
+          _gemmlowp_output_stage(), _fp_mixed_precision(false)
     {
     }
     /** Constructor
@@ -1607,12 +1608,13 @@ public:
      *                                        to perform 1x1 convolutions with the NHWC data layout)
      * @param[in] retain_internal_weights     (Optional) Retain the weights tensor from previous run
      * @param[in] gemmlowp_output_stage       (Optional) GEMMLowp Output stage info
+     * @param[in] fp_mixed_precision          (Optional) Use wider accumulators (32 bit instead of 16 for FP16) to improve accuracy.
      *
      */
     GEMMInfo(bool is_a_reshaped, bool is_b_reshaped, bool reshape_b_only_on_first_run, int depth_output_gemm3d = 0, bool reinterpret_input_as_3d = false, bool retain_internal_weights = false,
-             GEMMLowpOutputStageInfo gemmlowp_output_stage = GEMMLowpOutputStageInfo())
+             GEMMLowpOutputStageInfo gemmlowp_output_stage = GEMMLowpOutputStageInfo(), bool fp_mixed_precision = false)
         : _is_a_reshaped(is_a_reshaped), _is_b_reshaped(is_b_reshaped), _reshape_b_only_on_first_run(reshape_b_only_on_first_run), _depth_output_gemm3d(depth_output_gemm3d),
-          _reinterpret_input_as_3d(reinterpret_input_as_3d), _retain_internal_weights(retain_internal_weights), _gemmlowp_output_stage(gemmlowp_output_stage)
+          _reinterpret_input_as_3d(reinterpret_input_as_3d), _retain_internal_weights(retain_internal_weights), _gemmlowp_output_stage(gemmlowp_output_stage), _fp_mixed_precision(fp_mixed_precision)
     {
     }
     /** Flag which specifies if the matrix A has been reshaped
@@ -1673,6 +1675,14 @@ public:
     {
         return _gemmlowp_output_stage;
     };
+    /** Flag which specifies if a wider accumulator should be used.
+     *
+     * @return True if a wider accumulator has to be used
+     */
+    bool fp_mixed_precision() const
+    {
+        return _fp_mixed_precision;
+    };
 
 private:
     const bool                    _is_a_reshaped;
@@ -1682,6 +1692,7 @@ private:
     const bool                    _reinterpret_input_as_3d;
     const bool                    _retain_internal_weights;
     const GEMMLowpOutputStageInfo _gemmlowp_output_stage;
+    const bool                    _fp_mixed_precision;
 };
 
 /** Winograd information */