commit 553f6953fe3bdfad53c11c25f305a16d79d83b24
author Francesco Petrogalli <francesco.petrogalli@arm.com> Thu Jun 30 10:22:01 2022 +0000
committer Francesco Petrogalli <francesco.petrogalli@arm.com> Tue Jul 19 09:26:27 2022 +0000
tree 73642b948b79662096f593458c6138d2f7f48ec6
parent 99c46475daf277aa53e6747f9e41209f418fed33

[ONCPUML-951] Variable weight support for Convolution.

API changes for NEGEMMConvolutionLayer and CpuGemmConv2d

Built with:

    scons neon=1 opencl=0 os=linux arch=armv8.2-a multi_isa=1 \
        build=native -j32 Werror=false validation_tests=1 build_dir=opt \
        standalone=1 asserts=1 experimental_fixed_format_kernels=1 .

Tested with:

    ./build/opt/tests/arm_compute_validation

Hardware where the test executable was run:

Neoverse N1

Test coverage:

* NEGEMMConvolutionLayer, CpuGemmConv2d
* NHWC (the only one supported by the fixed-format kernels)
* F16, F32
* Shapes: RunSmall

Change-Id: I4fd3e495a7cbf61210ea02d37440ba9652934e99
Signed-off-by: Francesco Petrogalli <francesco.petrogalli@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7632
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

src/cpu/operators/internal/CpuGemmAssemblyDispatch.h

diff --git a/src/cpu/operators/internal/CpuGemmAssemblyDispatch.h b/src/cpu/operators/internal/CpuGemmAssemblyDispatch.h
index 3c25866..4ef108d 100644
--- a/src/cpu/operators/internal/CpuGemmAssemblyDispatch.h
+++ b/src/cpu/operators/internal/CpuGemmAssemblyDispatch.h
@@ -53,6 +53,7 @@
     float                   padding_value{ 0.f };
     bool                    fast_mode{ false };
     bool                    fixed_format{ false };
+    arm_gemm::WeightFormat  weight_format{ arm_gemm::WeightFormat::UNSPECIFIED };
 };
 
 /** Assembly kernel glue */
@@ -73,6 +74,7 @@
         virtual void                             prepare(ITensorPack &tensors) = 0;
         virtual experimental::MemoryRequirements workspace() const             = 0;
         virtual bool                             is_configured() const         = 0;
+        virtual bool                             isVarWeightsKernel() const    = 0;
         virtual ~IFallback()                                                   = default;
     };
 
@@ -101,15 +103,14 @@
 
     /** Indicates whether or not there is an optimal assembly implementation that can be used to process the given parameters.
      *
-     * @param[in] a    Input tensor info (Matrix A)
-     * @param[in] b    Input tensor info (Matrix B)
-     * @param[in] c    Input tensor info (Matrix C) used to pass the bias for quantized calculations
-     * @param[in] d    Output tensor to store the result of matrix multiplication. Data type supported: same as @p input0.
-     * @param[in] info GEMM meta-data
+     * This method has the same use of @ref
+     * NEGEMMConvolutionLayer::has_opt_impl, with the only caveat that
+     * the value of arm_gemm::WeightFormat need to be passed via the
+     * parameter info.
      *
      * @return a status.
      */
-    static Status has_opt_impl(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *d, const AsmGemmInfo &info);
+    static Status has_opt_impl(arm_gemm::WeightFormat &weight_format, const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *d, const AsmGemmInfo &info);
     /** Checks if activation is supported by the gemm assembly dispatcher
      *
      * @param[in] activation Activation to check
@@ -122,6 +123,14 @@
      * @return True if the function is configured and ready to run
      */
     bool is_configured() const;
+    /** Indicates if the convolution executes in variable weights mode.
+     *
+     * Similar to @ref CpuGemm::isVarWeightsKernel
+     */
+    bool isVarWeightsKernel() const
+    {
+        return _arm_gemm && _arm_gemm->isVarWeightsKernel();
+    }
 
     // Inherited methods overridden:
     void                             prepare(ITensorPack &tensors) override;