COMPMID-1687: Optimize CLGEMMMatrixMultiplyKernel for Mali-G76 - Part1

The current implementation is limited just to FP32

Change-Id: I185ab57e483e879d7c301e9cc3033efc8b41e244
Reviewed-on: https://review.mlplatform.org/389
Reviewed-by: Anthony Barbier <Anthony.barbier@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>

4 files changed, 343 insertions, 4 deletions

diff --git a/tests/framework/Macros.h b/tests/framework/Macros.h
index deca1ef51a..591b80e9d8 100644
--- a/tests/framework/Macros.h
+++ b/tests/framework/Macros.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -49,8 +49,8 @@
 
 #define CONCAT(ARG0, ARG1) ARG0##ARG1
 
-#define VARIADIC_SIZE_IMPL(e0, e1, e2, e3, e4, e5, e6, e7, e8, e9, size, ...) size
-#define VARIADIC_SIZE(...) VARIADIC_SIZE_IMPL(__VA_ARGS__, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0)
+#define VARIADIC_SIZE_IMPL(e0, e1, e2, e3, e4, e5, e6, e7, e8, e9, e10, size, ...) size
+#define VARIADIC_SIZE(...) VARIADIC_SIZE_IMPL(__VA_ARGS__, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0)
 
 #define JOIN_PARAM1(OP, param) OP(0, param)
 #define JOIN_PARAM2(OP, param, ...) \
@@ -80,6 +80,9 @@
 #define JOIN_PARAM10(OP, param, ...) \
     OP(9, param)                     \
     , JOIN_PARAM9(OP, __VA_ARGS__)
+#define JOIN_PARAM11(OP, param, ...) \
+    OP(10, param)                    \
+    , JOIN_PARAM10(OP, __VA_ARGS__)
 #define JOIN_PARAM(OP, NUM, ...) \
     CONCAT(JOIN_PARAM, NUM)      \
     (OP, __VA_ARGS__)
@@ -264,4 +267,4 @@
 //
 // TEST CASE MACROS END
 //
-#endif /* ARM_COMPUTE_TEST_FRAMEWORK_MACROS */
+#endif /* ARM_COMPUTE_TEST_FRAMEWORK_MACROS */
\ No newline at end of file
diff --git a/tests/validation/CL/GEMMMatrixMultiplyReshaped.cpp b/tests/validation/CL/GEMMMatrixMultiplyReshaped.cpp
new file mode 100644
index 0000000000..e2fa194765
--- /dev/null
+++ b/tests/validation/CL/GEMMMatrixMultiplyReshaped.cpp
@@ -0,0 +1,224 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.h"
+#include "arm_compute/core/CL/kernels/CLGEMMReshapeLHSMatrixKernel.h"
+#include "arm_compute/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "arm_compute/runtime/CL/CLTensor.h"
+#include "arm_compute/runtime/CL/CLTensorAllocator.h"
+#include "tests/CL/CLAccessor.h"
+#include "tests/CL/Helper.h"
+#include "tests/PaddingCalculator.h"
+#include "tests/datasets/ShapeDatasets.h"
+#include "tests/framework/Asserts.h"
+#include "tests/framework/Macros.h"
+#include "tests/framework/datasets/Datasets.h"
+#include "tests/validation/Validation.h"
+#include "tests/validation/fixtures/GEMMFixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace
+{
+// *INDENT-OFF*
+// clang-format off
+RelativeTolerance<float> tolerance_f32(0.001f);
+constexpr float          abs_tolerance_f32(0.0001f);
+
+/** M values to test */
+const auto m_values = framework::dataset::make("M", 37);
+
+/** N values to test */
+const auto n_values = framework::dataset::make("N", 51);
+
+/** K values to test */
+const auto k_values = framework::dataset::make("K", 43);
+
+/** Batch size values to test */
+const auto b_values = framework::dataset::make("batch_size", 1, 3);
+
+/** M0 values to test - Precommit */
+const auto m0_values_precommit = framework::dataset::make("M0", {4, 5, 6});
+
+/** N0 values to test - Precommit */
+const auto n0_values_precommit = framework::dataset::make("N0", { 2, 4, 8 });
+
+/** K0 values to test - Precommit */
+const auto k0_values_precommit = framework::dataset::make("K0", { 4, 8 });
+
+/** V0 values to test - Precommit */
+const auto v0_values_precommit = framework::dataset::make("V0", 1, 3);
+
+/** H0 values to test - Precommit */
+const auto h0_values_precommit = framework::dataset::make("H0", 1, 3);
+
+/** M0 values to test - Nightly */
+const auto m0_values_nightly = framework::dataset::make("M0", 2, 8);
+
+/** N0 values to test - Nightly */
+const auto n0_values_nightly = framework::dataset::make("N0", { 2, 4, 8, 16 });
+
+/** K0 values to test - Nightly */
+const auto k0_values_nightly = framework::dataset::make("K0", { 4, 8, 16 });
+
+/** V0 values to test - Nightly */
+const auto v0_values_nightly = framework::dataset::make("V0", 1, 4);
+
+/** H0 values to test - Nightly */
+const auto h0_values_nightly = framework::dataset::make("H0", 1, 4);
+
+/** Interleave values to test with LHS matrix */
+const auto i_values_lhs = framework::dataset::make("interleave_lhs", { true, false });
+
+/** Interleave values to test with RHS matrix */
+const auto i_values_rhs = framework::dataset::make("interleave_rhs", { true, false });
+
+} // namespace
+
+using namespace arm_compute::misc::shape_calculator;
+
+// Create function for CLGEMMReshapeLHSMatrixKernel
+using CLGEMMReshapeLHSMatrix = CLSynthetizeFunctionInitOutputWithZeroAndWithZeroConstantBorder<CLGEMMReshapeLHSMatrixKernel, 16>;
+
+// Create function for CLGEMMReshapeRHSMatrixKernel
+using CLGEMMReshapeRHSMatrix = CLSynthetizeFunctionInitOutputWithZeroAndWithZeroConstantBorder<CLGEMMReshapeRHSMatrixKernel, 16>;
+
+// Create function for CLGEMMMatrixMultiplyReshapedKernel
+using CLGEMMMatrixMultiplyReshaped = CLSynthetizeFunction<CLGEMMMatrixMultiplyReshapedKernel>;
+
+// Fixture for CLGEMMMatrixMultiplyReshaped
+using CLGEMMMatrixMultiplyReshapedFixture = GEMMMatrixMultiplyReshapedValidationFixture<CLTensor, CLAccessor, CLGEMMReshapeLHSMatrix, CLGEMMReshapeRHSMatrix, CLGEMMMatrixMultiplyReshaped>;
+
+TEST_SUITE(CL)
+TEST_SUITE(GEMMMatrixMultiplyReshaped)
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_values,
+                                                                   n_values),
+                                                                   k_values),
+                                                                   framework::dataset::make("batch_size", 1)),
+                                                                   m0_values_precommit),
+                                                                   n0_values_precommit),
+                                                                   k0_values_precommit),
+                                                                   v0_values_precommit),
+                                                                   h0_values_precommit),
+                                                                   i_values_lhs),
+                                                                   i_values_rhs),
+m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, v0_value, h0_value, i_value_lhs, i_value_rhs)
+{
+    const unsigned int M = m_value;
+    const unsigned int N = n_value;
+    const unsigned int K = k_value;
+
+    GEMMLHSMatrixInfo lhs_info;
+    lhs_info.m0         = m0_value;
+    lhs_info.k0         = k0_value;
+    lhs_info.v0         = v0_value;
+    lhs_info.interleave = i_value_lhs;
+    lhs_info.transpose  = false;
+
+    GEMMRHSMatrixInfo rhs_info;
+    rhs_info.n0         = n0_value;
+    rhs_info.k0         = k0_value;
+    rhs_info.h0         = h0_value;
+    rhs_info.interleave = i_value_rhs;
+    rhs_info.transpose  = true;
+
+    GEMMReshapeInfo gemm_info(M, N, K);
+
+    const TensorShape lhs_shape(K, M, b_value);
+    const TensorShape lhs_shape_reshaped = compute_lhs_reshaped_shape(TensorInfo(lhs_shape, 1, DataType::F32),
+                                                                      lhs_info,
+                                                                      false);
+
+    const TensorShape rhs_shape(N, K, b_value);
+    const TensorShape rhs_shape_reshaped = compute_rhs_reshaped_shape(TensorInfo(rhs_shape, 1, DataType::F32),
+                                                                      rhs_info);
+
+    const TensorShape dst_shape = compute_mm_shape(TensorInfo(lhs_shape_reshaped, 1, DataType::F32),
+                                                   TensorInfo(rhs_shape_reshaped, 1, DataType::F32),
+                                                   gemm_info);
+
+    // Create tensors
+    CLTensor lhs_reshaped = create_tensor<CLTensor>(lhs_shape_reshaped, DataType::F32);
+    CLTensor rhs_reshaped = create_tensor<CLTensor>(rhs_shape_reshaped, DataType::F32);
+    CLTensor dst          = create_tensor<CLTensor>(dst_shape, DataType::F32);
+
+    ARM_COMPUTE_EXPECT(lhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(rhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+    // Create and configure function
+    CLGEMMMatrixMultiplyReshaped gemm;
+    gemm.configure(&lhs_reshaped, &rhs_reshaped, &dst, 1.0f, lhs_info, rhs_info, gemm_info);
+}
+
+TEST_SUITE(Float)
+TEST_SUITE(FP32)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedFixture, framework::DatasetMode::ALL,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_values,
+                                                                   n_values),
+                                                                   k_values),
+                                                                   b_values),
+                                                                   m0_values_precommit),
+                                                                   n0_values_precommit),
+                                                                   k0_values_precommit),
+                                                                   v0_values_precommit),
+                                                                   h0_values_precommit),
+                                                                   i_values_lhs),
+                                                                   i_values_rhs))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyReshapedFixture, framework::DatasetMode::NIGHTLY,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_values,
+                                                                   n_values),
+                                                                   k_values),
+                                                                   b_values),
+                                                                   m0_values_nightly),
+                                                                   n0_values_nightly),
+                                                                   k0_values_nightly),
+                                                                   v0_values_nightly),
+                                                                   h0_values_nightly),
+                                                                   i_values_lhs),
+                                                                   i_values_rhs))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32);
+}
+TEST_SUITE_END() // Float
+TEST_SUITE_END() // FP32
+TEST_SUITE_END() // GEMMMatrixMulipltyReshaped
+TEST_SUITE_END() // CL
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
\ No newline at end of file
diff --git a/tests/validation/CL/Im2Col.cpp b/tests/validation/CL/Im2Col.cpp
index ebf2331e5e..432b3b239a 100644
--- a/tests/validation/CL/Im2Col.cpp
+++ b/tests/validation/CL/Im2Col.cpp
@@ -49,6 +49,7 @@ const auto conv_filter_sizes = framework::dataset::make("KernelDims", { Size2D(3
                                                                         Size2D(1U, 3U),
                                                                         Size2D(5U, 3U),
                                                                         Size2D(1U, 1U),
+                                                                        Size2D(9U, 9U),
                                                                         Size2D(11U, 11U)} );
 const auto padstrides        = framework::dataset::make("PadStride", { PadStrideInfo(1U, 1U, 0U, 0U),
                                                                        PadStrideInfo(1U, 1U, 1U, 1U),
diff --git a/tests/validation/fixtures/GEMMFixture.h b/tests/validation/fixtures/GEMMFixture.h
index 0083abffb5..ce2b177ce9 100644
--- a/tests/validation/fixtures/GEMMFixture.h
+++ b/tests/validation/fixtures/GEMMFixture.h
@@ -151,6 +151,117 @@ protected:
     SimpleTensor<T> _reference{};
 };
 
+template <typename TensorType, typename AccessorType, typename ReshapeLHSFunctionType, typename ReshapeRHSFunctionType, typename GEMMFunctionType>
+class GEMMMatrixMultiplyReshapedValidationFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int v0, unsigned int h0, bool interleave_lhs,
+               bool interleave_rhs)
+    {
+        GEMMLHSMatrixInfo lhs_info;
+        lhs_info.m0         = m0;
+        lhs_info.k0         = k0;
+        lhs_info.v0         = v0;
+        lhs_info.interleave = interleave_lhs;
+        lhs_info.transpose  = false;
+
+        GEMMRHSMatrixInfo rhs_info;
+        rhs_info.n0         = n0;
+        rhs_info.k0         = k0;
+        rhs_info.h0         = h0;
+        rhs_info.interleave = interleave_rhs;
+        rhs_info.transpose  = true;
+
+        // Set the tensor shapes for LHS and RHS matrices
+        const TensorShape lhs_shape(k, m, batch_size);
+        const TensorShape rhs_shape(n, k, batch_size);
+
+        _target    = compute_target(lhs_shape, rhs_shape, lhs_info, rhs_info);
+        _reference = compute_reference(lhs_shape, rhs_shape);
+    }
+
+protected:
+    template <typename U>
+    void fill(U &&tensor, int i)
+    {
+        std::uniform_real_distribution<> distribution(-1.0f, 1.0f);
+        library->fill(tensor, distribution, i);
+    }
+
+    TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info)
+    {
+        // Create tensors
+        TensorType lhs = create_tensor<TensorType>(lhs_shape, DataType::F32, 1);
+        TensorType rhs = create_tensor<TensorType>(rhs_shape, DataType::F32, 1);
+        TensorType lhs_reshaped;
+        TensorType rhs_reshaped;
+        TensorType dst;
+
+        const unsigned int M = lhs_shape[1];
+        const unsigned int N = rhs_shape[0];
+        const unsigned int K = lhs_shape[0];
+
+        // The output tensor will be auto-initialized within the function
+
+        // Create and configure function
+        ReshapeLHSFunctionType reshape_lhs;
+        ReshapeRHSFunctionType reshape_rhs;
+        GEMMFunctionType       gemm;
+        reshape_lhs.configure(&lhs, &lhs_reshaped, lhs_info);
+        reshape_rhs.configure(&rhs, &rhs_reshaped, rhs_info);
+        gemm.configure(&lhs_reshaped, &rhs_reshaped, &dst, 1.0f, lhs_info, rhs_info, GEMMReshapeInfo(M, N, K));
+
+        ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Allocate tensors
+        lhs.allocator()->allocate();
+        rhs.allocator()->allocate();
+        lhs_reshaped.allocator()->allocate();
+        rhs_reshaped.allocator()->allocate();
+        dst.allocator()->allocate();
+
+        ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!lhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!rhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Fill tensors
+        fill(AccessorType(lhs), 0);
+        fill(AccessorType(rhs), 1);
+
+        // Compute GEMM
+        reshape_lhs.run();
+        reshape_rhs.run();
+        gemm.run();
+
+        return dst;
+    }
+
+    SimpleTensor<float> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape)
+    {
+        TensorShape dst_shape = lhs_shape;
+        dst_shape[0]          = rhs_shape[0];
+        dst_shape[1]          = lhs_shape[1];
+
+        // Create reference
+        SimpleTensor<float> lhs{ lhs_shape, DataType::F32, 1 };
+        SimpleTensor<float> rhs{ rhs_shape, DataType::F32, 1 };
+        SimpleTensor<float> c{ dst_shape, DataType::F32, 1 };
+
+        // Fill reference
+        fill(lhs, 0);
+        fill(rhs, 1);
+        fill(c, 2);
+
+        return reference::gemm<float>(lhs, rhs, c, 1.0f, 0.0f);
+    }
+
+    TensorType          _target{};
+    SimpleTensor<float> _reference{};
+};
 } // namespace validation
 } // namespace test
 } // namespace arm_compute