COMPMID-1413 - Improve the performance of GEMMLowp with 8 bit dot product on OpenCL

COMPMID-1424 - Add dot product support for CLDepthwise QASYMM8 3x3 NHWC non-unit stride

With this patch we are able to improve the performance of MobileNet v1-qasymm8 by 37 %
Tried to use the dot product instruction in CLDepthwise QASYMM8 3x3 NHWC non-unit stride
but I have not seen any benefit (maybe because we have few arithemtic operation and we
do not have more load instructions). However Depthwise convolution has been improved by
30%

Change-Id: Id768a99c2e53a04276707e427af5d0ec93419ada
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/155082
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

5 files changed, 10 insertions, 7 deletions

diff --git a/tests/benchmark/fixtures/GEMMLowpFixture.h b/tests/benchmark/fixtures/GEMMLowpFixture.h
index 46a2f5cc6a..33c6415d20 100644
--- a/tests/benchmark/fixtures/GEMMLowpFixture.h
+++ b/tests/benchmark/fixtures/GEMMLowpFixture.h
@@ -58,7 +58,7 @@ public:
         c = create_tensor<TensorType>(shape_dst, DataType::S32, 1, QuantizationInfo(1.0f / 255.0f, 0));
 
         // Create and configure function
-        gemmlowp.configure(&a, &b, &c);
+        gemmlowp.configure(&a, &b, nullptr, &c);
 
         // Allocate tensors
         a.allocator()->allocate();
diff --git a/tests/validate_examples/cl_gemm.cpp b/tests/validate_examples/cl_gemm.cpp
index cdaa33f31a..8b3a103db7 100644
--- a/tests/validate_examples/cl_gemm.cpp
+++ b/tests/validate_examples/cl_gemm.cpp
@@ -193,7 +193,7 @@ public:
             init_sgemm_output(tmp_dst, src0, src1, DataType::S32);
 
             // Configure GEMMlowp matrix multiply function
-            mm_gemmlowp.configure(&src0, &src1, &tmp_dst);
+            mm_gemmlowp.configure(&src0, &src1, nullptr, &tmp_dst);
 
             // Configure GEMMlowp output stage
             mm_gemmlowp_output_stage.configure(&tmp_dst, add_bias ? &biases : nullptr, &dst, dst_multiplier, dst_shift, offset_dst);
diff --git a/tests/validation/CL/GEMMLowp.cpp b/tests/validation/CL/GEMMLowp.cpp
index 42bb2123bf..f0f768dd1b 100644
--- a/tests/validation/CL/GEMMLowp.cpp
+++ b/tests/validation/CL/GEMMLowp.cpp
@@ -67,7 +67,8 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, framework::dataset::c
 
     // Create and configure function
     CLGEMMLowpMatrixMultiplyCore gemmlowp_mm;
-    gemmlowp_mm.configure(&a, &b, &c);
+    // TODO (giaiod01) COMPMID-1672 - Extending the test to validate add bias in offset contribution
+    gemmlowp_mm.configure(&a, &b, nullptr, &c);
 }
 
 FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpMatrixMultiplyCoreFixture, framework::DatasetMode::ALL, datasets::SmallGEMMLowpDataset())
@@ -155,7 +156,7 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::da
     }
 
     // Validate padding
-    const PaddingSize padding = PaddingCalculator(shape.x(), 16).required_padding();
+    const PaddingSize padding = PaddingCalculator(shape.x(), 4).required_padding();
     validate(in.info()->padding(), padding);
     validate(out.info()->padding(), padding);
 
@@ -238,7 +239,7 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::da
     }
 
     // Validate padding
-    const PaddingSize padding = PaddingCalculator(shape.x(), 16).required_padding();
+    const PaddingSize padding = PaddingCalculator(shape.x(), 4).required_padding();
     validate(in.info()->padding(), padding);
     validate(out.info()->padding(), padding);
 
diff --git a/tests/validation/NEON/GEMMLowp.cpp b/tests/validation/NEON/GEMMLowp.cpp
index 9eba3c85c1..1458c9fdc3 100644
--- a/tests/validation/NEON/GEMMLowp.cpp
+++ b/tests/validation/NEON/GEMMLowp.cpp
@@ -95,7 +95,7 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, framework::dataset::c
 
     // Create and configure function
     NEGEMMLowpMatrixMultiplyCore gemmlowp_mm;
-    gemmlowp_mm.configure(&a, &b, &c);
+    gemmlowp_mm.configure(&a, &b, nullptr, &c);
 }
 
 // *INDENT-OFF*
@@ -125,6 +125,7 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(
     // Lock tensors
     Status status =  NEGEMMLowpMatrixMultiplyCore::validate(&a_info.clone()->set_is_resizable(false),
                                                             &b_info.clone()->set_is_resizable(false),
+                                                            nullptr,
                                                             &output_info.clone()->set_is_resizable(false));
     ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS);
 }
diff --git a/tests/validation/fixtures/GEMMLowpFixture.h b/tests/validation/fixtures/GEMMLowpFixture.h
index 73cb8328ea..b61b4eca38 100644
--- a/tests/validation/fixtures/GEMMLowpFixture.h
+++ b/tests/validation/fixtures/GEMMLowpFixture.h
@@ -75,7 +75,8 @@ protected:
         // Create and configure function
         // The GEMMinfo includes the values of the depth in case of reinterpreted 3d input/output
         FunctionType gemmlowp;
-        gemmlowp.configure(&a, &b, &c, GEMMInfo(false, false, false, (reinterpret_output_as_3d ? shape_c[2] : 1), reinterpret_input_as_3d));
+        // TODO (COMPMID-1672) - Extending the test to validate add bias in offset contribution
+        gemmlowp.configure(&a, &b, nullptr, &c, GEMMInfo(false, false, false, (reinterpret_output_as_3d ? shape_c[2] : 1), reinterpret_input_as_3d));
 
         ARM_COMPUTE_EXPECT(a.info()->is_resizable(), framework::LogLevel::ERRORS);
         ARM_COMPUTE_EXPECT(b.info()->is_resizable(), framework::LogLevel::ERRORS);