IVGCVSW-4743: Fix CpuAcc Hal 1.3 Comparison Failures

Broadcast for QASYMM8_SIGNED was not handled.

Change-Id: Id5dbb0dce78838319218de94551bba52d697f4a4
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3131
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

3 files changed, 102 insertions, 46 deletions

diff --git a/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp b/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp
index 0579dc67f4..3fd5f39e9f 100644
--- a/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp
+++ b/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp
@@ -116,24 +116,6 @@ void store_quantized_signed(int8_t *output_ptr, const float32x4x4_t &rf, const f
     store_quantized_signed(output_ptr, out);
 }
 
-float32x4x4_t dup_quantized(qasymm8_t broadcast_value, int offset, float scale)
-{
-    const qasymm8x16_t broadcast_value_vec = vdupq_n_u8(broadcast_value);
-    const int32x4_t    voffset             = vdupq_n_s32(offset);
-    const float32x4_t  vscale              = vdupq_n_f32(scale);
-
-    const float32x4x4_t broadcast_vector =
-    {
-        {
-            vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(broadcast_value_vec))))), voffset)), vscale),
-            vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(broadcast_value_vec))))), voffset)), vscale),
-            vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(broadcast_value_vec))))), voffset)), vscale),
-            vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_high_u8(broadcast_value_vec))))), voffset)), vscale),
-        }
-    };
-    return broadcast_vector;
-}
-
 template <ArithmeticOperation op, typename ScalarType>
 inline ScalarType elementwise_arithm_op_scalar(const ScalarType &a, const ScalarType &b)
 {
@@ -596,6 +578,23 @@ inline int elementwise_comp_op_quantized_broadcast_loop(int window_start_x, int
     return x;
 }
 
+template <ComparisonOperation op>
+inline int elementwise_comp_op_quantized_signed_broadcast_loop(int window_start_x, int window_end_x, int window_step_x,
+                                                               const int8_t *non_broadcast_input_ptr, float32x4x4_t broadcast_vector, uint8_t *output_ptr,
+                                                               int32x4_t voffset_non_broadcast, float32x4_t vscale_non_broadcast,
+                                                               float32x4_t voffseto, float32x4_t invvscaleo, bool reorder)
+{
+    ARM_COMPUTE_UNUSED(voffseto, invvscaleo);
+    int x = window_start_x;
+    for(; x <= (window_end_x - window_step_x); x += window_step_x)
+    {
+        const float32x4x4_t af = load_quantized_signed(non_broadcast_input_ptr + x, voffset_non_broadcast, vscale_non_broadcast);
+        const uint32x4x4_t  rf = elementwise_comp_op<op>(reorder ? broadcast_vector : af, reorder ? af : broadcast_vector);
+        store_quantized(output_ptr + x, rf);
+    }
+    return x;
+}
+
 template <typename InputScalarType, typename OutputScalarType, typename InputVectorType>
 void elementwise_op(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window,
                     OutputScalarType (*scalar_func)(const InputScalarType &, const InputScalarType &),
@@ -728,7 +727,7 @@ void elementwise_op_quantized(const ITensor *in1, const ITensor *in2, ITensor *o
             const auto output_ptr              = reinterpret_cast<uint8_t *>(output.ptr());
 
             const uint8_t       broadcast_value  = *reinterpret_cast<const uint8_t *>(broadcast_input.ptr());
-            const float32x4x4_t broadcast_vector = dup_quantized(broadcast_value, broadcast_qinfo.offset, broadcast_qinfo.scale);
+            const float32x4x4_t broadcast_vector = vdequantize(vdupq_n_u8(broadcast_value), broadcast_qinfo);
 
             int x = (*broadcast_func)(window_start_x, window_end_x, window_step_x, non_broadcast_input_ptr, broadcast_vector, output_ptr,
                                       voffset_non_broadcast, vscale_non_broadcast, voffseto, invvscaleo, !is_broadcast_input_2);
@@ -783,6 +782,8 @@ void elementwise_op_quantized(const ITensor *in1, const ITensor *in2, ITensor *o
 
 void elementwise_comp_quantized_signed(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window,
                                        uint8_t (*scalar_func)(const float &, const float &, UniformQuantizationInfo),
+                                       int (*broadcast_func)(int, int, int, const int8_t *, float32x4x4_t, uint8_t *, int32x4_t, float32x4_t,
+                                                             float32x4_t, float32x4_t, const bool),
                                        int (*neon_func)(int, int, int, const int8_t *, const int8_t *, uint8_t *,
                                                         int32x4_t, int32x4_t, float32x4_t, float32x4_t,
                                                         float32x4_t, float32x4_t))
@@ -795,13 +796,58 @@ void elementwise_comp_quantized_signed(const ITensor *in1, const ITensor *in2, I
     Window win = window;
     win.set(Window::DimX, Window::Dimension(0, 1, 1));
 
-    const int                     window_step_x  = 16;
-    const auto                    window_start_x = static_cast<int>(window.x().start());
-    const auto                    window_end_x   = static_cast<int>(window.x().end());
-    const UniformQuantizationInfo output_qinfo   = out->info()->quantization_info().uniform();
+    const int  window_step_x         = 16;
+    const auto window_start_x        = static_cast<int>(window.x().start());
+    const auto window_end_x          = static_cast<int>(window.x().end());
+    const bool is_broadcast_across_x = (input1_win.x().step() == 0) || (input2_win.x().step() == 0);
+
+    const UniformQuantizationInfo output_qinfo = out->info()->quantization_info().uniform();
 
     const float32x4_t voffseto   = vdupq_n_f32(output_qinfo.offset);
     const float32x4_t invvscaleo = vdupq_n_f32(1.f / output_qinfo.scale);
+
+    if(is_broadcast_across_x)
+    {
+        // Select the broadcast input on the X axis
+        const bool     is_broadcast_input_2 = input2_win.x().step() == 0;
+        Window         broadcast_win        = is_broadcast_input_2 ? input2_win : input1_win;
+        Window         non_broadcast_win    = !is_broadcast_input_2 ? input2_win : input1_win;
+        const ITensor *broadcast_tensor     = is_broadcast_input_2 ? in2 : in1;
+        const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1;
+
+        const UniformQuantizationInfo broadcast_qinfo     = broadcast_tensor->info()->quantization_info().uniform();
+        const UniformQuantizationInfo non_broadcast_qinfo = non_broadcast_tensor->info()->quantization_info().uniform();
+
+        const int32x4_t   voffset_non_broadcast = vdupq_n_s32(non_broadcast_qinfo.offset);
+        const float32x4_t vscale_non_broadcast  = vdupq_n_f32(non_broadcast_qinfo.scale);
+
+        // Clear X Dimension on execution window as we handle manually
+        non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator broadcast_input(broadcast_tensor, broadcast_win);
+        Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto non_broadcast_input_ptr = reinterpret_cast<const int8_t *>(non_broadcast_input.ptr());
+            const auto output_ptr              = reinterpret_cast<uint8_t *>(output.ptr());
+
+            const int8_t        broadcast_value  = *reinterpret_cast<const int8_t *>(broadcast_input.ptr());
+            const float32x4x4_t broadcast_vector = vdequantize(vdupq_n_s8(broadcast_value), broadcast_qinfo);
+
+            int x = (*broadcast_func)(window_start_x, window_end_x, window_step_x, non_broadcast_input_ptr, broadcast_vector, output_ptr,
+                                      voffset_non_broadcast, vscale_non_broadcast, voffseto, invvscaleo, !is_broadcast_input_2);
+            for(; x < window_end_x; ++x)
+            {
+                const float afs   = dequantize_qasymm8_signed(*(non_broadcast_input_ptr + x), non_broadcast_qinfo);
+                const float bfs   = dequantize_qasymm8_signed(broadcast_value, broadcast_qinfo);
+                *(output_ptr + x) = (*scalar_func)(!is_broadcast_input_2 ? bfs : afs, !is_broadcast_input_2 ? afs : bfs, output_qinfo);
+            }
+        },
+        broadcast_input, non_broadcast_input, output);
+    }
+    else
     {
         const UniformQuantizationInfo input1_qinfo = in1->info()->quantization_info().uniform();
         const UniformQuantizationInfo input2_qinfo = in2->info()->quantization_info().uniform();
@@ -895,7 +941,7 @@ void elementwise_op_quantized_signed(const ITensor *in1, const ITensor *in2, ITe
             const auto output_ptr              = reinterpret_cast<int8_t *>(output.ptr());
 
             const int8_t        broadcast_value  = *reinterpret_cast<const int8_t *>(broadcast_input.ptr());
-            const float32x4x4_t broadcast_vector = dup_quantized(broadcast_value, broadcast_qinfo.offset, broadcast_qinfo.scale);
+            const float32x4x4_t broadcast_vector = vdequantize(vdupq_n_u8(broadcast_value), broadcast_qinfo);
 
             int x = (*broadcast_func)(window_start_x, window_end_x, window_step_x, non_broadcast_input_ptr, broadcast_vector, output_ptr,
                                       voffset_non_broadcast, vscale_non_broadcast, voffseto, invvscaleo, !is_broadcast_input_2);
@@ -1003,7 +1049,9 @@ void elementwise_comp_op_quantized(const ITensor *in1, const ITensor *in2, ITens
 template <ComparisonOperation op>
 void elementwise_comp_op_quantized_signed(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
 {
-    elementwise_comp_quantized_signed(in1, in2, out, window, &elementwise_comp_op_quantized_scalar<op>, &elementwise_comp_op_quantized_signed_loop<op>);
+    elementwise_comp_quantized_signed(in1, in2, out, window, &elementwise_comp_op_quantized_scalar<op>,
+                                      &elementwise_comp_op_quantized_signed_broadcast_loop<op>,
+                                      &elementwise_comp_op_quantized_signed_loop<op>);
 }
 
 std::function<void(const ITensor *, const ITensor *, ITensor *, const Window &)>
diff --git a/tests/validation/NEON/Comparisons.cpp b/tests/validation/NEON/Comparisons.cpp
index 38e440e649..f080c834e5 100644
--- a/tests/validation/NEON/Comparisons.cpp
+++ b/tests/validation/NEON/Comparisons.cpp
@@ -52,8 +52,9 @@ const auto configure_dataset = combine(datasets::SmallShapes(),
                                                                               DataType::F32
                                                                             }));
 
-const auto run_small_dataset = combine(datasets::ComparisonOperations(), datasets::SmallShapes());
-const auto run_large_dataset = combine(datasets::ComparisonOperations(), datasets::LargeShapes());
+const auto run_small_dataset           = combine(datasets::ComparisonOperations(), datasets::SmallShapes());
+const auto run_small_broadcast_dataset = combine(datasets::ComparisonOperations(), datasets::SmallShapesBroadcast());
+const auto run_large_dataset           = combine(datasets::ComparisonOperations(), datasets::LargeShapes());
 
 } // namespace
 
@@ -90,23 +91,6 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(
 // clang-format on
 // *INDENT-ON*
 
-DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, configure_dataset,
-               shape, data_type)
-{
-    // Create tensors
-    Tensor ref_src1 = create_tensor<Tensor>(shape, data_type);
-    Tensor ref_src2 = create_tensor<Tensor>(shape, data_type);
-    Tensor dst      = create_tensor<Tensor>(shape, DataType::U8);
-
-    // Create and Configure function
-    NEElementwiseComparison compare;
-    compare.configure(&ref_src1, &ref_src2, &dst, ComparisonOperation::Equal);
-
-    // Validate valid region
-    const ValidRegion valid_region = shape_to_valid_region(shape);
-    validate(dst.info()->valid_region(), valid_region);
-}
-
 template <typename T>
 using NEComparisonFixture = ComparisonValidationFixture<Tensor, Accessor, NEElementwiseComparison, T>;
 
@@ -154,6 +138,8 @@ TEST_SUITE_END() // Float
 
 template <typename T>
 using NEComparisonQuantizedFixture = ComparisonValidationQuantizedFixture<Tensor, Accessor, NEElementwiseComparison, T>;
+template <typename T>
+using NEComparisonQuantizedBroadcastFixture = ComparisonQuantizedBroadcastValidationFixture<Tensor, Accessor, NEElementwiseComparison, T>;
 
 TEST_SUITE(Quantized)
 TEST_SUITE(QASYMM8)
@@ -169,11 +155,22 @@ FIXTURE_DATA_TEST_CASE(RunSmall,
 }
 TEST_SUITE_END()
 TEST_SUITE(QASYMM8_SIGNED)
+FIXTURE_DATA_TEST_CASE(RunSmallBroadcast,
+                       NEComparisonQuantizedBroadcastFixture<int8_t>,
+                       framework::DatasetMode::ALL,
+                       combine(combine(combine(run_small_broadcast_dataset, framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)),
+                                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.1, -30) })),
+                               framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.3f, 2) })))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+
 FIXTURE_DATA_TEST_CASE(RunSmall,
                        NEComparisonQuantizedFixture<int8_t>,
                        framework::DatasetMode::PRECOMMIT,
                        combine(combine(combine(run_small_dataset, framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)),
-                                       framework::dataset::make("QuantizationInfo", { QuantizationInfo() })),
+                                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.1, -30) })),
                                framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.3f, 2) })))
 {
     // Validate output
diff --git a/tests/validation/fixtures/ComparisonFixture.h b/tests/validation/fixtures/ComparisonFixture.h
index b2fe42de26..d1e1a539c7 100644
--- a/tests/validation/fixtures/ComparisonFixture.h
+++ b/tests/validation/fixtures/ComparisonFixture.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -146,6 +146,17 @@ public:
         ComparisonValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(op, shape, shape, data_type, qinfo0, qinfo1);
     }
 };
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class ComparisonQuantizedBroadcastValidationFixture : public ComparisonValidationGenericFixture<TensorType, AccessorType, FunctionType, T>
+{
+public:
+    template <typename...>
+    void setup(ComparisonOperation op, const TensorShape &shape0, const TensorShape &shape1, DataType data_type, QuantizationInfo qinfo0, QuantizationInfo qinfo1)
+    {
+        ComparisonValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(op, shape0, shape1, data_type, qinfo0, qinfo1);
+    }
+};
 } // namespace validation
 } // namespace test
 } // namespace arm_compute