COMPMID-3784 Add broadcast support to S32 NEPixelwiseMultiplication

Signed-off-by: SiCong Li <sicong.li@arm.com>
Change-Id: Ifae31c74eb46c561225394a387fc15332423bfa9
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4030
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>

4 files changed, 172 insertions, 88 deletions

diff --git a/arm_compute/core/NEON/kernels/NEPixelWiseMultiplicationKernel.h b/arm_compute/core/NEON/kernels/NEPixelWiseMultiplicationKernel.h
index c65f788091..6221d61f49 100644
--- a/arm_compute/core/NEON/kernels/NEPixelWiseMultiplicationKernel.h
+++ b/arm_compute/core/NEON/kernels/NEPixelWiseMultiplicationKernel.h
@@ -55,16 +55,17 @@ public:
      *
      * Valid configurations (Input1,Input2) -> Output :
      *
-     *   - (U8,U8)                         -> U8, S16
-     *   - (U8,S16)                        -> S16
-     *   - (S16,U8)                        -> S16
-     *   - (S16,S16)                       -> S16
-     *   - (S32,S32)                       -> S32
-     *   - (F16,F16)                       -> F16
-     *   - (F32,F32)                       -> F32
-     *   - (QASYMM8,QASYMM8)               -> QASYMM8
-     *   - (QASYMM8_SIGNED,QASYMM8_SIGNED) -> QASYMM8_SIGNED
-     *   - (QSYMM16,QSYMM16)               -> QSYMM16, S32
+     *                                                       Support: Broadcast? Scale=1/255?
+     *   - (U8,U8)                         -> U8, S16                 N          Y
+     *   - (U8,S16)                        -> S16                     N          Y
+     *   - (S16,U8)                        -> S16                     N          Y
+     *   - (S16,S16)                       -> S16                     N          Y
+     *   - (S32,S32)                       -> S32                     Y          N
+     *   - (F16,F16)                       -> F16                     N          Y
+     *   - (F32,F32)                       -> F32                     Y          Y
+     *   - (QASYMM8,QASYMM8)               -> QASYMM8                 Y          Y
+     *   - (QASYMM8_SIGNED,QASYMM8_SIGNED) -> QASYMM8_SIGNED          Y          Y
+     *   - (QSYMM16,QSYMM16)               -> QSYMM16, S32            N          Y
      *
      * @note For @p scale equal to 1/255 only round to nearest even (implemented as round half up) is supported.
      *       For all other scale values only round to zero (implemented as round towards minus infinity) is supported.
@@ -82,17 +83,17 @@ public:
     /** Static function to check if given info will lead to a valid configuration of @ref NEPixelWiseMultiplicationKernel
      *
      * Valid configurations (Input1,Input2) -> Output :
-     *
-     *   - (U8,U8)                         -> U8, S16
-     *   - (U8,S16)                        -> S16
-     *   - (S16,U8)                        -> S16
-     *   - (S16,S16)                       -> S16
-     *   - (S32,S32)                       -> S32
-     *   - (F16,F16)                       -> F16
-     *   - (F32,F32)                       -> F32
-     *   - (QASYMM8,QASYMM8)               -> QASYMM8
-     *   - (QASYMM8_SIGNED,QASYMM8_SIGNED) -> QASYMM8_SIGNED
-     *   - (QSYMM16,QSYMM16)               -> QSYMM16, S32
+     *                                                       Support: Broadcast? Scale=1/255?
+     *   - (U8,U8)                         -> U8, S16                 N          Y
+     *   - (U8,S16)                        -> S16                     N          Y
+     *   - (S16,U8)                        -> S16                     N          Y
+     *   - (S16,S16)                       -> S16                     N          Y
+     *   - (S32,S32)                       -> S32                     Y          N
+     *   - (F16,F16)                       -> F16                     N          Y
+     *   - (F32,F32)                       -> F32                     Y          Y
+     *   - (QASYMM8,QASYMM8)               -> QASYMM8                 Y          Y
+     *   - (QASYMM8_SIGNED,QASYMM8_SIGNED) -> QASYMM8_SIGNED          Y          Y
+     *   - (QSYMM16,QSYMM16)               -> QSYMM16, S32            N          Y
      *
      * @note For @p scale equal to 1/255 only round to nearest even (implemented as round half up) is supported.
      *       For all other scale values only round to zero (implemented as round towards minus infinity) is supported.
diff --git a/arm_compute/runtime/NEON/functions/NEPixelWiseMultiplication.h b/arm_compute/runtime/NEON/functions/NEPixelWiseMultiplication.h
index 4ff7f1d112..e1072980cf 100644
--- a/arm_compute/runtime/NEON/functions/NEPixelWiseMultiplication.h
+++ b/arm_compute/runtime/NEON/functions/NEPixelWiseMultiplication.h
@@ -42,16 +42,17 @@ public:
      *
      * Valid configurations (Input1,Input2) -> Output :
      *
-     *   - (U8,U8)                         -> U8, S16
-     *   - (U8,S16)                        -> S16
-     *   - (S16,U8)                        -> S16
-     *   - (S16,S16)                       -> S16
-     *   - (S32,S32)                       -> S32
-     *   - (F16,F16)                       -> F16
-     *   - (F32,F32)                       -> F32
-     *   - (QASYMM8,QASYMM8)               -> QASYMM8
-     *   - (QASYMM8_SIGNED,QASYMM8_SIGNED) -> QASYMM8_SIGNED
-     *   - (QSYMM16,QSYMM16)               -> QSYMM16, S32
+     *                                                       Support: Broadcast? Scale=1/255?
+     *   - (U8,U8)                         -> U8, S16                 N          Y
+     *   - (U8,S16)                        -> S16                     N          Y
+     *   - (S16,U8)                        -> S16                     N          Y
+     *   - (S16,S16)                       -> S16                     N          Y
+     *   - (S32,S32)                       -> S32                     Y          N
+     *   - (F16,F16)                       -> F16                     N          Y
+     *   - (F32,F32)                       -> F32                     Y          Y
+     *   - (QASYMM8,QASYMM8)               -> QASYMM8                 Y          Y
+     *   - (QASYMM8_SIGNED,QASYMM8_SIGNED) -> QASYMM8_SIGNED          Y          Y
+     *   - (QSYMM16,QSYMM16)               -> QSYMM16, S32            N          Y
      *
      * @note For @p scale equal to 1/255 only round to nearest even (implemented as round half up) is supported.
      *       For all other scale values only round to zero (implemented as round towards minus infinity) is supported.
@@ -74,16 +75,17 @@ public:
      *
      * Valid configurations (Input1,Input2) -> Output :
      *
-     *   - (U8,U8)                         -> U8, S16
-     *   - (U8,S16)                        -> S16
-     *   - (S16,U8)                        -> S16
-     *   - (S16,S16)                       -> S16
-     *   - (S32,S32)                       -> S32
-     *   - (F16,F16)                       -> F16
-     *   - (F32,F32)                       -> F32
-     *   - (QASYMM8,QASYMM8)               -> QASYMM8
-     *   - (QASYMM8_SIGNED,QASYMM8_SIGNED) -> QASYMM8_SIGNED
-     *   - (QSYMM16,QSYMM16)               -> QSYMM16, S32
+     *                                                       Support: Broadcast? Scale=1/255?
+     *   - (U8,U8)                         -> U8, S16                 N          Y
+     *   - (U8,S16)                        -> S16                     N          Y
+     *   - (S16,U8)                        -> S16                     N          Y
+     *   - (S16,S16)                       -> S16                     N          Y
+     *   - (S32,S32)                       -> S32                     Y          N
+     *   - (F16,F16)                       -> F16                     N          Y
+     *   - (F32,F32)                       -> F32                     Y          Y
+     *   - (QASYMM8,QASYMM8)               -> QASYMM8                 Y          Y
+     *   - (QASYMM8_SIGNED,QASYMM8_SIGNED) -> QASYMM8_SIGNED          Y          Y
+     *   - (QSYMM16,QSYMM16)               -> QSYMM16, S32            N          Y
      *
      * @note For @p scale equal to 1/255 only round to nearest even (implemented as round half up) is supported.
      *       For all other scale values only round to zero (implemented as round towards minus infinity) is supported.
diff --git a/src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp b/src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp
index 302ee7694f..84683ea69f 100644
--- a/src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp
+++ b/src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp
@@ -773,75 +773,151 @@ template <bool is_sat>
 void mul_S32_S32_S32(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window, int n)
 {
     // Create input windows
-    Window win        = window;
     Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
     Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
 
     // Clear X Dimension on execution window as we handle manually
+    Window win = window;
     win.set(Window::DimX, Window::Dimension(0, 1, 1));
-    input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
-    input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
-
-    Iterator input1(in1, input1_win);
-    Iterator input2(in2, input2_win);
-    Iterator output(out, win);
 
-    const int  window_step_x  = 8;
-    const auto window_start_x = static_cast<int>(window.x().start());
-    const auto window_end_x   = static_cast<int>(window.x().end());
+    const int  window_step_x         = 8;
+    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);
 
-    execute_window_loop(win, [&](const Coordinates &)
+    if(is_broadcast_across_x)
     {
-        const auto input1_ptr = reinterpret_cast<const int32_t *>(input1.ptr());
-        const auto input2_ptr = reinterpret_cast<const int32_t *>(input2.ptr());
-        const auto output_ptr = reinterpret_cast<int32_t *>(output.ptr());
+        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;
 
-        // Compute window_step_x elements per iteration
-        int x = window_start_x;
-        for(; x <= (window_end_x - window_step_x); x += window_step_x)
+        // 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 int32x4x2_t ta1 =
+            const auto non_broadcast_input_ptr = reinterpret_cast<const int32_t *>(non_broadcast_input.ptr());
+            const auto output_ptr              = reinterpret_cast<int32_t *>(output.ptr());
+
+            const int32_t broadcast_value     = *reinterpret_cast<const int32_t *>(broadcast_input.ptr());
+            const auto    broadcast_value_vec = vdupq_n_s32(broadcast_value);
+
+            // Compute window_step_x elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
             {
+                const int32x4x2_t broadcast_v =
                 {
-                    vld1q_s32(input1_ptr + x),
-                    vld1q_s32(input1_ptr + x + 4),
-                }
-            };
-            const int32x4x2_t ta2 =
+                    {
+                        broadcast_value_vec,
+                        broadcast_value_vec,
+                    }
+                };
+                const int32x4x2_t non_broadcast_v =
+                {
+                    {
+                        vld1q_s32(non_broadcast_input_ptr + x),
+                        vld1q_s32(non_broadcast_input_ptr + x + 4),
+                    }
+                };
+                const int32x4x2_t result = mul_S32_S32_S32_n_k<is_sat>(broadcast_v, non_broadcast_v, n);
+
+                vst1q_s32(output_ptr + x, result.val[0]);
+                vst1q_s32(output_ptr + x + 4, result.val[1]);
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
             {
+                int64_t tmp = static_cast<int64_t>(broadcast_value) * static_cast<int64_t>(*(non_broadcast_input_ptr + x));
+
+                if(tmp >= 0)
                 {
-                    vld1q_s32(input2_ptr + x),
-                    vld1q_s32(input2_ptr + x + 4),
+                    tmp >>= n;
                 }
-            };
-            const int32x4x2_t result = mul_S32_S32_S32_n_k<is_sat>(ta1, ta2, n);
+                else
+                {
+                    uint64_t mask = (1u << n) - 1;
+                    tmp           = (tmp + static_cast<int64_t>(mask)) >> n;
+                }
+                if(is_sat)
+                {
+                    tmp = utility::clamp<int64_t, int32_t>(tmp);
+                }
+                *(output_ptr + x) = static_cast<int32_t>(tmp);
+            }
+        },
+        broadcast_input, non_broadcast_input, output);
+    }
+    else
+    {
+        // Clear X Dimension on execution window as we handle manually
+        input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+        input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
 
-            vst1q_s32(output_ptr + x, result.val[0]);
-            vst1q_s32(output_ptr + x + 4, result.val[1]);
-        }
+        Iterator input1(in1, input1_win);
+        Iterator input2(in2, input2_win);
+        Iterator output(out, win);
 
-        // Compute left-over elements
-        for(; x < window_end_x; ++x)
+        execute_window_loop(win, [&](const Coordinates &)
         {
-            int64_t tmp = static_cast<int64_t>(*(input1_ptr + x)) * static_cast<int64_t>(*(input2_ptr + x));
+            const auto input1_ptr = reinterpret_cast<const int32_t *>(input1.ptr());
+            const auto input2_ptr = reinterpret_cast<const int32_t *>(input2.ptr());
+            const auto output_ptr = reinterpret_cast<int32_t *>(output.ptr());
 
-            if(tmp >= 0)
-            {
-                tmp >>= n;
-            }
-            else
+            // Compute window_step_x elements per iteration
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
             {
-                uint64_t mask = (1u << n) - 1;
-                tmp           = (tmp + static_cast<int64_t>(mask)) >> n;
+                const int32x4x2_t ta1 =
+                {
+                    {
+                        vld1q_s32(input1_ptr + x),
+                        vld1q_s32(input1_ptr + x + 4),
+                    }
+                };
+                const int32x4x2_t ta2 =
+                {
+                    {
+                        vld1q_s32(input2_ptr + x),
+                        vld1q_s32(input2_ptr + x + 4),
+                    }
+                };
+                const int32x4x2_t result = mul_S32_S32_S32_n_k<is_sat>(ta1, ta2, n);
+
+                vst1q_s32(output_ptr + x, result.val[0]);
+                vst1q_s32(output_ptr + x + 4, result.val[1]);
             }
-            if(is_sat)
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
             {
-                tmp = (tmp > INT_MAX) ? INT_MAX : ((tmp < INT_MIN) ? INT_MIN : tmp);
+                int64_t tmp = static_cast<int64_t>(*(input1_ptr + x)) * static_cast<int64_t>(*(input2_ptr + x));
+
+                if(tmp >= 0)
+                {
+                    tmp >>= n;
+                }
+                else
+                {
+                    uint64_t mask = (1u << n) - 1;
+                    tmp           = (tmp + static_cast<int64_t>(mask)) >> n;
+                }
+                if(is_sat)
+                {
+                    tmp = utility::clamp<int64_t, int32_t>(tmp);
+                }
+                *(output_ptr + x) = static_cast<int32_t>(tmp);
             }
-            *(output_ptr + x) = static_cast<int32_t>(tmp);
-        }
-    },
-    input1, input2, output);
+        },
+        input1, input2, output);
+    }
 }
 
 void mul_F32_F32_F32(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window, float scale)
diff --git a/tests/validation/NEON/PixelWiseMultiplication.cpp b/tests/validation/NEON/PixelWiseMultiplication.cpp
index a66f6f192f..1bb0588919 100644
--- a/tests/validation/NEON/PixelWiseMultiplication.cpp
+++ b/tests/validation/NEON/PixelWiseMultiplication.cpp
@@ -118,7 +118,7 @@ template <typename T>
 using NEPixelWiseMultiplicationToF32Fixture     = PixelWiseMultiplicationValidationFixture<Tensor, Accessor, NEPixelWiseMultiplication, T, float>;
 using NEPixelWiseMultiplicationU8U8ToS16Fixture = PixelWiseMultiplicationValidationFixture<Tensor, Accessor, NEPixelWiseMultiplication, uint8_t, uint8_t, int16_t>;
 template <typename T>
-using NEPixelWiseMultiplicationBroadcastFixture              = PixelWiseMultiplicationBroadcastValidationFixture<Tensor, Accessor, NEPixelWiseMultiplication, T, float>;
+using NEPixelWiseMultiplicationBroadcastFixture              = PixelWiseMultiplicationBroadcastValidationFixture<Tensor, Accessor, NEPixelWiseMultiplication, T, T>;
 using NEPixelWiseMultiplicationBroadcastQASYMM8Fixture       = PixelWiseMultiplicationBroadcastValidationQuantizedFixture<Tensor, Accessor, NEPixelWiseMultiplication, uint8_t, uint8_t>;
 using NEPixelWiseMultiplicationBroadcastQASYMM8SignedFixture = PixelWiseMultiplicationBroadcastValidationQuantizedFixture<Tensor, Accessor, NEPixelWiseMultiplication, int8_t, int8_t>;
 
@@ -493,6 +493,11 @@ TEST_SUITE(ScaleOther)
 PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunSmall, ToS32Fixture<int32_t>, ALL, SmallShapes(), S32, S32, S32, scale_other, TO_ZERO, InPlaceDataSet, WRAP_VALIDATE(int32_t, 1))
 TEST_SUITE_END() // ScaleOther
 
+TEST_SUITE(Broadcast)
+PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunSmall, BroadcastFixture<int32_t>, ALL, SmallShapesBroadcast(), S32, S32, S32, scale_unity, TO_ZERO, framework::dataset::make("InPlace", { false }),
+                                                 WRAP_VALIDATE(int32_t, 1))
+TEST_SUITE_END() // Broadcast
+
 TEST_SUITE_END() // S32toS32
 
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC