Add LUT-based leaky relu for QASYMM8 on CPU

* Add LUT generation function for Leaky ReLU.
* Some additional changes in the existing LUT implementation:
  + Bring back the NEON implementation of hard swish for 32-bit
    build. Library size of 64-bit build is not affected.
  + Add some extra #ifdef to remove unnecessary code in 32-bit
    build.

Resolves: COMPMID-5386
Signed-off-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Change-Id: I1ea49611cc922765ee741e31138c888401d33e9b
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7845
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

5 files changed, 91 insertions, 22 deletions

diff --git a/arm_compute/core/QuantizationInfo.h b/arm_compute/core/QuantizationInfo.h
index 0bd0f21bc1..21d962d08b 100644
--- a/arm_compute/core/QuantizationInfo.h
+++ b/arm_compute/core/QuantizationInfo.h
@@ -409,6 +409,17 @@ inline qasymm8_t qasymm8_hard_swish(qasymm8_t                      in,
     return tmp;
 }
 
+inline qasymm8_t qasymm8_leaky_relu(qasymm8_t                      in,
+                                    const UniformQuantizationInfo &qi_in,
+                                    const UniformQuantizationInfo &qi_out,
+                                    float                          alpha)
+{
+    float tmp_f         = dequantize_qasymm8(in, qi_in);
+    tmp_f               = tmp_f > 0 ? tmp_f : tmp_f * alpha;
+    const qasymm8_t tmp = quantize_qasymm8(tmp_f, qi_out);
+    return tmp;
+}
+
 /** Dequantize a value given a 8-bit symmetric quantization scheme
  *
  * @param[in] value Value to dequantize
diff --git a/arm_compute/core/Types.h b/arm_compute/core/Types.h
index 4524976d6b..94fe1a07f4 100644
--- a/arm_compute/core/Types.h
+++ b/arm_compute/core/Types.h
@@ -1680,6 +1680,7 @@ public:
         return _enabled;
     }
 
+#ifdef __aarch64__
     const LookupTable256 &lut() const
     {
         return _lut;
@@ -1687,7 +1688,27 @@ public:
 
     void init_lut(const UniformQuantizationInfo &qi_in, const UniformQuantizationInfo &qi_out)
     {
-        qasymm8_hard_swish_populate_table(_lut, qi_in, qi_out);
+        if(_act == ActivationFunction::HARD_SWISH)
+        {
+            qasymm8_hard_swish_populate_table(_lut, qi_in, qi_out);
+        }
+        else if(_act == ActivationFunction::LEAKY_RELU)
+        {
+            qasymm8_leaky_relu_populate_table(_lut, qi_in, qi_out, _a);
+        }
+    }
+#endif // __aarch64__
+
+    static inline bool is_lut_supported(ActivationFunction act_func, DataType data_type)
+    {
+#ifdef __aarch64__
+        auto supported = (data_type == DataType::QASYMM8 && (act_func == ActivationFunction::HARD_SWISH || act_func == ActivationFunction::LEAKY_RELU));
+        return supported;
+#else  // __aarch64__
+        ARM_COMPUTE_UNUSED(act_func);
+        ARM_COMPUTE_UNUSED(data_type);
+        return false;
+#endif // __aarch64__
     }
 
 private:
@@ -1695,15 +1716,26 @@ private:
     float              _a       = {};
     float              _b       = {};
     bool               _enabled = { false };
-    LookupTable256     _lut     = {};
 
-    inline void qasymm8_hard_swish_populate_table(LookupTable256 &lut, const UniformQuantizationInfo &qi_in, const UniformQuantizationInfo &qi_out)
+#ifdef __aarch64__
+    LookupTable256 _lut = {};
+
+    static inline void qasymm8_hard_swish_populate_table(LookupTable256 &lut, const UniformQuantizationInfo &qi_in, const UniformQuantizationInfo &qi_out)
     {
         for(size_t i = 0; i < lut.size(); ++i)
         {
             lut[i] = qasymm8_hard_swish(i, qi_in, qi_out);
         }
     }
+
+    static inline void qasymm8_leaky_relu_populate_table(LookupTable256 &lut, const UniformQuantizationInfo &qi_in, const UniformQuantizationInfo &qi_out, float alpha)
+    {
+        for(size_t i = 0; i < lut.size(); ++i)
+        {
+            lut[i] = qasymm8_leaky_relu(i, qi_in, qi_out, alpha);
+        }
+    }
+#endif // __aarch64__
 };
 
 /** Fully connected layer info */
diff --git a/src/cpu/kernels/CpuActivationKernel.cpp b/src/cpu/kernels/CpuActivationKernel.cpp
index 330de1ece2..9eaf44af51 100644
--- a/src/cpu/kernels/CpuActivationKernel.cpp
+++ b/src/cpu/kernels/CpuActivationKernel.cpp
@@ -45,11 +45,13 @@ namespace
 {
 static const std::vector<CpuActivationKernel::ActivationKernel> available_kernels =
 {
-    { // neon LUT implementantion of HARD_SWISH takes precedence
-        "neon_qu8_activation_hardswish_lut",
-        [](const ActivationDataTypeISASelectorData & data) { return data.dt == DataType::QASYMM8 && data.f == ActivationLayerInfo::ActivationFunction::HARD_SWISH; },
-        REGISTER_QASYMM8_NEON(arm_compute::cpu::neon_qasymm8_hardswish_lut)
+#ifdef __aarch64__
+    { // Neon LUT implementantion takes precedence
+        "neon_qu8_activation_lut",
+        [](const ActivationDataTypeISASelectorData & data) { return ActivationLayerInfo::is_lut_supported(data.f, data.dt); },
+        REGISTER_QASYMM8_NEON(arm_compute::cpu::neon_qasymm8_activation_lut)
     },
+#endif // __aarch64__
     {
         "sve2_qu8_activation",
         [](const ActivationDataTypeISASelectorData & data) { return data.dt == DataType::QASYMM8 && data.isa.sve2; },
@@ -87,7 +89,7 @@ static const std::vector<CpuActivationKernel::ActivationKernel> available_kernel
     },
     {
         "neon_qu8_activation",
-        [](const ActivationDataTypeISASelectorData & data) { return data.dt == DataType::QASYMM8 && data.f != ActivationLayerInfo::ActivationFunction::HARD_SWISH; },
+        [](const ActivationDataTypeISASelectorData & data) { return data.dt == DataType::QASYMM8; },
         REGISTER_QASYMM8_NEON(arm_compute::cpu::neon_qasymm8_activation)
     },
     {
@@ -188,10 +190,12 @@ void CpuActivationKernel::configure(const ITensorInfo *src, ITensorInfo *dst, Ac
     _run_method = uk->ukernel;
     _name       = std::string("CpuActivationKernel").append("/").append(uk->name);
 
-    if(activation_info.activation() == ActivationLayerInfo::ActivationFunction::HARD_SWISH && src->data_type() == DataType::QASYMM8)
+#ifdef __aarch64__
+    if(ActivationLayerInfo::is_lut_supported(activation_info.activation(), src->data_type()))
     {
         activation_info.init_lut(src->quantization_info().uniform(),(dst)?dst->quantization_info().uniform():src->quantization_info().uniform());
     }
+#endif // __aarch64__
     _act_info = activation_info;
 
     // Configure kernel window
diff --git a/src/cpu/kernels/activation/generic/neon/qasymm8.cpp b/src/cpu/kernels/activation/generic/neon/qasymm8.cpp
index 29f5e6b376..5095ecf5bd 100644
--- a/src/cpu/kernels/activation/generic/neon/qasymm8.cpp
+++ b/src/cpu/kernels/activation/generic/neon/qasymm8.cpp
@@ -417,9 +417,9 @@ void substitute_bytes_neon(
 #endif // __aarch64__
 } // namespace
 
-void neon_qasymm8_hardswish_lut(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
+void neon_qasymm8_activation_lut(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
 {
-    ARM_COMPUTE_ERROR_ON(act_info.activation() != ActivationLayerInfo::ActivationFunction::HARD_SWISH);
+    ARM_COMPUTE_ERROR_ON(!ActivationLayerInfo::is_lut_supported(act_info.activation(), src->info()->data_type()));
 #ifdef __aarch64__
     const int window_step_x  = src->info()->tensor_shape().x();
     Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
@@ -472,6 +472,13 @@ void neon_qasymm8_activation(const ITensor *src, ITensor *dst, const ActivationL
     const float       a_f32  = act_info.a();
     const float       b_f32  = act_info.b();
 
+#ifndef __aarch64__
+    const auto const_6_f32     = vdupq_n_f32(6.f);
+    const auto const_0_f32     = vdupq_n_f32(0.f);
+    const auto const_3_f32     = vdupq_n_f32(3.f);
+    const auto const_inv_6_f32 = vdupq_n_f32(0.166666667f);
+#endif // __aarch64__
+
     // Initialise scale/offset for re-quantization
     float       s  = qi_in.scale / qi_out.scale;
     float       o  = -qi_in.offset * s + qi_out.offset;
@@ -545,21 +552,28 @@ void neon_qasymm8_activation(const ITensor *src, ITensor *dst, const ActivationL
                 // Re-quantize to new output space
                 tmp = vquantize(tmp_dep, qi_out);
             }
-            else if(act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU)
+#ifndef __aarch64__ // LUT-based implementation is used for aarch64 instead.
+            else if(act == ActivationLayerInfo::ActivationFunction::HARD_SWISH)
             {
+                // De-quantize
                 const auto vin_deq = vdequantize(vin, qi_in);
-
-#ifdef __aarch64__
-                const uint32x4x4_t pos_mask =
+                // Perform activation
+                const float32x4x4_t tmp_dep =
                 {
                     {
-                        wrapper::vcgtz(vin_deq.val[0]),
-                        wrapper::vcgtz(vin_deq.val[1]),
-                        wrapper::vcgtz(vin_deq.val[2]),
-                        wrapper::vcgtz(vin_deq.val[3]),
+                        wrapper::vmul(vin_deq.val[0], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[0], const_3_f32))))),
+                        wrapper::vmul(vin_deq.val[1], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[1], const_3_f32))))),
+                        wrapper::vmul(vin_deq.val[2], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[2], const_3_f32))))),
+                        wrapper::vmul(vin_deq.val[3], wrapper::vmul(const_inv_6_f32, wrapper::vmin(const_6_f32, wrapper::vmax(const_0_f32, wrapper::vadd(vin_deq.val[3], const_3_f32))))),
                     }
                 };
-#else  // __aarch64__
+                // Re-quantize to new output space
+                tmp = vquantize(tmp_dep, qi_out);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU)
+            {
+                const auto vin_deq = vdequantize(vin, qi_in);
+
                 const uint32x4x4_t pos_mask =
                 {
                     {
@@ -569,7 +583,6 @@ void neon_qasymm8_activation(const ITensor *src, ITensor *dst, const ActivationL
                         wrapper::vcgt(vin_deq.val[3], vconst_0_f32),
                     }
                 };
-#endif // __aarch64__
 
                 const float32x4x4_t tmp_dep =
                 {
@@ -583,6 +596,7 @@ void neon_qasymm8_activation(const ITensor *src, ITensor *dst, const ActivationL
 
                 tmp = vquantize(tmp_dep, qi_out);
             }
+#endif // __aarch64__
             else
             {
                 ARM_COMPUTE_ERROR("Unsupported activation function");
@@ -622,12 +636,20 @@ void neon_qasymm8_activation(const ITensor *src, ITensor *dst, const ActivationL
                 tmp_f       = a_f32 * std::tanh(b_f32 * tmp_f);
                 tmp         = quantize_qasymm8(tmp_f, qi_out);
             }
+#ifndef __aarch64__ // LUT-based implementation is used for aarch64 instead.
+            else if(act == ActivationLayerInfo::ActivationFunction::HARD_SWISH)
+            {
+                float tmp_f = dequantize_qasymm8(in, qi_in);
+                tmp_f       = tmp_f * ((std::min(std::max((tmp_f + 3), 0.0f), 6.0f)) * 0.166666667f);
+                tmp         = quantize_qasymm8(tmp_f, qi_out);
+            }
             else if(act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU)
             {
                 float tmp_f = dequantize_qasymm8(in, qi_in);
                 tmp_f       = tmp_f > 0 ? tmp_f : tmp_f * a_f32;
                 tmp         = quantize_qasymm8(tmp_f, qi_out);
             }
+#endif // __aarch64__
             else
             {
                 ARM_COMPUTE_ERROR("Unsupported activation function");
diff --git a/src/cpu/kernels/activation/list.h b/src/cpu/kernels/activation/list.h
index 7220d6cce1..b2322a6477 100644
--- a/src/cpu/kernels/activation/list.h
+++ b/src/cpu/kernels/activation/list.h
@@ -32,7 +32,7 @@ namespace cpu
     void func_name(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
 
 DECLARE_ACTIVATION_KERNEL(neon_qasymm8_activation);
-DECLARE_ACTIVATION_KERNEL(neon_qasymm8_hardswish_lut);
+DECLARE_ACTIVATION_KERNEL(neon_qasymm8_activation_lut);
 DECLARE_ACTIVATION_KERNEL(sve2_qasymm8_activation);
 DECLARE_ACTIVATION_KERNEL(neon_qasymm8_signed_activation);
 DECLARE_ACTIVATION_KERNEL(sve2_qasymm8_signed_activation);