commit 700b913ed9257c44147372378bc8a0dadcfd2ac2
author Pablo Marquez Tello <pablo.tello@arm.com> Wed Jun 22 11:23:53 2022 +0100
committer Pablo Marquez Tello <pablo.tello@arm.com> Thu Jun 23 13:45:28 2022 +0000
tree 4fc6c239eff89e4d229d05cc2c81efe8d58bf435
parent 3efdfb3979c429f3f0c8e2e7a47672de8deceac6

Select neon LUT Hard-Swish kernel on all devices

* Resolves COMPMID-5211

Change-Id: I560ab2992c6089774c7ebee3538847905521607d
Signed-off-by: Pablo Marquez Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7840
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>

src/cpu/kernels/CpuActivationKernel.cpp

diff --git a/src/cpu/kernels/CpuActivationKernel.cpp b/src/cpu/kernels/CpuActivationKernel.cpp
index 2ff54cd..330de1e 100644
--- a/src/cpu/kernels/CpuActivationKernel.cpp
+++ b/src/cpu/kernels/CpuActivationKernel.cpp
@@ -45,6 +45,11 @@
 {
 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)
+    },
     {
         "sve2_qu8_activation",
         [](const ActivationDataTypeISASelectorData & data) { return data.dt == DataType::QASYMM8 && data.isa.sve2; },
@@ -86,11 +91,6 @@
         REGISTER_QASYMM8_NEON(arm_compute::cpu::neon_qasymm8_activation)
     },
     {
-        "neon_qu8_activation_hardswish",
-        [](const ActivationDataTypeISASelectorData & data) { return data.dt == DataType::QASYMM8 && data.f == ActivationLayerInfo::ActivationFunction::HARD_SWISH; },
-        REGISTER_QASYMM8_NEON(arm_compute::cpu::neon_qasymm8_hardswish_lut)
-    },
-    {
         "neon_qs8_activation",
         [](const ActivationDataTypeISASelectorData & data) { return data.dt == DataType::QASYMM8_SIGNED; },
         REGISTER_QASYMM8_SIGNED_NEON(arm_compute::cpu::neon_qasymm8_signed_activation)

src/cpu/kernels/activation/generic/sve2/qasymm8.cpp

diff --git a/src/cpu/kernels/activation/generic/sve2/qasymm8.cpp b/src/cpu/kernels/activation/generic/sve2/qasymm8.cpp
index 2fa8dee..928a414 100644
--- a/src/cpu/kernels/activation/generic/sve2/qasymm8.cpp
+++ b/src/cpu/kernels/activation/generic/sve2/qasymm8.cpp
@@ -57,10 +57,7 @@
     const auto                    vconst_1        = svdup_n_f32(1.f);
     const auto                    va_f32          = svdup_n_f32(act_info.a());
     const auto                    vb_f32          = svdup_n_f32(act_info.b());
-    const auto                    const_6_f32     = svdup_n_f32(6.f);
-    const auto                    const_0_f32     = svdup_n_f32(0.f);
-    const auto                    const_3_f32     = svdup_n_f32(3.f);
-    const auto                    const_inv_6_f32 = svdup_n_f32(0.166666667f);
+
 
     // Initialise scale/offset for re-quantization
     bool requant = true;
@@ -146,19 +143,6 @@
                 // Re-quantize to new output space
                 tmp = svquantize_z(pg, tmp_dep, qi_out);
             }
-            else if(act == ActivationLayerInfo::ActivationFunction::HARD_SWISH)
-            {
-                // De-quantize
-                const auto vin_deq = svdequantize_z(pg, vin, qi_in);
-                // Perform activation
-                const svfloat32x4_t tmp_dep = svcreate4_f32(svmul_f32_z(pg, svget4_f32(vin_deq, 0), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg,
-                                                                                                                svget4_f32(vin_deq, 0), const_3_f32))))),
-                                                            svmul_f32_z(pg, svget4_f32(vin_deq, 1), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 1), const_3_f32))))),
-                                                            svmul_f32_z(pg, svget4_f32(vin_deq, 2), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 2), const_3_f32))))),
-                                                            svmul_f32_z(pg, svget4_f32(vin_deq, 3), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 3), const_3_f32))))));
-                // Re-quantize to new output space
-                tmp = svquantize_z(pg, tmp_dep, qi_out);
-            }
             else if(act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU)
             {
                 svbool_t    p0, p1, p2, p3;