COMPMID-1021: CPUInfo refactoring.

Removed CPUTarget in favor of the CPUModel type.
CPUInfo now holds a vector of N CPUs.
CPUInfo autoinitialise upon construction with 1 GENERIC CPU.
CPPScheduler fills CPUInfo's vector upon construction (runtime).
IScheduler has a single CPUInfo obj and ThreadInfo always gets a pointer to it (avoid copying the vector)

Change-Id: I30f293258c959c87f6bac5eac8b963beb6a4d365
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/124626
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

9 files changed, 429 insertions, 212 deletions

diff --git a/src/runtime/CPP/CPPScheduler.cpp b/src/runtime/CPP/CPPScheduler.cpp
index 168ed6e30f..92dce34c71 100644
--- a/src/runtime/CPP/CPPScheduler.cpp
+++ b/src/runtime/CPP/CPPScheduler.cpp
@@ -27,6 +27,7 @@
 #include "arm_compute/core/Error.h"
 #include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/Utils.h"
+#include "arm_compute/runtime/CPUUtils.h"
 
 #include <condition_variable>
 #include <iostream>
@@ -159,6 +160,7 @@ CPPScheduler::CPPScheduler()
     : _num_threads(num_threads_hint()),
       _threads(_num_threads - 1)
 {
+    get_cpu_configuration(_cpu_info);
 }
 
 void CPPScheduler::set_num_threads(unsigned int num_threads)
@@ -178,7 +180,7 @@ void CPPScheduler::schedule(ICPPKernel *kernel, unsigned int split_dimension)
 
     /** [Scheduler example] */
     ThreadInfo info;
-    info.cpu_info = _info;
+    info.cpu_info = &_cpu_info;
 
     const Window      &max_window     = kernel->window();
     const unsigned int num_iterations = max_window.num_iterations(split_dimension);
diff --git a/src/runtime/CPP/SingleThreadScheduler.cpp b/src/runtime/CPP/SingleThreadScheduler.cpp
index c8285b43a7..2adc14ce80 100644
--- a/src/runtime/CPP/SingleThreadScheduler.cpp
+++ b/src/runtime/CPP/SingleThreadScheduler.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -45,7 +45,7 @@ void SingleThreadScheduler::schedule(ICPPKernel *kernel, unsigned int split_dime
 {
     ARM_COMPUTE_UNUSED(split_dimension);
     ThreadInfo info;
-    info.cpu_info = cpu_info();
+    info.cpu_info = &_cpu_info;
     kernel->run(kernel->window(), info);
 }
 
diff --git a/src/runtime/CPUUtils.cpp b/src/runtime/CPUUtils.cpp
new file mode 100644
index 0000000000..7e8bf2bb3f
--- /dev/null
+++ b/src/runtime/CPUUtils.cpp
@@ -0,0 +1,404 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/runtime/CPUUtils.h"
+
+#include "arm_compute/core/CPP/CPPTypes.h"
+#include "arm_compute/core/Error.h"
+#include "support/ToolchainSupport.h"
+
+#include <array>
+#include <cstdlib>
+#include <cstring>
+#include <fcntl.h>
+#include <fstream>
+#include <map>
+#include <sched.h>
+#include <sys/stat.h>
+#include <sys/types.h>
+#include <unistd.h>
+
+#ifndef BARE_METAL
+#include <regex>
+#include <thread>
+#endif /* BARE_METAL */
+
+#if !defined(BARE_METAL) && (defined(__arm__) || defined(__aarch64__))
+#include <sys/auxv.h>
+
+/* Get HWCAP bits from asm/hwcap.h */
+#include <asm/hwcap.h>
+#endif /* !BARE_METAL */
+
+/* Make sure the bits we care about are defined, just in case asm/hwcap.h is
+ * out of date (or for bare metal mode) */
+#ifndef HWCAP_ASIMDHP
+#define HWCAP_ASIMDHP (1 << 10)
+#endif /* HWCAP_ASIMDHP */
+
+#ifndef HWCAP_CPUID
+#define HWCAP_CPUID (1 << 11)
+#endif /* HWCAP_CPUID */
+
+#ifndef HWCAP_ASIMDDP
+#define HWCAP_ASIMDDP (1 << 20)
+#endif /* HWCAP_ASIMDDP */
+
+namespace
+{
+using namespace arm_compute;
+
+#if !defined(BARE_METAL) && (defined(__arm__) || defined(__aarch64__))
+struct PerCPUData
+{
+    CPUModel     model     = CPUModel::GENERIC;
+    unsigned int midr      = 0;
+    bool         model_set = false;
+};
+
+/* Convert an MIDR register value to a CPUModel enum value. */
+CPUModel midr_to_model(const unsigned int midr)
+{
+    CPUModel model;
+
+    // Unpack variant and CPU ID
+    const int variant = (midr >> 20) & 0xF;
+    const int cpunum  = (midr >> 4) & 0xFFF;
+
+    // Only CPUs we have code paths for are detected.  All other CPUs can be safely classed as "GENERIC"
+    switch(cpunum)
+    {
+        case 0xd03:
+            model = CPUModel::A53;
+            break;
+
+        case 0xd05:
+            if(variant != 0)
+            {
+                model = CPUModel::A55r1;
+            }
+            else
+            {
+                model = CPUModel::A55r0;
+            }
+            break;
+
+        default:
+            model = CPUModel::GENERIC;
+            break;
+    }
+
+    return model;
+}
+
+void populate_models_cpuid(std::vector<PerCPUData> &cpusv)
+{
+    // If the CPUID capability is present, MIDR information is provided in /sys. Use that to populate the CPU model table.
+    uint32_t i = 0;
+    for(auto &c : cpusv)
+    {
+        std::stringstream str;
+        str << "/sys/devices/system/cpu/cpu" << i++ << "/regs/identification/midr_el1";
+        std::ifstream file;
+        file.open(str.str(), std::ios::in);
+        if(file.is_open())
+        {
+            std::string line;
+            if(bool(getline(file, line)))
+            {
+                const unsigned long midr = support::cpp11::stoul(line, nullptr, 16);
+                c.midr                   = (midr & 0xffffffff);
+                c.model                  = midr_to_model(c.midr);
+                c.model_set              = true;
+            }
+        }
+    }
+}
+
+void populate_models_cpuinfo(std::vector<PerCPUData> &cpusv)
+{
+    // If "long-form" cpuinfo is present, parse that to populate models.
+    std::regex proc_regex("^processor.*(\\d+)$");
+    std::regex imp_regex("^CPU implementer.*0x(..)$");
+    std::regex var_regex("^CPU variant.*0x(.)$");
+    std::regex part_regex("^CPU part.*0x(...)$");
+    std::regex rev_regex("^CPU revision.*(\\d+)$");
+
+    std::ifstream file;
+    file.open("/proc/cpuinfo", std::ios::in);
+
+    if(file.is_open())
+    {
+        std::string line;
+        int         midr   = 0;
+        int         curcpu = -1;
+
+        while(bool(getline(file, line)))
+        {
+            std::smatch match;
+
+            if(std::regex_match(line, match, proc_regex))
+            {
+                std::string id     = match[1];
+                int         newcpu = support::cpp11::stoi(id, nullptr, 0);
+
+                if(curcpu >= 0 && midr == 0)
+                {
+                    // Matched a new CPU ID without any description of the previous one - looks like old format.
+                    return;
+                }
+
+                if(curcpu >= 0)
+                {
+                    cpusv[curcpu].midr      = midr;
+                    cpusv[curcpu].model     = midr_to_model(midr);
+                    cpusv[curcpu].model_set = true;
+                }
+
+                midr   = 0;
+                curcpu = newcpu;
+
+                continue;
+            }
+
+            if(std::regex_match(line, match, imp_regex))
+            {
+                int impv = support::cpp11::stoi(match[1], nullptr, 16);
+                midr |= (impv << 24);
+                continue;
+            }
+
+            if(std::regex_match(line, match, var_regex))
+            {
+                int varv = support::cpp11::stoi(match[1], nullptr, 16);
+                midr |= (varv << 16);
+                continue;
+            }
+
+            if(std::regex_match(line, match, part_regex))
+            {
+                int partv = support::cpp11::stoi(match[1], nullptr, 16);
+                midr |= (partv << 4);
+                continue;
+            }
+
+            if(std::regex_match(line, match, rev_regex))
+            {
+                int regv = support::cpp11::stoi(match[1], nullptr, 10);
+                midr |= (regv);
+                midr |= (0xf << 16);
+                continue;
+            }
+        }
+
+        if(curcpu >= 0)
+        {
+            cpusv[curcpu].midr      = midr;
+            cpusv[curcpu].model     = midr_to_model(midr);
+            cpusv[curcpu].model_set = true;
+        }
+    }
+}
+
+int get_max_cpus()
+{
+    int max_cpus = 1;
+#if !defined(BARE_METAL) && (defined(__arm__) || defined(__aarch64__))
+    std::ifstream CPUspresent;
+    CPUspresent.open("/sys/devices/system/cpu/present", std::ios::in);
+    bool success = false;
+
+    if(CPUspresent.is_open())
+    {
+        std::string line;
+
+        if(bool(getline(CPUspresent, line)))
+        {
+            /* The content of this file is a list of ranges or single values, e.g.
+                 * 0-5, or 1-3,5,7 or similar.  As we are interested in the
+                 * max valid ID, we just need to find the last valid
+                 * delimiter ('-' or ',') and parse the integer immediately after that.
+                 */
+            auto startfrom = line.begin();
+
+            for(auto i = line.begin(); i < line.end(); ++i)
+            {
+                if(*i == '-' || *i == ',')
+                {
+                    startfrom = i + 1;
+                }
+            }
+
+            line.erase(line.begin(), startfrom);
+
+            max_cpus = support::cpp11::stoi(line, nullptr, 0) + 1;
+            success  = true;
+        }
+    }
+
+    // Return std::thread::hardware_concurrency() as a fallback.
+    if(!success)
+    {
+        max_cpus = std::thread::hardware_concurrency();
+    }
+#endif /* BARE_METAL */
+
+    return max_cpus;
+}
+#endif /* !defined(BARE_METAL) && (defined(__arm__) || defined(__aarch64__)) */
+
+} // namespace
+
+namespace arm_compute
+{
+void get_cpu_configuration(CPUInfo &cpuinfo)
+{
+#if !defined(BARE_METAL) && (defined(__arm__) || defined(__aarch64__))
+    bool cpuid        = false;
+    bool fp16_support = false;
+    bool dot_support  = false;
+
+    const uint32_t hwcaps = getauxval(AT_HWCAP);
+
+    if((hwcaps & HWCAP_CPUID) != 0)
+    {
+        cpuid = true;
+    }
+
+    if((hwcaps & HWCAP_ASIMDHP) != 0)
+    {
+        fp16_support = true;
+    }
+
+    if((hwcaps & HWCAP_ASIMDDP) != 0)
+    {
+        dot_support = true;
+    }
+
+#ifdef __aarch64__
+    /* Pre-4.15 kernels don't have the ASIMDDP bit.
+     *
+     * Although the CPUID bit allows us to read the feature register
+     * directly, the kernel quite sensibly masks this to only show
+     * features known by it to be safe to show to userspace.  As a
+     * result, pre-4.15 kernels won't show the relevant bit in the
+     * feature registers either.
+     *
+     * So for now, use a whitelist of CPUs known to support the feature.
+     */
+    if(!dot_support && cpuid)
+    {
+        /* List of CPUs with dot product support:         A55r1       A75r1       A75r2  */
+        const unsigned int dotprod_whitelist_masks[]  = { 0xfff0fff0, 0xfff0fff0, 0xfff0fff0, 0 };
+        const unsigned int dotprod_whitelist_values[] = { 0x4110d050, 0x4110d0a0, 0x4120d0a0, 0 };
+
+        unsigned long cpuid;
+
+        __asm __volatile(
+            "mrs %0, midr_el1\n"
+            : "=r"(cpuid)
+            :
+            : );
+
+        for(int i = 0; dotprod_whitelist_values[i] != 0; i++)
+        {
+            if((cpuid & dotprod_whitelist_masks[i]) == dotprod_whitelist_values[i])
+            {
+                dot_support = true;
+                break;
+            }
+        }
+    }
+#endif /* __aarch64__ */
+    const unsigned int max_cpus = get_max_cpus();
+    cpuinfo.set_cpu_num(max_cpus);
+    cpuinfo.set_fp16(fp16_support);
+    cpuinfo.set_dotprod(dot_support);
+    std::vector<PerCPUData> percpu(max_cpus);
+    if(cpuid)
+    {
+        populate_models_cpuid(percpu);
+    }
+    else
+    {
+        populate_models_cpuinfo(percpu);
+    }
+    int j(0);
+    for(const auto &v : percpu)
+    {
+        cpuinfo.set_cpu_model(j++, v.model);
+    }
+#else  /* !defined(BARE_METAL) && (defined(__arm__) || defined(__aarch64__)) */
+    ARM_COMPUTE_UNUSED(cpuinfo);
+#endif /* !defined(BARE_METAL) && (defined(__arm__) || defined(__aarch64__)) */
+}
+
+unsigned int get_threads_hint()
+{
+    unsigned int num_threads_hint = 1;
+
+#ifndef BARE_METAL
+    std::map<std::string, unsigned int> cpu_part_occurrence_map;
+
+    // CPU part regex
+    std::regex  cpu_part_rgx(R"(.*CPU part.+?(?=:).+?(?=\w+)(\w+).*)");
+    std::smatch cpu_part_match;
+
+    // Read cpuinfo and get occurrence of each core
+    std::ifstream cpuinfo;
+    cpuinfo.open("/proc/cpuinfo", std::ios::in);
+    if(cpuinfo.is_open())
+    {
+        std::string line;
+        while(bool(getline(cpuinfo, line)))
+        {
+            if(std::regex_search(line.cbegin(), line.cend(), cpu_part_match, cpu_part_rgx))
+            {
+                std::string cpu_part = cpu_part_match[1];
+                if(cpu_part_occurrence_map.find(cpu_part) != cpu_part_occurrence_map.end())
+                {
+                    cpu_part_occurrence_map[cpu_part]++;
+                }
+                else
+                {
+                    cpu_part_occurrence_map[cpu_part] = 1;
+                }
+            }
+        }
+    }
+
+    // Get min number of threads
+    auto min_common_cores = std::min_element(cpu_part_occurrence_map.begin(), cpu_part_occurrence_map.end(),
+                                             [](const std::pair<std::string, unsigned int> &p1, const std::pair<std::string, unsigned int> &p2)
+    {
+        return p1.second < p2.second;
+    });
+
+    // Set thread hint
+    num_threads_hint = cpu_part_occurrence_map.empty() ? std::thread::hardware_concurrency() : min_common_cores->second;
+#endif /* BARE_METAL */
+
+    return num_threads_hint;
+}
+
+} // namespace arm_compute
diff --git a/src/runtime/IScheduler.cpp b/src/runtime/IScheduler.cpp
index 583cb40eca..54a2bd2182 100644
--- a/src/runtime/IScheduler.cpp
+++ b/src/runtime/IScheduler.cpp
@@ -23,202 +23,20 @@
  */
 #include "arm_compute/runtime/IScheduler.h"
 
-#include <array>
-#include <cstdlib>
-#include <cstring>
-#include <fcntl.h>
-#include <fstream>
-#include <map>
-#include <sched.h>
-#include <sys/stat.h>
-#include <sys/types.h>
-#include <unistd.h>
-
-#ifndef BARE_METAL
-#include <regex>
-#include <thread>
-#endif /* BARE_METAL */
-
-namespace
-{
-unsigned int get_threads_hint()
-{
-    unsigned int num_threads_hint = 1;
-
-#ifndef BARE_METAL
-    std::map<std::string, unsigned int> cpu_part_occurrence_map;
-
-    // CPU part regex
-    std::regex  cpu_part_rgx(R"(.*CPU part.+?(?=:).+?(?=\w+)(\w+).*)");
-    std::smatch cpu_part_match;
-
-    // Read cpuinfo and get occurrence of each core
-    std::ifstream cpuinfo;
-    cpuinfo.open("/proc/cpuinfo", std::ios::in);
-    if(cpuinfo.is_open())
-    {
-        std::string line;
-        while(bool(getline(cpuinfo, line)))
-        {
-            if(std::regex_search(line.cbegin(), line.cend(), cpu_part_match, cpu_part_rgx))
-            {
-                std::string cpu_part = cpu_part_match[1];
-                if(cpu_part_occurrence_map.find(cpu_part) != cpu_part_occurrence_map.end())
-                {
-                    cpu_part_occurrence_map[cpu_part]++;
-                }
-                else
-                {
-                    cpu_part_occurrence_map[cpu_part] = 1;
-                }
-            }
-        }
-    }
-
-    // Get min number of threads
-    auto min_common_cores = std::min_element(cpu_part_occurrence_map.begin(), cpu_part_occurrence_map.end(),
-                                             [](const std::pair<std::string, unsigned int> &p1, const std::pair<std::string, unsigned int> &p2)
-    {
-        return p1.second < p2.second;
-    });
-
-    // Set thread hint
-    num_threads_hint = cpu_part_occurrence_map.empty() ? std::thread::hardware_concurrency() : min_common_cores->second;
-#endif /* BARE_METAL */
-
-    return num_threads_hint;
-}
-
-unsigned int get_cpu_impl()
-{
-#ifndef BARE_METAL
-    int fd = open("/proc/cpuinfo", 0); // NOLINT
-    std::array<char, 3000> buff{ {} };
-    char *pos     = nullptr;
-    char *end     = nullptr;
-    bool  foundid = false;
-
-    int cpu = sched_getcpu();
-
-    if(fd == -1)
-    {
-        return 0;
-    }
-
-    int charsread = read(fd, buff.data(), 3000);
-    pos           = buff.data();
-    end           = buff.data() + charsread;
-
-    close(fd);
-
-    /* So, to date I've encountered two formats for /proc/cpuinfo.
-     *
-     * One of them just lists processor : n  for each processor (with no
-     * other info), then at the end lists part information for the current
-     * CPU.
-     *
-     * The other has an entire clause (including part number info) for each
-     * CPU in the system, with "processor : n" headers.
-     *
-     * We can cope with either of these formats by waiting to see
-     * "processor: n" (where n = our CPU ID), and then looking for the next
-     * "CPU part" field.
-     */
-    while(pos < end)
-    {
-        if(foundid && strncmp(pos, "CPU part", 8) == 0)
-        {
-            /* Found part number */
-            pos += 11;
-
-            for(char *ch = pos; ch < end; ch++)
-            {
-                if(*ch == '\n')
-                {
-                    *ch = '\0';
-                    break;
-                }
-            }
-
-            return strtoul(pos, nullptr, 0);
-        }
-
-        if(strncmp(pos, "processor", 9) == 0)
-        {
-            /* Found processor ID, see if it's ours. */
-            pos += 11;
-
-            for(char *ch = pos; ch < end; ch++)
-            {
-                if(*ch == '\n')
-                {
-                    *ch = '\0';
-                    break;
-                }
-            }
-
-            int num = strtol(pos, nullptr, 0);
-
-            if(num == cpu)
-            {
-                foundid = true;
-            }
-        }
-
-        while(pos < end)
-        {
-            char ch = *pos++;
-            if(ch == '\n' || ch == '\0')
-            {
-                break;
-            }
-        }
-    }
-#endif /* BARE_METAL */
-
-    return 0;
-}
-} // namespace
+#include "arm_compute/runtime/CPUUtils.h"
 
 namespace arm_compute
 {
 IScheduler::IScheduler()
+    : _cpu_info()
 {
     // Work out the best possible number of execution threads
     _num_threads_hint = get_threads_hint();
-
-    // Work out the CPU implementation
-    switch(get_cpu_impl())
-    {
-        case 0xd0f:
-            _info.CPU = CPUTarget::A55_DOT;
-            break;
-        case 0xd03:
-            _info.CPU = CPUTarget::A53;
-            break;
-        default:
-#ifdef __arm__
-            _info.CPU = CPUTarget::ARMV7;
-#elif __aarch64__
-            _info.CPU = CPUTarget::ARMV8;
-#else  /* __arm__ || __aarch64__ */
-            _info.CPU = CPUTarget::INTRINSICS;
-#endif /* __arm__ || __aarch64__ */
-            break;
-    }
-
-    _info.L1_size = 31000;
-    _info.L2_size = 500000;
-}
-
-void IScheduler::set_target(CPUTarget target)
-{
-    _info.CPU = target;
 }
 
-CPUInfo IScheduler::cpu_info() const
+CPUInfo &IScheduler::cpu_info()
 {
-    return _info;
+    return _cpu_info;
 }
 
 unsigned int IScheduler::num_threads_hint() const
diff --git a/src/runtime/NEON/functions/NEGEMM.cpp b/src/runtime/NEON/functions/NEGEMM.cpp
index c8cba8a174..18e6e919c3 100644
--- a/src/runtime/NEON/functions/NEGEMM.cpp
+++ b/src/runtime/NEON/functions/NEGEMM.cpp
@@ -65,7 +65,8 @@ void NEGEMM::configure(const ITensor *a, const ITensor *b, const ITensor *c, ITe
     // Check if we need to reshape the matrix B only on the first run
     _reshape_b_only_on_first_run      = gemm_info.reshape_b_only_on_first_run();
     _run_vector_matrix_multiplication = a->info()->dimension(1) < 2;
-    const bool run_optimised          = setup_assembly_kernel(a, b, c, d, alpha, beta, _workspace, _memory_group, _asm_glue);
+
+    const bool run_optimised = a->info()->data_type() == DataType::F32 && (c == nullptr || beta == 0.f) && setup_assembly_kernel(a, b, d, alpha, beta, _workspace, _memory_group, _asm_glue);
 
     // Check if the first input tensor is a vector.
     // If so, all the kernels for reshaping the tensors can be skipped
diff --git a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
index b2dd0227a5..cdbd32373a 100644
--- a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
@@ -271,8 +271,7 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig
     const unsigned int fixed_point_position = input->info()->fixed_point_position();
     const ITensor     *biases_to_use        = (_append_bias) ? biases : nullptr;
 
-    bool run_optimised =
-        (NEScheduler::get().cpu_info().CPU == CPUTarget::ARMV7 && dt == DataType::F32) || (NEScheduler::get().cpu_info().CPU >= CPUTarget::ARMV8 && dt == DataType::F32);
+    bool run_optimised = dt == DataType::F32;
 
     // Reshape weights if needed
     if(run_optimised)
@@ -369,8 +368,10 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig
     // Configure matrix multiply
     if(run_optimised)
     {
-        run_optimised = setup_assembly_kernel(&_input_im2col_reshaped, weights, nullptr, &_gemm_output, 1.f, 0.f, _workspace, _memory_group, _asm_glue);
-        ARM_COMPUTE_ERROR_ON_MSG(run_optimised == false, "setup_assembly_kernel failed.");
+        if(!setup_assembly_kernel(&_input_im2col_reshaped, weights, &_gemm_output, 1.f, 0.f, _workspace, _memory_group, _asm_glue))
+        {
+            ARM_COMPUTE_ERROR("setup_assembly_kernel failed.");
+        }
     }
     else
     {
@@ -450,17 +451,10 @@ Status NEGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI
     std::unique_ptr<ITensorInfo> reshaped_weights = weights->clone();
     bool                         optimised_kernel = false;
 
-#if defined(__arm__)
-    if(NEScheduler::get().cpu_info().CPU == CPUTarget::ARMV7 && dt == DataType::F32)
-    {
-        optimised_kernel = true;
-    }
-#elif defined(__aarch64__)
-    if(NEScheduler::get().cpu_info().CPU >= CPUTarget::ARMV8 && dt == DataType::F32)
+    if(dt == DataType::F32)
     {
         optimised_kernel = true;
     }
-#endif /* defined(__arm__) || defined(__aarch64__) */
 
     // Reshape weights if needed
     if(optimised_kernel)
diff --git a/src/runtime/NEON/functions/NEGEMMLowpAssemblyMatrixMultiplyCore.cpp b/src/runtime/NEON/functions/NEGEMMLowpAssemblyMatrixMultiplyCore.cpp
index a02872c3d5..27dd6c51d7 100644
--- a/src/runtime/NEON/functions/NEGEMMLowpAssemblyMatrixMultiplyCore.cpp
+++ b/src/runtime/NEON/functions/NEGEMMLowpAssemblyMatrixMultiplyCore.cpp
@@ -58,13 +58,13 @@ void NEGEMMLowpAssemblyMatrixMultiplyCore::configure(const ITensor *a, const ITe
     {
         case DataType::S8:
         {
-            run_optimised = setup_assembly_kernel(a, b, nullptr, output, 1.f, 1.f, _workspace, _memory_group, _asm_glue_signed);
+            run_optimised = setup_assembly_kernel(a, b, output, 1.f, 1.f, _workspace, _memory_group, _asm_glue_signed);
             break;
         }
         case DataType::QASYMM8:
         case DataType::U8:
         {
-            run_optimised = setup_assembly_kernel(a, b, nullptr, output, 1.f, 1.f, _workspace, _memory_group, _asm_glue_unsigned);
+            run_optimised = setup_assembly_kernel(a, b, output, 1.f, 1.f, _workspace, _memory_group, _asm_glue_unsigned);
             break;
         }
         default:
diff --git a/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp b/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp
index b0a7bdd7bc..7372c6ca57 100644
--- a/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp
+++ b/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp
@@ -62,13 +62,13 @@ void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b,
     {
         case DataType::S8:
         {
-            _dot_product_path = setup_assembly_kernel(a, b, nullptr, output, 1.f, 1.f, _workspace, _memory_group, _asm_glue_signed);
+            _dot_product_path = setup_assembly_kernel(a, b, output, 1.f, 1.f, _workspace, _memory_group, _asm_glue_signed);
             break;
         }
         case DataType::QASYMM8:
         case DataType::U8:
         {
-            _dot_product_path = setup_assembly_kernel(a, b, nullptr, output, 1.f, 1.f, _workspace, _memory_group, _asm_glue_unsigned);
+            _dot_product_path = setup_assembly_kernel(a, b, output, 1.f, 1.f, _workspace, _memory_group, _asm_glue_unsigned);
             break;
         }
         default:
@@ -156,10 +156,6 @@ void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b,
         _tmp_a.allocator()->allocate();
         _tmp_b.allocator()->allocate();
     }
-    else
-    {
-        _workspace.allocator()->allocate();
-    }
 
     if(_a_offset != 0)
     {
diff --git a/src/runtime/OMP/OMPScheduler.cpp b/src/runtime/OMP/OMPScheduler.cpp
index 3b30f1e56b..795c96caf0 100644
--- a/src/runtime/OMP/OMPScheduler.cpp
+++ b/src/runtime/OMP/OMPScheduler.cpp
@@ -27,6 +27,7 @@
 #include "arm_compute/core/Error.h"
 #include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/Utils.h"
+#include "arm_compute/runtime/CPUUtils.h"
 
 #include <omp.h>
 
@@ -41,6 +42,7 @@ OMPScheduler &OMPScheduler::get()
 OMPScheduler::OMPScheduler() // NOLINT
     : _num_threads(omp_get_max_threads())
 {
+    get_cpu_configuration(_cpu_info);
 }
 
 unsigned int OMPScheduler::num_threads() const
@@ -59,7 +61,7 @@ void OMPScheduler::schedule(ICPPKernel *kernel, unsigned int split_dimension)
     ARM_COMPUTE_ERROR_ON_MSG(!kernel, "The child class didn't set the kernel");
 
     ThreadInfo info;
-    info.cpu_info = _info;
+    info.cpu_info = &_cpu_info;
 
     const Window      &max_window     = kernel->window();
     const unsigned int num_iterations = max_window.num_iterations(split_dimension);
@@ -71,7 +73,7 @@ void OMPScheduler::schedule(ICPPKernel *kernel, unsigned int split_dimension)
     }
     else
     {
-        #pragma omp parallel private(info) num_threads(info.num_threads)
+        #pragma omp parallel firstprivate(info) num_threads(info.num_threads)
         {
             const int tid  = omp_get_thread_num();
             Window    win  = max_window.split_window(split_dimension, tid, info.num_threads);