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>

3 files changed, 71 insertions, 63 deletions

diff --git a/arm_compute/runtime/CPUUtils.h b/arm_compute/runtime/CPUUtils.h
new file mode 100644
index 0000000000..70211a5817
--- /dev/null
+++ b/arm_compute/runtime/CPUUtils.h
@@ -0,0 +1,44 @@
+/*
+ * 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.
+ */
+#ifndef __ARM_COMPUTE_RUNTIME_CPU_UTILS_H__
+#define __ARM_COMPUTE_RUNTIME_CPU_UTILS_H__
+
+namespace arm_compute
+{
+class CPUInfo;
+/** This function will try to detect the CPU configuration on the system and will fill
+ *  the cpuinfo object accordingly to reflect this.
+ *
+ * @param[out] cpuinfo @ref CPUInfo to be used to hold the system's cpu configuration.
+ */
+void get_cpu_configuration(CPUInfo &cpuinfo);
+/** Some systems have both big and small cores, this fuction computes the minimum number of cores
+ *  that are exactly the same on the system. To maximize performance the library attempts to process
+ *  workloads concurrently using as many threads as big cores are available on the system.
+ *
+ * @return The minumum number of common cores.
+ */
+unsigned int get_threads_hint();
+}
+#endif /* __ARM_COMPUTE_RUNTIME_CPU_UTILS_H__ */
diff --git a/arm_compute/runtime/IScheduler.h b/arm_compute/runtime/IScheduler.h
index 1dd7c2cfb2..a0bcada722 100644
--- a/arm_compute/runtime/IScheduler.h
+++ b/arm_compute/runtime/IScheduler.h
@@ -59,17 +59,11 @@ public:
      */
     virtual void schedule(ICPPKernel *kernel, unsigned int split_dimension) = 0;
 
-    /** Sets the target CPU architecture.
-     *
-     * @param[in] target Target CPU.
-     */
-    void set_target(CPUTarget target);
-
     /** Get CPU info.
      *
      * @return CPU info.
      */
-    CPUInfo cpu_info() const;
+    CPUInfo &cpu_info();
     /** Get a hint for the best possible number of execution threads
      *
      * @warning In case we can't work out the best number of threads,
@@ -80,7 +74,7 @@ public:
     unsigned int num_threads_hint() const;
 
 protected:
-    CPUInfo _info{};
+    CPUInfo _cpu_info;
 
 private:
     unsigned int _num_threads_hint = {};
diff --git a/arm_compute/runtime/NEON/AssemblyHelper.h b/arm_compute/runtime/NEON/AssemblyHelper.h
index e2d27cf941..40f28587c2 100644
--- a/arm_compute/runtime/NEON/AssemblyHelper.h
+++ b/arm_compute/runtime/NEON/AssemblyHelper.h
@@ -127,70 +127,32 @@ inline void allocate_workspace(size_t workspace_size, Tensor &workspace, MemoryG
 
 /** Create a wrapper kernel.
  *
- * @param[in]  a     Input tensor A.
- * @param[in]  b     Input tensor B.
- * @param[in]  c     (Optional) Input tensor C.
- * @param[out] d     Output tensor.
- * @param[in]  alpha Alpha value.
- * @param[in]  beta  Beta value.
- *
- * @return the wrapper kernel.
- */
-template <typename T>
-std::unique_ptr<NEGEMMAssemblyWrapper<T>> create_wrapper_kernel(const ITensor *a, const ITensor *b, const ITensor *c, ITensor *d, float alpha, float beta)
-{
-    // rework this function, why are we checking data type and other things here ? should we create another function can_run_optimised_kernel() ?
-#if defined(__arm__)
-    if(NEScheduler::get().cpu_info().CPU == CPUTarget::ARMV7 && a->info()->data_type() == DataType::F32 && (c == nullptr || beta == 0.f))
-    {
-        return support::cpp14::make_unique<NEGEMMAssemblyWrapper<T>>();
-    }
-#elif defined(__aarch64__)
-    if(NEScheduler::get().cpu_info().CPU >= CPUTarget::ARMV8 && a->info()->data_type() == DataType::F32 && (c == nullptr || beta == 0.f))
-    {
-        return support::cpp14::make_unique<NEGEMMAssemblyWrapper<T>>();
-    }
-    else if(a->info()->data_type() == DataType::F16 && (c == nullptr || beta == 0.f))
-    {
-#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-        return support::cpp14::make_unique<NEGEMMAssemblyWrapper<T>>();
-#else  /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
-        ARM_COMPUTE_ERROR("Recompile the library with arch=arm64-v8.2-a to enable support for FP16.");
-#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
-    }
-#endif /* defined(__arm__) || defined(__aarch64__) */
-    return nullptr;
-}
-
-/** Setup assembly kernel.
- *
  * @param[in]  a            Input tensor A.
  * @param[in]  b            Input tensor B.
- * @param[in]  c            (Optional) Input tensor C.
- * @param[in]  d            Output tensor.
+ * @param[out] d            Output tensor.
  * @param[in]  alpha        Alpha value.
  * @param[in]  beta         Beta value.
  * @param[out] workspace    Workspace tensor
  * @param[in]  memory_group Tensor memory group.
  * @param[out] asm_glue     Assembly glue kernel.
  *
- * @return True if the assembly kernel is setup correctly.
+ * @return the wrapper kernel.
  */
 template <typename T>
-inline bool setup_assembly_kernel(const ITensor *a, const ITensor *b, const ITensor *c, ITensor *d, float alpha, float beta,
+inline bool setup_assembly_kernel(const ITensor *a, const ITensor *b, ITensor *d, float alpha, float beta,
                                   Tensor &workspace, MemoryGroup &memory_group, T &asm_glue)
 {
-    const ::CPUInfo *ci          = get_CPUInfo();
-    const int        M           = d->info()->tensor_shape().y();
-    const int        N           = d->info()->tensor_shape().x();
-    const int        K           = a->info()->tensor_shape().x();
-    unsigned int     num_threads = NEScheduler::get().num_threads();
+    const CPUInfo &ci          = NEScheduler::get().cpu_info();
+    const int      M           = d->info()->tensor_shape().y();
+    const int      N           = d->info()->tensor_shape().x();
+    const int      K           = a->info()->tensor_shape().x();
+    unsigned int   num_threads = NEScheduler::get().num_threads();
     // unique_ptr to a Gemm object
-    std::unique_ptr<typename T::AssemblyGemm> asm_gemm(arm_gemm::gemm<typename T::TypeOperator, typename T::TypeResult>(*ci, M, N, K, false, false, alpha, beta, num_threads,
-                                                                                                                        false));
-
+    std::unique_ptr<typename T::AssemblyGemm>
+    asm_gemm(arm_gemm::gemm<typename T::TypeOperator, typename T::TypeResult>(ci, M, N, K, false, false, alpha, beta, num_threads, false));
     // arm_compute wrapper for the Gemm object (see above)
-    std::unique_ptr<NEGEMMAssemblyWrapper<typename T::AssemblyGemm>> acl_gemm_wrapper = create_wrapper_kernel<typename T::AssemblyGemm>(a, b, c, d, alpha, beta);
+    std::unique_ptr<NEGEMMAssemblyWrapper<typename T::AssemblyGemm>>
+                                                                  acl_gemm_wrapper = support::cpp14::make_unique<NEGEMMAssemblyWrapper<typename T::AssemblyGemm>>();
     if(acl_gemm_wrapper != nullptr && asm_gemm != nullptr)
     {
         acl_gemm_wrapper->configure(asm_gemm.get());
@@ -198,15 +160,23 @@ inline bool setup_assembly_kernel(const ITensor *a, const ITensor *b, const ITen
         if(workspace_size)
         {
             // Allocate workspace
-            allocate_workspace(workspace_size, workspace, memory_group, 4096, num_threads);
+            const unsigned int alignment = 4096;
+            allocate_workspace(workspace_size, workspace, memory_group, alignment, num_threads);
+            ARM_COMPUTE_ERROR_ON_NULLPTR(workspace.buffer());
             asm_gemm->set_working_space(reinterpret_cast<typename T::TypeResult *>(workspace.buffer()));
         }
-        const unsigned int window_size = asm_gemm->get_window_size();
-        if(window_size < num_threads)
+
+        //if we disable this code below in brackets then ConvLayer deadlocks when threads > 1 and
+        //the shapes are In=1x1x1024 Weights=1x1x1024x1001 Biases=1001 Out=1x1x1001
         {
-            num_threads = window_size;
-            asm_gemm->set_nthreads(num_threads);
+            const unsigned int window_size = asm_gemm->get_window_size();
+            if(window_size < num_threads)
+            {
+                num_threads = window_size;
+                asm_gemm->set_nthreads(num_threads);
+            }
         }
+
         asm_glue._gemm_kernel_asm  = std::move(asm_gemm);
         asm_glue._optimised_kernel = std::move(acl_gemm_wrapper);
         // We need to setup the ptrs in the run() method