let control core binding

example/cpu_02_conv2d_fwd_bias_relu_add/cpu_conv2d_fwd_bias_relu_add.cpp

@@ -16,8 +16,8 @@
 
 #define TEST_LAYOUT_NHWC_KYXC_NHWK 0
 #define TEST_LAYOUT_NHWC_KYXCK8_NHWK 1
-#define TEST_LAYOUT_NHWC_YXCK_NHWK 1
-#define TEST_LAYOUT TEST_LAYOUT_NHWC_KYXC_NHWK
+#define TEST_LAYOUT_NHWC_YXCK_NHWK 2
+#define TEST_LAYOUT TEST_LAYOUT_NHWC_YXCK_NHWK
 
 using F32 = float;
 using F16 = ck::half_t;

include/ck/tensor_operation/cpu/grid/gridwise_gemm_avx2.hpp

@@ -9,6 +9,7 @@
 #include "threadwise_tensor_slice_transfer_avx2.hpp"
 #include "threadwise_tensor_slice_transfer_avx2_specialization.hpp"
 #include "dynamic_buffer_cpu.hpp"
+#include "envvar.hpp"
 #include <utility>
 #include <unistd.h>
 #include <omp.h>
@@ -303,8 +304,8 @@ struct GridwiseGemmAvx2_MxN
 
         int total_threads = omp_get_max_threads();
 
-#if 0
-        if(total_threads > 1){
+        if(total_threads > 1 && ck::getenv_int("CK_CPU_BIND_CORE", 1) != 0)
+        {
 #pragma omp parallel
             {
                 int tid = omp_get_thread_num();
@@ -313,12 +314,12 @@ struct GridwiseGemmAvx2_MxN
 
                 CPU_SET(tid, &set);
 
-                if (sched_setaffinity(0, sizeof(set), &set) == -1) {
+                if(sched_setaffinity(0, sizeof(set), &set) == -1)
+                {
                     throw std::runtime_error("wrong! fail to set thread affinity");
                 }
             }
         }
-#endif
 
         // TODO: openmp aware ordering
         //

include/ck/tensor_operation/cpu/grid/gridwise_gemm_bias_activation_add_avx2.hpp

@@ -9,6 +9,7 @@
 #include "threadwise_tensor_slice_transfer_avx2.hpp"
 #include "threadwise_tensor_slice_transfer_avx2_specialization.hpp"
 #include "dynamic_buffer_cpu.hpp"
+#include "envvar.hpp"
 #include <utility>
 #include <unistd.h>
 #include <omp.h>
@@ -329,8 +330,8 @@ struct GridwiseGemmBiasActivationAddAvx2_MxN
 
         int total_threads = omp_get_max_threads();
 
-#if 0
-        if(total_threads > 1){
+        if(total_threads > 1 && ck::getenv_int("CK_CPU_BIND_CORE", 1) != 0)
+        {
 #pragma omp parallel
             {
                 int tid = omp_get_thread_num();
@@ -339,12 +340,12 @@ struct GridwiseGemmBiasActivationAddAvx2_MxN
 
                 CPU_SET(tid, &set);
 
-                if (sched_setaffinity(0, sizeof(set), &set) == -1) {
+                if(sched_setaffinity(0, sizeof(set), &set) == -1)
+                {
                     throw std::runtime_error("wrong! fail to set thread affinity");
                 }
             }
         }
-#endif
 
         // TODO: openmp aware ordering
         //

include/ck/utility/envvar.hpp

+#ifndef __ENVVAR_HPP
+#define __ENVVAR_HPP
+
+#include <cstdlib>
+
+namespace ck {
+
+static inline int getenv_int(const char* var_name, int default_int)
+{
+    char* v = ::getenv(var_name);
+    int r   = default_int;
+    if(v)
+        r = ::atoi(v);
+    return r;
+}
+
+static inline char* getenv_str(const char* var_name, char* default_str)
+{
+    char* v = ::getenv(var_name);
+    if(v)
+        return v;
+    return default_str;
+}
+
+} // namespace ck
+#endif