Merge pull request #13421 from seiko2plus/core_improve_infa_runtime

ENH: improve runtime detection of CPU features

10 files changed, 643 insertions, 114 deletions

diff --git a/numpy/core/code_generators/generate_umath.py b/numpy/core/code_generators/generate_umath.py
index 1fd08241d..f22380ccb 100644
--- a/numpy/core/code_generators/generate_umath.py
+++ b/numpy/core/code_generators/generate_umath.py
@@ -1014,7 +1014,7 @@ def make_arrays(funcdict):
                     for vt in t.simd:
                         code2list.append(textwrap.dedent("""\
                         #ifdef HAVE_ATTRIBUTE_TARGET_{ISA}
-                        if (npy_cpu_supports("{isa}")) {{
+                        if (NPY_CPU_HAVE({ISA})) {{
                             {fname}_functions[{idx}] = {type}_{fname}_{isa};
                         }}
                         #endif
@@ -1138,7 +1138,6 @@ def make_code(funcdict, filename):
 
         Please make changes to the code generator program (%s)
     **/
-    #include "cpuid.h"
     #include "ufunc_object.h"
     #include "ufunc_type_resolution.h"
     #include "loops.h"
diff --git a/numpy/core/setup.py b/numpy/core/setup.py
index 66c1b782e..e15cbf7c2 100644
--- a/numpy/core/setup.py
+++ b/numpy/core/setup.py
@@ -745,6 +745,7 @@ def configuration(parent_package='',top_path=None):
             join('src', 'common', 'ucsnarrow.c'),
             join('src', 'common', 'ufunc_override.c'),
             join('src', 'common', 'numpyos.c'),
+            join('src', 'common', 'npy_cpu_features.c.src'),
             ]
 
     if os.environ.get('NPY_USE_BLAS_ILP64', "0") != "0":
@@ -898,7 +899,6 @@ def configuration(parent_package='',top_path=None):
             join('src', 'umath', 'clip.c.src'),
             join('src', 'umath', 'ufunc_object.c'),
             join('src', 'umath', 'extobj.c'),
-            join('src', 'umath', 'cpuid.c'),
             join('src', 'umath', 'scalarmath.c.src'),
             join('src', 'umath', 'ufunc_type_resolution.c'),
             join('src', 'umath', 'override.c'),
diff --git a/numpy/core/setup_common.py b/numpy/core/setup_common.py
index 7cac66e61..63c4a76a9 100644
--- a/numpy/core/setup_common.py
+++ b/numpy/core/setup_common.py
@@ -132,11 +132,6 @@ OPTIONAL_INTRINSICS = [("__builtin_isnan", '5.'),
                        ("__builtin_bswap64", '5u'),
                        ("__builtin_expect", '5, 0'),
                        ("__builtin_mul_overflow", '5, 5, (int*)5'),
-                       # broken on OSX 10.11, make sure its not optimized away
-                       ("volatile int r = __builtin_cpu_supports", '"sse"',
-                        "stdio.h", "__BUILTIN_CPU_SUPPORTS"),
-                       ("volatile int r = __builtin_cpu_supports", '"avx512f"',
-                        "stdio.h", "__BUILTIN_CPU_SUPPORTS_AVX512F"),
                        # MMX only needed for icc, but some clangs don't have it
                        ("_m_from_int64", '0', "emmintrin.h"),
                        ("_mm_load_ps", '(float*)0', "xmmintrin.h"),  # SSE
diff --git a/numpy/core/src/common/npy_config.h b/numpy/core/src/common/npy_config.h
index eedfbe364..aebe241a5 100644
--- a/numpy/core/src/common/npy_config.h
+++ b/numpy/core/src/common/npy_config.h
@@ -2,6 +2,7 @@
 #define _NPY_NPY_CONFIG_H_
 
 #include "config.h"
+#include "npy_cpu_features.h"
 #include "numpy/numpyconfig.h"
 #include "numpy/npy_cpu.h"
 #include "numpy/npy_os.h"
diff --git a/numpy/core/src/common/npy_cpu_features.c.src b/numpy/core/src/common/npy_cpu_features.c.src
new file mode 100644
index 000000000..cbd99827b
--- /dev/null
+++ b/numpy/core/src/common/npy_cpu_features.c.src
@@ -0,0 +1,404 @@
+#include "npy_cpu_features.h"
+#include "numpy/npy_common.h" // for NPY_INLINE
+#include "numpy/npy_cpu.h" // To guarantee of having CPU definitions in scope.
+
+/******************** Private Definitions *********************/
+
+// Hold all CPU features boolean values
+static unsigned char npy__cpu_have[NPY_CPU_FEATURE_MAX];
+
+/******************** Private Declarations *********************/
+
+// Almost detect all CPU features in runtime
+static void
+npy__cpu_init_features(void);
+
+/******************** Public Definitions *********************/
+
+NPY_VISIBILITY_HIDDEN int
+npy_cpu_have(int feature_id)
+{
+    if (feature_id <= NPY_CPU_FEATURE_NONE || feature_id >= NPY_CPU_FEATURE_MAX)
+        return 0;
+    return npy__cpu_have[feature_id];
+}
+
+NPY_VISIBILITY_HIDDEN int
+npy_cpu_init(void)
+{
+    npy__cpu_init_features();
+    return 0;
+}
+
+NPY_VISIBILITY_HIDDEN PyObject *
+npy_cpu_features_dict(void)
+{
+    PyObject *dict = PyDict_New();
+    if (dict) {
+    /**begin repeat
+     * #feature = MMX, SSE, SSE2, SSE3, SSSE3, SSE41, POPCNT, SSE42,
+     *            AVX, F16C, XOP, FMA4, FMA3, AVX2, AVX512F,
+     *            AVX512CD, AVX512ER, AVX512PF, AVX5124FMAPS, AVX5124VNNIW,
+     *            AVX512VPOPCNTDQ, AVX512VL, AVX512BW, AVX512DQ, AVX512VNNI,
+     *            AVX512IFMA, AVX512VBMI, AVX512VBMI2, AVX512BITALG,
+     *            AVX512_KNL, AVX512_KNM, AVX512_SKX, AVX512_CLX, AVX512_CNL, AVX512_ICL,
+     *            VSX, VSX2, VSX3,
+     *            NEON, NEON_FP16, NEON_VFPV4, ASIMD, FPHP, ASIMDHP, ASIMDDP, ASIMDFHM#
+    */
+        if (PyDict_SetItemString(dict, "@feature@",
+            npy__cpu_have[NPY_CPU_FEATURE_@feature@] ? Py_True : Py_False) < 0) {
+            Py_DECREF(dict);
+            return NULL;
+        }
+    /**end repeat**/
+    }
+    return dict;
+}
+
+/****************************************************************
+ * This section is reserved to defining @npy__cpu_init_features
+ * for each CPU architecture, please try to keep it clean. Ty
+ ****************************************************************/
+
+/***************** X86 ******************/
+
+#if defined(NPY_CPU_AMD64) || defined(NPY_CPU_X86)
+
+#ifdef _MSC_VER
+    #include <intrin.h>
+#elif defined(__INTEL_COMPILER)
+    #include <immintrin.h>
+#endif
+
+static int
+npy__cpu_getxcr0(void)
+{
+#if defined(_MSC_VER) || defined (__INTEL_COMPILER)
+    return _xgetbv(0);
+#elif defined(__GNUC__) || defined(__clang__)
+    unsigned int eax, edx;
+    __asm__("xgetbv" : "=a" (eax), "=d" (edx) : "c" (0));
+    return (eax | (unsigned long long)edx << 32);
+#else
+    // TODO: handle other x86 compilers
+    return 0;
+#endif
+}
+
+static void
+npy__cpu_cpuid(int reg[4], int func_id)
+{
+#if defined(_MSC_VER)
+    __cpuidex(reg, func_id, 0);
+#elif defined(__INTEL_COMPILER)
+    __cpuid(reg, func_id);
+#elif defined(__GNUC__) || defined(__clang__)
+    #if defined(NPY_CPU_X86) && defined(__PIC__)
+    // %ebx may be the PIC register
+        #define NPY__CPUID_ASM \
+            "xchg{l}\t{%%}ebx, %1\n\t" \
+            "cpuid\n\t"                \
+            "xchg{l}\t{%%}ebx, %1\n\t"
+    #else
+        #define NPY__CPUID_ASM "cpuid"
+    #endif
+    __asm__(NPY__CPUID_ASM : "=a" (reg[0]), "=b" (reg[1]), "=c" (reg[2]), "=d" (reg[3]) : "a" (func_id), "c" (0) : );
+#else
+    // TODO: handle other x86 compilers
+    reg[0] = 0;
+#endif
+}
+
+static void
+npy__cpu_init_features(void)
+{
+    memset(npy__cpu_have, 0, sizeof(npy__cpu_have[0]) * NPY_CPU_FEATURE_MAX);
+
+    // validate platform support
+    int reg[] = {0, 0, 0, 0};
+    npy__cpu_cpuid(reg, 0);
+    if (reg[0] == 0)
+        return;
+
+    npy__cpu_cpuid(reg, 1);
+    npy__cpu_have[NPY_CPU_FEATURE_MMX]    = (reg[3] & (1 << 23)) != 0;
+    npy__cpu_have[NPY_CPU_FEATURE_SSE]    = (reg[3] & (1 << 25)) != 0;
+    npy__cpu_have[NPY_CPU_FEATURE_SSE2]   = (reg[3] & (1 << 26)) != 0;
+    npy__cpu_have[NPY_CPU_FEATURE_SSE3]   = (reg[2] & (1 << 0))  != 0;
+    npy__cpu_have[NPY_CPU_FEATURE_SSSE3]  = (reg[2] & (1 << 9))  != 0;
+    npy__cpu_have[NPY_CPU_FEATURE_SSE41]  = (reg[2] & (1 << 19)) != 0;
+    npy__cpu_have[NPY_CPU_FEATURE_POPCNT] = (reg[2] & (1 << 23)) != 0;
+    npy__cpu_have[NPY_CPU_FEATURE_SSE42]  = (reg[2] & (1 << 20)) != 0;
+    npy__cpu_have[NPY_CPU_FEATURE_F16C]   = (reg[2] & (1 << 29)) != 0;
+
+    // check OSXSAVE
+    if ((reg[2] & (1 << 27)) == 0)
+        return;
+    // check AVX OS support
+    int xcr = npy__cpu_getxcr0();
+    if ((xcr & 6) != 6)
+        return;
+    npy__cpu_have[NPY_CPU_FEATURE_AVX]    = (reg[2] & (1 << 28)) != 0;
+    if (!npy__cpu_have[NPY_CPU_FEATURE_AVX])
+        return;
+    npy__cpu_have[NPY_CPU_FEATURE_FMA3]   = (reg[2] & (1 << 12)) != 0;
+
+    // second call to the cpuid to get extended AMD feature bits
+    npy__cpu_cpuid(reg, 0x80000001);
+    npy__cpu_have[NPY_CPU_FEATURE_XOP]    = (reg[2] & (1 << 11)) != 0;
+    npy__cpu_have[NPY_CPU_FEATURE_FMA4]   = (reg[2] & (1 << 16)) != 0;
+
+    // third call to the cpuid to get extended AVX2 & AVX512 feature bits
+    npy__cpu_cpuid(reg, 7);
+    npy__cpu_have[NPY_CPU_FEATURE_AVX2]   = (reg[1] & (1 << 5))  != 0;
+    if (!npy__cpu_have[NPY_CPU_FEATURE_AVX2])
+        return;
+    // detect AVX2 & FMA3
+    npy__cpu_have[NPY_CPU_FEATURE_FMA]    = npy__cpu_have[NPY_CPU_FEATURE_FMA3];
+
+    // check AVX512 OS support
+    if ((xcr & 0xe6) != 0xe6)
+        return;
+    npy__cpu_have[NPY_CPU_FEATURE_AVX512F]  = (reg[1] & (1 << 16)) != 0;
+    npy__cpu_have[NPY_CPU_FEATURE_AVX512CD] = (reg[1] & (1 << 28)) != 0;
+    if (npy__cpu_have[NPY_CPU_FEATURE_AVX512F] && npy__cpu_have[NPY_CPU_FEATURE_AVX512CD]) {
+        // Knights Landing
+        npy__cpu_have[NPY_CPU_FEATURE_AVX512PF]        = (reg[1] & (1 << 26)) != 0;
+        npy__cpu_have[NPY_CPU_FEATURE_AVX512ER]        = (reg[1] & (1 << 27)) != 0;
+        npy__cpu_have[NPY_CPU_FEATURE_AVX512_KNL]      = npy__cpu_have[NPY_CPU_FEATURE_AVX512ER] &&
+                                                         npy__cpu_have[NPY_CPU_FEATURE_AVX512PF];
+        // Knights Mill
+        npy__cpu_have[NPY_CPU_FEATURE_AVX512VPOPCNTDQ] = (reg[2] & (1 << 14)) != 0;
+        npy__cpu_have[NPY_CPU_FEATURE_AVX5124VNNIW]    = (reg[3] & (1 << 2))  != 0;
+        npy__cpu_have[NPY_CPU_FEATURE_AVX5124FMAPS]    = (reg[3] & (1 << 3))  != 0;
+        npy__cpu_have[NPY_CPU_FEATURE_AVX512_KNM]      = npy__cpu_have[NPY_CPU_FEATURE_AVX512_KNL] &&
+                                                         npy__cpu_have[NPY_CPU_FEATURE_AVX5124FMAPS] &&
+                                                         npy__cpu_have[NPY_CPU_FEATURE_AVX5124VNNIW] &&
+                                                         npy__cpu_have[NPY_CPU_FEATURE_AVX512VPOPCNTDQ];
+
+        // Skylake-X
+        npy__cpu_have[NPY_CPU_FEATURE_AVX512DQ]        = (reg[1] & (1 << 17)) != 0;
+        npy__cpu_have[NPY_CPU_FEATURE_AVX512BW]        = (reg[1] & (1 << 30)) != 0;
+        npy__cpu_have[NPY_CPU_FEATURE_AVX512VL]        = (reg[1] & (1 << 31)) != 0;
+        npy__cpu_have[NPY_CPU_FEATURE_AVX512_SKX]      = npy__cpu_have[NPY_CPU_FEATURE_AVX512BW] &&
+                                                         npy__cpu_have[NPY_CPU_FEATURE_AVX512DQ] &&
+                                                         npy__cpu_have[NPY_CPU_FEATURE_AVX512VL];
+        // Cascade Lake
+        npy__cpu_have[NPY_CPU_FEATURE_AVX512VNNI]      = (reg[2] & (1 << 11)) != 0;
+        npy__cpu_have[NPY_CPU_FEATURE_AVX512_CLX]      = npy__cpu_have[NPY_CPU_FEATURE_AVX512_SKX] &&
+                                                         npy__cpu_have[NPY_CPU_FEATURE_AVX512VNNI];
+
+        // Cannon Lake
+        npy__cpu_have[NPY_CPU_FEATURE_AVX512IFMA]      = (reg[1] & (1 << 21)) != 0;
+        npy__cpu_have[NPY_CPU_FEATURE_AVX512VBMI]      = (reg[2] & (1 << 1))  != 0;
+        npy__cpu_have[NPY_CPU_FEATURE_AVX512_CNL]      = npy__cpu_have[NPY_CPU_FEATURE_AVX512_SKX] &&
+                                                         npy__cpu_have[NPY_CPU_FEATURE_AVX512IFMA] &&
+                                                         npy__cpu_have[NPY_CPU_FEATURE_AVX512VBMI];
+        // Ice Lake
+        npy__cpu_have[NPY_CPU_FEATURE_AVX512VBMI2]     = (reg[2] & (1 << 6))  != 0;
+        npy__cpu_have[NPY_CPU_FEATURE_AVX512BITALG]    = (reg[2] & (1 << 12)) != 0;
+        npy__cpu_have[NPY_CPU_FEATURE_AVX512_ICL]      = npy__cpu_have[NPY_CPU_FEATURE_AVX512_CLX] &&
+                                                         npy__cpu_have[NPY_CPU_FEATURE_AVX512_CNL] &&
+                                                         npy__cpu_have[NPY_CPU_FEATURE_AVX512VBMI2] &&
+                                                         npy__cpu_have[NPY_CPU_FEATURE_AVX512BITALG] &&
+                                                         npy__cpu_have[NPY_CPU_FEATURE_AVX512VPOPCNTDQ];
+    }
+}
+
+/***************** POWER ******************/
+
+#elif defined(NPY_CPU_PPC64) || defined(NPY_CPU_PPC64LE)
+
+#ifdef __linux__
+    #include <sys/auxv.h>
+    #ifndef AT_HWCAP2
+        #define AT_HWCAP2 26
+    #endif
+    #ifndef PPC_FEATURE2_ARCH_3_00
+        #define PPC_FEATURE2_ARCH_3_00 0x00800000
+    #endif
+#endif
+
+static void
+npy__cpu_init_features(void)
+{
+    memset(npy__cpu_have, 0, sizeof(npy__cpu_have[0]) * NPY_CPU_FEATURE_MAX);
+#ifdef __linux__
+    unsigned int hwcap = getauxval(AT_HWCAP);
+    if ((hwcap & PPC_FEATURE_HAS_VSX) == 0)
+        return;
+
+    hwcap = getauxval(AT_HWCAP2);
+    if (hwcap & PPC_FEATURE2_ARCH_3_00)
+    {
+        npy__cpu_have[NPY_CPU_FEATURE_VSX]  =
+        npy__cpu_have[NPY_CPU_FEATURE_VSX2] =
+        npy__cpu_have[NPY_CPU_FEATURE_VSX3] = 1;
+        return;
+    }
+    npy__cpu_have[NPY_CPU_FEATURE_VSX2] = (hwcap & PPC_FEATURE2_ARCH_2_07) != 0;
+    npy__cpu_have[NPY_CPU_FEATURE_VSX]  = 1;
+// TODO: AIX, FreeBSD
+#else
+    npy__cpu_have[NPY_CPU_FEATURE_VSX]  = 1;
+    #if defined(NPY_CPU_PPC64LE) || defined(NPY_HAVE_VSX2)
+    npy__cpu_have[NPY_CPU_FEATURE_VSX2] = 1;
+    #endif
+    #ifdef NPY_HAVE_VSX3
+    npy__cpu_have[NPY_CPU_FEATURE_VSX3] = 1;
+    #endif
+#endif
+}
+
+/***************** ARM ******************/
+
+#elif defined(__arm__) || defined(__aarch64__)
+
+static NPY_INLINE void
+npy__cpu_init_features_arm8(void)
+{
+    npy__cpu_have[NPY_CPU_FEATURE_NEON]       =
+    npy__cpu_have[NPY_CPU_FEATURE_NEON_FP16]  =
+    npy__cpu_have[NPY_CPU_FEATURE_NEON_VFPV4] =
+    npy__cpu_have[NPY_CPU_FEATURE_ASIMD]      = 1;
+}
+
+#ifdef __linux__
+/*
+ * we aren't sure of what kind kernel or clib we deal with
+ * so we play it safe
+*/
+#include <stdio.h>
+#include <fcntl.h>
+
+#define NPY__HWCAP  16
+#define NPY__HWCAP2 26
+
+// arch/arm/include/uapi/asm/hwcap.h
+#define NPY__HWCAP_HALF   (1 << 1)
+#define NPY__HWCAP_NEON   (1 << 12)
+#define NPY__HWCAP_VFPv3  (1 << 13)
+#define NPY__HWCAP_VFPv4  (1 << 16)
+#define NPY__HWCAP2_AES   (1 << 0)
+#define NPY__HWCAP2_PMULL (1 << 1)
+#define NPY__HWCAP2_SHA1  (1 << 2)
+#define NPY__HWCAP2_SHA2  (1 << 3)
+#define NPY__HWCAP2_CRC32 (1 << 4)
+// arch/arm64/include/uapi/asm/hwcap.h
+#define NPY__HWCAP_FP       (1 << 0)
+#define NPY__HWCAP_ASIMD    (1 << 1)
+#define NPY__HWCAP_FPHP     (1 << 9)
+#define NPY__HWCAP_ASIMDHP  (1 << 10)
+#define NPY__HWCAP_ASIMDDP  (1 << 20)
+#define NPY__HWCAP_ASIMDFHM (1 << 23)
+
+__attribute__((weak)) unsigned long getauxval(unsigned long); // linker should handle it
+static int
+npy__cpu_init_features_linux(void)
+{
+    unsigned long hwcap = 0, hwcap2 = 0;
+    if (getauxval != 0) {
+        hwcap = getauxval(NPY__HWCAP);
+    #ifdef __arm__
+        hwcap2 = getauxval(NPY__HWCAP2);
+    #endif
+    } else {
+        unsigned long auxv[2];
+        int fd = open("/proc/self/auxv", O_RDONLY);
+        if (fd >= 0) {
+            while (read(fd, &auxv, sizeof(auxv)) == sizeof(auxv)) {
+                if (auxv[0] == NPY__HWCAP) {
+                    hwcap = auxv[1];
+                }
+            #ifdef __arm__
+                else if (auxv[0] == NPY__HWCAP2) {
+                    hwcap2 = auxv[1];
+                }
+            #endif
+                // detect the end
+                else if (auxv[0] == 0 && auxv[1] == 0) {
+                    break;
+                }
+            }
+            close(fd);
+        }
+    }
+    if (hwcap == 0 && hwcap2 == 0) {
+        /*
+         * FIXME: failback to compiler definitions,
+         * BTW we can parse /proc/cpuinfo for badly patched kernels
+        */
+        return 0;
+    }
+#ifdef __arm__
+    // Detect Arm8 (aarch32 state)
+    if ((hwcap2 & NPY__HWCAP2_AES)  || (hwcap2 & NPY__HWCAP2_SHA1)  ||
+        (hwcap2 & NPY__HWCAP2_SHA2) || (hwcap2 & NPY__HWCAP2_PMULL) ||
+        (hwcap2 & NPY__HWCAP2_CRC32))
+#else
+    if (1)
+#endif
+    {
+        if (!(hwcap & (NPY__HWCAP_FP | NPY__HWCAP_ASIMD))) {
+            // Is this could happen? maybe disabled by kernel
+            // BTW this will break the baseline of AARCH64
+            return 1;
+        }
+        npy__cpu_have[NPY_CPU_FEATURE_FPHP]       = (hwcap & NPY__HWCAP_FPHP)     != 0;
+        npy__cpu_have[NPY_CPU_FEATURE_ASIMDHP]    = (hwcap & NPY__HWCAP_ASIMDHP)  != 0;
+        npy__cpu_have[NPY_CPU_FEATURE_ASIMDDP]    = (hwcap & NPY__HWCAP_ASIMDDP)  != 0;
+        npy__cpu_have[NPY_CPU_FEATURE_ASIMDFHM]   = (hwcap & NPY__HWCAP_ASIMDFHM) != 0;
+        npy__cpu_init_features_arm8();
+    } else {
+        npy__cpu_have[NPY_CPU_FEATURE_NEON]       = (hwcap & NPY__HWCAP_NEON)   != 0;
+        npy__cpu_have[NPY_CPU_FEATURE_NEON_FP16]  = (hwcap & (NPY__HWCAP_NEON | NPY__HWCAP_VFPv3 |
+                                                              NPY__HWCAP_HALF)) != 0;
+        npy__cpu_have[NPY_CPU_FEATURE_NEON_VFPV4] = (hwcap & (NPY__HWCAP_NEON | NPY__HWCAP_VFPv4)) != 0;
+    }
+    return 1;
+}
+#endif
+
+static void
+npy__cpu_init_features(void)
+{
+    memset(npy__cpu_have, 0, sizeof(npy__cpu_have[0]) * NPY_CPU_FEATURE_MAX);
+#ifdef __linux__
+    if (npy__cpu_init_features_linux())
+        return;
+#endif
+    // We have nothing else todo
+#if defined(NPY_HAVE_NEON_ARM8) || defined(__aarch64__) || (defined(__ARM_ARCH) && __ARM_ARCH >= 8)
+    #if defined(NPY_HAVE_FPHP) || defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+    npy__cpu_have[NPY_CPU_FEATURE_FPHP] = 1;
+    #endif
+    #if defined(NPY_HAVE_ASIMDHP) || defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+    npy__cpu_have[NPY_CPU_FEATURE_ASIMDHP] = 1;
+    #endif
+    #if defined(NPY_HAVE_ASIMDDP) || defined(__ARM_FEATURE_DOTPROD)
+    npy__cpu_have[NPY_CPU_FEATURE_ASIMDDP] = 1;
+    #endif
+    #if defined(NPY_HAVE_ASIMDFHM) || defined(__ARM_FEATURE_FP16FML)
+    npy__cpu_have[NPY_CPU_FEATURE_ASIMDFHM] = 1;
+    #endif
+    npy__cpu_init_features_arm8();
+#else
+    #if defined(NPY_HAVE_NEON) || defined(__ARM_NEON__)
+        npy__cpu_have[NPY_CPU_FEATURE_NEON] = 1;
+    #endif
+    #if defined(NPY_HAVE_NEON_FP16) || defined(__ARM_FP16_FORMAT_IEEE) || (defined(__ARM_FP) && (__ARM_FP & 2))
+        npy__cpu_have[NPY_CPU_FEATURE_NEON_FP16] = npy__cpu_have[NPY_CPU_FEATURE_NEON];
+    #endif
+    #if defined(NPY_HAVE_NEON_VFPV4) || defined(__ARM_FEATURE_FMA)
+        npy__cpu_have[NPY_CPU_FEATURE_NEON_VFPV4] = npy__cpu_have[NPY_CPU_FEATURE_NEON];
+    #endif
+#endif
+}
+
+/*********** Unsupported ARCH ***********/
+#else
+static void
+npy__cpu_init_features(void)
+{
+}
+#endif
diff --git a/numpy/core/src/common/npy_cpu_features.h b/numpy/core/src/common/npy_cpu_features.h
new file mode 100644
index 000000000..0e8901328
--- /dev/null
+++ b/numpy/core/src/common/npy_cpu_features.h
@@ -0,0 +1,117 @@
+#ifndef _NPY_CPU_FEATURES_H_
+#define _NPY_CPU_FEATURES_H_
+
+#include "numpy/numpyconfig.h" // for NPY_VISIBILITY_HIDDEN
+#include <Python.h> // for PyObject
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+enum npy_cpu_features
+{
+    NPY_CPU_FEATURE_NONE = 0,
+    // X86
+    NPY_CPU_FEATURE_MMX               = 1,
+    NPY_CPU_FEATURE_SSE               = 2,
+    NPY_CPU_FEATURE_SSE2              = 3,
+    NPY_CPU_FEATURE_SSE3              = 4,
+    NPY_CPU_FEATURE_SSSE3             = 5,
+    NPY_CPU_FEATURE_SSE41             = 6,
+    NPY_CPU_FEATURE_POPCNT            = 7,
+    NPY_CPU_FEATURE_SSE42             = 8,
+    NPY_CPU_FEATURE_AVX               = 9,
+    NPY_CPU_FEATURE_F16C              = 10,
+    NPY_CPU_FEATURE_XOP               = 11,
+    NPY_CPU_FEATURE_FMA4              = 12,
+    NPY_CPU_FEATURE_FMA3              = 13,
+    NPY_CPU_FEATURE_AVX2              = 14,
+    NPY_CPU_FEATURE_FMA               = 15, // AVX2 & FMA3, provides backward compatibility
+
+    NPY_CPU_FEATURE_AVX512F           = 30,
+    NPY_CPU_FEATURE_AVX512CD          = 31,
+    NPY_CPU_FEATURE_AVX512ER          = 32,
+    NPY_CPU_FEATURE_AVX512PF          = 33,
+    NPY_CPU_FEATURE_AVX5124FMAPS      = 34,
+    NPY_CPU_FEATURE_AVX5124VNNIW      = 35,
+    NPY_CPU_FEATURE_AVX512VPOPCNTDQ   = 36,
+    NPY_CPU_FEATURE_AVX512BW          = 37,
+    NPY_CPU_FEATURE_AVX512DQ          = 38,
+    NPY_CPU_FEATURE_AVX512VL          = 39,
+    NPY_CPU_FEATURE_AVX512IFMA        = 40,
+    NPY_CPU_FEATURE_AVX512VBMI        = 41,
+    NPY_CPU_FEATURE_AVX512VNNI        = 42,
+    NPY_CPU_FEATURE_AVX512VBMI2       = 43,
+    NPY_CPU_FEATURE_AVX512BITALG      = 44,
+
+    // X86 CPU Groups
+    // Knights Landing (F,CD,ER,PF)
+    NPY_CPU_FEATURE_AVX512_KNL        = 101,
+    // Knights Mill    (F,CD,ER,PF,4FMAPS,4VNNIW,VPOPCNTDQ)
+    NPY_CPU_FEATURE_AVX512_KNM        = 102,
+    // Skylake-X       (F,CD,BW,DQ,VL)
+    NPY_CPU_FEATURE_AVX512_SKX        = 103,
+    // Cascade Lake    (F,CD,BW,DQ,VL,VNNI)
+    NPY_CPU_FEATURE_AVX512_CLX        = 104,
+    // Cannon Lake     (F,CD,BW,DQ,VL,IFMA,VBMI)
+    NPY_CPU_FEATURE_AVX512_CNL        = 105,
+    // Ice Lake        (F,CD,BW,DQ,VL,IFMA,VBMI,VNNI,VBMI2,BITALG,VPOPCNTDQ)
+    NPY_CPU_FEATURE_AVX512_ICL        = 106,
+
+    // IBM/POWER VSX
+    // POWER7
+    NPY_CPU_FEATURE_VSX               = 200,
+    // POWER8
+    NPY_CPU_FEATURE_VSX2              = 201,
+    // POWER9
+    NPY_CPU_FEATURE_VSX3              = 202,
+
+    // ARM
+    NPY_CPU_FEATURE_NEON              = 300,
+    NPY_CPU_FEATURE_NEON_FP16         = 301,
+    // FMA
+    NPY_CPU_FEATURE_NEON_VFPV4        = 302,
+    // Advanced SIMD
+    NPY_CPU_FEATURE_ASIMD             = 303,
+    // ARMv8.2 half-precision
+    NPY_CPU_FEATURE_FPHP              = 304,
+    // ARMv8.2 half-precision vector arithm
+    NPY_CPU_FEATURE_ASIMDHP           = 305,
+    // ARMv8.2 dot product
+    NPY_CPU_FEATURE_ASIMDDP           = 306,
+    // ARMv8.2 single&half-precision multiply
+    NPY_CPU_FEATURE_ASIMDFHM          = 307,
+
+    NPY_CPU_FEATURE_MAX
+};
+
+/*
+ * Initialize CPU features
+ * return 0 on success otherwise return -1
+*/
+NPY_VISIBILITY_HIDDEN int
+npy_cpu_init(void);
+
+/*
+ * return 0 if CPU feature isn't available
+ * note: `npy_cpu_init` must be called first otherwise it will always return 0
+*/
+NPY_VISIBILITY_HIDDEN int
+npy_cpu_have(int feature_id);
+
+#define NPY_CPU_HAVE(FEATURE_NAME) \
+npy_cpu_have(NPY_CPU_FEATURE_##FEATURE_NAME)
+
+/*
+ * return a new dictionary contains CPU feature names
+ * with runtime availability.
+ * same as npy_cpu_have, `npy_cpu_init` must be called first.
+ */
+NPY_VISIBILITY_HIDDEN PyObject *
+npy_cpu_features_dict(void);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif // _NPY_CPU_FEATURES_H_
diff --git a/numpy/core/src/multiarray/multiarraymodule.c b/numpy/core/src/multiarray/multiarraymodule.c
index c2e597385..7792fcdcb 100644
--- a/numpy/core/src/multiarray/multiarraymodule.c
+++ b/numpy/core/src/multiarray/multiarraymodule.c
@@ -4388,6 +4388,11 @@ PyMODINIT_FUNC PyInit__multiarray_umath(void) {
     PyObject *m, *d, *s;
     PyObject *c_api;
 
+    /* Initialize CPU features */
+    if (npy_cpu_init() < 0) {
+        goto err;
+    }
+
     /* Create the module and add the functions */
     m = PyModule_Create(&moduledef);
     if (!m) {
@@ -4513,6 +4518,16 @@ PyMODINIT_FUNC PyInit__multiarray_umath(void) {
     PyDict_SetItemString(d, "__version__", s);
     Py_DECREF(s);
 
+    s = npy_cpu_features_dict();
+    if (s == NULL) {
+        goto err;
+    }
+    if (PyDict_SetItemString(d, "__cpu_features__", s) < 0) {
+        Py_DECREF(s);
+        goto err;
+    }
+    Py_DECREF(s);
+
     s = NpyCapsule_FromVoidPtr((void *)_datetime_strings, NULL);
     if (s == NULL) {
         goto err;
diff --git a/numpy/core/src/umath/cpuid.c b/numpy/core/src/umath/cpuid.c
deleted file mode 100644
index 72c6493e8..000000000
--- a/numpy/core/src/umath/cpuid.c
+++ /dev/null
@@ -1,97 +0,0 @@
-#define _UMATHMODULE
-#define _MULTIARRAYMODULE
-#define NPY_NO_DEPRECATED_API NPY_API_VERSION
-
-#include <Python.h>
-
-#include "npy_config.h"
-
-#include "cpuid.h"
-
-#define XCR_XFEATURE_ENABLED_MASK 0x0
-#define XSTATE_SSE 0x2
-#define XSTATE_YMM 0x4
-#define XSTATE_ZMM 0x70
-
-/*
- * verify the OS supports avx instructions
- * it can be disabled in some OS, e.g. with the nosavex boot option of linux
- */
-static NPY_INLINE
-int os_avx_support(void)
-{
-#if HAVE_XGETBV
-    /*
-     * use bytes for xgetbv to avoid issues with compiler not knowing the
-     * instruction
-     */
-    unsigned int eax, edx;
-    unsigned int ecx = XCR_XFEATURE_ENABLED_MASK;
-    __asm__("xgetbv" : "=a" (eax), "=d" (edx) : "c" (ecx));
-    return (eax & (XSTATE_SSE | XSTATE_YMM)) == (XSTATE_SSE | XSTATE_YMM);
-#else
-    return 0;
-#endif
-}
-
-static NPY_INLINE
-int os_avx512_support(void)
-{
-#if HAVE_XGETBV
-    unsigned int eax, edx;
-    unsigned int ecx = XCR_XFEATURE_ENABLED_MASK;
-    unsigned int xcr0 = XSTATE_ZMM | XSTATE_YMM | XSTATE_SSE;
-    __asm__("xgetbv" : "=a" (eax), "=d" (edx) : "c" (ecx));
-    return (eax & xcr0) == xcr0;
-#else
-    return 0;
-#endif
-}
-
-static NPY_INLINE
-int cpu_supports_fma(void)
-{
-#ifdef __x86_64__
-    unsigned int feature = 0x01;
-    unsigned int a, b, c, d;
-    __asm__ volatile (
-        "cpuid"				"\n\t"
-	: "=a" (a), "=b" (b), "=c" (c), "=d" (d)
-	: "a" (feature));
-    /*
-     * FMA is the 12th bit of ECX
-     */
-    return (c >> 12) & 1;
-#else
-    return 0;
-#endif
-}
-
-/*
- * Primitive cpu feature detect function
- * Currently only supports checking for avx on gcc compatible compilers.
- */
-NPY_NO_EXPORT int
-npy_cpu_supports(const char * feature)
-{
-#ifdef HAVE___BUILTIN_CPU_SUPPORTS
-    if (strcmp(feature, "avx512f") == 0) {
-#ifdef HAVE___BUILTIN_CPU_SUPPORTS_AVX512F
-        return __builtin_cpu_supports("avx512f") && os_avx512_support();
-#else
-        return 0;
-#endif
-    }
-    else if (strcmp(feature, "fma") == 0) {
-        return cpu_supports_fma() && __builtin_cpu_supports("avx2") && os_avx_support();
-    }
-    else if (strcmp(feature, "avx2") == 0) {
-        return __builtin_cpu_supports("avx2") && os_avx_support();
-    }
-    else if (strcmp(feature, "avx") == 0) {
-        return __builtin_cpu_supports("avx") && os_avx_support();
-    }
-#endif
-
-    return 0;
-}
diff --git a/numpy/core/src/umath/cpuid.h b/numpy/core/src/umath/cpuid.h
deleted file mode 100644
index 33702ed41..000000000
--- a/numpy/core/src/umath/cpuid.h
+++ /dev/null
@@ -1,9 +0,0 @@
-#ifndef _NPY_PRIVATE__CPUID_H_
-#define _NPY_PRIVATE__CPUID_H_
-
-#include <numpy/ndarraytypes.h>  /* for NPY_NO_EXPORT */
-
-NPY_NO_EXPORT int
-npy_cpu_supports(const char * feature);
-
-#endif
diff --git a/numpy/core/tests/test_cpu_features.py b/numpy/core/tests/test_cpu_features.py
new file mode 100644
index 000000000..3b5cb3157
--- /dev/null
+++ b/numpy/core/tests/test_cpu_features.py
@@ -0,0 +1,104 @@
+import sys, platform, re, pytest
+
+from numpy.testing import assert_equal
+from numpy.core._multiarray_umath import __cpu_features__
+
+class AbstractTest(object):
+    features = []
+    features_groups = {}
+    features_map = {}
+    features_flags = set()
+
+    def load_flags(self):
+        # a hook
+        pass
+
+    def test_features(self):
+        self.load_flags()
+        for gname, features in self.features_groups.items():
+            test_features = [self.features_map.get(f, f) in self.features_flags for f in features]
+            assert_equal(__cpu_features__.get(gname), all(test_features))
+
+        for feature_name in self.features:
+            map_name = self.features_map.get(feature_name, feature_name)
+            cpu_have = map_name in self.features_flags
+            npy_have = __cpu_features__.get(feature_name)
+            assert_equal(npy_have, cpu_have)
+
+    def load_flags_proc(self, magic_key):
+        with open('/proc/cpuinfo') as fd:
+            for line in fd:
+                if not line.startswith(magic_key):
+                    continue
+                flags_value = [s.strip() for s in line.split(':', 1)]
+                if len(flags_value) == 2:
+                    self.features_flags = self.features_flags.union(flags_value[1].upper().split())
+
+    def load_flags_auxv(self):
+        import subprocess
+        auxv = subprocess.check_output(['/bin/true'], env=dict(LD_SHOW_AUXV="1"))
+        for at in auxv.split(b'\n'):
+            if not at.startswith(b"AT_HWCAP"):
+                continue
+            hwcap_value = [s.strip() for s in at.split(b':', 1)]
+            if len(hwcap_value) == 2:
+                self.features_flags = self.features_flags.union(
+                    hwcap_value[1].upper().decode().split()
+                )
+
+is_linux = sys.platform.startswith('linux')
+machine  = platform.machine()
+is_x86   = re.match("^(amd64|x86|i386|i686)", machine, re.IGNORECASE)
+@pytest.mark.skipif(not is_linux or not is_x86, reason="Only for Linux and x86")
+class Test_X86_Features(AbstractTest):
+    features = [
+        "MMX", "SSE", "SSE2", "SSE3", "SSSE3", "SSE41", "POPCNT", "SSE42",
+        "AVX", "F16C", "XOP", "FMA4", "FMA3", "AVX2", "AVX512F", "AVX512CD",
+        "AVX512ER", "AVX512PF", "AVX5124FMAPS", "AVX5124VNNIW", "AVX512VPOPCNTDQ",
+        "AVX512VL", "AVX512BW", "AVX512DQ", "AVX512VNNI", "AVX512IFMA",
+        "AVX512VBMI", "AVX512VBMI2", "AVX512BITALG",
+    ]
+    features_groups = dict(
+        AVX512_KNL = ["AVX512F", "AVX512CD", "AVX512ER", "AVX512PF"],
+        AVX512_KNM = ["AVX512F", "AVX512CD", "AVX512ER", "AVX512PF", "AVX5124FMAPS",
+                      "AVX5124VNNIW", "AVX512VPOPCNTDQ"],
+        AVX512_SKX = ["AVX512F", "AVX512CD", "AVX512BW", "AVX512DQ", "AVX512VL"],
+        AVX512_CLX = ["AVX512F", "AVX512CD", "AVX512BW", "AVX512DQ", "AVX512VL", "AVX512VNNI"],
+        AVX512_CNL = ["AVX512F", "AVX512CD", "AVX512BW", "AVX512DQ", "AVX512VL", "AVX512IFMA",
+                      "AVX512VBMI"],
+        AVX512_ICL = ["AVX512F", "AVX512CD", "AVX512BW", "AVX512DQ", "AVX512VL", "AVX512IFMA",
+                      "AVX512VBMI", "AVX512VNNI", "AVX512VBMI2", "AVX512BITALG", "AVX512VPOPCNTDQ"],
+    )
+    features_map = dict(
+        SSE3="PNI", SSE41="SSE4_1", SSE42="SSE4_2", FMA3="FMA",
+        AVX512VNNI="AVX512_VNNI", AVX512BITALG="AVX512_BITALG", AVX512VBMI2="AVX512_VBMI2",
+        AVX5124FMAPS="AVX512_4FMAPS", AVX5124VNNIW="AVX512_4VNNIW", AVX512VPOPCNTDQ="AVX512_VPOPCNTDQ",
+    )
+    def load_flags(self):
+        self.load_flags_proc("flags")
+
+is_power = re.match("^(powerpc|ppc)64", machine, re.IGNORECASE)
+@pytest.mark.skipif(not is_linux or not is_power, reason="Only for Linux and Power")
+class Test_POWER_Features(AbstractTest):
+    features = ["VSX", "VSX2", "VSX3"]
+    features_map = dict(VSX2="ARCH_2_07", VSX3="ARCH_3_00")
+
+    def load_flags(self):
+        self.load_flags_auxv()
+
+is_arm = re.match("^(arm|aarch64)", machine, re.IGNORECASE)
+@pytest.mark.skipif(not is_linux or not is_arm, reason="Only for Linux and ARM")
+class Test_ARM_Features(AbstractTest):
+    features = [
+        "NEON", "ASIMD", "FPHP", "ASIMDHP", "ASIMDDP", "ASIMDFHM"
+    ]
+    features_groups = dict(
+        NEON_FP16  = ["NEON", "HALF"],
+        NEON_VFPV4 = ["NEON", "VFPV4"],
+    )
+    def load_flags(self):
+        self.load_flags_proc("Features")
+        if re.match("^(aarch64|AARCH64)", platform.machine()):
+            self.features_map = dict(
+                NEON="ASIMD", HALF="ASIMD", VFPV4="ASIMD"
+            )