util: add util function buffer_zero_avx512()

And intialize buffer_is_zero() with it, when Intel AVX512F is
available on host.

This function utilizes Intel AVX512 fundamental instructions which
is faster than its implementation with AVX2 (in my unit test, with
4K buffer, on CascadeLake SP, ~36% faster, buffer_zero_avx512() V.S.
buffer_zero_avx2()).

Signed-off-by: Robert Hoo <robert.hu@linux.intel.com>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

include/qemu/cpuid.h

@@ -45,6 +45,9 @@
 #ifndef bit_AVX2
 #define bit_AVX2        (1 << 5)
 #endif
+#ifndef bit_AVX512F
+#define bit_AVX512F        (1 << 16)
+#endif
 #ifndef bit_BMI2
 #define bit_BMI2        (1 << 8)
 #endif

util/bufferiszero.c

@@ -63,11 +63,11 @@ buffer_zero_int(const void *buf, size_t len)
     }
 }
 
-#if defined(CONFIG_AVX2_OPT) || defined(__SSE2__)
+#if defined(CONFIG_AVX512F_OPT) || defined(CONFIG_AVX2_OPT) || defined(__SSE2__)
 /* Do not use push_options pragmas unnecessarily, because clang
  * does not support them.
  */
-#ifdef CONFIG_AVX2_OPT
+#if defined(CONFIG_AVX512F_OPT) || defined(CONFIG_AVX2_OPT)
 #pragma GCC push_options
 #pragma GCC target("sse2")
 #endif
@@ -104,7 +104,7 @@ buffer_zero_sse2(const void *buf, size_t len)
 
     return _mm_movemask_epi8(_mm_cmpeq_epi8(t, zero)) == 0xFFFF;
 }
-#ifdef CONFIG_AVX2_OPT
+#if defined(CONFIG_AVX512F_OPT) || defined(CONFIG_AVX2_OPT)
 #pragma GCC pop_options
 #endif
 
@@ -187,18 +187,54 @@ buffer_zero_avx2(const void *buf, size_t len)
 #pragma GCC pop_options
 #endif /* CONFIG_AVX2_OPT */
 
+#ifdef CONFIG_AVX512F_OPT
+#pragma GCC push_options
+#pragma GCC target("avx512f")
+#include <immintrin.h>
+
+static bool
+buffer_zero_avx512(const void *buf, size_t len)
+{
+    /* Begin with an unaligned head of 64 bytes.  */
+    __m512i t = _mm512_loadu_si512(buf);
+    __m512i *p = (__m512i *)(((uintptr_t)buf + 5 * 64) & -64);
+    __m512i *e = (__m512i *)(((uintptr_t)buf + len) & -64);
+
+    /* Loop over 64-byte aligned blocks of 256.  */
+    while (p <= e) {
+        __builtin_prefetch(p);
+        if (unlikely(_mm512_test_epi64_mask(t, t))) {
+            return false;
+        }
+        t = p[-4] | p[-3] | p[-2] | p[-1];
+        p += 4;
+    }
+
+    t |= _mm512_loadu_si512(buf + len - 4 * 64);
+    t |= _mm512_loadu_si512(buf + len - 3 * 64);
+    t |= _mm512_loadu_si512(buf + len - 2 * 64);
+    t |= _mm512_loadu_si512(buf + len - 1 * 64);
+
+    return !_mm512_test_epi64_mask(t, t);
+
+}
+#pragma GCC pop_options
+#endif
+
+
 /* Note that for test_buffer_is_zero_next_accel, the most preferred
  * ISA must have the least significant bit.
  */
-#define CACHE_AVX2    1
-#define CACHE_SSE4    2
-#define CACHE_SSE2    4
+#define CACHE_AVX512F 1
+#define CACHE_AVX2    2
+#define CACHE_SSE4    4
+#define CACHE_SSE2    8
 
 /* Make sure that these variables are appropriately initialized when
  * SSE2 is enabled on the compiler command-line, but the compiler is
  * too old to support CONFIG_AVX2_OPT.
  */
-#ifdef CONFIG_AVX2_OPT
+#if defined(CONFIG_AVX512F_OPT) || defined(CONFIG_AVX2_OPT)
 # define INIT_CACHE 0
 # define INIT_ACCEL buffer_zero_int
 #else
@@ -211,6 +247,7 @@ buffer_zero_avx2(const void *buf, size_t len)
 
 static unsigned cpuid_cache = INIT_CACHE;
 static bool (*buffer_accel)(const void *, size_t) = INIT_ACCEL;
+static int length_to_accel = 64;
 
 static void init_accel(unsigned cache)
 {
@@ -225,11 +262,17 @@ static void init_accel(unsigned cache)
     if (cache & CACHE_AVX2) {
         fn = buffer_zero_avx2;
     }
+#endif
+#ifdef CONFIG_AVX512F_OPT
+    if (cache & CACHE_AVX512F) {
+        fn = buffer_zero_avx512;
+        length_to_accel = 256;
+    }
 #endif
     buffer_accel = fn;
 }
 
-#ifdef CONFIG_AVX2_OPT
+#if defined(CONFIG_AVX512F_OPT) || defined(CONFIG_AVX2_OPT)
 #include "qemu/cpuid.h"
 
 static void __attribute__((constructor)) init_cpuid_cache(void)
@@ -252,9 +295,17 @@ static void __attribute__((constructor)) init_cpuid_cache(void)
             int bv;
             __asm("xgetbv" : "=a"(bv), "=d"(d) : "c"(0));
             __cpuid_count(7, 0, a, b, c, d);
-            if ((bv & 6) == 6 && (b & bit_AVX2)) {
+            if ((bv & 0x6) == 0x6 && (b & bit_AVX2)) {
                 cache |= CACHE_AVX2;
             }
+            /* 0xe6:
+            *  XCR0[7:5] = 111b (OPMASK state, upper 256-bit of ZMM0-ZMM15
+            *                    and ZMM16-ZMM31 state are enabled by OS)
+            *  XCR0[2:1] = 11b (XMM state and YMM state are enabled by OS)
+            */
+            if ((bv & 0xe6) == 0xe6 && (b & bit_AVX512F)) {
+                cache |= CACHE_AVX512F;
+            }
         }
     }
     cpuid_cache = cache;
@@ -277,7 +328,7 @@ bool test_buffer_is_zero_next_accel(void)
 
 static bool select_accel_fn(const void *buf, size_t len)
 {
-    if (likely(len >= 64)) {
+    if (likely(len >= length_to_accel)) {
         return buffer_accel(buf, len);
     }
     return buffer_zero_int(buf, len);