[common] add runtime detection and selection of AVX/AVX2 support

common/include/common/cpuinfo.h

@@ -29,4 +29,19 @@ bool cpuInfo_get(char * model, size_t modelSize, int * procs, int * cores,
 
 void cpuInfo_log(void);
 
+typedef struct
+{
+  bool sse, sse2, sse3, ssse3;
+  bool fma;
+  bool sse4_1, sse4_2;
+  bool popcnt;
+  bool aes;
+  bool xsave, osxsave;
+  bool avx, avx2;
+  bool bmi1, bmi2;
+}
+CPUInfoFeatures;
+
+const CPUInfoFeatures * cpuInfo_getFeatures(void);
+
 #endif

common/include/common/framebuffer.h

@@ -70,7 +70,8 @@ void framebuffer_prepare(FrameBuffer * frame);
 /**
  * Write data from the src buffer into the KVMFRFrame
  */
-bool framebuffer_write(FrameBuffer * frame, const void * src, size_t size);
+extern bool (*framebuffer_write)(FrameBuffer * frame,
+    const void * restrict src, size_t size);
 
 /**
  * Gets the underlying data buffer of the framebuffer.

common/include/common/rects.h

@@ -27,18 +27,8 @@
 #include "common/framebuffer.h"
 #include "common/types.h"
 
-inline static void rectCopyUnaligned(uint8_t * dst, const uint8_t * src,
+extern void (*rectCopyUnaligned)(uint8_t * dst, const uint8_t * src,
-    int ystart, int yend, int dx, int dstPitch, int srcPitch, int width)
+    int ystart, int yend, int dx, int dstPitch, int srcPitch, int width);
-{
-  src += ystart * srcPitch + dx;
-  dst += ystart * dstPitch + dx;
-  for (int i = ystart; i < yend; ++i)
-  {
-    memcpy(dst, src, width);
-    src += srcPitch;
-    dst += dstPitch;
- }
-}
 
 void rectsBufferToFramebuffer(FrameDamageRect * rects, int count, int bpp,
   FrameBuffer * frame, int dstPitch, int height,

common/src/cpuinfo.c

@@ -20,6 +20,7 @@
 
 #include "common/cpuinfo.h"
 #include "common/debug.h"
+#include "common/util.h"
 
 void cpuInfo_log(void)
 {
@@ -37,3 +38,73 @@ void cpuInfo_log(void)
   DEBUG_INFO("CPU Model: %s", model);
   DEBUG_INFO("CPU: %d sockets, %d cores, %d threads", sockets, cores, procs);
 }
+
+const CPUInfoFeatures * cpuInfo_getFeatures(void)
+{
+  static bool initialized = false;
+  static CPUInfoFeatures features;
+
+  if (likely(initialized))
+    return &features;
+
+  int cpuid[4] = {0};
+
+  // leaf1
+  asm volatile
+  (
+    "cpuid;"
+    : "=a" (cpuid[0]),
+      "=b" (cpuid[1]),
+      "=c" (cpuid[2]),
+      "=d" (cpuid[3])
+    : "a" (1)
+  );
+
+  features.sse     = cpuid[3] & (1 << 25);
+  features.sse2    = cpuid[3] & (1 << 26);
+  features.sse3    = cpuid[2] & (1 <<  0);
+  features.ssse3   = cpuid[2] & (1 <<  9);
+  features.fma     = cpuid[2] & (1 << 12);
+  features.sse4_1  = cpuid[2] & (1 << 19);
+  features.sse4_2  = cpuid[2] & (1 << 20);
+  features.popcnt  = cpuid[2] & (1 << 23);
+  features.aes     = cpuid[2] & (1 << 25);
+  features.xsave   = cpuid[2] & (1 << 26);
+  features.osxsave = cpuid[2] & (1 << 27);
+  features.avx     = cpuid[2] & (1 << 28);
+
+  // leaf7
+  asm volatile
+  (
+    "cpuid;"
+    : "=a" (cpuid[0]),
+      "=b" (cpuid[1]),
+      "=c" (cpuid[2]),
+      "=d" (cpuid[3])
+    : "a" (7), "c" (0)
+  );
+
+  features.avx2 = cpuid[1] & (1 << 5);
+  features.bmi1 = cpuid[2] & (1 << 3);
+  features.bmi2 = cpuid[2] & (1 << 8);
+
+  if (features.osxsave && features.avx)
+  {
+    int xgetbv = 0;
+    asm volatile
+    (
+      "xgetbv;"
+      : "=a" (xgetbv)
+      : "c" (0)
+      : "edx"
+    );
+
+    if (!(xgetbv & 0x6))
+    {
+      features.avx  = false;
+      features.avx2 = false;
+    }
+  }
+
+  return &features;
+};

common/src/framebuffer.c

@@ -19,6 +19,7 @@
  */
 
 #include "common/framebuffer.h"
+#include "common/cpuinfo.h"
 #include "common/debug.h"
 
 //#define FB_PROFILE
@@ -29,6 +30,7 @@
 #include <string.h>
 #include <emmintrin.h>
 #include <smmintrin.h>
+#include <immintrin.h>
 #include <unistd.h>
 
 bool framebuffer_wait(const FrameBuffer * frame, size_t size)
@@ -165,7 +167,8 @@ void framebuffer_prepare(FrameBuffer * frame)
   atomic_store_explicit(&frame->wp, 0, memory_order_release);
 }
 
-bool framebuffer_write(FrameBuffer * frame, const void * restrict src, size_t size)
+static bool framebuffer_write_sse4_1(FrameBuffer * frame,
+    const void * restrict src, size_t size)
 {
 #ifdef FB_PROFILE
   static RunningAvg ra = NULL;
@@ -222,6 +225,100 @@ bool framebuffer_write(FrameBuffer * frame, const void * restrict src, size_t si
   return true;
 }
 
+#pragma GCC push_options
+#pragma GCC target ("avx2")
+bool framebuffer_write_avx2(FrameBuffer * frame,
+    const void * restrict src, size_t size)
+{
+#ifdef FB_PROFILE
+    static RunningAvg ra = NULL;
+    static int raCount = 0;
+    const uint64_t ts = microtime();
+    if (!ra)
+        ra = runningavg_new(100);
+#endif
+
+  __m256i *restrict s = (__m256i *)src;
+  __m256i *restrict d = (__m256i *)frame->data;
+  size_t wp = 0;
+
+  _mm_mfence();
+
+  /* copy in chunks */
+  while (size > 127)
+  {
+    __m256i *_d = (__m256i *)d;
+    __m256i *_s = (__m256i *)s;
+    __m256i v1 = _mm256_stream_load_si256(_s + 0);
+    __m256i v2 = _mm256_stream_load_si256(_s + 1);
+    __m256i v3 = _mm256_stream_load_si256(_s + 2);
+    __m256i v4 = _mm256_stream_load_si256(_s + 3);
+
+    _mm256_stream_si256(_d + 0, v1);
+    _mm256_stream_si256(_d + 1, v2);
+    _mm256_stream_si256(_d + 2, v3);
+    _mm256_stream_si256(_d + 3, v4);
+
+    s += 4;
+    d += 4;
+    size -= 128;
+    wp += 128;
+
+    if (wp % FB_CHUNK_SIZE == 0)
+      atomic_store_explicit(&frame->wp, wp, memory_order_release);
+  }
+
+  if (size > 63)
+  {
+    __m256i *_d = (__m256i *)d;
+    __m256i *_s = (__m256i *)s;
+    __m256i v1 = _mm256_stream_load_si256(_s);
+    __m256i v2 = _mm256_stream_load_si256(_s + 1);
+
+    _mm256_stream_si256(_d, v1);
+    _mm256_stream_si256(_d + 1, v2);
+
+    s += 2;
+    d += 2;
+    size -= 64;
+    wp += 64;
+
+    if (wp % FB_CHUNK_SIZE == 0)
+      atomic_store_explicit(&frame->wp, wp, memory_order_release);
+  }
+
+  if (size)
+  {
+    memcpy(frame->data + wp, s, size);
+    wp += size;
+  }
+
+  atomic_store_explicit(&frame->wp, wp, memory_order_release);
+
+#ifdef FB_PROFILE
+  runningavg_push(ra, microtime() - ts);
+  if (++raCount % 100 == 0)
+    DEBUG_INFO("Average Copy Time: %.2fμs", runningavg_calc(ra));
+#endif
+
+  return true;
+}
+#pragma GCC pop_options
+
+static bool _framebuffer_write(FrameBuffer * frame,
+    const void * restrict src, size_t size)
+{
+  if (cpuInfo_getFeatures()->avx2)
+    framebuffer_write = &framebuffer_write_avx2;
+  else
+    framebuffer_write = &framebuffer_write_sse4_1;
+
+  return framebuffer_write(frame, src, size);
+}
+
+bool (*framebuffer_write)(FrameBuffer * frame,
+  const void * restrict src, size_t size) = &_framebuffer_write;
+
 const uint8_t * framebuffer_get_buffer(const FrameBuffer * frame)
 {
   return frame->data;

common/src/rects.c

@@ -20,8 +20,10 @@
 
 #include "common/rects.h"
 #include "common/util.h"
+#include "common/cpuinfo.h"
 
 #include <stdlib.h>
+#include <immintrin.h>
 
 struct Corner
 {
@@ -298,3 +300,58 @@ int rectsRejectContained(FrameDamageRect * rects, int count)
 
   return removeRects(rects, count, removed);
 }
+
+static void rectCopyUnaligned_memcpy(uint8_t * dst, const uint8_t * src,
+    int ystart, int yend, int dx, int dstPitch, int srcPitch, int width)
+{
+  src += ystart * srcPitch + dx;
+  dst += ystart * dstPitch + dx;
+  for (int i = ystart; i < yend; ++i)
+  {
+    memcpy(dst, src, width);
+    src += srcPitch;
+    dst += dstPitch;
+  }
+}
+
+#pragma GCC push_options
+#pragma GCC target ("avx2")
+static void rectCopyUnaligned_avx(uint8_t * dst, const uint8_t * src,
+    int ystart, int yend, int dx, int dstPitch, int srcPitch, int width)
+{
+  src += ystart * srcPitch + dx;
+  dst += ystart * dstPitch + dx;
+  for (int i = ystart; i < yend; ++i)
+  {
+    int col;
+    for(col = 0; col <= width - 32; col += 32)
+    {
+      _mm_prefetch(src + col + 256, _MM_HINT_T0);
+      __m256i srcData = _mm256_loadu_si256((__m256i*)(src + col));
+      _mm256_storeu_si256((__m256i*)(dst + col), srcData);
+    }
+
+    for(; col < width; ++col)
+      dst[col] = src[col];
+
+    src += srcPitch;
+    dst += dstPitch;
+ }
+}
+#pragma GCC pop_options
+
+static void _rectCopyUnaligned(uint8_t * dst, const uint8_t * src,
+    int ystart, int yend, int dx, int dstPitch, int srcPitch, int width)
+{
+  if (cpuInfo_getFeatures()->avx)
+    rectCopyUnaligned = &rectCopyUnaligned_avx;
+  else
+    rectCopyUnaligned = &rectCopyUnaligned_memcpy;
+
+  return rectCopyUnaligned(
+      dst, src, ystart, yend, dx, dstPitch, srcPitch, width);
+}
+
+void (*rectCopyUnaligned)(uint8_t * dst, const uint8_t * src,
+    int ystart, int yend, int dx, int dstPitch, int srcPitch, int width) =
+  &_rectCopyUnaligned;