AVX512 VMBI version of ARGBToRGB24

Use VMBI instructions but on AVX2 registers to avoid clockrate change.

Bug: libyuv:778
Test: LibYUVConvertTest.NV21ToRGB24_Opt
Change-Id: Id4f8ad1e0e142a380c8a46c5eab90ce145a10edd
Reviewed-on: https://chromium-review.googlesource.com/956609Reviewed-by: richard winterton <rrwinterton@gmail.com>

Android.bp

@@ -43,7 +43,6 @@ cc_library {
         "source/scale_neon.cc",
         "source/scale_neon64.cc",
         "source/video_common.cc",
-
         "source/convert_jpeg.cc",
         "source/mjpeg_decoder.cc",
         "source/mjpeg_validate.cc",

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 1703
+Version: 1704
 License: BSD
 License File: LICENSE
 

include/libyuv/row.h

@@ -277,6 +277,7 @@ extern "C" {
 #define HAS_ARGBTOAR30ROW_AVX2
 #define HAS_ARGBTORAWROW_AVX2
 #define HAS_ARGBTORGB24ROW_AVX2
+#define HAS_ARGBTORGB24ROW_AVX512VBMI
 #define HAS_CONVERT16TO8ROW_AVX2
 #define HAS_CONVERT8TO16ROW_AVX2
 #define HAS_I210TOAR30ROW_AVX2
@@ -1706,6 +1707,8 @@ void ARGBToAR30Row_SSSE3(const uint8_t* src, uint8_t* dst, int width);
 void ARGBToRAWRow_AVX2(const uint8_t* src, uint8_t* dst, int width);
 void ARGBToRGB24Row_AVX2(const uint8_t* src, uint8_t* dst, int width);
 
+void ARGBToRGB24Row_AVX512VBMI(const uint8_t* src, uint8_t* dst, int width);
+
 void ARGBToRGB565DitherRow_C(const uint8_t* src_argb,
                              uint8_t* dst_rgb,
                              const uint32_t dither4,
@@ -2497,12 +2500,13 @@ void ABGRToAR30Row_Any_SSSE3(const uint8_t* src_ptr,
 void ARGBToAR30Row_Any_SSSE3(const uint8_t* src_ptr,
                              uint8_t* dst_ptr,
                              int width);
-void ARGBToRAWRow_Any_AVX2(const uint8_t* src_ptr,
-                             uint8_t* dst_ptr,
-                             int width);
+void ARGBToRAWRow_Any_AVX2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void ARGBToRGB24Row_Any_AVX2(const uint8_t* src_ptr,
                              uint8_t* dst_ptr,
                              int width);
+void ARGBToRGB24Row_Any_AVX512VBMI(const uint8_t* src_ptr,
+                                   uint8_t* dst_ptr,
+                                   int width);
 void ARGBToRGB565DitherRow_Any_SSE2(const uint8_t* src_ptr,
                                     uint8_t* dst_ptr,
                                     const uint32_t param,

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 1703
+#define LIBYUV_VERSION 1704
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_

source/convert_from_argb.cc

@@ -887,6 +887,14 @@ int ARGBToRGB24(const uint8_t* src_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBTORGB24ROW_AVX512VBMI)
+  if (TestCpuFlag(kCpuHasAVX512VBMI)) {
+    ARGBToRGB24Row = ARGBToRGB24Row_Any_AVX512VBMI;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToRGB24Row = ARGBToRGB24Row_AVX512VBMI;
+    }
+  }
+#endif
 #if defined(HAS_ARGBTORGB24ROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToRGB24Row = ARGBToRGB24Row_Any_NEON;

source/row_any.cc

@@ -447,6 +447,9 @@ ANY11(ARGBToARGB4444Row_Any_SSE2, ARGBToARGB4444Row_SSE2, 0, 4, 2, 3)
 #if defined(HAS_ARGBTORGB24ROW_AVX2)
 ANY11(ARGBToRGB24Row_Any_AVX2, ARGBToRGB24Row_AVX2, 0, 4, 3, 31)
 #endif
+#if defined(HAS_ARGBTORGB24ROW_AVX512VBMI)
+ANY11(ARGBToRGB24Row_Any_AVX512VBMI, ARGBToRGB24Row_AVX512VBMI, 0, 4, 3, 31)
+#endif
 #if defined(HAS_ARGBTORAWROW_AVX2)
 ANY11(ARGBToRAWRow_Any_AVX2, ARGBToRAWRow_AVX2, 0, 4, 3, 31)
 #endif

source/row_gcc.cc

@@ -552,6 +552,54 @@ void ARGBToRGB24Row_AVX2(const uint8_t* src, uint8_t* dst, int width) {
         "xmm7");
 }
 
+// TODO(fbarchard): Detect compiler can di avx512 and add ifdefs
+
+// Shuffle table for converting ARGBToRGB24
+static const ulvec8 kPermARGBToRGB24_0 = {
+    0u,  1u,  2u,  4u,  5u,  6u,  8u,  9u,  10u, 12u, 13u,
+    14u, 16u, 17u, 18u, 20u, 21u, 22u, 24u, 25u, 26u, 28u,
+    29u, 30u, 32u, 33u, 34u, 36u, 37u, 38u, 40u, 41u};
+static const ulvec8 kPermARGBToRGB24_1 = {
+    10u, 12u, 13u, 14u, 16u, 17u, 18u, 20u, 21u, 22u, 24u,
+    25u, 26u, 28u, 29u, 30u, 32u, 33u, 34u, 36u, 37u, 38u,
+    40u, 41u, 42u, 44u, 45u, 46u, 48u, 49u, 50u, 52u};
+static const ulvec8 kPermARGBToRGB24_2 = {
+    21u, 22u, 24u, 25u, 26u, 28u, 29u, 30u, 32u, 33u, 34u,
+    36u, 37u, 38u, 40u, 41u, 42u, 44u, 45u, 46u, 48u, 49u,
+    50u, 52u, 53u, 54u, 56u, 57u, 58u, 60u, 61u, 62u};
+
+void ARGBToRGB24Row_AVX512VBMI(const uint8_t* src, uint8_t* dst, int width) {
+  asm volatile(
+      "vmovdqa    %3,%%ymm5                      \n"
+      "vmovdqa    %4,%%ymm6                      \n"
+      "vmovdqa    %5,%%ymm7                      \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu    (%0),%%ymm0                    \n"
+      "vmovdqu    0x20(%0),%%ymm1                \n"
+      "vmovdqu    0x40(%0),%%ymm2                \n"
+      "vmovdqu    0x60(%0),%%ymm3                \n"
+      "lea        0x80(%0),%0                    \n"
+      "vpermt2b   %%ymm1,%%ymm5,%%ymm0           \n"
+      "vpermt2b   %%ymm2,%%ymm6,%%ymm1           \n"
+      "vpermt2b   %%ymm3,%%ymm7,%%ymm2           \n"
+      "vmovdqu    %%ymm0,(%1)                    \n"
+      "vmovdqu    %%ymm1,0x20(%1)                \n"
+      "vmovdqu    %%ymm2,0x40(%1)                \n"
+      "lea        0x60(%1),%1                    \n"
+      "sub        $0x20,%2                       \n"
+      "jg         1b                             \n"
+      "vzeroupper                                \n"
+      : "+r"(src),                // %0
+        "+r"(dst),                // %1
+        "+r"(width)               // %2
+      : "m"(kPermARGBToRGB24_0),  // %3
+        "m"(kPermARGBToRGB24_1),  // %4
+        "m"(kPermARGBToRGB24_2)   // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5", "xmm6", "xmm7");
+}
+
 void ARGBToRAWRow_AVX2(const uint8_t* src, uint8_t* dst, int width) {
   asm volatile(
       "vbroadcastf128 %3,%%ymm6                  \n"
@@ -587,11 +635,11 @@ void ARGBToRAWRow_AVX2(const uint8_t* src, uint8_t* dst, int width) {
       "sub        $0x20,%2                       \n"
       "jg         1b                             \n"
       "vzeroupper                                \n"
-      : "+r"(src),                     // %0
-        "+r"(dst),                     // %1
-        "+r"(width)                    // %2
-      : "m"(kShuffleMaskARGBToRAW),    // %3
-        "m"(kPermdRGB24_AVX)           // %4
+      : "+r"(src),                   // %0
+        "+r"(dst),                   // %1
+        "+r"(width)                  // %2
+      : "m"(kShuffleMaskARGBToRAW),  // %3
+        "m"(kPermdRGB24_AVX)         // %4
       : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
         "xmm7");
 }