ARM Skia NEON patches - 31 - Xfermode: xfer16

src/opts/SkXfermode_opts_arm_neon.cpp

 #include "SkXfermode.h"
 #include "SkXfermode_proccoeff.h"
 #include "SkColorPriv.h"
 
 #include <arm_neon.h>
 #include "SkColor_opts_neon.h"
 #include "SkXfermode_opts_arm_neon.h"
 
 #define SkAlphaMulAlpha(a, b)   SkMulDiv255Round(a, b)
 
@@ skipping 562 matching lines @@
     SkASSERT(dst && src && count >= 0);
 
     SkXfermodeProc proc = this->getProc();
     SkXfermodeProcSIMD procSIMD = reinterpret_cast<SkXfermodeProcSIMD>(fProcSIMD);
 
     if (NULL == aa) {
         // Unrolled NEON code
         while (count >= 8) {
             uint8x8x4_t vsrc, vdst, vres;
 
-#if (__GNUC__ == 4) && (__GNUC_MINOR__ > 6)
+#if (__GNUC__ > 4) || ((__GNUC__ == 4) && (__GNUC_MINOR__ > 6))
             asm volatile (
                 "vld4.u8    %h[vsrc], [%[src]]!  \t\n"
                 "vld4.u8    %h[vdst], [%[dst]]   \t\n"
                 : [vsrc] "=w" (vsrc), [vdst] "=w" (vdst), [src] "+&r" (src)
                 : [dst] "r" (dst)
                 :
             );
 #else
             register uint8x8_t d0 asm("d0");
             register uint8x8_t d1 asm("d1");
@@ skipping 38 matching lines @@
                 SkPMColor C = proc(src[i], dstC);
                 if (a != 0xFF) {
                     C = SkFourByteInterp(C, dstC, a);
                 }
                 dst[i] = C;
             }
         }
     }
 }
 
+void SkNEONProcCoeffXfermode::xfer16(uint16_t* SK_RESTRICT dst,
+                                     const SkPMColor* SK_RESTRICT src, int count,
+                                     const SkAlpha* SK_RESTRICT aa) const {
+    SkASSERT(dst && src && count >= 0);
+
+    SkXfermodeProc proc = this->getProc();
+    SkXfermodeProcSIMD procSIMD = reinterpret_cast<SkXfermodeProcSIMD>(fProcSIMD);

[mtklein, 2013/11/06 16:04:17]

SkASSERT procSIMD != NULL?

[kevin.petit.not.used.account, 2013/11/06 16:50:05]

Why not but it isn't possible to reach this code with a NULL fProcSIMD. You can
have a look at SkPlatformXfermodeFactory_impl_neon() to convince yourself.

If the intent is to document that and or protect ourselves from future
modifications, then yes (plus modifying the above function I guess).

[mtklein, 2013/11/06 17:18:41]

Yeah, the intention is just documentation.

[kevin.petit.not.used.account, 2013/11/07 11:26:20]

Done.

+
+    if (NULL == aa) {
+        while(count >= 8) {
+            uint16x8_t vdst, vres16;
+            uint8x8x4_t vdst32, vsrc, vres;
+
+            vdst = vld1q_u16(dst);
+
+#if (__GNUC__ > 4) || ((__GNUC__ == 4) && (__GNUC_MINOR__ > 6))

[mtklein, 2013/11/06 16:04:17]

Is Clang any better at generating code from intrinsics?  I'm now officially in
your camp that GCC just barfs out garbage.  (Even up to 4.8 as far as I've
seen).

[kevin.petit.not.used.account, 2013/11/06 16:50:05]

To be honest I haven't done a detailed analysis of the code generated by Clang.

I know that gcc, starting from 4.7 generates pretty good code for most of the
intrinsics as long as the pressure on registers is reasonable (read as long as
the function isn't too long and doesn't use simultaneously too many NEON
variables). If that's not the case it starts doing lots of stupid things.

However vld4 is a remarkable counter-example that even 4.8 doesn't always handle
in an optimal way.

If you have examples of bad code generated by gcc or clang, I'd be more than
interested.

[mtklein, 2013/11/06 17:18:41]

Ah, this may be a coincidence then.  I was indeed thinking of the bizarre code
GCC generates for loads.  I see things like, load into d16-d17, followed
immediately by storing d16 and d17 back to the stack.

[kevin.petit.not.used.account, 2013/11/07 11:26:20]

Yes, that's a typical pattern. Could you point me to the code producing this
assembly please?

As a general remark, don't hesitate to drop me a line with code snippets when
you stumble upon bad assembly produced by gcc or clang. Our compiler team and
myself are very keen on getting that kind of feedback.

+            asm volatile (
+                "vld4.u8    %h[vsrc], [%[src]]!  \t\n"
+                : [vsrc] "=w" (vsrc), [src] "+&r" (src)
+                : :
+            );
+#else
+            register uint8x8_t d0 asm("d0");
+            register uint8x8_t d1 asm("d1");
+            register uint8x8_t d2 asm("d2");
+            register uint8x8_t d3 asm("d3");
+
+            asm volatile (
+                "vld4.u8    {d0-d3},[%[src]]!;"
+                : "=w" (d0), "=w" (d1), "=w" (d2), "=w" (d3),
+                  [src] "+&r" (src)
+                : :
+            );
+            vsrc.val[0] = d0;
+            vsrc.val[1] = d1;
+            vsrc.val[2] = d2;
+            vsrc.val[3] = d3;
+#endif
+
+            vdst32 = SkPixel16ToPixel32_neon8(vdst);
+            vres = procSIMD(vsrc, vdst32);
+            vres16 = SkPixel32ToPixel16_neon8(vres);
+
+            vst1q_u16(dst, vres16);
+
+            count -= 8;
+            dst += 8;
+        }
+        for (int i = 0; i < count; i++) {
+            SkPMColor dstC = SkPixel16ToPixel32(dst[i]);
+            dst[i] = SkPixel32ToPixel16_ToU16(proc(src[i], dstC));
+        }
+    } else {
+        for (int i = count - 1; i >= 0; --i) {
+            unsigned a = aa[i];
+            if (0 != a) {
+                SkPMColor dstC = SkPixel16ToPixel32(dst[i]);
+                SkPMColor C = proc(src[i], dstC);
+                if (0xFF != a) {
+                    C = SkFourByteInterp(C, dstC, a);
+                }
+                dst[i] = SkPixel32ToPixel16_ToU16(C);
+            }
+        }
+    }
+}
+
 #ifdef SK_DEVELOPER
 void SkNEONProcCoeffXfermode::toString(SkString* str) const {
     this->INHERITED::toString(str);
 }
 #endif
 
 ////////////////////////////////////////////////////////////////////////////////
 
 SkXfermodeProcSIMD gNEONXfermodeProcs[] = {
     NULL, // kClear_Mode
@@ skipping 37 matching lines @@
 SkProcCoeffXfermode* SkPlatformXfermodeFactory_impl_neon(const ProcCoeff& rec,
                                                          SkXfermode::Mode mode) {
 
     void* procSIMD = reinterpret_cast<void*>(gNEONXfermodeProcs[mode]);
 
     if (procSIMD != NULL) {
         return SkNEW_ARGS(SkNEONProcCoeffXfermode, (rec, mode, procSIMD));
     }
     return NULL;
 }