ARM Skia NEON patches - 32 - Xfermode: 1-pixel NEON modeprocs

In some cases, it's easy to provide a NEON version of the 1-pixel modeprocs.
Combined with https://codereview.chromium.org/23724013/ (merged) it allows
up to 35% speed improvement on Xfermodes when aa is non-NULL.

Signed-off-by: Kévin PETIT <kevin.petit@arm.com>

BUG=
R=djsollen@google.com, reed@google.com, mtklein@google.com, luisjoseromeroesclusa@hotmail.com

Author: kevin.petit.arm@gmail.com

Review URL: https://codereview.chromium.org/104883004

git-svn-id: http://skia.googlecode.com/svn/trunk@12525 2bbb7eff-a529-9590-31e7-b0007b416f81

5 files changed, 201 insertions, 2 deletions

diff --git a/src/core/SkXfermode.cpp b/src/core/SkXfermode.cpp
index 313e2aef8c..8cb79c2dbd 100644
--- a/src/core/SkXfermode.cpp
+++ b/src/core/SkXfermode.cpp
@@ -1669,6 +1669,7 @@ void SkXfermode::Term() {
 
 extern SkProcCoeffXfermode* SkPlatformXfermodeFactory(const ProcCoeff& rec,
                                                       SkXfermode::Mode mode);
+extern SkXfermodeProc SkPlatformXfermodeProcFactory(SkXfermode::Mode mode);
 
 SkXfermode* SkXfermode::Create(Mode mode) {
     SkASSERT(SK_ARRAY_COUNT(gProcCoeffs) == kModeCount);
@@ -1690,7 +1691,13 @@ SkXfermode* SkXfermode::Create(Mode mode) {
 
     SkXfermode* xfer = gCachedXfermodes[mode];
     if (NULL == xfer) {
-        const ProcCoeff& rec = gProcCoeffs[mode];
+        ProcCoeff rec = gProcCoeffs[mode];
+
+        SkXfermodeProc pp = SkPlatformXfermodeProcFactory(mode);
+
+        if (pp != NULL) {
+            rec.fProc = pp;
+        }
 
         // check if we have a platform optim for that
         SkProcCoeffXfermode* xfm = SkPlatformXfermodeFactory(rec, mode);
diff --git a/src/opts/SkXfermode_opts_arm.cpp b/src/opts/SkXfermode_opts_arm.cpp
index eb3b3016e3..2a796d6f6a 100644
--- a/src/opts/SkXfermode_opts_arm.cpp
+++ b/src/opts/SkXfermode_opts_arm.cpp
@@ -5,12 +5,22 @@
 extern SkProcCoeffXfermode* SkPlatformXfermodeFactory_impl_neon(const ProcCoeff& rec,
                                                                 SkXfermode::Mode mode);
 
+extern SkXfermodeProc SkPlatformXfermodeProcFactory_impl_neon(SkXfermode::Mode mode);
+
 SkProcCoeffXfermode* SkPlatformXfermodeFactory_impl(const ProcCoeff& rec,
                                                     SkXfermode::Mode mode) {
     return NULL;
 }
 
+SkXfermodeProc SkPlatformXfermodeProcFactory_impl(SkXfermode::Mode mode) {
+    return NULL;
+}
+
 SkProcCoeffXfermode* SkPlatformXfermodeFactory(const ProcCoeff& rec,
                                                SkXfermode::Mode mode) {
     return SK_ARM_NEON_WRAP(SkPlatformXfermodeFactory_impl)(rec, mode);
 }
+
+SkXfermodeProc SkPlatformXfermodeProcFactory(SkXfermode::Mode mode) {
+    return SK_ARM_NEON_WRAP(SkPlatformXfermodeProcFactory_impl)(mode);
+}
diff --git a/src/opts/SkXfermode_opts_arm_neon.cpp b/src/opts/SkXfermode_opts_arm_neon.cpp
index 7435dd44de..6a79b73726 100644
--- a/src/opts/SkXfermode_opts_arm_neon.cpp
+++ b/src/opts/SkXfermode_opts_arm_neon.cpp
@@ -93,6 +93,133 @@ static inline uint8x8_t clamp_div255round_simd8_32(int32x4_t val1, int32x4_t val
 }
 
 ////////////////////////////////////////////////////////////////////////////////
+// 1 pixel modeprocs
+////////////////////////////////////////////////////////////////////////////////
+
+//  kSrcATop_Mode,  //!< [Da, Sc * Da + (1 - Sa) * Dc]
+SkPMColor srcatop_modeproc_neon(SkPMColor src, SkPMColor dst) {
+    unsigned sa = SkGetPackedA32(src);
+    unsigned da = SkGetPackedA32(dst);
+    unsigned isa = 255 - sa;
+
+    uint8x8_t vda, visa, vsrc, vdst;
+
+    vda = vdup_n_u8(da);
+    visa = vdup_n_u8(isa);
+
+    uint16x8_t vsrc_wide, vdst_wide;
+    vsrc_wide = vmull_u8(vda, vreinterpret_u8_u32(vdup_n_u32(src)));
+    vdst_wide = vmull_u8(visa, vreinterpret_u8_u32(vdup_n_u32(dst)));
+
+    vsrc_wide += vdupq_n_u16(128);
+    vsrc_wide += vshrq_n_u16(vsrc_wide, 8);
+
+    vdst_wide += vdupq_n_u16(128);
+    vdst_wide += vshrq_n_u16(vdst_wide, 8);
+
+    vsrc = vshrn_n_u16(vsrc_wide, 8);
+    vdst = vshrn_n_u16(vdst_wide, 8);
+
+    vsrc += vdst;
+    vsrc = vset_lane_u8(da, vsrc, 3);
+
+    return vget_lane_u32(vreinterpret_u32_u8(vsrc), 0);
+}
+
+//  kDstATop_Mode,  //!< [Sa, Sa * Dc + Sc * (1 - Da)]
+SkPMColor dstatop_modeproc_neon(SkPMColor src, SkPMColor dst) {
+    unsigned sa = SkGetPackedA32(src);
+    unsigned da = SkGetPackedA32(dst);
+    unsigned ida = 255 - da;
+
+    uint8x8_t vsa, vida, vsrc, vdst;
+
+    vsa = vdup_n_u8(sa);
+    vida = vdup_n_u8(ida);
+
+    uint16x8_t vsrc_wide, vdst_wide;
+    vsrc_wide = vmull_u8(vida, vreinterpret_u8_u32(vdup_n_u32(src)));
+    vdst_wide = vmull_u8(vsa, vreinterpret_u8_u32(vdup_n_u32(dst)));
+
+    vsrc_wide += vdupq_n_u16(128);
+    vsrc_wide += vshrq_n_u16(vsrc_wide, 8);
+
+    vdst_wide += vdupq_n_u16(128);
+    vdst_wide += vshrq_n_u16(vdst_wide, 8);
+
+    vsrc = vshrn_n_u16(vsrc_wide, 8);
+    vdst = vshrn_n_u16(vdst_wide, 8);
+
+    vsrc += vdst;
+    vsrc = vset_lane_u8(sa, vsrc, 3);
+
+    return vget_lane_u32(vreinterpret_u32_u8(vsrc), 0);
+}
+
+//  kXor_Mode   [Sa + Da - 2 * Sa * Da, Sc * (1 - Da) + (1 - Sa) * Dc]
+SkPMColor xor_modeproc_neon(SkPMColor src, SkPMColor dst) {
+    unsigned sa = SkGetPackedA32(src);
+    unsigned da = SkGetPackedA32(dst);
+    unsigned ret_alpha = sa + da - (SkAlphaMulAlpha(sa, da) << 1);
+    unsigned isa = 255 - sa;
+    unsigned ida = 255 - da;
+
+    uint8x8_t vsrc, vdst, visa, vida;
+    uint16x8_t vsrc_wide, vdst_wide;
+
+    visa = vdup_n_u8(isa);
+    vida = vdup_n_u8(ida);
+    vsrc = vreinterpret_u8_u32(vdup_n_u32(src));
+    vdst = vreinterpret_u8_u32(vdup_n_u32(dst));
+
+    vsrc_wide = vmull_u8(vsrc, vida);
+    vdst_wide = vmull_u8(vdst, visa);
+
+    vsrc_wide += vdupq_n_u16(128);
+    vsrc_wide += vshrq_n_u16(vsrc_wide, 8);
+
+    vdst_wide += vdupq_n_u16(128);
+    vdst_wide += vshrq_n_u16(vdst_wide, 8);
+
+    vsrc = vshrn_n_u16(vsrc_wide, 8);
+    vdst = vshrn_n_u16(vdst_wide, 8);
+
+    vsrc += vdst;
+
+    vsrc = vset_lane_u8(ret_alpha, vsrc, 3);
+
+    return vget_lane_u32(vreinterpret_u32_u8(vsrc), 0);
+}
+
+// kPlus_Mode
+SkPMColor plus_modeproc_neon(SkPMColor src, SkPMColor dst) {
+    uint8x8_t vsrc, vdst;
+    vsrc = vreinterpret_u8_u32(vdup_n_u32(src));
+    vdst = vreinterpret_u8_u32(vdup_n_u32(dst));
+    vsrc = vqadd_u8(vsrc, vdst);
+
+    return vget_lane_u32(vreinterpret_u32_u8(vsrc), 0);
+}
+
+// kModulate_Mode
+SkPMColor modulate_modeproc_neon(SkPMColor src, SkPMColor dst) {
+    uint8x8_t vsrc, vdst, vres;
+    uint16x8_t vres_wide;
+
+    vsrc = vreinterpret_u8_u32(vdup_n_u32(src));
+    vdst = vreinterpret_u8_u32(vdup_n_u32(dst));
+
+    vres_wide = vmull_u8(vsrc, vdst);
+
+    vres_wide += vdupq_n_u16(128);
+    vres_wide += vshrq_n_u16(vres_wide, 8);
+
+    vres = vshrn_n_u16(vres_wide, 8);
+
+    return vget_lane_u32(vreinterpret_u32_u8(vres), 0);
+}
+
+////////////////////////////////////////////////////////////////////////////////
 // 8 pixels modeprocs
 ////////////////////////////////////////////////////////////////////////////////
 
@@ -755,6 +882,45 @@ SK_COMPILE_ASSERT(
     mode_count_arm
 );
 
+SkXfermodeProc gNEONXfermodeProcs1[] = {
+    NULL, // kClear_Mode
+    NULL, // kSrc_Mode
+    NULL, // kDst_Mode
+    NULL, // kSrcOver_Mode
+    NULL, // kDstOver_Mode
+    NULL, // kSrcIn_Mode
+    NULL, // kDstIn_Mode
+    NULL, // kSrcOut_Mode
+    NULL, // kDstOut_Mode
+    srcatop_modeproc_neon,
+    dstatop_modeproc_neon,
+    xor_modeproc_neon,
+    plus_modeproc_neon,
+    modulate_modeproc_neon,
+    NULL, // kScreen_Mode
+
+    NULL, // kOverlay_Mode
+    NULL, // kDarken_Mode
+    NULL, // kLighten_Mode
+    NULL, // kColorDodge_Mode
+    NULL, // kColorBurn_Mode
+    NULL, // kHardLight_Mode
+    NULL, // kSoftLight_Mode
+    NULL, // kDifference_Mode
+    NULL, // kExclusion_Mode
+    NULL, // kMultiply_Mode
+
+    NULL, // kHue_Mode
+    NULL, // kSaturation_Mode
+    NULL, // kColor_Mode
+    NULL, // kLuminosity_Mode
+};
+
+SK_COMPILE_ASSERT(
+    SK_ARRAY_COUNT(gNEONXfermodeProcs1) == SkXfermode::kLastMode + 1,
+    mode1_count_arm
+);
+
 SkProcCoeffXfermode* SkPlatformXfermodeFactory_impl_neon(const ProcCoeff& rec,
                                                          SkXfermode::Mode mode) {
 
@@ -765,3 +931,7 @@ SkProcCoeffXfermode* SkPlatformXfermodeFactory_impl_neon(const ProcCoeff& rec,
     }
     return NULL;
 }
+
+SkXfermodeProc SkPlatformXfermodeProcFactory_impl_neon(SkXfermode::Mode mode) {
+    return gNEONXfermodeProcs1[mode];
+}
diff --git a/src/opts/SkXfermode_opts_arm_neon.h b/src/opts/SkXfermode_opts_arm_neon.h
index 4c88fc7a63..a8d438195e 100644
--- a/src/opts/SkXfermode_opts_arm_neon.h
+++ b/src/opts/SkXfermode_opts_arm_neon.h
@@ -26,4 +26,10 @@ private:
     typedef SkProcCoeffXfermode INHERITED;
 };
 
+extern SkPMColor srcatop_modeproc_neon(SkPMColor src, SkPMColor dst);
+extern SkPMColor dstatop_modeproc_neon(SkPMColor src, SkPMColor dst);
+extern SkPMColor xor_modeproc_neon(SkPMColor src, SkPMColor dst);
+extern SkPMColor plus_modeproc_neon(SkPMColor src, SkPMColor dst);
+extern SkPMColor modulate_modeproc_neon(SkPMColor src, SkPMColor dst);
+
 #endif //#ifdef SkXfermode_opts_arm_neon_DEFINED
diff --git a/src/opts/SkXfermode_opts_none.cpp b/src/opts/SkXfermode_opts_none.cpp
index ca53fa0dd0..7c46fdd93c 100644
--- a/src/opts/SkXfermode_opts_none.cpp
+++ b/src/opts/SkXfermode_opts_none.cpp
@@ -1,11 +1,17 @@
 #include "SkXfermode.h"
 #include "SkXfermode_proccoeff.h"
 
-// The prototype below is for Clang
+// The prototypes below are for Clang
 extern SkProcCoeffXfermode* SkPlatformXfermodeFactory(const ProcCoeff& rec,
                                                       SkXfermode::Mode mode);
 
+extern SkXfermodeProc SkPlatformXfermodeProcFactory(SkXfermode::Mode mode);
+
 SkProcCoeffXfermode* SkPlatformXfermodeFactory(const ProcCoeff& rec,
                                                SkXfermode::Mode mode) {
     return NULL;
 }
+
+SkXfermodeProc SkPlatformXfermodeProcFactory(SkXfermode::Mode mode) {
+    return NULL;
+}