Support NEON anti-aliasing blit function in SKIA

original function has calculation are called repetitively, and
Processed individual pixels only. So Function call overhead was reduced.
And In contrast to the serial data processing in the existing code,
the Neon code processes 8 or 4 pixels in parallel. In contrast to the
serial data processing in the existing code, the Neon code processes
8 or 4 pixels in parallel

Change-Id: Ib54aeb6bacfd4ed6d257324e8146b211f90b68e9
Signed-off-by: Lee Hwangjae <hj-yo.lee@samsung.com>

2 files changed, 182 insertions, 1 deletions

diff --git a/src/core/SkBlitter_RGB16.cpp b/src/core/SkBlitter_RGB16.cpp
index 8a4d454..c3fb3cf 100644
--- a/src/core/SkBlitter_RGB16.cpp
+++ b/src/core/SkBlitter_RGB16.cpp
@@ -579,13 +579,22 @@ void SkRGB16_Blitter::blitH(int x, int y, int width) {
     blend32_16_row(fSrcColor32, device, width);
 }
 
+#ifdef NEON_BLITANTIH
+extern "C" void blitAntiH_NEON(const SkAlpha* SK_RESTRICT antialias,
+                               uint16_t * SK_RESTRICT device,
+                               const int16_t* SK_RESTRICT runs,
+                               uint32_t srcExpanded, unsigned scale);
+#endif
+
 void SkRGB16_Blitter::blitAntiH(int x, int y,
                                 const SkAlpha* SK_RESTRICT antialias,
                                 const int16_t* SK_RESTRICT runs) {
     uint16_t* SK_RESTRICT device = fDevice.getAddr16(x, y);
     uint32_t    srcExpanded = fExpandedRaw16;
     unsigned    scale = fScale;
-
+#ifdef NEON_BLITANTIH
+    blitAntiH_NEON(antialias, device, runs, srcExpanded, scale);
+#else
     // TODO: respect fDoDither
     for (;;) {
         int count = runs[0];
@@ -609,6 +618,7 @@ void SkRGB16_Blitter::blitAntiH(int x, int y,
         }
         device += count;
     }
+#endif
 }
 
 static inline void blend_8_pixels(U8CPU bw, uint16_t dst[], unsigned dst_scale,
diff --git a/src/core/asm/SkBlitter_RGB16_NEON.S b/src/core/asm/SkBlitter_RGB16_NEON.S
new file mode 100644
index 0000000..715887f
--- /dev/null
+++ b/src/core/asm/SkBlitter_RGB16_NEON.S
@@ -0,0 +1,171 @@
+    .arch armv7-a
+    .text
+    .global blitAntiH_NEON
+    .type   blitAntiH_NEON, %function
+blitAntiH_NEON:
+
+    .fnstart
+
+    @r0     antialias
+    @r1     device
+    @r2     runs
+    @r3     srcExpanded
+    @r4     scale
+
+
+    @r5     count
+    @r6     temp
+    @r7     0x7E0
+    @r8     ~(0x7E0)
+    @r9     scale5
+    @r10    src32
+    @r11    temp2
+    @r12    temp3
+    @r14    count
+
+    @d16    = {0x07E0, 0x07E0, 0x07E0, 0x07E0}
+    @d17    = {0xF81F, 0XF81F, 0XF81F, 0XF81F}
+    @q12    = d16
+    @q13    = d17
+
+
+    stmfd       sp!, {r4-r12,r14}
+    mov         r14, r13
+    add         r14, #40
+    ldr         r4, [r14]
+
+    mov        r5, #0x7E
+    vdup.16    d16, r5
+    vshl.u16   d16, #4
+    vmov.u16   d17, d16
+
+    vmvn.u16   d18, d17
+    vmov.u16   d19, d18
+
+    vmovl.u16  q12, d16
+    vmvn.u32  q13, q12
+
+default:
+    ldrh        r5, [r2]            @r5 = runs[0]
+    cmp         r5, #0              @if (r5 <=0)
+    bls         end_prog            @return
+
+    add         r2, r5, lsl #1      @ runs += (count*2)
+    ldrb        r6, [r0]            @ r6 = antialias[0]
+    add         r0, r5              @ antialias += count
+    cmp         r6, #0              @ if(r6 == 0)
+    addeq       r1, r5, lsl #1      @ device += (count*2)
+    beq         default             @ goto default
+
+    add         r6, #1              @ antialias[0]++
+    mul         r7, r6, r4          @ antialias[0] * scale = r7
+
+    lsr         r8, r7, #11         @ SkAlpha255To256(aa) * scale >> 11
+    mul         r10, r3, r8         @ r10 = scale5 * srcExpanded
+    rsb         r9, r8, #32         @ r9 = 32 - scale5
+    vdup.32     q10, r9             @ q10 = scale5
+    vdup.32     q11, r10            @ q11 = src32
+
+    mov         r7, #0x7E
+    lsl         r7, #4              @ r7 = 0x7E0
+    mvn         r8, r7              @ r8 = ~(0x7E0)
+
+    mov         r14, r5, lsr #3
+    cmp         r14, #0
+    beq         start_fourbytes
+
+eightbytes:
+    vld1.16      {q0}, [r1]       @device load
+    vand.u16    q1, q0, q8     @ c & 0X7E0
+    vand.u16    q2, q0, q9     @ c &~(0x7E0)
+
+    vmovl.u16   q3, d2
+    vmovl.u16   q4, d4         @ conversion from 16bits to 32bits
+    vshl.u32    q5, q3, #16    @ q5 = (c & 0x7E0) << 16
+    vorr.u32    q6, q4, q5     @ q6 = (c & 0x7E0) << 16 | (c & ~(0x7E0))
+    vmul.u32    q7, q6, q10    @ dst32(q7) = scale5 * SkExpand_rgb_16(*device)
+    vadd.u32    q14, q7, q11   @ q14 = dst32 + src32
+    vshr.u32    q15, q14, #5   @ q15 = (src32 + dst32) >> 5
+    vshr.u32    q5, q15, #16   @ q5 = ( (src32 + dst32) >> 5 ) >> 16
+    vand.u32    q6, q5, q12    @ q6 = (c >> 16) & (0x7E0)
+    vand.u32    q7, q15, q13   @ q7 = (c & ~(0x7E0))
+    vorr.u32    q14, q6, q7    @ q14 = ((c >> 16) & 0x7E0) | (c & ~(0x7E0))
+    vmovn.u32   d0, q14
+
+    vmovl.u16   q3, d3
+    vmovl.u16   q4, d5         @ conversion from 16bits to 32bits
+    vshl.u32    q5, q3, #16    @ q5 = (c & 0x7E0) << 16
+    vorr.u32    q6, q4, q5     @ q6 = (c & 0x7E0) << 16 | (c & ~(0x7E0))
+    vmul.u32    q7, q6, q10    @ dst32(q7) = scale5 * SkExpand_rgb_16(*device)
+    vadd.u32    q14, q7, q11   @ q14 = dst32 + src32
+    vshr.u32    q15, q14, #5   @ q15 = (src32 + dst32) >> 5
+    vshr.u32    q5, q15, #16   @ q5 = ( (src32 + dst32) >> 5 ) >> 16
+    vand.u32    q6, q5, q12    @ q6 = (c >> 16) & (0x7E0)
+    vand.u32    q7, q15, q13   @ q7 = (c & ~(0x7E0))
+    vorr.u32    q14, q6, q7    @ q14 = ((c >> 16) & 0x7E0) | (c & ~(0x7E0))
+    vmovn.u32   d1, q14
+
+    vst1.16     {q0}, [r1]
+
+    add         r1, #16                  @ device++
+    subs        r14, r14, #1
+    bhi         eightbytes
+
+start_fourbytes:
+    and         r14, r5, #4
+    cmp         r14, #0
+    beq         start_onebyte
+
+fourbytes:
+    vld1.16     d0, [r1]            @ device load
+    vand.u16    d1, d0, d16         @ c & 0x7E0
+    vand.u16    d2, d0, d18         @ c & ~(0x7E0)
+    vmovl.u16   q3, d1
+    vmovl.u16   q4, d2              @ conversion from 16bits to 32bits
+
+    vshl.u32    q5, q3, #16         @ q5 = (c & 0x7E0) << 16
+    vorr.u32    q6, q4, q5          @ q6 = (c & 0x7E0) << 16 | (c & ~(0x7E0))
+    vmul.u32    q7, q6, q10         @ dst32(q7) = scale5 * SkExpand_rgb_16(*device)
+    vadd.u32    q3, q11, q7         @ q3 = dst32 + src32
+    vshr.u32    q1, q3, #5          @ q1 = (src32 + dst32) >> 5
+    vshr.u32    q2, q1, #16         @ q2 = ( (src32 + dst32) >> 5 ) >> 16
+
+    vand.u32    q3, q2, q12         @ q3 = (c >> 16) & (0x7E0)
+    vand.u32    q4, q1, q13         @ q4 = (c & ~(0x7E0))
+
+    vorr.u32    q5, q4, q3          @ q6 = ((c >> 16) & 0x7E0) | (c & ~(0x7E0))
+    vmovn.u32   d0, q5
+    vst1.16     d0, [r1]
+    add         r1, #8
+
+
+start_onebyte:
+    and         r14, r5, #3
+    cmp         r14, #0
+    beq         default
+onebyte:
+    ldrh        r6, [r1]            @ device load
+    and         r5, r6, r7          @ r5 = c & 0x7E0
+    and         r11, r6, r8         @ r11 = c & (~0x7E0)
+    orr         r12, r11, r5, lsl #16  @ r12 = (c & (~0x7E0) ) | ((c & 0x7E0) << 16)
+
+    mul         r6, r9, r12             @dst32(r5) = scale5 * SkExpand_rgb_16(*device)
+
+    add         r5, r6, r10             @ src32 + dst32
+    lsr         r6, r5, #5              @ (src32 + dst32) >> 5
+    and         r12, r7, r6, lsr #16    @ (c>>16) & 2016
+    and         r11, r8, r6             @ (c & ~(2016))
+    orr         r5, r11, r12            @ (c & ~(2016) | ( (c>>16) & 2016)
+
+    strh        r5, [r1]                @ *device = r5
+    add         r1, #2                  @ device++
+    subs        r14, r14, #1
+    bhi         onebyte
+    b           default
+
+
+end_prog:
+    ldmfd       sp!, {r4-r12, r14}
+    mov         pc, lr
+
+.fnend