ARM Skia NEON patches - 27 - S32A_D565_Blend

BlitRow565: new intrinsics version of S32A_D565_Blend

This new version is basically a rewrite of the existing code with
a few speed and accuracy improvements. There is a switch to enable
pixel perfect results at the cost of a (quite big) decrease of
performances (disabled in this patch).

Here are the benchmark results (speedup vs. existing code):

+-------+------------+------------+
| count | Cortex -A9 | Cortex-A15 |
+-------+------------+------------+
| 1     | +103.6%    | +12%       |
+-------+------------+------------+
| 2     | +3.6%      | +21.6%     |
+-------+------------+------------+
| 4     | +0.8%      | -0.8%      |
+-------+------------+------------+
| 8     | +3.9%      | -1%        |
+-------+------------+------------+
| 16    | +14.7%     | +5.7%      |
+-------+------------+------------+
| 64    | +18.1%     | +13.2%     |
+-------+------------+------------+
| 256   | +16.3%     | +27.4%     |
+-------+------------+------------+
| 1024  | +78.2%     | +17.4%     |
+-------+------------+------------+

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

BUG=skia:
R=djsollen@google.com, mtklein@google.com, halcanary@google.com

Author: kevin.petit@arm.com

Review URL: https://codereview.chromium.org/156113005

git-svn-id: http://skia.googlecode.com/svn/trunk@13438 2bbb7eff-a529-9590-31e7-b0007b416f81
This commit is contained in:
commit-bot@chromium.org 2014-02-13 18:37:35 +00:00
parent 35d4872f9a
commit be233d63ca
2 changed files with 113 additions and 98 deletions

View File

@ -37,3 +37,8 @@
# deprecated calling pattern.
# https://codereview.chromium.org/154163002/
extractbitmap
# Added by kevin.petit@arm.com for https://codereview.chromium.org/156113005/
shadertext3
gradients_view_perspective
drawbitmaprect

View File

@ -230,113 +230,123 @@ void S32A_D565_Opaque_neon(uint16_t* SK_RESTRICT dst,
}
}
static inline uint16x8_t SkDiv255Round_neon8(uint16x8_t prod) {
prod += vdupq_n_u16(128);
prod += vshrq_n_u16(prod, 8);
return vshrq_n_u16(prod, 8);
}
void S32A_D565_Blend_neon(uint16_t* SK_RESTRICT dst,
const SkPMColor* SK_RESTRICT src, int count,
U8CPU alpha, int /*x*/, int /*y*/) {
SkASSERT(255 > alpha);
U8CPU alpha_for_asm = alpha;
asm volatile (
/* This code implements a Neon version of S32A_D565_Blend. The output differs from
* the original in two respects:
* 1. The results have a few mismatches compared to the original code. These mismatches
* never exceed 1. It's possible to improve accuracy vs. a floating point
* implementation by introducing rounding right shifts (vrshr) for the final stage.
* Rounding is not present in the code below, because although results would be closer
* to a floating point implementation, the number of mismatches compared to the
* original code would be far greater.
* 2. On certain inputs, the original code can overflow, causing colour channels to
* mix. Although the Neon code can also overflow, it doesn't allow one colour channel
* to affect another.
/* This code implements a Neon version of S32A_D565_Blend. The results have
* a few mismatches compared to the original code. These mismatches never
* exceed 1.
*/
#if 1
/* reflects SkAlpha255To256()'s change from a+a>>7 to a+1 */
"add %[alpha], %[alpha], #1 \n\t" // adjust range of alpha 0-256
#else
"add %[alpha], %[alpha], %[alpha], lsr #7 \n\t" // adjust range of alpha 0-256
#endif
"vmov.u16 q3, #255 \n\t" // set up constant
"movs r4, %[count], lsr #3 \n\t" // calc. count>>3
"vmov.u16 d2[0], %[alpha] \n\t" // move alpha to Neon
"beq 2f \n\t" // if count8 == 0, exit
"vmov.u16 q15, #0x1f \n\t" // set up blue mask
if (count >= 8) {
uint16x8_t valpha_max, vmask_blue;
uint8x8_t valpha;
"1: \n\t"
"vld1.u16 {d0, d1}, [%[dst]] \n\t" // load eight dst RGB565 pixels
"subs r4, r4, #1 \n\t" // decrement loop counter
"vld4.u8 {d24, d25, d26, d27}, [%[src]]! \n\t" // load eight src ABGR32 pixels
// and deinterleave
// prepare constants
valpha_max = vmovq_n_u16(255);
valpha = vdup_n_u8(alpha);
vmask_blue = vmovq_n_u16(SK_B16_MASK);
"vshl.u16 q9, q0, #5 \n\t" // shift green to top of lanes
"vand q10, q0, q15 \n\t" // extract blue
"vshr.u16 q8, q0, #11 \n\t" // extract red
"vshr.u16 q9, q9, #10 \n\t" // extract green
// dstrgb = {q8, q9, q10}
"vshr.u8 d24, d24, #3 \n\t" // shift red to 565 range
"vshr.u8 d25, d25, #2 \n\t" // shift green to 565 range
"vshr.u8 d26, d26, #3 \n\t" // shift blue to 565 range
"vmovl.u8 q11, d24 \n\t" // widen red to 16 bits
"vmovl.u8 q12, d25 \n\t" // widen green to 16 bits
"vmovl.u8 q14, d27 \n\t" // widen alpha to 16 bits
"vmovl.u8 q13, d26 \n\t" // widen blue to 16 bits
// srcrgba = {q11, q12, q13, q14}
"vmul.u16 q2, q14, d2[0] \n\t" // sa * src_scale
"vmul.u16 q11, q11, d2[0] \n\t" // red result = src_red * src_scale
"vmul.u16 q12, q12, d2[0] \n\t" // grn result = src_grn * src_scale
"vmul.u16 q13, q13, d2[0] \n\t" // blu result = src_blu * src_scale
"vshr.u16 q2, q2, #8 \n\t" // sa * src_scale >> 8
"vsub.u16 q2, q3, q2 \n\t" // 255 - (sa * src_scale >> 8)
// dst_scale = q2
"vmla.u16 q11, q8, q2 \n\t" // red result += dst_red * dst_scale
"vmla.u16 q12, q9, q2 \n\t" // grn result += dst_grn * dst_scale
"vmla.u16 q13, q10, q2 \n\t" // blu result += dst_blu * dst_scale
#if 1
// trying for a better match with SkDiv255Round(a)
// C alg is: a+=128; (a+a>>8)>>8
// we'll use just a rounding shift [q2 is available for scratch]
"vrshr.u16 q11, q11, #8 \n\t" // shift down red
"vrshr.u16 q12, q12, #8 \n\t" // shift down green
"vrshr.u16 q13, q13, #8 \n\t" // shift down blue
#else
// arm's original "truncating divide by 256"
"vshr.u16 q11, q11, #8 \n\t" // shift down red
"vshr.u16 q12, q12, #8 \n\t" // shift down green
"vshr.u16 q13, q13, #8 \n\t" // shift down blue
#endif
"vsli.u16 q13, q12, #5 \n\t" // insert green into blue
"vsli.u16 q13, q11, #11 \n\t" // insert red into green/blue
"vst1.16 {d26, d27}, [%[dst]]! \n\t" // write pixel back to dst, update ptr
"bne 1b \n\t" // if counter != 0, loop
"2: \n\t" // exit
: [src] "+r" (src), [dst] "+r" (dst), [count] "+r" (count), [alpha] "+r" (alpha_for_asm)
:
: "cc", "memory", "r4", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "d16", "d17", "d18", "d19", "d20", "d21", "d22", "d23", "d24", "d25", "d26", "d27", "d28", "d29", "d30", "d31"
);
count &= 7;
if (count > 0) {
do {
SkPMColor sc = *src++;
if (sc) {
uint16_t dc = *dst;
unsigned dst_scale = 255 - SkMulDiv255Round(SkGetPackedA32(sc), alpha);
unsigned dr = SkMulS16(SkPacked32ToR16(sc), alpha) + SkMulS16(SkGetPackedR16(dc), dst_scale);
unsigned dg = SkMulS16(SkPacked32ToG16(sc), alpha) + SkMulS16(SkGetPackedG16(dc), dst_scale);
unsigned db = SkMulS16(SkPacked32ToB16(sc), alpha) + SkMulS16(SkGetPackedB16(dc), dst_scale);
*dst = SkPackRGB16(SkDiv255Round(dr), SkDiv255Round(dg), SkDiv255Round(db));
}
dst += 1;
} while (--count != 0);
uint16x8_t vdst, vdst_r, vdst_g, vdst_b;
uint16x8_t vres_a, vres_r, vres_g, vres_b;
uint8x8x4_t vsrc;
// load pixels
vdst = vld1q_u16(dst);
#if (__GNUC__ > 4) || ((__GNUC__ == 4) && (__GNUC_MINOR__ > 6))
asm (
"vld4.u8 %h[vsrc], [%[src]]!"
: [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
// deinterleave dst
vdst_g = vshlq_n_u16(vdst, SK_R16_BITS); // shift green to top of lanes
vdst_b = vdst & vmask_blue; // extract blue
vdst_r = vshrq_n_u16(vdst, SK_R16_SHIFT); // extract red
vdst_g = vshrq_n_u16(vdst_g, SK_R16_BITS + SK_B16_BITS); // extract green
// shift src to 565
vsrc.val[NEON_R] = vshr_n_u8(vsrc.val[NEON_R], 8 - SK_R16_BITS);
vsrc.val[NEON_G] = vshr_n_u8(vsrc.val[NEON_G], 8 - SK_G16_BITS);
vsrc.val[NEON_B] = vshr_n_u8(vsrc.val[NEON_B], 8 - SK_B16_BITS);
// calc src * src_scale
vres_a = vmull_u8(vsrc.val[NEON_A], valpha);
vres_r = vmull_u8(vsrc.val[NEON_R], valpha);
vres_g = vmull_u8(vsrc.val[NEON_G], valpha);
vres_b = vmull_u8(vsrc.val[NEON_B], valpha);
// prepare dst_scale
vres_a = SkDiv255Round_neon8(vres_a);
vres_a = valpha_max - vres_a; // 255 - (sa * src_scale) / 255
// add dst * dst_scale to previous result
vres_r = vmlaq_u16(vres_r, vdst_r, vres_a);
vres_g = vmlaq_u16(vres_g, vdst_g, vres_a);
vres_b = vmlaq_u16(vres_b, vdst_b, vres_a);
#ifdef S32A_D565_BLEND_EXACT
// It is possible to get exact results with this but it is slow,
// even slower than C code in some cases
vres_r = SkDiv255Round_neon8(vres_r);
vres_g = SkDiv255Round_neon8(vres_g);
vres_b = SkDiv255Round_neon8(vres_b);
#else
vres_r = vrshrq_n_u16(vres_r, 8);
vres_g = vrshrq_n_u16(vres_g, 8);
vres_b = vrshrq_n_u16(vres_b, 8);
#endif
// pack result
vres_b = vsliq_n_u16(vres_b, vres_g, SK_G16_SHIFT); // insert green into blue
vres_b = vsliq_n_u16(vres_b, vres_r, SK_R16_SHIFT); // insert red into green/blue
// store
vst1q_u16(dst, vres_b);
dst += 8;
count -= 8;
} while (count >= 8);
}
// leftovers
while (count-- > 0) {
SkPMColor sc = *src++;
if (sc) {
uint16_t dc = *dst;
unsigned dst_scale = 255 - SkMulDiv255Round(SkGetPackedA32(sc), alpha);
unsigned dr = SkMulS16(SkPacked32ToR16(sc), alpha) + SkMulS16(SkGetPackedR16(dc), dst_scale);
unsigned dg = SkMulS16(SkPacked32ToG16(sc), alpha) + SkMulS16(SkGetPackedG16(dc), dst_scale);
unsigned db = SkMulS16(SkPacked32ToB16(sc), alpha) + SkMulS16(SkGetPackedB16(dc), dst_scale);
*dst = SkPackRGB16(SkDiv255Round(dr), SkDiv255Round(dg), SkDiv255Round(db));
}
dst += 1;
}
}