Databases
StackOverflow2013
Postsdbo
POOptimizing RGBA8888 to RGB565 conversion with NEON
Body
 

Note that there are some explanatory texts on larger screens.

plurals
  1. POOptimizing RGBA8888 to RGB565 conversion with NEON
    text
    copied!<p>I'm trying to optimize an image format conversion on iOS using the NEON vector instruction set. I assumed this would map well to that because it processes a bunch of similar data.</p> <p>My attempts haven't gone that well, though, achieving only a marginal speedup vs the naive c implementation:</p> <pre><code>for(int i = 0; i &lt; pixelCount; ++i, ++inPixel32) { const unsigned int r = ((*inPixel32 &gt;&gt; 0 ) &amp; 0xFF); const unsigned int g = ((*inPixel32 &gt;&gt; 8 ) &amp; 0xFF); const unsigned int b = ((*inPixel32 &gt;&gt; 16) &amp; 0xFF); *outPixel16++ = ((r &gt;&gt; 3) &lt;&lt; 11) | ((g &gt;&gt; 2) &lt;&lt; 5) | ((b &gt;&gt; 3) &lt;&lt; 0); } </code></pre> <p>1 megapixel image array on iPad 2:</p> <p>format is [min avg max n=number of timer samples] in milliseconds</p> <p>C: [14.446 14.632 18.405 n=1000]ms</p> <p>NEON: [11.920 12.032 15.336 n=1000]ms</p> <p>My attempt at a NEON implementation is below:</p> <pre><code> int i; const int pixelsPerLoop = 8; for(i = 0; i &lt; pixelCount; i += pixelsPerLoop, inPixel32 += pixelsPerLoop, outPixel16 += pixelsPerLoop) { //Read all r,g,b pixels into 3 registers uint8x8x4_t rgba = vld4_u8(inPixel32); //Right-shift r,g,b as appropriate uint8x8_t r = vshr_n_u8(rgba.val[0], 3); uint8x8_t g = vshr_n_u8(rgba.val[1], 2); uint8x8_t b = vshr_n_u8(rgba.val[2], 3); //Widen b uint16x8_t r5_g6_b5 = vmovl_u8(b); //Widen r uint16x8_t r16 = vmovl_u8(r); //Left shift into position within 16-bit int r16 = vshlq_n_u16(r16, 11); r5_g6_b5 |= r16; //Widen g uint16x8_t g16 = vmovl_u8(g); //Left shift into position within 16-bit int g16 = vshlq_n_u16(g16, 5); r5_g6_b5 |= g16; //Now write back to memory vst1q_u16(outPixel16, r5_g6_b5); } //Do the remainder on normal flt hardware </code></pre> <p>Code was compiled via LLVM 3.0 into the following (.loc and extra labels removed):</p> <pre><code>_DNConvert_ARGB8888toRGB565: push {r4, r5, r7, lr} mov r9, r1 mov.w r12, #0 add r7, sp, #8 cmp r2, #0 mov.w r1, #0 it ne movne r1, #1 cmp r0, #0 mov.w r3, #0 it ne movne r3, #1 cmp.w r9, #0 mov.w r4, #0 it ne movne r4, #1 tst.w r9, #3 bne LBB0_8 ands r1, r3 ands r1, r4 cmp r1, #1 bne LBB0_8 movs r1, #0 lsr.w lr, r9, #2 cmp.w r1, r9, lsr #2 bne LBB0_9 mov r3, r2 mov r5, r0 b LBB0_5 LBB0_4: movw r1, #65528 add.w r0, lr, #7 movt r1, #32767 ands r1, r0 LBB0_5: mov.w r12, #1 cmp r1, lr bhs LBB0_8 rsb r0, r1, r9, lsr #2 mov.w r9, #63488 mov.w lr, #2016 mov.w r12, #1 LBB0_7: ldr r2, [r5], #4 subs r0, #1 and.w r1, r9, r2, lsl #8 and.w r4, lr, r2, lsr #5 ubfx r2, r2, #19, #5 orr.w r2, r2, r4 orr.w r1, r1, r2 strh r1, [r3], #2 bne LBB0_7 LBB0_8: mov r0, r12 pop {r4, r5, r7, pc} LBB0_9: sub.w r1, lr, #1 movs r3, #32 add.w r3, r3, r1, lsl #2 bic r3, r3, #31 adds r5, r0, r3 movs r3, #16 add.w r1, r3, r1, lsl #1 bic r1, r1, #15 adds r3, r2, r1 movs r1, #0 LBB0_10: vld4.8 {d16, d17, d18, d19}, [r0]! adds r1, #8 cmp r1, lr vshr.u8 d20, d16, #3 vshr.u8 d21, d17, #2 vshr.u8 d16, d18, #3 vmovl.u8 q11, d20 vmovl.u8 q9, d21 vmovl.u8 q8, d16 vshl.i16 q10, q11, #11 vshl.i16 q9, q9, #5 vorr q8, q8, q10 vorr q8, q8, q9 vst1.16 {d16, d17}, [r2]! Ltmp28: blo LBB0_10 b LBB0_4 </code></pre> <p>Full code is available at <a href="https://github.com/darknoon/DNImageConvert" rel="noreferrer">https://github.com/darknoon/DNImageConvert</a> I would appreciate any help, thanks!</p>
 

Querying!

 
Guidance
An individual column

Larger individual text columns get their own page to allow for proper reading.

SQuiL has stopped working due to an internal error.

If you are curious you may find further information in the browser console, which is accessible through the devtools (F12).

Reload