Issue 2381493006: HalfFloat_SSE2 port from C algorithm to SSE2

	Unified diffs	Side-by-side diffs	Delta from patch set	Stats (+69 lines, -3 lines)			Patch
M	README.chromium	View	1 2 3 4 5 6	1 chunk	+1 line, -1 line	0 comments	Download
M	include/libyuv/row.h	View	1 2 3 4 5	2 chunks	+10 lines, -0 lines	0 comments	Download
M	include/libyuv/version.h	View	1 2 3 4 5 6	1 chunk	+1 line, -1 line	0 comments	Download
M	source/planar_functions.cc	View	1 2 3 4 5 6	2 chunks	+16 lines, -0 lines	0 comments	Download
M	source/row_any.cc	View		1 chunk	+3 lines, -0 lines	0 comments	Download
M	source/row_gcc.cc	View	1 2 3 4	2 chunks	+38 lines, -1 line	0 comments	Download

Messages

Total messages: 8 (1 generated)

Expand Messages | Collapse Messages | Show Generated Messages | Hide Generated Messages

hubbe

LGTM Not really qualified to review the SSE2 code though. Is it fast? Does it ...

4 years, 2 months ago (2016-09-30 00:14:10 UTC) #2

fbarchard1

For 10k iterations SSE2 TestHalfFloatPlane (4638 ms) C TestHalfFloatPlane (7479 ms) SSE2 code is: 0000000000000000 ...

4 years, 2 months ago (2016-09-30 00:54:13 UTC) #4

For 10k iterations
SSE2 TestHalfFloatPlane (4638 ms)
C TestHalfFloatPlane (7479 ms)

SSE2 code is:
0000000000000000 <HalfFloatRow_SSE2>:
   0:	f3 0f 59 05 00 00 00 	mulss  0x0(%rip),%xmm0        # 8
<HalfFloatRow_SSE2+0x8>
   7:	00 
   8:	f3 0f 11 44 24 fc    	movss  %xmm0,-0x4(%rsp)
   e:	66 0f 6e 64 24 fc    	movd   -0x4(%rsp),%xmm4
  14:	66 0f 70 e4 00       	pshufd $0x0,%xmm4,%xmm4
  19:	66 0f ef ed          	pxor   %xmm5,%xmm5
  1d:	f3 0f 6f 07          	movdqu (%rdi),%xmm0
  21:	48 8d 7f 10          	lea    0x10(%rdi),%rdi
  25:	66 0f 6f c8          	movdqa %xmm0,%xmm1
  29:	66 0f 61 c5          	punpcklwd %xmm5,%xmm0
  2d:	66 0f 69 cd          	punpckhwd %xmm5,%xmm1
  31:	0f 5b c0             	cvtdq2ps %xmm0,%xmm0
  34:	0f 5b c9             	cvtdq2ps %xmm1,%xmm1
  37:	0f 59 c4             	mulps  %xmm4,%xmm0
  3a:	0f 59 cc             	mulps  %xmm4,%xmm1
  3d:	66 0f 72 d0 0d       	psrld  $0xd,%xmm0
  42:	66 0f 72 d1 0d       	psrld  $0xd,%xmm1
  47:	66 0f 6b c1          	packssdw %xmm1,%xmm0
  4b:	f3 0f 7f 06          	movdqu %xmm0,(%rsi)
  4f:	48 8d 76 10          	lea    0x10(%rsi),%rsi
  53:	83 ea 08             	sub    $0x8,%edx
  56:	7f c5                	jg     1d <HalfFloatRow_SSE2+0x1d>
  58:	c3                   	retq   

C code is:

0000000000000000 <HalfFloatRow_C>:
   0:	85 d2                	test   %edx,%edx
   2:	0f 8e 07 02 00 00    	jle    20f <HalfFloatRow_C+0x20f>
   8:	f3 0f 59 05 00 00 00 	mulss  0x0(%rip),%xmm0        # 10
<HalfFloatRow_C+0x10>
   f:	00 
  10:	8d 4a ff             	lea    -0x1(%rdx),%ecx
  13:	4c 8d 49 01          	lea    0x1(%rcx),%r9
  17:	45 31 db             	xor    %r11d,%r11d
  1a:	49 83 f9 08          	cmp    $0x8,%r9
  1e:	0f 82 60 01 00 00    	jb     184 <HalfFloatRow_C+0x184>
  24:	45 31 db             	xor    %r11d,%r11d
  27:	49 b8 f8 ff ff ff 01 	movabs $0x1fffffff8,%r8
  2e:	00 00 00 
  31:	4d 21 c8             	and    %r9,%r8
  34:	0f 84 4a 01 00 00    	je     184 <HalfFloatRow_C+0x184>
  3a:	4c 8d 14 4f          	lea    (%rdi,%rcx,2),%r10
  3e:	45 31 db             	xor    %r11d,%r11d
  41:	49 39 f2             	cmp    %rsi,%r10
  44:	72 0d                	jb     53 <HalfFloatRow_C+0x53>
  46:	48 8d 0c 4e          	lea    (%rsi,%rcx,2),%rcx
  4a:	48 39 f9             	cmp    %rdi,%rcx
  4d:	0f 83 31 01 00 00    	jae    184 <HalfFloatRow_C+0x184>
  53:	0f 28 c8             	movaps %xmm0,%xmm1
  56:	0f c6 c9 00          	shufps $0x0,%xmm1,%xmm1
  5a:	49 8d 40 f8          	lea    -0x8(%r8),%rax
  5e:	49 89 c2             	mov    %rax,%r10
  61:	49 c1 ea 03          	shr    $0x3,%r10
  65:	31 c9                	xor    %ecx,%ecx
  67:	48 0f ba e0 03       	bt     $0x3,%rax
  6c:	72 4b                	jb     b9 <HalfFloatRow_C+0xb9>
  6e:	f3 0f 6f 17          	movdqu (%rdi),%xmm2
  72:	66 0f ef db          	pxor   %xmm3,%xmm3
  76:	66 0f 6f e2          	movdqa %xmm2,%xmm4
  7a:	66 0f 61 e3          	punpcklwd %xmm3,%xmm4
  7e:	0f 5b e4             	cvtdq2ps %xmm4,%xmm4
  81:	66 0f 69 d3          	punpckhwd %xmm3,%xmm2
  85:	0f 5b d2             	cvtdq2ps %xmm2,%xmm2
  88:	0f 59 d1             	mulps  %xmm1,%xmm2
  8b:	0f 59 e1             	mulps  %xmm1,%xmm4
  8e:	66 0f 72 d4 0d       	psrld  $0xd,%xmm4
  93:	66 0f 72 d2 0d       	psrld  $0xd,%xmm2
  98:	66 0f 72 f2 10       	pslld  $0x10,%xmm2
  9d:	66 0f 72 e2 10       	psrad  $0x10,%xmm2
  a2:	66 0f 72 f4 10       	pslld  $0x10,%xmm4
  a7:	66 0f 72 e4 10       	psrad  $0x10,%xmm4
  ac:	66 0f 6b e2          	packssdw %xmm2,%xmm4
  b0:	f3 0f 7f 26          	movdqu %xmm4,(%rsi)
  b4:	b9 08 00 00 00       	mov    $0x8,%ecx
  b9:	4d 85 d2             	test   %r10,%r10
  bc:	0f 84 b6 00 00 00    	je     178 <HalfFloatRow_C+0x178>
  c2:	4d 89 c2             	mov    %r8,%r10
  c5:	49 29 ca             	sub    %rcx,%r10
  c8:	48 8d 44 4f 10       	lea    0x10(%rdi,%rcx,2),%rax
  cd:	48 8d 4c 4e 10       	lea    0x10(%rsi,%rcx,2),%rcx
  d2:	66 0f ef d2          	pxor   %xmm2,%xmm2
  d6:	66 2e 0f 1f 84 00 00 	nopw   %cs:0x0(%rax,%rax,1)
  dd:	00 00 00 
  e0:	f3 0f 6f 58 f0       	movdqu -0x10(%rax),%xmm3
  e5:	66 0f 6f e3          	movdqa %xmm3,%xmm4
  e9:	66 0f 61 e2          	punpcklwd %xmm2,%xmm4
  ed:	0f 5b e4             	cvtdq2ps %xmm4,%xmm4
  f0:	66 0f 69 da          	punpckhwd %xmm2,%xmm3
  f4:	0f 5b db             	cvtdq2ps %xmm3,%xmm3
  f7:	0f 59 d9             	mulps  %xmm1,%xmm3
  fa:	0f 59 e1             	mulps  %xmm1,%xmm4
  fd:	66 0f 72 d4 0d       	psrld  $0xd,%xmm4
 102:	66 0f 72 d3 0d       	psrld  $0xd,%xmm3
 107:	66 0f 72 f3 10       	pslld  $0x10,%xmm3
 10c:	66 0f 72 e3 10       	psrad  $0x10,%xmm3
 111:	66 0f 72 f4 10       	pslld  $0x10,%xmm4
 116:	66 0f 72 e4 10       	psrad  $0x10,%xmm4
 11b:	66 0f 6b e3          	packssdw %xmm3,%xmm4
 11f:	f3 0f 7f 61 f0       	movdqu %xmm4,-0x10(%rcx)
 124:	f3 0f 6f 18          	movdqu (%rax),%xmm3
 128:	66 0f 6f e3          	movdqa %xmm3,%xmm4
 12c:	66 0f 61 e2          	punpcklwd %xmm2,%xmm4
 130:	0f 5b e4             	cvtdq2ps %xmm4,%xmm4
 133:	66 0f 69 da          	punpckhwd %xmm2,%xmm3
 137:	0f 5b db             	cvtdq2ps %xmm3,%xmm3
 13a:	0f 59 d9             	mulps  %xmm1,%xmm3
 13d:	0f 59 e1             	mulps  %xmm1,%xmm4
 140:	66 0f 72 d4 0d       	psrld  $0xd,%xmm4
 145:	66 0f 72 d3 0d       	psrld  $0xd,%xmm3
 14a:	66 0f 72 f3 10       	pslld  $0x10,%xmm3
 14f:	66 0f 72 e3 10       	psrad  $0x10,%xmm3
 154:	66 0f 72 f4 10       	pslld  $0x10,%xmm4
 159:	66 0f 72 e4 10       	psrad  $0x10,%xmm4
 15e:	66 0f 6b e3          	packssdw %xmm3,%xmm4
 162:	f3 0f 7f 21          	movdqu %xmm4,(%rcx)
 166:	48 83 c0 20          	add    $0x20,%rax
 16a:	48 83 c1 20          	add    $0x20,%rcx
 16e:	49 83 c2 f0          	add    $0xfffffffffffffff0,%r10
 172:	0f 85 68 ff ff ff    	jne    e0 <HalfFloatRow_C+0xe0>
 178:	4d 39 c1             	cmp    %r8,%r9
 17b:	4d 89 c3             	mov    %r8,%r11
 17e:	0f 84 8b 00 00 00    	je     20f <HalfFloatRow_C+0x20f>
 184:	8d 42 01             	lea    0x1(%rdx),%eax
 187:	44 89 d9             	mov    %r11d,%ecx
 18a:	83 c9 01             	or     $0x1,%ecx
 18d:	29 c8                	sub    %ecx,%eax
 18f:	a8 01                	test   $0x1,%al
 191:	74 20                	je     1b3 <HalfFloatRow_C+0x1b3>
 193:	42 0f b7 04 5f       	movzwl (%rdi,%r11,2),%eax
 198:	0f 57 c9             	xorps  %xmm1,%xmm1
 19b:	f3 0f 2a c8          	cvtsi2ss %eax,%xmm1
 19f:	f3 0f 59 c8          	mulss  %xmm0,%xmm1
 1a3:	66 0f 7e c8          	movd   %xmm1,%eax
 1a7:	c1 e8 0d             	shr    $0xd,%eax
 1aa:	66 42 89 04 5e       	mov    %ax,(%rsi,%r11,2)
 1af:	49 83 cb 01          	or     $0x1,%r11
 1b3:	39 d1                	cmp    %edx,%ecx
 1b5:	74 58                	je     20f <HalfFloatRow_C+0x20f>
 1b7:	44 29 da             	sub    %r11d,%edx
 1ba:	4a 8d 4c 5f 02       	lea    0x2(%rdi,%r11,2),%rcx
 1bf:	4a 8d 44 5e 02       	lea    0x2(%rsi,%r11,2),%rax
 1c4:	66 66 66 2e 0f 1f 84 	data32 data32 nopw %cs:0x0(%rax,%rax,1)
 1cb:	00 00 00 00 00 
 1d0:	0f b7 71 fe          	movzwl -0x2(%rcx),%esi
 1d4:	0f 57 c9             	xorps  %xmm1,%xmm1
 1d7:	f3 0f 2a ce          	cvtsi2ss %esi,%xmm1
 1db:	f3 0f 59 c8          	mulss  %xmm0,%xmm1
 1df:	66 0f 7e ce          	movd   %xmm1,%esi
 1e3:	c1 ee 0d             	shr    $0xd,%esi
 1e6:	66 89 70 fe          	mov    %si,-0x2(%rax)
 1ea:	0f b7 31             	movzwl (%rcx),%esi
 1ed:	0f 57 c9             	xorps  %xmm1,%xmm1
 1f0:	f3 0f 2a ce          	cvtsi2ss %esi,%xmm1
 1f4:	f3 0f 59 c8          	mulss  %xmm0,%xmm1
 1f8:	66 0f 7e ce          	movd   %xmm1,%esi
 1fc:	c1 ee 0d             	shr    $0xd,%esi
 1ff:	66 89 30             	mov    %si,(%rax)
 202:	48 83 c1 04          	add    $0x4,%rcx
 206:	48 83 c0 04          	add    $0x4,%rax
 20a:	83 c2 fe             	add    $0xfffffffe,%edx
 20d:	75 c1                	jne    1d0 <HalfFloatRow_C+0x1d0>
 20f:	c3                   	retq

fbarchard1

note the mulps could be replaced with an integer subtract of 112 from the exponent. ...

4 years, 2 months ago (2016-09-30 00:59:58 UTC) #5

fbarchard1

tested on prescott emultator /usr/local/google/home/fbarchard/intelsde/sde -p4p -- out/Release/libyuv_unittest --gtest_filter=*Half* Note: Google Test filter = *Half* ...

4 years, 2 months ago (2016-09-30 01:06:52 UTC) #6

fbarchard1

4 years, 2 months ago (2016-09-30 16:47:21 UTC) #8

Message was sent while issue was closed.

Committed patchset #7 (id:120001) manually as
4a14cb2e81235ecd656e799aecaaf139db8ce4a2 (presubmit successful).

Expand Messages | Collapse Messages | Show Generated Messages | Hide Generated Messages