HalfFloat Neon for ARMv7.

HalfFloat Neon for ARMv7.

64 bit version made similar to 32 bit with registers 1 for load and store results, and 2 and 3 as expanded float temporary values.

TEST=out/Release/libyuv_unittest --gtest_filter=*Half*

BUG=libyuv:560
R=wangcheng@google.com

Committed: https://chromium.googlesource.com/libyuv/libyuv/+/eca08525cb0bda9b21da786240cf841340fbe575

[fbarchard1, 2016-11-01 02:34:21]

Was
TestHalfFloatPlane_16bit_denormal (2912 ms)
TestHalfFloatPlane_10bit_Opt (2922 ms)
TestHalfFloatPlane_9bit_Opt (2943 ms)
TestHalfFloatPlane_Opt (2916 ms)
TestHalfFloatPlane_One (2935 ms)
TestHalfFloatPlane_Offby1 (2949 ms)

Now
TestHalfFloatPlane_16bit_denormal (1125 ms)
TestHalfFloatPlane_10bit_Opt (1036 ms)
TestHalfFloatPlane_9bit_Opt (1035 ms)
TestHalfFloatPlane_Opt (1019 ms)
TestHalfFloatPlane_One (972 ms)
TestHalfFloatPlane_Offby1 (1026 ms)

[fbarchard1, 2016-11-01 02:43:43]

ios build complains:

../../third_party/llvm-build/Release+Asserts/bin/clang++ -MMD -MF
obj/libyuv_neon/row_neon.o.d -DV8_DEPRECATION_WARNINGS -DDCHECK_ALWAYS_ON=1
-DNO_TCMALLOC -DDISABLE_NACL -DCHROMIUM_BUILD -DFIELDTRIAL_TESTING_ENABLED
-DCR_CLANG_REVISION=277962-1 -D__STDC_CONSTANT_MACROS -D__STDC_FORMAT_MACROS
-D_DEBUG -DDYNAMIC_ANNOTATIONS_ENABLED=1 -DWTF_USE_DYNAMIC_ANNOTATIONS=1 -I../..
-Igen -I../../include -mfpu=neon -fno-strict-aliasing --param=ssp-buffer-size=4
-fstack-protector -fcolor-diagnostics -arch armv7 -Wall -Werror -Wextra
-Wno-missing-field-initializers -Wno-unused-parameter -Wno-c++11-narrowing
-Wno-covered-switch-default -Wno-deprecated-register
-Wno-unneeded-internal-declaration -Wno-inconsistent-missing-override
-Wno-shift-negative-value -Wno-undefined-var-template
-Wno-nonportable-include-path -O0 -gdwarf-2 -isysroot
/Applications/Xcode7.0.app/Contents/Developer/Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS9.0.sdk
-stdlib=libc++ -miphoneos-version-min=9.0 -fvisibility=hidden -Xclang -load
-Xclang
../../third_party/llvm-build/Release+Asserts/lib/libFindBadConstructs.dylib
-Xclang -add-plugin -Xclang find-bad-constructs -Xclang
-plugin-arg-find-bad-constructs -Xclang check-templates -Xclang
-plugin-arg-find-bad-constructs -Xclang follow-macro-expansion -Xclang
-plugin-arg-find-bad-constructs -Xclang check-implicit-copy-ctors
-Wheader-hygiene -Wstring-conversion -fno-threadsafe-statics
-fvisibility-inlines-hidden -Wno-undefined-bool-conversion
-Wno-tautological-undefined-compare -std=c++11 -fno-rtti -fno-exceptions -c
../../source/row_neon.cc -o obj/libyuv_neon/row_neon.o

../../source/row_neon.cc:2747:6: error: instruction requires: half-float
conversions
    "vcvt.f16.f32  d2, q2                      \n"  // 8 half floats
     ^

So switch code for scaling by 1

Now on q820:
TestHalfFloatPlane_16bit_denormal (1270 ms)
TestHalfFloatPlane_10bit_Opt (1189 ms)
TestHalfFloatPlane_9bit_Opt (1156 ms)
TestHalfFloatPlane_Opt (1143 ms)
TestHalfFloatPlane_One (1094 ms)
TestHalfFloatPlane_Offby1 (1137 ms)

[fbarchard1, 2016-11-01 02:46:30]

disassembly to verify parameter passing overhead

Disassembly of section .text.HalfFloat1Row_NEON:

00000000 <HalfFloat1Row_NEON>:
   0:	ed2d 8b04 	vpush	{d8-d9}
   4:	f04f 62f0 	mov.w	r2, #125829120	; 0x7800000
   8:	eea8 2b10 	vdup.32	q4, r2
   c:	f920 2a0d 	vld1.8	{d2-d3}, [r0]!
  10:	3b08      	subs	r3, #8
  12:	ff88 4a12 	vmovl.u8	q2, d2
  16:	ff88 6a13 	vmovl.u8	q3, d3
  1a:	ffbb 46c4 	vcvt.f32.u32	q2, q2
  1e:	ffbb 66c6 	vcvt.f32.u32	q3, q3
  22:	ff04 4d58 	vmul.f32	q2, q2, q4
  26:	ff06 6d58 	vmul.f32	q3, q3, q4
  2a:	ff93 2914 	vqshrn.u32	d2, q2, #13
  2e:	ff93 3916 	vqshrn.u32	d3, q3, #13
  32:	f900 2a0d 	vst1.8	{d2-d3}, [r0]!
  36:	dce9      	bgt.n	c <HalfFloat1Row_NEON+0xc>
  38:	ecbd 8b04 	vpop	{d8-d9}
  3c:	4770      	bx	lr
  3e:	bf00      	nop

Disassembly of section .text.HalfFloatRow_NEON:

00000000 <HalfFloatRow_NEON>:
   0:	ee07 2a10 	vmov	s14, r2
   4:	eddf 7a11 	vldr	s15, [pc, #68]	; 4c <HalfFloatRow_NEON+0x4c>
   8:	ee67 7a27 	vmul.f32	s15, s14, s15
   c:	ed2d 8b04 	vpush	{d8-d9}
  10:	ee17 2a90 	vmov	r2, s15
  14:	eea8 2b10 	vdup.32	q4, r2
  18:	f920 2a0d 	vld1.8	{d2-d3}, [r0]!
  1c:	3b08      	subs	r3, #8
  1e:	ff88 4a12 	vmovl.u8	q2, d2
  22:	ff88 6a13 	vmovl.u8	q3, d3
  26:	ffbb 46c4 	vcvt.f32.u32	q2, q2
  2a:	ffbb 66c6 	vcvt.f32.u32	q3, q3
  2e:	ff04 4d58 	vmul.f32	q2, q2, q4
  32:	ff06 6d58 	vmul.f32	q3, q3, q4
  36:	ff93 2914 	vqshrn.u32	d2, q2, #13
  3a:	ff93 3916 	vqshrn.u32	d3, q3, #13
  3e:	f900 2a0d 	vst1.8	{d2-d3}, [r0]!
  42:	dce9      	bgt.n	18 <HalfFloatRow_NEON+0x18>
  44:	ecbd 8b04 	vpop	{d8-d9}
  48:	4770      	bx	lr
  4a:	bf00      	nop
  4c:	07800000 	.word	0x07800000

[fbarchard1, 2016-11-01 02:50:24]

avoiding q4 to removed vpush/vpop:

Disassembly of section .text.HalfFloat1Row_NEON:

00000000 <HalfFloat1Row_NEON>:
   0:	f04f 62f0 	mov.w	r2, #125829120	; 0x7800000
   4:	eea0 2b10 	vdup.32	q0, r2
   8:	f920 2a0d 	vld1.8	{d2-d3}, [r0]!
   c:	3b08      	subs	r3, #8
   e:	ff88 4a12 	vmovl.u8	q2, d2
  12:	ff88 6a13 	vmovl.u8	q3, d3
  16:	ffbb 46c4 	vcvt.f32.u32	q2, q2
  1a:	ffbb 66c6 	vcvt.f32.u32	q3, q3
  1e:	ff04 4d50 	vmul.f32	q2, q2, q0
  22:	ff06 6d50 	vmul.f32	q3, q3, q0
  26:	ff93 2914 	vqshrn.u32	d2, q2, #13
  2a:	ff93 3916 	vqshrn.u32	d3, q3, #13
  2e:	f900 2a0d 	vst1.8	{d2-d3}, [r0]!
  32:	dce9      	bgt.n	8 <HalfFloat1Row_NEON+0x8>
  34:	4770      	bx	lr
  36:	bf00      	nop

Disassembly of section .text.HalfFloatRow_NEON:

00000000 <HalfFloatRow_NEON>:
   0:	ee07 2a10 	vmov	s14, r2
   4:	eddf 7a0f 	vldr	s15, [pc, #60]	; 44 <HalfFloatRow_NEON+0x44>
   8:	ee67 7a27 	vmul.f32	s15, s14, s15
   c:	ee17 2a90 	vmov	r2, s15
  10:	eea0 2b10 	vdup.32	q0, r2
  14:	f920 2a0d 	vld1.8	{d2-d3}, [r0]!
  18:	3b08      	subs	r3, #8
  1a:	ff88 4a12 	vmovl.u8	q2, d2
  1e:	ff88 6a13 	vmovl.u8	q3, d3
  22:	ffbb 46c4 	vcvt.f32.u32	q2, q2
  26:	ffbb 66c6 	vcvt.f32.u32	q3, q3
  2a:	ff04 4d50 	vmul.f32	q2, q2, q0
  2e:	ff06 6d50 	vmul.f32	q3, q3, q0
  32:	ff93 2914 	vqshrn.u32	d2, q2, #13
  36:	ff93 3916 	vqshrn.u32	d3, q3, #13
  3a:	f900 2a0d 	vst1.8	{d2-d3}, [r0]!
  3e:	dce9      	bgt.n	14 <HalfFloatRow_NEON+0x14>
  40:	4770      	bx	lr
  42:	bf00      	nop
  44:	07800000 	.word	0x07800000

Now:
TestHalfFloatPlane_16bit_denormal (1183 ms)
TestHalfFloatPlane_10bit_Opt (1240 ms)
TestHalfFloatPlane_9bit_Opt (1247 ms)
TestHalfFloatPlane_Opt (1199 ms)
TestHalfFloatPlane_One (1139 ms)
TestHalfFloatPlane_Offby1 (1068 ms)

[fbarchard1, 2016-11-01 02:50:40]

Description was changed from

==========
HalfFloat Neon for ARMv7.

TEST=out/Release/libyuv_unittest --gtest_filter=*Half*

BUG=libyuv:560
R=wangcheng@google.com
==========

to

==========
HalfFloat Neon for ARMv7.

TEST=out/Release/libyuv_unittest --gtest_filter=*Half*

BUG=libyuv:560
R=wangcheng@google.com
==========

[fbarchard1, 2016-11-01 03:02:04]

Original C code was

TestHalfFloatPlane (5483 ms)

void HalfFloatRow_C(const uint16* src, uint16* dst, float scale, int width) {
  int i;
  float mult = 1.9259299444e-34f * scale;
  for (i = 0; i < width; ++i) {
    float value = src[i] * mult;
    dst[i] = (uint16)((*(uint32_t*)&value) >> 13);
  }
}

64 bit version
TestHalfFloatPlane_16bit_denormal (905 ms)
TestHalfFloatPlane_10bit_Opt (922 ms)
TestHalfFloatPlane_9bit_Opt (932 ms)
TestHalfFloatPlane_Opt (936 ms)
TestHalfFloatPlane_One (841 ms)
TestHalfFloatPlane_Offby1 (892 ms)


Disassembly of section .text.HalfFloat1Row_NEON:

0000000000000000 <HalfFloat1Row_NEON>:
   0:	4cdf7001 	ld1	{v1.16b}, [x0], #16
   4:	71002042 	subs	w2, w2, #0x8
   8:	2f10a422 	uxtl	v2.4s, v1.4h
   c:	6f10a423 	uxtl2	v3.4s, v1.8h
  10:	4e21d842 	scvtf	v2.4s, v2.4s
  14:	4e21d863 	scvtf	v3.4s, v3.4s
  18:	0e216841 	fcvtn	v1.4h, v2.4s
  1c:	4e216861 	fcvtn2	v1.8h, v3.4s
  20:	4c9f7021 	st1	{v1.16b}, [x1], #16
  24:	54fffeec 	b.gt	0 <HalfFloat1Row_NEON>
  28:	d65f03c0 	ret

Disassembly of section .text.HalfFloatRow_NEON:

0000000000000000 <HalfFloatRow_NEON>:
   0:	1c0001e1 	ldr	s1, 3c <HalfFloatRow_NEON+0x3c>
   4:	1e210800 	fmul	s0, s0, s1
   8:	4cdf7001 	ld1	{v1.16b}, [x0], #16
   c:	71002042 	subs	w2, w2, #0x8
  10:	2f10a422 	uxtl	v2.4s, v1.4h
  14:	6f10a423 	uxtl2	v3.4s, v1.8h
  18:	4e21d842 	scvtf	v2.4s, v2.4s
  1c:	4e21d863 	scvtf	v3.4s, v3.4s
  20:	4f809042 	fmul	v2.4s, v2.4s, v0.s[0]
  24:	4f809063 	fmul	v3.4s, v3.4s, v0.s[0]
  28:	2f139441 	uqshrn	v1.4h, v2.4s, #13
  2c:	6f139461 	uqshrn2	v1.8h, v3.4s, #13
  30:	4c9f7021 	st1	{v1.16b}, [x1], #16
  34:	54fffeac 	b.gt	8 <HalfFloatRow_NEON+0x8>
  38:	d65f03c0 	ret
  3c:	07800000 	.word	0x07800000

[fbarchard1, 2016-11-01 03:02:05]

Original C code was

TestHalfFloatPlane (5483 ms)

void HalfFloatRow_C(const uint16* src, uint16* dst, float scale, int width) {
  int i;
  float mult = 1.9259299444e-34f * scale;
  for (i = 0; i < width; ++i) {
    float value = src[i] * mult;
    dst[i] = (uint16)((*(uint32_t*)&value) >> 13);
  }
}

64 bit version
TestHalfFloatPlane_16bit_denormal (905 ms)
TestHalfFloatPlane_10bit_Opt (922 ms)
TestHalfFloatPlane_9bit_Opt (932 ms)
TestHalfFloatPlane_Opt (936 ms)
TestHalfFloatPlane_One (841 ms)
TestHalfFloatPlane_Offby1 (892 ms)


Disassembly of section .text.HalfFloat1Row_NEON:

0000000000000000 <HalfFloat1Row_NEON>:
   0:	4cdf7001 	ld1	{v1.16b}, [x0], #16
   4:	71002042 	subs	w2, w2, #0x8
   8:	2f10a422 	uxtl	v2.4s, v1.4h
   c:	6f10a423 	uxtl2	v3.4s, v1.8h
  10:	4e21d842 	scvtf	v2.4s, v2.4s
  14:	4e21d863 	scvtf	v3.4s, v3.4s
  18:	0e216841 	fcvtn	v1.4h, v2.4s
  1c:	4e216861 	fcvtn2	v1.8h, v3.4s
  20:	4c9f7021 	st1	{v1.16b}, [x1], #16
  24:	54fffeec 	b.gt	0 <HalfFloat1Row_NEON>
  28:	d65f03c0 	ret

Disassembly of section .text.HalfFloatRow_NEON:

0000000000000000 <HalfFloatRow_NEON>:
   0:	1c0001e1 	ldr	s1, 3c <HalfFloatRow_NEON+0x3c>
   4:	1e210800 	fmul	s0, s0, s1
   8:	4cdf7001 	ld1	{v1.16b}, [x0], #16
   c:	71002042 	subs	w2, w2, #0x8
  10:	2f10a422 	uxtl	v2.4s, v1.4h
  14:	6f10a423 	uxtl2	v3.4s, v1.8h
  18:	4e21d842 	scvtf	v2.4s, v2.4s
  1c:	4e21d863 	scvtf	v3.4s, v3.4s
  20:	4f809042 	fmul	v2.4s, v2.4s, v0.s[0]
  24:	4f809063 	fmul	v3.4s, v3.4s, v0.s[0]
  28:	2f139441 	uqshrn	v1.4h, v2.4s, #13
  2c:	6f139461 	uqshrn2	v1.8h, v3.4s, #13
  30:	4c9f7021 	st1	{v1.16b}, [x1], #16
  34:	54fffeac 	b.gt	8 <HalfFloatRow_NEON+0x8>
  38:	d65f03c0 	ret
  3c:	07800000 	.word	0x07800000

[fbarchard1, 2016-11-01 17:31:58]

Description was changed from

==========
HalfFloat Neon for ARMv7.

TEST=out/Release/libyuv_unittest --gtest_filter=*Half*

BUG=libyuv:560
R=wangcheng@google.com
==========

to

==========
HalfFloat Neon for ARMv7.

64 bit version made similar to 32 bit with registers 1 for load and store
results, and 2 and 3 as expanded float temporary values.

TEST=out/Release/libyuv_unittest --gtest_filter=*Half*

BUG=libyuv:560
R=wangcheng@google.com
==========

[wangcheng, 2016-11-01 17:59:15]

lgtm

[fbarchard1, 2016-11-01 18:36:53]

Description was changed from

==========
HalfFloat Neon for ARMv7.

64 bit version made similar to 32 bit with registers 1 for load and store
results, and 2 and 3 as expanded float temporary values.

TEST=out/Release/libyuv_unittest --gtest_filter=*Half*

BUG=libyuv:560
R=wangcheng@google.com
==========

to

==========
HalfFloat Neon for ARMv7.

64 bit version made similar to 32 bit with registers 1 for load and store
results, and 2 and 3 as expanded float temporary values.

TEST=out/Release/libyuv_unittest --gtest_filter=*Half*

BUG=libyuv:560
R=wangcheng@google.com

Committed:
https://chromium.googlesource.com/libyuv/libyuv/+/eca08525cb0bda9b21da786240c...
==========

[fbarchard1, 2016-11-01 18:36:54]

Committed patchset #5 (id:80001) manually as
eca08525cb0bda9b21da786240cf841340fbe575 (presubmit successful).