SoC 2009/ARM NEON Optimization - Revision history

Conrad: /* Goals */

2009-07-01T19:27:50Z

Goals

← Older revision		Revision as of 19:27, 1 July 2009
Line 18:		Line 18:
	\|5 \|\| Chroma MC \|\|style="background: #ffff00"\| In progress \|\| Ported ffmpeg's and the general case works, but not the special cases		\|5 \|\| Chroma MC \|\|style="background: #ffff00"\| In progress \|\| Ported ffmpeg's and the general case works, but not the special cases
	\|-		\|-
	\|6 \|\| (I)DCT \|\|style="background: #~~ffff00~~"\| ~~In Progress~~ \|\| 4x4 (I)DCT and 8x8 IDCT done, 8x8 DCT not started		\|6 \|\| (I)DCT \|\|style="background: #00ff00"\| Done \|\| 4x4 (I)DCT and 8x8 IDCT done, 8x8 DCT not started
	\|-		\|-
	\|7 \|\| (De)quant \|\|style="background: #ffff00"\| In Progress \|\| 4x4 quant done, rest not started		\|7 \|\| (De)quant \|\|style="background: #ffff00"\| In Progress \|\| 4x4 quant done, rest not started
Line 34:		Line 34:
	\|13 \|\| pixel variance \|\| style="background: #00ff00"\| Done \|\|		\|13 \|\| pixel variance \|\| style="background: #00ff00"\| Done \|\|
	\|-		\|-
	\|14 \|\| lowres_init \|\| ~~Not Started~~ \|\|		\|14 \|\| lowres_init \|\|style="background: #00ff00"\| Done \|\|
	\|-		\|-
	\|15 \|\| hpel_filter \|\| Not Started \|\|		\|15 \|\| hpel_filter \|\| Not Started \|\|

Conrad: /* Goals */

2009-06-11T05:47:04Z

Goals

← Older revision		Revision as of 05:47, 11 June 2009
Line 30:		Line 30:
	\|11 \|\| Intra Prediction \|\| style="background: #ffff00"\| In Progress \|\| Simpler functions done		\|11 \|\| Intra Prediction \|\| style="background: #ffff00"\| In Progress \|\| Simpler functions done
	\|-		\|-
	\|12 \|\| sa8d \|\| style~~="backgroundstyle~~="background: #00ff00"\| Done \|\|		\|12 \|\| sa8d \|\| style="background: #00ff00"\| Done \|\|
	\|-		\|-
	\|13 \|\| pixel variance \|\| style="background: #00ff00"\| Done \|\|		\|13 \|\| pixel variance \|\| style="background: #00ff00"\| Done \|\|

Conrad at 05:45, 11 June 2009

2009-06-11T05:45:52Z

← Older revision		Revision as of 05:45, 11 June 2009
Line 30:		Line 30:
	\|11 \|\| Intra Prediction \|\| style="background: #ffff00"\| In Progress \|\| Simpler functions done		\|11 \|\| Intra Prediction \|\| style="background: #ffff00"\| In Progress \|\| Simpler functions done
	\|-		\|-
	\|12 \|\| sa8d \|\| style="background: #~~ffff00~~"\| ~~In Progress~~ \|\| ~~Should work but doesn't~~		\|12 \|\| sa8d \|\| style="backgroundstyle="background: #00ff00"\| Done \|\|
	\|-		\|-
	\|13 \|\| pixel variance \|\| style="background: #00ff00"\| Done \|\|		\|13 \|\| pixel variance \|\| style="background: #00ff00"\| Done \|\|
Line 73:		Line 73:
	''' May 23''' Program Starts		''' May 23''' Program Starts

	''' May 23 - July 13''' Code		''' May 23 - June 14''' Code

			''' June 15 - June 21''' Vacation on Chincoteague

			''' June 22 - July 12''' Code

	''' July 13''' Mid Term deadline. Specific goals pending, but I'd like to have none of the functions without asm equivalents to be more than about 2% of common profiles. In other words, all the major functions would have been written in NEON ASM		''' July 13''' Mid Term deadline. Specific goals pending, but I'd like to have none of the functions without asm equivalents to be more than about 2% of common profiles. In other words, all the major functions would have been written in NEON ASM

Conrad: /* Goals */

2009-06-09T00:22:39Z

Goals

← Older revision		Revision as of 00:22, 9 June 2009
Line 26:		Line 26:
	\|9 \|\| Zigzag \|\|style="background: #ffff00"\| In Progress \|\| 4x4 scan/sub frame done, rest in progress		\|9 \|\| Zigzag \|\|style="background: #ffff00"\| In Progress \|\| 4x4 scan/sub frame done, rest in progress
	\|-		\|-
	\|10 \|\| Hadamard AC \|\| ~~Not Started~~ \|\|		\|10 \|\| Hadamard AC \|\| style="background: #00ff00"\| Done \|\|
	\|-		\|-
	\|11 \|\| Intra Prediction \|\| ~~Not Started~~ \|\|		\|11 \|\| Intra Prediction \|\| style="background: #ffff00"\| In Progress \|\| Simpler functions done
	\|-		\|-
	\|12 \|\| sa8d \|\| ~~Not Started~~ \|\|		\|12 \|\| sa8d \|\| style="background: #ffff00"\| In Progress \|\| Should work but doesn't
	\|-		\|-
	\|13 \|\| pixel variance \|\| ~~Not Started~~ \|\|		\|13 \|\| pixel variance \|\| style="background: #00ff00"\| Done \|\|
	\|-		\|-
	\|14 \|\| lowres_init \|\| Not Started \|\|		\|14 \|\| lowres_init \|\| Not Started \|\|
	\|-		\|-
	\|15 \|\| ~~Loop Filter~~ \|\| Not Started \|\|		\|15 \|\| hpel_filter \|\| Not Started \|\|
	\|-		\|-
	\|16 \|\| ~~intra sad/satd/sa8d~~ \|\| Not Started \|\|		\|16 \|\| Loop Filter \|\| Not Started \|\|
	\|-		\|-
	\|17 \|\| ~~Integral init~~ \|\| Not Started \|\|		\|17 \|\| intra sad/satd/sa8d \|\| Not Started \|\|
	\|-		\|-
	\|18 \|\| ~~Denoise~~ \|\| Not Started \|\|		\|18 \|\| Integral init \|\| Not Started \|\|
	\|-		\|-
	\|19 \|\| ~~Decimate score~~ \|\| Not Started \|\|		\|19 \|\| Denoise \|\| Not Started \|\|
	\|-		\|-
	\|20 \|\| ~~Coeff last/level_run~~ \|\| Not Started \|\|		\|20 \|\| Decimate score \|\| Not Started \|\|
	\|-		\|-
	\|21 \|\| ~~cabac~~ \|\| Not Started \|\|		\|21 \|\| Coeff last/level_run \|\| Not Started \|\|
	\|-		\|-
	\|22 \|\| ~~ARMv6 optimizations~~ \|\|~~style="background: #ffff00"~~\| ~~In progress~~ \|~~\| For smaller functions, write code using ARMv6's simd extensions. For instance, ARMv6 SAD width 4 is faster on Cortex-A8.~~		\|22 \|\| cabac \|\| Not Started \|\|
	\|-		\|-
	\|23 \|\| ~~Cortex-A9~~ optimizations \|\|style="background: #ffff00"\| In progress \|\| ~~Primarily consists of running checkasm --bench on an A9~~, ~~noting whether~~ ARMv6 ~~or NEON~~ is faster~~, figuring out why, and ensuring that the faster function is selected~~.		\|23 \|\| ARMv6 optimizations \|\|style="background: #ffff00"\| In progress \|\| For smaller functions, write code using ARMv6's simd extensions. For instance, ARMv6 SAD width 4 is faster on Cortex-A8.
	\|-		\|-
	\|24 \|\| ~~gcc optimization~~ \|\| ~~Not Started~~ \|\| ~~Investigate what gcc produces for important C functions~~ and ~~attempt to optimize it for ARM~~		\|24 \|\| Cortex-A9 optimizations \|\|style="background: #ffff00"\| In progress \|\| Primarily consists of running checkasm --bench on an A9, noting whether ARMv6 or NEON is faster, figuring out why, and ensuring that the faster function is selected.
	\|-		\|-
	\|25 \|\| Performance Counter Optimization \|\| Not Started \|\| Utilize the ARMv7 performance counters to optimize stalls, cache usage, NEON->ARM data hazards, and more		\|25 \|\| gcc optimization \|\| Not Started \|\| Investigate what gcc produces for important C functions and attempt to optimize it for ARM
			\|-
			\|26 \|\| Performance Counter Optimization \|\| Not Started \|\| Utilize the ARMv7 performance counters to optimize stalls, cache usage, NEON->ARM data hazards, and more
	\|}		\|}

Conrad at 00:24, 1 June 2009

2009-06-01T00:24:10Z

← Older revision		Revision as of 00:24, 1 June 2009
Line 18:		Line 18:
	\|5 \|\| Chroma MC \|\|style="background: #ffff00"\| In progress \|\| Ported ffmpeg's and the general case works, but not the special cases		\|5 \|\| Chroma MC \|\|style="background: #ffff00"\| In progress \|\| Ported ffmpeg's and the general case works, but not the special cases
	\|-		\|-
	\|6 \|\| (I)DCT \|\|style="background: #ffff00"\| In Progress \|\| 4x4 (I)DCT done, 8x8 not started		\|6 \|\| (I)DCT \|\|style="background: #ffff00"\| In Progress \|\| 4x4 (I)DCT and 8x8 IDCT done, 8x8 DCT not started
	\|-		\|-
	\|7 \|\| (De)quant \|\|style="background: #ffff00"\| In Progress \|\| 4x4 quant done, rest not started		\|7 \|\| (De)quant \|\|style="background: #ffff00"\| In Progress \|\| 4x4 quant done, rest not started
	\|-		\|-
	\|8 \|\| SSIM \|\|style="background: #~~ffff00~~"\| ~~In Progress~~ \|\| ~~core done, end in progress~~		\|8 \|\| SSIM \|\|style="background: #00ff00"\| Done \|\|
	\|-		\|-
	\|9 \|\| Zigzag \|\|style="background: #ffff00"\| In Progress \|\| 4x4 scan/sub frame done, rest in progress		\|9 \|\| Zigzag \|\|style="background: #ffff00"\| In Progress \|\| 4x4 scan/sub frame done, rest in progress

Conrad: /* Timeline */

2009-05-07T23:29:26Z

Timeline

← Older revision		Revision as of 23:29, 7 May 2009
Line 71:		Line 71:
	''' May 23''' Program Starts		''' May 23''' Program Starts

	''' May 23 ~~- July 13 ''' Work on writing NEON asm for~~		''' May 23 - July 13''' Code

	~~''' June 13~~ - July 13''' Code

	''' July 13''' Mid Term deadline. Specific goals pending, but I'd like to have none of the functions without asm equivalents to be more than about 2% of common profiles. In other words, all the major functions would have been written in NEON ASM		''' July 13''' Mid Term deadline. Specific goals pending, but I'd like to have none of the functions without asm equivalents to be more than about 2% of common profiles. In other words, all the major functions would have been written in NEON ASM

Conrad at 22:37, 7 May 2009

2009-05-07T22:37:47Z

← Older revision		Revision as of 22:37, 7 May 2009
Line 58:		Line 58:
	\|25 \|\| Performance Counter Optimization \|\| Not Started \|\| Utilize the ARMv7 performance counters to optimize stalls, cache usage, NEON->ARM data hazards, and more		\|25 \|\| Performance Counter Optimization \|\| Not Started \|\| Utilize the ARMv7 performance counters to optimize stalls, cache usage, NEON->ARM data hazards, and more
	\|}		\|}

			== Repository ==
			I'm currently using gitorious to publish my work [http://gitorious.org/projects/x264-arm]

	== Timeline ==		== Timeline ==

Conrad: New page: {{SoCProject|year=2009|student=David Conrad|mentor=Holger Lubitz}} == Abstract == Formerly used primarily in PDAs and cellphones, ARM is starting to target netbooks and laptops with its n...

2009-05-07T22:34:38Z

New page: {{SoCProject|year=2009|student=David Conrad|mentor=Holger Lubitz}} == Abstract == Formerly used primarily in PDAs and cellphones, ARM is starting to target netbooks and laptops with its n...

New page

{{SoCProject|year=2009|student=David Conrad|mentor=Holger Lubitz}}

== Abstract ==
Formerly used primarily in PDAs and cellphones, ARM is starting to target netbooks and laptops with its new Cortex-A series of processors. One key feature of these new processors is the NEON SIMD unit, the use of which will massively boost the performance of many multimedia applications. This project will consist of writing NEON/ARMv6 SIMD assembly for all of the major DSP functions in x264, ideally speeding x264 up by over 4-5 times.

== Goals ==
{| class="wikitable"
! No. !! Task Description !! Status !! Comment
|-
|1 || ARM framework || style="background: #00ff00"| Mostly Done || Waiting on more info on CPUID emulation
|-
|2 || SAD/SATD||style="background: #00ff00"| Done || Qualifying task
|-
|3 || SSD ||style="background: #00ff00"| Done ||
|-
|4 || Luma MC||style="background: #ffff00"| In progress || get_ref_neon complete
|-
|5 || Chroma MC ||style="background: #ffff00"| In progress || Ported ffmpeg's and the general case works, but not the special cases
|-
|6 || (I)DCT ||style="background: #ffff00"| In Progress || 4x4 (I)DCT done, 8x8 not started
|-
|7 || (De)quant ||style="background: #ffff00"| In Progress || 4x4 quant done, rest not started
|-
|8 || SSIM ||style="background: #ffff00"| In Progress || core done, end in progress
|-
|9 || Zigzag ||style="background: #ffff00"| In Progress || 4x4 scan/sub frame done, rest in progress
|-
|10 || Hadamard AC || Not Started ||
|-
|11 || Intra Prediction || Not Started ||
|-
|12 || sa8d || Not Started ||
|-
|13 || pixel variance || Not Started ||
|-
|14 || lowres_init || Not Started ||
|-
|15 || Loop Filter || Not Started ||
|-
|16 || intra sad/satd/sa8d || Not Started ||
|-
|17 || Integral init || Not Started ||
|-
|18 || Denoise || Not Started ||
|-
|19 || Decimate score || Not Started ||
|-
|20 || Coeff last/level_run || Not Started ||
|-
|21 || cabac || Not Started ||
|-
|22 || ARMv6 optimizations ||style="background: #ffff00"| In progress || For smaller functions, write code using ARMv6's simd extensions. For instance, ARMv6 SAD width 4 is faster on Cortex-A8.
|-
|23 || Cortex-A9 optimizations ||style="background: #ffff00"| In progress || Primarily consists of running checkasm --bench on an A9, noting whether ARMv6 or NEON is faster, figuring out why, and ensuring that the faster function is selected.
|-
|24 || gcc optimization || Not Started || Investigate what gcc produces for important C functions and attempt to optimize it for ARM
|-
|25 || Performance Counter Optimization || Not Started || Utilize the ARMv7 performance counters to optimize stalls, cache usage, NEON->ARM data hazards, and more
|}

== Timeline ==
This is an estimation.

'''March-April''' Qualifying task, work on the more important (>5% in profiles) functions

'''May''' Final projects and exams

''' May 23''' Program Starts

''' May 23 - July 13 ''' Work on writing NEON asm for

''' June 13 - July 13''' Code

''' July 13''' Mid Term deadline. Specific goals pending, but I'd like to have none of the functions without asm equivalents to be more than about 2% of common profiles. In other words, all the major functions would have been written in NEON ASM

''' July 13 - August 17''' Code

'''August 17''' Program end