[Ffmpeg-devel] [patch] move av_xiphlacing in avutil

[Jindrich Makovicka]

Michael Niedermayer wrote:
> Hi
> 
> On Wed, Nov 23, 2005 at 07:52:29AM +0100, Jindrich Makovicka wrote:
> 
>>Michael Niedermayer wrote:
>>
>>>hmm, IMHO if the code is small and generic it should be ok (a single crc
>>>calculation function for all 8,16 and 32bit, and no hardcoded tables, but
>>>tables generated with some init_crc() function)
>>>unless there are objections of course ...
>>
>>This should allow arbitrary 8-32 big & little endian crc. The code is
>>currently quite a lot slower than a hardcoded table, about 30% slower
>>for 16bit crc, slightly better for 8 and 32 bit.
> 
> 
> could you post the generated asm code (gcc -S -O3 -march...) of the slower
> code and of the current fast code

CFLAGS="-S -O3 -march=athlon-4 -mtune=athlon-4"

uint16_t crc16_block(uint8_t *data,uint32_t num_bytes);
crc16_block:
	pushl	%ebp
	xorl	%eax, %eax
	movl	%esp, %ebp
	pushl	%edi
	movl	8(%ebp), %edi
	pushl	%esi
	movl	12(%ebp), %esi
	pushl	%ebx
	testl	%esi, %esi
	je	.L13
	xorl	%ebx, %ebx
	xorl	%ecx, %ecx
	.p2align 4,,7
.L14:
	movzbl	(%edi,%ebx), %eax
	movzbl	%ch, %edx
	incl	%ebx
	sall	$8, %ecx
	xorl	%edx, %eax
	xorw	crc_lut(%eax,%eax), %cx
	cmpl	%ebx, %esi
	jne	.L14
	movzwl	%cx, %eax
.L13:
	popl	%ebx
	popl	%esi
	popl	%edi
	leave
	ret

uint32_t av_crc(AVCRC *ctx, uint32_t start_crc, uint8_t *buffer, size_t 
length) - .L61 is relevant for crc16
av_crc:
	pushl	%ebp
	movl	%esp, %ebp
	pushl	%edi
	pushl	%esi
	pushl	%ebx
	subl	$12, %esp
	movl	8(%ebp), %esi
	movl	16(%ebp), %ebx
	movl	20(%ebp), %edi
	movl	8(%esi), %eax
	movl	%eax, %edx
	movl	%eax, -20(%ebp)
	movl	4(%esi), %eax
	andl	12(%ebp), %edx
	cmpl	$8, %eax
	jg	.L60
	xorl	%ecx, %ecx
	testl	%edi, %edi
	je	.L43
	.p2align 4,,7
.L42:
	movzbl	(%ebx), %eax
	incl	%ecx
	incl	%ebx
	xorl	%edx, %eax
	cmpl	%ecx, %edi
	movzbl	%al, %eax
	movl	12(%esi,%eax,4), %edx
	jne	.L42
	.p2align 4,,7
.L43:
	andl	-20(%ebp), %edx
	addl	$12, %esp
	popl	%ebx
	popl	%esi
	popl	%edi
	leave
	movl	%edx, %eax
	ret
	.p2align 4,,7
.L60:
	movl	(%esi), %ecx
	testl	%ecx, %ecx
	je	.L61
	xorl	%ecx, %ecx
	testl	%edi, %edi
	je	.L43
	.p2align 4,,7
.L47:
	movzbl	(%ebx), %eax
	incl	%ecx
	incl	%ebx
	xorl	%edx, %eax
	shrl	$8, %edx
	movzbl	%al, %eax
	xorl	12(%esi,%eax,4), %edx
	cmpl	%ecx, %edi
	je	.L43
	movzbl	(%ebx), %eax
	incl	%ecx
	incl	%ebx
	xorl	%edx, %eax
	shrl	$8, %edx
	movzbl	%al, %eax
	xorl	12(%esi,%eax,4), %edx
	cmpl	%ecx, %edi
	jne	.L47
	jmp	.L43
	.p2align 4,,7
.L61:
	testl	%edi, %edi
	je	.L43
	subl	$8, %eax
	movl	$0, -24(%ebp)
	movl	%eax, -16(%ebp)
	.p2align 4,,7
.L48:
	movzbl	-16(%ebp), %ecx
	movl	%edx, %eax
	sall	$8, %edx
	shrl	%cl, %eax
	xorb	(%ebx), %al
	incl	-24(%ebp)
	incl	%ebx
	movzbl	%al, %eax
	xorl	12(%esi,%eax,4), %edx
	cmpl	-24(%ebp), %edi
	jne	.L48
	andl	-20(%ebp), %edx
	addl	$12, %esp
	popl	%ebx
	popl	%esi
	popl	%edi
	leave
	movl	%edx, %eax
	ret

> 
> hmm, cant the table and start_crc be modified so that the ctx->le case is always useable?
> 

I am not sure, I'll try to look at other implementations.

-- 
Jindrich Makovicka