[PATCH 7/8] drm/format-helper: Optimize 32-to-16-bpp conversion

[Thomas Zimmermann]

Hi

Am 26.03.25 um 11:53 schrieb Jani Nikula:
> On Tue, 25 Mar 2025, Thomas Zimmermann <tzimmermann at suse.de> wrote:
>> For ease of implementation, existing line-conversion functions
>> for 16-bit formats write each pixel individually. Optimize the
>> performance by writing mulitple pixels in single 64-bit and 32-bit
>> stores.
>>
>> Signed-off-by: Thomas Zimmermann <tzimmermann at suse.de>
>> ---
>>   drivers/gpu/drm/drm_format_helper.c | 40 ++++++++++++++++++++++++++++-
>>   1 file changed, 39 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/gpu/drm/drm_format_helper.c b/drivers/gpu/drm/drm_format_helper.c
>> index b9c9c712aa9c..66137df85725 100644
>> --- a/drivers/gpu/drm/drm_format_helper.c
>> +++ b/drivers/gpu/drm/drm_format_helper.c
>> @@ -262,10 +262,48 @@ static __always_inline void drm_fb_xfrm_line_32to16(void *dbuf, const void *sbuf
>>   						    unsigned int pixels,
>>   						    u32 (*xfrm_pixel)(u32))
>>   {
>> -	__le16 *dbuf16 = dbuf;
>> +	__le64 *dbuf64 = dbuf;
>> +	__le32 *dbuf32;
>> +	__le16 *dbuf16;
>>   	const __le32 *sbuf32 = sbuf;
>>   	const __le32 *send32 = sbuf32 + pixels;
>>   
>> +#if defined(CONFIG_64BIT)
>> +	/* write 4 pixels at once */
>> +	send32 -= pixels & GENMASK(1, 0);
>> +	while (sbuf32 < send32) {
> I find the adjusting of send32 before and after the loop with different
> masks a bit confusing. Would it not suffice to:
>
> 	while (sbuf32 < ALIGN_DOWN(send32, 4))
>
> and leave send32 untouched? With different alignments for 2 pixels at a
> time.

Makes sense.

Best regards
Thomas

>
>
> BR,
> Jani.
>
>
>> +		u32 pix[4] = {
>> +			le32_to_cpup(sbuf32++),
>> +			le32_to_cpup(sbuf32++),
>> +			le32_to_cpup(sbuf32++),
>> +			le32_to_cpup(sbuf32++),
>> +		};
>> +		/* write output bytes in reverse order for little endianness */
>> +		u64 val64 = ((u64)xfrm_pixel(pix[0])) |
>> +			    ((u64)xfrm_pixel(pix[1]) << 16) |
>> +			    ((u64)xfrm_pixel(pix[2]) << 32) |
>> +			    ((u64)xfrm_pixel(pix[3]) << 48);
>> +		*dbuf64++ = cpu_to_le64(val64);
>> +	}
>> +	send32 += pixels & GENMASK(1, 1);
>> +#endif
>> +
>> +	/* write 2 pixels at once */
>> +	dbuf32 = (__le32 __force *)dbuf64;
>> +	while (sbuf32 < send32) {
>> +		u32 pix[2] = {
>> +			le32_to_cpup(sbuf32++),
>> +			le32_to_cpup(sbuf32++),
>> +		};
>> +		/* write output bytes in reverse order for little endianness */
>> +		u32 val32 = xfrm_pixel(pix[0]) |
>> +			   (xfrm_pixel(pix[1]) << 16);
>> +		*dbuf32++ = cpu_to_le32(val32);
>> +	}
>> +	send32 += pixels & GENMASK(0, 0);
>> +
>> +	/* write trailing pixel */
>> +	dbuf16 = (__le16 __force *)dbuf32;
>>   	while (sbuf32 < send32)
>>   		*dbuf16++ = cpu_to_le16(xfrm_pixel(le32_to_cpup(sbuf32++)));
>>   }

-- 
--
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Frankenstrasse 146, 90461 Nuernberg, Germany
GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman
HRB 36809 (AG Nuernberg)