[PATCH i-g-t 1/2] lib/gpgpu_shader: Add x_offset parameter for gpgpu_shader__write_on_exception

Hajda, Andrzej andrzej.hajda at intel.com
Mon Oct 14 13:49:27 UTC 2024


Somehow I missed this.

W dniu 01.10.2024 o 17:07, Zbigniew Kempczyński pisze:
> On Wed, Sep 25, 2024 at 03:30:02PM +0200, Christoph Manszewski wrote:
>> Currently gpgpu_shader__write_on_exception always writes the first
>> column equal to thread group ID x. Make it possible to specify an offset
>> from that value.
>>
>> Signed-off-by: Christoph Manszewski <christoph.manszewski at intel.com>
>> Cc: Zbigniew Kempczyński <zbigniew.kempczynski at intel.com>
>> ---
>>   lib/gpgpu_shader.c          | 29 +++++++-------
>>   lib/gpgpu_shader.h          |  4 +-
>>   lib/iga64_generated_codes.c | 77 ++++++++++++++++++++-----------------
>>   tests/intel/xe_exec_sip.c   |  6 +--
>>   4 files changed, 62 insertions(+), 54 deletions(-)
>>
>> diff --git a/lib/gpgpu_shader.c b/lib/gpgpu_shader.c
>> index 6d8c7ebb8..8df8d9612 100644
>> --- a/lib/gpgpu_shader.c
>> +++ b/lib/gpgpu_shader.c
>> @@ -670,6 +670,7 @@ void gpgpu_shader__set_exception(struct gpgpu_shader *shdr, uint32_t value)
>>    * gpgpu_shader__write_on_exception:
>>    * @shdr: shader to be modified
>>    * @value: dword to be written
>> + * @x_offset: write target offset within the surface in columns added to the 'thread group id x'
>>    * @y_offset: write target offset within the surface in rows
>>    * @mask: mask to be applied on exception register
>>    * @expected: expected value of exception register with @mask applied
>> @@ -678,45 +679,47 @@ void gpgpu_shader__set_exception(struct gpgpu_shader *shdr, uint32_t value)
>>    * to provided ones: cr0.1 & @mask == @expected,
>>    * if yes fill dword in (row, column/dword) == (tg_id_y + @y_offset, tg_id_x).
> tg_id_x + @x_offset
>
>>    */
>> -void gpgpu_shader__write_on_exception(struct gpgpu_shader *shdr, uint32_t value,
>> +void gpgpu_shader__write_on_exception(struct gpgpu_shader *shdr, uint32_t value, uint32_t x_offset,
>>   				      uint32_t y_offset, uint32_t mask, uint32_t expected)
>>   {
>>   	emit_iga64_code(shdr, write_on_exception, "					\n\
>>   	// Clear message header								\n\
>>   (W)	mov (16|M0)              r4.0<1>:ud    0x0:ud					\n\
>>   	// Payload									\n\
>> -(W)	mov (1|M0)               r5.0<1>:ud    ARG(3):ud				\n\
>> +(W)	mov (1|M0)               r5.0<1>:ud    ARG(4):ud				\n\
>>   #if GEN_VER < 2000 // prepare Media Block Write						\n\
>>   	// X offset of the block in bytes := (thread group id X << ARG(0))		\n\
>> -(W)	shl (1|M0)               r4.0<1>:ud    r0.1<0;1,0>:ud    ARG(0):ud		\n\
>> +(W)	add (1|M0)               r4.0<1>:ud    r0.1<0;1,0>:ud	 ARG(1):ud		\n\
>> +(W)	shl (1|M0)               r4.0<1>:ud    r4.0<0;1,0>:ud    ARG(0):ud		\n\
>>   	// Y offset of the block in rows := thread group id Y				\n\
>>   (W)	mov (1|M0)               r4.1<1>:ud    r0.6<0;1,0>:ud				\n\
>> -(W)	add (1|M0)               r4.1<1>:ud    r4.1<0;1,0>:ud   ARG(1):ud		\n\
>> +(W)	add (1|M0)               r4.1<1>:ud    r4.1<0;1,0>:ud   ARG(2):ud		\n\
> You may drop mov and below should be enough:
>
> 	add (1|M0)               r4.1<1>:ud    r0.6<0;1,0>:ud    ARG(2):ud
>
>>   	// block width [0,63] representing 1 to 64 bytes				\n\
>> -(W)	mov (1|M0)               r4.2<1>:ud    ARG(2):ud				\n\
>> +(W)	mov (1|M0)               r4.2<1>:ud    ARG(3):ud				\n\
>>   	// FFTID := FFTID from R0 header						\n\
>>   (W)	mov (1|M0)               r4.4<1>:ud    r0.5<0;1,0>:ud				\n\
>>   #else // prepare Typed 2D Block Store							\n\
>> -	// Load r2.0-3 with tg id X << ARG(0)						\n\
>> -(W)	shl (1|M0)               r2.0<1>:ud    r0.1<0;1,0>:ud    ARG(0):ud		\n\
>> -	// Load r2.4-7 with tg id Y + ARG(1):ud						\n\
>> +	// Load r2.0 with tg id (X + ARG(1)) << ARG(0)					\n\
>> +(W)	add (1|M0)               r2.0<1>:ud    r0.1<0;1,0>:ud	 ARG(1):ud		\n\
>> +(W)	shl (1|M0)               r2.0<1>:ud    r2.0<0;1,0>:ud    ARG(0):ud		\n\
>> +	// Load r2.4-7 with tg id Y + ARG(2):ud						\n\
>>   (W)	mov (1|M0)               r2.1<1>:ud    r0.6<0;1,0>:ud				\n\
>> -(W)	add (1|M0)               r2.1<1>:ud    r2.1<0;1,0>:ud    ARG(1):ud		\n\
>> +(W)	add (1|M0)               r2.1<1>:ud    r2.1<0;1,0>:ud    ARG(2):ud		\n\
> Same here.
>
>>   	// Store X and Y block start (160:191 and 192:223)				\n\
>>   (W)	mov (2|M0)               r4.5<1>:ud    r2.0<2;2,1>:ud				\n\
>>   	// Store X and Y block max_size (224:231 and 232:239)				\n\
>> -(W)	mov (1|M0)               r4.7<1>:ud    ARG(2):ud				\n\
>> +(W)	mov (1|M0)               r4.7<1>:ud    ARG(3):ud				\n\
>>   #endif											\n\
>>   	// Check if masked exception is equal to provided value and write conditionally \n\
>> -(W)      and (1|M0)              r3.0<1>:ud     cr0.1<0;1,0>:ud ARG(4):ud		\n\
>> +(W)      and (1|M0)              r3.0<1>:ud     cr0.1<0;1,0>:ud ARG(5):ud		\n\
>>   (W)      mov (1|M0)              f0.0<1>:ud     0x0:ud					\n\
>> -(W)      cmp (1|M0)     (eq)f0.0 null:ud        r3.0<0;1,0>:ud  ARG(5):ud		\n\
>> +(W)      cmp (1|M0)     (eq)f0.0 null:ud        r3.0<0;1,0>:ud  ARG(6):ud		\n\
> Adding argument in the middle of the list makes a lot of noise in ARG(x)
> changing. What a pity we don't have possibility of making definitions
> like:
>
> #define x_offset ARG(1)
> #define y_offset ARG(2)
>
> in the code. Changing definitions would keep code clear.

Why not:

        emit_iga64_code(shdr, end_system_routine_step_if_eq, "   \n\

#define x_offset ARG(1)						\n\
#define y_offset ARG(2)						\n\
...
	mov (1|M0)               r30.6<1>:ud    y_offset:ud     \n\
...

Regards
Andrzej



More information about the igt-dev mailing list