[PATCH 2/3] lib/gpgpu_shader: add function for computing max number of threads per tg

[Hajda, Andrzej]

W dniu 01.07.2025 o 15:53, Jan Maslak pisze:
> Adds function compute_max_threads_per_tg() that computes the maximum
> allowed amount of threads in a thread group for XE2-3 platforms.
>
> Signed-off-by: Jan Maslak <jan.maslak at intel.com>
> ---
>   lib/gpgpu_shader.c | 64 ++++++++++++++++++++++++++++++++++++++++++++++
>   lib/gpgpu_shader.h |  9 +++++++
>   2 files changed, 73 insertions(+)
>
> diff --git a/lib/gpgpu_shader.c b/lib/gpgpu_shader.c
> index b83e645e3..7e1d73d72 100644
> --- a/lib/gpgpu_shader.c
> +++ b/lib/gpgpu_shader.c
> @@ -919,3 +919,67 @@ void gpgpu_shader__read_a64_d32(struct gpgpu_shader *shdr, uint64_t ppgtt_addr)
>   #endif										\n\
>   	", lower_32_bits(addr), upper_32_bits(addr));
>   }
> +
> +static uint32_t compute_max_threads_in_tg_no_vrt(bool large_grf_mode,
> +						 uint32_t simd_size)
> +{
> +	if (large_grf_mode) {
> +		return 32;
> +	} else {
> +		if (simd_size == 16)
> +			return 64;
> +		else if (simd_size == 32)
> +			return 32;
> +	}
> +
> +	igt_warn("Unsupported SIMD size %d\n", simd_size);
> +	return 1;
> +}
> +
> +static uint32_t compute_max_threads_in_tg_vrt(enum gpgpu_shader_vrt_modes vrt,
> +					      uint32_t simd_size,
> +					      bool hw_local_id_generation)
> +{
> +	if (vrt <= VRT_128) {
> +		if (simd_size == 16)
> +			return 64;
> +		else if (simd_size == 32) {
> +			if (hw_local_id_generation)
> +				return 32;
> +			else
> +				return 64;
> +		}
> +	} else if (vrt == VRT_160) {
> +		if (simd_size == 16)
> +			return 48;
> +		else if (simd_size == 32) {
> +			if (hw_local_id_generation)
> +				return 32;
> +			else
> +				return 48;
> +		}
> +	} else if (vrt == VRT_192) {
> +		if (simd_size == 16)
> +			return 40;
> +		else if (simd_size == 32) {
> +			if (hw_local_id_generation)
> +				return 32;
> +			else
> +				return 40;
> +		}
> +	}

Quite complicated logic, could you document it with bspec number.


> +
> +	igt_warn("Unsupported VRT mode %d or SIMD size %d\n", vrt, simd_size);
> +	return 1;
> +}
> +
> +uint32_t compute_max_threads_in_tg(enum gpgpu_shader_vrt_modes vrt, bool large_grf_mode,
> +				   uint32_t simd_size, bool local_id_generation)
> +{
> +	if (vrt == VRT_DISABLED) {
> +		return compute_max_threads_in_tg_no_vrt(large_grf_mode, simd_size);
> +	} else {
> +		return compute_max_threads_in_tg_vrt(vrt, simd_size,
> +						     local_id_generation);
> +	}
> +}
> diff --git a/lib/gpgpu_shader.h b/lib/gpgpu_shader.h
> index 411ad6292..b4db48b32 100644
> --- a/lib/gpgpu_shader.h
> +++ b/lib/gpgpu_shader.h
> @@ -14,7 +14,12 @@ struct intel_bb;
>   struct intel_buf;
>   
>   enum gpgpu_shader_vrt_modes {
> +	VRT_32 = 0x0,
> +	VRT_64 = 0x1,
>   	VRT_96 = 0x2,
> +	VRT_128 = 0x3,
> +	VRT_160 = 0x4,
> +	VRT_192 = 0x5,
>   	VRT_DISABLED,
>   };
>   
> @@ -103,4 +108,8 @@ void gpgpu_shader__jump_neq(struct gpgpu_shader *shdr, int label_id,
>   			    uint32_t dw_offset, uint32_t value);
>   void gpgpu_shader__loop_begin(struct gpgpu_shader *shdr, int label_id);
>   void gpgpu_shader__loop_end(struct gpgpu_shader *shdr, int label_id, uint32_t iter);
> +
> +uint32_t compute_max_threads_in_tg(enum gpgpu_shader_vrt_modes vrt, bool large_grf_mode,
> +				   uint32_t simd_size, bool local_id_generation);

I wonder if it wouldn't be better to pass shader as parameter and put 
all parameters into struct gpgpu_shader?

Also for public function please add prefix, maybe: 
gpgpu_shader_get_max_threads_in_tg(struct gpgpu_shader shader);


Regards

Andrzej



> +
>   #endif /* GPGPU_SHADER_H */