[PATCH] drm/amdgpu: introduce a new parameter to configure how many KCQ we want(v3)

Felix Kuehling felix.kuehling at amd.com
Tue Jul 28 14:38:19 UTC 2020


Am 2020-07-28 um 5:00 a.m. schrieb Monk Liu:
> what:
> the MQD's save and restore of KCQ (kernel compute queue)
> cost lots of clocks during world switch which impacts a lot
> to multi-VF performance
>
> how:
> introduce a paramter to control the number of KCQ to avoid
> performance drop if there is no kernel compute queue needed
>
> notes:
> this paramter only affects gfx 8/9/10
>
> v2:
> refine namings
>
> v3:
> choose queues for each ring to that try best to cross pipes evenly.

Thanks. Some more suggestions for simplifications inline.


>
> TODO:
> in the future we will let hypervisor driver to set this paramter
> automatically thus no need for user to configure it through
> modprobe in virtual machine
>
> Signed-off-by: Monk Liu <Monk.Liu at amd.com>
> ---
>  drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  1 +
>  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c |  5 +++
>  drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c    |  4 +++
>  drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c    | 58 +++++++++++++++---------------
>  drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c     | 30 ++++++++--------
>  drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c      | 29 +++++++--------
>  drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c      | 31 ++++++++--------
>  7 files changed, 87 insertions(+), 71 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> index e97c088..de11136 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> @@ -201,6 +201,7 @@ extern int amdgpu_si_support;
>  #ifdef CONFIG_DRM_AMDGPU_CIK
>  extern int amdgpu_cik_support;
>  #endif
> +extern int amdgpu_num_kcq;
>  
>  #define AMDGPU_VM_MAX_NUM_CTX			4096
>  #define AMDGPU_SG_THRESHOLD			(256*1024*1024)
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> index 62ecac9..cf445bab 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> @@ -1199,6 +1199,11 @@ static int amdgpu_device_check_arguments(struct amdgpu_device *adev)
>  
>  	amdgpu_gmc_tmz_set(adev);
>  
> +	if (amdgpu_num_kcq > 8 || amdgpu_num_kcq < 0) {
> +		amdgpu_num_kcq = 8;
> +		dev_warn(adev->dev, "set kernel compute queue number to 8 due to invalid paramter provided by user\n");
> +	}
> +
>  	return 0;
>  }
>  
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
> index 6291f5f..b545c40 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
> @@ -150,6 +150,7 @@ int amdgpu_noretry;
>  int amdgpu_force_asic_type = -1;
>  int amdgpu_tmz = 0;
>  int amdgpu_reset_method = -1; /* auto */
> +int amdgpu_num_kcq = -1;
>  
>  struct amdgpu_mgpu_info mgpu_info = {
>  	.mutex = __MUTEX_INITIALIZER(mgpu_info.mutex),
> @@ -765,6 +766,9 @@ module_param_named(tmz, amdgpu_tmz, int, 0444);
>  MODULE_PARM_DESC(reset_method, "GPU reset method (-1 = auto (default), 0 = legacy, 1 = mode0, 2 = mode1, 3 = mode2, 4 = baco)");
>  module_param_named(reset_method, amdgpu_reset_method, int, 0444);
>  
> +MODULE_PARM_DESC(num_kcq, "number of kernel compute queue user want to setup (8 if set to greater than 8 or less than 0, only affect gfx 8+)");
> +module_param_named(num_kcq, amdgpu_num_kcq, int, 0444);
> +
>  static const struct pci_device_id pciidlist[] = {
>  #ifdef  CONFIG_DRM_AMDGPU_SI
>  	{0x1002, 0x6780, PCI_ANY_ID, PCI_ANY_ID, 0, 0, CHIP_TAHITI},
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c
> index 8eff017..f83a9a7 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c
> @@ -202,40 +202,42 @@ bool amdgpu_gfx_is_high_priority_compute_queue(struct amdgpu_device *adev,
>  
>  void amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev)
>  {
> -	int i, queue, pipe, mec;
> +	int i, queue, pipe;
>  	bool multipipe_policy = amdgpu_gfx_is_multipipe_capable(adev);
> +	int max_queues_per_mec = min(adev->gfx.mec.num_pipe_per_mec *
> +								 adev->gfx.mec.num_queue_per_pipe,
> +								 adev->gfx.num_compute_rings);

Indentation looks wrong. Did you use the wrong TAB size?


> +
> +	if (multipipe_policy) {
> +		/* policy: make queues evenly cross all pipes on MEC1 only */
> +		for (i = 0; i < max_queues_per_mec; i++) {
> +			pipe = i % adev->gfx.mec.num_pipe_per_mec;
> +			queue = (i / adev->gfx.mec.num_pipe_per_mec) %
> +				adev->gfx.mec.num_queue_per_pipe;
> +
> +			set_bit(pipe * adev->gfx.mec.num_queue_per_pipe + queue,
> +					adev->gfx.mec.queue_bitmap);
> +		}
> +	} else {
> +		int mec;
>  
> -	/* policy for amdgpu compute queue ownership */
> -	for (i = 0; i < AMDGPU_MAX_COMPUTE_QUEUES; ++i) {
> -		queue = i % adev->gfx.mec.num_queue_per_pipe;
> -		pipe = (i / adev->gfx.mec.num_queue_per_pipe)
> -			% adev->gfx.mec.num_pipe_per_mec;
> -		mec = (i / adev->gfx.mec.num_queue_per_pipe)
> -			/ adev->gfx.mec.num_pipe_per_mec;
> -
> -		/* we've run out of HW */
> -		if (mec >= adev->gfx.mec.num_mec)
> -			break;
> +		/* policy: amdgpu owns all queues in the given pipe */
> +		for (i = 0; i < adev->gfx.num_compute_rings; ++i) {

You could also use i < max_queues_per_mec here.


> +			queue = i % adev->gfx.mec.num_queue_per_pipe;
> +			pipe = (i / adev->gfx.mec.num_queue_per_pipe)
> +				% adev->gfx.mec.num_pipe_per_mec;
> +			mec = (i / adev->gfx.mec.num_queue_per_pipe)
> +				/ adev->gfx.mec.num_pipe_per_mec;

Then mec will always be 0 and you can eliminate that variable.


>  
> -		if (multipipe_policy) {
> -			/* policy: amdgpu owns the first two queues of the first MEC */
> -			if (mec == 0 && queue < 2)
> -				set_bit(i, adev->gfx.mec.queue_bitmap);
> -		} else {
> -			/* policy: amdgpu owns all queues in the first pipe */
> -			if (mec == 0 && pipe == 0)
> -				set_bit(i, adev->gfx.mec.queue_bitmap);
> +			/* we've run out of HW */
> +			if (mec >= adev->gfx.mec.num_mec)
> +				break;

And you won't need this if (...) break; any more.

It'll make the two policy cases look more similar, so it's easier to see
how they are actually different.

Regards,
  Felix


> +
> +			set_bit(i, adev->gfx.mec.queue_bitmap);
>  		}
>  	}
>  
> -	/* update the number of active compute rings */
> -	adev->gfx.num_compute_rings =
> -		bitmap_weight(adev->gfx.mec.queue_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
> -
> -	/* If you hit this case and edited the policy, you probably just
> -	 * need to increase AMDGPU_MAX_COMPUTE_RINGS */
> -	if (WARN_ON(adev->gfx.num_compute_rings > AMDGPU_MAX_COMPUTE_RINGS))
> -		adev->gfx.num_compute_rings = AMDGPU_MAX_COMPUTE_RINGS;
> +	dev_info(adev->dev, "mec queue bitmap weight=%d\n", bitmap_weight(adev->gfx.mec.queue_bitmap, AMDGPU_MAX_COMPUTE_QUEUES));
>  }
>  
>  void amdgpu_gfx_graphics_queue_acquire(struct amdgpu_device *adev)
> diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c
> index db9f1e8..3a93b3c 100644
> --- a/drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c
> +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c
> @@ -4022,21 +4022,23 @@ static int gfx_v10_0_mec_init(struct amdgpu_device *adev)
>  	amdgpu_gfx_compute_queue_acquire(adev);
>  	mec_hpd_size = adev->gfx.num_compute_rings * GFX10_MEC_HPD_SIZE;
>  
> -	r = amdgpu_bo_create_reserved(adev, mec_hpd_size, PAGE_SIZE,
> -				      AMDGPU_GEM_DOMAIN_GTT,
> -				      &adev->gfx.mec.hpd_eop_obj,
> -				      &adev->gfx.mec.hpd_eop_gpu_addr,
> -				      (void **)&hpd);
> -	if (r) {
> -		dev_warn(adev->dev, "(%d) create HDP EOP bo failed\n", r);
> -		gfx_v10_0_mec_fini(adev);
> -		return r;
> -	}
> +	if (mec_hpd_size) {
> +		r = amdgpu_bo_create_reserved(adev, mec_hpd_size, PAGE_SIZE,
> +									  AMDGPU_GEM_DOMAIN_GTT,
> +									  &adev->gfx.mec.hpd_eop_obj,
> +									  &adev->gfx.mec.hpd_eop_gpu_addr,
> +									  (void **)&hpd);
> +		if (r) {
> +			dev_warn(adev->dev, "(%d) create HDP EOP bo failed\n", r);
> +			gfx_v10_0_mec_fini(adev);
> +			return r;
> +		}
>  
> -	memset(hpd, 0, mec_hpd_size);
> +		memset(hpd, 0, mec_hpd_size);
>  
> -	amdgpu_bo_kunmap(adev->gfx.mec.hpd_eop_obj);
> -	amdgpu_bo_unreserve(adev->gfx.mec.hpd_eop_obj);
> +		amdgpu_bo_kunmap(adev->gfx.mec.hpd_eop_obj);
> +		amdgpu_bo_unreserve(adev->gfx.mec.hpd_eop_obj);
> +	}
>  
>  	if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
>  		mec_hdr = (const struct gfx_firmware_header_v1_0 *)adev->gfx.mec_fw->data;
> @@ -7159,7 +7161,7 @@ static int gfx_v10_0_early_init(void *handle)
>  		break;
>  	}
>  
> -	adev->gfx.num_compute_rings = AMDGPU_MAX_COMPUTE_RINGS;
> +	adev->gfx.num_compute_rings = amdgpu_num_kcq;
>  
>  	gfx_v10_0_set_kiq_pm4_funcs(adev);
>  	gfx_v10_0_set_ring_funcs(adev);
> diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c
> index 8d72089..eb4b812 100644
> --- a/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c
> +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c
> @@ -1343,21 +1343,22 @@ static int gfx_v8_0_mec_init(struct amdgpu_device *adev)
>  	amdgpu_gfx_compute_queue_acquire(adev);
>  
>  	mec_hpd_size = adev->gfx.num_compute_rings * GFX8_MEC_HPD_SIZE;
> +	if (mec_hpd_size) {
> +		r = amdgpu_bo_create_reserved(adev, mec_hpd_size, PAGE_SIZE,
> +									  AMDGPU_GEM_DOMAIN_VRAM,
> +									  &adev->gfx.mec.hpd_eop_obj,
> +									  &adev->gfx.mec.hpd_eop_gpu_addr,
> +									  (void **)&hpd);
> +		if (r) {
> +			dev_warn(adev->dev, "(%d) create HDP EOP bo failed\n", r);
> +			return r;
> +		}
>  
> -	r = amdgpu_bo_create_reserved(adev, mec_hpd_size, PAGE_SIZE,
> -				      AMDGPU_GEM_DOMAIN_VRAM,
> -				      &adev->gfx.mec.hpd_eop_obj,
> -				      &adev->gfx.mec.hpd_eop_gpu_addr,
> -				      (void **)&hpd);
> -	if (r) {
> -		dev_warn(adev->dev, "(%d) create HDP EOP bo failed\n", r);
> -		return r;
> -	}
> -
> -	memset(hpd, 0, mec_hpd_size);
> +		memset(hpd, 0, mec_hpd_size);
>  
> -	amdgpu_bo_kunmap(adev->gfx.mec.hpd_eop_obj);
> -	amdgpu_bo_unreserve(adev->gfx.mec.hpd_eop_obj);
> +		amdgpu_bo_kunmap(adev->gfx.mec.hpd_eop_obj);
> +		amdgpu_bo_unreserve(adev->gfx.mec.hpd_eop_obj);
> +	}
>  
>  	return 0;
>  }
> @@ -5294,7 +5295,7 @@ static int gfx_v8_0_early_init(void *handle)
>  	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
>  
>  	adev->gfx.num_gfx_rings = GFX8_NUM_GFX_RINGS;
> -	adev->gfx.num_compute_rings = AMDGPU_MAX_COMPUTE_RINGS;
> +	adev->gfx.num_compute_rings = amdgpu_num_kcq;
>  	adev->gfx.funcs = &gfx_v8_0_gfx_funcs;
>  	gfx_v8_0_set_ring_funcs(adev);
>  	gfx_v8_0_set_irq_funcs(adev);
> diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c
> index e4e751f..43ad044 100644
> --- a/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c
> +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c
> @@ -1938,22 +1938,23 @@ static int gfx_v9_0_mec_init(struct amdgpu_device *adev)
>  	/* take ownership of the relevant compute queues */
>  	amdgpu_gfx_compute_queue_acquire(adev);
>  	mec_hpd_size = adev->gfx.num_compute_rings * GFX9_MEC_HPD_SIZE;
> +	if (mec_hpd_size) {
> +		r = amdgpu_bo_create_reserved(adev, mec_hpd_size, PAGE_SIZE,
> +									  AMDGPU_GEM_DOMAIN_VRAM,
> +									  &adev->gfx.mec.hpd_eop_obj,
> +									  &adev->gfx.mec.hpd_eop_gpu_addr,
> +									  (void **)&hpd);
> +		if (r) {
> +			dev_warn(adev->dev, "(%d) create HDP EOP bo failed\n", r);
> +			gfx_v9_0_mec_fini(adev);
> +			return r;
> +		}
>  
> -	r = amdgpu_bo_create_reserved(adev, mec_hpd_size, PAGE_SIZE,
> -				      AMDGPU_GEM_DOMAIN_VRAM,
> -				      &adev->gfx.mec.hpd_eop_obj,
> -				      &adev->gfx.mec.hpd_eop_gpu_addr,
> -				      (void **)&hpd);
> -	if (r) {
> -		dev_warn(adev->dev, "(%d) create HDP EOP bo failed\n", r);
> -		gfx_v9_0_mec_fini(adev);
> -		return r;
> -	}
> -
> -	memset(hpd, 0, mec_hpd_size);
> +		memset(hpd, 0, mec_hpd_size);
>  
> -	amdgpu_bo_kunmap(adev->gfx.mec.hpd_eop_obj);
> -	amdgpu_bo_unreserve(adev->gfx.mec.hpd_eop_obj);
> +		amdgpu_bo_kunmap(adev->gfx.mec.hpd_eop_obj);
> +		amdgpu_bo_unreserve(adev->gfx.mec.hpd_eop_obj);
> +	}
>  
>  	mec_hdr = (const struct gfx_firmware_header_v1_0 *)adev->gfx.mec_fw->data;
>  
> @@ -4625,7 +4626,7 @@ static int gfx_v9_0_early_init(void *handle)
>  		adev->gfx.num_gfx_rings = 0;
>  	else
>  		adev->gfx.num_gfx_rings = GFX9_NUM_GFX_RINGS;
> -	adev->gfx.num_compute_rings = AMDGPU_MAX_COMPUTE_RINGS;
> +	adev->gfx.num_compute_rings = amdgpu_num_kcq;
>  	gfx_v9_0_set_kiq_pm4_funcs(adev);
>  	gfx_v9_0_set_ring_funcs(adev);
>  	gfx_v9_0_set_irq_funcs(adev);


More information about the amd-gfx mailing list