[PATCH] drm/amdgpu: introduce a new parameter to configure how many KCQ we want(v3)
Liu, Monk
Monk.Liu at amd.com
Fri Jul 31 02:11:33 UTC 2020
[AMD Official Use Only - Internal Distribution Only]
>>> Indentation looks wrong. Did you use the wrong TAB size?
My TAB size is 4 space, and I don't know why it looks strange , but I can use space to replace the TABs anyway
Will send v4 patch against other suggestions from your side soon
_____________________________________
Monk Liu|GPU Virtualization Team |AMD
-----Original Message-----
From: Kuehling, Felix <Felix.Kuehling at amd.com>
Sent: Tuesday, July 28, 2020 10:38 PM
To: Liu, Monk <Monk.Liu at amd.com>; amd-gfx at lists.freedesktop.org
Subject: Re: [PATCH] drm/amdgpu: introduce a new parameter to configure how many KCQ we want(v3)
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_CTX4096
> #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