drm/amdgpu: introduce a new parameter to configure how many KCQ we want(v2)
Subsystems (2)
PATCH 1/1 drm/amdgpu: introduce a new parameter to configure how many KCQ we want(v2)
what:
the MQD's save and restore of kernel compute queues cost lots of clocks during world switch which impacts a lot to multi-VF performance
how:
introduce a paramter to control the number of kernel compute queues to avoid performance drop if there is no kernel compute queue needed
notes:
this paramter only affects gfx 8/9/10
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
--- 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 | 27 +++++++++++++------------- 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, 71 insertions(+), 56 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +1 -0
@@ -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_user_set;#define AMDGPU_VM_MAX_NUM_CTX 4096#define AMDGPU_SG_THRESHOLD (256*1024*1024)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +5 -0
@@ -1199,6 +1199,11 @@ static int amdgpu_device_check_arguments(struct amdgpu_device *adev)amdgpu_gmc_tmz_set (adev); +if (amdgpu_num_kcq_user_set >8 || amdgpu_num_kcq_user_set <0 ) { + amdgpu_num_kcq_user_set =8 ; +dev_warn (adev-dev,"set KCQ number to 8 due to invalid paramter provided by user\n" ); + } +return 0 ; }
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +4 -0
@@ -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_user_set =8 ;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 KCQ 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_user_set, 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},
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +14 -13
@@ -202,7 +202,7 @@ 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, mec, j =0 ;bool multipipe_policy =amdgpu_gfx_is_multipipe_capable (adev);/* policy for amdgpu compute queue ownership */
@@ -219,23 +219,24 @@ void amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev)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); +if (mec ==0 && queue <2 ) { +if (j++ < adev->gfx.num_compute_rings) +set_bit (i, adev->gfx.mec.queue_bitmap); +else +break ; + } }else {/* policy: amdgpu owns all queues in the first pipe */ -if (mec ==0 && pipe ==0 ) -set_bit (i, adev->gfx.mec.queue_bitmap); +if (mec ==0 && pipe ==0 ) { +if (j++ < adev->gfx.num_compute_rings) +set_bit (i, adev->gfx.mec.queue_bitmap); +else +break ; + } } } -/* 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)
drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c +16 -14
@@ -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_user_set;gfx_v10_0_set_kiq_pm4_funcs (adev);gfx_v10_0_set_ring_funcs (adev);
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +15 -14
@@ -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_user_set; adev->gfx.funcs = &gfx_v8_0_gfx_funcs;gfx_v8_0_set_ring_funcs (adev);gfx_v8_0_set_irq_funcs (adev);
drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c +16 -15
@@ -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_user_set;gfx_v9_0_set_kiq_pm4_funcs (adev);gfx_v9_0_set_ring_funcs (adev);gfx_v9_0_set_irq_funcs (adev);
Signature · 6 lines
2.7.4 _______________________________________________ amd-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Am 2020-07-27 um 6:47 a.m. schrieb Monk Liu:
Quoted 91 lines
what:
the MQD's save and restore of kernel compute queues cost lots of clocks during world switch which impacts a lot to multi-VF performancehow:
introduce a paramter to control the number of kernel compute queues to avoid performance drop if there is no kernel compute queue needednotes:
this paramter only affects gfx 8/9/10TODO:
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 machineSigned-off-by Monk Liu <[email protected]>--- 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 | 27 +++++++++++++------------- 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, 71 insertions(+), 56 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu.h +1
-0@@ -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_user_set;#define AMDGPU_VM_MAX_NUM_CTX 4096#define AMDGPU_SG_THRESHOLD (256*1024*1024)drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +5
-0@@ -1199,6 +1199,11 @@ static int amdgpu_device_check_arguments(struct amdgpu_device *adev)amdgpu_gmc_tmz_set (adev); +if (amdgpu_num_kcq_user_set >8 || amdgpu_num_kcq_user_set <0 ) { + amdgpu_num_kcq_user_set =8 ; +dev_warn (adev-dev,"set KCQ number to 8 due to invalid paramter provided by user\n" ); + } +return 0 ; }drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +4
-0@@ -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_user_set =8 ;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 KCQ 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_user_set, 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},drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +1
-1@@ -202,7 +202,7 @@ 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, mec, j =0 ;bool multipipe_policy =amdgpu_gfx_is_multipipe_capable (adev);/* policy for amdgpu compute queue ownership */
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c void amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev) · line 223
Quoted 7 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +2
-2@@ -219,23 +219,24 @@ void amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev)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); +if (mec ==0 && queue <2 ) { +if (j++ < adev->gfx.num_compute_rings)
This is not ideal, because it wouldn't distribute the queues evenly across pipes if there are fewer than 7. I would change how queue and pipe are calculated from i for the multipipe_policy case:
if (multipipe_policy) {
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;
} else {
/* previous way */
}Then your change will work better.
Regards,
Felix
Quoted 171 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +11
-10+set_bit (i, adev->gfx.mec.queue_bitmap); + else + break; + } }else {/* policy: amdgpu owns all queues in the first pipe */ -if (mec ==0 && pipe ==0 ) -set_bit (i, adev->gfx.mec.queue_bitmap); +if (mec ==0 && pipe ==0 ) { +if (j++ < adev->gfx.num_compute_rings) +set_bit (i, adev->gfx.mec.queue_bitmap); +else +break ; + } } } -/* 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)drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c +16
-14@@ -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_user_set;gfx_v10_0_set_kiq_pm4_funcs (adev);gfx_v10_0_set_ring_funcs (adev);drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +15
-14@@ -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_user_set; adev->gfx.funcs = &gfx_v8_0_gfx_funcs;gfx_v8_0_set_ring_funcs (adev);gfx_v8_0_set_irq_funcs (adev);drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c +15
-14@@ -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;
drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c static int gfx_v9_0_early_init(void *handle) · line 4632
Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c +1
-1@@ -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_user_set;gfx_v9_0_set_kiq_pm4_funcs (adev);gfx_v9_0_set_ring_funcs (adev);gfx_v9_0_set_irq_funcs (adev);
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
[AMD Official Use Only - Internal Distribution Only]
Thanks Felix
I reworked my patch with your suggestion and I can get queues evenly cross pipes, e.g.: modprobe amdgpu num_kcq=6
[ 409.878557] amdgpu 0000:00:07.0: amdgpu: ring comp_1.0.0 uses VM inv eng 1 on hub 0 [ 409.878559] amdgpu 0000:00:07.0: amdgpu: ring comp_1.1.0 uses VM inv eng 4 on hub 0 [ 409.878561] amdgpu 0000:00:07.0: amdgpu: ring comp_1.2.0 uses VM inv eng 5 on hub 0 [ 409.878563] amdgpu 0000:00:07.0: amdgpu: ring comp_1.3.0 uses VM inv eng 6 on hub 0 [ 409.878565] amdgpu 0000:00:07.0: amdgpu: ring comp_1.0.1 uses VM inv eng 7 on hub 0 [ 409.878567] amdgpu 0000:00:07.0: amdgpu: ring comp_1.1.1 uses VM inv eng 8 on hub 0 [ 409.878568] amdgpu 0000:00:07.0: amdgpu: ring kiq_2.1.0 uses VM inv eng 9 on hub 0
Please review my patch upcoming
_____________________________________
Monk Liu|GPU Virtualization Team |AMD
-----Original Message-----
From: Kuehling, Felix <[email protected]>
Sent: Tuesday, July 28, 2020 7:33 AM
To: [email protected]; Liu, Monk <[email protected]>
Subject: Re: [PATCH] drm/amdgpu: introduce a new parameter to configure how many KCQ we want(v2)
Am 2020-07-27 um 6:47 a.m. schrieb Monk Liu:
Quoted 97 lines
what:
the MQD's save and restore of kernel compute queues cost lots of clocks during world switch which impacts a lot to multi-VF performancehow:
introduce a paramter to control the number of kernel compute queues to avoid performance drop if there is no kernel compute queue needednotes:
this paramter only affects gfx 8/9/10TODO:
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 machineSigned-off-by Monk Liu <[email protected]>--- 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 | 27 +++++++++++++------------- 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, 71 insertions(+), 56 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu.h +0
-0@@ -201,6 +201,7 @@ extern int amdgpu_si_support; #ifdef
CONFIG_DRM_AMDGPU_CIKextern intamdgpu_cik_support; #endifdrivers/gpu/drm/amd/amdgpu/amdgpu.h +1
-0+extern int amdgpu_num_kcq_user_set;#define AMDGPU_VM_MAX_NUM_CTX4096 #define AMDGPU_SG_THRESHOLD(256*1024*1024)drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +5
-0@@ -1199,6 +1199,11 @@ static int amdgpu_device_check_arguments(struct amdgpu_device *adev)amdgpu_gmc_tmz_set (adev ); +if (amdgpu_num_kcq_user_set >8 || amdgpu_num_kcq_user_set <0 ) { +amdgpu_num_kcq_user_set =8 ; +dev_warn (adev-dev,"set KCQ number to 8 due to invalid paramter provided by user\n" ); +} +return 0 ; }drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +5
-1@@ -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_user_set =8 ;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 KCQ 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_user_set, 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
--gita/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.cb/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c
index 8eff017..0b59049 100644drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +1
-1@@ -202,7 +202,7 @@ 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, mec, j = 0;bool multipipe_policy =amdgpu_gfx_is_multipipe_capable (adev); /* policy for amdgpu compute queue ownership */ @@ -219,23 +219,24@@ void
amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev)if (
multipipe_policy) {
/* policy: amdgpu owns the first two queues of the first MEC */
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c void amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev) · line 223
Quoted 4 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +2
-2-if (mec ==0 && queue <2 ) -set_bit (i, adev->gfx.mec.queue_bitmap); +if (mec ==0 && queue <2 ) { +if (j++ < adev->gfx.num_compute_rings)
This is not ideal, because it wouldn't distribute the queues evenly across pipes if there are fewer than 7. I would change how queue and pipe are calculated from i for the multipipe_policy case:
if (multipipe_policy) {
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;
} else {
/* previous way */
}
Then your change will work better.
Regards,
Felix
Quoted 177 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +13
-10+set_bit (i, adev->gfx.mec.queue_bitmap); +else +break; +} }else {/* policy: amdgpu owns all queues in the first pipe */ -if (mec ==0 && pipe ==0 ) -set_bit (i, adev->gfx.mec.queue_bitmap); +if (mec ==0 && pipe ==0 ) { +if (j++ < adev->gfx.num_compute_rings) +set_bit (i, adev->gfx.mec.queue_bitmap); +else +break ; +} } } -/* 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)drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c +15
-13@@ -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;
}drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c +1
-1-adev->gfx.num_compute_rings =AMDGPU_MAX_COMPUTE_RINGS ; +adev->gfx.num_compute_rings = amdgpu_num_kcq_user_set;gfx_v10_0_set_kiq_pm4_funcs (adev);gfx_v10_0_set_ring_funcs (adev);drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +15
-14@@ -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_user_set; adev->gfx.funcs = &gfx_v8_0_gfx_funcs;gfx_v8_0_set_ring_funcs (adev);gfx_v8_0_set_irq_funcs (adev);drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c +15
-14@@ -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;
drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c static int gfx_v9_0_early_init(void *handle) · line 4632
Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c +1
-1@@ -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_user_set;gfx_v9_0_set_kiq_pm4_funcs (adev);gfx_v9_0_set_ring_funcs (adev);gfx_v9_0_set_irq_funcs (adev);
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx