drm/amdgpu: Add a helper macro to align mqd size
Subsystems (3)
PATCH 1/4 drm/amdgpu: Add a helper macro to align mqd size
MES FW uses addr(mqd_addr + sizeof(struct mqd) + 3*sizeof(uint32_t)) as fence address and writes a 32 bit fence value to this address.
Driver needs to allocate some extra memory(at least 4 DWs) in addition to sizeof(struct mqd) as mqd memory.
For gfx11/12, sizeof(struct mqd) < PAGE_SIZE, allocate mqd memory with PAGE_SIZE aligned works. For gfx12.1, sizeof(struct mqd) == PAGE_SIZE, it doesn't work.
KFD mqd manager hardcodes mqd size to PAGE_SIZE/MQD_SIZE.
Let's use AMDGPU_MQD_SIZE_ALIGN to avoid hardcoding in differnet place and across different IP version. It is used in two place.
1. mqd memory alloction
2. mqd stride initialization
--- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 5 +++++ 1 file changed, 5 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +5 -0
@@ -1239,6 +1239,11 @@ struct amdgpu_device {struct amdgpu_kfd_dev kfd; }; +/* +* MES will use memory beyond struct MQD size, 5 DWs currently +*/ +#define AMDGPU_MQD_SIZE_ALIGN (mqd_size) ALIGN(((mqd_size) + 20), PAGE_SIZE) +static inline uint32_t amdgpu_ip_version (const struct amdgpu_device *adev,uint8_t ip,uint8_t inst) {
Signature · 1 line
2.34.1
PATCH 2/4 drm/amdgpu: Use AMDGPU_MQD_SIZE_ALIGN in KGD
Use AMDGPU_MQD_SIZE_ALIGN for both kernel and user queue.
--- drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c | 18 ++++++++++-------- drivers/gpu/drm/amd/amdgpu/gfx_v12_1.c | 2 +- drivers/gpu/drm/amd/amdgpu/mes_userqueue.c | 3 ++- 3 files changed, 13 insertions(+), 10 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +10 -8
@@ -385,6 +385,8 @@ int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];struct amdgpu_ring *ring = &kiq->ring;u32 domain =AMDGPU_GEM_DOMAIN_GTT ; +u32 gfx_mqd_size =max (adev->mqds[AMDGPU_HW_IP_GFX ].mqd_size, mqd_size); +u32 compute_mqd_size =max (adev->mqds[AMDGPU_HW_IP_COMPUTE ].mqd_size, mqd_size);#if !defined(CONFIG_ARM ) && !defined(CONFIG_ARM64 )/* Only enable on gfx10 and 11 for now to avoid changing behavior on older chips */
@@ -424,17 +426,17 @@ int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,for (i =0 ; i < adev->gfx.num_gfx_rings; i++) { ring = &adev->gfx.gfx_ring[i];if (!ring->mqd_obj) { - r =amdgpu_bo_create_kernel (adev, mqd_size,PAGE_SIZE , - domain, &ring->mqd_obj, + r =amdgpu_bo_create_kernel (adev,AMDGPU_MQD_SIZE_ALIGN (gfx_mqd_size), +PAGE_SIZE , domain, &ring->mqd_obj, &ring->mqd_gpu_addr, &ring->mqd_ptr);if (r) {dev_warn (adev->dev,"failed to create ring mqd bo (%d)" , r);return r; } - ring->mqd_size = mqd_size; + ring->mqd_size = gfx_mqd_size;/* prepare MQD backup */ - adev->gfx.me.mqd_backup[i] =kzalloc (mqd_size,GFP_KERNEL ); + adev->gfx.me.mqd_backup[i] =kzalloc (gfx_mqd_size,GFP_KERNEL );if (!adev->gfx.me.mqd_backup[i]) {dev_warn (adev->dev,"no memory to create MQD backup for ring %s\n" , ring->name);return -ENOMEM ;
@@ -448,17 +450,17 @@ int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev, j = i + xcc_id * adev->gfx.num_compute_rings; ring = &adev->gfx.compute_ring[j];if (!ring->mqd_obj) { - r =amdgpu_bo_create_kernel (adev, mqd_size,PAGE_SIZE , - domain, &ring->mqd_obj, + r =amdgpu_bo_create_kernel (adev,AMDGPU_MQD_SIZE_ALIGN (compute_mqd_size), +PAGE_SIZE , domain, &ring->mqd_obj, &ring->mqd_gpu_addr, &ring->mqd_ptr);if (r) {dev_warn (adev->dev,"failed to create ring mqd bo (%d)" , r);return r; } - ring->mqd_size = mqd_size; + ring->mqd_size = compute_mqd_size;/* prepare MQD backup */ - adev->gfx.mec.mqd_backup[j] =kzalloc (mqd_size,GFP_KERNEL ); + adev->gfx.mec.mqd_backup[j] =kzalloc (compute_mqd_size,GFP_KERNEL );if (!adev->gfx.mec.mqd_backup[j]) {dev_warn (adev->dev,"no memory to create MQD backup for ring %s\n" , ring->name);return -ENOMEM ;
drivers/gpu/drm/amd/amdgpu/gfx_v12_1.c +1 -1
@@ -2235,7 +2235,7 @@ static int gfx_v12_1_compute_mqd_init(struct amdgpu_device *adev, void *m, mqd->cp_hqd_queue_priority = prop->hqd_queue_priority; mqd->cp_mqd_stride_size = prop->mqd_stride_size ? prop->mqd_stride_size : -sizeof (struct v12_1_compute_mqd ); +AMDGPU_MQD_SIZE_ALIGN (adev->mqds[AMDGPU_HW_IP_COMPUTE ].mqd_size); mqd->cp_hqd_active = prop->hqd_active;
drivers/gpu/drm/amd/amdgpu/mes_userqueue.c +2 -1
@@ -289,7 +289,8 @@ static int mes_userq_mqd_create(struct amdgpu_usermode_queue *queue,return -ENOMEM ; } - r =amdgpu_userq_create_object (uq_mgr, &queue->mqd, mqd_hw_default->mqd_size); + r =amdgpu_userq_create_object (uq_mgr, &queue->mqd, +AMDGPU_MQD_SIZE_ALIGN (mqd_hw_default->mqd_size));if (r) {DRM_ERROR ("Failed to create MQD object for userqueue\n" ); goto free_props;
Signature · 1 line
2.34.1
PATCH 3/4 drm/amdkfd: Adjust parameter of allocate_mqd
In reply to Lang YuMake allocate_mqd consistent with other callbacks. Prepare for next patch to use mqd_manager->mqd_size.
--- drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 4 ++-- drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c | 6 ++++-- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h | 6 +++--- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c | 3 ++- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c | 3 ++- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c | 3 ++- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12.c | 3 ++- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12_1.c | 3 ++- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c | 3 ++- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c | 3 ++- 11 files changed, 24 insertions(+), 15 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +2 -2
@@ -676,7 +676,7 @@ static int create_queue_nocpsch(struct device_queue_manager *dqm,/* Temporarily release dqm lock to avoid a circular lock dependency */ dqm_unlock (dqm); - q->mqd_mem_obj = mqd_mgr->allocate_mqd (mqd_mgr->dev, &q->properties); + q->mqd_mem_obj = mqd_mgr->allocate_mqd (mqd_mgr, &q->properties);dqm_lock (dqm);if (!q->mqd_mem_obj) {
@@ -2002,7 +2002,7 @@ static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q, dqm->asic_ops.init_sdma_vm (dqm, q, qpd); q->properties.tba_addr = qpd->tba_addr; q->properties.tma_addr = qpd->tma_addr; - q->mqd_mem_obj = mqd_mgr->allocate_mqd (mqd_mgr->dev, &q->properties); + q->mqd_mem_obj = mqd_mgr->allocate_mqd (mqd_mgr, &q->properties);if (!q->mqd_mem_obj) { retval = -ENOMEM ; goto out_deallocate_doorbell;
drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c +1 -1
@@ -130,7 +130,7 @@ static bool kq_initialize(struct kernel_queue *kq, struct kfd_node *dev, kq->queue->device = dev; - kq->queue->mqd_mem_obj = kq->mqd_mgr->allocate_mqd (kq->mqd_mgr->dev, + kq->queue->mqd_mem_obj = kq->mqd_mgr->allocate_mqd (kq->mqd_mgr, &kq->queue->properties);if (!kq->queue->mqd_mem_obj) goto err_allocate_mqd;
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c +4 -2
@@ -46,9 +46,10 @@ int pipe_priority_map[] = { KFD_PIPE_PRIORITY_CS_HIGH }; -struct kfd_mem_obj *allocate_hiq_mqd (struct kfd_node *dev,struct queue_properties *q) +struct kfd_mem_obj *allocate_hiq_mqd (struct mqd_manager *mm,struct queue_properties *q) {struct kfd_mem_obj *mqd_mem_obj; +struct kfd_node *dev = mm->dev; mqd_mem_obj =kzalloc (sizeof (struct kfd_mem_obj ),GFP_KERNEL );if (!mqd_mem_obj)
@@ -61,10 +62,11 @@ struct kfd_mem_obj *allocate_hiq_mqd(struct kfd_node *dev, struct queue_propertireturn mqd_mem_obj; } -struct kfd_mem_obj *allocate_sdma_mqd (struct kfd_node *dev, +struct kfd_mem_obj *allocate_sdma_mqd (struct mqd_manager *mm,struct queue_properties *q) {struct kfd_mem_obj *mqd_mem_obj; +struct kfd_node *dev = mm->dev;uint64_t offset; mqd_mem_obj =kzalloc (sizeof (struct kfd_mem_obj ),GFP_KERNEL );
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h +3 -3
@@ -68,7 +68,7 @@*/ extern int pipe_priority_map[];struct mqd_manager { -struct kfd_mem_obj * (*allocate_mqd)(struct kfd_node *kfd, +struct kfd_mem_obj * (*allocate_mqd)(struct mqd_manager *mm,struct queue_properties *q); void (*init_mqd)(struct mqd_manager *mm, void **mqd,
@@ -153,10 +153,10 @@ struct mqd_user_context_save_area_header {uint32_t wave_state_size; }; -struct kfd_mem_obj *allocate_hiq_mqd (struct kfd_node *dev, +struct kfd_mem_obj *allocate_hiq_mqd (struct mqd_manager *mm,struct queue_properties *q); -struct kfd_mem_obj *allocate_sdma_mqd (struct kfd_node *dev, +struct kfd_mem_obj *allocate_sdma_mqd (struct mqd_manager *mm,struct queue_properties *q);void free_mqd_hiq_sdma (struct mqd_manager *mm,void *mqd,struct kfd_mem_obj *mqd_mem_obj);
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c +2 -1
@@ -73,9 +73,10 @@ static void set_priority(struct cik_mqd *m, struct queue_properties *q) m->cp_hqd_queue_priority = q->priority; } -static struct kfd_mem_obj *allocate_mqd (struct kfd_node *kfd, +static struct kfd_mem_obj *allocate_mqd (struct mqd_manager *mm,struct queue_properties *q) { +struct kfd_node *kfd = mm->dev;struct kfd_mem_obj *mqd_mem_obj; if (kfd_gtt_sa_allocate(kfd, sizeof(struct cik_mqd),
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c +2 -1
@@ -73,9 +73,10 @@ static void set_priority(struct v10_compute_mqd *m, struct queue_properties *q) m->cp_hqd_queue_priority = q->priority; } -static struct kfd_mem_obj *allocate_mqd (struct kfd_node *kfd, +static struct kfd_mem_obj *allocate_mqd (struct mqd_manager *mm,struct queue_properties *q) { +struct kfd_node *kfd = mm->dev;struct kfd_mem_obj *mqd_mem_obj; if (kfd_gtt_sa_allocate(kfd, sizeof(struct v10_compute_mqd),
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c +2 -1
@@ -99,9 +99,10 @@ static void set_priority(struct v11_compute_mqd *m, struct queue_properties *q) m->cp_hqd_queue_priority = q->priority; } -static struct kfd_mem_obj *allocate_mqd (struct kfd_node *node, +static struct kfd_mem_obj *allocate_mqd (struct mqd_manager *mm,struct queue_properties *q) { +struct kfd_node *node = mm->dev;struct kfd_mem_obj *mqd_mem_obj;int size;
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12.c +2 -1
@@ -80,9 +80,10 @@ static void set_priority(struct v12_compute_mqd *m, struct queue_properties *q) m->cp_hqd_queue_priority = q->priority; } -static struct kfd_mem_obj *allocate_mqd (struct kfd_node *node, +static struct kfd_mem_obj *allocate_mqd (struct mqd_manager *mm,struct queue_properties *q) { +struct kfd_node *node = mm->dev;struct kfd_mem_obj *mqd_mem_obj;/*
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12_1.c +2 -1
@@ -145,9 +145,10 @@ static void set_priority(struct v12_1_compute_mqd *m, struct queue_properties *q m->cp_hqd_queue_priority = q->priority; } -static struct kfd_mem_obj *allocate_mqd (struct kfd_node *node, +static struct kfd_mem_obj *allocate_mqd (struct mqd_manager *mm,struct queue_properties *q) { +struct kfd_node *node = mm->dev;struct kfd_mem_obj *mqd_mem_obj;unsigned int size;
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c +2 -1
@@ -120,10 +120,11 @@ static bool mqd_on_vram(struct amdgpu_device *adev) } } -static struct kfd_mem_obj *allocate_mqd (struct kfd_node *node, +static struct kfd_mem_obj *allocate_mqd (struct mqd_manager *mm,struct queue_properties *q) {int retval; +struct kfd_node *node = mm->dev;struct kfd_mem_obj *mqd_mem_obj =NULL ;/* For V9 only, due to a HW bug, the control stack of a user mode
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c +2 -1
@@ -76,9 +76,10 @@ static void set_priority(struct vi_mqd *m, struct queue_properties *q) m->cp_hqd_queue_priority = q->priority; } -static struct kfd_mem_obj *allocate_mqd (struct kfd_node *kfd, +static struct kfd_mem_obj *allocate_mqd (struct mqd_manager *mm,struct queue_properties *q) { +struct kfd_node *kfd = mm->dev;struct kfd_mem_obj *mqd_mem_obj; if (kfd_gtt_sa_allocate(kfd, sizeof(struct vi_mqd),
Signature · 1 line
2.34.1
PATCH 4/4 drm/amdkfd: Use AMDGPU_MQD_SIZE_ALIGN in gfx11+ kfd mqd manager
In reply to Lang YuMES is enabled by default from gfx11+.
--- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c | 3 ++ .../gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c | 22 +++----------- .../gpu/drm/amd/amdkfd/kfd_mqd_manager_v12.c | 10 +++---- .../drm/amd/amdkfd/kfd_mqd_manager_v12_1.c | 29 ++++--------------- 4 files changed, 17 insertions(+), 47 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c +3 -0
@@ -292,6 +292,9 @@ void kfd_get_hiq_xcc_mqd(struct kfd_node *dev, struct kfd_mem_obj *mqd_mem_obj,uint64_t kfd_mqd_stride (struct mqd_manager *mm,struct queue_properties *q) { +if (KFD_GC_VERSION (mm->dev) >=IP_VERSION (11 ,0 ,0 )) +return AMDGPU_MQD_SIZE_ALIGN (mm->mqd_size); +return mm->mqd_size; }
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c +4 -18
@@ -102,20 +102,11 @@ static void set_priority(struct v11_compute_mqd *m, struct queue_properties *q)static struct kfd_mem_obj *allocate_mqd (struct mqd_manager *mm,struct queue_properties *q) { +u32 mqd_size =AMDGPU_MQD_SIZE_ALIGN (mm->mqd_size);struct kfd_node *node = mm->dev;struct kfd_mem_obj *mqd_mem_obj; -int size; - -/* -* MES write to areas beyond MQD size. So allocate -* 1 PAGE_SIZE memory for MQD is MES is enabled. -*/ -if (node->kfd->shared_resources.enable_mes) - size =PAGE_SIZE ; -else - size =sizeof (struct v11_compute_mqd ); -if (kfd_gtt_sa_allocate (node, size, &mqd_mem_obj)) +if (kfd_gtt_sa_allocate (node, mqd_size, &mqd_mem_obj))return NULL ;return mqd_mem_obj;
@@ -127,18 +118,13 @@ static void init_mqd(struct mqd_manager *mm, void **mqd, {uint64_t addr;struct v11_compute_mqd *m; -int size; +u32 mqd_size =AMDGPU_MQD_SIZE_ALIGN (mm->mqd_size);uint32_t wa_mask = q->is_dbg_wa ?0xffff :0xffffffff ; m = (struct v11_compute_mqd *) mqd_mem_obj->cpu_ptr; addr = mqd_mem_obj->gpu_addr; -if (mm->dev->kfd->shared_resources.enable_mes) - size =PAGE_SIZE ; -else - size =sizeof (struct v11_compute_mqd ); - -memset (m,0 , size); +memset (m,0 , mqd_size); m->header =0xC0310800 ; m->compute_pipelinestat_enable =1 ;
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12.c +4 -6
@@ -83,14 +83,11 @@ static void set_priority(struct v12_compute_mqd *m, struct queue_properties *q)static struct kfd_mem_obj *allocate_mqd (struct mqd_manager *mm,struct queue_properties *q) { +u32 mqd_size =AMDGPU_MQD_SIZE_ALIGN (mm->mqd_size);struct kfd_node *node = mm->dev;struct kfd_mem_obj *mqd_mem_obj; -/* -* Allocate one PAGE_SIZE memory for MQD as MES writes to areas beyond -* struct MQD size. -*/ -if (kfd_gtt_sa_allocate (node,PAGE_SIZE , &mqd_mem_obj)) +if (kfd_gtt_sa_allocate (node, mqd_size, &mqd_mem_obj))return NULL ;return mqd_mem_obj;
@@ -102,11 +99,12 @@ static void init_mqd(struct mqd_manager *mm, void **mqd, {uint64_t addr;struct v12_compute_mqd *m; +u32 mqd_size =AMDGPU_MQD_SIZE_ALIGN (mm->mqd_size); m = (struct v12_compute_mqd *) mqd_mem_obj->cpu_ptr; addr = mqd_mem_obj->gpu_addr; -memset (m,0 ,PAGE_SIZE ); +memset (m,0 , mqd_size); m->header =0xC0310800 ; m->compute_pipelinestat_enable =1 ;
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12_1.c +6 -23
@@ -32,17 +32,6 @@#include "amdgpu_amdkfd.h" #include "kfd_device_queue_manager.h" -#define MQD_SIZE (2 * PAGE_SIZE) - -static uint64_t mqd_stride_v12_1 (struct mqd_manager *mm, -struct queue_properties *q) -{ -if (q->type ==KFD_QUEUE_TYPE_COMPUTE ) -return MQD_SIZE ; -else -return PAGE_SIZE ; -} -static inline struct v12_1_compute_mqd *get_mqd (void *mqd) {return (struct v12_1_compute_mqd *)mqd;
@@ -148,21 +137,14 @@ static void set_priority(struct v12_1_compute_mqd *m, struct queue_properties *qstatic struct kfd_mem_obj *allocate_mqd (struct mqd_manager *mm,struct queue_properties *q) { +u32 mqd_size =AMDGPU_MQD_SIZE_ALIGN (mm->mqd_size);struct kfd_node *node = mm->dev;struct kfd_mem_obj *mqd_mem_obj; -unsigned int size; -/* -* Allocate two PAGE_SIZE memory for Compute MQD as MES writes to areas beyond -* struct MQD size. Size of the Compute MQD is 1 PAGE_SIZE. -* For SDMA MQD, we allocate 1 Page_size. -*/ if (q->type ==KFD_QUEUE_TYPE_COMPUTE ) - size =MQD_SIZE *NUM_XCC (node->xcc_mask); -else - size =PAGE_SIZE ; + mqd_size *=NUM_XCC (node->xcc_mask); -if (kfd_gtt_sa_allocate (node, size, &mqd_mem_obj)) +if (kfd_gtt_sa_allocate (node, mqd_size, &mqd_mem_obj))return NULL ;return mqd_mem_obj;
@@ -174,11 +156,12 @@ static void init_mqd(struct mqd_manager *mm, void **mqd, {uint64_t addr;struct v12_1_compute_mqd *m; +u32 mqd_size =AMDGPU_MQD_SIZE_ALIGN (mm->mqd_size); m = (struct v12_1_compute_mqd *) mqd_mem_obj->cpu_ptr; addr = mqd_mem_obj->gpu_addr; -memset (m,0 ,MQD_SIZE ); +memset (m,0 , mqd_size); m->header =0xC0310800 ; m->compute_pipelinestat_enable =1 ;
@@ -681,7 +664,7 @@ struct mqd_manager *mqd_manager_init_v12_1(enum KFD_MQD_TYPE type, mqd->is_occupied = kfd_is_occupied_cp; mqd->mqd_size =sizeof (struct v12_1_compute_mqd ); mqd->get_wave_state = get_wave_state_v12_1; - mqd->mqd_stride = mqd_stride_v12_1; + mqd->mqd_stride = kfd_mqd_stride;#if defined(CONFIG_DEBUG_FS ) mqd->debugfs_show_mqd = debugfs_show_mqd;#endif
Signature · 1 line
2.34.1
Re: [PATCH 4/4] drm/amdkfd: Use AMDGPU_MQD_SIZE_ALIGN in gfx11+ kfd mqd manager
Full set reviewed.
Quoted 177 lines · Lang Yu
MES is enabled by default from gfx11+.
Signed-off-by Lang Yu <[email protected]>--- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c | 3 ++ .../gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c | 22 +++----------- .../gpu/drm/amd/amdkfd/kfd_mqd_manager_v12.c | 10 +++---- .../drm/amd/amdkfd/kfd_mqd_manager_v12_1.c | 29 ++++--------------- 4 files changed, 17 insertions(+), 47 deletions(-)drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c +3
-0@@ -292,6 +292,9 @@ void kfd_get_hiq_xcc_mqd(struct kfd_node *dev, struct kfd_mem_obj *mqd_mem_obj,uint64_t kfd_mqd_stride (struct mqd_manager *mm,struct queue_properties *q) { +if (KFD_GC_VERSION (mm->dev) >=IP_VERSION (11 ,0 ,0 )) +return AMDGPU_MQD_SIZE_ALIGN (mm->mqd_size); +return mm->mqd_size; }drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c +4
-18@@ -102,20 +102,11 @@ static void set_priority(struct v11_compute_mqd *m, struct queue_properties *q)static struct kfd_mem_obj *allocate_mqd (struct mqd_manager *mm,struct queue_properties *q) { +u32 mqd_size =AMDGPU_MQD_SIZE_ALIGN (mm->mqd_size);struct kfd_node *node = mm->dev;struct kfd_mem_obj *mqd_mem_obj; -int size; - -/* -* MES write to areas beyond MQD size. So allocate -* 1 PAGE_SIZE memory for MQD is MES is enabled. -*/ -if (node->kfd->shared_resources.enable_mes) - size =PAGE_SIZE ; -else - size =sizeof (struct v11_compute_mqd ); -if (kfd_gtt_sa_allocate (node, size, &mqd_mem_obj)) +if (kfd_gtt_sa_allocate (node, mqd_size, &mqd_mem_obj))return NULL ;return mqd_mem_obj;@@ -127,18 +118,13 @@ static void init_mqd(struct mqd_manager *mm, void **mqd, {uint64_t addr;struct v11_compute_mqd *m; -int size; +u32 mqd_size =AMDGPU_MQD_SIZE_ALIGN (mm->mqd_size);uint32_t wa_mask = q->is_dbg_wa ?0xffff :0xffffffff ; m = (struct v11_compute_mqd *) mqd_mem_obj->cpu_ptr; addr = mqd_mem_obj->gpu_addr; -if (mm->dev->kfd->shared_resources.enable_mes) - size =PAGE_SIZE ; -else - size =sizeof (struct v11_compute_mqd ); - -memset (m,0 , size); +memset (m,0 , mqd_size); m->header =0xC0310800 ; m->compute_pipelinestat_enable =1 ;drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12.c +4
-6@@ -83,14 +83,11 @@ static void set_priority(struct v12_compute_mqd *m, struct queue_properties *q)static struct kfd_mem_obj *allocate_mqd (struct mqd_manager *mm,struct queue_properties *q) { +u32 mqd_size =AMDGPU_MQD_SIZE_ALIGN (mm->mqd_size);struct kfd_node *node = mm->dev;struct kfd_mem_obj *mqd_mem_obj; -/* -* Allocate one PAGE_SIZE memory for MQD as MES writes to areas beyond -* struct MQD size. -*/ -if (kfd_gtt_sa_allocate (node,PAGE_SIZE , &mqd_mem_obj)) +if (kfd_gtt_sa_allocate (node, mqd_size, &mqd_mem_obj))return NULL ;return mqd_mem_obj;@@ -102,11 +99,12 @@ static void init_mqd(struct mqd_manager *mm, void **mqd, {uint64_t addr;struct v12_compute_mqd *m; +u32 mqd_size =AMDGPU_MQD_SIZE_ALIGN (mm->mqd_size); m = (struct v12_compute_mqd *) mqd_mem_obj->cpu_ptr; addr = mqd_mem_obj->gpu_addr; -memset (m,0 ,PAGE_SIZE ); +memset (m,0 , mqd_size); m->header =0xC0310800 ; m->compute_pipelinestat_enable =1 ;drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12_1.c +6
-23@@ -32,17 +32,6 @@#include "amdgpu_amdkfd.h" #include "kfd_device_queue_manager.h" -#define MQD_SIZE (2 * PAGE_SIZE) - -static uint64_t mqd_stride_v12_1 (struct mqd_manager *mm, -struct queue_properties *q) -{ -if (q->type ==KFD_QUEUE_TYPE_COMPUTE ) -return MQD_SIZE ; -else -return PAGE_SIZE ; -} -static inline struct v12_1_compute_mqd *get_mqd (void *mqd) {return (struct v12_1_compute_mqd *)mqd;@@ -148,21 +137,14 @@ static void set_priority(struct v12_1_compute_mqd *m, struct queue_properties *qstatic struct kfd_mem_obj *allocate_mqd (struct mqd_manager *mm,struct queue_properties *q) { +u32 mqd_size =AMDGPU_MQD_SIZE_ALIGN (mm->mqd_size);struct kfd_node *node = mm->dev;struct kfd_mem_obj *mqd_mem_obj; -unsigned int size; -/* -* Allocate two PAGE_SIZE memory for Compute MQD as MES writes to areas beyond -* struct MQD size. Size of the Compute MQD is 1 PAGE_SIZE. -* For SDMA MQD, we allocate 1 Page_size. -*/ if (q->type ==KFD_QUEUE_TYPE_COMPUTE ) - size =MQD_SIZE *NUM_XCC (node->xcc_mask); -else - size =PAGE_SIZE ; + mqd_size *=NUM_XCC (node->xcc_mask); -if (kfd_gtt_sa_allocate (node, size, &mqd_mem_obj)) +if (kfd_gtt_sa_allocate (node, mqd_size, &mqd_mem_obj))return NULL ;return mqd_mem_obj;@@ -174,11 +156,12 @@ static void init_mqd(struct mqd_manager *mm, void **mqd, {uint64_t addr;struct v12_1_compute_mqd *m; +u32 mqd_size =AMDGPU_MQD_SIZE_ALIGN (mm->mqd_size); m = (struct v12_1_compute_mqd *) mqd_mem_obj->cpu_ptr; addr = mqd_mem_obj->gpu_addr; -memset (m,0 ,MQD_SIZE ); +memset (m,0 , mqd_size); m->header =0xC0310800 ; m->compute_pipelinestat_enable =1 ;@@ -681,7 +664,7 @@ struct mqd_manager *mqd_manager_init_v12_1(enum KFD_MQD_TYPE type, mqd->is_occupied = kfd_is_occupied_cp; mqd->mqd_size =sizeof (struct v12_1_compute_mqd ); mqd->get_wave_state = get_wave_state_v12_1; - mqd->mqd_stride = mqd_stride_v12_1; + mqd->mqd_stride = kfd_mqd_stride;#if defined(CONFIG_DEBUG_FS ) mqd->debugfs_show_mqd = debugfs_show_mqd;#endif
RE: [PATCH 4/4] drm/amdkfd: Use AMDGPU_MQD_SIZE_ALIGN in gfx11+ kfd mqd manager
In reply to Lang Yu[AMD Official Use Only - AMD Internal Distribution Only]
Quoted 160 lines
-----Original Message-----
From: Yu, Lang <[email protected]>
Sent: Monday, January 26, 2026 5:25 AM
To: [email protected]
Cc: Zhang, Hawking <[email protected]>; Deucher, Alexander <[email protected]>; Joshi, Mukul <[email protected]>; Belanger, David <[email protected]>; Yu, Lang <[email protected]>
Subject: [PATCH 4/4] drm/amdkfd: UseAMDGPU_MQD_SIZE_ALIGNin gfx11+ kfd mqd managerMES is enabled by default from gfx11+.
Signed-off-by Lang Yu <[email protected]>--- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c | 3 ++ .../gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c | 22 +++----------- .../gpu/drm/amd/amdkfd/kfd_mqd_manager_v12.c | 10 +++---- .../drm/amd/amdkfd/kfd_mqd_manager_v12_1.c | 29 ++++--------------- 4 files changed, 17 insertions(+), 47 deletions(-)drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c +0
-0@@ -292,6 +292,9 @@ void kfd_get_hiq_xcc_mqd(struct kfd_node *dev,struct
kfd_mem_obj*mqd_mem_obj,uint64_tkfd_mqd_stride(structmqd_manager*mm,
structqueue_properties*q)
{drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c +3
-0+if (KFD_GC_VERSION (mm->dev) >=IP_VERSION (11 ,0 ,0 )) +return AMDGPU_MQD_SIZE_ALIGN (mm->mqd_size); +return mm->mqd_size; }drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c +2
-11@@ -102,20 +102,11 @@ static void set_priority(struct v11_compute_mqd *m, struct queue_properties *q) static struct kfd_mem_obj *allocate_mqd (struct mqd_manager *mm,struct queue_properties *q) { +u32 mqd_size =AMDGPU_MQD_SIZE_ALIGN (mm->mqd_size);struct kfd_node *node = mm->dev;struct kfd_mem_obj *mqd_mem_obj; -int size; - -/* -* MES write to areas beyond MQD size. So allocate -* 1 PAGE_SIZE memory for MQD is MES is enabled. -*/ -if (node->kfd->shared_resources.enable_mes) - size =PAGE_SIZE ; -else - size =sizeof (struct v11_compute_mqd ); -if (kfd_gtt_sa_allocate (node, size, &mqd_mem_obj)) +if (kfd_gtt_sa_allocate (node, mqd_size, &mqd_mem_obj))return NULL ;return mqd_mem_obj;drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c +2
-7@@ -127,18 +118,13 @@ static void init_mqd(struct mqd_manager *mm, void **mqd, { uint64_t addr;struct v11_compute_mqd *m; -int size; +u32 mqd_size =AMDGPU_MQD_SIZE_ALIGN (mm->mqd_size);uint32_t wa_mask = q->is_dbg_wa ?0xffff :0xffffffff ; m = (struct v11_compute_mqd *) mqd_mem_obj->cpu_ptr; addr = mqd_mem_obj->gpu_addr; -if (mm->dev->kfd->shared_resources.enable_mes) - size =PAGE_SIZE ; -else - size =sizeof (struct v11_compute_mqd ); - -memset (m,0 , size); +memset (m,0 , mqd_size); m->header =0xC0310800 ; m->compute_pipelinestat_enable =1 ;drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12.c +0
-0@@ -83,14 +83,11 @@ static void set_priority(struct v12_compute_mqd *m,struct
queue_properties*q) static structkfd_mem_obj*allocate_mqd(structmqd_manager*mm,
structqueue_properties*q)
{drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12.c +1
-2+u32 mqd_size =AMDGPU_MQD_SIZE_ALIGN (mm->mqd_size);struct kfd_node *node = mm->dev;struct kfd_mem_obj *mqd_mem_obj; -/* -* Allocate one PAGE_SIZE memory for MQD as MES writes to areas beyond
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12.c +3
-4-* struct MQD size. -*/ -if (kfd_gtt_sa_allocate (node,PAGE_SIZE , &mqd_mem_obj)) +if (kfd_gtt_sa_allocate (node, mqd_size, &mqd_mem_obj))return NULL ;return mqd_mem_obj;@@ -102,11 +99,12 @@ static void init_mqd(struct mqd_manager *mm, void **mqd, {uint64_t addr;struct v12_compute_mqd *m; +u32 mqd_size =AMDGPU_MQD_SIZE_ALIGN (mm->mqd_size); m = (struct v12_compute_mqd *) mqd_mem_obj->cpu_ptr; addr = mqd_mem_obj->gpu_addr; -memset (m,0 ,PAGE_SIZE ); +memset (m,0 , mqd_size); m->header =0xC0310800 ; m->compute_pipelinestat_enable =1 ;drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12_1.c +0
-11@@ -32,17 +32,6 @@#include "amdgpu_amdkfd.h" #include "kfd_device_queue_manager.h" -#define MQD_SIZE (2 * PAGE_SIZE) - -static uint64_t mqd_stride_v12_1 (struct mqd_manager *mm, -struct queue_properties *q) -{ -if (q->type ==KFD_QUEUE_TYPE_COMPUTE ) -return MQD_SIZE ; -else -return PAGE_SIZE ; -} -static inline struct v12_1_compute_mqd *get_mqd (void *mqd) { return (struct v12_1_compute_mqd *)mqd; @@ -148,21 +137,14 @@static void set_priority(struct
v12_1_compute_mqd*m, structqueue_properties*q static structkfd_mem_obj*allocate_mqd(structmqd_manager*mm,
structqueue_properties*q)
{drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12_1.c +1
-3+u32 mqd_size =AMDGPU_MQD_SIZE_ALIGN (mm->mqd_size);struct kfd_node *node = mm->dev;struct kfd_mem_obj *mqd_mem_obj; -unsigned int size; -/* -* Allocate two PAGE_SIZE memory for Compute MQD as MES writes to areas beyond
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12_1.c static void set_priority(struct v12_1_compute_mqd *m, struct queue_properties *q · line 145
Quoted 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12_1.c +1
-6-* struct MQD size. Size of the Compute MQD is 1 PAGE_SIZE. -* For SDMA MQD, we allocate 1 Page_size. -*/ if (q->type ==KFD_QUEUE_TYPE_COMPUTE ) - size =MQD_SIZE *NUM_XCC (node->xcc_mask); -else - size =PAGE_SIZE ; + mqd_size *=NUM_XCC (node->xcc_mask);
SDMA MQD allocation is not per XCC in the partition. This wastes a ton of memory for SDMA queues.
Regards,
Mukul
Quoted 30 lines
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12_1.c +4
-4-if (kfd_gtt_sa_allocate (node, size, &mqd_mem_obj)) +if (kfd_gtt_sa_allocate (node, mqd_size, &mqd_mem_obj))return NULL ;return mqd_mem_obj;@@ -174,11 +156,12 @@ static void init_mqd(struct mqd_manager *mm, void **mqd, { uint64_t addr;struct v12_1_compute_mqd *m; +u32 mqd_size =AMDGPU_MQD_SIZE_ALIGN (mm->mqd_size); m = (struct v12_1_compute_mqd *) mqd_mem_obj->cpu_ptr; addr = mqd_mem_obj->gpu_addr; -memset (m,0 ,MQD_SIZE ); +memset (m,0 , mqd_size); m->header =0xC0310800 ; m->compute_pipelinestat_enable =1 ;@@ -681,7 +664,7 @@ struct mqd_manager *mqd_manager_init_v12_1(enum KFD_MQD_TYPE type, mqd->is_occupied = kfd_is_occupied_cp; mqd->mqd_size =sizeof (struct v12_1_compute_mqd ); mqd->get_wave_state = get_wave_state_v12_1; - mqd->mqd_stride = mqd_stride_v12_1; + mqd->mqd_stride = kfd_mqd_stride;#if defined(CONFIG_DEBUG_FS ) mqd->debugfs_show_mqd = debugfs_show_mqd;#endif --2.34.1
RE: [PATCH 4/4] drm/amdkfd: Use AMDGPU_MQD_SIZE_ALIGN in gfx11+ kfd mqd manager
[Public]
Quoted 163 lines
-----Original Message-----
From: Joshi, Mukul <[email protected]>
Sent: Wednesday, January 28, 2026 12:23 AM
To: Yu, Lang <[email protected]>; [email protected]
Cc: Zhang, Hawking <[email protected]>; Deucher, Alexander <[email protected]>; Belanger, David <[email protected]>
Subject: RE: [PATCH 4/4] drm/amdkfd: UseAMDGPU_MQD_SIZE_ALIGNin gfx11+ kfd mqd manager[AMD Official Use Only - AMD Internal Distribution Only]
Quoted 152 lines
-----Original Message-----
From: Yu, Lang <[email protected]>
Sent: Monday, January 26, 2026 5:25 AM
To: [email protected]
Cc: Zhang, Hawking <[email protected]>; Deucher, Alexander <[email protected]>; Joshi, Mukul <[email protected]>; Belanger, David <[email protected]>; Yu, Lang <[email protected]>
Subject: [PATCH 4/4] drm/amdkfd: UseAMDGPU_MQD_SIZE_ALIGNin gfx11+ kfd mqd managerMES is enabled by default from gfx11+.
Signed-off-by Lang Yu <[email protected]>--- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c | 3 ++ .../gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c | 22 +++----------- .../gpu/drm/amd/amdkfd/kfd_mqd_manager_v12.c | 10 +++---- .../drm/amd/amdkfd/kfd_mqd_manager_v12_1.c | 29 ++++--------------- 4 files changed, 17 insertions(+), 47 deletions(-)drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c +0
-0@@ -292,6 +292,9 @@ void kfd_get_hiq_xcc_mqd(struct kfd_node *dev,struct
kfd_mem_obj*mqd_mem_obj,uint64_tkfd_mqd_stride(structmqd_manager*mm,struct queue_properties *q) {drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c +3
-0+if (KFD_GC_VERSION (mm->dev) >=IP_VERSION (11 ,0 ,0 )) +return AMDGPU_MQD_SIZE_ALIGN (mm->mqd_size); +return mm->mqd_size; }drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c +4
-18@@ -102,20 +102,11 @@ static void set_priority(struct v11_compute_mqd *m, struct queue_properties *q) static struct kfd_mem_obj *allocate_mqd (struct mqd_manager *mm,struct queue_properties *q) { +u32 mqd_size =AMDGPU_MQD_SIZE_ALIGN (mm->mqd_size);struct kfd_node *node = mm->dev;struct kfd_mem_obj *mqd_mem_obj; -int size; - -/* -* MES write to areas beyond MQD size. So allocate -* 1 PAGE_SIZE memory for MQD is MES is enabled. -*/ -if (node->kfd->shared_resources.enable_mes) - size =PAGE_SIZE ; -else - size =sizeof (struct v11_compute_mqd ); -if (kfd_gtt_sa_allocate (node, size, &mqd_mem_obj)) +if (kfd_gtt_sa_allocate (node, mqd_size, &mqd_mem_obj))return NULL ;return mqd_mem_obj;@@ -127,18 +118,13 @@ static void init_mqd(struct mqd_manager *mm, void **mqd, { uint64_t addr;struct v11_compute_mqd *m; -int size; +u32 mqd_size =AMDGPU_MQD_SIZE_ALIGN (mm->mqd_size);uint32_t wa_mask = q->is_dbg_wa ?0xffff :0xffffffff ; m = (struct v11_compute_mqd *) mqd_mem_obj->cpu_ptr; addr = mqd_mem_obj->gpu_addr; -if (mm->dev->kfd->shared_resources.enable_mes) - size =PAGE_SIZE ; -else - size =sizeof (struct v11_compute_mqd ); - -memset (m,0 , size); +memset (m,0 , mqd_size); m->header =0xC0310800 ; m->compute_pipelinestat_enable = 1; diff --git
a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12.cb/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12.cindexb7ac2dea8775..a51f217329db100644drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12.c +1
-2@@ -83,14 +83,11 @@ static void set_priority(struct v12_compute_mqd *m, struct queue_properties *q) static struct kfd_mem_obj *allocate_mqd (struct mqd_manager *mm,struct queue_properties *q) { +u32 mqd_size =AMDGPU_MQD_SIZE_ALIGN (mm->mqd_size);struct kfd_node *node = mm->dev;struct kfd_mem_obj *mqd_mem_obj; -/* -* Allocate one PAGE_SIZE memory for MQD as MES writes to areas beyond
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12.c +3
-4-* struct MQD size. -*/ -if (kfd_gtt_sa_allocate (node,PAGE_SIZE , &mqd_mem_obj)) +if (kfd_gtt_sa_allocate (node, mqd_size, &mqd_mem_obj))return NULL ;return mqd_mem_obj;@@ -102,11 +99,12 @@ static void init_mqd(struct mqd_manager *mm, void **mqd, {uint64_t addr;struct v12_compute_mqd *m; +u32 mqd_size =AMDGPU_MQD_SIZE_ALIGN (mm->mqd_size); m = (struct v12_compute_mqd *) mqd_mem_obj->cpu_ptr; addr = mqd_mem_obj->gpu_addr; -memset (m,0 ,PAGE_SIZE ); +memset (m,0 , mqd_size); m->header =0xC0310800 ; m->compute_pipelinestat_enable = 1; diff --git
a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12_1.cb/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12_1.cindex0b0d802a0917..d0776ba2cc99100644drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12_1.c +0
-11@@ -32,17 +32,6 @@#include "amdgpu_amdkfd.h" #include "kfd_device_queue_manager.h" -#define MQD_SIZE (2 * PAGE_SIZE) - -static uint64_t mqd_stride_v12_1 (struct mqd_manager *mm, -struct queue_properties *q) -{ -if (q->type ==KFD_QUEUE_TYPE_COMPUTE ) -return MQD_SIZE ; -else -return PAGE_SIZE ; -} -static inline struct v12_1_compute_mqd *get_mqd (void *mqd) { return (struct v12_1_compute_mqd *)mqd; @@ -148,21 +137,14 @@static void set_priority(struct
v12_1_compute_mqd*m, structqueue_properties*q static structkfd_mem_obj*allocate_mqd(structmqd_manager*mm,struct queue_properties *q) {drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12_1.c +1
-1+u32 mqd_size =AMDGPU_MQD_SIZE_ALIGN (mm->mqd_size);struct kfd_node *node = mm->dev;struct kfd_mem_obj *mqd_mem_obj; -unsigned int size;
Quoted 15 lines
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12_1.c static void set_priority(struct v12_1_compute_mqd *m, struct queue_properties *q · line 145
Quoted 12 lines
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12_1.c +0
-2-/* -* Allocate two PAGE_SIZE memory for Compute MQD as MES writes to areas beyond
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12_1.c +1
-6-* struct MQD size. Size of the Compute MQD is 1 PAGE_SIZE. -* For SDMA MQD, we allocate 1 Page_size. -*/ if (q->type ==KFD_QUEUE_TYPE_COMPUTE ) - size =MQD_SIZE *NUM_XCC (node->xcc_mask); -else - size =PAGE_SIZE ; + mqd_size *=NUM_XCC (node->xcc_mask);SDMA MQD allocation is not per XCC in the partition. This wastes a ton of memory for SDMA queues.
Look carefully. " if (q->type == KFD_QUEUE_TYPE_COMPUTE)" is still there. Only compute queue will be applied.
Regards,
Lang
Quoted 33 lines
Regards,
MukulQuoted 29 lines
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12_1.c +3
-2-if (kfd_gtt_sa_allocate (node, size, &mqd_mem_obj)) +if (kfd_gtt_sa_allocate (node, mqd_size, &mqd_mem_obj))return NULL ;return mqd_mem_obj;@@ -174,11 +156,12 @@ static void init_mqd(struct mqd_manager *mm, void **mqd, { uint64_t addr;struct v12_1_compute_mqd *m; +u32 mqd_size =AMDGPU_MQD_SIZE_ALIGN (mm->mqd_size); m = (struct v12_1_compute_mqd *) mqd_mem_obj->cpu_ptr; addr = mqd_mem_obj->gpu_addr; -memset (m,0 ,MQD_SIZE ); +memset (m,0 , mqd_size); m->header =0xC0310800 ; m->compute_pipelinestat_enable = 1; @@ -681,7 +664,7 @@ struct
mqd_manager*mqd_manager_init_v12_1(enumKFD_MQD_TYPEtype,mqd->is_occupied = kfd_is_occupied_cp; mqd->mqd_size = sizeof(struct v12_1_compute_mqd); mqd->get_wave_state = get_wave_state_v12_1;drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12_1.c +1
-2- mqd->mqd_stride = mqd_stride_v12_1; + mqd->mqd_stride = kfd_mqd_stride;#if defined(CONFIG_DEBUG_FS ) mqd->debugfs_show_mqd = debugfs_show_mqd;#endif --2.34.1
[AMD Official Use Only - AMD Internal Distribution Only]
Quoted 37 lines
-----Original Message-----
From: Yu, Lang <[email protected]>
Sent: Monday, January 26, 2026 5:25 AM
To: [email protected]
Cc: Zhang, Hawking <[email protected]>; Deucher, Alexander <[email protected]>; Joshi, Mukul <[email protected]>; Belanger, David <[email protected]>; Yu, Lang <[email protected]>
Subject: [PATCH 1/4] drm/amdgpu: Add a helper macro to align mqd sizeMES FW uses
addr(mqd_addr + sizeof(struct mqd) + 3*sizeof(uint32_t))as fence address and writes a 32 bit fence value to this address.Driver needs to allocate some extra memory(at least 4 DWs) in addition to sizeof(struct mqd) as mqd memory.
For gfx11/12, sizeof(struct mqd) <
PAGE_SIZE, allocate mqd memory withPAGE_SIZEaligned works. For gfx12.1, sizeof(struct mqd) ==PAGE_SIZE, it doesn't work.KFD mqd manager hardcodes mqd size to PAGE_SIZE/MQD_SIZE.
Let's use
AMDGPU_MQD_SIZE_ALIGNto avoid hardcoding in differnet place and across different IP version. It is used in two place.
1. mqd memory alloction
2. mqd stride initializationSigned-off-by Lang Yu <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 5 +++++ 1 file changed, 5 insertions(+)drivers/gpu/drm/amd/amdgpu/amdgpu.h +0
-0
drivers/gpu/drm/amd/amdgpu/amdgpu.h struct amdgpu_device { · line 1245
Quoted 7 lines
drivers/gpu/drm/amd/amdgpu/amdgpu.h +4
-0@@ -1239,6 +1239,11 @@ struct amdgpu_device {struct amdgpu_kfd_dev kfd; }; +/* +* MES will use memory beyond struct MQD size, 5 DWs currently */ +#define AMDGPU_MQD_SIZE_ALIGN (mqd_size) ALIGN(((mqd_size) + 20), +PAGE_SIZE)
Should we be using GPU Page size instead of CPU page size here? Maybe this should be AMDGPU_GPU_PAGE_SIZE instead of PAGE_SIZE or you can use AMDGPU_GPU_PAGE_ALIGN macro instead of ALIGN here.
Regards,
Mukul
Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/amdgpu.h +1
-1+static inline uint32_t amdgpu_ip_version (const struct amdgpu_device *adev,uint8_t ip,uint8_t inst) { --2.34.1
[Public]
Quoted 50 lines
-----Original Message-----
From: Joshi, Mukul <[email protected]>
Sent: Wednesday, January 28, 2026 12:26 AM
To: Yu, Lang <[email protected]>; [email protected]
Cc: Zhang, Hawking <[email protected]>; Deucher, Alexander <[email protected]>; Belanger, David <[email protected]>
Subject: RE: [PATCH 1/4] drm/amdgpu: Add a helper macro to align mqd size[AMD Official Use Only - AMD Internal Distribution Only]
Quoted 40 lines
-----Original Message-----
From: Yu, Lang <[email protected]>
Sent: Monday, January 26, 2026 5:25 AM
To: [email protected]
Cc: Zhang, Hawking <[email protected]>; Deucher, Alexander <[email protected]>; Joshi, Mukul <[email protected]>; Belanger, David <[email protected]>; Yu, Lang <[email protected]>
Subject: [PATCH 1/4] drm/amdgpu: Add a helper macro to align mqd sizeMES FW uses
addr(mqd_addr + sizeof(struct mqd) + 3*sizeof(uint32_t))as fence address and writes a 32 bit fence value to this address.Driver needs to allocate some extra memory(at least 4 DWs) in addition to sizeof(struct mqd) as mqd memory.
For gfx11/12, sizeof(struct mqd) <
PAGE_SIZE, allocate mqd memory withPAGE_SIZEaligned works. For gfx12.1, sizeof(struct mqd) ==PAGE_SIZE, it doesn't work.KFD mqd manager hardcodes mqd size to PAGE_SIZE/MQD_SIZE.
Let's use
AMDGPU_MQD_SIZE_ALIGNto avoid hardcoding in differnet place and across different IP version. It is used in two place.
1. mqd memory alloction
2. mqd stride initializationSigned-off-by Lang Yu <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 5 +++++ 1 file changed, 5 insertions(+)drivers/gpu/drm/amd/amdgpu/amdgpu.h +0
-0@@ -1239,6 +1239,11 @@ struct amdgpu_device {struct amdgpu_kfd_dev kfd; };
Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu.h struct amdgpu_device { · line 1245
Quoted 4 lines
drivers/gpu/drm/amd/amdgpu/amdgpu.h +4
-0+/* +* MES will use memory beyond struct MQD size, 5 DWs currently */ +#define AMDGPU_MQD_SIZE_ALIGN (mqd_size) ALIGN(((mqd_size) + 20), +PAGE_SIZE)Should we be using GPU Page size instead of CPU page size here? Maybe this should be
AMDGPU_GPU_PAGE_SIZEinstead ofPAGE_SIZEor you can useAMDGPU_GPU_PAGE_ALIGNmacro instead of ALIGN here.
Make sense. Thanks.
Regards,
Lang
Quoted 8 lines
Regards,
MukulQuoted 5 lines
drivers/gpu/drm/amd/amdgpu/amdgpu.h +1
-1+static inline uint32_t amdgpu_ip_version (const struct amdgpu_device *adev,uint8_t ip,uint8_t inst) { --2.34.1