[PATCH 2/2] drm/amdgpu: cleanup VCN IB generation

Andrey.Grodzovsky@xxxxxxx (Andrey Grodzovsky) · Fri, 9 Feb 2018 07:38:26 -0500

On 02/09/2018 07:31 AM, Christian KÃ¶nig wrote:
> Am 09.02.2018 um 13:20 schrieb Andrey Grodzovsky:
>>
>>
>> On 02/07/2018 02:48 PM, Christian KÃ¶nig wrote:
>>> Start to use amdgpu_bo_create_reserved.
>>>
>>> Signed-off-by: Christian KÃ¶nig <christian.koenig at amd.com>
>>> ---
>>> Â  drivers/gpu/drm/amd/amdgpu/amdgpu_vcn.c | 91 
>>> ++++++++-------------------------
>>> Â  1 file changed, 21 insertions(+), 70 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vcn.c 
>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_vcn.c
>>> index e86d0b2e9b7a..97bdde9a739e 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vcn.c
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vcn.c
>>> @@ -270,43 +270,28 @@ int amdgpu_vcn_dec_ring_test_ring(struct 
>>> amdgpu_ring *ring)
>>> Â Â Â Â Â  return r;
>>> Â  }
>>> Â  -static int amdgpu_vcn_dec_send_msg(struct amdgpu_ring *ring, 
>>> struct amdgpu_bo *bo,
>>> -Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  bool direct, struct dma_fence **fence)
>>> +static int amdgpu_vcn_dec_send_msg(struct amdgpu_ring *ring,
>>> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  struct amdgpu_bo *bo, bool direct,
>>> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  struct dma_fence **fence)
>>> Â  {
>>> -Â Â Â  struct ttm_operation_ctx ctx = { true, false };
>>> -Â Â Â  struct ttm_validate_buffer tv;
>>> -Â Â Â  struct ww_acquire_ctx ticket;
>>> -Â Â Â  struct list_head head;
>>> +Â Â Â  struct amdgpu_device *adev = ring->adev;
>>> +Â Â Â  struct dma_fence *f = NULL;
>>> Â Â Â Â Â  struct amdgpu_job *job;
>>> Â Â Â Â Â  struct amdgpu_ib *ib;
>>> -Â Â Â  struct dma_fence *f = NULL;
>>> -Â Â Â  struct amdgpu_device *adev = ring->adev;
>>> Â Â Â Â Â  uint64_t addr;
>>> Â Â Â Â Â  int i, r;
>>> Â  -Â Â Â  memset(&tv, 0, sizeof(tv));
>>> -Â Â Â  tv.bo = &bo->tbo;
>>> -
>>> -Â Â Â  INIT_LIST_HEAD(&head);
>>> -Â Â Â  list_add(&tv.head, &head);
>>> -
>>> -Â Â Â  r = ttm_eu_reserve_buffers(&ticket, &head, true, NULL);
>>> -Â Â Â  if (r)
>>> -Â Â Â Â Â Â Â  return r;
>>> -
>>> -Â Â Â  r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
>>> -Â Â Â  if (r)
>>> -Â Â Â Â Â Â Â  goto err;
>>> -
>>> Â Â Â Â Â  r = amdgpu_job_alloc_with_ib(adev, 64, &job);
>>> Â Â Â Â Â  if (r)
>>> Â Â Â Â Â Â Â Â Â  goto err;
>>> Â  Â Â Â Â Â  ib = &job->ibs[0];
>>> Â Â Â Â Â  addr = amdgpu_bo_gpu_offset(bo);
>>> -Â Â Â  ib->ptr[0] = PACKET0(SOC15_REG_OFFSET(UVD, 0, 
>>> mmUVD_GPCOM_VCPU_DATA0), 0);
>>> +Â Â Â  ib->ptr[0] = PACKET0(SOC15_REG_OFFSET(UVD, 0, 
>>> mmUVD_GPCOM_VCPU_DATA0),
>>> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  0);
>>
>> last argumentÂ  slipped to new line
>>
>>> Â Â Â Â Â  ib->ptr[1] = addr;
>>> -Â Â Â  ib->ptr[2] = PACKET0(SOC15_REG_OFFSET(UVD, 0, 
>>> mmUVD_GPCOM_VCPU_DATA1), 0);
>>> +Â Â Â  ib->ptr[2] = PACKET0(SOC15_REG_OFFSET(UVD, 0, 
>>> mmUVD_GPCOM_VCPU_DATA1),
>>> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  0);
>>
>> Same as above
>
> Mhm, actually correct according to the coding style (it makes the line 
> less than 80 chars long), but I agree that it looks ugly. I'm going to 
> remove it.
>
> Question is rather can somebody please test this on Raven?
>
> Thanks,
> Christian.

Today working from home so no raven at hand, if by Monday nobody tests 
it I can give it a try then with libdrm's VCN and UVD tests.

Thanks,
Andrey

>
>>
>> Thanks,
>> Andrey
>>
>>> Â Â Â Â Â  ib->ptr[3] = addr >> 32;
>>> Â Â Â Â Â  ib->ptr[4] = PACKET0(SOC15_REG_OFFSET(UVD, 0, 
>>> mmUVD_GPCOM_VCPU_CMD), 0);
>>> Â Â Â Â Â  ib->ptr[5] = 0;
>>> @@ -330,11 +315,12 @@ static int amdgpu_vcn_dec_send_msg(struct 
>>> amdgpu_ring *ring, struct amdgpu_bo *b
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â  goto err_free;
>>> Â Â Â Â Â  }
>>> Â  -Â Â Â  ttm_eu_fence_buffer_objects(&ticket, &head, f);
>>> +Â Â Â  amdgpu_bo_fence(bo, f, false);
>>> +Â Â Â  amdgpu_bo_unreserve(bo);
>>> +Â Â Â  amdgpu_bo_unref(&bo);
>>> Â  Â Â Â Â Â  if (fence)
>>> Â Â Â Â Â Â Â Â Â  *fence = dma_fence_get(f);
>>> -Â Â Â  amdgpu_bo_unref(&bo);
>>> Â Â Â Â Â  dma_fence_put(f);
>>> Â  Â Â Â Â Â  return 0;
>>> @@ -343,7 +329,8 @@ static int amdgpu_vcn_dec_send_msg(struct 
>>> amdgpu_ring *ring, struct amdgpu_bo *b
>>> Â Â Â Â Â  amdgpu_job_free(job);
>>> Â  Â  err:
>>> -Â Â Â  ttm_eu_backoff_reservation(&ticket, &head);
>>> +Â Â Â  amdgpu_bo_unreserve(bo);
>>> +Â Â Â  amdgpu_bo_unref(&bo);
>>> Â Â Â Â Â  return r;
>>> Â  }
>>> Â  @@ -351,31 +338,16 @@ static int 
>>> amdgpu_vcn_dec_get_create_msg(struct amdgpu_ring *ring, uint32_t hand
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  struct dma_fence **fence)
>>> Â  {
>>> Â Â Â Â Â  struct amdgpu_device *adev = ring->adev;
>>> -Â Â Â  struct amdgpu_bo *bo;
>>> +Â Â Â  struct amdgpu_bo *bo = NULL;
>>> Â Â Â Â Â  uint32_t *msg;
>>> Â Â Â Â Â  int r, i;
>>> Â  -Â Â Â  r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
>>> -Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  AMDGPU_GEM_DOMAIN_VRAM,
>>> -Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
>>> -Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS,
>>> -Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  NULL, NULL, &bo);
>>> +Â Â Â  r = amdgpu_bo_create_reserved(adev, 1024, PAGE_SIZE,
>>> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  AMDGPU_GEM_DOMAIN_VRAM,
>>> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  &bo, NULL, (void **)&msg);
>>> Â Â Â Â Â  if (r)
>>> Â Â Â Â Â Â Â Â Â  return r;
>>> Â  -Â Â Â  r = amdgpu_bo_reserve(bo, false);
>>> -Â Â Â  if (r) {
>>> -Â Â Â Â Â Â Â  amdgpu_bo_unref(&bo);
>>> -Â Â Â Â Â Â Â  return r;
>>> -Â Â Â  }
>>> -
>>> -Â Â Â  r = amdgpu_bo_kmap(bo, (void **)&msg);
>>> -Â Â Â  if (r) {
>>> -Â Â Â Â Â Â Â  amdgpu_bo_unreserve(bo);
>>> -Â Â Â Â Â Â Â  amdgpu_bo_unref(&bo);
>>> -Â Â Â Â Â Â Â  return r;
>>> -Â Â Â  }
>>> -
>>> Â Â Â Â Â  msg[0] = cpu_to_le32(0x00000028);
>>> Â Â Â Â Â  msg[1] = cpu_to_le32(0x00000038);
>>> Â Â Â Â Â  msg[2] = cpu_to_le32(0x00000001);
>>> @@ -393,9 +365,6 @@ static int amdgpu_vcn_dec_get_create_msg(struct 
>>> amdgpu_ring *ring, uint32_t hand
>>> Â Â Â Â Â  for (i = 14; i < 1024; ++i)
>>> Â Â Â Â Â Â Â Â Â  msg[i] = cpu_to_le32(0x0);
>>> Â  -Â Â Â  amdgpu_bo_kunmap(bo);
>>> -Â Â Â  amdgpu_bo_unreserve(bo);
>>> -
>>> Â Â Â Â Â  return amdgpu_vcn_dec_send_msg(ring, bo, true, fence);
>>> Â  }
>>> Â  @@ -407,27 +376,12 @@ static int 
>>> amdgpu_vcn_dec_get_destroy_msg(struct amdgpu_ring *ring, uint32_t han
>>> Â Â Â Â Â  uint32_t *msg;
>>> Â Â Â Â Â  int r, i;
>>> Â  -Â Â Â  r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
>>> -Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  AMDGPU_GEM_DOMAIN_VRAM,
>>> -Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
>>> -Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS,
>>> -Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  NULL, NULL, &bo);
>>> +Â Â Â  r = amdgpu_bo_create_reserved(adev, 1024, PAGE_SIZE,
>>> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  AMDGPU_GEM_DOMAIN_VRAM,
>>> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  &bo, NULL, (void **)&msg);
>>> Â Â Â Â Â  if (r)
>>> Â Â Â Â Â Â Â Â Â  return r;
>>> Â  -Â Â Â  r = amdgpu_bo_reserve(bo, false);
>>> -Â Â Â  if (r) {
>>> -Â Â Â Â Â Â Â  amdgpu_bo_unref(&bo);
>>> -Â Â Â Â Â Â Â  return r;
>>> -Â Â Â  }
>>> -
>>> -Â Â Â  r = amdgpu_bo_kmap(bo, (void **)&msg);
>>> -Â Â Â  if (r) {
>>> -Â Â Â Â Â Â Â  amdgpu_bo_unreserve(bo);
>>> -Â Â Â Â Â Â Â  amdgpu_bo_unref(&bo);
>>> -Â Â Â Â Â Â Â  return r;
>>> -Â Â Â  }
>>> -
>>> Â Â Â Â Â  msg[0] = cpu_to_le32(0x00000028);
>>> Â Â Â Â Â  msg[1] = cpu_to_le32(0x00000018);
>>> Â Â Â Â Â  msg[2] = cpu_to_le32(0x00000000);
>>> @@ -437,9 +391,6 @@ static int amdgpu_vcn_dec_get_destroy_msg(struct 
>>> amdgpu_ring *ring, uint32_t han
>>> Â Â Â Â Â  for (i = 6; i < 1024; ++i)
>>> Â Â Â Â Â Â Â Â Â  msg[i] = cpu_to_le32(0x0);
>>> Â  -Â Â Â  amdgpu_bo_kunmap(bo);
>>> -Â Â Â  amdgpu_bo_unreserve(bo);
>>> -
>>> Â Â Â Â Â  return amdgpu_vcn_dec_send_msg(ring, bo, direct, fence);
>>> Â  }
>>
>