From b8866c26ec072f1924f5cf601ebea33ca9823326 Mon Sep 17 00:00:00 2001 From: Andres Rodriguez Date: Fri, 28 Apr 2017 20:05:51 -0400 Subject: drm/amdgpu: implement ring set_priority for gfx_v8 compute v9 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Programming CP_HQD_QUEUE_PRIORITY enables a queue to take priority over other queues on the same pipe. Multiple queues on a pipe are timesliced so this gives us full precedence over other queues. Programming CP_HQD_PIPE_PRIORITY changes the SPI_ARB_PRIORITY of the wave as follows: 0x2: CS_H 0x1: CS_M 0x0: CS_L The SPI block will then dispatch work according to the policy set by SPI_ARB_PRIORITY. In the current policy CS_H is higher priority than gfx. In order to prevent getting stuck in loops of resources bouncing between GFX and high priority compute and introducing further latency, we statically reserve a portion of the pipe. v2: fix srbm_select to ring->queue and use ring->funcs->type v3: use AMD_SCHED_PRIORITY_* instead of AMDGPU_CTX_PRIORITY_* v4: switch int to enum amd_sched_priority v5: corresponding changes for srbm_lock v6: change CU reservation to PIPE_PERCENT allocation v7: use kiq instead of MMIO v8: back to MMIO, and make the implementation sleep safe. v9: corresponding changes for splitting HIGH into _HW/_SW Acked-by: Christian König Signed-off-by: Andres Rodriguez Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c | 99 +++++++++++++++++++++++++++++++++++ 1 file changed, 99 insertions(+) (limited to 'drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c') diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c index 147e92b3a959..51896b7353b6 100644 --- a/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c @@ -6394,6 +6394,104 @@ static void gfx_v8_0_ring_set_wptr_compute(struct amdgpu_ring *ring) WDOORBELL32(ring->doorbell_index, lower_32_bits(ring->wptr)); } +static void gfx_v8_0_ring_set_pipe_percent(struct amdgpu_ring *ring, + bool acquire) +{ + struct amdgpu_device *adev = ring->adev; + int pipe_num, tmp, reg; + int pipe_percent = acquire ? SPI_WCL_PIPE_PERCENT_GFX__VALUE_MASK : 0x1; + + pipe_num = ring->me * adev->gfx.mec.num_pipe_per_mec + ring->pipe; + + /* first me only has 2 entries, GFX and HP3D */ + if (ring->me > 0) + pipe_num -= 2; + + reg = mmSPI_WCL_PIPE_PERCENT_GFX + pipe_num; + tmp = RREG32(reg); + tmp = REG_SET_FIELD(tmp, SPI_WCL_PIPE_PERCENT_GFX, VALUE, pipe_percent); + WREG32(reg, tmp); +} + +static void gfx_v8_0_pipe_reserve_resources(struct amdgpu_device *adev, + struct amdgpu_ring *ring, + bool acquire) +{ + int i, pipe; + bool reserve; + struct amdgpu_ring *iring; + + mutex_lock(&adev->gfx.pipe_reserve_mutex); + pipe = amdgpu_gfx_queue_to_bit(adev, ring->me, ring->pipe, 0); + if (acquire) + set_bit(pipe, adev->gfx.pipe_reserve_bitmap); + else + clear_bit(pipe, adev->gfx.pipe_reserve_bitmap); + + if (!bitmap_weight(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES)) { + /* Clear all reservations - everyone reacquires all resources */ + for (i = 0; i < adev->gfx.num_gfx_rings; ++i) + gfx_v8_0_ring_set_pipe_percent(&adev->gfx.gfx_ring[i], + true); + + for (i = 0; i < adev->gfx.num_compute_rings; ++i) + gfx_v8_0_ring_set_pipe_percent(&adev->gfx.compute_ring[i], + true); + } else { + /* Lower all pipes without a current reservation */ + for (i = 0; i < adev->gfx.num_gfx_rings; ++i) { + iring = &adev->gfx.gfx_ring[i]; + pipe = amdgpu_gfx_queue_to_bit(adev, + iring->me, + iring->pipe, + 0); + reserve = test_bit(pipe, adev->gfx.pipe_reserve_bitmap); + gfx_v8_0_ring_set_pipe_percent(iring, reserve); + } + + for (i = 0; i < adev->gfx.num_compute_rings; ++i) { + iring = &adev->gfx.compute_ring[i]; + pipe = amdgpu_gfx_queue_to_bit(adev, + iring->me, + iring->pipe, + 0); + reserve = test_bit(pipe, adev->gfx.pipe_reserve_bitmap); + gfx_v8_0_ring_set_pipe_percent(iring, reserve); + } + } + + mutex_unlock(&adev->gfx.pipe_reserve_mutex); +} + +static void gfx_v8_0_hqd_set_priority(struct amdgpu_device *adev, + struct amdgpu_ring *ring, + bool acquire) +{ + uint32_t pipe_priority = acquire ? 0x2 : 0x0; + uint32_t queue_priority = acquire ? 0xf : 0x0; + + mutex_lock(&adev->srbm_mutex); + vi_srbm_select(adev, ring->me, ring->pipe, ring->queue, 0); + + WREG32(mmCP_HQD_PIPE_PRIORITY, pipe_priority); + WREG32(mmCP_HQD_QUEUE_PRIORITY, queue_priority); + + vi_srbm_select(adev, 0, 0, 0, 0); + mutex_unlock(&adev->srbm_mutex); +} +static void gfx_v8_0_ring_set_priority_compute(struct amdgpu_ring *ring, + enum amd_sched_priority priority) +{ + struct amdgpu_device *adev = ring->adev; + bool acquire = priority == AMD_SCHED_PRIORITY_HIGH_HW; + + if (ring->funcs->type != AMDGPU_RING_TYPE_COMPUTE) + return; + + gfx_v8_0_hqd_set_priority(adev, ring, acquire); + gfx_v8_0_pipe_reserve_resources(adev, ring, acquire); +} + static void gfx_v8_0_ring_emit_fence_compute(struct amdgpu_ring *ring, u64 addr, u64 seq, unsigned flags) @@ -6839,6 +6937,7 @@ static const struct amdgpu_ring_funcs gfx_v8_0_ring_funcs_compute = { .test_ib = gfx_v8_0_ring_test_ib, .insert_nop = amdgpu_ring_insert_nop, .pad_ib = amdgpu_ring_generic_pad_ib, + .set_priority = gfx_v8_0_ring_set_priority_compute, }; static const struct amdgpu_ring_funcs gfx_v8_0_ring_funcs_kiq = { -- cgit From c1b24a1405ed8720f5837fc77fbc52fd008cbb42 Mon Sep 17 00:00:00 2001 From: Jérémy Lefaure Date: Sun, 15 Oct 2017 22:29:23 -0400 Subject: drm/amdgpu: use ARRAY_SIZE MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Using the ARRAY_SIZE macro improves the readability of the code. Found with Coccinelle with the following semantic patch: @r depends on (org || report)@ type T; T[] E; position p; @@ ( (sizeof(E)@p /sizeof(*E)) | (sizeof(E)@p /sizeof(E[...])) | (sizeof(E)@p /sizeof(T)) ) Reviewed-by: Christian König Reviewed-by: Thierry Reding Signed-off-by: Jérémy Lefaure Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c | 9 +++++---- drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c | 9 +++++---- 2 files changed, 10 insertions(+), 8 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c') diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c index 51896b7353b6..b8002ac3e536 100644 --- a/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c @@ -20,6 +20,7 @@ * OTHER DEALINGS IN THE SOFTWARE. * */ +#include #include #include #include "amdgpu.h" @@ -3952,10 +3953,10 @@ static int gfx_v8_0_init_save_restore_list(struct amdgpu_device *adev) adev->gfx.rlc.reg_list_format_size_bytes >> 2, unique_indices, &indices_count, - sizeof(unique_indices) / sizeof(int), + ARRAY_SIZE(unique_indices), indirect_start_offsets, &offset_count, - sizeof(indirect_start_offsets)/sizeof(int)); + ARRAY_SIZE(indirect_start_offsets)); /* save and restore list */ WREG32_FIELD(RLC_SRM_CNTL, AUTO_INCR_ADDR, 1); @@ -3977,14 +3978,14 @@ static int gfx_v8_0_init_save_restore_list(struct amdgpu_device *adev) /* starting offsets starts */ WREG32(mmRLC_GPM_SCRATCH_ADDR, adev->gfx.rlc.starting_offsets_start); - for (i = 0; i < sizeof(indirect_start_offsets)/sizeof(int); i++) + for (i = 0; i < ARRAY_SIZE(indirect_start_offsets); i++) WREG32(mmRLC_GPM_SCRATCH_DATA, indirect_start_offsets[i]); /* unique indices */ temp = mmRLC_SRM_INDEX_CNTL_ADDR_0; data = mmRLC_SRM_INDEX_CNTL_DATA_0; - for (i = 0; i < sizeof(unique_indices) / sizeof(int); i++) { + for (i = 0; i < ARRAY_SIZE(unique_indices); i++) { if (unique_indices[i] != 0) { WREG32(temp + i, unique_indices[i] & 0x3FFFF); WREG32(data + i, unique_indices[i] >> 20); diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c index 99a5b3b92e8e..7f15bb2c5233 100644 --- a/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c @@ -20,6 +20,7 @@ * OTHER DEALINGS IN THE SOFTWARE. * */ +#include #include #include #include "amdgpu.h" @@ -1730,10 +1731,10 @@ static int gfx_v9_0_init_rlc_save_restore_list(struct amdgpu_device *adev) adev->gfx.rlc.reg_list_format_size_bytes >> 2, unique_indirect_regs, &unique_indirect_reg_count, - sizeof(unique_indirect_regs)/sizeof(int), + ARRAY_SIZE(unique_indirect_regs), indirect_start_offsets, &indirect_start_offsets_count, - sizeof(indirect_start_offsets)/sizeof(int)); + ARRAY_SIZE(indirect_start_offsets)); /* enable auto inc in case it is disabled */ tmp = RREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_SRM_CNTL)); @@ -1770,12 +1771,12 @@ static int gfx_v9_0_init_rlc_save_restore_list(struct amdgpu_device *adev) /* write the starting offsets to RLC scratch ram */ WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_GPM_SCRATCH_ADDR), adev->gfx.rlc.starting_offsets_start); - for (i = 0; i < sizeof(indirect_start_offsets)/sizeof(int); i++) + for (i = 0; i < ARRAY_SIZE(indirect_start_offsets); i++) WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_GPM_SCRATCH_DATA), indirect_start_offsets[i]); /* load unique indirect regs*/ - for (i = 0; i < sizeof(unique_indirect_regs)/sizeof(int); i++) { + for (i = 0; i < ARRAY_SIZE(unique_indirect_regs); i++) { WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_SRM_INDEX_CNTL_ADDR_0) + i, unique_indirect_regs[i] & 0x3FFFF); WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_SRM_INDEX_CNTL_DATA_0) + i, -- cgit