AlexIndustrial/mesa
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

panvk: Select next iter scoreboard early

Browse Source 
This closely match what NEXT_SB_ENTRY gives us on v11+ and should also
reduce the number of instruction by low margin.

Signed-off-by: Mary Guillemard <mary.guillemard@collabora.com>
Reviewed-by: Lars-Ivar Hesselberg Simonsen <lars-ivar.simonsen@arm.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35307>
This commit is contained in:
Mary Guillemard
2025-06-03 10:09:38 +02:00
committed by Marge Bot
parent e23612d0a9
commit 9d1d6c75fe
6 changed files with 36 additions and 42 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/panfrost/vulkan/csf/panvk_cmd_buffer.h
+3 -11
View File
@@ -386,15 +386,6 @@ struct panvk_cmd_buffer {
VK_DEFINE_HANDLE_CASTS(panvk_cmd_buffer, vk.base, VkCommandBuffer,
VK_OBJECT_TYPE_COMMAND_BUFFER)
static inline uint32_t
next_iter_sb(struct panvk_cmd_buffer *cmdbuf, uint32_t iter_sb_idx)
{
struct panvk_device *dev = to_panvk_device(cmdbuf->vk.base.device);
return iter_sb_idx + 1 < dev->csf.sb.iter_count ? SB_ITER(iter_sb_idx + 1)
: SB_ITER(0);
}
static bool
inherits_render_ctx(struct panvk_cmd_buffer *cmdbuf)
{
@@ -431,8 +422,9 @@ extern const struct vk_command_buffer_ops panvk_per_arch(cmd_buffer_ops);
void panvk_per_arch(cmd_flush_draws)(struct panvk_cmd_buffer *cmdbuf);
void panvk_per_arch(cs_pick_iter_sb)(struct panvk_cmd_buffer *cmdbuf,
enum panvk_subqueue_id subqueue);
void panvk_per_arch(cs_next_iter_sb)(struct panvk_cmd_buffer *cmdbuf,
enum panvk_subqueue_id subqueue,
struct cs_index scratch_regs);
void panvk_per_arch(get_cs_deps)(struct panvk_cmd_buffer *cmdbuf,
const VkDependencyInfo *in,
src/panfrost/vulkan/csf/panvk_vX_cmd_buffer.c
+17 -4
View File
@@ -618,16 +618,25 @@ panvk_per_arch(CmdPipelineBarrier2)(VkCommandBuffer commandBuffer,
}
void
panvk_per_arch(cs_pick_iter_sb)(struct panvk_cmd_buffer *cmdbuf,
enum panvk_subqueue_id subqueue)
panvk_per_arch(cs_next_iter_sb)(struct panvk_cmd_buffer *cmdbuf,
enum panvk_subqueue_id subqueue,
struct cs_index scratch_regs)
{
struct panvk_device *dev = to_panvk_device(cmdbuf->vk.base.device);
struct cs_builder *b = panvk_get_cs_builder(cmdbuf, subqueue);
struct cs_index iter_sb = cs_scratch_reg32(b, 0);
struct cs_index cmp_scratch = cs_scratch_reg32(b, 1);
struct cs_index iter_sb = cs_extract32(b, scratch_regs, 0);
struct cs_index cmp_scratch = cs_extract32(b, scratch_regs, 1);
cs_load32_to(b, iter_sb, cs_subqueue_ctx_reg(b),
offsetof(struct panvk_cs_subqueue_context, iter_sb));
/* Select next scoreboard entry and wrap around if we get past the limit */
cs_add32(b, iter_sb, iter_sb, 1);
cs_add32(b, cmp_scratch, iter_sb, -SB_ITER(dev->csf.sb.iter_count));
cs_if(b, MALI_CS_CONDITION_GEQUAL, cmp_scratch) {
cs_move32_to(b, iter_sb, SB_ITER(0));
}
cs_match(b, iter_sb, cmp_scratch) {
#define CASE(x) \
cs_case(b, SB_ITER(x)) { \
@@ -642,6 +651,10 @@ panvk_per_arch(cs_pick_iter_sb)(struct panvk_cmd_buffer *cmdbuf,
CASE(4)
#undef CASE
}
cs_store32(b, iter_sb, cs_subqueue_ctx_reg(b),
offsetof(struct panvk_cs_subqueue_context, iter_sb));
cs_flush_stores(b);
}
static struct cs_buffer
src/panfrost/vulkan/csf/panvk_vX_cmd_dispatch.c
+3 -6
View File
@@ -272,7 +272,9 @@ cmd_dispatch(struct panvk_cmd_buffer *cmdbuf, struct panvk_dispatch_info *info)
}
}
panvk_per_arch(cs_pick_iter_sb)(cmdbuf, PANVK_SUBQUEUE_COMPUTE);
struct cs_index next_iter_sb_scratch = cs_scratch_reg_tuple(b, 0, 2);
panvk_per_arch(cs_next_iter_sb)(cmdbuf, PANVK_SUBQUEUE_COMPUTE,
next_iter_sb_scratch);
if (indirect) {
/* Use run_compute with a set task axis instead of run_compute_indirect as
@@ -314,7 +316,6 @@ cmd_dispatch(struct panvk_cmd_buffer *cmdbuf, struct panvk_dispatch_info *info)
cs_case(b, SB_ITER(x)) { \
cs_sync64_add(b, true, MALI_CS_SYNC_SCOPE_CSG, add_val, sync_addr, \
cs_defer(SB_WAIT_ITER(x), SB_ID(DEFERRED_SYNC))); \
cs_move32_to(b, iter_sb, next_iter_sb(cmdbuf, x)); \
}
CASE(0)
@@ -325,10 +326,6 @@ cmd_dispatch(struct panvk_cmd_buffer *cmdbuf, struct panvk_dispatch_info *info)
#undef CASE
}
cs_store32(b, iter_sb, cs_subqueue_ctx_reg(b),
offsetof(struct panvk_cs_subqueue_context, iter_sb));
cs_flush_stores(b);
++cmdbuf->state.cs[PANVK_SUBQUEUE_COMPUTE].relative_sync_point;
clear_dirty_after_dispatch(cmdbuf);
}
src/panfrost/vulkan/csf/panvk_vX_cmd_draw.c
+6 -14
View File
@@ -1169,8 +1169,9 @@ get_tiler_desc(struct panvk_cmd_buffer *cmdbuf)
/* Flush all stores to tiler_ctx_addr. */
cs_flush_stores(b);
/* Then we change the scoreboard slot used for iterators. */
panvk_per_arch(cs_pick_iter_sb)(cmdbuf, PANVK_SUBQUEUE_VERTEX_TILER);
struct cs_index next_iter_sb_scratch = cs_scratch_reg_tuple(b, 0, 2);
panvk_per_arch(cs_next_iter_sb)(cmdbuf, PANVK_SUBQUEUE_VERTEX_TILER,
next_iter_sb_scratch);
cs_heap_operation(b, MALI_CS_HEAP_OPERATION_VERTEX_TILER_STARTED, cs_now());
return VK_SUCCESS;
@@ -2875,7 +2876,6 @@ flush_tiling(struct panvk_cmd_buffer *cmdbuf)
cs_defer(SB_WAIT_ITER(x), SB_ID(DEFERRED_SYNC))); \
cs_sync64_add(b, true, MALI_CS_SYNC_SCOPE_CSG, add_val, sync_addr, \
cs_defer(SB_WAIT_ITER(x), SB_ID(DEFERRED_SYNC))); \
cs_move32_to(b, iter_sb, next_iter_sb(cmdbuf, x)); \
}
CASE(0)
@@ -2886,10 +2886,6 @@ flush_tiling(struct panvk_cmd_buffer *cmdbuf)
#undef CASE
}
cs_store32(b, iter_sb, cs_subqueue_ctx_reg(b),
offsetof(struct panvk_cs_subqueue_context, iter_sb));
cs_flush_stores(b);
/* Update the vertex seqno. */
++cmdbuf->state.cs[PANVK_SUBQUEUE_VERTEX_TILER].relative_sync_point;
}
@@ -2978,8 +2974,9 @@ issue_fragment_jobs(struct panvk_cmd_buffer *cmdbuf)
struct cs_builder *b = panvk_get_cs_builder(cmdbuf, PANVK_SUBQUEUE_FRAGMENT);
bool has_oq_chain = cmdbuf->state.gfx.render.oq.chain != 0;
/* Reserve a scoreboard for the fragment job. */
panvk_per_arch(cs_pick_iter_sb)(cmdbuf, PANVK_SUBQUEUE_FRAGMENT);
struct cs_index next_iter_sb_scratch = cs_scratch_reg_tuple(b, 0, 2);
panvk_per_arch(cs_next_iter_sb)(cmdbuf, PANVK_SUBQUEUE_FRAGMENT,
next_iter_sb_scratch);
/* Now initialize the fragment bits. */
cs_update_frag_ctx(b) {
@@ -3178,7 +3175,6 @@ issue_fragment_jobs(struct panvk_cmd_buffer *cmdbuf)
} \
cs_sync64_add(b, true, MALI_CS_SYNC_SCOPE_CSG, add_val, sync_addr, \
async); \
cs_move32_to(b, iter_sb, next_iter_sb(cmdbuf, x)); \
}
CASE(0)
@@ -3189,10 +3185,6 @@ issue_fragment_jobs(struct panvk_cmd_buffer *cmdbuf)
#undef CASE
}
cs_store32(b, iter_sb, cs_subqueue_ctx_reg(b),
offsetof(struct panvk_cs_subqueue_context, iter_sb));
cs_flush_stores(b);
/* Update the ring buffer position. */
if (free_render_descs) {
cs_render_desc_ringbuf_move_ptr(b, calc_render_descs_size(cmdbuf),
src/panfrost/vulkan/csf/panvk_vX_cmd_precomp.c
+3 -6
View File
@@ -104,7 +104,9 @@ panvk_per_arch(dispatch_precomp)(struct panvk_precomp_ctx *ctx,
cs_move32_to(b, cs_sr_reg32(b, COMPUTE, JOB_SIZE_Z), grid.count[2]);
}
panvk_per_arch(cs_pick_iter_sb)(cmdbuf, PANVK_SUBQUEUE_COMPUTE);
struct cs_index next_iter_sb_scratch = cs_scratch_reg_tuple(b, 0, 2);
panvk_per_arch(cs_next_iter_sb)(cmdbuf, PANVK_SUBQUEUE_COMPUTE,
next_iter_sb_scratch);
unsigned task_axis = MALI_TASK_AXIS_X;
unsigned task_increment = 0;
@@ -132,7 +134,6 @@ panvk_per_arch(dispatch_precomp)(struct panvk_precomp_ctx *ctx,
cs_case(b, SB_ITER(x)) { \
cs_sync64_add(b, true, MALI_CS_SYNC_SCOPE_CSG, add_val, sync_addr, \
cs_defer(SB_WAIT_ITER(x), SB_ID(DEFERRED_SYNC))); \
cs_move32_to(b, iter_sb, next_iter_sb(cmdbuf, x)); \
}
CASE(0)
@@ -143,10 +144,6 @@ panvk_per_arch(dispatch_precomp)(struct panvk_precomp_ctx *ctx,
#undef CASE
}
cs_store32(b, iter_sb, cs_subqueue_ctx_reg(b),
offsetof(struct panvk_cs_subqueue_context, iter_sb));
cs_flush_stores(b);
++cmdbuf->state.cs[PANVK_SUBQUEUE_COMPUTE].relative_sync_point;
/* XXX: clobbers the registers instead to avoid recreating them when calling
src/panfrost/vulkan/csf/panvk_vX_queue.c
+4 -1
View File
@@ -395,7 +395,10 @@ init_subqueue(struct panvk_queue *queue, enum panvk_subqueue_id subqueue)
.syncobjs = panvk_priv_mem_dev_addr(queue->syncobjs),
.debug.syncobjs = panvk_priv_mem_dev_addr(queue->debug_syncobjs),
.debug.tracebuf.cs = subq->tracebuf.addr.dev,
.iter_sb = SB_ITER(0),
/* Iterator scoreboard will be picked in CS and wrap back to SB_ITER(0) on
first RUN_* so we ensure an invalid value here that is handled by our
partial modulo implementation */
.iter_sb = SB_ITER(dev->csf.sb.iter_count),
.reg_dump_addr = panvk_priv_mem_dev_addr(subq->regs_save),
};