panvk/csf: add more utrace tracepoints

Add all newly defined tracepoints, including meta, render,
dispatch/dispatch_indirect, barrier, and sync_wait.

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Christoph Pillmayer <christoph.pillmayer@arm.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32693>

src/panfrost/vulkan/csf/panvk_vX_cmd_buffer.c

@@ -585,6 +585,7 @@ panvk_per_arch(emit_barrier)(struct panvk_cmd_buffer *cmdbuf,
                              struct panvk_cs_deps deps)
 {
    uint32_t wait_subqueue_mask = 0;
+   uint32_t utrace_subqueue_mask = 0;
    for (uint32_t i = 0; i < PANVK_SUBQUEUE_COUNT; i++) {
       /* no need to perform both types of waits on the same subqueue */
       if (deps.src[i].wait_sb_mask)
@@ -592,10 +593,16 @@ panvk_per_arch(emit_barrier)(struct panvk_cmd_buffer *cmdbuf,
       assert(!(deps.dst[i].wait_subqueue_mask & BITFIELD_BIT(i)));
 
       wait_subqueue_mask |= deps.dst[i].wait_subqueue_mask;
+
+      if (deps.src[i].wait_sb_mask || deps.dst[i].wait_subqueue_mask ||
+          !panvk_cache_flush_is_nop(&deps.src[i].cache_flush))
+         utrace_subqueue_mask |= BITFIELD_BIT(i);
    }
 
-   for (uint32_t i = 0; i < PANVK_SUBQUEUE_COUNT; i++) {
+   u_foreach_bit(i, utrace_subqueue_mask)
+      trace_begin_barrier(&cmdbuf->utrace.uts[i], cmdbuf);
 
+   for (uint32_t i = 0; i < PANVK_SUBQUEUE_COUNT; i++) {
       struct cs_builder *b = panvk_get_cs_builder(cmdbuf, i);
       struct panvk_cs_state *cs_state = &cmdbuf->state.cs[i];
 
@@ -643,6 +650,13 @@ panvk_per_arch(emit_barrier)(struct panvk_cmd_buffer *cmdbuf,
          emit_barrier_insert_waits(b, cmdbuf, &deps, i, tmp_regs);
       }
    }
+
+   u_foreach_bit(i, utrace_subqueue_mask) {
+      trace_end_barrier(
+         &cmdbuf->utrace.uts[i], cmdbuf, deps.src[i].wait_sb_mask,
+         deps.dst[i].wait_subqueue_mask, deps.src[i].cache_flush.l2,
+         deps.src[i].cache_flush.lsc, deps.src[i].cache_flush.others);
+   }
 }
 
 VKAPI_ATTR void VKAPI_CALL

src/panfrost/vulkan/csf/panvk_vX_cmd_dispatch.c

@@ -23,6 +23,7 @@
 #include "panvk_macros.h"
 #include "panvk_meta.h"
 #include "panvk_physical_device.h"
+#include "panvk_tracepoints.h"
 
 #include "pan_desc.h"
 #include "pan_encoder.h"
@@ -351,11 +352,20 @@ panvk_per_arch(CmdDispatchBase)(VkCommandBuffer commandBuffer,
                                 uint32_t groupCountY, uint32_t groupCountZ)
 {
    VK_FROM_HANDLE(panvk_cmd_buffer, cmdbuf, commandBuffer);
+   const struct panvk_shader *shader = cmdbuf->state.compute.shader;
    struct panvk_dispatch_info info = {
       .wg_base = {baseGroupX, baseGroupY, baseGroupZ},
       .direct.wg_count = {groupCountX, groupCountY, groupCountZ},
    };
+
+   trace_begin_dispatch(&cmdbuf->utrace.uts[PANVK_SUBQUEUE_COMPUTE], cmdbuf);
+
    cmd_dispatch(cmdbuf, &info);
+
+   trace_end_dispatch(&cmdbuf->utrace.uts[PANVK_SUBQUEUE_COMPUTE], cmdbuf,
+                      baseGroupX, baseGroupY, baseGroupZ, groupCountX,
+                      groupCountY, groupCountZ, shader->cs.local_size.x,
+                      shader->cs.local_size.y, shader->cs.local_size.z);
 }
 
 VKAPI_ATTR void VKAPI_CALL
@@ -368,5 +378,13 @@ panvk_per_arch(CmdDispatchIndirect)(VkCommandBuffer commandBuffer,
    struct panvk_dispatch_info info = {
       .indirect.buffer_dev_addr = buffer_gpu,
    };
+
+   trace_begin_dispatch_indirect(&cmdbuf->utrace.uts[PANVK_SUBQUEUE_COMPUTE],
+                                 cmdbuf);
+
    cmd_dispatch(cmdbuf, &info);
+
+   trace_end_dispatch_indirect(&cmdbuf->utrace.uts[PANVK_SUBQUEUE_COMPUTE],
+                               cmdbuf,
+                               (struct u_trace_address){.offset = buffer_gpu});
 }

src/panfrost/vulkan/csf/panvk_vX_cmd_draw.c

@@ -31,6 +31,7 @@
 #include "panvk_priv_bo.h"
 #include "panvk_query_pool.h"
 #include "panvk_shader.h"
+#include "panvk_tracepoints.h"
 
 #include "pan_desc.h"
 #include "pan_earlyzs.h"
@@ -2829,6 +2830,9 @@ panvk_per_arch(CmdBeginRendering)(VkCommandBuffer commandBuffer,
    /* If we're not resuming, the FBD should be NULL. */
    assert(!state->render.fbds.gpu || resuming);
 
+   trace_begin_render(&cmdbuf->utrace.uts[PANVK_SUBQUEUE_VERTEX_TILER], cmdbuf);
+   trace_begin_render(&cmdbuf->utrace.uts[PANVK_SUBQUEUE_FRAGMENT], cmdbuf);
+
    if (!resuming)
       panvk_per_arch(cmd_preload_render_area_border)(cmdbuf, pRenderingInfo);
 }
@@ -3031,7 +3035,9 @@ issue_fragment_jobs(struct panvk_cmd_buffer *cmdbuf)
                             length_reg);
 
    /* Wait for the tiling to be done before submitting the fragment job. */
+   trace_begin_sync_wait(&cmdbuf->utrace.uts[PANVK_SUBQUEUE_FRAGMENT], cmdbuf);
    wait_finish_tiling(cmdbuf);
+   trace_end_sync_wait(&cmdbuf->utrace.uts[PANVK_SUBQUEUE_FRAGMENT], cmdbuf);
 
    /* Disable the oom handler once the vertex/tiler work has finished.
     * We need to disable the handler at this point as the vertex/tiler subqueue
@@ -3403,4 +3409,11 @@ panvk_per_arch(CmdEndRendering)(VkCommandBuffer commandBuffer)
    /* If we're not suspending, we need to resolve attachments. */
    if (!suspending)
       panvk_per_arch(cmd_resolve_attachments)(cmdbuf);
+
+   trace_end_render(&cmdbuf->utrace.uts[PANVK_SUBQUEUE_VERTEX_TILER], cmdbuf,
+                    cmdbuf->state.gfx.render.flags,
+                    &cmdbuf->state.gfx.render.fb.info);
+   trace_end_render(&cmdbuf->utrace.uts[PANVK_SUBQUEUE_FRAGMENT], cmdbuf,
+                    cmdbuf->state.gfx.render.flags,
+                    &cmdbuf->state.gfx.render.fb.info);
 }

src/panfrost/vulkan/panvk_vX_cmd_meta.c

@@ -6,6 +6,7 @@
 
 #include "panvk_cmd_meta.h"
 #include "panvk_entrypoints.h"
+#include "panvk_tracepoints.h"
 
 static bool
 copy_to_image_use_gfx_pipeline(struct panvk_device *dev,
@@ -45,6 +46,10 @@ panvk_per_arch(cmd_meta_compute_start)(
    save_ctx->push_constants = cmdbuf->state.push_constants;
    save_ctx->cs.shader = cmdbuf->state.compute.shader;
    save_ctx->cs.desc = cmdbuf->state.compute.cs.desc;
+
+#if PAN_ARCH >= 10
+   trace_begin_meta(&cmdbuf->utrace.uts[PANVK_SUBQUEUE_COMPUTE], cmdbuf);
+#endif
 }
 
 void
@@ -55,6 +60,10 @@ panvk_per_arch(cmd_meta_compute_end)(
    struct panvk_descriptor_set *push_set0 =
       cmdbuf->state.compute.desc_state.push_sets[0];
 
+#if PAN_ARCH >= 10
+   trace_end_meta(&cmdbuf->utrace.uts[PANVK_SUBQUEUE_COMPUTE], cmdbuf);
+#endif
+
    cmdbuf->state.compute.desc_state.sets[0] = save_ctx->set0;
    if (save_ctx->push_set0.desc_count) {
       memcpy(push_set0->descs.host, save_ctx->push_set0.desc_storage,
@@ -108,6 +117,11 @@ panvk_per_arch(cmd_meta_gfx_start)(
    gfx_state_set_dirty(cmdbuf, OQ);
 
    cmdbuf->state.gfx.vk_meta = true;
+
+#if PAN_ARCH >= 10
+   trace_begin_meta(&cmdbuf->utrace.uts[PANVK_SUBQUEUE_VERTEX_TILER], cmdbuf);
+   trace_begin_meta(&cmdbuf->utrace.uts[PANVK_SUBQUEUE_FRAGMENT], cmdbuf);
+#endif
 }
 
 void
@@ -118,6 +132,11 @@ panvk_per_arch(cmd_meta_gfx_end)(
    struct panvk_descriptor_set *push_set0 =
       cmdbuf->state.gfx.desc_state.push_sets[0];
 
+#if PAN_ARCH >= 10
+   trace_end_meta(&cmdbuf->utrace.uts[PANVK_SUBQUEUE_VERTEX_TILER], cmdbuf);
+   trace_end_meta(&cmdbuf->utrace.uts[PANVK_SUBQUEUE_FRAGMENT], cmdbuf);
+#endif
+
    cmdbuf->state.gfx.desc_state.sets[0] = save_ctx->set0;
    if (save_ctx->push_set0.desc_count) {
       memcpy(push_set0->descs.host, save_ctx->push_set0.desc_storage,