tu: Add tracepoints around draws, with shader sha1s.

Like for CSes, this lets you associate draw call times with the associated
shaders for pinpointing which pipelines are the most expensive.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35759>

src/freedreno/vulkan/tu_cmd_buffer.cc

@@ -6595,6 +6595,14 @@ tu6_draw_common(struct tu_cmd_buffer *cmd,
    const struct tu_program_state *program = &cmd->state.program;
    struct tu_render_pass_state *rp = &cmd->state.rp;
 
+   trace_start_draw(
+      &cmd->trace, &cmd->draw_cs, cmd, draw_count,
+      cmd->state.program.stage_sha1[MESA_SHADER_VERTEX],
+      cmd->state.program.stage_sha1[MESA_SHADER_TESS_CTRL],
+      cmd->state.program.stage_sha1[MESA_SHADER_TESS_EVAL],
+      cmd->state.program.stage_sha1[MESA_SHADER_GEOMETRY],
+      cmd->state.program.stage_sha1[MESA_SHADER_FRAGMENT]);
+
    /* Emit state first, because it's needed for bandwidth calculations */
    uint32_t dynamic_draw_state_dirty = 0;
    if (!BITSET_IS_EMPTY(cmd->vk.dynamic_graphics_state.dirty) ||
@@ -7035,6 +7043,8 @@ tu_CmdDraw(VkCommandBuffer commandBuffer,
    tu_cs_emit(cs, tu_draw_initiator(cmd, DI_SRC_SEL_AUTO_INDEX));
    tu_cs_emit(cs, instanceCount);
    tu_cs_emit(cs, vertexCount);
+
+   trace_end_draw(&cmd->trace, cs);
 }
 TU_GENX(tu_CmdDraw);
 
@@ -7081,6 +7091,9 @@ tu_CmdDrawMultiEXT(VkCommandBuffer commandBuffer,
       tu_cs_emit(cs, instanceCount);
       tu_cs_emit(cs, draw->vertexCount);
    }
+
+   if (i != 0)
+      trace_end_draw(&cmd->trace, cs);
 }
 TU_GENX(tu_CmdDrawMultiEXT);
 
@@ -7107,6 +7120,8 @@ tu_CmdDrawIndexed(VkCommandBuffer commandBuffer,
    tu_cs_emit(cs, firstIndex);
    tu_cs_emit_qw(cs, cmd->state.index_va);
    tu_cs_emit(cs, cmd->state.max_index_count);
+
+   trace_end_draw(&cmd->trace, cs);
 }
 TU_GENX(tu_CmdDrawIndexed);
 
@@ -7158,6 +7173,9 @@ tu_CmdDrawMultiIndexedEXT(VkCommandBuffer commandBuffer,
       tu_cs_emit_qw(cs, cmd->state.index_va);
       tu_cs_emit(cs, cmd->state.max_index_count);
    }
+
+   if (i != 0)
+      trace_end_draw(&cmd->trace, cs);
 }
 TU_GENX(tu_CmdDrawMultiIndexedEXT);
 
@@ -7201,6 +7219,8 @@ tu_CmdDrawIndirect(VkCommandBuffer commandBuffer,
    tu_cs_emit(cs, drawCount);
    tu_cs_emit_qw(cs, vk_buffer_address(&buf->vk, offset));
    tu_cs_emit(cs, stride);
+
+   trace_end_draw(&cmd->trace, cs);
 }
 TU_GENX(tu_CmdDrawIndirect);
 
@@ -7232,6 +7252,8 @@ tu_CmdDrawIndexedIndirect(VkCommandBuffer commandBuffer,
    tu_cs_emit(cs, cmd->state.max_index_count);
    tu_cs_emit_qw(cs, vk_buffer_address(&buf->vk, offset));
    tu_cs_emit(cs, stride);
+
+   trace_end_draw(&cmd->trace, cs);
 }
 TU_GENX(tu_CmdDrawIndexedIndirect);
 
@@ -7269,6 +7291,8 @@ tu_CmdDrawIndirectCount(VkCommandBuffer commandBuffer,
    tu_cs_emit_qw(cs, vk_buffer_address(&buf->vk, offset));
    tu_cs_emit_qw(cs, vk_buffer_address(&count_buf->vk, countBufferOffset));
    tu_cs_emit(cs, stride);
+
+   trace_end_draw(&cmd->trace, cs);
 }
 TU_GENX(tu_CmdDrawIndirectCount);
 
@@ -7303,6 +7327,8 @@ tu_CmdDrawIndexedIndirectCount(VkCommandBuffer commandBuffer,
    tu_cs_emit_qw(cs, vk_buffer_address(&buf->vk, offset));
    tu_cs_emit_qw(cs, vk_buffer_address(&count_buf->vk, countBufferOffset));
    tu_cs_emit(cs, stride);
+
+   trace_end_draw(&cmd->trace, cs);
 }
 TU_GENX(tu_CmdDrawIndexedIndirectCount);
 
@@ -7345,6 +7371,8 @@ tu_CmdDrawIndirectByteCountEXT(VkCommandBuffer commandBuffer,
    tu_cs_emit_qw(cs, vk_buffer_address(&buf->vk, counterBufferOffset));
    tu_cs_emit(cs, counterOffset);
    tu_cs_emit(cs, vertexStride);
+
+   trace_end_draw(&cmd->trace, cs);
 }
 TU_GENX(tu_CmdDrawIndirectByteCountEXT);
 

src/freedreno/vulkan/tu_cmd_buffer.h

@@ -417,6 +417,8 @@ enum tu_suspend_resume_state
    SR_IN_CHAIN_AFTER_PRE_CHAIN,
 };
 
+typedef char tu_sha1_str[SHA1_DIGEST_STRING_LENGTH];
+
 struct tu_cmd_state
 {
    uint32_t dirty;
@@ -547,6 +549,9 @@ struct tu_cmd_state
 
    bool occlusion_query_may_be_running;
 
+   bool trace_draws_enabled;
+   enum tu_pipeline_type trace_draws_pipeline_type;
+
    enum tu_suspend_resume_state suspend_resume;
 
    bool suspending, resuming;

src/freedreno/vulkan/tu_perfetto.cc

@@ -15,6 +15,7 @@
 #include "util/perf/u_perfetto.h"
 #include "util/perf/u_perfetto_renderpass.h"
 
+#include "tu_cmd_buffer.h"
 #include "tu_tracepoints.h"
 #include "tu_tracepoints_perfetto.h"
 #include "vk_object.h"
@@ -54,6 +55,7 @@ enum tu_stage_id {
    GMEM_STAGE_ID,
    BYPASS_STAGE_ID,
    BLIT_STAGE_ID,
+   DRAW_STAGE_ID,
    COMPUTE_STAGE_ID,
    CLEAR_SYSMEM_STAGE_ID,
    CLEAR_GMEM_STAGE_ID,
@@ -84,6 +86,7 @@ static const struct {
    [GMEM_STAGE_ID]           = { "GMEM", "Rendering to GMEM" },
    [BYPASS_STAGE_ID]         = { "Bypass", "Rendering to system memory" },
    [BLIT_STAGE_ID]           = { "Blit", "Performing a Blit operation" },
+   [DRAW_STAGE_ID]           = { "Draw", "Performing a graphics-pipeline draw" },
    [COMPUTE_STAGE_ID]        = { "Compute", "Compute job" },
    [CLEAR_SYSMEM_STAGE_ID]   = { "Clear Sysmem", "" },
    [CLEAR_GMEM_STAGE_ID]     = { "Clear GMEM", "Per-tile (GMEM) clear" },
@@ -525,6 +528,7 @@ CREATE_EVENT_CALLBACK(binning_ib, BINNING_STAGE_ID)
 CREATE_EVENT_CALLBACK(draw_ib_gmem, GMEM_STAGE_ID)
 CREATE_EVENT_CALLBACK(draw_ib_sysmem, BYPASS_STAGE_ID)
 CREATE_EVENT_CALLBACK(blit, BLIT_STAGE_ID)
+CREATE_EVENT_CALLBACK(draw, DRAW_STAGE_ID)
 CREATE_EVENT_CALLBACK(compute, COMPUTE_STAGE_ID)
 CREATE_EVENT_CALLBACK(compute_indirect, COMPUTE_STAGE_ID)
 CREATE_EVENT_CALLBACK(generic_clear, GENERIC_CLEAR_STAGE_ID)

src/freedreno/vulkan/tu_pipeline.cc

@@ -2230,6 +2230,11 @@ tu_emit_program_state(struct tu_cs *sub_cs,
                   shaders[stage]->dynamic_descriptor_sizes[i];
             }
          }
+
+         if (variants[stage]) {
+            memcpy(prog->stage_sha1[stage], variants[stage]->sha1_str,
+                   sizeof(variants[stage]->sha1_str));
+         }
       }
    }
 

src/freedreno/vulkan/tu_pipeline.h

@@ -105,6 +105,8 @@ struct tu_program_state
 
       struct tu_program_descriptor_linkage link[MESA_SHADER_STAGES];
 
+      char stage_sha1[MESA_SHADER_STAGES][SHA1_DIGEST_STRING_LENGTH];
+
       unsigned dynamic_descriptor_offsets[MAX_SETS];
 
       bool per_view_viewport;

src/freedreno/vulkan/tu_tracepoints.py

@@ -120,6 +120,19 @@ begin_end_tp('render_pass',
               Arg(type='int32_t',                               var='lrzWriteDisabledAtDraw',                               c_format='%d'),
               Arg(type='uint32_t',                              var='lrzStatus', c_format='%s', to_prim_type='(fd_lrz_gpu_dir_to_str((enum fd_lrz_gpu_dir)({} & 0xff)))', is_indirect=True),])
 
+begin_end_tp('draw',
+             [Arg(type='uint32_t', var='count', c_format='%u'),
+              Arg(type='tu_sha1_str', var='vs_sha1',  c_format='%s',
+                  copy_func='strcpy'),
+              Arg(type='tu_sha1_str', var='tcs_sha1',  c_format='%s',
+                  copy_func='strcpy'),
+              Arg(type='tu_sha1_str', var='tes_sha1',  c_format='%s',
+                  copy_func='strcpy'),
+              Arg(type='tu_sha1_str', var='gs_sha1',  c_format='%s',
+                  copy_func='strcpy'),
+              Arg(type='tu_sha1_str', var='fs_sha1',  c_format='%s',
+                  copy_func='strcpy'),
+              ], tp_default_enabled=False)
 
 begin_end_tp('binning_ib')
 begin_end_tp('draw_ib_sysmem')

src/util/perf/u_trace.py

@@ -588,6 +588,8 @@ void __trace_${trace_name}(
  % for arg in trace.tp_struct:
   % if arg.copy_func is None:
    __entry->${arg.name} = ${arg.var};
+  % elif arg.length_arg is None:
+     ${arg.copy_func}(__entry->${arg.name}, ${arg.var});
   % else:
    ${arg.copy_func}(__entry->${arg.name}, ${arg.var}, ${arg.length_arg});
   % endif