turnip: Treating non-d/s-write pipelines as not having d/s feedback loops.

A subpass in gfxbench has the depth buffer present, but not written to,
for a render pass using the depth buffer as an input attachment.  We can
skip single-prim-mode and the associated "oh no don't use sysmem" in that
case.

Improves gfxbench vk-5-normal perf by 1.56193% +/- 0.0743035% (n=14).
Part of #6327.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18241>

src/freedreno/vulkan/tu_autotune.c

@@ -513,18 +513,13 @@ tu_autotune_use_bypass(struct tu_autotune *at,
    const struct tu_render_pass *pass = cmd_buffer->state.pass;
    const struct tu_framebuffer *framebuffer = cmd_buffer->state.framebuffer;
 
-   for (unsigned i = 0; i < pass->subpass_count; i++) {
-      const struct tu_subpass *subpass = &pass->subpasses[i];
-      /* GMEM works much faster in this case */
-      if (subpass->raster_order_attachment_access)
-         return false;
-
-      /* Would be very slow in sysmem mode because we have to enable
-       * SINGLE_PRIM_MODE(FLUSH_PER_OVERLAP_AND_OVERWRITE)
-       */
-      if (subpass->feedback_loop_color || subpass->feedback_loop_ds)
-         return false;
-   }
+   /* If a feedback loop in the subpass caused one of the pipelines used to set
+    * SINGLE_PRIM_MODE(FLUSH_PER_OVERLAP_AND_OVERWRITE) or even
+    * SINGLE_PRIM_MODE(FLUSH), then that should cause significantly increased
+    * sysmem bandwidth (though we haven't quantified it).
+    */
+   if (cmd_buffer->state.rp.sysmem_single_prim_mode)
+      return false;
 
    /* For VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT buffers
     * we would have to allocate GPU memory at the submit time and copy

src/freedreno/vulkan/tu_cmd_buffer.c

@@ -2489,6 +2489,7 @@ tu_CmdBindPipeline(VkCommandBuffer commandBuffer,
        */
       cmd->state.rp.disable_gmem = true;
    }
+   cmd->state.rp.sysmem_single_prim_mode |= pipeline->sysmem_single_prim_mode;
 
    struct tu_cs *cs = &cmd->draw_cs;
 
@@ -3440,6 +3441,7 @@ tu_render_pass_state_merge(struct tu_render_pass_state *dst,
    dst->has_tess |= src->has_tess;
    dst->has_prim_generated_query_in_rp |= src->has_prim_generated_query_in_rp;
    dst->disable_gmem |= src->disable_gmem;
+   dst->sysmem_single_prim_mode |= src->sysmem_single_prim_mode;
    dst->draw_cs_writes_to_cond_pred |= src->draw_cs_writes_to_cond_pred;
 
    dst->drawcall_count += src->drawcall_count;

src/freedreno/vulkan/tu_cmd_buffer.h

@@ -262,6 +262,7 @@ struct tu_render_pass_state
    bool has_tess;
    bool has_prim_generated_query_in_rp;
    bool disable_gmem;
+   bool sysmem_single_prim_mode;
 
    /* Track whether conditional predicate for COND_REG_EXEC is changed in draw_cs */
    bool draw_cs_writes_to_cond_pred;

src/freedreno/vulkan/tu_pipeline.c

@@ -3840,6 +3840,7 @@ tu_pipeline_builder_parse_rasterization_order(
        */
       sysmem_prim_mode = FLUSH_PER_OVERLAP_AND_OVERWRITE;
       gmem_prim_mode = FLUSH_PER_OVERLAP;
+      pipeline->sysmem_single_prim_mode = true;
    } else {
       /* If there is a feedback loop, then the shader can read the previous value
        * of a pixel being written out. It can also write some components and then
@@ -3850,8 +3851,10 @@ tu_pipeline_builder_parse_rasterization_order(
        * for advanced_blend in sysmem mode if a feedback loop is detected.
        */
       if (builder->subpass_feedback_loop_color ||
-          builder->subpass_feedback_loop_ds) {
+          (builder->subpass_feedback_loop_ds &&
+           (ds_info->depthWriteEnable || ds_info->stencilTestEnable))) {
          sysmem_prim_mode = FLUSH_PER_OVERLAP_AND_OVERWRITE;
+         pipeline->sysmem_single_prim_mode = true;
       }
    }
 

src/freedreno/vulkan/tu_pipeline.h

@@ -192,6 +192,8 @@ struct tu_pipeline
    bool raster_order_attachment_access;
    bool subpass_feedback_loop_ds;
    bool feedback_loop_may_involve_textures;
+   /* If the pipeline sets SINGLE_PRIM_MODE for sysmem. */
+   bool sysmem_single_prim_mode;
 
    bool z_negative_one_to_one;