tu: Fix feedback loops in sysmem mode

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12360>

diff --git a/src/freedreno/vulkan/tu_clear_blit.c b/src/freedreno/vulkan/tu_clear_blit.c
index 5ecf11c..130720a 100644 (file)
--- a/src/freedreno/vulkan/tu_clear_blit.c
+++ b/src/freedreno/vulkan/tu_clear_blit.c
@@ -1045,6 +1045,9 @@ r3d_setup(struct tu_cmd_buffer *cmd,
    tu_cs_emit_regs(cs, A6XX_GRAS_LRZ_CNTL(0));
    tu_cs_emit_regs(cs, A6XX_RB_LRZ_CNTL(0));
 
+   tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_SC_CNTL,
+                        A6XX_GRAS_SC_CNTL_CCUSINGLECACHELINESIZE(2));
+
    if (cmd->state.predication_active) {
       tu_cs_emit_pkt7(cs, CP_DRAW_PRED_ENABLE_LOCAL, 1);
       tu_cs_emit(cs, 0);

diff --git a/src/freedreno/vulkan/tu_cmd_buffer.c b/src/freedreno/vulkan/tu_cmd_buffer.c
index bd80aab..d29ec1d 100644 (file)
--- a/src/freedreno/vulkan/tu_cmd_buffer.c
+++ b/src/freedreno/vulkan/tu_cmd_buffer.c
@@ -279,6 +279,26 @@ tu6_emit_mrt(struct tu_cmd_buffer *cmd,
 
    unsigned layers = MAX2(fb->layers, util_logbase2(subpass->multiview_mask) + 1);
    tu_cs_emit_regs(cs, A6XX_GRAS_MAX_LAYER_INDEX(layers - 1));
+
+   tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_SC_CNTL,
+                        A6XX_GRAS_SC_CNTL_CCUSINGLECACHELINESIZE(2));
+
+   /* 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
+    * read different components without a barrier in between. This is a
+    * problem in sysmem mode with UBWC, because the main buffer and flags
+    * buffer can get out-of-sync if only one is flushed. We fix this by
+    * setting the SINGLE_PRIM_MODE field to the same value that the blob does
+    * for advanced_blend in sysmem mode if a feedback loop is detected.
+    */
+   if (subpass->feedback) {
+      tu_cond_exec_start(cs, CP_COND_EXEC_0_RENDER_MODE_SYSMEM);
+      tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_SC_CNTL,
+                           A6XX_GRAS_SC_CNTL_CCUSINGLECACHELINESIZE(2) |
+                           A6XX_GRAS_SC_CNTL_SINGLE_PRIM_MODE(
+                              FLUSH_PER_OVERLAP_AND_OVERWRITE));
+      tu_cond_exec_end(cs);
+   }
 }
 
 void
@@ -783,8 +803,6 @@ tu6_init_hw(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
    tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_B183, 0);
 
    tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_SU_CONSERVATIVE_RAS_CNTL, 0);
-   tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_SC_CNTL,
-                        A6XX_GRAS_SC_CNTL_CCUSINGLECACHELINESIZE(2));
    tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_80AF, 0);
    tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9210, 0);
    tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9211, 0);

diff --git a/src/freedreno/vulkan/tu_pass.c b/src/freedreno/vulkan/tu_pass.c
index a76abe9..c67b3e6 100644 (file)
--- a/src/freedreno/vulkan/tu_pass.c
+++ b/src/freedreno/vulkan/tu_pass.c
@@ -403,6 +403,36 @@ tu_render_pass_patch_input_gmem(struct tu_render_pass *pass)
    }
 }
 
+static void
+tu_render_pass_check_feedback_loop(struct tu_render_pass *pass)
+{
+   for (unsigned i = 0; i < pass->subpass_count; i++) {
+      struct tu_subpass *subpass = &pass->subpasses[i];
+
+      for (unsigned j = 0; j < subpass->color_count; j++) {
+         uint32_t a = subpass->color_attachments[j].attachment;
+         if (a == VK_ATTACHMENT_UNUSED)
+            continue;
+         for (unsigned k = 0; k < subpass->input_count; k++) {
+            if (subpass->input_attachments[k].attachment == a) {
+               subpass->feedback = true;
+               break;
+            }
+         }
+      }
+
+      if (subpass->depth_stencil_attachment.attachment != VK_ATTACHMENT_UNUSED) {
+         for (unsigned k = 0; k < subpass->input_count; k++) {
+            if (subpass->input_attachments[k].attachment ==
+                subpass->depth_stencil_attachment.attachment) {
+               subpass->feedback = true;
+               break;
+            }
+         }
+      }
+   }
+}
+
 static void update_samples(struct tu_subpass *subpass,
                            VkSampleCountFlagBits samples)
 {
@@ -708,6 +738,8 @@ tu_CreateRenderPass2(VkDevice _device,
 
    tu_render_pass_patch_input_gmem(pass);
 
+   tu_render_pass_check_feedback_loop(pass);
+
    /* disable unused attachments */
    for (uint32_t i = 0; i < pass->attachment_count; i++) {
       struct tu_render_pass_attachment *att = &pass->attachments[i];

diff --git a/src/freedreno/vulkan/tu_private.h b/src/freedreno/vulkan/tu_private.h
index eaadf02..d87d64a 100644 (file)
--- a/src/freedreno/vulkan/tu_private.h
+++ b/src/freedreno/vulkan/tu_private.h
@@ -1566,6 +1566,9 @@ struct tu_subpass
    uint32_t resolve_count;
    bool resolve_depth_stencil;
 
+   /* True if there is any feedback loop at all. */
+   bool feedback;
+
    /* True if we must invalidate UCHE thanks to a feedback loop. */
    bool feedback_invalidate;