tu: Implement extendedDynamicState3ColorBlendEnable

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18912>

src/freedreno/vulkan/tu_cmd_buffer.c

@@ -2712,8 +2712,9 @@ tu_CmdBindPipeline(VkCommandBuffer commandBuffer,
       cmd->state.pipeline_color_write_enable = pipeline->blend.color_write_enable;
       cmd->state.dirty |= TU_CMD_DIRTY_BLEND;
    }
-   if (cmd->state.pipeline_blend_enable != pipeline->blend.blend_enable) {
+   if (!(pipeline->dynamic_state_mask & BIT(TU_DYNAMIC_STATE_BLEND_ENABLE)) &&
-      cmd->state.pipeline_blend_enable = pipeline->blend.blend_enable;
+       cmd->state.blend_enable != pipeline->blend.blend_enable) {
+      cmd->state.blend_enable = pipeline->blend.blend_enable;
       cmd->state.dirty |= TU_CMD_DIRTY_BLEND;
    }
    if (!(pipeline->dynamic_state_mask & BIT(TU_DYNAMIC_STATE_LOGIC_OP_ENABLE)) &&
@@ -3314,6 +3315,29 @@ tu_CmdSetProvokingVertexModeEXT(VkCommandBuffer commandBuffer,
       provokingVertexMode == VK_PROVOKING_VERTEX_MODE_LAST_VERTEX_EXT;
 }
 
+VKAPI_ATTR void VKAPI_CALL
+tu_CmdSetColorBlendEnableEXT(VkCommandBuffer commandBuffer,
+                             uint32_t firstAttachment,
+                             uint32_t attachmentCount,
+                             const VkBool32 *pColorBlendEnables)
+{
+   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+
+   for (unsigned i = 0; i < attachmentCount; i++) {
+      unsigned att = i + firstAttachment;
+      cmd->state.blend_enable =
+         (cmd->state.blend_enable & ~BIT(att)) |
+         COND(pColorBlendEnables[i], BIT(att));
+      const uint32_t blend_enable =
+         A6XX_RB_MRT_CONTROL_BLEND | A6XX_RB_MRT_CONTROL_BLEND2;
+      cmd->state.rb_mrt_control[i] =
+         (cmd->state.rb_mrt_control[i] & ~blend_enable) |
+         COND(pColorBlendEnables[i], blend_enable);
+   }
+
+   cmd->state.dirty |= TU_CMD_DIRTY_BLEND;
+}
+
 static void
 tu_flush_for_access(struct tu_cache_state *cache,
                     enum tu_cmd_access_mask src_mask,
@@ -4609,7 +4633,7 @@ tu6_emit_blend(struct tu_cs *cs, struct tu_cmd_buffer *cmd)
 
    uint32_t blend_enable_mask = color_write_enable;
    if (!(cmd->state.logic_op_enabled && cmd->state.rop_reads_dst))
-      blend_enable_mask &= cmd->state.pipeline_blend_enable;
+      blend_enable_mask &= cmd->state.blend_enable;
 
    tu_cs_emit_regs(cs, A6XX_SP_FS_OUTPUT_CNTL1(.mrt = num_rts));
    tu_cs_emit_regs(cs, A6XX_RB_FS_OUTPUT_CNTL1(.mrt = num_rts));

src/freedreno/vulkan/tu_cmd_buffer.h

@@ -341,7 +341,7 @@ struct tu_cmd_state
    uint32_t rb_mrt_control[MAX_RTS], rb_mrt_blend_control[MAX_RTS];
    uint32_t rb_mrt_control_rop;
    uint32_t rb_blend_cntl, sp_blend_cntl;
-   uint32_t pipeline_color_write_enable, pipeline_blend_enable;
+   uint32_t pipeline_color_write_enable, blend_enable;
    uint32_t color_write_enable;
    bool logic_op_enabled;
    bool rop_reads_dst;

src/freedreno/vulkan/tu_device.c

@@ -778,7 +778,7 @@ tu_GetPhysicalDeviceFeatures2(VkPhysicalDevice physicalDevice,
          features->extendedDynamicState3LineStippleEnable = false;
          features->extendedDynamicState3ProvokingVertexMode = true;
          features->extendedDynamicState3SampleLocationsEnable = true;
-         features->extendedDynamicState3ColorBlendEnable = false;
+         features->extendedDynamicState3ColorBlendEnable = true;
          features->extendedDynamicState3ColorBlendEquation = false;
          features->extendedDynamicState3ColorWriteMask = false;
          features->extendedDynamicState3ViewportWScalingEnable = false;

src/freedreno/vulkan/tu_lrz.c

@@ -673,6 +673,7 @@ tu6_calculate_lrz_state(struct tu_cmd_buffer *cmd,
    gras_lrz_cntl.dir_write = cmd->state.lrz.gpu_dir_tracking;
    gras_lrz_cntl.disable_on_wrong_dir = cmd->state.lrz.gpu_dir_tracking;
 
+
    /* See comment in tu_pipeline about disabling LRZ write for blending. */
    if ((cmd->state.pipeline->dynamic_state_mask & BIT(TU_DYNAMIC_STATE_LOGIC_OP)) &&
        cmd->state.logic_op_enabled && cmd->state.rop_reads_dst) {
@@ -681,6 +682,13 @@ tu6_calculate_lrz_state(struct tu_cmd_buffer *cmd,
       gras_lrz_cntl.lrz_write = false;
    }
 
+   if ((cmd->state.pipeline->dynamic_state_mask & BIT(TU_DYNAMIC_STATE_BLEND_ENABLE)) &&
+       cmd->state.blend_enable) {
+      if (gras_lrz_cntl.lrz_write)
+         perf_debug(cmd->device, "disabling lrz write due to dynamic blend");
+      gras_lrz_cntl.lrz_write = false;
+   }
+
    if ((cmd->state.pipeline->dynamic_state_mask &
         BIT(TU_DYNAMIC_STATE_COLOR_WRITE_ENABLE)) &&
        (cmd->state.color_write_enable &

src/freedreno/vulkan/tu_pipeline.c

@@ -2319,6 +2319,8 @@ tu6_emit_rb_mrt_controls(struct tu_pipeline *pipeline,
          total_bpp += write_bpp;
 
          pipeline->blend.color_write_enable |= BIT(i);
+
+         if (!(pipeline->dynamic_state_mask & BIT(TU_DYNAMIC_STATE_BLEND_ENABLE))) {
             if (att->blendEnable)
                pipeline->blend.blend_enable |= BIT(i);
 
@@ -2326,6 +2328,7 @@ tu6_emit_rb_mrt_controls(struct tu_pipeline *pipeline,
                total_bpp += write_bpp;
             }
          }
+      }
 
       pipeline->blend.rb_mrt_control[i] = rb_mrt_control & pipeline->blend.rb_mrt_control_mask;
       pipeline->blend.rb_mrt_blend_control[i] = rb_mrt_blend_control;
@@ -3778,6 +3781,15 @@ tu_pipeline_builder_parse_dynamic(struct tu_pipeline_builder *builder,
       case VK_DYNAMIC_STATE_PROVOKING_VERTEX_MODE_EXT:
          pipeline->dynamic_state_mask |= BIT(TU_DYNAMIC_STATE_PROVOKING_VTX);
          break;
+      case VK_DYNAMIC_STATE_COLOR_BLEND_ENABLE_EXT:
+         pipeline->dynamic_state_mask |=
+            BIT(TU_DYNAMIC_STATE_BLEND) |
+            BIT(TU_DYNAMIC_STATE_BLEND_ENABLE);
+         pipeline->blend.rb_mrt_control_mask &=
+            ~(A6XX_RB_MRT_CONTROL_BLEND | A6XX_RB_MRT_CONTROL_BLEND2);
+         pipeline->blend.sp_blend_cntl_mask &= ~A6XX_SP_BLEND_CNTL_ENABLE_BLEND__MASK;
+         pipeline->blend.rb_blend_cntl_mask &= ~A6XX_RB_BLEND_CNTL_ENABLE_BLEND__MASK;
+         break;
       default:
          assert(!"unsupported dynamic state");
          break;
@@ -3881,6 +3893,7 @@ tu_pipeline_builder_parse_libraries(struct tu_pipeline_builder *builder,
             BIT(TU_DYNAMIC_STATE_SAMPLE_LOCATIONS) |
             BIT(TU_DYNAMIC_STATE_SAMPLE_LOCATIONS_ENABLE) |
             BIT(TU_DYNAMIC_STATE_BLEND) |
+            BIT(TU_DYNAMIC_STATE_BLEND_ENABLE) |
             BIT(TU_DYNAMIC_STATE_LOGIC_OP) |
             BIT(TU_DYNAMIC_STATE_LOGIC_OP_ENABLE) |
             BIT(TU_DYNAMIC_STATE_COLOR_WRITE_ENABLE) |
@@ -4561,7 +4574,10 @@ tu_pipeline_builder_parse_multisample_and_color_blend(
     * Therefore, we need to emit it in a separate draw state. We keep
     * it disabled for sysmem path as well for the moment.
     */
-   if (blend_enable_mask)
+   if (blend_enable_mask &&
+       !(pipeline->dynamic_state_mask &
+        (BIT(TU_DYNAMIC_STATE_LOGIC_OP) |
+         BIT(TU_DYNAMIC_STATE_BLEND_ENABLE))))
       pipeline->lrz.force_disable_mask |= TU_LRZ_FORCE_DISABLE_WRITE;
 
    for (int i = 0; i < blend_info->attachmentCount; i++) {

src/freedreno/vulkan/tu_pipeline.h

@@ -44,6 +44,7 @@ enum tu_dynamic_state
    TU_DYNAMIC_STATE_VIEWPORT_RANGE,
    TU_DYNAMIC_STATE_LINE_MODE,
    TU_DYNAMIC_STATE_PROVOKING_VTX,
+   TU_DYNAMIC_STATE_BLEND_ENABLE,
    /* re-use the line width enum as it uses GRAS_SU_CNTL: */
    TU_DYNAMIC_STATE_RAST = VK_DYNAMIC_STATE_LINE_WIDTH,
 };