nvk: implement VK_EXT_conservative_rasterization

This change is built on top of work originally done by Benjamin Lee. Implement conservative rasterization on GPUs that support it. This is done through a MME method on pre-Volta, and through SET_CONSERVATIVE_RASTER* (newly published) on more recent GPUs. primitiveUnderestimation and fullyCoveredFragmentShaderInputVariable will be supported later as they require SPH and compiler work. Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/9627 Signed-off-by: Arthur Huillet <ahuillet@nvidia.com> Reviewed-by: Mary Guillemard <mary.guillemard@collabora.com> Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/28937>
2025-12-24 19:40:10 +01:00 · 2024-04-26 15:29:39 +02:00 · 2024-04-26 15:29:39 +02:00 · 715f2f1425
commit 715f2f1425
parent 4dd97b1d72
4 changed files with 95 additions and 23 deletions
--- a/src/nouveau/vulkan/nvk_cmd_draw.c
+++ b/src/nouveau/vulkan/nvk_cmd_draw.c
@ -39,19 +39,21 @@ nvk_cmd_buffer_3d_cls(struct nvk_cmd_buffer *cmd)
   return pdev->info.cls_eng3d;
 }

-void
-nvk_mme_set_priv_reg(struct mme_builder *b)
-{
+static void
+mme_set_priv_reg(struct mme_builder *b,
+                 struct mme_value value,
+                 struct mme_value mask,
+                 struct mme_value reg) {
   mme_mthd(b, NV9097_WAIT_FOR_IDLE);
   mme_emit(b, mme_zero());

   mme_mthd(b, NV9097_SET_MME_SHADOW_SCRATCH(0));
   mme_emit(b, mme_zero());
-   mme_emit(b, mme_load(b));
-   mme_emit(b, mme_load(b));
+   mme_emit(b, value);
+   mme_emit(b, mask);

   mme_mthd(b, NV9097_SET_FALCON04);
-   mme_emit(b, mme_load(b));
+   mme_emit(b, reg);

   struct mme_value loop_cond = mme_mov(b, mme_zero());
   mme_while(b, ine, loop_cond, mme_imm(1)) {
@ -61,6 +63,27 @@ nvk_mme_set_priv_reg(struct mme_builder *b)
   };
 }

+void
+nvk_mme_set_priv_reg(struct mme_builder *b)
+{
+   struct mme_value value = mme_load(b);
+   struct mme_value mask = mme_load(b);
+   struct mme_value reg = mme_load(b);
+
+   mme_set_priv_reg(b, value, mask, reg);
+}
+
+void
+nvk_mme_set_conservative_raster_state(struct mme_builder *b) {
+   struct mme_value new_state = mme_load(b);
+   struct mme_value old_state = nvk_mme_load_scratch(b, CONSERVATIVE_RASTER_STATE);
+
+   mme_if(b, ine, new_state, old_state) {
+      nvk_mme_store_scratch(b, CONSERVATIVE_RASTER_STATE, new_state);
+      mme_set_priv_reg(b, new_state, mme_imm(BITFIELD_RANGE(23, 2)), mme_imm(0x418800));
+   }
+}
+
 VkResult
 nvk_push_draw_state_init(struct nvk_device *dev, struct nv_push *p)
 {
@ -171,6 +194,12 @@ nvk_push_draw_state_init(struct nvk_device *dev, struct nv_push *p)
      P_INLINE_DATA(p, BITFIELD_BIT(14));
      P_INLINE_DATA(p, reg);
   }
+   
+   /* Set CONSERVATIVE_RASTER_STATE to an invalid value, to ensure the
+    * hardware reg is always set the first time conservative rasterization
+    * is enabled */
+   P_IMMD(p, NV9097, SET_MME_SHADOW_SCRATCH(NVK_MME_SCRATCH_CONSERVATIVE_RASTER_STATE),
+                     ~0);

   P_IMMD(p, NV9097, SET_RENDER_ENABLE_C, MODE_TRUE);

@ -1497,7 +1526,7 @@ vk_to_nv9097_provoking_vertex(VkProvokingVertexModeEXT vk_mode)
 static void
 nvk_flush_rs_state(struct nvk_cmd_buffer *cmd)
 {
-   struct nv_push *p = nvk_cmd_buffer_push(cmd, 40);
+   struct nv_push *p = nvk_cmd_buffer_push(cmd, 44);

   const struct vk_dynamic_graphics_state *dyn =
      &cmd->vk.dynamic_graphics_state;
@ -1650,6 +1679,31 @@ nvk_flush_rs_state(struct nvk_cmd_buffer *cmd)

   if (BITSET_TEST(dyn->dirty, MESA_VK_DYNAMIC_RS_RASTERIZATION_STREAM))
      P_IMMD(p, NV9097, SET_RASTER_INPUT, dyn->rs.rasterization_stream);
+
+   if (BITSET_TEST(dyn->dirty, MESA_VK_DYNAMIC_RS_CONSERVATIVE_MODE) ||
+       BITSET_TEST(dyn->dirty, MESA_VK_DYNAMIC_RS_EXTRA_PRIMITIVE_OVERESTIMATION_SIZE)) {
+      if (dyn->rs.conservative_mode == VK_CONSERVATIVE_RASTERIZATION_MODE_DISABLED_EXT) {
+         P_IMMD(p, NVB197, SET_CONSERVATIVE_RASTER, ENABLE_FALSE);
+      } else {
+         uint32_t extra_overestimate =
+            MIN2(3, dyn->rs.extra_primitive_overestimation_size * 4);
+
+         if (nvk_cmd_buffer_3d_cls(cmd)  < VOLTA_A) {
+            P_1INC(p, NVB197, CALL_MME_MACRO(NVK_MME_SET_CONSERVATIVE_RASTER_STATE));
+            P_INLINE_DATA(p, extra_overestimate << 23);
+         } else {
+            P_IMMD(p, NVC397, SET_CONSERVATIVE_RASTER_CONTROL, {
+                      .extra_prim_bloat = extra_overestimate,
+                      .copy_inner_to_outer = 
+                        (dyn->rs.conservative_mode == VK_CONSERVATIVE_RASTERIZATION_MODE_UNDERESTIMATE_EXT),
+                      .triangle_snap_mode = TRIANGLE_SNAP_MODE_MODE_PRE_SNAP,
+                      .line_and_point_snap_mode = LINE_AND_POINT_SNAP_MODE_MODE_PRE_SNAP,
+                      .uncertainty_region_size = UNCERTAINTY_REGION_SIZE_SIZE_512,
+                     });
+         }
+         P_IMMD(p, NVB197, SET_CONSERVATIVE_RASTER, ENABLE_TRUE);
+      }
+   }
 }

 static VkSampleLocationEXT
--- a/src/nouveau/vulkan/nvk_mme.c
+++ b/src/nouveau/vulkan/nvk_mme.c
@ -7,22 +7,23 @@
 #include "nvk_private.h"

 static const nvk_mme_builder_func mme_builders[NVK_MME_COUNT] = {
-   [NVK_MME_BIND_CBUF_DESC]              = nvk_mme_bind_cbuf_desc,
-   [NVK_MME_CLEAR]                       = nvk_mme_clear,
-   [NVK_MME_DRAW]                        = nvk_mme_draw,
-   [NVK_MME_DRAW_INDEXED]                = nvk_mme_draw_indexed,
-   [NVK_MME_DRAW_INDIRECT]               = nvk_mme_draw_indirect,
-   [NVK_MME_DRAW_INDEXED_INDIRECT]       = nvk_mme_draw_indexed_indirect,
-   [NVK_MME_DRAW_INDIRECT_COUNT]         = nvk_mme_draw_indirect_count,
-   [NVK_MME_DRAW_INDEXED_INDIRECT_COUNT] = nvk_mme_draw_indexed_indirect_count,
-   [NVK_MME_ADD_CS_INVOCATIONS]          = nvk_mme_add_cs_invocations,
-   [NVK_MME_DISPATCH_INDIRECT]           = nvk_mme_dispatch_indirect,
-   [NVK_MME_WRITE_CS_INVOCATIONS]        = nvk_mme_write_cs_invocations,
-   [NVK_MME_COPY_QUERIES]                = nvk_mme_copy_queries,
-   [NVK_MME_XFB_COUNTER_LOAD]            = nvk_mme_xfb_counter_load,
-   [NVK_MME_XFB_DRAW_INDIRECT]           = nvk_mme_xfb_draw_indirect,
-   [NVK_MME_SET_PRIV_REG]                = nvk_mme_set_priv_reg,
-   [NVK_MME_SET_WRITE_MASK]              = nvk_mme_set_write_mask,
+   [NVK_MME_BIND_CBUF_DESC]                = nvk_mme_bind_cbuf_desc,
+   [NVK_MME_CLEAR]                         = nvk_mme_clear,
+   [NVK_MME_DRAW]                          = nvk_mme_draw,
+   [NVK_MME_DRAW_INDEXED]                  = nvk_mme_draw_indexed,
+   [NVK_MME_DRAW_INDIRECT]                 = nvk_mme_draw_indirect,
+   [NVK_MME_DRAW_INDEXED_INDIRECT]         = nvk_mme_draw_indexed_indirect,
+   [NVK_MME_DRAW_INDIRECT_COUNT]           = nvk_mme_draw_indirect_count,
+   [NVK_MME_DRAW_INDEXED_INDIRECT_COUNT]   = nvk_mme_draw_indexed_indirect_count,
+   [NVK_MME_ADD_CS_INVOCATIONS]            = nvk_mme_add_cs_invocations,
+   [NVK_MME_DISPATCH_INDIRECT]             = nvk_mme_dispatch_indirect,
+   [NVK_MME_WRITE_CS_INVOCATIONS]          = nvk_mme_write_cs_invocations,
+   [NVK_MME_COPY_QUERIES]                  = nvk_mme_copy_queries,
+   [NVK_MME_XFB_COUNTER_LOAD]              = nvk_mme_xfb_counter_load,
+   [NVK_MME_XFB_DRAW_INDIRECT]             = nvk_mme_xfb_draw_indirect,
+   [NVK_MME_SET_PRIV_REG]                  = nvk_mme_set_priv_reg,
+   [NVK_MME_SET_WRITE_MASK]                = nvk_mme_set_write_mask,
+   [NVK_MME_SET_CONSERVATIVE_RASTER_STATE] = nvk_mme_set_conservative_raster_state,
 };

 uint32_t *
--- a/src/nouveau/vulkan/nvk_mme.h
+++ b/src/nouveau/vulkan/nvk_mme.h
@ -26,6 +26,7 @@ enum nvk_mme {
   NVK_MME_XFB_DRAW_INDIRECT,
   NVK_MME_SET_PRIV_REG,
   NVK_MME_SET_WRITE_MASK,
+   NVK_MME_SET_CONSERVATIVE_RASTER_STATE,
   NVK_MME_COUNT,
 };

@ -37,6 +38,9 @@ enum nvk_mme_scratch {
   NVK_MME_SCRATCH_DRAW_PAD_DW,
   NVK_MME_SCRATCH_DRAW_IDX,
   NVK_MME_SCRATCH_VIEW_MASK,
+   NVK_MME_SCRATCH_WRITE_MASK_DYN,
+   NVK_MME_SCRATCH_WRITE_MASK_PIPELINE,
+   NVK_MME_SCRATCH_CONSERVATIVE_RASTER_STATE,

   /* Must be at the end */
   NVK_MME_NUM_SCRATCH,
@ -126,5 +130,6 @@ void nvk_mme_xfb_counter_load(struct mme_builder *b);
 void nvk_mme_xfb_draw_indirect(struct mme_builder *b);
 void nvk_mme_set_priv_reg(struct mme_builder *b);
 void nvk_mme_set_write_mask(struct mme_builder *b);
+void nvk_mme_set_conservative_raster_state(struct mme_builder *b);

 #endif /* NVK_MME_H */
--- a/src/nouveau/vulkan/nvk_physical_device.c
+++ b/src/nouveau/vulkan/nvk_physical_device.c
@ -174,6 +174,7 @@ nvk_get_device_extensions(const struct nvk_instance *instance,
      .EXT_buffer_device_address = true,
      .EXT_calibrated_timestamps = true,
      .EXT_conditional_rendering = true,
+      .EXT_conservative_rasterization = info->cls_eng3d >= MAXWELL_B,
      .EXT_color_write_enable = true,
      .EXT_custom_border_color = true,
      .EXT_depth_bias_control = true,
@ -868,6 +869,17 @@ nvk_get_device_properties(const struct nvk_instance *instance,
      /* VK_KHR_push_descriptor */
      .maxPushDescriptors = NVK_MAX_PUSH_DESCRIPTORS,

+      /* VK_EXT_conservative_rasterization */
+      .primitiveOverestimationSize = info->cls_eng3d >= VOLTA_A ? 1.0f / 512.0f : 0.0,
+      .maxExtraPrimitiveOverestimationSize = 0.75,
+      .extraPrimitiveOverestimationSizeGranularity = 0.25,
+      .primitiveUnderestimation = false,
+      .conservativePointAndLineRasterization = true,
+      .degenerateLinesRasterized = info->cls_eng3d >= VOLTA_A,
+      .degenerateTrianglesRasterized = info->cls_eng3d >= PASCAL_A,
+      .fullyCoveredFragmentShaderInputVariable = false,
+      .conservativeRasterizationPostDepthCoverage = true,
+
      /* VK_EXT_custom_border_color */
      .maxCustomBorderColorSamplers = 4000,