hk,libagx: handle adjacency without a GS

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32759>

src/asahi/lib/agx_linker.h

@@ -62,6 +62,9 @@ struct agx_vs_prolog_key {
    /* If !hw and the draw call is indexed, the index size */
    uint8_t sw_index_size_B;
 
+   /* Adjacency primitive to emulate */
+   enum mesa_prim adjacency;
+
    /* Robustness settings for the vertex fetch */
    struct agx_robustness robustness;
 };

src/asahi/lib/agx_nir_lower_gs.h

@@ -23,6 +23,9 @@ struct nir_def *agx_load_per_vertex_input(struct nir_builder *b,
                                           nir_intrinsic_instr *intr,
                                           struct nir_def *vertex);
 
+nir_def *agx_nir_load_vertex_id(struct nir_builder *b, nir_def *id,
+                                unsigned index_size_B);
+
 bool agx_nir_lower_sw_vs(struct nir_shader *s, unsigned index_size_B);
 
 bool agx_nir_lower_vs_before_gs(struct nir_shader *vs,

src/asahi/lib/agx_nir_lower_ia.c

@@ -14,11 +14,9 @@
  * vertex shaders, as part of geometry/tessellation lowering. It does not apply
  * the topology, which happens in the geometry shader.
  */
-static nir_def *
-load_vertex_id(nir_builder *b, unsigned index_size_B)
+nir_def *
+agx_nir_load_vertex_id(nir_builder *b, nir_def *id, unsigned index_size_B)
 {
-   nir_def *id = nir_channel(b, nir_load_global_invocation_id(b, 32), 0);
-
    /* If drawing with an index buffer, pull the vertex ID. Otherwise, the
     * vertex ID is just the index as-is.
     */
@@ -40,7 +38,8 @@ lower(nir_builder *b, nir_intrinsic_instr *intr, void *data)
    b->cursor = nir_before_instr(&intr->instr);
 
    if (intr->intrinsic == nir_intrinsic_load_vertex_id) {
-      nir_def_replace(&intr->def, load_vertex_id(b, *index_size_B));
+      nir_def *id = nir_channel(b, nir_load_global_invocation_id(b, 32), 0);
+      nir_def_replace(&intr->def, agx_nir_load_vertex_id(b, id, *index_size_B));
       return true;
    } else if (intr->intrinsic == nir_intrinsic_load_instance_id) {
       nir_def_replace(&intr->def,

src/asahi/lib/agx_nir_prolog_epilog.c

@@ -13,6 +13,7 @@
 #include "agx_nir_lower_vbo.h"
 #include "agx_pack.h"
 #include "agx_tilebuffer.h"
+#include "libagx.h"
 #include "nir.h"
 #include "nir_builder.h"
 #include "nir_builder_opcodes.h"
@@ -138,6 +139,36 @@ lower_non_monolithic_uniforms(nir_builder *b, nir_intrinsic_instr *intr,
    }
 }
 
+static bool
+lower_adjacency(nir_builder *b, nir_intrinsic_instr *intr, void *data)
+{
+   const struct agx_vs_prolog_key *key = data;
+   b->cursor = nir_before_instr(&intr->instr);
+
+   if (intr->intrinsic != nir_intrinsic_load_vertex_id)
+      return false;
+
+   nir_def *id = nir_load_vertex_id(b);
+
+   if (key->adjacency == MESA_PRIM_LINES_ADJACENCY) {
+      id = libagx_map_to_line_adj(b, id);
+   } else if (key->adjacency == MESA_PRIM_TRIANGLE_STRIP_ADJACENCY) {
+      id = libagx_map_to_tri_strip_adj(b, id);
+   } else if (key->adjacency == MESA_PRIM_LINE_STRIP_ADJACENCY) {
+      id = libagx_map_to_line_strip_adj(b, id);
+   } else if (key->adjacency == MESA_PRIM_TRIANGLES_ADJACENCY) {
+      /* Sequence (0, 2, 4), (6, 8, 10), ... */
+      id = nir_imul_imm(b, id, 2);
+   } else {
+      unreachable("unknown");
+   }
+
+   id = agx_nir_load_vertex_id(b, id, key->sw_index_size_B);
+
+   nir_def_replace(&intr->def, id);
+   return true;
+}
+
 void
 agx_nir_vs_prolog(nir_builder *b, const void *key_)
 {
@@ -169,8 +200,17 @@ agx_nir_vs_prolog(nir_builder *b, const void *key_)
    /* Now lower the resulting program using the key */
    lower_vbo(b->shader, key->attribs, key->robustness);
 
+   /* Clean up redundant vertex ID loads */
+   if (!key->hw || key->adjacency) {
+      NIR_PASS(_, b->shader, nir_opt_cse);
+      NIR_PASS(_, b->shader, nir_opt_dce);
+   }
+
    if (!key->hw) {
       agx_nir_lower_sw_vs(b->shader, key->sw_index_size_B);
+   } else if (key->adjacency) {
+      nir_shader_intrinsics_pass(b->shader, lower_adjacency,
+                                 nir_metadata_control_flow, (void *)key);
    }
 
    /* Finally, lower uniforms according to our ABI */

src/asahi/libagx/draws.cl

@@ -28,6 +28,34 @@ libagx_predicate_indirect(global uint32_t *out, constant uint32_t *in,
    }
 }
 
+/*
+ * Indexing/offseting is in software if necessary so we strip all
+ * indexing/offset information.
+ */
+KERNEL(1)
+libagx_draw_without_adj(global VkDrawIndirectCommand *out,
+                        global VkDrawIndirectCommand *in,
+                        global struct agx_ia_state *ia, uint64_t index_buffer,
+                        uint64_t index_buffer_range_el, int index_size_B,
+                        enum mesa_prim prim)
+{
+   *out = (VkDrawIndirectCommand){
+      .vertexCount = libagx_remap_adj_count(in->vertexCount, prim),
+      .instanceCount = in->instanceCount,
+   };
+
+   /* TODO: Deduplicate */
+   if (index_size_B) {
+      uint offs = in->firstVertex;
+
+      ia->index_buffer = libagx_index_buffer(
+         index_buffer, index_buffer_range_el, offs, index_size_B, 0);
+
+      ia->index_buffer_range_el =
+         libagx_index_buffer_range_el(index_buffer_range_el, offs);
+   }
+}
+
 /* Precondition: len must be < the group size */
 static void
 libagx_memcpy_small(global uchar *dst, constant uchar *src, uint len, uint tid)

src/asahi/libagx/geometry.cl

@@ -220,6 +220,50 @@ vertex_id_for_topology(enum mesa_prim mode, bool flatshade_first, uint prim,
    }
 }
 
+uint
+libagx_map_to_line_adj(uint id)
+{
+   /* Sequence (1, 2), (5, 6), (9, 10), ... */
+   return ((id & ~1) * 2) + (id & 1) + 1;
+}
+
+uint
+libagx_map_to_line_strip_adj(uint id)
+{
+   /* Sequence (1, 2), (2, 3), (4, 5), .. */
+   uint prim = id / 2;
+   uint vert = id & 1;
+   return prim + vert + 1;
+}
+
+uint
+libagx_map_to_tri_strip_adj(uint id)
+{
+   /* Sequence (0, 2, 4), (2, 6, 4), (4, 6, 8), (6, 10, 8)
+    *
+    * Although tri strips with adjacency have 6 cases in general, after
+    * disregarding the vertices only available in a geometry shader, there are
+    * only even/odd cases. In other words, it's just a triangle strip subject to
+    * extra padding.
+    *
+    * Dividing through by two, the sequence is:
+    *
+    *   (0, 1, 2), (1, 3, 2), (2, 3, 4), (3, 5, 4)
+    */
+   uint prim = id / 3;
+   uint vtx = id % 3;
+
+   /* Flip the winding order of odd triangles */
+   if ((prim % 2) == 1) {
+      if (vtx == 1)
+         vtx = 2;
+      else if (vtx == 2)
+         vtx = 1;
+   }
+
+   return 2 * (prim + vtx);
+}
+
 static void
 store_index(uintptr_t index_buffer, uint index_size_B, uint id, uint value)
 {

src/asahi/libagx/libagx_dgc.h

@@ -6,6 +6,7 @@
 #pragma once
 
 #include "compiler/libcl/libcl.h"
+#include "compiler/shader_enums.h"
 #include "agx_pack.h"
 
 #define agx_push(ptr, T, cfg)                                                  \
@@ -496,10 +497,6 @@ struct agx_usc_builder {
 #endif
 } PACKED;
 
-#ifdef __OPENCL_VERSION__
-static_assert(sizeof(struct agx_usc_builder) == 8);
-#endif
-
 static struct agx_usc_builder
 agx_usc_builder(GLOBAL void *out, ASSERTED size_t size)
 {
@@ -577,3 +574,18 @@ libagx_draw_robust_index_vdm_size()
    struct agx_draw draw = agx_draw_indexed(0, 0, 0, 0, 0, 0, 0, 0, 0);
    return agx_vdm_draw_size(0, draw);
 }
+
+static inline unsigned
+libagx_remap_adj_count(unsigned count, enum mesa_prim prim)
+{
+   if (prim == MESA_PRIM_TRIANGLE_STRIP_ADJACENCY) {
+      /* Spec gives formula for # of primitives in a tri strip adj */
+      unsigned c4 = count >= 4 ? count - 4 : 0;
+      return 3 * (c4 / 2);
+   } else if (prim == MESA_PRIM_LINE_STRIP_ADJACENCY) {
+      return 2 * (count >= 3 ? count - 3 : 0);
+   } else {
+      /* Adjacency lists just drop half the vertices. */
+      return count / 2;
+   }
+}

src/asahi/vulkan/hk_cmd_buffer.c

@@ -637,7 +637,6 @@ hk_upload_usc_words(struct hk_cmd_buffer *cmd, struct hk_shader *s,
    struct hk_device *dev = hk_cmd_buffer_device(cmd);
 
    enum pipe_shader_type sw_stage = s->info.stage;
-   enum pipe_shader_type hw_stage = s->b.info.stage;
 
    unsigned constant_push_ranges = DIV_ROUND_UP(s->b.info.rodata.size_16, 64);
    unsigned push_ranges = 2;
@@ -683,7 +682,7 @@ hk_upload_usc_words(struct hk_cmd_buffer *cmd, struct hk_shader *s,
       if (cmd->state.gfx.draw_id_ptr)
          agx_usc_uniform(&b, (6 * count) + 4, 1, cmd->state.gfx.draw_id_ptr);
 
-      if (hw_stage == MESA_SHADER_COMPUTE) {
+      if (linked->sw_indexing) {
          agx_usc_uniform(
             &b, (6 * count) + 8, 4,
             root_ptr + hk_root_descriptor_offset(draw.input_assembly));

src/asahi/vulkan/hk_cmd_draw.c

@@ -1098,11 +1098,22 @@ hk_rast_prim(struct hk_cmd_buffer *cmd)
 {
    struct hk_graphics_state *gfx = &cmd->state.gfx;
    struct hk_api_shader *gs = gfx->shaders[MESA_SHADER_GEOMETRY];
+   struct vk_dynamic_graphics_state *dyn = &cmd->vk.dynamic_graphics_state;
 
-   if (gs != NULL)
+   if (gs != NULL) {
       return gs->variants[HK_GS_VARIANT_RAST].info.gs.out_prim;
-   else
+   } else {
+      switch (dyn->ia.primitive_topology) {
+      case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY:
+      case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY:
+         return MESA_PRIM_LINES;
+      case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY:
+      case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY:
+         return MESA_PRIM_TRIANGLES;
+      default:
          return hk_gs_in_prim(cmd);
+      }
+   }
 }
 
 static uint64_t
@@ -1721,7 +1732,8 @@ hk_get_fast_linked_locked_vs(struct hk_device *dev, struct hk_shader *shader,
       ralloc_memdup(shader->linked.ht, key, sizeof(*key));
 
    /* XXX: Fix this higher up the stack */
-   linked->b.uses_base_param |= !key->prolog.hw;
+   linked->sw_indexing = !key->prolog.hw || key->prolog.adjacency;
+   linked->b.uses_base_param |= linked->sw_indexing;
 
    _mesa_hash_table_insert(shader->linked.ht, key_clone, linked);
    return linked;
@@ -2545,7 +2557,17 @@ hk_flush_dynamic_state(struct hk_cmd_buffer *cmd, struct hk_cs *cs,
          .prolog.robustness.soft_fault = agx_has_soft_fault(&dev->dev),
       };
 
-      if (!key.prolog.hw) {
+      enum mesa_prim prim = vk_conv_topology(dyn->ia.primitive_topology);
+
+      if (mesa_prim_has_adjacency(prim)) {
+         if (draw.restart) {
+            prim = u_decomposed_prim(prim);
+         }
+
+         key.prolog.adjacency = prim;
+      }
+
+      if (key.prolog.adjacency || !key.prolog.hw) {
          key.prolog.sw_index_size_B =
             draw.indexed ? agx_index_size_to_B(draw.index_size) : 0;
       }
@@ -2886,6 +2908,17 @@ hk_flush_dynamic_state(struct hk_cmd_buffer *cmd, struct hk_cs *cs,
       gfx->dirty |= HK_DIRTY_VARYINGS;
    }
 
+   if (gfx->shaders[MESA_SHADER_TESS_EVAL] ||
+       gfx->shaders[MESA_SHADER_GEOMETRY] || linked_vs->sw_indexing) {
+      /* XXX: We should deduplicate this logic */
+      bool indirect = agx_is_indirect(draw.b) || draw.restart;
+
+      desc->root.draw.input_assembly =
+         indirect ? hk_pool_alloc(cmd, sizeof(struct agx_ia_state), 4).gpu
+                  : hk_upload_ia_params(cmd, draw);
+      desc->root_dirty = true;
+   }
+
    if (gfx->shaders[MESA_SHADER_TESS_EVAL] ||
        gfx->shaders[MESA_SHADER_GEOMETRY]) {
 
@@ -2895,10 +2928,6 @@ hk_flush_dynamic_state(struct hk_cmd_buffer *cmd, struct hk_cs *cs,
       /* XXX: We should deduplicate this logic */
       bool indirect = agx_is_indirect(draw.b) || draw.restart;
 
-      desc->root.draw.input_assembly =
-         indirect ? hk_pool_alloc(cmd, sizeof(struct agx_ia_state), 4).gpu
-                  : hk_upload_ia_params(cmd, draw);
-
       if (!indirect) {
          uint32_t verts = draw.b.count[0], instances = draw.b.count[1];
          unsigned vb_size =
@@ -3075,18 +3104,9 @@ hk_handle_passthrough_gs(struct hk_cmd_buffer *cmd, struct agx_draw draw)
       return;
 
    /* Determine if we need a geometry shader to emulate XFB or adjacency */
-   struct vk_dynamic_graphics_state *dyn = &cmd->vk.dynamic_graphics_state;
    struct hk_shader *last_sw = hk_bound_sw_vs_before_gs(gfx);
    uint32_t xfb_outputs = last_sw->info.xfb_info.output_count;
-
-   VkPrimitiveTopology topology = dyn->ia.primitive_topology;
-   bool adjacency =
-      (topology == VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY) ||
-      (topology == VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY) ||
-      (topology == VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY) ||
-      (topology == VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY);
-
-   bool needs_gs = xfb_outputs || adjacency;
+   bool needs_gs = xfb_outputs;
 
    /* Various pipeline statistics are implemented in the pre-GS shader. TODO:
     * This could easily be optimized.
@@ -3137,9 +3157,8 @@ hk_handle_passthrough_gs(struct hk_cmd_buffer *cmd, struct agx_draw draw)
    }
 
    struct hk_device *dev = hk_cmd_buffer_device(cmd);
-   perf_debug(dev, "Binding passthrough GS for%s%s%s\n",
-              xfb_outputs ? " XFB" : "", adjacency ? " adjacency" : "",
-              ia_stats ? " statistics" : "");
+   perf_debug(dev, "Binding passthrough GS for%s%s\n",
+              xfb_outputs ? " XFB" : "", ia_stats ? " statistics" : "");
 
    gs = hk_meta_shader(dev, hk_nir_passthrough_gs, key, key_size);
    gs->is_passthrough = true;
@@ -3376,9 +3395,15 @@ hk_draw(struct hk_cmd_buffer *cmd, uint16_t draw_id, struct agx_draw draw_)
       if (!cs)
          return;
 
+      struct vk_dynamic_graphics_state *dyn = &cmd->vk.dynamic_graphics_state;
       bool geom = cmd->state.gfx.shaders[MESA_SHADER_GEOMETRY];
       bool tess = cmd->state.gfx.shaders[MESA_SHADER_TESS_EVAL];
       bool needs_idx_robust = hk_needs_index_robustness(cmd, &draw);
+      bool adj =
+         mesa_prim_has_adjacency(vk_conv_topology(dyn->ia.primitive_topology));
+      adj &= !geom;
+      needs_idx_robust &= !adj;
+
       struct hk_cs *ccs = NULL;
       uint8_t *out = cs->current;
       assert(cs->current + 0x1000 < cs->end);
@@ -3388,7 +3413,9 @@ hk_draw(struct hk_cmd_buffer *cmd, uint16_t draw_id, struct agx_draw draw_)
 
       cs->stats.calls++;
 
-      if (geom || tess || ia_stats || needs_idx_robust) {
+      if (geom || tess || ia_stats || needs_idx_robust ||
+          (adj && (agx_is_indirect(draw.b) || draw.restart))) {
+
          ccs =
             hk_cmd_buffer_get_cs_general(cmd, &cmd->current_cs.pre_gfx, true);
          if (!ccs)
@@ -3415,9 +3442,42 @@ hk_draw(struct hk_cmd_buffer *cmd, uint16_t draw_id, struct agx_draw draw_)
             continue;
       }
 
+      if (adj) {
+         assert(!geom && "geometry shaders handle adj directly");
+         enum mesa_prim prim = vk_conv_topology(dyn->ia.primitive_topology);
+
+         if (draw.restart) {
+            draw = hk_draw_without_restart(cmd, ccs, draw, 1);
+            prim = u_decomposed_prim(prim);
+         }
+
+         if (agx_is_indirect(draw.b)) {
+            const size_t size = sizeof(VkDrawIndexedIndirectCommand);
+            static_assert(size > sizeof(VkDrawIndirectCommand),
+                          "allocation size is conservative");
+
+            uint64_t out_draw = hk_pool_alloc(cmd, size, 4).gpu;
+            struct hk_descriptor_state *desc = &cmd->state.gfx.descriptors;
+
+            libagx_draw_without_adj(
+               ccs, agx_1d(1), out_draw, draw.b.ptr,
+               desc->root.draw.input_assembly, draw.index_buffer,
+               draw.indexed ? agx_draw_index_range_el(draw) : 0,
+               draw.indexed ? agx_index_size_to_B(draw.index_size) : 0, prim);
+
+            draw = agx_draw_indirect(out_draw);
+         } else {
+            unsigned count = libagx_remap_adj_count(draw.b.count[0], prim);
+
+            draw = (struct agx_draw){
+               .b = agx_3d(count, draw.b.count[1], 1),
+            };
+         }
+      }
+
       enum agx_primitive topology = cmd->state.gfx.topology;
       if (needs_idx_robust) {
-         assert(!geom && !tess);
+         assert(!geom && !tess && !adj);
          perf_debug(dev, "lowering robust index buffer");
 
          cs->current = out;

src/asahi/vulkan/hk_shader.h

@@ -313,6 +313,11 @@ hk_pre_gs_variant(struct hk_api_shader *obj, bool rast_disc)
 struct hk_linked_shader {
    struct agx_linked_shader b;
 
+   /* True if the VS prolog uses software indexing, either for geom/tess or
+    * adjacency primitives.
+    */
+   bool sw_indexing;
+
    /* Distinct from hk_shader::counts due to addition of cf_binding_count, which
     * is delayed since it depends on cull distance.
     */