poly,asahi: Fetch directly from poly_vertex_state::output_buffer in GS

We have access to the poly_vertex_state from the GS so we might as well
use it.  Asahi uses a single poly_vertex_state for VS and TCS and just
assumes the tessellator stalls before we update it for TCS.  If a driver
wants to use two separate poly_vertex_state buffers, it will be the
driver's responsibility to make the system values return the right one.

Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Reviewed-by: Mary Guillemard <mary@mary.zone>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38404>

src/asahi/vulkan/hk_cmd_draw.c

@@ -1100,9 +1100,7 @@ hk_rast_prim(struct hk_cmd_buffer *cmd)
 }
 
 static uint64_t
-hk_upload_geometry_params(struct hk_cmd_buffer *cmd,
-                          uint64_t vertex_output_buffer,
-                          struct agx_draw draw)
+hk_upload_geometry_params(struct hk_cmd_buffer *cmd, struct agx_draw draw)
 {
    struct hk_device *dev = hk_cmd_buffer_device(cmd);
    struct hk_descriptor_state *desc = &cmd->state.gfx.descriptors;
@@ -1127,7 +1125,6 @@ hk_upload_geometry_params(struct hk_cmd_buffer *cmd,
       /* Overriden by the indirect setup kernel. As tess->GS is always indirect,
        * we can assume here that we're VS->GS.
        */
-      .input_buffer = vertex_output_buffer,
       .input_mask = desc->root.draw.vertex_outputs,
    };
 
@@ -3091,7 +3088,7 @@ hk_flush_dynamic_state(struct hk_cmd_buffer *cmd, struct hk_cs *cs,
 
    if (gfx->shaders[MESA_SHADER_GEOMETRY]) {
       gfx->descriptors.root.draw.geometry_params =
-         hk_upload_geometry_params(cmd, vertex_output_buffer, draw);
+         hk_upload_geometry_params(cmd, draw);
 
       gfx->descriptors.root_dirty = true;
    }

src/compiler/nir/nir_intrinsics.py

@@ -1436,6 +1436,10 @@ system_value("input_topology_poly", 1)
 system_value("vs_outputs_poly", 1, bit_sizes=[64])
 
 # Address of poly_vertex_param for vertex and tessellation evaluation
+# shaders.  In tessellation control shaders, this is the poly_vertex_param
+# for the vertex shader.  In geometry shaders, this is the poly_vertex_param
+# For whichever of vertex or tessellation evaluation shader preceeded the
+# geometry shader.
 system_value("vertex_param_buffer_poly", 1, bit_sizes=[64])
 
 # Address of the parameter buffer for poly geometry shaders

src/gallium/drivers/asahi/agx_state.c

@@ -4025,7 +4025,6 @@ agx_batch_geometry_params(struct agx_batch *batch, uint64_t input_index_buffer,
       if (vb_size) {
          vp.output_buffer =
             agx_pool_alloc_aligned(&batch->pool, vb_size, 4).gpu;
-         params.input_buffer = vp.output_buffer;
       }
 
       struct poly_gs_info *gsi = &batch->ctx->gs->gs;

src/poly/cl/geometry.cl

@@ -364,19 +364,18 @@ poly_vertex_output_buffer(constant struct poly_vertex_params *p)
 }
 
 uintptr_t
-poly_vertex_output_address(uintptr_t buffer, uint64_t mask, uint vtx,
-                           gl_varying_slot location)
+poly_vertex_output_address(constant struct poly_vertex_params *p,
+                           uint64_t mask, uint vtx, gl_varying_slot location)
 {
    /* Written like this to let address arithmetic work */
-   return buffer + ((uintptr_t)poly_tcs_in_offs_el(vtx, location, mask)) * 16;
+   return p->output_buffer +
+      ((uintptr_t)poly_tcs_in_offs_el(vtx, location, mask)) * 16;
 }
 
-uintptr_t
-poly_geometry_input_address(constant struct poly_geometry_params *p, uint vtx,
-                            gl_varying_slot location)
+uint64_t
+poly_geometry_input_mask(constant struct poly_geometry_params *p)
 {
-   return poly_vertex_output_address(p->input_buffer, p->input_mask, vtx,
-                                     location);
+   return p->input_mask;
 }
 
 unsigned

src/poly/geometry.h

@@ -234,7 +234,6 @@ struct poly_geometry_params {
     * reflect the vertex shader for VS->GS or instead the tessellation
     * evaluation shader for TES->GS.
     */
-   uint64_t input_buffer;
    uint64_t input_mask;
 
    /* Location-indexed mask of flat outputs, used for lowering GL edge flags. */
@@ -278,7 +277,7 @@ struct poly_geometry_params {
     */
    uint32_t input_topology;
 } PACKED;
-static_assert(sizeof(struct poly_geometry_params) == 86 * 4);
+static_assert(sizeof(struct poly_geometry_params) == 84 * 4);
 
 /* TCS shared memory layout:
  *
@@ -576,10 +575,8 @@ poly_gs_setup_indirect(uint64_t index_buffer, constant uint *draw,
          heap, p->input_primitives * p->count_buffer_stride);
    }
 
-   const uintptr_t vertex_buffer =
+   vp->output_buffer =
       (uintptr_t)poly_heap_alloc_nonatomic(heap, vertex_buffer_size);
-   vp->output_buffer = vertex_buffer;
-   p->input_buffer = vertex_buffer;
 
    p->input_mask = vs_outputs;
 

src/poly/nir/poly_nir_lower_gs.c

@@ -285,20 +285,21 @@ poly_load_per_vertex_input(nir_builder *b, nir_intrinsic_instr *intr,
    nir_def *location = nir_iadd_imm(b, intr->src[1].ssa, sem.location);
    nir_def *addr;
 
+   nir_def *input_mask;
    if (b->shader->info.stage == MESA_SHADER_GEOMETRY) {
       /* GS may be preceded by VS or TES so specified as param */
-      addr = poly_geometry_input_address(
-         b, nir_load_geometry_param_buffer_poly(b), vertex, location);
+      input_mask = poly_geometry_input_mask(
+         b, nir_load_geometry_param_buffer_poly(b));
    } else {
       assert(b->shader->info.stage == MESA_SHADER_TESS_CTRL);
 
       /* TCS always preceded by VS so we use the VS state directly */
-      nir_def *vp = nir_load_vertex_param_buffer_poly(b);
-      addr = poly_vertex_output_address(b, poly_vertex_output_buffer(b, vp),
-                                        nir_load_vs_outputs_poly(b), vertex,
-                                        location);
+      input_mask = nir_load_vs_outputs_poly(b);
    }
 
+   nir_def *vp = nir_load_vertex_param_buffer_poly(b);
+   addr = poly_vertex_output_address(b, vp, input_mask, vertex, location);
+
    addr = nir_iadd_imm(b, addr, 4 * nir_intrinsic_component(intr));
    return nir_load_global_constant(b, intr->def.num_components,
                                    intr->def.bit_size, addr, .align_mul = 4);
@@ -1403,7 +1404,6 @@ lower_vs_before_gs(nir_builder *b, nir_intrinsic_instr *intr, void *data)
    nir_def *location = nir_iadd_imm(b, intr->src[1].ssa, sem.location);
 
    nir_def *vp = nir_load_vertex_param_buffer_poly(b);
-   nir_def *buffer = poly_vertex_output_buffer(b, vp);
 
    /* Instancing is unrolled during tessellation so nr_verts is ignored. */
    nir_def *nr_verts = b->shader->info.stage == MESA_SHADER_VERTEX ?
@@ -1423,7 +1423,7 @@ lower_vs_before_gs(nir_builder *b, nir_intrinsic_instr *intr, void *data)
       nir_iadd(b, nir_imul(b, instance_id, nr_verts), primitive_id);
 
    nir_def *addr = poly_vertex_output_address(
-      b, buffer, nir_imm_int64(b, b->shader->info.outputs_written), linear_id,
+      b, vp, nir_imm_int64(b, b->shader->info.outputs_written), linear_id,
       location);
 
    assert(nir_src_bit_size(intr->src[0]) == 32);