pan,nir: introduce load_attribute_pan

load_attribute_pan is a panfrost-specific intrinsic for loading
vertex attributes. Takes explicit vertex and instance IDs which
we need in order to implement vertex attribute divisor with
non-zero base instance on v9+.

Passes which are used by panvk are modified to be aware of
load_attribute_pan.

Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Mary Guillemard <mary.guillemard@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32039>

src/compiler/nir/nir_divergence_analysis.c

@@ -384,6 +384,12 @@ visit_intrinsic(nir_intrinsic_instr *instr, struct divergence_state *state)
          is_divergent = true;
       }
       break;
+   case nir_intrinsic_load_attribute_pan:
+      assert(stage == MESA_SHADER_VERTEX);
+      is_divergent = src_divergent(instr->src[0], state) ||
+                     src_divergent(instr->src[1], state) ||
+                     src_divergent(instr->src[2], state);
+      break;
    case nir_intrinsic_load_per_vertex_input:
       is_divergent = src_divergent(instr->src[0], state) ||
                      src_divergent(instr->src[1], state);

src/compiler/nir/nir_gather_info.c

@@ -543,6 +543,7 @@ gather_intrinsic_info(nir_intrinsic_instr *instr, nir_shader *shader,
    case nir_intrinsic_load_input_vertex:
    case nir_intrinsic_load_interpolated_input:
    case nir_intrinsic_load_per_primitive_input:
+   case nir_intrinsic_load_attribute_pan:
       if (shader->info.stage == MESA_SHADER_TESS_EVAL &&
           instr->intrinsic == nir_intrinsic_load_input &&
           !is_patch_special) {

src/compiler/nir/nir_intrinsics.py

@@ -1476,6 +1476,12 @@ intrinsic("prefetch_sam_ir3", [1, 1], flags=[CAN_REORDER])
 intrinsic("prefetch_tex_ir3", [1], flags=[CAN_REORDER])
 intrinsic("prefetch_ubo_ir3", [1], flags=[CAN_REORDER])
 
+# Panfrost-specific intrinsic for loading vertex attributes. Takes explicit
+# vertex and instance IDs which we need in order to implement vertex attribute
+# divisor with non-zero base instance on v9+.
+# src[] = { vertex_id, instance_id, offset }
+load("attribute_pan", [1, 1, 1], [BASE, COMPONENT, DEST_TYPE, IO_SEMANTICS], [CAN_ELIMINATE, CAN_REORDER])
+
 # Intrinsics used by the Midgard/Bifrost blend pipeline. These are defined
 # within a blend shader to read/write the raw value from the tile buffer,
 # without applying any format conversion in the process. If the shader needs

src/compiler/nir/nir_lower_io.c

@@ -2870,6 +2870,7 @@ nir_get_io_offset_src_number(const nir_intrinsic_instr *instr)
    case nir_intrinsic_store_per_vertex_output:
    case nir_intrinsic_store_per_view_output:
    case nir_intrinsic_store_per_primitive_output:
+   case nir_intrinsic_load_attribute_pan:
       return 2;
    default:
       return -1;

src/compiler/nir/nir_opt_shrink_vectors.c

@@ -332,7 +332,8 @@ opt_shrink_vectors_intrinsic(nir_builder *b, nir_intrinsic_instr *instr,
    case nir_intrinsic_load_global:
    case nir_intrinsic_load_global_constant:
    case nir_intrinsic_load_kernel_input:
-   case nir_intrinsic_load_scratch: {
+   case nir_intrinsic_load_scratch:
+   case nir_intrinsic_load_attribute_pan: {
       /* Must be a vectorized intrinsic that we can resize. */
       assert(instr->num_components != 0);
 

src/compiler/nir/nir_opt_sink.c

@@ -109,6 +109,7 @@ can_sink_instr(nir_instr *instr, nir_move_options options, bool *can_mov_out_of_
       case nir_intrinsic_load_frag_coord:
       case nir_intrinsic_load_frag_coord_zw:
       case nir_intrinsic_load_pixel_coord:
+      case nir_intrinsic_load_attribute_pan:
          return options & nir_move_load_input;
       case nir_intrinsic_load_uniform:
       case nir_intrinsic_load_kernel_input:

src/compiler/nir/nir_print.c

@@ -1364,6 +1364,7 @@ print_intrinsic_instr(nir_intrinsic_instr *instr, print_state *state)
          case nir_intrinsic_load_per_vertex_input:
          case nir_intrinsic_load_input_vertex:
          case nir_intrinsic_load_coefficients_agx:
+         case nir_intrinsic_load_attribute_pan:
             mode = nir_var_shader_in;
             break;
 

src/compiler/nir/nir_validate.c

@@ -621,6 +621,7 @@ validate_intrinsic_instr(nir_intrinsic_instr *instr, validate_state *state)
    case nir_intrinsic_load_per_view_output:
    case nir_intrinsic_load_per_primitive_output:
    case nir_intrinsic_load_push_constant:
+   case nir_intrinsic_load_attribute_pan:
       /* All memory load operations must load at least a byte */
       validate_assert(state, instr->def.bit_size >= 8);
       break;

src/panfrost/compiler/bi_lower_divergent_indirects.c

@@ -48,6 +48,7 @@ bi_lower_divergent_indirects_impl(nir_builder *b, nir_intrinsic_instr *intr,
    switch (intr->intrinsic) {
    case nir_intrinsic_load_input:
    case nir_intrinsic_load_interpolated_input:
+   case nir_intrinsic_load_attribute_pan:
       /* Attributes and varyings */
       offset = nir_get_io_offset_src(intr);
       break;