nir: Make image_size a variable-width intrinsic
authorJason Ekstrand <jason.ekstrand@intel.com>
Fri, 3 Mar 2017 05:42:06 +0000 (21:42 -0800)
committerJason Ekstrand <jason.ekstrand@intel.com>
Tue, 14 Mar 2017 14:36:40 +0000 (07:36 -0700)
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
src/compiler/glsl/glsl_to_nir.cpp
src/compiler/nir/nir_intrinsics.h
src/compiler/spirv/spirv_to_nir.c

index fc2a2c4..3a71afd 100644 (file)
@@ -865,10 +865,11 @@ nir_visitor::visit(ir_call *ir)
 
          /* Set the intrinsic destination. */
          if (ir->return_deref) {
-            const nir_intrinsic_info *info =
-                    &nir_intrinsic_infos[instr->intrinsic];
+            unsigned num_components = ir->return_deref->type->vector_elements;
+            if (instr->intrinsic == nir_intrinsic_image_size)
+               instr->num_components = num_components;
             nir_ssa_dest_init(&instr->instr, &instr->dest,
-                              info->dest_components, 32, NULL);
+                              num_components, 32, NULL);
          }
 
          if (op == nir_intrinsic_image_size ||
index 5c8f283..105c56f 100644 (file)
@@ -186,7 +186,7 @@ INTRINSIC(image_atomic_or, 3, ARR(4, 1, 1), true, 1, 1, 0, xx, xx, xx, 0)
 INTRINSIC(image_atomic_xor, 3, ARR(4, 1, 1), true, 1, 1, 0, xx, xx, xx, 0)
 INTRINSIC(image_atomic_exchange, 3, ARR(4, 1, 1), true, 1, 1, 0, xx, xx, xx, 0)
 INTRINSIC(image_atomic_comp_swap, 4, ARR(4, 1, 1, 1), true, 1, 1, 0, xx, xx, xx, 0)
-INTRINSIC(image_size, 0, ARR(0), true, 4, 1, 0, xx, xx, xx,
+INTRINSIC(image_size, 0, ARR(0), true, 0, 1, 0, xx, xx, xx,
           NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
 INTRINSIC(image_samples, 0, ARR(0), true, 1, 1, 0, xx, xx, xx,
           NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
index 5b7d8f9..d3ad2d1 100644 (file)
@@ -1998,17 +1998,21 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode,
    if (opcode != SpvOpImageWrite) {
       struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
       struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
-      nir_ssa_dest_init(&intrin->instr, &intrin->dest, 4, 32, NULL);
+
+      unsigned dest_components =
+         nir_intrinsic_infos[intrin->intrinsic].dest_components;
+      if (intrin->intrinsic == nir_intrinsic_image_size) {
+         dest_components = intrin->num_components =
+            glsl_get_vector_elements(type->type);
+      }
+
+      nir_ssa_dest_init(&intrin->instr, &intrin->dest,
+                        dest_components, 32, NULL);
 
       nir_builder_instr_insert(&b->nb, &intrin->instr);
 
-      /* The image intrinsics always return 4 channels but we may not want
-       * that many.  Emit a mov to trim it down.
-       */
-      unsigned swiz[4] = {0, 1, 2, 3};
       val->ssa = vtn_create_ssa_value(b, type->type);
-      val->ssa->def = nir_swizzle(&b->nb, &intrin->dest.ssa, swiz,
-                                  glsl_get_vector_elements(type->type), false);
+      val->ssa->def = &intrin->dest.ssa;
    } else {
       nir_builder_instr_insert(&b->nb, &intrin->instr);
    }