[Mesa-dev] [PATCH] i965: fix double-precision vertex inputs measurement

[Juan A. Suarez Romero]

For double-precision vertex inputs we need to measure them in dvec4
terms, and for single-precision vertex inputs we need to measure them in
vec4 terms.

For the later case, we use type_size_vec4() function. For the former
case, we had a wrong implementation based on type_size_vec4().

This commit introduces a proper type_size_dvec4() function, that we use
to measure vertex inputs.

Measuring double-precision vertex inputs as dvec4 is required because
ARB_vertex_attrib_64bit states that these uses the same number of
locations than the single-precision version. That is, two consecutives
dvec4 would be located in location "x" and location "x+1", not "x+2".
---
 src/compiler/nir/nir.h                         |  2 +-
 src/mesa/drivers/dri/i965/brw_fs.cpp           |  2 +-
 src/mesa/drivers/dri/i965/brw_shader.h         |  1 +
 src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp | 66 ++++++++++++++++++++------
 4 files changed, 54 insertions(+), 17 deletions(-)

diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
index a21a7bd..9681f7e 100644
--- a/src/compiler/nir/nir.h
+++ b/src/compiler/nir/nir.h
@@ -1,4 +1,4 @@
-/*
+o/*
  * Copyright © 2014 Connor Abbott
  *
  * Permission is hereby granted, free of charge, to any person obtaining a
diff --git a/src/mesa/drivers/dri/i965/brw_fs.cpp b/src/mesa/drivers/dri/i965/brw_fs.cpp
index 65b64b6..3028168 100644
--- a/src/mesa/drivers/dri/i965/brw_fs.cpp
+++ b/src/mesa/drivers/dri/i965/brw_fs.cpp
@@ -552,7 +552,7 @@ extern "C" int
 type_size_vs_input(const struct glsl_type *type)
 {
    if (type->is_double()) {
-      return type_size_vec4(type) / 2;
+      return type_size_dvec4(type);
    } else {
       return type_size_vec4(type);
    }
diff --git a/src/mesa/drivers/dri/i965/brw_shader.h b/src/mesa/drivers/dri/i965/brw_shader.h
index 60f3b5f..656dc89 100644
--- a/src/mesa/drivers/dri/i965/brw_shader.h
+++ b/src/mesa/drivers/dri/i965/brw_shader.h
@@ -293,6 +293,7 @@ struct gl_shader *brw_new_shader(struct gl_context *ctx, GLuint name, GLuint typ
 
 int type_size_scalar(const struct glsl_type *type);
 int type_size_vec4(const struct glsl_type *type);
+int type_size_dvec4(const struct glsl_type *type);
 int type_size_vec4_times_4(const struct glsl_type *type);
 int type_size_vs_input(const struct glsl_type *type);
 
diff --git a/src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp b/src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp
index f73d678..4b5dfe6 100644
--- a/src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp
+++ b/src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp
@@ -566,18 +566,12 @@ vec4_visitor::emit_pack_snorm_4x8(const dst_reg &dst, const src_reg &src0)
    emit(VEC4_OPCODE_PACK_BYTES, dst, bytes);
 }
 
-/**
- * Returns the minimum number of vec4 elements needed to pack a type.
- *
- * For simple types, it will return 1 (a single vec4); for matrices, the
- * number of columns; for array and struct, the sum of the vec4_size of
- * each of its elements; and for sampler and atomic, zero.
- *
- * This method is useful to calculate how much register space is needed to
- * store a particular type.
+/*
+ * Returns the minimum number of vec4 (as_vec4 == true) or dvec4 (as_vec4 ==
+ * false) elements needed to pack a type.
  */
-extern "C" int
-type_size_vec4(const struct glsl_type *type)
+static int
+type_size_xvec4(const struct glsl_type *type, bool as_vec4)
 {
    unsigned int i;
    int size;
@@ -590,7 +584,8 @@ type_size_vec4(const struct glsl_type *type)
    case GLSL_TYPE_DOUBLE:
       if (type->is_matrix()) {
          const glsl_type *col_type = type->column_type();
-         unsigned col_slots = col_type->is_dual_slot_double() ? 2 : 1;
+         unsigned col_slots =
+            (as_vec4 && col_type->is_dual_slot_double()) ? 2 : 1;
          return type->matrix_columns * col_slots;
       } else {
          /* Regardless of size of vector, it gets a vec4. This is bad
@@ -598,15 +593,15 @@ type_size_vec4(const struct glsl_type *type)
           * mess.  Hopefully a later pass over the code can pack scalars
           * down if appropriate.
           */
-         return type->is_dual_slot_double() ? 2 : 1;
+         return (as_vec4 && type->is_dual_slot_double()) ? 2 : 1;
       }
    case GLSL_TYPE_ARRAY:
       assert(type->length > 0);
-      return type_size_vec4(type->fields.array) * type->length;
+      return type_size_xvec4(type->fields.array, as_vec4) * type->length;
    case GLSL_TYPE_STRUCT:
       size = 0;
       for (i = 0; i < type->length; i++) {
-	 size += type_size_vec4(type->fields.structure[i].type);
+	 size += type_size_xvec4(type->fields.structure[i].type, as_vec4);
       }
       return size;
    case GLSL_TYPE_SUBROUTINE:
@@ -631,6 +626,47 @@ type_size_vec4(const struct glsl_type *type)
    return 0;
 }
 
+/**
+ * Returns the minimum number of vec4 elements needed to pack a type.
+ *
+ * For simple types, it will return 1 (a single vec4); for matrices, the
+ * number of columns; for array and struct, the sum of the vec4_size of
+ * each of its elements; and for sampler and atomic, zero.
+ *
+ * This method is useful to calculate how much register space is needed to
+ * store a particular type.
+ */
+extern "C" int
+type_size_vec4(const struct glsl_type *type)
+{
+   return type_size_xvec4(type, true);
+}
+
+/**
+ * Returns the minimum number of dvec4 elements needed to pack a type.
+ *
+ * For simple types, it will return 1 (a single dvec4); for matrices, the
+ * number of columns; for array and struct, the sum of the dvec4_size of
+ * each of its elements; and for sampler and atomic, zero.
+ *
+ * This method is useful to calculate how much register space is needed to
+ * store a particular type.
+ *
+ * Measuring double-precision vertex inputs as dvec4 is required because
+ * ARB_vertex_attrib_64bit states that these uses the same number of locations
+ * than the single-precision version. That is, two consecutives dvec4 would be
+ * located in location "x" and location "x+1", not "x+2".
+ *
+ * In order to map vec4/dvec4 vertex inputs in the proper ATTRs,
+ * remap_vs_attrs() will take in account both the location and also if the
+ * type fits in one or two vec4 slots.
+ */
+extern "C" int
+type_size_dvec4(const struct glsl_type *type)
+{
+   return type_size_xvec4(type, false);
+}
+
 src_reg::src_reg(class vec4_visitor *v, const struct glsl_type *type)
 {
    init();
-- 
2.5.5