d3d12: Lower cube images to 2D arrays via existing int cubemap lowering pass
authorJesse Natalie <jenatali@microsoft.com>
Wed, 29 Dec 2021 17:59:33 +0000 (09:59 -0800)
committerMarge Bot <emma+marge@anholt.net>
Fri, 7 Jan 2022 03:31:16 +0000 (03:31 +0000)
Note that the coordinates are not lowered, just types.

Reviewed-by: Sil Vilerino <sivileri@microsoft.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14342>

src/gallium/drivers/d3d12/d3d12_lower_int_cubemap_to_array.c

index 630da4f..819dc51 100644 (file)
 #include "nir_builtin_builder.h"
 
 static bool
-lower_int_cubmap_to_array_filter(const nir_instr *instr,
-                                 UNUSED const void *_options)
+type_needs_lowering(const struct glsl_type *type)
 {
-   if (instr->type != nir_instr_type_tex)
+   type = glsl_without_array(type);
+   if (!glsl_type_is_image(type) && !glsl_type_is_sampler(type))
       return false;
-
-   nir_tex_instr *tex = nir_instr_as_tex(instr);
-
-   if (tex->sampler_dim != GLSL_SAMPLER_DIM_CUBE)
+   if (glsl_get_sampler_dim(type) != GLSL_SAMPLER_DIM_CUBE)
       return false;
+   if (glsl_type_is_image(type))
+      return true;
+   return glsl_base_type_is_integer(glsl_get_sampler_result_type(type));
+}
 
-   switch (tex->op) {
-   case nir_texop_tex:
-   case nir_texop_txb:
-   case nir_texop_txd:
-   case nir_texop_txl:
-   case nir_texop_txs:
-   case nir_texop_lod:
-   case nir_texop_tg4:
-      break;
-   default:
-      return false;
+static bool
+lower_int_cubmap_to_array_filter(const nir_instr *instr,
+                                 UNUSED const void *_options)
+{
+   if (instr->type == nir_instr_type_intrinsic) {
+      nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
+      switch (intr->intrinsic) {
+      case nir_intrinsic_image_atomic_add:
+      case nir_intrinsic_image_atomic_and:
+      case nir_intrinsic_image_atomic_comp_swap:
+      case nir_intrinsic_image_atomic_dec_wrap:
+      case nir_intrinsic_image_atomic_exchange:
+      case nir_intrinsic_image_atomic_fadd:
+      case nir_intrinsic_image_atomic_fmax:
+      case nir_intrinsic_image_atomic_fmin:
+      case nir_intrinsic_image_atomic_imax:
+      case nir_intrinsic_image_atomic_imin:
+      case nir_intrinsic_image_atomic_inc_wrap:
+      case nir_intrinsic_image_atomic_or:
+      case nir_intrinsic_image_atomic_umax:
+      case nir_intrinsic_image_atomic_umin:
+      case nir_intrinsic_image_atomic_xor:
+      case nir_intrinsic_image_load:
+      case nir_intrinsic_image_size:
+      case nir_intrinsic_image_store:
+      case nir_intrinsic_image_deref_atomic_add:
+      case nir_intrinsic_image_deref_atomic_and:
+      case nir_intrinsic_image_deref_atomic_comp_swap:
+      case nir_intrinsic_image_deref_atomic_dec_wrap:
+      case nir_intrinsic_image_deref_atomic_exchange:
+      case nir_intrinsic_image_deref_atomic_fadd:
+      case nir_intrinsic_image_deref_atomic_fmax:
+      case nir_intrinsic_image_deref_atomic_fmin:
+      case nir_intrinsic_image_deref_atomic_imax:
+      case nir_intrinsic_image_deref_atomic_imin:
+      case nir_intrinsic_image_deref_atomic_inc_wrap:
+      case nir_intrinsic_image_deref_atomic_or:
+      case nir_intrinsic_image_deref_atomic_umax:
+      case nir_intrinsic_image_deref_atomic_umin:
+      case nir_intrinsic_image_deref_atomic_xor:
+      case nir_intrinsic_image_deref_load:
+      case nir_intrinsic_image_deref_size:
+      case nir_intrinsic_image_deref_store:
+         return nir_intrinsic_image_dim(intr) == GLSL_SAMPLER_DIM_CUBE;
+      default:
+         return false;
+      }
+   } else if (instr->type == nir_instr_type_deref) {
+      nir_deref_instr *deref = nir_instr_as_deref(instr);
+      return type_needs_lowering(deref->type);
+   } else if (instr->type == nir_instr_type_tex) {
+      nir_tex_instr *tex = nir_instr_as_tex(instr);
+
+      if (tex->sampler_dim != GLSL_SAMPLER_DIM_CUBE)
+         return false;
+
+      switch (tex->op) {
+      case nir_texop_tex:
+      case nir_texop_txb:
+      case nir_texop_txd:
+      case nir_texop_txl:
+      case nir_texop_txs:
+      case nir_texop_lod:
+      case nir_texop_tg4:
+         break;
+      default:
+         return false;
+      }
+
+      int sampler_deref = nir_tex_instr_src_index(tex, nir_tex_src_sampler_deref);
+      assert(sampler_deref >= 0);
+      nir_deref_instr *deref = nir_instr_as_deref(tex->src[sampler_deref].src.ssa->parent_instr);
+      nir_variable *cube = nir_deref_instr_get_variable(deref);
+      return glsl_base_type_is_integer(glsl_get_sampler_result_type(cube->type));
    }
 
-   int sampler_deref = nir_tex_instr_src_index(tex, nir_tex_src_sampler_deref);
-   assert(sampler_deref >= 0);
-   nir_deref_instr *deref = nir_instr_as_deref(tex->src[sampler_deref].src.ssa->parent_instr);
-   nir_variable *cube = nir_deref_instr_get_variable(deref);
-   return glsl_base_type_is_integer(glsl_get_sampler_result_type(cube->type));
+   return false;
 }
 
 typedef struct {
+   bool image;
    nir_ssa_def *rx;
    nir_ssa_def *ry;
    nir_ssa_def *rz;
@@ -84,7 +145,9 @@ evaluate_face_x(nir_builder *b, coord_t *coord)
    if (coord->array)
       face = nir_fadd(b, face, coord->array);
 
-   return nir_vec3(b, x,y, face);
+   return coord->image ?
+      nir_vec4(b, x,y, face, nir_ssa_undef(b, 1, 32)) :
+      nir_vec3(b, x,y, face);
 }
 
 static nir_ssa_def *
@@ -100,8 +163,10 @@ evaluate_face_y(nir_builder *b, coord_t *coord)
 
    if (coord->array)
       face = nir_fadd(b, face, coord->array);
-
-   return nir_vec3(b, x,y, face);
+   
+   return coord->image ?
+      nir_vec4(b, x,y, face, nir_ssa_undef(b, 1, 32)) :
+      nir_vec3(b, x,y, face);
 }
 
 static nir_ssa_def *
@@ -117,8 +182,10 @@ evaluate_face_z(nir_builder *b, coord_t *coord)
 
    if (coord->array)
       face = nir_fadd(b, face, coord->array);
-
-   return nir_vec3(b, x,y, face);
+   
+   return coord->image ?
+      nir_vec4(b, x,y, face, nir_ssa_undef(b, 1, 32)) :
+      nir_vec3(b, x,y, face);
 }
 
 static nir_ssa_def *
@@ -311,15 +378,10 @@ handle_cube_gather(nir_builder *b, nir_tex_instr *tex, nir_ssa_def *coord)
 }
 
 static nir_ssa_def *
-lower_cube_sample(nir_builder *b, nir_tex_instr *tex)
+lower_cube_coords(nir_builder *b, nir_ssa_def *coord, bool is_array, bool is_image)
 {
-   int coord_index = nir_tex_instr_src_index(tex, nir_tex_src_coord);
-   assert(coord_index >= 0);
-
-   /* Evaluate the face and the xy coordinates for a 2D tex op */
-   nir_ssa_def *coord = tex->src[coord_index].src.ssa;
-
    coord_t coords;
+   coords.image = is_image;
    coords.rx = nir_channel(b, coord, 0);
    coords.ry = nir_channel(b, coord, 1);
    coords.rz = nir_channel(b, coord, 2);
@@ -327,7 +389,7 @@ lower_cube_sample(nir_builder *b, nir_tex_instr *tex)
    coords.ary = nir_fabs(b, coords.ry);
    coords.arz = nir_fabs(b, coords.rz);
    coords.array = NULL;
-   if (tex->is_array)
+   if (is_array)
       coords.array = nir_fmul(b, nir_channel(b, coord, 3), nir_imm_float(b, 6.0f));
 
    nir_ssa_def *use_face_x = nir_iand(b,
@@ -355,6 +417,19 @@ lower_cube_sample(nir_builder *b, nir_tex_instr *tex)
    // This contains in xy the normalized sample coordinates, and in z the face index
    nir_ssa_def *coord_and_face = nir_if_phi(b, face_x_coord, face_y_or_z_coord);
 
+   return coord_and_face;
+}
+
+static nir_ssa_def *
+lower_cube_sample(nir_builder *b, nir_tex_instr *tex)
+{
+   int coord_index = nir_tex_instr_src_index(tex, nir_tex_src_coord);
+   assert(coord_index >= 0);
+
+   /* Evaluate the face and the xy coordinates for a 2D tex op */
+   nir_ssa_def *coord = tex->src[coord_index].src.ssa;
+   nir_ssa_def *coord_and_face = lower_cube_coords(b, coord, tex->is_array, false);
+
    if (tex->op == nir_texop_tg4)
       return handle_cube_gather(b, tex, coord_and_face);
    else
@@ -362,6 +437,16 @@ lower_cube_sample(nir_builder *b, nir_tex_instr *tex)
 }
 
 static nir_ssa_def *
+lower_cube_image_load_store_atomic(nir_builder *b, nir_intrinsic_instr *intr)
+{
+   b->cursor = nir_before_instr(&intr->instr);
+   nir_intrinsic_set_image_array(intr, true);
+   nir_intrinsic_set_image_dim(intr, GLSL_SAMPLER_DIM_2D);
+
+   return NIR_LOWER_INSTR_PROGRESS;
+}
+
+static nir_ssa_def *
 lower_cube_txs(nir_builder *b, nir_tex_instr *tex)
 {
    b->cursor = nir_after_instr(&tex->instr);
@@ -375,8 +460,22 @@ lower_cube_txs(nir_builder *b, nir_tex_instr *tex)
                       cube_array_dim);
 }
 
+static nir_ssa_def *
+lower_cube_image_size(nir_builder *b, nir_intrinsic_instr *intr)
+{
+   b->cursor = nir_after_instr(&intr->instr);
+   if (!nir_intrinsic_image_array(intr))
+      return nir_channels(b, &intr->dest.ssa, 3);
+
+   nir_ssa_def *array_dim = nir_channel(b, &intr->dest.ssa, 2);
+   nir_ssa_def *cube_array_dim = nir_idiv(b, array_dim, nir_imm_int(b, 6));
+   return nir_vec3(b, nir_channel(b, &intr->dest.ssa, 0),
+                      nir_channel(b, &intr->dest.ssa, 1),
+                      cube_array_dim);
+}
+
 static const struct glsl_type *
-make_2darray_from_cubemap(const struct glsl_type *type)
+make_2darray_sampler_from_cubemap(const struct glsl_type *type)
 {
    return  glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_CUBE ?
             glsl_sampler_type(
@@ -386,31 +485,31 @@ make_2darray_from_cubemap(const struct glsl_type *type)
 }
 
 static const struct glsl_type *
-make_2darray_from_cubemap_with_array(const struct glsl_type *type)
+make_2darray_image_from_cubemap(const struct glsl_type *type)
+{
+   return  glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_CUBE ?
+            glsl_image_type(
+               GLSL_SAMPLER_DIM_2D,
+               true,
+               glsl_get_sampler_result_type(type)) : type;
+}
+
+static const struct glsl_type *
+make_2darray_from_cubemap_with_array(const struct glsl_type *type, bool is_image)
 {
    if (glsl_type_is_array(type)) {
       const struct glsl_type *new_type = glsl_without_array(type);
-      return new_type != type ? glsl_array_type(make_2darray_from_cubemap(glsl_without_array(type)),
+      return new_type != type ? glsl_array_type(make_2darray_from_cubemap_with_array(glsl_without_array(type), is_image),
                                                 glsl_get_length(type), 0) : type;
-   } else
-      return make_2darray_from_cubemap(type);
+   } else if (is_image)
+      return make_2darray_image_from_cubemap(type);
+   else
+      return make_2darray_sampler_from_cubemap(type);
 }
 
 static nir_ssa_def *
-lower_int_cubmap_to_array_impl(nir_builder *b, nir_instr *instr,
-                               UNUSED void *_options)
+lower_int_cubemap_to_array_tex(nir_builder *b, nir_tex_instr *tex)
 {
-   nir_tex_instr *tex = nir_instr_as_tex(instr);
-
-   int sampler_index = nir_tex_instr_src_index(tex, nir_tex_src_sampler_deref);
-   assert(sampler_index >= 0);
-
-   nir_deref_instr *sampler_deref = nir_instr_as_deref(tex->src[sampler_index].src.ssa->parent_instr);
-   nir_variable *sampler = nir_deref_instr_get_variable(sampler_deref);
-
-   sampler->type = make_2darray_from_cubemap_with_array(sampler->type);
-   sampler_deref->type = sampler->type;
-
    switch (tex->op) {
    case nir_texop_tex:
    case nir_texop_txb:
@@ -426,6 +525,38 @@ lower_int_cubmap_to_array_impl(nir_builder *b, nir_instr *instr,
    }
 }
 
+static nir_ssa_def *
+lower_cube_intrinsic(nir_builder *b, nir_intrinsic_instr *intr)
+{
+   if (intr->intrinsic == nir_intrinsic_image_size ||
+      intr->intrinsic == nir_intrinsic_image_deref_size)
+      return lower_cube_image_size(b, intr);
+   else
+      return lower_cube_image_load_store_atomic(b, intr);
+}
+
+static nir_ssa_def *
+lower_cube_deref(nir_builder *b, nir_deref_instr *deref)
+{
+   deref->type = make_2darray_from_cubemap_with_array(
+      deref->type,
+      glsl_type_is_image(glsl_without_array(deref->type)));
+   return NIR_LOWER_INSTR_PROGRESS;
+}
+
+static nir_ssa_def *
+lower_int_cubmap_to_array_impl(nir_builder *b, nir_instr *instr,
+                               UNUSED void *_options)
+{
+   if (instr->type == nir_instr_type_tex)
+      return lower_int_cubemap_to_array_tex(b, nir_instr_as_tex(instr));
+   else if (instr->type == nir_instr_type_intrinsic)
+      return lower_cube_intrinsic(b, nir_instr_as_intrinsic(instr));
+   else if (instr->type == nir_instr_type_deref)
+      return lower_cube_deref(b, nir_instr_as_deref(instr));
+   return NULL;
+}
+
 bool
 d3d12_lower_int_cubmap_to_array(nir_shader *s)
 {
@@ -436,13 +567,11 @@ d3d12_lower_int_cubmap_to_array(nir_shader *s)
                                        NULL);
 
    if (result) {
-      nir_foreach_variable_with_modes_safe(var, s, nir_var_uniform) {
-         if (glsl_type_is_sampler(var->type)) {
-            if (glsl_get_sampler_dim(var->type) == GLSL_SAMPLER_DIM_CUBE &&
-                (glsl_base_type_is_integer(glsl_get_sampler_result_type(var->type)))) {
-               var->type = make_2darray_from_cubemap_with_array(var->type);
-            }
-         }
+      nir_foreach_variable_with_modes_safe(var, s, nir_var_uniform | nir_var_image) {
+         if (!type_needs_lowering(var->type))
+            continue;
+         bool is_image = glsl_type_is_image(glsl_without_array(var->type));
+         var->type = make_2darray_from_cubemap_with_array(var->type, is_image);
       }
    }
    return result;