Fix T73133: UDIM texture count in Eevee is limited by OpenGL

Based on @fclem's suggestion in D6421, this commit implements support for
storing all tiles of a UDIM texture in a single 2D array texture on the GPU.

Previously, Eevee was binding one OpenGL texture per tile, quickly running
into hardware limits with nontrivial UDIM texture sets.
Workbench meanwhile had no UDIM support at all, as reusing the per-tile
approach would require splitting the mesh by tile as well as texture.

With this commit, both Workbench as well as Eevee now support huge numbers
of tiles, with the eventual limits being GPU memory and ultimately
GL_MAX_ARRAY_TEXTURE_LAYERS, which tends to be in the 1000s on modern GPUs.

Initially my plan was to have one array texture per unique size, but managing
the different textures and keeping everything consistent ended up being way
too complex.

Therefore, we now use a simpler version that allocates a texture that
is large enough to fit the largest tile and then packs all tiles into as many
layers as necessary.

As a result, each UDIM texture only binds two textures (one for the actual
images, one for metadata) regardless of how many tiles are used.

Note that this rolls back per-tile GPUTextures, meaning that we again have
per-Image GPUTextures like we did before the original UDIM commit,
but now with four instead of two types.

Reviewed By: fclem

Differential Revision: https://developer.blender.org/D6456

6 files changed, 572 insertions, 142 deletions

diff --git a/source/blender/gpu/GPU_material.h b/source/blender/gpu/GPU_material.h
index a5363c7a42c..8c166ed6b64 100644
--- a/source/blender/gpu/GPU_material.h
+++ b/source/blender/gpu/GPU_material.h
@@ -71,12 +71,13 @@ typedef enum eGPUType {
   /* Values not in GPU_DATATYPE_STR */
   GPU_TEX1D_ARRAY = 1001,
   GPU_TEX2D = 1002,
-  GPU_TEX3D = 1003,
-  GPU_SHADOW2D = 1004,
-  GPU_TEXCUBE = 1005,
+  GPU_TEX2D_ARRAY = 1003,
+  GPU_TEX3D = 1004,
+  GPU_SHADOW2D = 1005,
+  GPU_TEXCUBE = 1006,
 
   /* GLSL Struct types */
-  GPU_CLOSURE = 1006,
+  GPU_CLOSURE = 1007,
 
   /* Opengl Attributes */
   GPU_ATTR = 3001,
@@ -142,7 +143,8 @@ typedef enum eGPUMaterialStatus {
 GPUNodeLink *GPU_attribute(CustomDataType type, const char *name);
 GPUNodeLink *GPU_constant(float *num);
 GPUNodeLink *GPU_uniform(float *num);
-GPUNodeLink *GPU_image(struct Image *ima, struct ImageUser *iuser, int tile);
+GPUNodeLink *GPU_image(struct Image *ima, struct ImageUser *iuser);
+GPUNodeLink *GPU_image_tilemap(struct Image *ima, struct ImageUser *iuser);
 GPUNodeLink *GPU_color_band(GPUMaterial *mat, int size, float *pixels, float *layer);
 GPUNodeLink *GPU_builtin(eGPUBuiltin builtin);
 
diff --git a/source/blender/gpu/intern/gpu_codegen.c b/source/blender/gpu/intern/gpu_codegen.c
index ca804a26acd..1da7274b2cd 100644
--- a/source/blender/gpu/intern/gpu_codegen.c
+++ b/source/blender/gpu/intern/gpu_codegen.c
@@ -279,6 +279,9 @@ static void gpu_parse_material_library(GHash *hash, GPUMaterialLibrary *library)
       if (!type && gpu_str_prefix(code, "sampler1DArray")) {
         type = GPU_TEX1D_ARRAY;
       }
+      if (!type && gpu_str_prefix(code, "sampler2DArray")) {
+        type = GPU_TEX2D_ARRAY;
+      }
       if (!type && gpu_str_prefix(code, "sampler2D")) {
         type = GPU_TEX2D;
       }
@@ -618,7 +621,7 @@ static void codegen_set_unique_ids(ListBase *nodes)
         input->bindtex = false;
         if (input->ima) {
           /* input is texture from image */
-          codegen_set_texid(bindhash, input, &texid, input->ima, input->image_tile);
+          codegen_set_texid(bindhash, input, &texid, input->ima, input->type);
         }
         else if (input->coba) {
           /* input is color band texture, check coba pointer */
@@ -657,10 +660,18 @@ static int codegen_process_uniforms_functions(GPUMaterial *material, DynStr *ds,
       if (input->source == GPU_SOURCE_TEX) {
         /* create exactly one sampler for each texture */
         if (codegen_input_has_texture(input) && input->bindtex) {
-          BLI_dynstr_appendf(ds,
-                             "uniform %s samp%d;\n",
-                             (input->coba) ? "sampler1DArray" : "sampler2D",
-                             input->texid);
+          const char *type;
+          if (input->coba || input->type == GPU_TEX1D_ARRAY) {
+            type = "sampler1DArray";
+          }
+          else if (input->type == GPU_TEX2D_ARRAY) {
+            type = "sampler2DArray";
+          }
+          else {
+            BLI_assert(input->type == GPU_TEX2D);
+            type = "sampler2D";
+          }
+          BLI_dynstr_appendf(ds, "uniform %s samp%d;\n", type, input->texid);
         }
       }
       else if (input->source == GPU_SOURCE_BUILTIN) {
@@ -1544,10 +1555,10 @@ static void gpu_node_input_link(GPUNode *node, GPUNodeLink *link, const eGPUType
       input->coba = link->coba;
       break;
     case GPU_NODE_LINK_IMAGE_BLENDER:
+    case GPU_NODE_LINK_IMAGE_TILEMAP:
       input->source = GPU_SOURCE_TEX;
       input->ima = link->ima;
       input->iuser = link->iuser;
-      input->image_tile = link->image_tile;
       break;
     case GPU_NODE_LINK_ATTR:
       input->source = GPU_SOURCE_ATTR;
@@ -1792,13 +1803,12 @@ GPUNodeLink *GPU_uniform(float *num)
   return link;
 }
 
-GPUNodeLink *GPU_image(Image *ima, ImageUser *iuser, int tile)
+GPUNodeLink *GPU_image(Image *ima, ImageUser *iuser)
 {
   GPUNodeLink *link = GPU_node_link_create();
   link->link_type = GPU_NODE_LINK_IMAGE_BLENDER;
   link->ima = ima;
   link->iuser = iuser;
-  link->image_tile = tile;
   return link;
 }
 
diff --git a/source/blender/gpu/intern/gpu_codegen.h b/source/blender/gpu/intern/gpu_codegen.h
index 0e6982c603e..f8e1b76580f 100644
--- a/source/blender/gpu/intern/gpu_codegen.h
+++ b/source/blender/gpu/intern/gpu_codegen.h
@@ -59,6 +59,7 @@ typedef enum {
   GPU_NODE_LINK_COLORBAND,
   GPU_NODE_LINK_CONSTANT,
   GPU_NODE_LINK_IMAGE_BLENDER,
+  GPU_NODE_LINK_IMAGE_TILEMAP,
   GPU_NODE_LINK_OUTPUT,
   GPU_NODE_LINK_UNIFORM,
 } GPUNodeLinkType;
@@ -95,11 +96,10 @@ struct GPUNodeLink {
       const char *attr_name;
       CustomDataType attr_type;
     };
-    /* GPU_NODE_LINK_IMAGE_BLENDER */
+    /* GPU_NODE_LINK_IMAGE_BLENDER | GPU_NODE_LINK_IMAGE_TILEMAP */
     struct {
       struct Image *ima;
       struct ImageUser *iuser;
-      int image_tile;
     };
   };
 };
@@ -139,7 +139,6 @@ typedef struct GPUInput {
       struct ImageUser *iuser;  /* image user */
       bool bindtex;             /* input is responsible for binding the texture? */
       int texid;                /* number for multitexture, starting from zero */
-      int image_tile;           /* image tile */
       eGPUType textype;         /* texture type (2D, 1D Array ...) */
     };
     /* GPU_SOURCE_ATTR */
diff --git a/source/blender/gpu/intern/gpu_draw.c b/source/blender/gpu/intern/gpu_draw.c
index 95738bb1a95..d2008680098 100644
--- a/source/blender/gpu/intern/gpu_draw.c
+++ b/source/blender/gpu/intern/gpu_draw.c
@@ -31,6 +31,7 @@
 #include <string.h>
 
 #include "BLI_blenlib.h"
+#include "BLI_boxpack_2d.h"
 #include "BLI_linklist.h"
 #include "BLI_math.h"
 #include "BLI_threads.h"
@@ -71,7 +72,7 @@ static bool is_power_of_2_resolution(int w, int h)
 
 static bool is_over_resolution_limit(GLenum textarget, int w, int h)
 {
-  int size = (textarget == GL_TEXTURE_2D) ? GPU_max_texture_size() : GPU_max_cube_map_size();
+  int size = (textarget == GL_TEXTURE_CUBE_MAP) ? GPU_max_cube_map_size() : GPU_max_texture_size();
   int reslimit = (U.glreslimit != 0) ? min_ii(U.glreslimit, size) : size;
 
   return (w > reslimit || h > reslimit);
@@ -179,18 +180,294 @@ float GPU_get_anisotropic(void)
 
 /* Set OpenGL state for an MTFace */
 
-static GPUTexture **gpu_get_tile_gputexture(ImageTile *tile, GLenum textarget)
+static GPUTexture **gpu_get_image_gputexture(Image *ima, GLenum textarget)
 {
   if (textarget == GL_TEXTURE_2D) {
-    return &tile->gputexture[TEXTARGET_TEXTURE_2D];
+    return &ima->gputexture[TEXTARGET_TEXTURE_2D];
   }
   else if (textarget == GL_TEXTURE_CUBE_MAP) {
-    return &tile->gputexture[TEXTARGET_TEXTURE_CUBE_MAP];
+    return &ima->gputexture[TEXTARGET_TEXTURE_CUBE_MAP];
+  }
+  else if (textarget == GL_TEXTURE_2D_ARRAY) {
+    return &ima->gputexture[TEXTARGET_TEXTURE_2D_ARRAY];
+  }
+  else if (textarget == GL_TEXTURE_1D_ARRAY) {
+    return &ima->gputexture[TEXTARGET_TEXTURE_TILE_MAPPING];
   }
 
   return NULL;
 }
 
+static uint gpu_texture_create_tile_mapping(Image *ima)
+{
+  GPUTexture *tilearray = ima->gputexture[TEXTARGET_TEXTURE_2D_ARRAY];
+  if (tilearray == NULL) {
+    return 0;
+  }
+
+  float array_w = GPU_texture_width(tilearray);
+  float array_h = GPU_texture_height(tilearray);
+
+  ImageTile *last_tile = ima->tiles.last;
+  /* Tiles are sorted by number. */
+  int max_tile = last_tile->tile_number - 1001;
+
+  /* create image */
+  int bindcode;
+  glGenTextures(1, (GLuint *)&bindcode);
+  glBindTexture(GL_TEXTURE_1D_ARRAY, bindcode);
+
+  int width = max_tile + 1;
+  float *data = MEM_callocN(width * 8 * sizeof(float), __func__);
+  for (int i = 0; i < width; i++) {
+    data[4 * i] = -1.0f;
+  }
+  LISTBASE_FOREACH (ImageTile *, tile, &ima->tiles) {
+    int i = tile->tile_number - 1001;
+    data[4 * i] = tile->runtime.tilearray_layer;
+
+    float *tile_info = &data[4 * width + 4 * i];
+    tile_info[0] = tile->runtime.tilearray_offset[0] / array_w;
+    tile_info[1] = tile->runtime.tilearray_offset[1] / array_h;
+    tile_info[2] = tile->runtime.tilearray_size[0] / array_w;
+    tile_info[3] = tile->runtime.tilearray_size[1] / array_h;
+  }
+
+  glTexImage2D(GL_TEXTURE_1D_ARRAY, 0, GL_RGBA32F, width, 2, 0, GL_RGBA, GL_FLOAT, data);
+  MEM_freeN(data);
+
+  glTexParameteri(GL_TEXTURE_1D_ARRAY, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+  glTexParameteri(GL_TEXTURE_1D_ARRAY, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+
+  glBindTexture(GL_TEXTURE_1D_ARRAY, 0);
+
+  return bindcode;
+}
+
+typedef struct PackTile {
+  FixedSizeBoxPack boxpack;
+  ImageTile *tile;
+  float pack_score;
+} PackTile;
+
+static int compare_packtile(const void *a, const void *b)
+{
+  const PackTile *tile_a = a;
+  const PackTile *tile_b = b;
+
+  return tile_a->pack_score < tile_b->pack_score;
+}
+
+static uint gpu_texture_create_tile_array(Image *ima, ImBuf *main_ibuf)
+{
+  int arraywidth = 0, arrayheight = 0;
+
+  ListBase boxes = {NULL};
+
+  LISTBASE_FOREACH (ImageTile *, tile, &ima->tiles) {
+    ImageUser iuser;
+    BKE_imageuser_default(&iuser);
+    iuser.tile = tile->tile_number;
+    ImBuf *ibuf = BKE_image_acquire_ibuf(ima, &iuser, NULL);
+
+    if (ibuf) {
+      PackTile *packtile = MEM_callocN(sizeof(PackTile), __func__);
+      packtile->tile = tile;
+      packtile->boxpack.w = ibuf->x;
+      packtile->boxpack.h = ibuf->y;
+
+      if (is_over_resolution_limit(
+              GL_TEXTURE_2D_ARRAY, packtile->boxpack.w, packtile->boxpack.h)) {
+        packtile->boxpack.w = smaller_power_of_2_limit(packtile->boxpack.w);
+        packtile->boxpack.h = smaller_power_of_2_limit(packtile->boxpack.h);
+      }
+      arraywidth = max_ii(arraywidth, packtile->boxpack.w);
+      arrayheight = max_ii(arrayheight, packtile->boxpack.h);
+
+      /* We sort the tiles by decreasing size, with an additional penalty term
+       * for high aspect ratios. This improves packing efficiency. */
+      float w = packtile->boxpack.w, h = packtile->boxpack.h;
+      packtile->pack_score = max_ff(w, h) / min_ff(w, h) * w * h;
+
+      BKE_image_release_ibuf(ima, ibuf, NULL);
+      BLI_addtail(&boxes, packtile);
+    }
+  }
+
+  BLI_assert(arraywidth > 0 && arrayheight > 0);
+
+  BLI_listbase_sort(&boxes, compare_packtile);
+  int arraylayers = 0;
+  /* Keep adding layers until all tiles are packed. */
+  while (boxes.first != NULL) {
+    ListBase packed = {NULL};
+    BLI_box_pack_2d_fixedarea(&boxes, arraywidth, arrayheight, &packed);
+    BLI_assert(packed.first != NULL);
+
+    LISTBASE_FOREACH (PackTile *, packtile, &packed) {
+      ImageTile *tile = packtile->tile;
+      int *tileoffset = tile->runtime.tilearray_offset;
+      int *tilesize = tile->runtime.tilearray_size;
+
+      tileoffset[0] = packtile->boxpack.x;
+      tileoffset[1] = packtile->boxpack.y;
+      tilesize[0] = packtile->boxpack.w;
+      tilesize[1] = packtile->boxpack.h;
+      tile->runtime.tilearray_layer = arraylayers;
+    }
+
+    BLI_freelistN(&packed);
+    arraylayers++;
+  }
+
+  /* create image */
+  int bindcode;
+  glGenTextures(1, (GLuint *)&bindcode);
+  glBindTexture(GL_TEXTURE_2D_ARRAY, bindcode);
+
+  GLenum data_type, internal_format;
+  if (main_ibuf->rect_float) {
+    data_type = GL_FLOAT;
+    internal_format = GL_RGBA16F;
+  }
+  else {
+    data_type = GL_UNSIGNED_BYTE;
+    internal_format = GL_RGBA8;
+    if (!IMB_colormanagement_space_is_data(main_ibuf->rect_colorspace) &&
+        !IMB_colormanagement_space_is_scene_linear(main_ibuf->rect_colorspace)) {
+      internal_format = GL_SRGB8_ALPHA8;
+    }
+  }
+
+  glTexImage3D(GL_TEXTURE_2D_ARRAY,
+               0,
+               internal_format,
+               arraywidth,
+               arrayheight,
+               arraylayers,
+               0,
+               GL_RGBA,
+               data_type,
+               NULL);
+
+  LISTBASE_FOREACH (ImageTile *, tile, &ima->tiles) {
+    int tilelayer = tile->runtime.tilearray_layer;
+    int *tileoffset = tile->runtime.tilearray_offset;
+    int *tilesize = tile->runtime.tilearray_size;
+
+    if (tilesize[0] == 0 || tilesize[1] == 0) {
+      continue;
+    }
+
+    ImageUser iuser;
+    BKE_imageuser_default(&iuser);
+    iuser.tile = tile->tile_number;
+    ImBuf *ibuf = BKE_image_acquire_ibuf(ima, &iuser, NULL);
+    BLI_assert(ibuf != NULL);
+
+    bool needs_scale = (ibuf->x != tilesize[0] || ibuf->y != tilesize[1]);
+
+    ImBuf *scale_ibuf = NULL;
+    if (ibuf->rect_float) {
+      float *rect_float = ibuf->rect_float;
+
+      const bool store_premultiplied = ima->alpha_mode != IMA_ALPHA_STRAIGHT;
+      if (ibuf->channels != 4 || !store_premultiplied) {
+        rect_float = MEM_mallocN(sizeof(float) * 4 * ibuf->x * ibuf->y, __func__);
+        IMB_colormanagement_imbuf_to_float_texture(
+            rect_float, 0, 0, ibuf->x, ibuf->y, ibuf, store_premultiplied);
+      }
+
+      float *pixeldata = rect_float;
+      if (needs_scale) {
+        scale_ibuf = IMB_allocFromBuffer(NULL, rect_float, ibuf->x, ibuf->y, 4);
+        IMB_scaleImBuf(scale_ibuf, tilesize[0], tilesize[1]);
+        pixeldata = scale_ibuf->rect_float;
+      }
+
+      glTexSubImage3D(GL_TEXTURE_2D_ARRAY,
+                      0,
+                      tileoffset[0],
+                      tileoffset[1],
+                      tilelayer,
+                      tilesize[0],
+                      tilesize[1],
+                      1,
+                      GL_RGBA,
+                      GL_FLOAT,
+                      pixeldata);
+
+      if (rect_float != ibuf->rect_float) {
+        MEM_freeN(rect_float);
+      }
+    }
+    else {
+      unsigned int *rect = ibuf->rect;
+
+      if (!IMB_colormanagement_space_is_data(ibuf->rect_colorspace)) {
+        rect = MEM_mallocN(sizeof(uchar) * 4 * ibuf->x * ibuf->y, __func__);
+        IMB_colormanagement_imbuf_to_byte_texture((uchar *)rect,
+                                                  0,
+                                                  0,
+                                                  ibuf->x,
+                                                  ibuf->y,
+                                                  ibuf,
+                                                  internal_format == GL_SRGB8_ALPHA8,
+                                                  ima->alpha_mode == IMA_ALPHA_PREMUL);
+      }
+
+      unsigned int *pixeldata = rect;
+      if (needs_scale) {
+        scale_ibuf = IMB_allocFromBuffer(rect, NULL, ibuf->x, ibuf->y, 4);
+        IMB_scaleImBuf(scale_ibuf, tilesize[0], tilesize[1]);
+        pixeldata = scale_ibuf->rect;
+      }
+      glTexSubImage3D(GL_TEXTURE_2D_ARRAY,
+                      0,
+                      tileoffset[0],
+                      tileoffset[1],
+                      tilelayer,
+                      tilesize[0],
+                      tilesize[1],
+                      1,
+                      GL_RGBA,
+                      GL_UNSIGNED_BYTE,
+                      pixeldata);
+
+      if (rect != ibuf->rect) {
+        MEM_freeN(rect);
+      }
+    }
+
+    if (scale_ibuf != NULL) {
+      IMB_freeImBuf(scale_ibuf);
+    }
+
+    BKE_image_release_ibuf(ima, ibuf, NULL);
+  }
+
+  glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MAG_FILTER, gpu_get_mipmap_filter(1));
+
+  if (GPU_get_mipmap()) {
+    glGenerateMipmap(GL_TEXTURE_2D_ARRAY);
+    glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MIN_FILTER, gpu_get_mipmap_filter(0));
+    if (ima) {
+      ima->gpuflag |= IMA_GPU_MIPMAP_COMPLETE;
+    }
+  }
+  else {
+    glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
+  }
+
+  if (GLEW_EXT_texture_filter_anisotropic) {
+    glTexParameterf(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MAX_ANISOTROPY_EXT, GPU_get_anisotropic());
+  }
+
+  glBindTexture(GL_TEXTURE_2D_ARRAY, 0);
+
+  return bindcode;
+}
+
 static uint gpu_texture_create_from_ibuf(Image *ima, ImBuf *ibuf, int textarget)
 {
   uint bindcode = 0;
@@ -305,48 +582,105 @@ static GPUTexture **gpu_get_movieclip_gputexture(MovieClip *clip,
   return NULL;
 }
 
-static void gpu_texture_update_scaled(
-    uchar *rect, float *rect_float, int full_w, int full_h, int x, int y, int w, int h)
+static ImBuf *update_do_scale(uchar *rect,
+                              float *rect_float,
+                              int *x,
+                              int *y,
+                              int *w,
+                              int *h,
+                              int limit_w,
+                              int limit_h,
+                              int full_w,
+                              int full_h)
 {
   /* Partial update with scaling. */
-  int limit_w = smaller_power_of_2_limit(full_w);
-  int limit_h = smaller_power_of_2_limit(full_h);
   float xratio = limit_w / (float)full_w;
   float yratio = limit_h / (float)full_h;
 
+  int part_w = *w, part_h = *h;
+
   /* Find sub coordinates in scaled image. Take ceiling because we will be
    * losing 1 pixel due to rounding errors in x,y. */
-  int sub_x = x * xratio;
-  int sub_y = y * yratio;
-  int sub_w = (int)ceil(xratio * w);
-  int sub_h = (int)ceil(yratio * h);
+  *x *= xratio;
+  *y *= yratio;
+  *w = (int)ceil(xratio * (*w));
+  *h = (int)ceil(yratio * (*h));
 
   /* ...but take back if we are over the limit! */
-  if (sub_w + sub_x > limit_w) {
-    sub_w--;
+  if (*x + *w > limit_w) {
+    (*w)--;
   }
-  if (sub_h + sub_y > limit_h) {
-    sub_h--;
+  if (*y + *h > limit_h) {
+    (*h)--;
   }
 
   /* Scale pixels. */
-  ImBuf *ibuf = IMB_allocFromBuffer((uint *)rect, rect_float, w, h, 4);
-  IMB_scaleImBuf(ibuf, sub_w, sub_h);
+  ImBuf *ibuf = IMB_allocFromBuffer((uint *)rect, rect_float, part_w, part_h, 4);
+  IMB_scaleImBuf(ibuf, *w, *h);
+
+  return ibuf;
+}
+
+static void gpu_texture_update_scaled_array(uchar *rect,
+                                            float *rect_float,
+                                            int full_w,
+                                            int full_h,
+                                            int x,
+                                            int y,
+                                            int layer,
+                                            const int *tile_offset,
+                                            const int *tile_size,
+                                            int w,
+                                            int h)
+{
+  ImBuf *ibuf = update_do_scale(
+      rect, rect_float, &x, &y, &w, &h, tile_size[0], tile_size[1], full_w, full_h);
+
+  /* Shift to account for tile packing. */
+  x += tile_offset[0];
+  y += tile_offset[1];
 
   if (ibuf->rect_float) {
-    glTexSubImage2D(
-        GL_TEXTURE_2D, 0, sub_x, sub_y, sub_w, sub_h, GL_RGBA, GL_FLOAT, ibuf->rect_float);
+    glTexSubImage3D(
+        GL_TEXTURE_2D_ARRAY, 0, x, y, layer, w, h, 1, GL_RGBA, GL_FLOAT, ibuf->rect_float);
   }
   else {
-    glTexSubImage2D(
-        GL_TEXTURE_2D, 0, sub_x, sub_y, sub_w, sub_h, GL_RGBA, GL_UNSIGNED_BYTE, ibuf->rect);
+    glTexSubImage3D(
+        GL_TEXTURE_2D_ARRAY, 0, x, y, layer, w, h, 1, GL_RGBA, GL_UNSIGNED_BYTE, ibuf->rect);
   }
 
   IMB_freeImBuf(ibuf);
 }
 
-static void gpu_texture_update_unscaled(
-    uchar *rect, float *rect_float, int x, int y, int w, int h, GLint tex_stride, GLint tex_offset)
+static void gpu_texture_update_scaled(
+    uchar *rect, float *rect_float, int full_w, int full_h, int x, int y, int w, int h)
+{
+  /* Partial update with scaling. */
+  int limit_w = smaller_power_of_2_limit(full_w);
+  int limit_h = smaller_power_of_2_limit(full_h);
+
+  ImBuf *ibuf = update_do_scale(
+      rect, rect_float, &x, &y, &w, &h, limit_w, limit_h, full_w, full_h);
+
+  if (ibuf->rect_float) {
+    glTexSubImage2D(GL_TEXTURE_2D, 0, x, y, w, h, GL_RGBA, GL_FLOAT, ibuf->rect_float);
+  }
+  else {
+    glTexSubImage2D(GL_TEXTURE_2D, 0, x, y, w, h, GL_RGBA, GL_UNSIGNED_BYTE, ibuf->rect);
+  }
+
+  IMB_freeImBuf(ibuf);
+}
+
+static void gpu_texture_update_unscaled(uchar *rect,
+                                        float *rect_float,
+                                        int x,
+                                        int y,
+                                        int layer,
+                                        int w,
+                                        int h,
+                                        GLint tex_stride,
+                                        GLint tex_offset)
 {
   /* Partial update without scaling. Stride and offset are used to copy only a
    * subset of a possible larger buffer than what we are updating. */
@@ -354,22 +688,61 @@ static void gpu_texture_update_unscaled(
   glGetIntegerv(GL_UNPACK_ROW_LENGTH, &row_length);
   glPixelStorei(GL_UNPACK_ROW_LENGTH, tex_stride);
 
-  if (rect_float == NULL) {
-    glTexSubImage2D(GL_TEXTURE_2D, 0, x, y, w, h, GL_RGBA, GL_UNSIGNED_BYTE, rect + tex_offset);
+  if (layer >= 0) {
+    if (rect_float == NULL) {
+      glTexSubImage3D(GL_TEXTURE_2D_ARRAY,
+                      0,
+                      x,
+                      y,
+                      layer,
+                      w,
+                      h,
+                      1,
+                      GL_RGBA,
+                      GL_UNSIGNED_BYTE,
+                      rect + tex_offset);
+    }
+    else {
+      glTexSubImage3D(GL_TEXTURE_2D_ARRAY,
+                      0,
+                      x,
+                      y,
+                      layer,
+                      w,
+                      h,
+                      1,
+                      GL_RGBA,
+                      GL_FLOAT,
+                      rect_float + tex_offset);
+    }
   }
   else {
-    glTexSubImage2D(GL_TEXTURE_2D, 0, x, y, w, h, GL_RGBA, GL_FLOAT, rect_float + tex_offset);
+    if (rect_float == NULL) {
+      glTexSubImage2D(GL_TEXTURE_2D, 0, x, y, w, h, GL_RGBA, GL_UNSIGNED_BYTE, rect + tex_offset);
+    }
+    else {
+      glTexSubImage2D(GL_TEXTURE_2D, 0, x, y, w, h, GL_RGBA, GL_FLOAT, rect_float + tex_offset);
+    }
   }
 
   glPixelStorei(GL_UNPACK_ROW_LENGTH, row_length);
 }
 
-static void gpu_texture_update_from_ibuf(Image *ima, ImBuf *ibuf, int x, int y, int w, int h)
+static void gpu_texture_update_from_ibuf(
+    GPUTexture *tex, Image *ima, ImBuf *ibuf, ImageTile *tile, int x, int y, int w, int h)
 {
   /* Partial update of texture for texture painting. This is often much
-   * quicker than fully updating the texture for high resolution images.
-   * Assumes the OpenGL texture is bound to 0. */
-  const bool scaled = is_over_resolution_limit(GL_TEXTURE_2D, ibuf->x, ibuf->y);
+   * quicker than fully updating the texture for high resolution images. */
+  GPU_texture_bind(tex, 0);
+
+  bool scaled;
+  if (tile != NULL) {
+    int *tilesize = tile->runtime.tilearray_size;
+    scaled = (ibuf->x != tilesize[0]) || (ibuf->y != tilesize[1]);
+  }
+  else {
+    scaled = is_over_resolution_limit(GL_TEXTURE_2D, ibuf->x, ibuf->y);
+  }
 
   if (scaled) {
     /* Extra padding to account for bleed from neighboring pixels. */
@@ -429,11 +802,35 @@ static void gpu_texture_update_from_ibuf(Image *ima, ImBuf *ibuf, int x, int y,
 
   if (scaled) {
     /* Slower update where we first have to scale the input pixels. */
-    gpu_texture_update_scaled(rect, rect_float, ibuf->x, ibuf->y, x, y, w, h);
+    if (tile != NULL) {
+      int *tileoffset = tile->runtime.tilearray_offset;
+      int *tilesize = tile->runtime.tilearray_size;
+      int tilelayer = tile->runtime.tilearray_layer;
+      gpu_texture_update_scaled_array(
+          rect, rect_float, ibuf->x, ibuf->y, x, y, tilelayer, tileoffset, tilesize, w, h);
+    }
+    else {
+      gpu_texture_update_scaled(rect, rect_float, ibuf->x, ibuf->y, x, y, w, h);
+    }
   }
   else {
     /* Fast update at same resolution. */
-    gpu_texture_update_unscaled(rect, rect_float, x, y, w, h, tex_stride, tex_offset);
+    if (tile != NULL) {
+      int *tileoffset = tile->runtime.tilearray_offset;
+      int tilelayer = tile->runtime.tilearray_layer;
+      gpu_texture_update_unscaled(rect,
+                                  rect_float,
+                                  x + tileoffset[0],
+                                  y + tileoffset[1],
+                                  tilelayer,
+                                  w,
+                                  h,
+                                  tex_stride,
+                                  tex_offset);
+    }
+    else {
+      gpu_texture_update_unscaled(rect, rect_float, x, y, -1, w, h, tex_stride, tex_offset);
+    }
   }
 
   /* Free buffers if needed. */
@@ -443,6 +840,15 @@ static void gpu_texture_update_from_ibuf(Image *ima, ImBuf *ibuf, int x, int y,
   if (rect_float && rect_float != ibuf->rect_float) {
     MEM_freeN(rect_float);
   }
+
+  if (GPU_get_mipmap()) {
+    glGenerateMipmap((tile != NULL) ? GL_TEXTURE_2D_ARRAY : GL_TEXTURE_2D);
+  }
+  else {
+    ima->gpuflag &= ~IMA_GPU_MIPMAP_COMPLETE;
+  }
+
+  GPU_texture_unbind(tex);
 }
 
 GPUTexture *GPU_texture_from_blender(Image *ima, ImageUser *iuser, int textarget)
@@ -460,19 +866,8 @@ GPUTexture *GPU_texture_from_blender(Image *ima, ImageUser *iuser, int textarget
   /* Tag as in active use for garbage collector. */
   BKE_image_tag_time(ima);
 
-  ImageTile *tile = BKE_image_get_tile_from_iuser(ima, iuser);
-
-  if (tile == NULL) {
-    /* TODO(lukas): When a tile gets deleted, the materials using the image
-     * aren't rebuilt and therefore continue to use it.
-     * This workaround isn't ideal, the result should be a pink color
-     * (for a missing tile). With the current behavior, new tiles also won't
-     * be detected. */
-    tile = BKE_image_get_tile(ima, 0);
-  }
-
   /* Test if we already have a texture. */
-  GPUTexture **tex = gpu_get_tile_gputexture(tile, textarget);
+  GPUTexture **tex = gpu_get_image_gputexture(ima, textarget);
   if (*tex) {
     return *tex;
   }
@@ -480,6 +875,7 @@ GPUTexture *GPU_texture_from_blender(Image *ima, ImageUser *iuser, int textarget
   /* Check if we have a valid image. If not, we return a dummy
    * texture with zero bindcode so we don't keep trying. */
   uint bindcode = 0;
+  ImageTile *tile = BKE_image_get_tile(ima, 0);
   if (tile->ok == 0) {
     *tex = GPU_texture_from_bindcode(textarget, bindcode);
     return *tex;
@@ -492,7 +888,15 @@ GPUTexture *GPU_texture_from_blender(Image *ima, ImageUser *iuser, int textarget
     return *tex;
   }
 
-  bindcode = gpu_texture_create_from_ibuf(ima, ibuf, textarget);
+  if (textarget == GL_TEXTURE_2D_ARRAY) {
+    bindcode = gpu_texture_create_tile_array(ima, ibuf);
+  }
+  else if (textarget == GL_TEXTURE_1D_ARRAY) {
+    bindcode = gpu_texture_create_tile_mapping(ima);
+  }
+  else {
+    bindcode = gpu_texture_create_from_ibuf(ima, ibuf, textarget);
+  }
 
   BKE_image_release_ibuf(ima, ibuf, NULL);
 
@@ -856,13 +1260,15 @@ void GPU_paint_set_mipmap(Main *bmain, bool mipmap)
     for (Image *ima = bmain->images.first; ima; ima = ima->id.next) {
       if (BKE_image_has_opengl_texture(ima)) {
         if (ima->gpuflag & IMA_GPU_MIPMAP_COMPLETE) {
-          LISTBASE_FOREACH (ImageTile *, tile, &ima->tiles) {
-            GPUTexture *tex = tile->gputexture[TEXTARGET_TEXTURE_2D];
-            if (tex != NULL) {
-              GPU_texture_bind(tex, 0);
-              glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, gpu_get_mipmap_filter(0));
-              glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, gpu_get_mipmap_filter(1));
-              GPU_texture_unbind(tex);
+          for (int a = 0; a < TEXTARGET_COUNT; a++) {
+            if (ELEM(a, TEXTARGET_TEXTURE_2D, TEXTARGET_TEXTURE_2D_ARRAY)) {
+              GPUTexture *tex = ima->gputexture[a];
+              if (tex != NULL) {
+                GPU_texture_bind(tex, 0);
+                glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, gpu_get_mipmap_filter(0));
+                glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, gpu_get_mipmap_filter(1));
+                GPU_texture_unbind(tex);
+              }
             }
           }
         }
@@ -878,13 +1284,15 @@ void GPU_paint_set_mipmap(Main *bmain, bool mipmap)
   else {
     for (Image *ima = bmain->images.first; ima; ima = ima->id.next) {
       if (BKE_image_has_opengl_texture(ima)) {
-        LISTBASE_FOREACH (ImageTile *, tile, &ima->tiles) {
-          GPUTexture *tex = tile->gputexture[TEXTARGET_TEXTURE_2D];
-          if (tex != NULL) {
-            GPU_texture_bind(tex, 0);
-            glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
-            glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, gpu_get_mipmap_filter(1));
-            GPU_texture_unbind(tex);
+        for (int a = 0; a < TEXTARGET_COUNT; a++) {
+          if (ELEM(a, TEXTARGET_TEXTURE_2D, TEXTARGET_TEXTURE_2D_ARRAY)) {
+            GPUTexture *tex = ima->gputexture[a];
+            if (tex != NULL) {
+              GPU_texture_bind(tex, 0);
+              glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
+              glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, gpu_get_mipmap_filter(1));
+              GPU_texture_unbind(tex);
+            }
           }
         }
       }
@@ -899,26 +1307,22 @@ void GPU_paint_update_image(Image *ima, ImageUser *iuser, int x, int y, int w, i
 {
   ImBuf *ibuf = BKE_image_acquire_ibuf(ima, iuser, NULL);
   ImageTile *tile = BKE_image_get_tile_from_iuser(ima, iuser);
-  GPUTexture *tex = tile->gputexture[TEXTARGET_TEXTURE_2D];
 
-  if ((tex == NULL) || (ibuf == NULL) || (w == 0) || (h == 0)) {
+  if ((ibuf == NULL) || (w == 0) || (h == 0)) {
     /* Full reload of texture. */
     GPU_free_image(ima);
   }
-  else {
-    /* Partial update of texture. */
-    GPU_texture_bind(tex, 0);
 
-    gpu_texture_update_from_ibuf(ima, ibuf, x, y, w, h);
-
-    if (GPU_get_mipmap()) {
-      glGenerateMipmap(GL_TEXTURE_2D);
-    }
-    else {
-      ima->gpuflag &= ~IMA_GPU_MIPMAP_COMPLETE;
-    }
+  GPUTexture *tex = ima->gputexture[TEXTARGET_TEXTURE_2D];
+  /* Check if we need to update the main gputexture. */
+  if (tex != NULL && tile == ima->tiles.first) {
+    gpu_texture_update_from_ibuf(tex, ima, ibuf, NULL, x, y, w, h);
+  }
 
-    GPU_texture_unbind(tex);
+  /* Check if we need to update the array gputexture. */
+  tex = ima->gputexture[TEXTARGET_TEXTURE_2D_ARRAY];
+  if (tex != NULL) {
+    gpu_texture_update_from_ibuf(tex, ima, ibuf, tile, x, y, w, h);
   }
 
   BKE_image_release_ibuf(ima, ibuf, NULL);
@@ -960,13 +1364,11 @@ void GPU_free_unused_buffers(Main *bmain)
 
 static void gpu_free_image_immediate(Image *ima)
 {
-  LISTBASE_FOREACH (ImageTile *, tile, &ima->tiles) {
-    for (int i = 0; i < TEXTARGET_COUNT; i++) {
-      /* free glsl image binding */
-      if (tile->gputexture[i] != NULL) {
-        GPU_texture_free(tile->gputexture[i]);
-        tile->gputexture[i] = NULL;
-      }
+  for (int i = 0; i < TEXTARGET_COUNT; i++) {
+    /* free glsl image binding */
+    if (ima->gputexture[i] != NULL) {
+      GPU_texture_free(ima->gputexture[i]);
+      ima->gputexture[i] = NULL;
     }
   }
 
diff --git a/source/blender/gpu/intern/gpu_texture.c b/source/blender/gpu/intern/gpu_texture.c
index 497fc13a2c8..201194232db 100644
--- a/source/blender/gpu/intern/gpu_texture.c
+++ b/source/blender/gpu/intern/gpu_texture.c
@@ -1032,10 +1032,6 @@ GPUTexture *GPU_texture_create_buffer(eGPUTextureFormat tex_format, const GLuint
 
 GPUTexture *GPU_texture_from_bindcode(int textarget, int bindcode)
 {
-  /* see GPUInput::textarget: it can take two values - GL_TEXTURE_2D and GL_TEXTURE_CUBE_MAP
-   * these values are correct for glDisable, so textarget can be safely used in
-   * GPU_texture_bind/GPU_texture_unbind through tex->target_base */
-  /* (is any of this obsolete now that we don't glEnable/Disable textures?) */
   GPUTexture *tex = MEM_callocN(sizeof(GPUTexture), "GPUTexture");
   tex->bindcode = bindcode;
   tex->number = -1;
@@ -1052,12 +1048,8 @@ GPUTexture *GPU_texture_from_bindcode(int textarget, int bindcode)
   else {
     GLint w, h;
 
-    GLenum gettarget;
-
-    if (textarget == GL_TEXTURE_2D) {
-      gettarget = GL_TEXTURE_2D;
-    }
-    else {
+    GLenum gettarget = textarget;
+    if (textarget == GL_TEXTURE_CUBE_MAP) {
       gettarget = GL_TEXTURE_CUBE_MAP_POSITIVE_X;
     }
 
diff --git a/source/blender/gpu/shaders/material/gpu_shader_material_tex_image.glsl b/source/blender/gpu/shaders/material/gpu_shader_material_tex_image.glsl
index fadb3b92df4..4633e59fde1 100644
--- a/source/blender/gpu/shaders/material/gpu_shader_material_tex_image.glsl
+++ b/source/blender/gpu/shaders/material/gpu_shader_material_tex_image.glsl
@@ -354,67 +354,92 @@ void node_tex_image_empty(vec3 co, out vec4 color, out float alpha)
   alpha = 0.0;
 }
 
-void node_tex_tile_map(vec3 co, out vec4 color, out vec3 map)
+bool node_tex_tile_lookup(inout vec3 co, sampler2DArray ima, sampler1DArray map)
 {
-  float tx = floor(co.x);
-  float ty = floor(co.y);
+  vec2 tile_pos = floor(co.xy);
 
-  if (tx < 0 || ty < 0 || tx >= 10)
-    map = vec3(0, 0, -1);
-  else
-    map = vec3(co.x - tx, co.y - ty, 1001 + 10 * ty + tx);
+  if (tile_pos.x < 0 || tile_pos.y < 0 || tile_pos.x >= 10)
+    return false;
 
-  color = vec4(1.0, 0.0, 1.0, 1.0);
+  float tile = 10 * tile_pos.y + tile_pos.x;
+  if (tile >= textureSize(map, 0).x)
+    return false;
+
+  /* Fetch tile information. */
+  float tile_layer = texelFetch(map, ivec2(tile, 0), 0).x;
+  if (tile_layer < 0)
+    return false;
+
+  vec4 tile_info = texelFetch(map, ivec2(tile, 1), 0);
+
+  co = vec3(((co.xy - tile_pos) * tile_info.zw) + tile_info.xy, tile_layer);
+  return true;
 }
 
 void node_tex_tile_linear(
-    vec3 map, float tile_id, sampler2D ima, vec4 in_color, out vec4 color, out float alpha)
+    vec3 co, sampler2DArray ima, sampler1DArray map, out vec4 color, out float alpha)
 {
-  if (map.z == tile_id) {
-    vec3 co = map.xyy;
-    node_tex_image_linear(co, ima, color, alpha);
+  if (node_tex_tile_lookup(co, ima, map)) {
+    color = safe_color(texture(ima, co));
   }
   else {
-    color = in_color;
-    alpha = color.a;
+    color = vec4(1.0, 0.0, 1.0, 1.0);
   }
+
+  alpha = color.a;
 }
 
 void node_tex_tile_nearest(
-    vec3 map, float tile_id, sampler2D ima, vec4 in_color, out vec4 color, out float alpha)
+    vec3 co, sampler2DArray ima, sampler1DArray map, out vec4 color, out float alpha)
 {
-  if (map.z == tile_id) {
-    vec3 co = map.xyy;
-    node_tex_image_nearest(co, ima, color, alpha);
+  if (node_tex_tile_lookup(co, ima, map)) {
+    ivec3 pix = ivec3(fract(co.xy) * textureSize(ima, 0).xy, co.z);
+    color = safe_color(texelFetch(ima, pix, 0));
   }
   else {
-    color = in_color;
-    alpha = color.a;
+    color = vec4(1.0, 0.0, 1.0, 1.0);
   }
+
+  alpha = color.a;
 }
 
 void node_tex_tile_cubic(
-    vec3 map, float tile_id, sampler2D ima, vec4 in_color, out vec4 color, out float alpha)
+    vec3 co, sampler2DArray ima, sampler1DArray map, out vec4 color, out float alpha)
 {
-  if (map.z == tile_id) {
-    vec3 co = map.xyy;
-    node_tex_image_cubic(co, ima, color, alpha);
+  if (node_tex_tile_lookup(co, ima, map)) {
+    vec2 tex_size = vec2(textureSize(ima, 0).xy);
+
+    co.xy *= tex_size;
+    /* texel center */
+    vec2 tc = floor(co.xy - 0.5) + 0.5;
+    vec2 w0, w1, w2, w3;
+    cubic_bspline_coefs(co.xy - tc, w0, w1, w2, w3);
+
+    vec2 s0 = w0 + w1;
+    vec2 s1 = w2 + w3;
+
+    vec2 f0 = w1 / (w0 + w1);
+    vec2 f1 = w3 / (w2 + w3);
+
+    vec4 final_co;
+    final_co.xy = tc - 1.0 + f0;
+    final_co.zw = tc + 1.0 + f1;
+    final_co /= tex_size.xyxy;
+
+    color = safe_color(textureLod(ima, vec3(final_co.xy, co.z), 0.0)) * s0.x * s0.y;
+    color += safe_color(textureLod(ima, vec3(final_co.zy, co.z), 0.0)) * s1.x * s0.y;
+    color += safe_color(textureLod(ima, vec3(final_co.xw, co.z), 0.0)) * s0.x * s1.y;
+    color += safe_color(textureLod(ima, vec3(final_co.zw, co.z), 0.0)) * s1.x * s1.y;
   }
   else {
-    color = in_color;
-    alpha = color.a;
+    color = vec4(1.0, 0.0, 1.0, 1.0);
   }
+
+  alpha = color.a;
 }
 
 void node_tex_tile_smart(
-    vec3 map, float tile_id, sampler2D ima, vec4 in_color, out vec4 color, out float alpha)
+    vec3 co, sampler2DArray ima, sampler1DArray map, out vec4 color, out float alpha)
 {
-  if (map.z == tile_id) {
-    vec3 co = map.xyy;
-    node_tex_image_smart(co, ima, color, alpha);
-  }
-  else {
-    color = in_color;
-    alpha = color.a;
-  }
+  node_tex_tile_cubic(co, ima, map, color, alpha);
 }