Merge remote-tracking branch 'public/master' into vulkan
[mesa.git] / src / mesa / state_tracker / st_cb_texture.c
index a8b19a1f3232bda551870ad00c6b89042a220358..460c1790663a00b61779b168e83fd15ef8ddf391 100644 (file)
@@ -60,6 +60,7 @@
 #include "pipe/p_context.h"
 #include "pipe/p_defines.h"
 #include "util/u_inlines.h"
+#include "util/u_upload_mgr.h"
 #include "pipe/p_shader_tokens.h"
 #include "util/u_tile.h"
 #include "util/u_format.h"
@@ -67,6 +68,9 @@
 #include "util/u_sampler.h"
 #include "util/u_math.h"
 #include "util/u_box.h"
+#include "util/u_simple_shaders.h"
+#include "cso_cache/cso_context.h"
+#include "tgsi/tgsi_ureg.h"
 
 #define DBG if (0) printf
 
@@ -123,7 +127,7 @@ gl_target_to_pipe(GLenum target)
 static struct gl_texture_image *
 st_NewTextureImage(struct gl_context * ctx)
 {
-   DBG("%s\n", __FUNCTION__);
+   DBG("%s\n", __func__);
    (void) ctx;
    return (struct gl_texture_image *) ST_CALLOC_STRUCT(st_texture_image);
 }
@@ -144,7 +148,7 @@ st_NewTextureObject(struct gl_context * ctx, GLuint name, GLenum target)
 {
    struct st_texture_object *obj = ST_CALLOC_STRUCT(st_texture_object);
 
-   DBG("%s\n", __FUNCTION__);
+   DBG("%s\n", __func__);
    _mesa_initialize_texture_object(ctx, &obj->base, name, target);
 
    return &obj->base;
@@ -172,15 +176,12 @@ st_FreeTextureImageBuffer(struct gl_context *ctx,
 {
    struct st_texture_image *stImage = st_texture_image(texImage);
 
-   DBG("%s\n", __FUNCTION__);
+   DBG("%s\n", __func__);
 
    if (stImage->pt) {
       pipe_resource_reference(&stImage->pt, NULL);
    }
 
-   _mesa_align_free(stImage->TexData);
-   stImage->TexData = NULL;
-
    free(stImage->transfer);
    stImage->transfer = NULL;
    stImage->num_transfers = 0;
@@ -390,6 +391,57 @@ guess_base_level_size(GLenum target,
 }
 
 
+/**
+ * Try to determine whether we should allocate memory for a full texture
+ * mipmap.  The problem is when we get a glTexImage(level=0) call, we
+ * can't immediately know if other mipmap levels are coming next.  Here
+ * we try to guess whether to allocate memory for a mipmap or just the
+ * 0th level.
+ *
+ * If we guess incorrectly here we'll later reallocate the right amount of
+ * memory either in st_AllocTextureImageBuffer() or st_finalize_texture().
+ *
+ * \param stObj  the texture object we're going to allocate memory for.
+ * \param stImage  describes the incoming image which we need to store.
+ */
+static boolean
+allocate_full_mipmap(const struct st_texture_object *stObj,
+                     const struct st_texture_image *stImage)
+{
+   switch (stObj->base.Target) {
+   case GL_TEXTURE_RECTANGLE_NV:
+   case GL_TEXTURE_BUFFER:
+   case GL_TEXTURE_EXTERNAL_OES:
+   case GL_TEXTURE_2D_MULTISAMPLE:
+   case GL_TEXTURE_2D_MULTISAMPLE_ARRAY:
+      /* these texture types cannot be mipmapped */
+      return FALSE;
+   }
+
+   if (stImage->base.Level > 0 || stObj->base.GenerateMipmap)
+      return TRUE;
+
+   if (stImage->base._BaseFormat == GL_DEPTH_COMPONENT ||
+       stImage->base._BaseFormat == GL_DEPTH_STENCIL_EXT)
+      /* depth/stencil textures are seldom mipmapped */
+      return FALSE;
+
+   if (stObj->base.BaseLevel == 0 && stObj->base.MaxLevel == 0)
+      return FALSE;
+
+   if (stObj->base.Sampler.MinFilter == GL_NEAREST ||
+       stObj->base.Sampler.MinFilter == GL_LINEAR)
+      /* not a mipmap minification filter */
+      return FALSE;
+
+   if (stObj->base.Target == GL_TEXTURE_3D)
+      /* 3D textures are seldom mipmapped */
+      return FALSE;
+
+   return TRUE;
+}
+
+
 /**
  * Try to allocate a pipe_resource object for the given st_texture_object.
  *
@@ -408,7 +460,7 @@ guess_and_alloc_texture(struct st_context *st,
    GLuint ptWidth, ptHeight, ptDepth, ptLayers;
    enum pipe_format fmt;
 
-   DBG("%s\n", __FUNCTION__);
+   DBG("%s\n", __func__);
 
    assert(!stObj->pt);
 
@@ -434,22 +486,15 @@ guess_and_alloc_texture(struct st_context *st,
     * to re-allocating a texture buffer with space for more (or fewer)
     * mipmap levels later.
     */
-   if ((stObj->base.Sampler.MinFilter == GL_NEAREST ||
-        stObj->base.Sampler.MinFilter == GL_LINEAR ||
-        (stObj->base.BaseLevel == 0 &&
-         stObj->base.MaxLevel == 0) ||
-        stImage->base._BaseFormat == GL_DEPTH_COMPONENT ||
-        stImage->base._BaseFormat == GL_DEPTH_STENCIL_EXT) &&
-       !stObj->base.GenerateMipmap &&
-       stImage->base.Level == 0) {
-      /* only alloc space for a single mipmap level */
-      lastLevel = 0;
-   }
-   else {
+   if (allocate_full_mipmap(stObj, stImage)) {
       /* alloc space for a full mipmap */
       lastLevel = _mesa_get_tex_max_num_levels(stObj->base.Target,
                                                width, height, depth) - 1;
    }
+   else {
+      /* only alloc space for a single mipmap level */
+      lastLevel = 0;
+   }
 
    /* Save the level=0 dimensions */
    stObj->width0 = width;
@@ -476,7 +521,7 @@ guess_and_alloc_texture(struct st_context *st,
 
    stObj->lastLevel = lastLevel;
 
-   DBG("%s returning %d\n", __FUNCTION__, (stObj->pt != NULL));
+   DBG("%s returning %d\n", __func__, (stObj->pt != NULL));
 
    return stObj->pt != NULL;
 }
@@ -499,9 +544,8 @@ st_AllocTextureImageBuffer(struct gl_context *ctx,
    GLuint height = texImage->Height;
    GLuint depth = texImage->Depth;
 
-   DBG("%s\n", __FUNCTION__);
+   DBG("%s\n", __func__);
 
-   assert(!stImage->TexData);
    assert(!stImage->pt); /* xxx this might be wrong */
 
    /* Look if the parent texture object has space for this image */
@@ -646,6 +690,989 @@ st_get_blit_mask(GLenum srcFormat, GLenum dstFormat)
    }
 }
 
+void
+st_init_pbo_upload(struct st_context *st)
+{
+   struct pipe_context *pipe = st->pipe;
+   struct pipe_screen *screen = pipe->screen;
+
+   st->pbo_upload.enabled =
+      screen->get_param(screen, PIPE_CAP_TEXTURE_BUFFER_OBJECTS) &&
+      screen->get_param(screen, PIPE_CAP_TEXTURE_BUFFER_OFFSET_ALIGNMENT) >= 1 &&
+      screen->get_shader_param(screen, PIPE_SHADER_FRAGMENT, PIPE_SHADER_CAP_INTEGERS);
+   if (!st->pbo_upload.enabled)
+      return;
+
+   st->pbo_upload.rgba_only =
+      screen->get_param(screen, PIPE_CAP_BUFFER_SAMPLER_VIEW_RGBA_ONLY);
+
+   if (screen->get_param(screen, PIPE_CAP_TGSI_INSTANCEID)) {
+      if (screen->get_param(screen, PIPE_CAP_TGSI_VS_LAYER_VIEWPORT)) {
+         st->pbo_upload.upload_layers = true;
+      } else if (screen->get_param(screen, PIPE_CAP_MAX_GEOMETRY_OUTPUT_VERTICES) >= 3) {
+         st->pbo_upload.upload_layers = true;
+         st->pbo_upload.use_gs = true;
+      }
+   }
+
+   /* Blend state */
+   memset(&st->pbo_upload.blend, 0, sizeof(struct pipe_blend_state));
+   st->pbo_upload.blend.rt[0].colormask = PIPE_MASK_RGBA;
+
+   /* Rasterizer state */
+   memset(&st->pbo_upload.raster, 0, sizeof(struct pipe_rasterizer_state));
+   st->pbo_upload.raster.half_pixel_center = 1;
+}
+
+void
+st_destroy_pbo_upload(struct st_context *st)
+{
+   if (st->pbo_upload.fs) {
+      cso_delete_fragment_shader(st->cso_context, st->pbo_upload.fs);
+      st->pbo_upload.fs = NULL;
+   }
+
+   if (st->pbo_upload.gs) {
+      cso_delete_geometry_shader(st->cso_context, st->pbo_upload.gs);
+      st->pbo_upload.gs = NULL;
+   }
+
+   if (st->pbo_upload.vs) {
+      cso_delete_vertex_shader(st->cso_context, st->pbo_upload.vs);
+      st->pbo_upload.vs = NULL;
+   }
+}
+
+/**
+ * Converts format to a format with the same components, types
+ * and sizes, but with the components in RGBA order.
+ */
+static enum pipe_format
+unswizzle_format(enum pipe_format format)
+{
+   switch (format)
+   {
+   case PIPE_FORMAT_B8G8R8A8_UNORM:
+   case PIPE_FORMAT_A8R8G8B8_UNORM:
+   case PIPE_FORMAT_A8B8G8R8_UNORM:
+      return PIPE_FORMAT_R8G8B8A8_UNORM;
+
+   case PIPE_FORMAT_B10G10R10A2_UNORM:
+      return PIPE_FORMAT_R10G10B10A2_UNORM;
+
+   case PIPE_FORMAT_B10G10R10A2_SNORM:
+      return PIPE_FORMAT_R10G10B10A2_SNORM;
+
+   case PIPE_FORMAT_B10G10R10A2_UINT:
+      return PIPE_FORMAT_R10G10B10A2_UINT;
+
+   default:
+      return format;
+   }
+}
+
+/**
+ * Converts PIPE_FORMAT_A* to PIPE_FORMAT_R*.
+ */
+static enum pipe_format
+alpha_to_red(enum pipe_format format)
+{
+   switch (format)
+   {
+   case PIPE_FORMAT_A8_UNORM:
+      return PIPE_FORMAT_R8_UNORM;
+   case PIPE_FORMAT_A8_SNORM:
+      return PIPE_FORMAT_R8_SNORM;
+   case PIPE_FORMAT_A8_UINT:
+      return PIPE_FORMAT_R8_UINT;
+   case PIPE_FORMAT_A8_SINT:
+      return PIPE_FORMAT_R8_SINT;
+
+   case PIPE_FORMAT_A16_UNORM:
+      return PIPE_FORMAT_R16_UNORM;
+   case PIPE_FORMAT_A16_SNORM:
+      return PIPE_FORMAT_R16_SNORM;
+   case PIPE_FORMAT_A16_UINT:
+      return PIPE_FORMAT_R16_UINT;
+   case PIPE_FORMAT_A16_SINT:
+      return PIPE_FORMAT_R16_SINT;
+   case PIPE_FORMAT_A16_FLOAT:
+      return PIPE_FORMAT_R16_FLOAT;
+
+   case PIPE_FORMAT_A32_UINT:
+      return PIPE_FORMAT_R32_UINT;
+   case PIPE_FORMAT_A32_SINT:
+      return PIPE_FORMAT_R32_SINT;
+   case PIPE_FORMAT_A32_FLOAT:
+      return PIPE_FORMAT_R32_FLOAT;
+
+   default:
+      return format;
+   }
+}
+
+/**
+ * Converts PIPE_FORMAT_R*A* to PIPE_FORMAT_R*G*.
+ */
+static enum pipe_format
+red_alpha_to_red_green(enum pipe_format format)
+{
+   switch (format)
+   {
+   case PIPE_FORMAT_R8A8_UNORM:
+      return PIPE_FORMAT_R8G8_UNORM;
+   case PIPE_FORMAT_R8A8_SNORM:
+      return PIPE_FORMAT_R8G8_SNORM;
+   case PIPE_FORMAT_R8A8_UINT:
+      return PIPE_FORMAT_R8G8_UINT;
+   case PIPE_FORMAT_R8A8_SINT:
+      return PIPE_FORMAT_R8G8_SINT;
+
+   case PIPE_FORMAT_R16A16_UNORM:
+      return PIPE_FORMAT_R16G16_UNORM;
+   case PIPE_FORMAT_R16A16_SNORM:
+      return PIPE_FORMAT_R16G16_SNORM;
+   case PIPE_FORMAT_R16A16_UINT:
+      return PIPE_FORMAT_R16G16_UINT;
+   case PIPE_FORMAT_R16A16_SINT:
+      return PIPE_FORMAT_R16G16_SINT;
+   case PIPE_FORMAT_R16A16_FLOAT:
+      return PIPE_FORMAT_R16G16_FLOAT;
+
+   case PIPE_FORMAT_R32A32_UINT:
+      return PIPE_FORMAT_R32G32_UINT;
+   case PIPE_FORMAT_R32A32_SINT:
+      return PIPE_FORMAT_R32G32_SINT;
+   case PIPE_FORMAT_R32A32_FLOAT:
+      return PIPE_FORMAT_R32G32_FLOAT;
+
+   default:
+       return format;
+   }
+}
+
+/**
+ * Converts PIPE_FORMAT_L*A* to PIPE_FORMAT_R*G*.
+ */
+static enum pipe_format
+luminance_alpha_to_red_green(enum pipe_format format)
+{
+   switch (format)
+   {
+   case PIPE_FORMAT_L8A8_UNORM:
+      return PIPE_FORMAT_R8G8_UNORM;
+   case PIPE_FORMAT_L8A8_SNORM:
+      return PIPE_FORMAT_R8G8_SNORM;
+   case PIPE_FORMAT_L8A8_UINT:
+      return PIPE_FORMAT_R8G8_UINT;
+   case PIPE_FORMAT_L8A8_SINT:
+      return PIPE_FORMAT_R8G8_SINT;
+
+   case PIPE_FORMAT_L16A16_UNORM:
+      return PIPE_FORMAT_R16G16_UNORM;
+   case PIPE_FORMAT_L16A16_SNORM:
+      return PIPE_FORMAT_R16G16_SNORM;
+   case PIPE_FORMAT_L16A16_UINT:
+      return PIPE_FORMAT_R16G16_UINT;
+   case PIPE_FORMAT_L16A16_SINT:
+      return PIPE_FORMAT_R16G16_SINT;
+   case PIPE_FORMAT_L16A16_FLOAT:
+      return PIPE_FORMAT_R16G16_FLOAT;
+
+   case PIPE_FORMAT_L32A32_UINT:
+      return PIPE_FORMAT_R32G32_UINT;
+   case PIPE_FORMAT_L32A32_SINT:
+      return PIPE_FORMAT_R32G32_SINT;
+   case PIPE_FORMAT_L32A32_FLOAT:
+      return PIPE_FORMAT_R32G32_FLOAT;
+
+   default:
+       return format;
+   }
+}
+
+/**
+ * Returns true if format is a PIPE_FORMAT_A* format, and false otherwise.
+ */
+static bool
+format_is_alpha(enum pipe_format format)
+{
+   const struct util_format_description *desc = util_format_description(format);
+
+   if (desc->nr_channels == 1 &&
+       desc->swizzle[0] == UTIL_FORMAT_SWIZZLE_0 &&
+       desc->swizzle[1] == UTIL_FORMAT_SWIZZLE_0 &&
+       desc->swizzle[2] == UTIL_FORMAT_SWIZZLE_0 &&
+       desc->swizzle[3] == UTIL_FORMAT_SWIZZLE_X)
+      return true;
+
+   return false;
+}
+
+/**
+ * Returns true if format is a PIPE_FORMAT_R* format, and false otherwise.
+ */
+static bool
+format_is_red(enum pipe_format format)
+{
+   const struct util_format_description *desc = util_format_description(format);
+
+   if (desc->nr_channels == 1 &&
+       desc->swizzle[0] == UTIL_FORMAT_SWIZZLE_X &&
+       desc->swizzle[1] == UTIL_FORMAT_SWIZZLE_0 &&
+       desc->swizzle[2] == UTIL_FORMAT_SWIZZLE_0 &&
+       desc->swizzle[3] == UTIL_FORMAT_SWIZZLE_1)
+      return true;
+
+   return false;
+}
+
+
+/**
+ * Returns true if format is a PIPE_FORMAT_L* format, and false otherwise.
+ */
+static bool
+format_is_luminance(enum pipe_format format)
+{
+   const struct util_format_description *desc = util_format_description(format);
+
+   if (desc->nr_channels == 1 &&
+       desc->swizzle[0] == UTIL_FORMAT_SWIZZLE_X &&
+       desc->swizzle[1] == UTIL_FORMAT_SWIZZLE_X &&
+       desc->swizzle[2] == UTIL_FORMAT_SWIZZLE_X &&
+       desc->swizzle[3] == UTIL_FORMAT_SWIZZLE_1)
+      return true;
+
+   return false;
+}
+
+/**
+ * Returns true if format is a PIPE_FORMAT_R*A* format, and false otherwise.
+ */
+static bool
+format_is_red_alpha(enum pipe_format format)
+{
+   const struct util_format_description *desc = util_format_description(format);
+
+   if (desc->nr_channels == 2 &&
+       desc->swizzle[0] == UTIL_FORMAT_SWIZZLE_X &&
+       desc->swizzle[1] == UTIL_FORMAT_SWIZZLE_0 &&
+       desc->swizzle[2] == UTIL_FORMAT_SWIZZLE_0 &&
+       desc->swizzle[3] == UTIL_FORMAT_SWIZZLE_Y)
+      return true;
+
+   return false;
+}
+
+static bool
+format_is_swizzled_rgba(enum pipe_format format)
+{
+    const struct util_format_description *desc = util_format_description(format);
+
+    if ((desc->swizzle[0] == TGSI_SWIZZLE_X || desc->swizzle[0] == UTIL_FORMAT_SWIZZLE_0) &&
+        (desc->swizzle[1] == TGSI_SWIZZLE_Y || desc->swizzle[1] == UTIL_FORMAT_SWIZZLE_0) &&
+        (desc->swizzle[2] == TGSI_SWIZZLE_Z || desc->swizzle[2] == UTIL_FORMAT_SWIZZLE_0) &&
+        (desc->swizzle[3] == TGSI_SWIZZLE_W || desc->swizzle[3] == UTIL_FORMAT_SWIZZLE_1))
+       return false;
+
+    return true;
+}
+
+struct format_table
+{
+   unsigned char swizzle[4];
+   enum pipe_format format;
+};
+
+static const struct format_table table_8888_unorm[] = {
+   { { 0, 1, 2, 3 }, PIPE_FORMAT_R8G8B8A8_UNORM },
+   { { 2, 1, 0, 3 }, PIPE_FORMAT_B8G8R8A8_UNORM },
+   { { 3, 0, 1, 2 }, PIPE_FORMAT_A8R8G8B8_UNORM },
+   { { 3, 2, 1, 0 }, PIPE_FORMAT_A8B8G8R8_UNORM }
+};
+
+static const struct format_table table_1010102_unorm[] = {
+   { { 0, 1, 2, 3 }, PIPE_FORMAT_R10G10B10A2_UNORM },
+   { { 2, 1, 0, 3 }, PIPE_FORMAT_B10G10R10A2_UNORM }
+};
+
+static const struct format_table table_1010102_snorm[] = {
+   { { 0, 1, 2, 3 }, PIPE_FORMAT_R10G10B10A2_SNORM },
+   { { 2, 1, 0, 3 }, PIPE_FORMAT_B10G10R10A2_SNORM }
+};
+
+static const struct format_table table_1010102_uint[] = {
+   { { 0, 1, 2, 3 }, PIPE_FORMAT_R10G10B10A2_UINT },
+   { { 2, 1, 0, 3 }, PIPE_FORMAT_B10G10R10A2_UINT }
+};
+
+static enum pipe_format
+swizzle_format(enum pipe_format format, const int * const swizzle)
+{
+   unsigned i;
+
+   switch (format) {
+   case PIPE_FORMAT_R8G8B8A8_UNORM:
+   case PIPE_FORMAT_B8G8R8A8_UNORM:
+   case PIPE_FORMAT_A8R8G8B8_UNORM:
+   case PIPE_FORMAT_A8B8G8R8_UNORM:
+      for (i = 0; i < ARRAY_SIZE(table_8888_unorm); i++) {
+         if (swizzle[0] == table_8888_unorm[i].swizzle[0] &&
+             swizzle[1] == table_8888_unorm[i].swizzle[1] &&
+             swizzle[2] == table_8888_unorm[i].swizzle[2] &&
+             swizzle[3] == table_8888_unorm[i].swizzle[3])
+            return table_8888_unorm[i].format;
+      }
+      break;
+
+   case PIPE_FORMAT_R10G10B10A2_UNORM:
+   case PIPE_FORMAT_B10G10R10A2_UNORM:
+      for (i = 0; i < ARRAY_SIZE(table_1010102_unorm); i++) {
+         if (swizzle[0] == table_1010102_unorm[i].swizzle[0] &&
+             swizzle[1] == table_1010102_unorm[i].swizzle[1] &&
+             swizzle[2] == table_1010102_unorm[i].swizzle[2] &&
+             swizzle[3] == table_1010102_unorm[i].swizzle[3])
+            return table_1010102_unorm[i].format;
+      }
+      break;
+
+   case PIPE_FORMAT_R10G10B10A2_SNORM:
+   case PIPE_FORMAT_B10G10R10A2_SNORM:
+      for (i = 0; i < ARRAY_SIZE(table_1010102_snorm); i++) {
+         if (swizzle[0] == table_1010102_snorm[i].swizzle[0] &&
+             swizzle[1] == table_1010102_snorm[i].swizzle[1] &&
+             swizzle[2] == table_1010102_snorm[i].swizzle[2] &&
+             swizzle[3] == table_1010102_snorm[i].swizzle[3])
+            return table_1010102_snorm[i].format;
+      }
+      break;
+
+   case PIPE_FORMAT_R10G10B10A2_UINT:
+   case PIPE_FORMAT_B10G10R10A2_UINT:
+      for (i = 0; i < ARRAY_SIZE(table_1010102_uint); i++) {
+         if (swizzle[0] == table_1010102_uint[i].swizzle[0] &&
+             swizzle[1] == table_1010102_uint[i].swizzle[1] &&
+             swizzle[2] == table_1010102_uint[i].swizzle[2] &&
+             swizzle[3] == table_1010102_uint[i].swizzle[3])
+            return table_1010102_uint[i].format;
+      }
+      break;
+
+   default:
+      break;
+   }
+
+   return PIPE_FORMAT_NONE;
+}
+
+static bool
+reinterpret_formats(enum pipe_format *src_format, enum pipe_format *dst_format)
+{
+   enum pipe_format src = *src_format;
+   enum pipe_format dst = *dst_format;
+
+   /* Note: dst_format has already been transformed from luminance/intensity
+    *       to red when this function is called.  The source format will never
+    *       be an intensity format, because GL_INTENSITY is not a legal value
+    *       for the format parameter in glTex(Sub)Image(). */
+
+   if (format_is_alpha(src)) {
+      if (!format_is_alpha(dst))
+         return false;
+
+      src = alpha_to_red(src);
+      dst = alpha_to_red(dst);
+   } else if (format_is_luminance(src)) {
+      if (!format_is_red(dst) && !format_is_red_alpha(dst))
+         return false;
+
+      src = util_format_luminance_to_red(src);
+   } else if (util_format_is_luminance_alpha(src)) {
+      src = luminance_alpha_to_red_green(src);
+
+      if (format_is_red_alpha(dst)) {
+         dst = red_alpha_to_red_green(dst);
+      } else if (!format_is_red(dst))
+         return false;
+   } else if (format_is_swizzled_rgba(src)) {
+      const struct util_format_description *src_desc = util_format_description(src);
+      const struct util_format_description *dst_desc = util_format_description(dst);
+      int swizzle[4];
+      unsigned i;
+
+      /* Make sure the format is an RGBA and not an RGBX format */
+      if (src_desc->nr_channels != 4 || src_desc->swizzle[3] == UTIL_FORMAT_SWIZZLE_1)
+         return false;
+
+      if (dst_desc->nr_channels != 4 || dst_desc->swizzle[3] == UTIL_FORMAT_SWIZZLE_1)
+         return false;
+
+      for (i = 0; i < 4; i++)
+         swizzle[i] = dst_desc->swizzle[src_desc->swizzle[i]];
+
+      dst = swizzle_format(dst, swizzle);
+      if (dst == PIPE_FORMAT_NONE)
+         return false;
+
+      src = unswizzle_format(src);
+   }
+
+   *src_format = src;
+   *dst_format = dst;
+   return true;
+}
+
+static void *
+create_pbo_upload_vs(struct st_context *st)
+{
+   struct ureg_program *ureg;
+   struct ureg_src in_pos;
+   struct ureg_src in_instanceid;
+   struct ureg_dst out_pos;
+   struct ureg_dst out_layer;
+
+   ureg = ureg_create(TGSI_PROCESSOR_VERTEX);
+   if (!ureg)
+      return NULL;
+
+   in_pos = ureg_DECL_vs_input(ureg, TGSI_SEMANTIC_POSITION);
+
+   out_pos = ureg_DECL_output(ureg, TGSI_SEMANTIC_POSITION, 0);
+
+   if (st->pbo_upload.upload_layers) {
+      in_instanceid = ureg_DECL_system_value(ureg, TGSI_SEMANTIC_INSTANCEID, 0);
+
+      if (!st->pbo_upload.use_gs)
+         out_layer = ureg_DECL_output(ureg, TGSI_SEMANTIC_LAYER, 0);
+   }
+
+   /* out_pos = in_pos */
+   ureg_MOV(ureg, out_pos, in_pos);
+
+   if (st->pbo_upload.upload_layers) {
+      if (st->pbo_upload.use_gs) {
+         /* out_pos.z = i2f(gl_InstanceID) */
+         ureg_I2F(ureg, ureg_writemask(out_pos, TGSI_WRITEMASK_Z),
+                        ureg_scalar(in_instanceid, TGSI_SWIZZLE_X));
+      } else {
+         /* out_layer = gl_InstanceID */
+         ureg_MOV(ureg, out_layer, in_instanceid);
+      }
+   }
+
+   ureg_END(ureg);
+
+   return ureg_create_shader_and_destroy(ureg, st->pipe);
+}
+
+static void *
+create_pbo_upload_gs(struct st_context *st)
+{
+   static const int zero = 0;
+   struct ureg_program *ureg;
+   struct ureg_dst out_pos;
+   struct ureg_dst out_layer;
+   struct ureg_src in_pos;
+   struct ureg_src imm;
+   unsigned i;
+
+   ureg = ureg_create(TGSI_PROCESSOR_GEOMETRY);
+   if (!ureg)
+      return NULL;
+
+   ureg_property(ureg, TGSI_PROPERTY_GS_INPUT_PRIM, PIPE_PRIM_TRIANGLES);
+   ureg_property(ureg, TGSI_PROPERTY_GS_OUTPUT_PRIM, PIPE_PRIM_TRIANGLE_STRIP);
+   ureg_property(ureg, TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES, 3);
+
+   out_pos = ureg_DECL_output(ureg, TGSI_SEMANTIC_POSITION, 0);
+   out_layer = ureg_DECL_output(ureg, TGSI_SEMANTIC_LAYER, 0);
+
+   in_pos = ureg_DECL_input(ureg, TGSI_SEMANTIC_POSITION, 0, 0, 1);
+
+   imm = ureg_DECL_immediate_int(ureg, &zero, 1);
+
+   for (i = 0; i < 3; ++i) {
+      struct ureg_src in_pos_vertex = ureg_src_dimension(in_pos, i);
+
+      /* out_pos = in_pos[i] */
+      ureg_MOV(ureg, out_pos, in_pos_vertex);
+
+      /* out_layer.x = f2i(in_pos[i].z) */
+      ureg_F2I(ureg, ureg_writemask(out_layer, TGSI_WRITEMASK_X),
+                     ureg_scalar(in_pos_vertex, TGSI_SWIZZLE_Z));
+
+      ureg_EMIT(ureg, ureg_scalar(imm, TGSI_SWIZZLE_X));
+   }
+
+   ureg_END(ureg);
+
+   return ureg_create_shader_and_destroy(ureg, st->pipe);
+}
+
+static void *
+create_pbo_upload_fs(struct st_context *st)
+{
+   struct pipe_context *pipe = st->pipe;
+   struct pipe_screen *screen = pipe->screen;
+   struct ureg_program *ureg;
+   struct ureg_dst out;
+   struct ureg_src sampler;
+   struct ureg_src pos;
+   struct ureg_src layer;
+   struct ureg_src const0;
+   struct ureg_dst temp0;
+
+   ureg = ureg_create(TGSI_PROCESSOR_FRAGMENT);
+   if (!ureg)
+      return NULL;
+
+   out     = ureg_DECL_output(ureg, TGSI_SEMANTIC_COLOR, 0);
+   sampler = ureg_DECL_sampler(ureg, 0);
+   if (screen->get_param(screen, PIPE_CAP_TGSI_FS_POSITION_IS_SYSVAL)) {
+      pos = ureg_DECL_system_value(ureg, TGSI_SEMANTIC_POSITION, 0);
+   } else {
+      pos = ureg_DECL_fs_input(ureg, TGSI_SEMANTIC_POSITION, 0,
+                               TGSI_INTERPOLATE_LINEAR);
+   }
+   if (st->pbo_upload.upload_layers) {
+      layer = ureg_DECL_fs_input(ureg, TGSI_SEMANTIC_LAYER, 0,
+                                       TGSI_INTERPOLATE_CONSTANT);
+   }
+   const0  = ureg_DECL_constant(ureg, 0);
+   temp0   = ureg_DECL_temporary(ureg);
+
+   /* Note: const0 = [ -xoffset + skip_pixels, -yoffset, stride, image_height ] */
+
+   /* temp0.xy = f2i(temp0.xy) */
+   ureg_F2I(ureg, ureg_writemask(temp0, TGSI_WRITEMASK_XY),
+                  ureg_swizzle(pos,
+                               TGSI_SWIZZLE_X, TGSI_SWIZZLE_Y,
+                               TGSI_SWIZZLE_Y, TGSI_SWIZZLE_Y));
+
+   /* temp0.xy = temp0.xy + const0.xy */
+   ureg_UADD(ureg, ureg_writemask(temp0, TGSI_WRITEMASK_XY),
+                   ureg_swizzle(ureg_src(temp0),
+                                TGSI_SWIZZLE_X, TGSI_SWIZZLE_Y,
+                                TGSI_SWIZZLE_Y, TGSI_SWIZZLE_Y),
+                   ureg_swizzle(const0,
+                                TGSI_SWIZZLE_X, TGSI_SWIZZLE_Y,
+                                TGSI_SWIZZLE_Y, TGSI_SWIZZLE_Y));
+
+   /* temp0.x = const0.z * temp0.y + temp0.x */
+   ureg_UMAD(ureg, ureg_writemask(temp0, TGSI_WRITEMASK_X),
+                   ureg_scalar(const0, TGSI_SWIZZLE_Z),
+                   ureg_scalar(ureg_src(temp0), TGSI_SWIZZLE_Y),
+                   ureg_scalar(ureg_src(temp0), TGSI_SWIZZLE_X));
+
+   if (st->pbo_upload.upload_layers) {
+      /* temp0.x = const0.w * layer + temp0.x */
+      ureg_UMAD(ureg, ureg_writemask(temp0, TGSI_WRITEMASK_X),
+                      ureg_scalar(const0, TGSI_SWIZZLE_W),
+                      ureg_scalar(layer, TGSI_SWIZZLE_X),
+                      ureg_scalar(ureg_src(temp0), TGSI_SWIZZLE_X));
+   }
+
+   /* temp0.w = 0 */
+   ureg_MOV(ureg, ureg_writemask(temp0, TGSI_WRITEMASK_W), ureg_imm1u(ureg, 0));
+
+   /* out = txf(sampler, temp0.x) */
+   ureg_TXF(ureg, out, TGSI_TEXTURE_BUFFER, ureg_src(temp0), sampler);
+
+   ureg_release_temporary(ureg, temp0);
+
+   ureg_END(ureg);
+
+   return ureg_create_shader_and_destroy(ureg, pipe);
+}
+
+static bool
+try_pbo_upload_common(struct gl_context *ctx,
+                      struct pipe_surface *surface,
+                      int xoffset, int yoffset,
+                      unsigned upload_width, unsigned upload_height,
+                      struct pipe_resource *buffer,
+                      enum pipe_format src_format,
+                      intptr_t buf_offset,
+                      unsigned bytes_per_pixel,
+                      unsigned stride,
+                      unsigned image_height)
+{
+   struct st_context *st = st_context(ctx);
+   struct cso_context *cso = st->cso_context;
+   struct pipe_context *pipe = st->pipe;
+   unsigned depth = surface->u.tex.last_layer - surface->u.tex.first_layer + 1;
+   unsigned skip_pixels = 0;
+   bool success = false;
+
+   /* Check alignment. */
+   {
+      unsigned ofs = (buf_offset * bytes_per_pixel) % ctx->Const.TextureBufferOffsetAlignment;
+      if (ofs != 0) {
+         if (ofs % bytes_per_pixel != 0)
+            return false;
+
+         skip_pixels = ofs / bytes_per_pixel;
+         buf_offset -= skip_pixels;
+      }
+   }
+
+   /* Create the shaders */
+   if (!st->pbo_upload.vs) {
+      st->pbo_upload.vs = create_pbo_upload_vs(st);
+      if (!st->pbo_upload.vs)
+         return false;
+   }
+
+   if (depth != 1 && st->pbo_upload.use_gs && !st->pbo_upload.gs) {
+      st->pbo_upload.gs = create_pbo_upload_gs(st);
+      if (!st->pbo_upload.gs)
+         return false;
+   }
+
+   if (!st->pbo_upload.fs) {
+      st->pbo_upload.fs = create_pbo_upload_fs(st);
+      if (!st->pbo_upload.fs)
+         return false;
+   }
+
+   cso_save_state(cso, (CSO_BIT_FRAGMENT_SAMPLER_VIEWS |
+                        CSO_BIT_FRAGMENT_SAMPLERS |
+                        CSO_BIT_VERTEX_ELEMENTS |
+                        CSO_BIT_AUX_VERTEX_BUFFER_SLOT |
+                        CSO_BIT_FRAMEBUFFER |
+                        CSO_BIT_VIEWPORT |
+                        CSO_BIT_BLEND |
+                        CSO_BIT_DEPTH_STENCIL_ALPHA |
+                        CSO_BIT_RASTERIZER |
+                        CSO_BIT_STREAM_OUTPUTS |
+                        CSO_BITS_ALL_SHADERS));
+   cso_save_constant_buffer_slot0(cso, PIPE_SHADER_FRAGMENT);
+
+
+   /* Set up the sampler_view */
+   {
+      unsigned first_element = buf_offset;
+      unsigned last_element = buf_offset + skip_pixels + upload_width - 1
+         + (upload_height - 1 + (depth - 1) * image_height) * stride;
+      struct pipe_sampler_view templ;
+      struct pipe_sampler_view *sampler_view;
+      struct pipe_sampler_state sampler = {0};
+      const struct pipe_sampler_state *samplers[1] = {&sampler};
+
+      /* This should be ensured by Mesa before calling our callbacks */
+      assert((last_element + 1) * bytes_per_pixel <= buffer->width0);
+
+      if (last_element - first_element > ctx->Const.MaxTextureBufferSize - 1)
+         goto fail;
+
+      memset(&templ, 0, sizeof(templ));
+      templ.target = PIPE_BUFFER;
+      templ.format = src_format;
+      templ.u.buf.first_element = first_element;
+      templ.u.buf.last_element = last_element;
+      templ.swizzle_r = PIPE_SWIZZLE_RED;
+      templ.swizzle_g = PIPE_SWIZZLE_GREEN;
+      templ.swizzle_b = PIPE_SWIZZLE_BLUE;
+      templ.swizzle_a = PIPE_SWIZZLE_ALPHA;
+
+      sampler_view = pipe->create_sampler_view(pipe, buffer, &templ);
+      if (sampler_view == NULL)
+         goto fail;
+
+      cso_set_sampler_views(cso, PIPE_SHADER_FRAGMENT, 1, &sampler_view);
+
+      pipe_sampler_view_reference(&sampler_view, NULL);
+
+      cso_set_samplers(cso, PIPE_SHADER_FRAGMENT, 1, samplers);
+   }
+
+   /* Upload vertices */
+   {
+      struct pipe_vertex_buffer vbo;
+      struct pipe_vertex_element velem;
+
+      float x0 = (float) xoffset / surface->width * 2.0f - 1.0f;
+      float y0 = (float) yoffset / surface->height * 2.0f - 1.0f;
+      float x1 = (float) (xoffset + upload_width) / surface->width * 2.0f - 1.0f;
+      float y1 = (float) (yoffset + upload_height) / surface->height * 2.0f - 1.0f;
+
+      float *verts = NULL;
+
+      vbo.user_buffer = NULL;
+      vbo.buffer = NULL;
+      vbo.stride = 2 * sizeof(float);
+
+      u_upload_alloc(st->uploader, 0, 8 * sizeof(float), 4,
+                     &vbo.buffer_offset, &vbo.buffer, (void **) &verts);
+      if (!verts)
+         goto fail;
+
+      verts[0] = x0;
+      verts[1] = y0;
+      verts[2] = x0;
+      verts[3] = y1;
+      verts[4] = x1;
+      verts[5] = y0;
+      verts[6] = x1;
+      verts[7] = y1;
+
+      u_upload_unmap(st->uploader);
+
+      velem.src_offset = 0;
+      velem.instance_divisor = 0;
+      velem.vertex_buffer_index = cso_get_aux_vertex_buffer_slot(cso);
+      velem.src_format = PIPE_FORMAT_R32G32_FLOAT;
+
+      cso_set_vertex_elements(cso, 1, &velem);
+
+      cso_set_vertex_buffers(cso, velem.vertex_buffer_index, 1, &vbo);
+
+      pipe_resource_reference(&vbo.buffer, NULL);
+   }
+
+   /* Upload constants */
+   /* Note: the user buffer must be valid until draw time */
+   struct {
+      int32_t xoffset;
+      int32_t yoffset;
+      int32_t stride;
+      int32_t image_size;
+   } constants;
+
+   {
+      struct pipe_constant_buffer cb;
+
+      constants.xoffset = -xoffset + skip_pixels;
+      constants.yoffset = -yoffset;
+      constants.stride = stride;
+      constants.image_size = stride * image_height;
+
+      if (st->constbuf_uploader) {
+         cb.buffer = NULL;
+         cb.user_buffer = NULL;
+         u_upload_data(st->constbuf_uploader, 0, sizeof(constants),
+                       ctx->Const.UniformBufferOffsetAlignment,
+                       &constants, &cb.buffer_offset, &cb.buffer);
+         if (!cb.buffer)
+            goto fail;
+
+         u_upload_unmap(st->constbuf_uploader);
+      } else {
+         cb.buffer = NULL;
+         cb.user_buffer = &constants;
+         cb.buffer_offset = 0;
+      }
+      cb.buffer_size = sizeof(constants);
+
+      cso_set_constant_buffer(cso, PIPE_SHADER_FRAGMENT, 0, &cb);
+
+      pipe_resource_reference(&cb.buffer, NULL);
+   }
+
+   /* Framebuffer_state */
+   {
+      struct pipe_framebuffer_state fb;
+      memset(&fb, 0, sizeof(fb));
+      fb.width = surface->width;
+      fb.height = surface->height;
+      fb.nr_cbufs = 1;
+      pipe_surface_reference(&fb.cbufs[0], surface);
+
+      cso_set_framebuffer(cso, &fb);
+
+      pipe_surface_reference(&fb.cbufs[0], NULL);
+   }
+
+   cso_set_viewport_dims(cso, surface->width, surface->height, FALSE);
+
+   /* Blend state */
+   cso_set_blend(cso, &st->pbo_upload.blend);
+
+   /* Depth/stencil/alpha state */
+   {
+      struct pipe_depth_stencil_alpha_state dsa;
+      memset(&dsa, 0, sizeof(dsa));
+      cso_set_depth_stencil_alpha(cso, &dsa);
+   }
+
+   /* Rasterizer state */
+   cso_set_rasterizer(cso, &st->pbo_upload.raster);
+
+   /* Set up the shaders */
+   cso_set_vertex_shader_handle(cso, st->pbo_upload.vs);
+
+   cso_set_geometry_shader_handle(cso, depth != 1 ? st->pbo_upload.gs : NULL);
+
+   cso_set_tessctrl_shader_handle(cso, NULL);
+
+   cso_set_tesseval_shader_handle(cso, NULL);
+
+   cso_set_fragment_shader_handle(cso, st->pbo_upload.fs);
+
+   /* Disable stream output */
+   cso_set_stream_outputs(cso, 0, NULL, 0);
+
+   if (depth == 1) {
+      cso_draw_arrays(cso, PIPE_PRIM_TRIANGLE_STRIP, 0, 4);
+   } else {
+      cso_draw_arrays_instanced(cso, PIPE_PRIM_TRIANGLE_STRIP,
+                                0, 4, 0, depth);
+   }
+
+   success = true;
+
+fail:
+   cso_restore_state(cso);
+   cso_restore_constant_buffer_slot0(cso, PIPE_SHADER_FRAGMENT);
+
+   return success;
+}
+
+static bool
+try_pbo_upload(struct gl_context *ctx, GLuint dims,
+               struct gl_texture_image *texImage,
+               GLenum format, GLenum type,
+               enum pipe_format dst_format,
+               GLint xoffset, GLint yoffset, GLint zoffset,
+               GLint width, GLint height, GLint depth,
+               const void *pixels,
+               const struct gl_pixelstore_attrib *unpack)
+{
+   struct st_context *st = st_context(ctx);
+   struct st_texture_image *stImage = st_texture_image(texImage);
+   struct st_texture_object *stObj = st_texture_object(texImage->TexObject);
+   struct pipe_resource *texture = stImage->pt;
+   struct pipe_context *pipe = st->pipe;
+   struct pipe_screen *screen = pipe->screen;
+   struct pipe_surface *surface = NULL;
+   enum pipe_format src_format;
+   const struct util_format_description *desc;
+   GLenum gl_target = texImage->TexObject->Target;
+   intptr_t buf_offset;
+   unsigned bytes_per_pixel;
+   unsigned stride, image_height;
+   bool success;
+
+   if (!st->pbo_upload.enabled)
+      return false;
+
+   /* From now on, we need the gallium representation of dimensions. */
+   if (gl_target == GL_TEXTURE_1D_ARRAY) {
+      depth = height;
+      height = 1;
+      zoffset = yoffset;
+      yoffset = 0;
+      image_height = 1;
+   } else {
+      image_height = unpack->ImageHeight > 0 ? unpack->ImageHeight : height;
+   }
+
+   if (depth != 1 && !st->pbo_upload.upload_layers)
+      return false;
+
+   /* Choose the source format. Initially, we do so without checking driver
+    * support at all because of the remapping we later perform and because
+    * at least the Radeon driver actually supports some formats for texture
+    * buffers which it doesn't support for regular textures. */
+   src_format = st_choose_matching_format(st, 0, format, type, unpack->SwapBytes);
+   if (!src_format) {
+      return false;
+   }
+
+   src_format = util_format_linear(src_format);
+   desc = util_format_description(src_format);
+
+   if (desc->layout != UTIL_FORMAT_LAYOUT_PLAIN)
+      return false;
+
+   if (desc->colorspace != UTIL_FORMAT_COLORSPACE_RGB)
+      return false;
+
+   if (st->pbo_upload.rgba_only) {
+      enum pipe_format orig_dst_format = dst_format;
+
+      if (!reinterpret_formats(&src_format, &dst_format)) {
+         return false;
+      }
+
+      if (dst_format != orig_dst_format &&
+          !screen->is_format_supported(screen, dst_format, PIPE_TEXTURE_2D, 0,
+                                       PIPE_BIND_RENDER_TARGET)) {
+         return false;
+      }
+   }
+
+   if (!src_format ||
+       !screen->is_format_supported(screen, src_format, PIPE_BUFFER, 0,
+                                    PIPE_BIND_SAMPLER_VIEW)) {
+      return false;
+   }
+
+   /* Check if the offset satisfies the alignment requirements */
+   buf_offset = (intptr_t) pixels;
+   bytes_per_pixel = desc->block.bits / 8;
+
+   if (buf_offset % bytes_per_pixel) {
+      return false;
+   }
+
+   /* Convert to texels */
+   buf_offset = buf_offset / bytes_per_pixel;
+
+   /* Compute the stride, taking unpack->Alignment into account */
+   {
+       unsigned pixels_per_row = unpack->RowLength > 0 ?
+                           unpack->RowLength : width;
+       unsigned bytes_per_row = pixels_per_row * bytes_per_pixel;
+       unsigned remainder = bytes_per_row % unpack->Alignment;
+       unsigned offset_rows;
+
+       if (remainder > 0)
+          bytes_per_row += (unpack->Alignment - remainder);
+
+       if (bytes_per_row % bytes_per_pixel) {
+          return false;
+       }
+
+       stride = bytes_per_row / bytes_per_pixel;
+
+       offset_rows = unpack->SkipRows;
+       if (dims == 3)
+          offset_rows += image_height * unpack->SkipImages;
+
+       buf_offset += unpack->SkipPixels + stride * offset_rows;
+   }
+
+   /* Set up the surface */
+   {
+      unsigned level = stObj->pt != stImage->pt ? 0 : texImage->TexObject->MinLevel + texImage->Level;
+      unsigned max_layer = util_max_layer(texture, level);
+
+      zoffset += texImage->Face + texImage->TexObject->MinLayer;
+
+      struct pipe_surface templ;
+      memset(&templ, 0, sizeof(templ));
+      templ.format = dst_format;
+      templ.u.tex.level = level;
+      templ.u.tex.first_layer = MIN2(zoffset, max_layer);
+      templ.u.tex.last_layer = MIN2(zoffset + depth - 1, max_layer);
+
+      surface = pipe->create_surface(pipe, texture, &templ);
+      if (!surface)
+         return false;
+   }
+
+   success = try_pbo_upload_common(ctx,  surface,
+                                   xoffset, yoffset, width, height,
+                                   st_buffer_object(unpack->BufferObj)->buffer,
+                                   src_format,
+                                   buf_offset,
+                                   bytes_per_pixel, stride, image_height);
+
+   pipe_surface_reference(&surface, NULL);
+
+   return success;
+}
 
 static void
 st_TexSubImage(struct gl_context *ctx, GLuint dims,
@@ -695,21 +1722,15 @@ st_TexSubImage(struct gl_context *ctx, GLuint dims,
       goto fallback;
    }
 
-   /* See if the texture format already matches the format and type,
-    * in which case the memcpy-based fast path will likely be used and
-    * we don't have to blit. */
-   if (_mesa_format_matches_format_and_type(texImage->TexFormat, format,
-                                            type, unpack->SwapBytes)) {
-      goto fallback;
-   }
 
+   /* See if the destination format is supported. */
    if (format == GL_DEPTH_COMPONENT || format == GL_DEPTH_STENCIL)
       bind = PIPE_BIND_DEPTH_STENCIL;
    else
       bind = PIPE_BIND_RENDER_TARGET;
 
-   /* See if the destination format is supported.
-    * For luminance and intensity, only the red channel is stored there. */
+   /* For luminance and intensity, only the red channel is stored
+    * in the destination. */
    dst_format = util_format_linear(dst->format);
    dst_format = util_format_luminance_to_red(dst_format);
    dst_format = util_format_intensity_to_red(dst_format);
@@ -720,6 +1741,21 @@ st_TexSubImage(struct gl_context *ctx, GLuint dims,
       goto fallback;
    }
 
+   if (_mesa_is_bufferobj(unpack->BufferObj)) {
+      if (try_pbo_upload(ctx, dims, texImage, format, type, dst_format,
+                         xoffset, yoffset, zoffset,
+                         width, height, depth, pixels, unpack))
+         return;
+   }
+
+   /* See if the texture format already matches the format and type,
+    * in which case the memcpy-based fast path will likely be used and
+    * we don't have to blit. */
+   if (_mesa_format_matches_format_and_type(texImage->TexFormat, format,
+                                            type, unpack->SwapBytes, NULL)) {
+      goto fallback;
+   }
+
    /* Choose the source format. */
    src_format = st_choose_matching_format(st, PIPE_BIND_SAMPLER_VIEW,
                                           format, type, unpack->SwapBytes);
@@ -742,6 +1778,11 @@ st_TexSubImage(struct gl_context *ctx, GLuint dims,
    if (gl_target == GL_TEXTURE_CUBE_MAP) {
       gl_target = GL_TEXTURE_2D;
    }
+   /* TexSubImage can specify subsets of cube map array faces
+    * so we need to upload via 2D array instead */
+   if (gl_target == GL_TEXTURE_CUBE_MAP_ARRAY) {
+      gl_target = GL_TEXTURE_2D_ARRAY;
+   }
 
    /* Initialize the source texture description. */
    memset(&src_templ, 0, sizeof(src_templ));
@@ -804,18 +1845,18 @@ st_TexSubImage(struct gl_context *ctx, GLuint dims,
             /* 1D array textures.
              * We need to convert gallium coords to GL coords.
              */
-            GLvoid *src = _mesa_image_address3d(unpack, pixels,
+            GLvoid *src = _mesa_image_address2d(unpack, pixels,
                                                 width, depth, format,
-                                                type, 0, slice, 0);
+                                                type, slice, 0);
             memcpy(map, src, bytesPerRow);
          }
          else {
             ubyte *slice_map = map;
 
             for (row = 0; row < (unsigned) height; row++) {
-               GLvoid *src = _mesa_image_address3d(unpack, pixels,
-                                                   width, height, format,
-                                                   type, slice, row, 0);
+               GLvoid *src = _mesa_image_address(dims, unpack, pixels,
+                                                 width, height, format,
+                                                 type, slice, row, 0);
                memcpy(slice_map, src, bytesPerRow);
                slice_map += transfer->stride;
             }
@@ -882,20 +1923,173 @@ st_TexImage(struct gl_context * ctx, GLuint dims,
 }
 
 
+static void
+st_CompressedTexSubImage(struct gl_context *ctx, GLuint dims,
+                         struct gl_texture_image *texImage,
+                         GLint x, GLint y, GLint z,
+                         GLsizei w, GLsizei h, GLsizei d,
+                         GLenum format, GLsizei imageSize, const GLvoid *data)
+{
+   struct st_context *st = st_context(ctx);
+   struct st_texture_image *stImage = st_texture_image(texImage);
+   struct st_texture_object *stObj = st_texture_object(texImage->TexObject);
+   struct pipe_resource *texture = stImage->pt;
+   struct pipe_context *pipe = st->pipe;
+   struct pipe_screen *screen = pipe->screen;
+   struct pipe_resource *dst = stImage->pt;
+   struct pipe_surface *surface = NULL;
+   struct compressed_pixelstore store;
+   enum pipe_format copy_format;
+   unsigned bytes_per_block;
+   unsigned bw, bh;
+   intptr_t buf_offset;
+   bool success = false;
+
+   /* Check basic pre-conditions for PBO upload */
+   if (!st->prefer_blit_based_texture_transfer) {
+      goto fallback;
+   }
+
+   if (!_mesa_is_bufferobj(ctx->Unpack.BufferObj))
+      goto fallback;
+
+   if ((_mesa_is_format_etc2(texImage->TexFormat) && !st->has_etc2) ||
+       (texImage->TexFormat == MESA_FORMAT_ETC1_RGB8 && !st->has_etc1)) {
+      /* ETC isn't supported and is represented by uncompressed formats. */
+      goto fallback;
+   }
+
+   if (!dst) {
+      goto fallback;
+   }
+
+   if (!st->pbo_upload.enabled ||
+       !screen->get_param(screen, PIPE_CAP_SURFACE_REINTERPRET_BLOCKS)) {
+      goto fallback;
+   }
+
+   /* Choose the pipe format for the upload. */
+   bytes_per_block = util_format_get_blocksize(dst->format);
+   bw = util_format_get_blockwidth(dst->format);
+   bh = util_format_get_blockheight(dst->format);
+
+   switch (bytes_per_block) {
+   case 8:
+      copy_format = PIPE_FORMAT_R16G16B16A16_UINT;
+      break;
+   case 16:
+      copy_format = PIPE_FORMAT_R32G32B32A32_UINT;
+      break;
+   default:
+      goto fallback;
+   }
+
+   if (!screen->is_format_supported(screen, copy_format, PIPE_BUFFER, 0,
+                                    PIPE_BIND_SAMPLER_VIEW)) {
+      goto fallback;
+   }
+
+   if (!screen->is_format_supported(screen, copy_format, dst->target,
+                                    dst->nr_samples, PIPE_BIND_RENDER_TARGET)) {
+      goto fallback;
+   }
+
+   /* Interpret the pixelstore settings. */
+   _mesa_compute_compressed_pixelstore(dims, texImage->TexFormat, w, h, d,
+                                       &ctx->Unpack, &store);
+   assert(store.CopyBytesPerRow % bytes_per_block == 0);
+   assert(store.SkipBytes % bytes_per_block == 0);
+
+   /* Compute the offset into the buffer */
+   buf_offset = (intptr_t)data + store.SkipBytes;
+
+   if (buf_offset % bytes_per_block) {
+      goto fallback;
+   }
+
+   buf_offset = buf_offset / bytes_per_block;
+
+   /* Set up the surface. */
+   {
+      unsigned level = stObj->pt != stImage->pt ? 0 : texImage->TexObject->MinLevel + texImage->Level;
+      unsigned max_layer = util_max_layer(texture, level);
+
+      z += texImage->Face + texImage->TexObject->MinLayer;
+
+      struct pipe_surface templ;
+      memset(&templ, 0, sizeof(templ));
+      templ.format = copy_format;
+      templ.u.tex.level = level;
+      templ.u.tex.first_layer = MIN2(z, max_layer);
+      templ.u.tex.last_layer = MIN2(z + d - 1, max_layer);
+
+      surface = pipe->create_surface(pipe, texture, &templ);
+      if (!surface)
+         goto fallback;
+   }
+
+   success = try_pbo_upload_common(ctx, surface,
+                                   x / bw, y / bh,
+                                   store.CopyBytesPerRow / bytes_per_block,
+                                   store.CopyRowsPerSlice,
+                                   st_buffer_object(ctx->Unpack.BufferObj)->buffer,
+                                   copy_format,
+                                   buf_offset,
+                                   bytes_per_block,
+                                   store.TotalBytesPerRow / bytes_per_block,
+                                   store.TotalRowsPerSlice);
+
+   pipe_surface_reference(&surface, NULL);
+
+   if (success)
+      return;
+
+fallback:
+   _mesa_store_compressed_texsubimage(ctx, dims, texImage,
+                                      x, y, z, w, h, d,
+                                      format, imageSize, data);
+}
+
 static void
 st_CompressedTexImage(struct gl_context *ctx, GLuint dims,
                       struct gl_texture_image *texImage,
                       GLsizei imageSize, const GLvoid *data)
 {
    prep_teximage(ctx, texImage, GL_NONE, GL_NONE);
-   _mesa_store_compressed_teximage(ctx, dims, texImage, imageSize, data);
+
+   /* only 2D and 3D compressed images are supported at this time */
+   if (dims == 1) {
+      _mesa_problem(ctx, "Unexpected glCompressedTexImage1D call");
+      return;
+   }
+
+   /* This is pretty simple, because unlike the general texstore path we don't
+    * have to worry about the usual image unpacking or image transfer
+    * operations.
+    */
+   assert(texImage);
+   assert(texImage->Width > 0);
+   assert(texImage->Height > 0);
+   assert(texImage->Depth > 0);
+
+   /* allocate storage for texture data */
+   if (!st_AllocTextureImageBuffer(ctx, texImage)) {
+      _mesa_error(ctx, GL_OUT_OF_MEMORY, "glCompressedTexImage%uD", dims);
+      return;
+   }
+
+   st_CompressedTexSubImage(ctx, dims, texImage,
+                            0, 0, 0,
+                            texImage->Width, texImage->Height, texImage->Depth,
+                            texImage->TexFormat,
+                            imageSize, data);
 }
 
 
 
 
 /**
- * Called via ctx->Driver.GetTexImage()
+ * Called via ctx->Driver.GetTexSubImage()
  *
  * This uses a blit to copy the texture to a texture format which matches
  * the format and type combo and then a fast read-back is done using memcpy.
@@ -909,16 +2103,15 @@ st_CompressedTexImage(struct gl_context *ctx, GLuint dims,
  *       we do here should be free in such cases.
  */
 static void
-st_GetTexImage(struct gl_context * ctx,
-               GLenum format, GLenum type, GLvoid * pixels,
-               struct gl_texture_image *texImage)
+st_GetTexSubImage(struct gl_context * ctx,
+                  GLint xoffset, GLint yoffset, GLint zoffset,
+                  GLsizei width, GLsizei height, GLint depth,
+                  GLenum format, GLenum type, GLvoid * pixels,
+                  struct gl_texture_image *texImage)
 {
    struct st_context *st = st_context(ctx);
    struct pipe_context *pipe = st->pipe;
    struct pipe_screen *screen = pipe->screen;
-   GLuint width = texImage->Width;
-   GLuint height = texImage->Height;
-   GLuint depth = texImage->Depth;
    struct st_texture_image *stImage = st_texture_image(texImage);
    struct st_texture_object *stObj = st_texture_object(texImage->TexObject);
    struct pipe_resource *src = stObj->pt;
@@ -949,7 +2142,7 @@ st_GetTexImage(struct gl_context * ctx,
 
    /* XXX Fallback to _mesa_GetTexImage_sw for depth-stencil formats
     * due to an incomplete stencil blit implementation in some drivers. */
-   if (format == GL_DEPTH_STENCIL) {
+   if (format == GL_DEPTH_STENCIL || format == GL_STENCIL_INDEX) {
       goto fallback;
    }
 
@@ -963,7 +2156,7 @@ st_GetTexImage(struct gl_context * ctx,
    /* See if the texture format already matches the format and type,
     * in which case the memcpy-based fast path will be used. */
    if (_mesa_format_matches_format_and_type(texImage->TexFormat, format,
-                                            type, ctx->Pack.SwapBytes)) {
+                                            type, ctx->Pack.SwapBytes, NULL)) {
       goto fallback;
    }
 
@@ -1053,7 +2246,7 @@ st_GetTexImage(struct gl_context * ctx,
       }
    }
 
-   /* create the destination texture */
+   /* create the destination texture of size (width X height X depth) */
    memset(&dst_templ, 0, sizeof(dst_templ));
    dst_templ.target = pipe_target;
    dst_templ.format = dst_format;
@@ -1071,10 +2264,16 @@ st_GetTexImage(struct gl_context * ctx,
 
    /* From now on, we need the gallium representation of dimensions. */
    if (gl_target == GL_TEXTURE_1D_ARRAY) {
+      zoffset = yoffset;
+      yoffset = 0;
       depth = height;
       height = 1;
    }
 
+   assert(texImage->Face == 0 ||
+          texImage->TexObject->MinLayer == 0 ||
+          zoffset == 0);
+
    memset(&blit, 0, sizeof(blit));
    blit.src.resource = src;
    blit.src.level = texImage->Level + texImage->TexObject->MinLevel;
@@ -1082,9 +2281,11 @@ st_GetTexImage(struct gl_context * ctx,
    blit.dst.resource = dst;
    blit.dst.level = 0;
    blit.dst.format = dst->format;
-   blit.src.box.x = blit.dst.box.x = 0;
-   blit.src.box.y = blit.dst.box.y = 0;
-   blit.src.box.z = texImage->Face + texImage->TexObject->MinLayer;
+   blit.src.box.x = xoffset;
+   blit.dst.box.x = 0;
+   blit.src.box.y = yoffset;
+   blit.dst.box.y = 0;
+   blit.src.box.z = texImage->Face + texImage->TexObject->MinLayer + zoffset;
    blit.dst.box.z = 0;
    blit.src.box.width = blit.dst.box.width = width;
    blit.src.box.height = blit.dst.box.height = height;
@@ -1108,7 +2309,7 @@ st_GetTexImage(struct gl_context * ctx,
 
    /* copy/pack data into user buffer */
    if (_mesa_format_matches_format_and_type(mesa_format, format, type,
-                                            ctx->Pack.SwapBytes)) {
+                                            ctx->Pack.SwapBytes, NULL)) {
       /* memcpy */
       const uint bytesPerRow = width * util_format_get_blocksize(dst_format);
       GLuint row, slice;
@@ -1152,7 +2353,7 @@ st_GetTexImage(struct gl_context * ctx,
       }
 
       if (ST_DEBUG & DEBUG_FALLBACK)
-         debug_printf("%s: fallback format translation\n", __FUNCTION__);
+         debug_printf("%s: fallback format translation\n", __func__);
 
       dstMesaFormat = _mesa_format_from_format_and_type(format, type);
       dstStride = _mesa_image_row_stride(&ctx->Pack, width, format, type);
@@ -1205,7 +2406,9 @@ end:
 
 fallback:
    if (!done) {
-      _mesa_GetTexImage_sw(ctx, format, type, pixels, texImage);
+      _mesa_GetTexSubImage_sw(ctx, xoffset, yoffset, zoffset,
+                              width, height, depth,
+                              format, type, pixels, texImage);
    }
 }
 
@@ -1238,7 +2441,7 @@ fallback_copy_texsubimage(struct gl_context *ctx,
    struct pipe_transfer *transfer;
 
    if (ST_DEBUG & DEBUG_FALLBACK)
-      debug_printf("%s: fallback processing\n", __FUNCTION__);
+      debug_printf("%s: fallback processing\n", __func__);
 
    if (st_fb_orientation(ctx->ReadBuffer) == Y_0_TOP) {
       srcY = strb->Base.Height - srcY - height;
@@ -1393,7 +2596,7 @@ st_CopyTexSubImage(struct gl_context *ctx, GLuint dims,
           texImage->TexFormat != MESA_FORMAT_ETC1_RGB8);
 
    if (!strb || !strb->surface || !stImage->pt) {
-      debug_printf("%s: null strb or stImage\n", __FUNCTION__);
+      debug_printf("%s: null strb or stImage\n", __func__);
       return;
    }
 
@@ -1521,23 +2724,6 @@ copy_image_data_to_texture(struct st_context *st,
 
       pipe_resource_reference(&stImage->pt, NULL);
    }
-   else if (stImage->TexData) {
-      /* Copy from malloc'd memory */
-      /* XXX this should be re-examined/tested with a compressed format */
-      GLuint blockSize = util_format_get_blocksize(stObj->pt->format);
-      GLuint srcRowStride = stImage->base.Width * blockSize;
-      GLuint srcSliceStride = stImage->base.Height * srcRowStride;
-      st_texture_image_data(st,
-                            stObj->pt,
-                            stImage->base.Face,
-                            dstLevel,
-                            stImage->TexData,
-                            srcRowStride,
-                            srcSliceStride);
-      _mesa_align_free(stImage->TexData);
-      stImage->TexData = NULL;
-   }
-
    pipe_resource_reference(&stImage->pt, stObj->pt);
 }
 
@@ -1554,7 +2740,7 @@ st_finalize_texture(struct gl_context *ctx,
 {
    struct st_context *st = st_context(ctx);
    struct st_texture_object *stObj = st_texture_object(tObj);
-   const GLuint nr_faces = (stObj->base.Target == GL_TEXTURE_CUBE_MAP) ? 6 : 1;
+   const GLuint nr_faces = _mesa_num_tex_faces(stObj->base.Target);
    GLuint face;
    const struct st_texture_image *firstImage;
    enum pipe_format firstImageFormat;
@@ -1633,6 +2819,12 @@ st_finalize_texture(struct gl_context *ctx,
          width = stObj->width0;
          height = stObj->height0;
          depth = stObj->depth0;
+      } else {
+         /* The width/height/depth may have been previously reset in
+          * guess_and_alloc_texture. */
+         stObj->width0 = width;
+         stObj->height0 = height;
+         stObj->depth0 = depth;
       }
       /* convert GL dims to Gallium dims */
       st_gl_texture_dims_to_pipe_dims(stObj->base.Target, width, height, depth,
@@ -1694,10 +2886,13 @@ st_finalize_texture(struct gl_context *ctx,
          /* Need to import images in main memory or held in other textures.
           */
          if (stImage && stObj->pt != stImage->pt) {
+            GLuint depth = stObj->depth0;
+            if (stObj->base.Target == GL_TEXTURE_3D)
+               depth = u_minify(depth, level);
             if (level == 0 ||
                 (stImage->base.Width == u_minify(stObj->width0, level) &&
                  stImage->base.Height == u_minify(stObj->height0, level) &&
-                 stImage->base.Depth == u_minify(stObj->depth0, level))) {
+                 stImage->base.Depth == depth)) {
                /* src image fits expected dest mipmap level size */
                copy_image_data_to_texture(st, stObj, level, stImage);
             }
@@ -1880,23 +3075,50 @@ st_TextureView(struct gl_context *ctx,
    return GL_TRUE;
 }
 
+static void
+st_ClearTexSubImage(struct gl_context *ctx,
+                    struct gl_texture_image *texImage,
+                    GLint xoffset, GLint yoffset, GLint zoffset,
+                    GLsizei width, GLsizei height, GLsizei depth,
+                    const GLvoid *clearValue)
+{
+   static const char zeros[16] = {0};
+   struct st_texture_image *stImage = st_texture_image(texImage);
+   struct pipe_resource *pt = stImage->pt;
+   struct st_context *st = st_context(ctx);
+   struct pipe_context *pipe = st->pipe;
+   unsigned level = texImage->Level;
+   struct pipe_box box;
+
+   if (!pt)
+      return;
+
+   u_box_3d(xoffset, yoffset, zoffset + texImage->Face,
+            width, height, depth, &box);
+   if (texImage->TexObject->Immutable) {
+      level += texImage->TexObject->MinLevel;
+      box.z += texImage->TexObject->MinLayer;
+   }
+
+   pipe->clear_texture(pipe, pt, level, &box, clearValue ? clearValue : zeros);
+}
 
 void
 st_init_texture_functions(struct dd_function_table *functions)
 {
    functions->ChooseTextureFormat = st_ChooseTextureFormat;
-   functions->QuerySamplesForFormat = st_QuerySamplesForFormat;
+   functions->QueryInternalFormat = st_QueryInternalFormat;
    functions->TexImage = st_TexImage;
    functions->TexSubImage = st_TexSubImage;
-   functions->CompressedTexSubImage = _mesa_store_compressed_texsubimage;
+   functions->CompressedTexSubImage = st_CompressedTexSubImage;
    functions->CopyTexSubImage = st_CopyTexSubImage;
    functions->GenerateMipmap = st_generate_mipmap;
 
-   functions->GetTexImage = st_GetTexImage;
+   functions->GetTexSubImage = st_GetTexSubImage;
 
    /* compressed texture functions */
    functions->CompressedTexImage = st_CompressedTexImage;
-   functions->GetCompressedTexImage = _mesa_GetCompressedTexImage_sw;
+   functions->GetCompressedTexSubImage = _mesa_GetCompressedTexSubImage_sw;
 
    functions->NewTextureObject = st_NewTextureObject;
    functions->NewTextureImage = st_NewTextureImage;
@@ -1912,4 +3134,5 @@ st_init_texture_functions(struct dd_function_table *functions)
 
    functions->AllocTextureStorage = st_AllocTextureStorage;
    functions->TextureView = st_TextureView;
+   functions->ClearTexSubImage = st_ClearTexSubImage;
 }