st/mesa: provide GL_OES_copy_image support by caching the original ETC data
[mesa.git] / src / mesa / state_tracker / st_cb_texture.c
index d1c820b18f812529abf59501e698b2bbb19c3603..c2f5fdc031a5f7ce7f4355e3b10479e3cc74e7cd 100644 (file)
@@ -48,6 +48,7 @@
 
 #include "state_tracker/st_debug.h"
 #include "state_tracker/st_context.h"
+#include "state_tracker/st_cb_bitmap.h"
 #include "state_tracker/st_cb_fbo.h"
 #include "state_tracker/st_cb_flush.h"
 #include "state_tracker/st_cb_texture.h"
@@ -186,8 +187,39 @@ st_FreeTextureImageBuffer(struct gl_context *ctx,
    free(stImage->transfer);
    stImage->transfer = NULL;
    stImage->num_transfers = 0;
+
+   if (stImage->etc_data) {
+      free(stImage->etc_data);
+      stImage->etc_data = NULL;
+   }
 }
 
+bool
+st_etc_fallback(struct st_context *st, struct gl_texture_image *texImage)
+{
+   return (_mesa_is_format_etc2(texImage->TexFormat) && !st->has_etc2) ||
+          (texImage->TexFormat == MESA_FORMAT_ETC1_RGB8 && !st->has_etc1);
+}
+
+static void
+etc_fallback_allocate(struct st_context *st, struct st_texture_image *stImage)
+{
+   struct gl_texture_image *texImage = &stImage->base;
+
+   if (!st_etc_fallback(st, texImage))
+      return;
+
+   if (stImage->etc_data)
+      free(stImage->etc_data);
+
+   unsigned data_size = _mesa_format_image_size(texImage->TexFormat,
+                                                texImage->Width2,
+                                                texImage->Height2,
+                                                texImage->Depth2);
+
+   stImage->etc_data =
+      malloc(data_size * _mesa_num_tex_faces(texImage->TexObject->Target));
+}
 
 /** called via ctx->Driver.MapTextureImage() */
 static void
@@ -214,26 +246,24 @@ st_MapTextureImage(struct gl_context *ctx,
    map = st_texture_image_map(st, stImage, pipeMode, x, y, slice, w, h, 1,
                               &transfer);
    if (map) {
-      if ((_mesa_is_format_etc2(texImage->TexFormat) && !st->has_etc2) ||
-          (texImage->TexFormat == MESA_FORMAT_ETC1_RGB8 && !st->has_etc1)) {
-         /* ETC isn't supported by gallium and it's represented
-          * by uncompressed formats. Only write transfers with precompressed
-          * data are supported by ES3, which makes this really simple.
+      if (st_etc_fallback(st, texImage)) {
+         /* ETC isn't supported by all gallium drivers, where it's represented
+          * by uncompressed formats. We store the compressed data (as it's
+          * needed for image copies in OES_copy_image), and decompress as
+          * necessary in Unmap.
           *
-          * Just create a temporary storage where the ETC texture will
-          * be stored. It will be decompressed in the Unmap function.
+          * Note: all ETC1/ETC2 formats have 4x4 block sizes.
           */
          unsigned z = transfer->box.z;
          struct st_texture_image_transfer *itransfer = &stImage->transfer[z];
 
-         itransfer->temp_data =
-            malloc(_mesa_format_image_size(texImage->TexFormat, w, h, 1));
-         itransfer->temp_stride =
-            _mesa_format_row_stride(texImage->TexFormat, w);
+         unsigned bytes = _mesa_get_format_bytes(texImage->TexFormat);
+         unsigned stride = *rowStrideOut = itransfer->temp_stride =
+            _mesa_format_row_stride(texImage->TexFormat, texImage->Width2);
+         *mapOut = itransfer->temp_data =
+            stImage->etc_data + ((x / 4) * bytes + (y / 4) * stride) +
+            z * stride * texImage->Height2 / 4;
          itransfer->map = map;
-
-         *mapOut = itransfer->temp_data;
-         *rowStrideOut = itransfer->temp_stride;
       }
       else {
          /* supported mapping */
@@ -257,8 +287,7 @@ st_UnmapTextureImage(struct gl_context *ctx,
    struct st_context *st = st_context(ctx);
    struct st_texture_image *stImage  = st_texture_image(texImage);
 
-   if ((_mesa_is_format_etc2(texImage->TexFormat) && !st->has_etc2) ||
-       (texImage->TexFormat == MESA_FORMAT_ETC1_RGB8 && !st->has_etc1)) {
+   if (st_etc_fallback(st, texImage)) {
       /* Decompress the ETC texture to the mapped one. */
       unsigned z = slice + stImage->base.Face;
       struct st_texture_image_transfer *itransfer = &stImage->transfer[z];
@@ -266,20 +295,21 @@ st_UnmapTextureImage(struct gl_context *ctx,
 
       assert(z == transfer->box.z);
 
-      if (texImage->TexFormat == MESA_FORMAT_ETC1_RGB8) {
-         _mesa_etc1_unpack_rgba8888(itransfer->map, transfer->stride,
-                                    itransfer->temp_data,
-                                    itransfer->temp_stride,
-                                    transfer->box.width, transfer->box.height);
-      }
-      else {
-         _mesa_unpack_etc2_format(itransfer->map, transfer->stride,
-                                  itransfer->temp_data, itransfer->temp_stride,
-                                  transfer->box.width, transfer->box.height,
-                                  texImage->TexFormat);
+      if (transfer->usage & PIPE_TRANSFER_WRITE) {
+         if (texImage->TexFormat == MESA_FORMAT_ETC1_RGB8) {
+            _mesa_etc1_unpack_rgba8888(itransfer->map, transfer->stride,
+                                       itransfer->temp_data,
+                                       itransfer->temp_stride,
+                                       transfer->box.width, transfer->box.height);
+         }
+         else {
+            _mesa_unpack_etc2_format(itransfer->map, transfer->stride,
+                                     itransfer->temp_data, itransfer->temp_stride,
+                                     transfer->box.width, transfer->box.height,
+                                     texImage->TexFormat);
+         }
       }
 
-      free(itransfer->temp_data);
       itransfer->temp_data = NULL;
       itransfer->temp_stride = 0;
       itransfer->map = 0;
@@ -456,23 +486,46 @@ guess_and_alloc_texture(struct st_context *st,
                        struct st_texture_object *stObj,
                        const struct st_texture_image *stImage)
 {
+   const struct gl_texture_image *firstImage;
    GLuint lastLevel, width, height, depth;
    GLuint bindings;
    GLuint ptWidth, ptHeight, ptDepth, ptLayers;
    enum pipe_format fmt;
+   bool guessed_box = false;
 
    DBG("%s\n", __func__);
 
    assert(!stObj->pt);
 
-   if (!guess_base_level_size(stObj->base.Target,
-                              stImage->base.Width2,
-                              stImage->base.Height2,
-                              stImage->base.Depth2,
-                              stImage->base.Level,
-                              &width, &height, &depth)) {
+   /* If a base level image with compatible size exists, use that as our guess.
+    */
+   firstImage = _mesa_base_tex_image(&stObj->base);
+   if (firstImage &&
+       firstImage->Width2 > 0 &&
+       firstImage->Height2 > 0 &&
+       firstImage->Depth2 > 0 &&
+       guess_base_level_size(stObj->base.Target,
+                             firstImage->Width2,
+                             firstImage->Height2,
+                             firstImage->Depth2,
+                             firstImage->Level,
+                             &width, &height, &depth)) {
+      if (stImage->base.Width2 == u_minify(width, stImage->base.Level) &&
+          stImage->base.Height2 == u_minify(height, stImage->base.Level) &&
+          stImage->base.Depth2 == u_minify(depth, stImage->base.Level))
+         guessed_box = true;
+   }
+
+   if (!guessed_box)
+      guessed_box = guess_base_level_size(stObj->base.Target,
+                                          stImage->base.Width2,
+                                          stImage->base.Height2,
+                                          stImage->base.Depth2,
+                                          stImage->base.Level,
+                                          &width, &height, &depth);
+
+   if (!guessed_box) {
       /* we can't determine the image size at level=0 */
-      stObj->width0 = stObj->height0 = stObj->depth0 = 0;
       /* this is not an out of memory error */
       return GL_TRUE;
    }
@@ -497,11 +550,6 @@ guess_and_alloc_texture(struct st_context *st,
       lastLevel = 0;
    }
 
-   /* Save the level=0 dimensions */
-   stObj->width0 = width;
-   stObj->height0 = height;
-   stObj->depth0 = depth;
-
    fmt = st_mesa_format_to_pipe_format(st, stImage->base.TexFormat);
 
    bindings = default_bindings(st, fmt);
@@ -549,6 +597,8 @@ st_AllocTextureImageBuffer(struct gl_context *ctx,
 
    assert(!stImage->pt); /* xxx this might be wrong */
 
+   etc_fallback_allocate(st, stImage);
+
    /* Look if the parent texture object has space for this image */
    if (stObj->pt &&
        level <= stObj->pt->last_level &&
@@ -1070,128 +1120,20 @@ reinterpret_formats(enum pipe_format *src_format, enum pipe_format *dst_format)
    return true;
 }
 
-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(PIPE_SHADER_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.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.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)
+                      const struct st_pbo_addresses *addr,
+                      enum pipe_format src_format)
 {
    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.vs) {
-      st->pbo.vs = st_pbo_create_vs(st);
-      if (!st->pbo.vs)
-         return false;
-   }
-
-   if (depth != 1 && st->pbo.use_gs && !st->pbo.gs) {
-      st->pbo.gs = st_pbo_create_gs(st);
-      if (!st->pbo.gs)
-         return false;
-   }
-
+   /* Create fragment shader */
    if (!st->pbo.upload_fs) {
-      st->pbo.upload_fs = create_pbo_upload_fs(st);
+      st->pbo.upload_fs = st_pbo_create_upload_fs(st);
       if (!st->pbo.upload_fs)
          return false;
    }
@@ -1213,31 +1155,22 @@ try_pbo_upload_common(struct gl_context *ctx,
 
    /* 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.u.buf.first_element = addr->first_element;
+      templ.u.buf.last_element = addr->last_element;
       templ.swizzle_r = PIPE_SWIZZLE_X;
       templ.swizzle_g = PIPE_SWIZZLE_Y;
       templ.swizzle_b = PIPE_SWIZZLE_Z;
       templ.swizzle_a = PIPE_SWIZZLE_W;
 
-      sampler_view = pipe->create_sampler_view(pipe, buffer, &templ);
+      sampler_view = pipe->create_sampler_view(pipe, addr->buffer, &templ);
       if (sampler_view == NULL)
          goto fail;
 
@@ -1248,89 +1181,6 @@ try_pbo_upload_common(struct gl_context *ctx,
       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;
@@ -1357,31 +1207,10 @@ try_pbo_upload_common(struct gl_context *ctx,
       cso_set_depth_stencil_alpha(cso, &dsa);
    }
 
-   /* Rasterizer state */
-   cso_set_rasterizer(cso, &st->pbo.raster);
-
-   /* Set up the shaders */
-   cso_set_vertex_shader_handle(cso, st->pbo.vs);
-
-   cso_set_geometry_shader_handle(cso, depth != 1 ? st->pbo.gs : NULL);
-
-   cso_set_tessctrl_shader_handle(cso, NULL);
-
-   cso_set_tesseval_shader_handle(cso, NULL);
-
+   /* Set up the fragment shader */
    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;
+   success = st_pbo_draw(st, addr, surface->width, surface->height);
 
 fail:
    cso_restore_state(cso);
@@ -1407,12 +1236,10 @@ try_pbo_upload(struct gl_context *ctx, GLuint dims,
    struct pipe_context *pipe = st->pipe;
    struct pipe_screen *screen = pipe->screen;
    struct pipe_surface *surface = NULL;
+   struct st_pbo_addresses addr;
    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)
@@ -1424,9 +1251,6 @@ try_pbo_upload(struct gl_context *ctx, GLuint dims,
       height = 1;
       zoffset = yoffset;
       yoffset = 0;
-      image_height = 1;
-   } else {
-      image_height = unpack->ImageHeight > 0 ? unpack->ImageHeight : height;
    }
 
    if (depth != 1 && !st->pbo.layers)
@@ -1470,40 +1294,17 @@ try_pbo_upload(struct gl_context *ctx, GLuint dims,
       return false;
    }
 
-   /* Check if the offset satisfies the alignment requirements */
-   buf_offset = (intptr_t) pixels;
-   bytes_per_pixel = desc->block.bits / 8;
+   /* Compute buffer addresses */
+   addr.xoffset = xoffset;
+   addr.yoffset = yoffset;
+   addr.width = width;
+   addr.height = height;
+   addr.depth = depth;
+   addr.bytes_per_pixel = desc->block.bits / 8;
 
-   if (buf_offset % bytes_per_pixel) {
+   if (!st_pbo_addresses_pixelstore(st, gl_target, dims == 3, unpack, pixels,
+                                    &addr))
       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 */
    {
@@ -1524,12 +1325,7 @@ try_pbo_upload(struct gl_context *ctx, GLuint dims,
          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);
+   success = try_pbo_upload_common(ctx, surface, &addr, src_format);
 
    pipe_surface_reference(&surface, NULL);
 
@@ -1562,6 +1358,9 @@ st_TexSubImage(struct gl_context *ctx, GLuint dims,
    unsigned dstz = texImage->Face + texImage->TexObject->MinLayer;
    unsigned dst_level = 0;
 
+   st_flush_bitmap_cache(st);
+   st_invalidate_readpix_cache(st);
+
    if (stObj->pt == stImage->pt)
       dst_level = texImage->TexObject->MinLevel + texImage->Level;
 
@@ -1571,7 +1370,7 @@ st_TexSubImage(struct gl_context *ctx, GLuint dims,
    if (!dst)
       goto fallback;
 
-   /* Try transfer_inline_write, which should be the fastest memcpy path. */
+   /* Try texture_subdata, which should be the fastest memcpy path. */
    if (pixels &&
        !_mesa_is_bufferobj(unpack->BufferObj) &&
        _mesa_texstore_can_use_memcpy(ctx, texImage->_BaseFormat,
@@ -1597,8 +1396,8 @@ st_TexSubImage(struct gl_context *ctx, GLuint dims,
       }
 
       u_box_3d(xoffset, yoffset, zoffset + dstz, width, height, depth, &box);
-      pipe->transfer_inline_write(pipe, dst, dst_level, 0,
-                                  &box, data, stride, layer_stride);
+      pipe->texture_subdata(pipe, dst, dst_level, 0,
+                            &box, data, stride, layer_stride);
       return;
    }
 
@@ -1836,8 +1635,8 @@ st_CompressedTexSubImage(struct gl_context *ctx, GLuint dims,
    struct pipe_resource *dst = stImage->pt;
    struct pipe_surface *surface = NULL;
    struct compressed_pixelstore store;
+   struct st_pbo_addresses addr;
    enum pipe_format copy_format;
-   unsigned bytes_per_block;
    unsigned bw, bh;
    intptr_t buf_offset;
    bool success = false;
@@ -1850,8 +1649,7 @@ st_CompressedTexSubImage(struct gl_context *ctx, GLuint dims,
    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)) {
+   if (st_etc_fallback(st, texImage)) {
       /* ETC isn't supported and is represented by uncompressed formats. */
       goto fallback;
    }
@@ -1866,11 +1664,11 @@ st_CompressedTexSubImage(struct gl_context *ctx, GLuint dims,
    }
 
    /* Choose the pipe format for the upload. */
-   bytes_per_block = util_format_get_blocksize(dst->format);
+   addr.bytes_per_pixel = util_format_get_blocksize(dst->format);
    bw = util_format_get_blockwidth(dst->format);
    bh = util_format_get_blockheight(dst->format);
 
-   switch (bytes_per_block) {
+   switch (addr.bytes_per_pixel) {
    case 8:
       copy_format = PIPE_FORMAT_R16G16B16A16_UINT;
       break;
@@ -1894,17 +1692,29 @@ st_CompressedTexSubImage(struct gl_context *ctx, GLuint dims,
    /* 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);
+   assert(store.CopyBytesPerRow % addr.bytes_per_pixel == 0);
+   assert(store.SkipBytes % addr.bytes_per_pixel == 0);
 
    /* Compute the offset into the buffer */
    buf_offset = (intptr_t)data + store.SkipBytes;
 
-   if (buf_offset % bytes_per_block) {
+   if (buf_offset % addr.bytes_per_pixel) {
       goto fallback;
    }
 
-   buf_offset = buf_offset / bytes_per_block;
+   buf_offset = buf_offset / addr.bytes_per_pixel;
+
+   addr.xoffset = x / bw;
+   addr.yoffset = y / bh;
+   addr.width = store.CopyBytesPerRow / addr.bytes_per_pixel;
+   addr.height = store.CopyRowsPerSlice;
+   addr.depth = d;
+   addr.pixels_per_row = store.TotalBytesPerRow / addr.bytes_per_pixel;
+   addr.image_height = store.TotalRowsPerSlice;
+
+   if (!st_pbo_addresses_setup(st, st_buffer_object(ctx->Unpack.BufferObj)->buffer,
+                               buf_offset, &addr))
+      goto fallback;
 
    /* Set up the surface. */
    {
@@ -1925,16 +1735,7 @@ st_CompressedTexSubImage(struct gl_context *ctx, GLuint dims,
          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);
+   success = try_pbo_upload_common(ctx, surface, &addr, copy_format);
 
    pipe_surface_reference(&surface, NULL);
 
@@ -2027,6 +1828,8 @@ st_GetTexSubImage(struct gl_context * ctx,
    assert(!_mesa_is_format_etc2(texImage->TexFormat) &&
           texImage->TexFormat != MESA_FORMAT_ETC1_RGB8);
 
+   st_flush_bitmap_cache(st);
+
    if (!st->prefer_blit_based_texture_transfer &&
        !_mesa_is_format_compressed(texImage->TexFormat)) {
       /* Try to avoid the fallback if we're doing texture decompression here */
@@ -2490,6 +2293,9 @@ st_CopyTexSubImage(struct gl_context *ctx, GLuint dims,
    unsigned bind;
    GLint srcY0, srcY1;
 
+   st_flush_bitmap_cache(st);
+   st_invalidate_readpix_cache(st);
+
    assert(!_mesa_is_format_etc2(texImage->TexFormat) &&
           texImage->TexFormat != MESA_FORMAT_ETC1_RGB8);
 
@@ -2673,9 +2479,6 @@ st_finalize_texture(struct gl_context *ctx,
       if (st_obj->buffer != stObj->pt) {
          pipe_resource_reference(&stObj->pt, st_obj->buffer);
          st_texture_release_all_sampler_views(st, stObj);
-         stObj->width0 = stObj->pt->width0 / _mesa_get_format_bytes(tObj->_BufferObjectFormat);
-         stObj->height0 = 1;
-         stObj->depth0 = 1;
       }
       return GL_TRUE;
 
@@ -2708,25 +2511,44 @@ st_finalize_texture(struct gl_context *ctx,
    /* Find size of level=0 Gallium mipmap image, plus number of texture layers */
    {
       GLuint width, height, depth;
-      if (!guess_base_level_size(stObj->base.Target,
-                                 firstImage->base.Width2,
-                                 firstImage->base.Height2,
-                                 firstImage->base.Depth2,
-                                 firstImage->base.Level,
-                                 &width, &height, &depth)) {
-         width = stObj->width0;
-         height = stObj->height0;
-         depth = stObj->depth0;
+
+      st_gl_texture_dims_to_pipe_dims(stObj->base.Target,
+                                      firstImage->base.Width2,
+                                      firstImage->base.Height2,
+                                      firstImage->base.Depth2,
+                                      &width, &height, &depth, &ptLayers);
+
+      /* If we previously allocated a pipe texture and its sizes are
+       * compatible, use them.
+       */
+      if (stObj->pt &&
+          u_minify(stObj->pt->width0, firstImage->base.Level) == width &&
+          u_minify(stObj->pt->height0, firstImage->base.Level) == height &&
+          u_minify(stObj->pt->depth0, firstImage->base.Level) == depth) {
+         ptWidth = stObj->pt->width0;
+         ptHeight = stObj->pt->height0;
+         ptDepth = stObj->pt->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;
+         /* Otherwise, compute a new level=0 size that is compatible with the
+          * base level image.
+          */
+         ptWidth = width > 1 ? width << firstImage->base.Level : 1;
+         ptHeight = height > 1 ? height << firstImage->base.Level : 1;
+         ptDepth = depth > 1 ? depth << firstImage->base.Level : 1;
+
+         /* If the base level image is 1x1x1, we still need to ensure that the
+          * resulting pipe texture ends up with the required number of levels
+          * in total.
+          */
+         if (ptWidth == 1 && ptHeight == 1 && ptDepth == 1) {
+            ptWidth <<= firstImage->base.Level;
+
+            if (stObj->base.Target == GL_TEXTURE_CUBE_MAP ||
+                stObj->base.Target == GL_TEXTURE_CUBE_MAP_ARRAY)
+               ptHeight = ptWidth;
+         }
       }
-      /* convert GL dims to Gallium dims */
-      st_gl_texture_dims_to_pipe_dims(stObj->base.Target, width, height, depth,
-                                      &ptWidth, &ptHeight, &ptDepth, &ptLayers);
+
       ptNumSamples = firstImage->base.NumSamples;
    }
 
@@ -2748,7 +2570,7 @@ st_finalize_texture(struct gl_context *ctx,
           */
          pipe_resource_reference(&stObj->pt, NULL);
          st_texture_release_all_sampler_views(st, stObj);
-         st->dirty.st |= ST_NEW_FRAMEBUFFER;
+         st->dirty |= ST_NEW_FRAMEBUFFER;
       }
    }
 
@@ -2784,16 +2606,23 @@ 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 height = stObj->height0;
-            GLuint depth = stObj->depth0;
+            GLuint height;
+            GLuint depth;
 
             if (stObj->base.Target != GL_TEXTURE_1D_ARRAY)
-               height = u_minify(height, level);
+               height = u_minify(ptHeight, level);
+            else
+               height = ptLayers;
+
             if (stObj->base.Target == GL_TEXTURE_3D)
-               depth = u_minify(depth, level);
+               depth = u_minify(ptDepth, level);
+            else if (stObj->base.Target == GL_TEXTURE_CUBE_MAP)
+               depth = 1;
+            else
+               depth = ptLayers;
 
             if (level == 0 ||
-                (stImage->base.Width == u_minify(stObj->width0, level) &&
+                (stImage->base.Width == u_minify(ptWidth, level) &&
                  stImage->base.Height == height &&
                  stImage->base.Depth == depth)) {
                /* src image fits expected dest mipmap level size */
@@ -2829,10 +2658,6 @@ st_AllocTextureStorage(struct gl_context *ctx,
 
    assert(levels > 0);
 
-   /* Save the level=0 dimensions */
-   stObj->width0 = width;
-   stObj->height0 = height;
-   stObj->depth0 = depth;
    stObj->lastLevel = levels - 1;
 
    fmt = st_mesa_format_to_pipe_format(st, texImage->TexFormat);
@@ -2882,6 +2707,8 @@ st_AllocTextureStorage(struct gl_context *ctx,
          struct st_texture_image *stImage =
             st_texture_image(texObj->Image[face][level]);
          pipe_resource_reference(&stImage->pt, stObj->pt);
+
+         etc_fallback_allocate(st, stImage);
       }
    }
 
@@ -2891,9 +2718,9 @@ st_AllocTextureStorage(struct gl_context *ctx,
 
 static GLboolean
 st_TestProxyTexImage(struct gl_context *ctx, GLenum target,
-                     GLint level, mesa_format format,
-                     GLint width, GLint height,
-                     GLint depth, GLint border)
+                     GLuint numLevels, GLint level,
+                     mesa_format format, GLuint numSamples,
+                     GLint width, GLint height, GLint depth)
 {
    struct st_context *st = st_context(ctx);
    struct pipe_context *pipe = st->pipe;
@@ -2915,14 +2742,19 @@ st_TestProxyTexImage(struct gl_context *ctx, GLenum target,
 
       pt.target = gl_target_to_pipe(target);
       pt.format = st_mesa_format_to_pipe_format(st, format);
+      pt.nr_samples = numSamples;
 
       st_gl_texture_dims_to_pipe_dims(target,
                                       width, height, depth,
                                       &pt.width0, &pt.height0,
                                       &pt.depth0, &pt.array_size);
 
-      if (level == 0 && (texObj->Sampler.MinFilter == GL_LINEAR ||
-                         texObj->Sampler.MinFilter == GL_NEAREST)) {
+      if (numLevels > 0) {
+         /* For immutable textures we know the final number of mip levels */
+         pt.last_level = numLevels - 1;
+      }
+      else if (level == 0 && (texObj->Sampler.MinFilter == GL_LINEAR ||
+                              texObj->Sampler.MinFilter == GL_NEAREST)) {
          /* assume just one mipmap level */
          pt.last_level = 0;
       }
@@ -2935,8 +2767,8 @@ st_TestProxyTexImage(struct gl_context *ctx, GLenum target,
    }
    else {
       /* Use core Mesa fallback */
-      return _mesa_test_proxy_teximage(ctx, target, level, format,
-                                       width, height, depth, border);
+      return _mesa_test_proxy_teximage(ctx, target, numLevels, level, format,
+                                       numSamples, width, height, depth);
    }
 }
 
@@ -2970,9 +2802,6 @@ st_TextureView(struct gl_context *ctx,
    tex->surface_format =
       st_mesa_format_to_pipe_format(st_context(ctx), image->TexFormat);
 
-   tex->width0 = image->Width;
-   tex->height0 = image->Height;
-   tex->depth0 = image->Depth;
    tex->lastLevel = numLevels - 1;
 
    return GL_TRUE;
@@ -2996,6 +2825,9 @@ st_ClearTexSubImage(struct gl_context *ctx,
    if (!pt)
       return;
 
+   st_flush_bitmap_cache(st);
+   st_invalidate_readpix_cache(st);
+
    u_box_3d(xoffset, yoffset, zoffset + texImage->Face,
             width, height, depth, &box);
    if (texImage->TexObject->Immutable) {