i965: Drop a bunch of downcasting and upcasting of gl_program pointers.
[mesa.git] / src / mesa / drivers / dri / i965 / intel_pixel_read.c
index 7074d7db9a24548d92c0d338c04431b6a1d1d524..4528d6d265a5dcd3fb1d7d06a8b51d3e3177cc34 100644 (file)
 #include "main/readpix.h"
 #include "main/state.h"
 #include "main/glformats.h"
+#include "program/prog_instruction.h"
 #include "drivers/common/meta.h"
 
 #include "brw_context.h"
+#include "brw_blorp.h"
 #include "intel_screen.h"
 #include "intel_batchbuffer.h"
 #include "intel_blit.h"
@@ -73,6 +75,7 @@ intel_readpixels_tiled_memcpy(struct gl_context * ctx,
 {
    struct brw_context *brw = brw_context(ctx);
    struct gl_renderbuffer *rb = ctx->ReadBuffer->_ColorReadBuffer;
+   const struct gen_device_info *devinfo = &brw->screen->devinfo;
 
    /* This path supports reading from color buffers only */
    if (rb == NULL)
@@ -91,7 +94,7 @@ intel_readpixels_tiled_memcpy(struct gl_context * ctx,
     * a 2D BGRA, RGBA, L8 or A8 texture. It could be generalized to support
     * more types.
     */
-   if (!brw->has_llc ||
+   if (!devinfo->has_llc ||
        !(type == GL_UNSIGNED_BYTE || type == GL_UNSIGNED_INT_8_8_8_8_REV) ||
        pixels == NULL ||
        _mesa_is_bufferobj(pack->BufferObj) ||
@@ -127,8 +130,8 @@ intel_readpixels_tiled_memcpy(struct gl_context * ctx,
       return false;
 
    if (!irb->mt ||
-       (irb->mt->tiling != I915_TILING_X &&
-       irb->mt->tiling != I915_TILING_Y)) {
+       (irb->mt->surf.tiling != ISL_TILING_X &&
+        irb->mt->surf.tiling != ISL_TILING_Y0)) {
       /* The algorithm is written only for X- or Y-tiled memory. */
       return false;
    }
@@ -141,13 +144,13 @@ intel_readpixels_tiled_memcpy(struct gl_context * ctx,
     * parts of the memory aren't swizzled at all. Userspace just can't handle
     * that.
     */
-   if (brw->gen < 5 && brw->has_swizzling)
+   if (devinfo->gen < 5 && brw->has_swizzling)
       return false;
 
    /* Since we are going to read raw data to the miptree, we need to resolve
     * any pending fast color clears before we start.
     */
-   intel_miptree_all_slices_resolve_color(brw, irb->mt, 0);
+   intel_miptree_access_raw(brw, irb->mt, irb->mt_level, irb->mt_layer, false);
 
    bo = irb->mt->bo;
 
@@ -162,8 +165,11 @@ intel_readpixels_tiled_memcpy(struct gl_context * ctx,
       return false;
    }
 
-   xoffset += irb->mt->level[irb->mt_level].slice[irb->mt_layer].x_offset;
-   yoffset += irb->mt->level[irb->mt_level].slice[irb->mt_layer].y_offset;
+   unsigned slice_offset_x, slice_offset_y;
+   intel_miptree_get_image_offset(irb->mt, irb->mt_level, irb->mt_layer,
+                                  &slice_offset_x, &slice_offset_y);
+   xoffset += slice_offset_x;
+   yoffset += slice_offset_y;
 
    dst_pitch = _mesa_image_row_stride(pack, width, format, type);
 
@@ -189,7 +195,7 @@ intel_readpixels_tiled_memcpy(struct gl_context * ctx,
        "mesa_format=0x%x tiling=%d "
        "pack=(alignment=%d row_length=%d skip_pixels=%d skip_rows=%d)\n",
        __func__, xoffset, yoffset, width, height,
-       format, type, rb->Format, irb->mt->tiling,
+       format, type, rb->Format, irb->mt->surf.tiling,
        pack->Alignment, pack->RowLength, pack->SkipPixels,
        pack->SkipRows);
 
@@ -198,9 +204,9 @@ intel_readpixels_tiled_memcpy(struct gl_context * ctx,
       yoffset, yoffset + height,
       pixels - (ptrdiff_t) yoffset * dst_pitch - (ptrdiff_t) xoffset * cpp,
       map + irb->mt->offset,
-      dst_pitch, irb->mt->pitch,
+      dst_pitch, irb->mt->surf.row_pitch,
       brw->has_swizzling,
-      irb->mt->tiling,
+      irb->mt->surf.tiling,
       mem_copy
    );
 
@@ -208,6 +214,45 @@ intel_readpixels_tiled_memcpy(struct gl_context * ctx,
    return true;
 }
 
+static bool
+intel_readpixels_blorp(struct gl_context *ctx,
+                       unsigned x, unsigned y,
+                       unsigned w, unsigned h,
+                       GLenum format, GLenum type, const void *pixels,
+                       const struct gl_pixelstore_attrib *packing)
+{
+   struct brw_context *brw = brw_context(ctx);
+   struct gl_renderbuffer *rb = ctx->ReadBuffer->_ColorReadBuffer;
+   if (!rb)
+      return false;
+
+   struct intel_renderbuffer *irb = intel_renderbuffer(rb);
+
+   /* _mesa_get_readpixels_transfer_ops() includes the cases of read
+    * color clamping along with the ctx->_ImageTransferState.
+    */
+   if (_mesa_get_readpixels_transfer_ops(ctx, rb->Format, format,
+                                         type, GL_FALSE))
+      return false;
+
+   GLenum dst_base_format = _mesa_unpack_format_to_base_format(format);
+   if (_mesa_need_rgb_to_luminance_conversion(rb->_BaseFormat,
+                                              dst_base_format))
+      return false;
+
+   unsigned swizzle;
+   if (irb->Base.Base._BaseFormat == GL_RGB) {
+      swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_ONE);
+   } else {
+      swizzle = SWIZZLE_XYZW;
+   }
+
+   return brw_blorp_download_miptree(brw, irb->mt, rb->Format, swizzle,
+                                     irb->mt_level, x, y, irb->mt_layer,
+                                     w, h, 1, GL_TEXTURE_2D, format, type,
+                                     rb->Name == 0, pixels, packing);
+}
+
 void
 intelReadPixels(struct gl_context * ctx,
                 GLint x, GLint y, GLsizei width, GLsizei height,
@@ -221,42 +266,21 @@ intelReadPixels(struct gl_context * ctx,
 
    DBG("%s\n", __func__);
 
+   /* Reading pixels wont dirty the front buffer, so reset the dirty
+    * flag after calling intel_prepare_render().
+    */
+   dirty = brw->front_buffer_dirty;
+   intel_prepare_render(brw);
+   brw->front_buffer_dirty = dirty;
+
    if (_mesa_is_bufferobj(pack->BufferObj)) {
-      if (_mesa_meta_pbo_GetTexSubImage(ctx, 2, NULL, x, y, 0, width, height, 1,
-                                        format, type, pixels, pack)) {
-         /* _mesa_meta_pbo_GetTexSubImage() implements PBO transfers by
-          * binding the user-provided BO as a fake framebuffer and rendering
-          * to it.  This breaks the invariant of the GL that nothing is able
-          * to render to a BO, causing nondeterministic corruption issues
-          * because the render cache is not coherent with a number of other
-          * caches that the BO could potentially be bound to afterwards.
-          *
-          * This could be solved in the same way that we guarantee texture
-          * coherency after a texture is attached to a framebuffer and
-          * rendered to, but that would involve checking *all* BOs bound to
-          * the pipeline for the case we need to emit a cache flush due to
-          * previous rendering to any of them -- Including vertex, index,
-          * uniform, atomic counter, shader image, transform feedback,
-          * indirect draw buffers, etc.
-          *
-          * That would increase the per-draw call overhead even though it's
-          * very unlikely that any of the BOs bound to the pipeline has been
-          * rendered to via a PBO at any point, so it seems better to just
-          * flush here unconditionally.
-          */
-         brw_emit_mi_flush(brw);
+      if (intel_readpixels_blorp(ctx, x, y, width, height,
+                                 format, type, pixels, pack))
          return;
-      }
 
       perf_debug("%s: fallback to CPU mapping in PBO case\n", __func__);
    }
 
-   /* Reading pixels wont dirty the front buffer, so reset the dirty
-    * flag after calling intel_prepare_render(). */
-   dirty = brw->front_buffer_dirty;
-   intel_prepare_render(brw);
-   brw->front_buffer_dirty = dirty;
-
    ok = intel_readpixels_tiled_memcpy(ctx, x, y, width, height,
                                       format, type, pixels, pack);
    if(ok)