Home | History | Annotate | Download | only in i965 
      1 /*
      2  * Copyright 2003 VMware, Inc.
      3  * All Rights Reserved.
      4  *
      5  * Permission is hereby granted, free of charge, to any person obtaining a
      6  * copy of this software and associated documentation files (the
      7  * "Software"), to deal in the Software without restriction, including
      8  * without limitation the rights to use, copy, modify, merge, publish,
      9  * distribute, sublicense, and/or sell copies of the Software, and to
     10  * permit persons to whom the Software is furnished to do so, subject to
     11  * the following conditions:
     12  *
     13  * The above copyright notice and this permission notice (including the
     14  * next paragraph) shall be included in all copies or substantial portions
     15  * of the Software.
     16  *
     17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
     18  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
     19  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
     20  * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
     21  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
     22  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
     23  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
     24  */
     25 
     26 #include "main/enums.h"
     27 #include "main/mtypes.h"
     28 #include "main/macros.h"
     29 #include "main/fbobject.h"
     30 #include "main/image.h"
     31 #include "main/bufferobj.h"
     32 #include "main/readpix.h"
     33 #include "main/state.h"
     34 #include "main/glformats.h"
     35 #include "drivers/common/meta.h"
     36 
     37 #include "brw_context.h"
     38 #include "intel_screen.h"
     39 #include "intel_batchbuffer.h"
     40 #include "intel_blit.h"
     41 #include "intel_buffers.h"
     42 #include "intel_fbo.h"
     43 #include "intel_mipmap_tree.h"
     44 #include "intel_pixel.h"
     45 #include "intel_buffer_objects.h"
     46 #include "intel_tiled_memcpy.h"
     47 
     48 #define FILE_DEBUG_FLAG DEBUG_PIXEL
     49 
     50 /**
     51  * \brief A fast path for glReadPixels
     52  *
     53  * This fast path is taken when the source format is BGRA, RGBA,
     54  * A or L and when the texture memory is X- or Y-tiled.  It downloads
     55  * the source data by directly mapping the memory without a GTT fence.
     56  * This then needs to be de-tiled on the CPU before presenting the data to
     57  * the user in the linear fasion.
     58  *
     59  * This is a performance win over the conventional texture download path.
     60  * In the conventional texture download path, the texture is either mapped
     61  * through the GTT or copied to a linear buffer with the blitter before
     62  * handing off to a software path.  This allows us to avoid round-tripping
     63  * through the GPU (in the case where we would be blitting) and do only a
     64  * single copy operation.
     65  */
     66 static bool
     67 intel_readpixels_tiled_memcpy(struct gl_context * ctx,
     68                               GLint xoffset, GLint yoffset,
     69                               GLsizei width, GLsizei height,
     70                               GLenum format, GLenum type,
     71                               GLvoid * pixels,
     72                               const struct gl_pixelstore_attrib *pack)
     73 {
     74    struct brw_context *brw = brw_context(ctx);
     75    struct gl_renderbuffer *rb = ctx->ReadBuffer->_ColorReadBuffer;
     76 
     77    /* This path supports reading from color buffers only */
     78    if (rb == NULL)
     79       return false;
     80 
     81    struct intel_renderbuffer *irb = intel_renderbuffer(rb);
     82    int dst_pitch;
     83 
     84    /* The miptree's buffer. */
     85    drm_intel_bo *bo;
     86 
     87    int error = 0;
     88 
     89    uint32_t cpp;
     90    mem_copy_fn mem_copy = NULL;
     91 
     92    /* This fastpath is restricted to specific renderbuffer types:
     93     * a 2D BGRA, RGBA, L8 or A8 texture. It could be generalized to support
     94     * more types.
     95     */
     96    if (!brw->has_llc ||
     97        !(type == GL_UNSIGNED_BYTE || type == GL_UNSIGNED_INT_8_8_8_8_REV) ||
     98        pixels == NULL ||
     99        _mesa_is_bufferobj(pack->BufferObj) ||
    100        pack->Alignment > 4 ||
    101        pack->SkipPixels > 0 ||
    102        pack->SkipRows > 0 ||
    103        (pack->RowLength != 0 && pack->RowLength != width) ||
    104        pack->SwapBytes ||
    105        pack->LsbFirst ||
    106        pack->Invert)
    107       return false;
    108 
    109    /* Only a simple blit, no scale, bias or other mapping. */
    110    if (ctx->_ImageTransferState)
    111       return false;
    112 
    113    /* It is possible that the renderbuffer (or underlying texture) is
    114     * multisampled.  Since ReadPixels from a multisampled buffer requires a
    115     * multisample resolve, we can't handle this here
    116     */
    117    if (rb->NumSamples > 1)
    118       return false;
    119 
    120    /* We can't handle copying from RGBX or BGRX because the tiled_memcpy
    121     * function doesn't set the last channel to 1. Note this checks BaseFormat
    122     * rather than TexFormat in case the RGBX format is being simulated with an
    123     * RGBA format.
    124     */
    125    if (rb->_BaseFormat == GL_RGB)
    126       return false;
    127 
    128    if (!intel_get_memcpy(rb->Format, format, type, &mem_copy, &cpp))
    129       return false;
    130 
    131    if (!irb->mt ||
    132        (irb->mt->tiling != I915_TILING_X &&
    133        irb->mt->tiling != I915_TILING_Y)) {
    134       /* The algorithm is written only for X- or Y-tiled memory. */
    135       return false;
    136    }
    137 
    138    /* Since we are going to read raw data to the miptree, we need to resolve
    139     * any pending fast color clears before we start.
    140     */
    141    intel_miptree_all_slices_resolve_color(brw, irb->mt, 0);
    142 
    143    bo = irb->mt->bo;
    144 
    145    if (drm_intel_bo_references(brw->batch.bo, bo)) {
    146       perf_debug("Flushing before mapping a referenced bo.\n");
    147       intel_batchbuffer_flush(brw);
    148    }
    149 
    150    error = brw_bo_map(brw, bo, false /* write enable */, "miptree");
    151    if (error) {
    152       DBG("%s: failed to map bo\n", __func__);
    153       return false;
    154    }
    155 
    156    xoffset += irb->mt->level[irb->mt_level].slice[irb->mt_layer].x_offset;
    157    yoffset += irb->mt->level[irb->mt_level].slice[irb->mt_layer].y_offset;
    158 
    159    dst_pitch = _mesa_image_row_stride(pack, width, format, type);
    160 
    161    /* For a window-system renderbuffer, the buffer is actually flipped
    162     * vertically, so we need to handle that.  Since the detiling function
    163     * can only really work in the forwards direction, we have to be a
    164     * little creative.  First, we compute the Y-offset of the first row of
    165     * the renderbuffer (in renderbuffer coordinates).  We then match that
    166     * with the last row of the client's data.  Finally, we give
    167     * tiled_to_linear a negative pitch so that it walks through the
    168     * client's data backwards as it walks through the renderbufer forwards.
    169     */
    170    if (rb->Name == 0) {
    171       yoffset = rb->Height - yoffset - height;
    172       pixels += (ptrdiff_t) (height - 1) * dst_pitch;
    173       dst_pitch = -dst_pitch;
    174    }
    175 
    176    /* We postponed printing this message until having committed to executing
    177     * the function.
    178     */
    179    DBG("%s: x,y=(%d,%d) (w,h)=(%d,%d) format=0x%x type=0x%x "
    180        "mesa_format=0x%x tiling=%d "
    181        "pack=(alignment=%d row_length=%d skip_pixels=%d skip_rows=%d)\n",
    182        __func__, xoffset, yoffset, width, height,
    183        format, type, rb->Format, irb->mt->tiling,
    184        pack->Alignment, pack->RowLength, pack->SkipPixels,
    185        pack->SkipRows);
    186 
    187    tiled_to_linear(
    188       xoffset * cpp, (xoffset + width) * cpp,
    189       yoffset, yoffset + height,
    190       pixels - (ptrdiff_t) yoffset * dst_pitch - (ptrdiff_t) xoffset * cpp,
    191       bo->virtual + irb->mt->offset,
    192       dst_pitch, irb->mt->pitch,
    193       brw->has_swizzling,
    194       irb->mt->tiling,
    195       mem_copy
    196    );
    197 
    198    drm_intel_bo_unmap(bo);
    199    return true;
    200 }
    201 
    202 void
    203 intelReadPixels(struct gl_context * ctx,
    204                 GLint x, GLint y, GLsizei width, GLsizei height,
    205                 GLenum format, GLenum type,
    206                 const struct gl_pixelstore_attrib *pack, GLvoid * pixels)
    207 {
    208    bool ok;
    209 
    210    struct brw_context *brw = brw_context(ctx);
    211    bool dirty;
    212 
    213    DBG("%s\n", __func__);
    214 
    215    if (_mesa_is_bufferobj(pack->BufferObj)) {
    216       if (_mesa_meta_pbo_GetTexSubImage(ctx, 2, NULL, x, y, 0, width, height, 1,
    217                                         format, type, pixels, pack)) {
    218          /* _mesa_meta_pbo_GetTexSubImage() implements PBO transfers by
    219           * binding the user-provided BO as a fake framebuffer and rendering
    220           * to it.  This breaks the invariant of the GL that nothing is able
    221           * to render to a BO, causing nondeterministic corruption issues
    222           * because the render cache is not coherent with a number of other
    223           * caches that the BO could potentially be bound to afterwards.
    224           *
    225           * This could be solved in the same way that we guarantee texture
    226           * coherency after a texture is attached to a framebuffer and
    227           * rendered to, but that would involve checking *all* BOs bound to
    228           * the pipeline for the case we need to emit a cache flush due to
    229           * previous rendering to any of them -- Including vertex, index,
    230           * uniform, atomic counter, shader image, transform feedback,
    231           * indirect draw buffers, etc.
    232           *
    233           * That would increase the per-draw call overhead even though it's
    234           * very unlikely that any of the BOs bound to the pipeline has been
    235           * rendered to via a PBO at any point, so it seems better to just
    236           * flush here unconditionally.
    237           */
    238          brw_emit_mi_flush(brw);
    239          return;
    240       }
    241 
    242       perf_debug("%s: fallback to CPU mapping in PBO case\n", __func__);
    243    }
    244 
    245    /* Reading pixels wont dirty the front buffer, so reset the dirty
    246     * flag after calling intel_prepare_render(). */
    247    dirty = brw->front_buffer_dirty;
    248    intel_prepare_render(brw);
    249    brw->front_buffer_dirty = dirty;
    250 
    251    ok = intel_readpixels_tiled_memcpy(ctx, x, y, width, height,
    252                                       format, type, pixels, pack);
    253    if(ok)
    254       return;
    255 
    256    /* Update Mesa state before calling _mesa_readpixels().
    257     * XXX this may not be needed since ReadPixels no longer uses the
    258     * span code.
    259     */
    260 
    261    if (ctx->NewState)
    262       _mesa_update_state(ctx);
    263 
    264    _mesa_readpixels(ctx, x, y, width, height, format, type, pack, pixels);
    265 
    266    /* There's an intel_prepare_render() call in intelSpanRenderStart(). */
    267    brw->front_buffer_dirty = dirty;
    268 }
    269