src/mesa/drivers/dri/i965/intel_tex_subimage.c

   1
   2 /**************************************************************************
   3  *
   4  * Copyright 2003 Tungsten Graphics, Inc., Cedar Park, Texas.
   5  * All Rights Reserved.
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a
   8  * copy of this software and associated documentation files (the
   9  * "Software"), to deal in the Software without restriction, including
  10  * without limitation the rights to use, copy, modify, merge, publish,
  11  * distribute, sub license, and/or sell copies of the Software, and to
  12  * permit persons to whom the Software is furnished to do so, subject to
  13  * the following conditions:
  14  *
  15  * The above copyright notice and this permission notice (including the
  16  * next paragraph) shall be included in all copies or substantial portions
  17  * of the Software.
  18  *
  19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  20  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  21  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  22  * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
  23  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  24  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  25  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  26  *
  27  **************************************************************************/
  28
  29 #include "main/bufferobj.h"
  30 #include "main/macros.h"
  31 #include "main/mtypes.h"
  32 #include "main/pbo.h"
  33 #include "main/texobj.h"
  34 #include "main/texstore.h"
  35 #include "main/texcompress.h"
  36 #include "main/enums.h"
  37
  38 #include "intel_batchbuffer.h"
  39 #include "intel_context.h"
  40 #include "intel_tex.h"
  41 #include "intel_mipmap_tree.h"
  42 #include "intel_blit.h"
  43
  44 #define FILE_DEBUG_FLAG DEBUG_TEXTURE
  45
  46 static bool
  47 intel_blit_texsubimage(struct gl_context * ctx,
  48                        struct gl_texture_image *texImage,
  49                        GLint xoffset, GLint yoffset,
  50                        GLint width, GLint height,
  51                        GLenum format, GLenum type, const void *pixels,
  52                        const struct gl_pixelstore_attrib *packing)
  53 {
  54    struct intel_context *intel = intel_context(ctx);
  55    struct intel_texture_image *intelImage = intel_texture_image(texImage);
  56
  57    /* Try to do a blit upload of the subimage if the texture is
  58     * currently busy.
  59     */
  60    if (!intelImage->mt)
  61       return false;
  62
  63    /* The blitter can't handle Y tiling */
  64    if (intelImage->mt->region->tiling == I915_TILING_Y)
  65       return false;
  66
  67    if (texImage->TexObject->Target != GL_TEXTURE_2D)
  68       return false;
  69
  70    /* On gen6, it's probably not worth swapping to the blit ring to do
  71     * this because of all the overhead involved.
  72     */
  73    if (intel->gen >= 6)
  74       return false;
  75
  76    if (!drm_intel_bo_busy(intelImage->mt->region->bo))
  77       return false;
  78
  79    DBG("BLT subimage %s target %s level %d offset %d,%d %dx%d\n",
  80        __FUNCTION__,
  81        _mesa_lookup_enum_by_nr(texImage->TexObject->Target),
  82        texImage->Level, xoffset, yoffset, width, height);
  83
  84    pixels = _mesa_validate_pbo_teximage(ctx, 2, width, height, 1,
  85                                         format, type, pixels, packing,
  86                                         "glTexSubImage");
  87    if (!pixels)
  88       return false;
  89
  90    struct intel_mipmap_tree *temp_mt =
  91       intel_miptree_create(intel, GL_TEXTURE_2D, texImage->TexFormat,
  92                            0, 0,
  93                            width, height, 1,
  94                            false, 0, INTEL_MIPTREE_TILING_NONE);
  95    if (!temp_mt)
  96       goto err;
  97
  98    GLubyte *dst = intel_miptree_map_raw(intel, temp_mt);
  99    if (!dst)
 100       goto err;
 101
 102    if (!_mesa_texstore(ctx, 2, texImage->_BaseFormat,
 103                        texImage->TexFormat,
 104                        temp_mt->region->pitch,
 105                        &dst,
 106                        width, height, 1,
 107                        format, type, pixels, packing)) {
 108       _mesa_error(ctx, GL_OUT_OF_MEMORY, "intelTexSubImage");
 109    }
 110
 111    intel_miptree_unmap_raw(intel, temp_mt);
 112
 113    bool ret;
 114
 115    ret = intel_miptree_blit(intel,
 116                             temp_mt, 0, 0,
 117                             0, 0, false,
 118                             intelImage->mt, texImage->Level, texImage->Face,
 119                             xoffset, yoffset, false,
 120                             width, height, GL_COPY);
 121    assert(ret);
 122
 123    intel_miptree_release(&temp_mt);
 124    _mesa_unmap_teximage_pbo(ctx, packing);
 125
 126    return ret;
 127
 128 err:
 129    _mesa_error(ctx, GL_OUT_OF_MEMORY, "intelTexSubImage");
 130    intel_miptree_release(&temp_mt);
 131    _mesa_unmap_teximage_pbo(ctx, packing);
 132    return false;
 133 }
 134
 135 /**
 136  * \brief A fast path for glTexImage and glTexSubImage.
 137  *
 138  * \param for_glTexImage Was this called from glTexImage or glTexSubImage?
 139  *
 140  * This fast path is taken when the hardware natively supports the texture
 141  * format (such as GL_BGRA) and when the texture memory is X-tiled. It uploads
 142  * the texture data by mapping the texture memory without a GTT fence, thus
 143  * acquiring a tiled view of the memory, and then memcpy'ing sucessive
 144  * subspans within each tile.
 145  *
 146  * This is a performance win over the conventional texture upload path because
 147  * it avoids the performance penalty of writing through the write-combine
 148  * buffer. In the conventional texture upload path,
 149  * texstore.c:store_texsubimage(), the texture memory is mapped through a GTT
 150  * fence, thus acquiring a linear view of the memory, then each row in the
 151  * image is memcpy'd. In this fast path, we replace each row's memcpy with
 152  * a sequence of memcpy's over each bit6 swizzle span in the row.
 153  *
 154  * This fast path's use case is Google Chrome's paint rectangles.  Chrome (as
 155  * of version 21) renders each page as a tiling of 256x256 GL_BGRA textures.
 156  * Each page's content is initially uploaded with glTexImage2D and damaged
 157  * regions are updated with glTexSubImage2D. On some workloads, the
 158  * performance gain of this fastpath on Sandybridge is over 5x.
 159  */
 160 bool
 161 intel_texsubimage_tiled_memcpy(struct gl_context * ctx,
 162                                GLuint dims,
 163                                struct gl_texture_image *texImage,
 164                                GLint xoffset, GLint yoffset, GLint zoffset,
 165                                GLsizei width, GLsizei height, GLsizei depth,
 166                                GLenum format, GLenum type,
 167                                const GLvoid *pixels,
 168                                const struct gl_pixelstore_attrib *packing,
 169                                bool for_glTexImage)
 170 {
 171    struct intel_context *intel = intel_context(ctx);
 172    struct intel_texture_image *image = intel_texture_image(texImage);
 173
 174    /* The miptree's buffer. */
 175    drm_intel_bo *bo;
 176
 177    int error = 0;
 178
 179    /* This fastpath is restricted to a specific texture type: level 0 of
 180     * a 2D BGRA texture. It could be generalized to support more types by
 181     * varying the arithmetic loop below.
 182     */
 183    if (!intel->has_llc ||
 184        format != GL_BGRA ||
 185        type != GL_UNSIGNED_BYTE ||
 186        texImage->TexFormat != MESA_FORMAT_ARGB8888 ||
 187        texImage->TexObject->Target != GL_TEXTURE_2D ||
 188        texImage->Level != 0 ||
 189        pixels == NULL ||
 190        _mesa_is_bufferobj(packing->BufferObj) ||
 191        packing->Alignment > 4 ||
 192        packing->SkipPixels > 0 ||
 193        packing->SkipRows > 0 ||
 194        (packing->RowLength != 0 && packing->RowLength != width) ||
 195        packing->SwapBytes ||
 196        packing->LsbFirst ||
 197        packing->Invert)
 198       return false;
 199
 200    if (for_glTexImage)
 201       ctx->Driver.AllocTextureImageBuffer(ctx, texImage);
 202
 203    if (!image->mt ||
 204        image->mt->region->tiling != I915_TILING_X) {
 205       /* The algorithm below is written only for X-tiled memory. */
 206       return false;
 207    }
 208
 209    /* Since we are going to write raw data to the miptree, we need to resolve
 210     * any pending fast color clears before we start.
 211     */
 212    intel_miptree_resolve_color(intel, image->mt);
 213
 214    bo = image->mt->region->bo;
 215
 216    if (drm_intel_bo_references(intel->batch.bo, bo)) {
 217       perf_debug("Flushing before mapping a referenced bo.\n");
 218       intel_batchbuffer_flush(intel);
 219    }
 220
 221    if (unlikely(intel->perf_debug)) {
 222       if (drm_intel_bo_busy(bo)) {
 223          perf_debug("Mapping a busy BO, causing a stall on the GPU.\n");
 224       }
 225    }
 226
 227    error = drm_intel_bo_map(bo, true /*write_enable*/);
 228    if (error || bo->virtual == NULL) {
 229       DBG("%s: failed to map bo\n", __FUNCTION__);
 230       return false;
 231    }
 232
 233    /* We postponed printing this message until having committed to executing
 234     * the function.
 235     */
 236    DBG("%s: level=%d offset=(%d,%d) (w,h)=(%d,%d)\n",
 237        __FUNCTION__, texImage->Level, xoffset, yoffset, width, height);
 238
 239    /* In the tiling algorithm below, some variables are in units of pixels,
 240     * others are in units of bytes, and others (such as height) are unitless.
 241     * Each variable name is suffixed with its units.
 242     */
 243
 244    const uint32_t x_max_pixels = xoffset + width;
 245    const uint32_t y_max_pixels = yoffset + height;
 246
 247    const uint32_t tile_size_bytes = 4096;
 248
 249    const uint32_t tile_width_bytes = 512;
 250    const uint32_t tile_width_pixels = 128;
 251
 252    const uint32_t tile_height = 8;
 253
 254    const uint32_t cpp = 4; /* chars per pixel of GL_BGRA */
 255    const uint32_t swizzle_width_pixels = 16;
 256
 257    const uint32_t stride_bytes = image->mt->region->pitch;
 258    const uint32_t width_tiles = stride_bytes / tile_width_bytes;
 259
 260    for (uint32_t y_pixels = yoffset; y_pixels < y_max_pixels; ++y_pixels) {
 261       const uint32_t y_offset_bytes = (y_pixels / tile_height) * width_tiles * tile_size_bytes
 262                                     + (y_pixels % tile_height) * tile_width_bytes;
 263
 264       for (uint32_t x_pixels = xoffset; x_pixels < x_max_pixels; x_pixels += swizzle_width_pixels) {
 265          const uint32_t x_offset_bytes = (x_pixels / tile_width_pixels) * tile_size_bytes
 266                                        + (x_pixels % tile_width_pixels) * cpp;
 267
 268          intptr_t offset_bytes = y_offset_bytes + x_offset_bytes;
 269          if (intel->has_swizzling) {
 270 #if 0
 271             /* Clear, unoptimized version. */
 272             bool bit6 = (offset_bytes >> 6) & 1;
 273             bool bit9 = (offset_bytes >> 9) & 1;
 274             bool bit10 = (offset_bytes >> 10) & 1;
 275
 276             if (bit9 ^ bit10)
 277                offset_bytes ^= (1 << 6);
 278 #else
 279             /* Optimized, obfuscated version. */
 280             offset_bytes ^= ((offset_bytes >> 3) ^ (offset_bytes >> 4))
 281                           & (1 << 6);
 282 #endif
 283          }
 284
 285          const uint32_t swizzle_bound_pixels = ALIGN(x_pixels + 1, swizzle_width_pixels);
 286          const uint32_t memcpy_bound_pixels = MIN2(x_max_pixels, swizzle_bound_pixels);
 287          const uint32_t copy_size = cpp * (memcpy_bound_pixels - x_pixels);
 288
 289          memcpy(bo->virtual + offset_bytes, pixels, copy_size);
 290          pixels += copy_size;
 291          x_pixels -= (x_pixels % swizzle_width_pixels);
 292       }
 293    }
 294
 295    drm_intel_bo_unmap(bo);
 296    return true;
 297 }
 298
 299 static void
 300 intelTexSubImage(struct gl_context * ctx,
 301                  GLuint dims,
 302                  struct gl_texture_image *texImage,
 303                  GLint xoffset, GLint yoffset, GLint zoffset,
 304                  GLsizei width, GLsizei height, GLsizei depth,
 305                  GLenum format, GLenum type,
 306                  const GLvoid * pixels,
 307                  const struct gl_pixelstore_attrib *packing)
 308 {
 309    bool ok;
 310
 311    ok = intel_texsubimage_tiled_memcpy(ctx, dims, texImage,
 312                                        xoffset, yoffset, zoffset,
 313                                        width, height, depth,
 314                                        format, type, pixels, packing,
 315                                        false /*for_glTexImage*/);
 316    if (ok)
 317      return;
 318
 319    /* The intel_blit_texsubimage() function only handles 2D images */
 320    if (dims != 2 || !intel_blit_texsubimage(ctx, texImage,
 321                                xoffset, yoffset,
 322                                width, height,
 323                                format, type, pixels, packing)) {
 324       _mesa_store_texsubimage(ctx, dims, texImage,
 325                               xoffset, yoffset, zoffset,
 326                               width, height, depth,
 327                               format, type, pixels, packing);
 328    }
 329 }
 330
 331 void
 332 intelInitTextureSubImageFuncs(struct dd_function_table *functions)
 333 {
 334    functions->TexSubImage = intelTexSubImage;
 335 }