2 /**************************************************************************
4 * Copyright 2003 Tungsten Graphics, Inc., Cedar Park, Texas.
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:
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
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.
27 **************************************************************************/
29 #include "main/macros.h"
30 #include "main/mtypes.h"
32 #include "main/texobj.h"
33 #include "main/texstore.h"
34 #include "main/texcompress.h"
35 #include "main/enums.h"
37 #include "intel_batchbuffer.h"
38 #include "intel_context.h"
39 #include "intel_tex.h"
40 #include "intel_mipmap_tree.h"
41 #include "intel_blit.h"
43 #define FILE_DEBUG_FLAG DEBUG_TEXTURE
46 intel_blit_texsubimage(struct gl_context
* ctx
,
47 struct gl_texture_image
*texImage
,
48 GLint xoffset
, GLint yoffset
,
49 GLint width
, GLint height
,
50 GLenum format
, GLenum type
, const void *pixels
,
51 const struct gl_pixelstore_attrib
*packing
)
53 struct intel_context
*intel
= intel_context(ctx
);
54 struct intel_texture_image
*intelImage
= intel_texture_image(texImage
);
55 GLuint dstRowStride
= 0;
56 drm_intel_bo
*temp_bo
= NULL
;
57 unsigned int blit_x
= 0, blit_y
= 0;
59 uint32_t tiling_mode
= I915_TILING_NONE
;
62 /* Try to do a blit upload of the subimage if the texture is
68 /* The blitter can't handle Y tiling */
69 if (intelImage
->mt
->region
->tiling
== I915_TILING_Y
)
72 if (texImage
->TexObject
->Target
!= GL_TEXTURE_2D
)
75 /* On gen6, it's probably not worth swapping to the blit ring to do
76 * this because of all the overhead involved.
81 if (!drm_intel_bo_busy(intelImage
->mt
->region
->bo
))
84 DBG("BLT subimage %s target %s level %d offset %d,%d %dx%d\n",
86 _mesa_lookup_enum_by_nr(texImage
->TexObject
->Target
),
87 texImage
->Level
, xoffset
, yoffset
, width
, height
);
89 pixels
= _mesa_validate_pbo_teximage(ctx
, 2, width
, height
, 1,
90 format
, type
, pixels
, packing
,
95 temp_bo
= drm_intel_bo_alloc_tiled(intel
->bufmgr
,
102 if (temp_bo
== NULL
) {
103 _mesa_error(ctx
, GL_OUT_OF_MEMORY
, "intelTexSubImage");
107 if (drm_intel_gem_bo_map_gtt(temp_bo
)) {
108 _mesa_error(ctx
, GL_OUT_OF_MEMORY
, "intelTexSubImage");
112 dstMap
= temp_bo
->virtual;
113 dstRowStride
= pitch
;
115 intel_miptree_get_image_offset(intelImage
->mt
, texImage
->Level
,
116 intelImage
->base
.Base
.Face
, 0,
121 if (!_mesa_texstore(ctx
, 2, texImage
->_BaseFormat
,
126 format
, type
, pixels
, packing
)) {
127 _mesa_error(ctx
, GL_OUT_OF_MEMORY
, "intelTexSubImage");
131 unsigned int dst_pitch
= intelImage
->mt
->region
->pitch
*
134 drm_intel_gem_bo_unmap_gtt(temp_bo
);
136 ret
= intelEmitCopyBlit(intel
,
138 dstRowStride
/ intelImage
->mt
->cpp
,
140 dst_pitch
/ intelImage
->mt
->cpp
,
141 intelImage
->mt
->region
->bo
, 0,
142 intelImage
->mt
->region
->tiling
,
143 0, 0, blit_x
, blit_y
, width
, height
,
147 drm_intel_bo_unreference(temp_bo
);
148 _mesa_unmap_teximage_pbo(ctx
, packing
);
154 * \brief A fast path for glTexImage and glTexSubImage.
156 * \param for_glTexImage Was this called from glTexImage or glTexSubImage?
158 * This fast path is taken when the hardware natively supports the texture
159 * format (such as GL_BGRA) and when the texture memory is X-tiled. It uploads
160 * the texture data by mapping the texture memory without a GTT fence, thus
161 * acquiring a tiled view of the memory, and then memcpy'ing sucessive
162 * subspans within each tile.
164 * This is a performance win over the conventional texture upload path because
165 * it avoids the performance penalty of writing through the write-combine
166 * buffer. In the conventional texture upload path,
167 * texstore.c:store_texsubimage(), the texture memory is mapped through a GTT
168 * fence, thus acquiring a linear view of the memory, then each row in the
169 * image is memcpy'd. In this fast path, we replace each row's memcpy with
170 * a sequence of memcpy's over each bit6 swizzle span in the row.
172 * This fast path's use case is Google Chrome's paint rectangles. Chrome (as
173 * of version 21) renders each page as a tiling of 256x256 GL_BGRA textures.
174 * Each page's content is initially uploaded with glTexImage2D and damaged
175 * regions are updated with glTexSubImage2D. On some workloads, the
176 * performance gain of this fastpath on Sandybridge is over 5x.
179 intel_texsubimage_tiled_memcpy(struct gl_context
* ctx
,
181 struct gl_texture_image
*texImage
,
182 GLint xoffset
, GLint yoffset
, GLint zoffset
,
183 GLsizei width
, GLsizei height
, GLsizei depth
,
184 GLenum format
, GLenum type
,
185 const GLvoid
*pixels
,
186 const struct gl_pixelstore_attrib
*packing
,
189 struct intel_context
*intel
= intel_context(ctx
);
190 struct intel_texture_image
*image
= intel_texture_image(texImage
);
192 /* The miptree's buffer. */
197 /* This fastpath is restricted to a specific texture type: level 0 of
198 * a 2D BGRA texture. It could be generalized to support more types by
199 * varying the arithmetic loop below.
201 if (!intel
->has_llc
||
203 type
!= GL_UNSIGNED_BYTE
||
204 texImage
->TexFormat
!= MESA_FORMAT_ARGB8888
||
205 texImage
->TexObject
->Target
!= GL_TEXTURE_2D
||
206 texImage
->Level
!= 0 ||
208 packing
->Alignment
> 4)
212 ctx
->Driver
.AllocTextureImageBuffer(ctx
, texImage
);
215 image
->mt
->region
->tiling
!= I915_TILING_X
) {
216 /* The algorithm below is written only for X-tiled memory. */
220 bo
= image
->mt
->region
->bo
;
222 if (drm_intel_bo_references(intel
->batch
.bo
, bo
)) {
223 perf_debug("Flushing before mapping a referenced bo.\n");
224 intel_batchbuffer_flush(intel
);
227 if (unlikely(INTEL_DEBUG
& DEBUG_PERF
)) {
228 if (drm_intel_bo_busy(bo
)) {
229 perf_debug("Mapping a busy BO, causing a stall on the GPU.\n");
233 error
= drm_intel_bo_map(bo
, true /*write_enable*/);
234 if (error
|| bo
->virtual == NULL
) {
235 DBG("%s: failed to map bo\n", __FUNCTION__
);
239 /* We postponed printing this message until having committed to executing
242 DBG("%s: level=%d offset=(%d,%d) (w,h)=(%d,%d)\n",
243 __FUNCTION__
, texImage
->Level
, xoffset
, yoffset
, width
, height
);
245 /* In the tiling algorithm below, some variables are in units of pixels,
246 * others are in units of bytes, and others (such as height) are unitless.
247 * Each variable name is suffixed with its units.
250 const uint32_t x_max_pixels
= xoffset
+ width
;
251 const uint32_t y_max_pixels
= yoffset
+ height
;
253 const uint32_t tile_size_bytes
= 4096;
255 const uint32_t tile_width_bytes
= 512;
256 const uint32_t tile_width_pixels
= 128;
258 const uint32_t tile_height
= 8;
260 const uint32_t cpp
= 4; /* chars per pixel of GL_BGRA */
261 const uint32_t swizzle_width_pixels
= 16;
263 const uint32_t stride_bytes
= image
->mt
->region
->pitch
* cpp
;
264 const uint32_t width_tiles
= stride_bytes
/ tile_width_bytes
;
266 for (uint32_t y_pixels
= yoffset
; y_pixels
< y_max_pixels
; ++y_pixels
) {
267 const uint32_t y_offset_bytes
= (y_pixels
/ tile_height
) * width_tiles
* tile_size_bytes
268 + (y_pixels
% tile_height
) * tile_width_bytes
;
270 for (uint32_t x_pixels
= xoffset
; x_pixels
< x_max_pixels
; x_pixels
+= swizzle_width_pixels
) {
271 const uint32_t x_offset_bytes
= (x_pixels
/ tile_width_pixels
) * tile_size_bytes
272 + (x_pixels
% tile_width_pixels
) * cpp
;
274 intptr_t offset_bytes
= y_offset_bytes
+ x_offset_bytes
;
275 if (intel
->has_swizzling
) {
277 /* Clear, unoptimized version. */
278 bool bit6
= (offset_bytes
>> 6) & 1;
279 bool bit9
= (offset_bytes
>> 9) & 1;
280 bool bit10
= (offset_bytes
>> 10) & 1;
283 offset_bytes
^= (1 << 6);
285 /* Optimized, obfuscated version. */
286 offset_bytes
^= ((offset_bytes
>> 3) ^ (offset_bytes
>> 4))
291 const uint32_t swizzle_bound_pixels
= ALIGN(x_pixels
+ 1, swizzle_width_pixels
);
292 const uint32_t memcpy_bound_pixels
= MIN2(x_max_pixels
, swizzle_bound_pixels
);
293 const uint32_t copy_size
= cpp
* (memcpy_bound_pixels
- x_pixels
);
295 memcpy(bo
->virtual + offset_bytes
, pixels
, copy_size
);
297 x_pixels
-= (x_pixels
% swizzle_width_pixels
);
301 drm_intel_bo_unmap(bo
);
306 intelTexSubImage(struct gl_context
* ctx
,
308 struct gl_texture_image
*texImage
,
309 GLint xoffset
, GLint yoffset
, GLint zoffset
,
310 GLsizei width
, GLsizei height
, GLsizei depth
,
311 GLenum format
, GLenum type
,
312 const GLvoid
* pixels
,
313 const struct gl_pixelstore_attrib
*packing
)
317 ok
= intel_texsubimage_tiled_memcpy(ctx
, dims
, texImage
,
318 xoffset
, yoffset
, zoffset
,
319 width
, height
, depth
,
320 format
, type
, pixels
, packing
,
321 false /*for_glTexImage*/);
325 /* The intel_blit_texsubimage() function only handles 2D images */
326 if (dims
!= 2 || !intel_blit_texsubimage(ctx
, texImage
,
329 format
, type
, pixels
, packing
)) {
330 _mesa_store_texsubimage(ctx
, dims
, texImage
,
331 xoffset
, yoffset
, zoffset
,
332 width
, height
, depth
,
333 format
, type
, pixels
, packing
);
338 intelInitTextureSubImageFuncs(struct dd_function_table
*functions
)
340 functions
->TexSubImage
= intelTexSubImage
;