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/mtypes.h"
27 #include "main/blit.h"
28 #include "main/context.h"
29 #include "main/enums.h"
30 #include "main/fbobject.h"
31 
32 #include "brw_context.h"
33 #include "brw_defines.h"
34 #include "intel_blit.h"
35 #include "intel_buffers.h"
36 #include "intel_fbo.h"
37 #include "intel_batchbuffer.h"
38 #include "intel_mipmap_tree.h"
39 
40 #define FILE_DEBUG_FLAG DEBUG_BLIT
41 
42 static void
43 intel_miptree_set_alpha_to_one(struct brw_context *brw,
44                                struct intel_mipmap_tree *mt,
45                                int x, int y, int width, int height);
46 
translate_raster_op(GLenum logicop)47 static GLuint translate_raster_op(GLenum logicop)
48 {
49    switch(logicop) {
50    case GL_CLEAR: return 0x00;
51    case GL_AND: return 0x88;
52    case GL_AND_REVERSE: return 0x44;
53    case GL_COPY: return 0xCC;
54    case GL_AND_INVERTED: return 0x22;
55    case GL_NOOP: return 0xAA;
56    case GL_XOR: return 0x66;
57    case GL_OR: return 0xEE;
58    case GL_NOR: return 0x11;
59    case GL_EQUIV: return 0x99;
60    case GL_INVERT: return 0x55;
61    case GL_OR_REVERSE: return 0xDD;
62    case GL_COPY_INVERTED: return 0x33;
63    case GL_OR_INVERTED: return 0xBB;
64    case GL_NAND: return 0x77;
65    case GL_SET: return 0xFF;
66    default: return 0;
67    }
68 }
69 
70 static uint32_t
br13_for_cpp(int cpp)71 br13_for_cpp(int cpp)
72 {
73    switch (cpp) {
74    case 16:
75       return BR13_32323232;
76    case 8:
77       return BR13_16161616;
78    case 4:
79       return BR13_8888;
80    case 2:
81       return BR13_565;
82    case 1:
83       return BR13_8;
84    default:
85       unreachable("not reached");
86    }
87 }
88 
89 /**
90  * Emits the packet for switching the blitter from X to Y tiled or back.
91  *
92  * This has to be called in a single BEGIN_BATCH_BLT_TILED() /
93  * ADVANCE_BATCH_TILED().  This is because BCS_SWCTRL is saved and restored as
94  * part of the power context, not a render context, and if the batchbuffer was
95  * to get flushed between setting and blitting, or blitting and restoring, our
96  * tiling state would leak into other unsuspecting applications (like the X
97  * server).
98  */
99 static uint32_t *
set_blitter_tiling(struct brw_context * brw,bool dst_y_tiled,bool src_y_tiled,uint32_t * __map)100 set_blitter_tiling(struct brw_context *brw,
101                    bool dst_y_tiled, bool src_y_tiled,
102                    uint32_t *__map)
103 {
104    const struct gen_device_info *devinfo = &brw->screen->devinfo;
105    const unsigned n_dwords = devinfo->gen >= 8 ? 5 : 4;
106    assert(devinfo->gen >= 6);
107 
108    /* Idle the blitter before we update how tiling is interpreted. */
109    OUT_BATCH(MI_FLUSH_DW | (n_dwords - 2));
110    OUT_BATCH(0);
111    OUT_BATCH(0);
112    OUT_BATCH(0);
113    if (n_dwords == 5)
114       OUT_BATCH(0);
115 
116    OUT_BATCH(MI_LOAD_REGISTER_IMM | (3 - 2));
117    OUT_BATCH(BCS_SWCTRL);
118    OUT_BATCH((BCS_SWCTRL_DST_Y | BCS_SWCTRL_SRC_Y) << 16 |
119              (dst_y_tiled ? BCS_SWCTRL_DST_Y : 0) |
120              (src_y_tiled ? BCS_SWCTRL_SRC_Y : 0));
121    return __map;
122 }
123 #define SET_BLITTER_TILING(...) __map = set_blitter_tiling(__VA_ARGS__, __map)
124 
125 #define BEGIN_BATCH_BLT_TILED(n, dst_y_tiled, src_y_tiled)              \
126       unsigned set_tiling_batch_size = 0;                               \
127       if (dst_y_tiled || src_y_tiled) {                                 \
128          if (devinfo->gen >= 8)                                         \
129             set_tiling_batch_size = 16;                                 \
130          else                                                           \
131             set_tiling_batch_size = 14;                                 \
132       }                                                                 \
133       BEGIN_BATCH_BLT(n + set_tiling_batch_size);                       \
134       if (dst_y_tiled || src_y_tiled)                                   \
135          SET_BLITTER_TILING(brw, dst_y_tiled, src_y_tiled)
136 
137 #define ADVANCE_BATCH_TILED(dst_y_tiled, src_y_tiled)                   \
138       if (dst_y_tiled || src_y_tiled)                                   \
139          SET_BLITTER_TILING(brw, false, false);                         \
140       ADVANCE_BATCH()
141 
142 static int
blt_pitch(struct intel_mipmap_tree * mt)143 blt_pitch(struct intel_mipmap_tree *mt)
144 {
145    int pitch = mt->surf.row_pitch;
146    if (mt->surf.tiling != ISL_TILING_LINEAR)
147       pitch /= 4;
148    return pitch;
149 }
150 
151 bool
intel_miptree_blit_compatible_formats(mesa_format src,mesa_format dst)152 intel_miptree_blit_compatible_formats(mesa_format src, mesa_format dst)
153 {
154    /* The BLT doesn't handle sRGB conversion */
155    assert(src == _mesa_get_srgb_format_linear(src));
156    assert(dst == _mesa_get_srgb_format_linear(dst));
157 
158    /* No swizzle or format conversions possible, except... */
159    if (src == dst)
160       return true;
161 
162    /* ...we can either discard the alpha channel when going from A->X,
163     * or we can fill the alpha channel with 0xff when going from X->A
164     */
165    if (src == MESA_FORMAT_B8G8R8A8_UNORM || src == MESA_FORMAT_B8G8R8X8_UNORM)
166       return (dst == MESA_FORMAT_B8G8R8A8_UNORM ||
167               dst == MESA_FORMAT_B8G8R8X8_UNORM);
168 
169    if (src == MESA_FORMAT_R8G8B8A8_UNORM || src == MESA_FORMAT_R8G8B8X8_UNORM)
170       return (dst == MESA_FORMAT_R8G8B8A8_UNORM ||
171               dst == MESA_FORMAT_R8G8B8X8_UNORM);
172 
173    /* We can also discard alpha when going from A2->X2 for 2 bit alpha,
174     * however we can't fill the alpha channel with two 1 bits when going
175     * from X2->A2, because intel_miptree_set_alpha_to_one() is not yet
176     * ready for this / can only handle 8 bit alpha.
177     */
178    if (src == MESA_FORMAT_B10G10R10A2_UNORM)
179       return (dst == MESA_FORMAT_B10G10R10A2_UNORM ||
180               dst == MESA_FORMAT_B10G10R10X2_UNORM);
181 
182    if (src == MESA_FORMAT_R10G10B10A2_UNORM)
183       return (dst == MESA_FORMAT_R10G10B10A2_UNORM ||
184               dst == MESA_FORMAT_R10G10B10X2_UNORM);
185 
186    return false;
187 }
188 
189 static void
get_blit_intratile_offset_el(const struct brw_context * brw,struct intel_mipmap_tree * mt,uint32_t total_x_offset_el,uint32_t total_y_offset_el,uint32_t * base_address_offset,uint32_t * x_offset_el,uint32_t * y_offset_el)190 get_blit_intratile_offset_el(const struct brw_context *brw,
191                              struct intel_mipmap_tree *mt,
192                              uint32_t total_x_offset_el,
193                              uint32_t total_y_offset_el,
194                              uint32_t *base_address_offset,
195                              uint32_t *x_offset_el,
196                              uint32_t *y_offset_el)
197 {
198    isl_tiling_get_intratile_offset_el(mt->surf.tiling,
199                                       mt->cpp * 8, mt->surf.row_pitch,
200                                       total_x_offset_el, total_y_offset_el,
201                                       base_address_offset,
202                                       x_offset_el, y_offset_el);
203    if (mt->surf.tiling == ISL_TILING_LINEAR) {
204       /* From the Broadwell PRM docs for XY_SRC_COPY_BLT::SourceBaseAddress:
205        *
206        *    "Base address of the destination surface: X=0, Y=0. Lower 32bits
207        *    of the 48bit addressing. When Src Tiling is enabled (Bit_15
208        *    enabled), this address must be 4KB-aligned. When Tiling is not
209        *    enabled, this address should be CL (64byte) aligned."
210        *
211        * The offsets we get from ISL in the tiled case are already aligned.
212        * In the linear case, we need to do some of our own aligning.
213        */
214       uint32_t delta = *base_address_offset & 63;
215       assert(delta % mt->cpp == 0);
216       *base_address_offset -= delta;
217       *x_offset_el += delta / mt->cpp;
218    } else {
219       assert(*base_address_offset % 4096 == 0);
220    }
221 }
222 
223 static bool
emit_miptree_blit(struct brw_context * brw,struct intel_mipmap_tree * src_mt,uint32_t src_x,uint32_t src_y,struct intel_mipmap_tree * dst_mt,uint32_t dst_x,uint32_t dst_y,uint32_t width,uint32_t height,bool reverse,GLenum logicop)224 emit_miptree_blit(struct brw_context *brw,
225                   struct intel_mipmap_tree *src_mt,
226                   uint32_t src_x, uint32_t src_y,
227                   struct intel_mipmap_tree *dst_mt,
228                   uint32_t dst_x, uint32_t dst_y,
229                   uint32_t width, uint32_t height,
230                   bool reverse, GLenum logicop)
231 {
232    /* According to the Ivy Bridge PRM, Vol1 Part4, section 1.2.1.2 (Graphics
233     * Data Size Limitations):
234     *
235     *    The BLT engine is capable of transferring very large quantities of
236     *    graphics data. Any graphics data read from and written to the
237     *    destination is permitted to represent a number of pixels that
238     *    occupies up to 65,536 scan lines and up to 32,768 bytes per scan line
239     *    at the destination. The maximum number of pixels that may be
240     *    represented per scan line’s worth of graphics data depends on the
241     *    color depth.
242     *
243     * The blitter's pitch is a signed 16-bit integer, but measured in bytes
244     * for linear surfaces and DWords for tiled surfaces.  So the maximum
245     * pitch is 32k linear and 128k tiled.
246     */
247    if (blt_pitch(src_mt) >= 32768 || blt_pitch(dst_mt) >= 32768) {
248       perf_debug("Falling back due to >= 32k/128k pitch\n");
249       return false;
250    }
251 
252    /* We need to split the blit into chunks that each fit within the blitter's
253     * restrictions.  We can't use a chunk size of 32768 because we need to
254     * ensure that src_tile_x + chunk_size fits.  We choose 16384 because it's
255     * a nice round power of two, big enough that performance won't suffer, and
256     * small enough to guarantee everything fits.
257     */
258    const uint32_t max_chunk_size = 16384;
259 
260    for (uint32_t chunk_x = 0; chunk_x < width; chunk_x += max_chunk_size) {
261       for (uint32_t chunk_y = 0; chunk_y < height; chunk_y += max_chunk_size) {
262          const uint32_t chunk_w = MIN2(max_chunk_size, width - chunk_x);
263          const uint32_t chunk_h = MIN2(max_chunk_size, height - chunk_y);
264 
265          uint32_t src_offset, src_tile_x, src_tile_y;
266          get_blit_intratile_offset_el(brw, src_mt,
267                                       src_x + chunk_x, src_y + chunk_y,
268                                       &src_offset, &src_tile_x, &src_tile_y);
269 
270          uint32_t dst_offset, dst_tile_x, dst_tile_y;
271          get_blit_intratile_offset_el(brw, dst_mt,
272                                       dst_x + chunk_x, dst_y + chunk_y,
273                                       &dst_offset, &dst_tile_x, &dst_tile_y);
274 
275          if (!intelEmitCopyBlit(brw,
276                                 src_mt->cpp,
277                                 reverse ? -src_mt->surf.row_pitch :
278                                            src_mt->surf.row_pitch,
279                                 src_mt->bo, src_mt->offset + src_offset,
280                                 src_mt->surf.tiling,
281                                 dst_mt->surf.row_pitch,
282                                 dst_mt->bo, dst_mt->offset + dst_offset,
283                                 dst_mt->surf.tiling,
284                                 src_tile_x, src_tile_y,
285                                 dst_tile_x, dst_tile_y,
286                                 chunk_w, chunk_h,
287                                 logicop)) {
288             /* If this is ever going to fail, it will fail on the first chunk */
289             assert(chunk_x == 0 && chunk_y == 0);
290             return false;
291          }
292       }
293    }
294 
295    return true;
296 }
297 
298 /**
299  * Implements a rectangular block transfer (blit) of pixels between two
300  * miptrees.
301  *
302  * Our blitter can operate on 1, 2, or 4-byte-per-pixel data, with generous,
303  * but limited, pitches and sizes allowed.
304  *
305  * The src/dst coordinates are relative to the given level/slice of the
306  * miptree.
307  *
308  * If @src_flip or @dst_flip is set, then the rectangle within that miptree
309  * will be inverted (including scanline order) when copying.  This is common
310  * in GL when copying between window system and user-created
311  * renderbuffers/textures.
312  */
313 bool
intel_miptree_blit(struct brw_context * brw,struct intel_mipmap_tree * src_mt,int src_level,int src_slice,uint32_t src_x,uint32_t src_y,bool src_flip,struct intel_mipmap_tree * dst_mt,int dst_level,int dst_slice,uint32_t dst_x,uint32_t dst_y,bool dst_flip,uint32_t width,uint32_t height,GLenum logicop)314 intel_miptree_blit(struct brw_context *brw,
315                    struct intel_mipmap_tree *src_mt,
316                    int src_level, int src_slice,
317                    uint32_t src_x, uint32_t src_y, bool src_flip,
318                    struct intel_mipmap_tree *dst_mt,
319                    int dst_level, int dst_slice,
320                    uint32_t dst_x, uint32_t dst_y, bool dst_flip,
321                    uint32_t width, uint32_t height,
322                    GLenum logicop)
323 {
324    /* The blitter doesn't understand multisampling at all. */
325    if (src_mt->surf.samples > 1 || dst_mt->surf.samples > 1)
326       return false;
327 
328    /* No sRGB decode or encode is done by the hardware blitter, which is
329     * consistent with what we want in many callers (glCopyTexSubImage(),
330     * texture validation, etc.).
331     */
332    mesa_format src_format = _mesa_get_srgb_format_linear(src_mt->format);
333    mesa_format dst_format = _mesa_get_srgb_format_linear(dst_mt->format);
334 
335    /* The blitter doesn't support doing any format conversions.  We do also
336     * support blitting ARGB8888 to XRGB8888 (trivial, the values dropped into
337     * the X channel don't matter), and XRGB8888 to ARGB8888 by setting the A
338     * channel to 1.0 at the end. Also trivially ARGB2101010 to XRGB2101010,
339     * but not XRGB2101010 to ARGB2101010 yet.
340     */
341    if (!intel_miptree_blit_compatible_formats(src_format, dst_format)) {
342       perf_debug("%s: Can't use hardware blitter from %s to %s, "
343                  "falling back.\n", __func__,
344                  _mesa_get_format_name(src_format),
345                  _mesa_get_format_name(dst_format));
346       return false;
347    }
348 
349    /* The blitter has no idea about HiZ or fast color clears, so we need to
350     * resolve the miptrees before we do anything.
351     */
352    intel_miptree_access_raw(brw, src_mt, src_level, src_slice, false);
353    intel_miptree_access_raw(brw, dst_mt, dst_level, dst_slice, true);
354 
355    if (src_flip) {
356       const unsigned h0 = src_mt->surf.phys_level0_sa.height;
357       src_y = minify(h0, src_level - src_mt->first_level) - src_y - height;
358    }
359 
360    if (dst_flip) {
361       const unsigned h0 = dst_mt->surf.phys_level0_sa.height;
362       dst_y = minify(h0, dst_level - dst_mt->first_level) - dst_y - height;
363    }
364 
365    uint32_t src_image_x, src_image_y, dst_image_x, dst_image_y;
366    intel_miptree_get_image_offset(src_mt, src_level, src_slice,
367                                   &src_image_x, &src_image_y);
368    intel_miptree_get_image_offset(dst_mt, dst_level, dst_slice,
369                                   &dst_image_x, &dst_image_y);
370    src_x += src_image_x;
371    src_y += src_image_y;
372    dst_x += dst_image_x;
373    dst_y += dst_image_y;
374 
375    if (!emit_miptree_blit(brw, src_mt, src_x, src_y,
376                           dst_mt, dst_x, dst_y, width, height,
377                           src_flip != dst_flip, logicop)) {
378       return false;
379    }
380 
381    /* XXX This could be done in a single pass using XY_FULL_MONO_PATTERN_BLT */
382    if (_mesa_get_format_bits(src_format, GL_ALPHA_BITS) == 0 &&
383        _mesa_get_format_bits(dst_format, GL_ALPHA_BITS) > 0) {
384       intel_miptree_set_alpha_to_one(brw, dst_mt,
385                                      dst_x, dst_y,
386                                      width, height);
387    }
388 
389    return true;
390 }
391 
392 bool
intel_miptree_copy(struct brw_context * brw,struct intel_mipmap_tree * src_mt,int src_level,int src_slice,uint32_t src_x,uint32_t src_y,struct intel_mipmap_tree * dst_mt,int dst_level,int dst_slice,uint32_t dst_x,uint32_t dst_y,uint32_t src_width,uint32_t src_height)393 intel_miptree_copy(struct brw_context *brw,
394                    struct intel_mipmap_tree *src_mt,
395                    int src_level, int src_slice,
396                    uint32_t src_x, uint32_t src_y,
397                    struct intel_mipmap_tree *dst_mt,
398                    int dst_level, int dst_slice,
399                    uint32_t dst_x, uint32_t dst_y,
400                    uint32_t src_width, uint32_t src_height)
401 {
402    /* The blitter doesn't understand multisampling at all. */
403    if (src_mt->surf.samples > 1 || dst_mt->surf.samples > 1)
404       return false;
405 
406    if (src_mt->format == MESA_FORMAT_S_UINT8)
407       return false;
408 
409    /* The blitter has no idea about HiZ or fast color clears, so we need to
410     * resolve the miptrees before we do anything.
411     */
412    intel_miptree_access_raw(brw, src_mt, src_level, src_slice, false);
413    intel_miptree_access_raw(brw, dst_mt, dst_level, dst_slice, true);
414 
415    uint32_t src_image_x, src_image_y;
416    intel_miptree_get_image_offset(src_mt, src_level, src_slice,
417                                   &src_image_x, &src_image_y);
418 
419    if (_mesa_is_format_compressed(src_mt->format)) {
420       GLuint bw, bh;
421       _mesa_get_format_block_size(src_mt->format, &bw, &bh);
422 
423       /* Compressed textures need not have dimensions that are a multiple of
424        * the block size.  Rectangles in compressed textures do need to be a
425        * multiple of the block size.  The one exception is that the right and
426        * bottom edges may be at the right or bottom edge of the miplevel even
427        * if it's not aligned.
428        */
429       assert(src_x % bw == 0);
430       assert(src_y % bh == 0);
431 
432       assert(src_width % bw == 0 ||
433              src_x + src_width ==
434              minify(src_mt->surf.logical_level0_px.width, src_level));
435       assert(src_height % bh == 0 ||
436              src_y + src_height ==
437              minify(src_mt->surf.logical_level0_px.height, src_level));
438 
439       src_x /= (int)bw;
440       src_y /= (int)bh;
441       src_width = DIV_ROUND_UP(src_width, (int)bw);
442       src_height = DIV_ROUND_UP(src_height, (int)bh);
443    }
444    src_x += src_image_x;
445    src_y += src_image_y;
446 
447    uint32_t dst_image_x, dst_image_y;
448    intel_miptree_get_image_offset(dst_mt, dst_level, dst_slice,
449                                   &dst_image_x, &dst_image_y);
450 
451    if (_mesa_is_format_compressed(dst_mt->format)) {
452       GLuint bw, bh;
453       _mesa_get_format_block_size(dst_mt->format, &bw, &bh);
454 
455       assert(dst_x % bw == 0);
456       assert(dst_y % bh == 0);
457 
458       dst_x /= (int)bw;
459       dst_y /= (int)bh;
460    }
461    dst_x += dst_image_x;
462    dst_y += dst_image_y;
463 
464    return emit_miptree_blit(brw, src_mt, src_x, src_y,
465                             dst_mt, dst_x, dst_y,
466                             src_width, src_height, false, GL_COPY);
467 }
468 
469 static bool
alignment_valid(struct brw_context * brw,unsigned offset,enum isl_tiling tiling)470 alignment_valid(struct brw_context *brw, unsigned offset,
471                 enum isl_tiling tiling)
472 {
473    const struct gen_device_info *devinfo = &brw->screen->devinfo;
474 
475    /* Tiled buffers must be page-aligned (4K). */
476    if (tiling != ISL_TILING_LINEAR)
477       return (offset & 4095) == 0;
478 
479    /* On Gen8+, linear buffers must be cacheline-aligned. */
480    if (devinfo->gen >= 8)
481       return (offset & 63) == 0;
482 
483    return true;
484 }
485 
486 static uint32_t
xy_blit_cmd(enum isl_tiling src_tiling,enum isl_tiling dst_tiling,uint32_t cpp)487 xy_blit_cmd(enum isl_tiling src_tiling, enum isl_tiling dst_tiling,
488             uint32_t cpp)
489 {
490    uint32_t CMD = 0;
491 
492    assert(cpp <= 4);
493    switch (cpp) {
494    case 1:
495    case 2:
496       CMD = XY_SRC_COPY_BLT_CMD;
497       break;
498    case 4:
499       CMD = XY_SRC_COPY_BLT_CMD | XY_BLT_WRITE_ALPHA | XY_BLT_WRITE_RGB;
500       break;
501    default:
502       unreachable("not reached");
503    }
504 
505    if (dst_tiling != ISL_TILING_LINEAR)
506       CMD |= XY_DST_TILED;
507 
508    if (src_tiling != ISL_TILING_LINEAR)
509       CMD |= XY_SRC_TILED;
510 
511    return CMD;
512 }
513 
514 /* Copy BitBlt
515  */
516 bool
intelEmitCopyBlit(struct brw_context * brw,GLuint cpp,int32_t src_pitch,struct brw_bo * src_buffer,GLuint src_offset,enum isl_tiling src_tiling,int32_t dst_pitch,struct brw_bo * dst_buffer,GLuint dst_offset,enum isl_tiling dst_tiling,GLshort src_x,GLshort src_y,GLshort dst_x,GLshort dst_y,GLshort w,GLshort h,GLenum logic_op)517 intelEmitCopyBlit(struct brw_context *brw,
518 		  GLuint cpp,
519 		  int32_t src_pitch,
520 		  struct brw_bo *src_buffer,
521 		  GLuint src_offset,
522 		  enum isl_tiling src_tiling,
523 		  int32_t dst_pitch,
524 		  struct brw_bo *dst_buffer,
525 		  GLuint dst_offset,
526 		  enum isl_tiling dst_tiling,
527 		  GLshort src_x, GLshort src_y,
528 		  GLshort dst_x, GLshort dst_y,
529 		  GLshort w, GLshort h,
530 		  GLenum logic_op)
531 {
532    const struct gen_device_info *devinfo = &brw->screen->devinfo;
533    GLuint CMD, BR13;
534    int dst_y2 = dst_y + h;
535    int dst_x2 = dst_x + w;
536    bool dst_y_tiled = dst_tiling == ISL_TILING_Y0;
537    bool src_y_tiled = src_tiling == ISL_TILING_Y0;
538    uint32_t src_tile_w, src_tile_h;
539    uint32_t dst_tile_w, dst_tile_h;
540 
541    if ((dst_y_tiled || src_y_tiled) && devinfo->gen < 6)
542       return false;
543 
544    const unsigned bo_sizes = dst_buffer->size + src_buffer->size;
545 
546    /* do space check before going any further */
547    if (!brw_batch_has_aperture_space(brw, bo_sizes))
548       intel_batchbuffer_flush(brw);
549 
550    if (!brw_batch_has_aperture_space(brw, bo_sizes))
551       return false;
552 
553    unsigned length = devinfo->gen >= 8 ? 10 : 8;
554 
555    intel_batchbuffer_require_space(brw, length * 4, BLT_RING);
556    DBG("%s src:buf(%p)/%d+%d %d,%d dst:buf(%p)/%d+%d %d,%d sz:%dx%d\n",
557        __func__,
558        src_buffer, src_pitch, src_offset, src_x, src_y,
559        dst_buffer, dst_pitch, dst_offset, dst_x, dst_y, w, h);
560 
561    intel_get_tile_dims(src_tiling, cpp, &src_tile_w, &src_tile_h);
562    intel_get_tile_dims(dst_tiling, cpp, &dst_tile_w, &dst_tile_h);
563 
564    /* For Tiled surfaces, the pitch has to be a multiple of the Tile width
565     * (X direction width of the Tile). This is ensured while allocating the
566     * buffer object.
567     */
568    assert(src_tiling == ISL_TILING_LINEAR || (src_pitch % src_tile_w) == 0);
569    assert(dst_tiling == ISL_TILING_LINEAR || (dst_pitch % dst_tile_w) == 0);
570 
571    /* For big formats (such as floating point), do the copy using 16 or
572     * 32bpp and multiply the coordinates.
573     */
574    if (cpp > 4) {
575       if (cpp % 4 == 2) {
576          dst_x *= cpp / 2;
577          dst_x2 *= cpp / 2;
578          src_x *= cpp / 2;
579          cpp = 2;
580       } else {
581          assert(cpp % 4 == 0);
582          dst_x *= cpp / 4;
583          dst_x2 *= cpp / 4;
584          src_x *= cpp / 4;
585          cpp = 4;
586       }
587    }
588 
589    if (!alignment_valid(brw, dst_offset, dst_tiling))
590       return false;
591    if (!alignment_valid(brw, src_offset, src_tiling))
592       return false;
593 
594    /* Blit pitch must be dword-aligned.  Otherwise, the hardware appears to drop
595     * the low bits.  Offsets must be naturally aligned.
596     */
597    if (src_pitch % 4 != 0 || src_offset % cpp != 0 ||
598        dst_pitch % 4 != 0 || dst_offset % cpp != 0)
599       return false;
600 
601    assert(cpp <= 4);
602    BR13 = br13_for_cpp(cpp) | translate_raster_op(logic_op) << 16;
603 
604    CMD = xy_blit_cmd(src_tiling, dst_tiling, cpp);
605 
606    /* For tiled source and destination, pitch value should be specified
607     * as a number of Dwords.
608     */
609    if (dst_tiling != ISL_TILING_LINEAR)
610       dst_pitch /= 4;
611 
612    if (src_tiling != ISL_TILING_LINEAR)
613       src_pitch /= 4;
614 
615    if (dst_y2 <= dst_y || dst_x2 <= dst_x)
616       return true;
617 
618    assert(dst_x < dst_x2);
619    assert(dst_y < dst_y2);
620 
621    BEGIN_BATCH_BLT_TILED(length, dst_y_tiled, src_y_tiled);
622    OUT_BATCH(CMD | (length - 2));
623    OUT_BATCH(BR13 | (uint16_t)dst_pitch);
624    OUT_BATCH(SET_FIELD(dst_y, BLT_Y) | SET_FIELD(dst_x, BLT_X));
625    OUT_BATCH(SET_FIELD(dst_y2, BLT_Y) | SET_FIELD(dst_x2, BLT_X));
626    if (devinfo->gen >= 8) {
627       OUT_RELOC64(dst_buffer, RELOC_WRITE, dst_offset);
628    } else {
629       OUT_RELOC(dst_buffer, RELOC_WRITE, dst_offset);
630    }
631    OUT_BATCH(SET_FIELD(src_y, BLT_Y) | SET_FIELD(src_x, BLT_X));
632    OUT_BATCH((uint16_t)src_pitch);
633    if (devinfo->gen >= 8) {
634       OUT_RELOC64(src_buffer, 0, src_offset);
635    } else {
636       OUT_RELOC(src_buffer, 0, src_offset);
637    }
638 
639    ADVANCE_BATCH_TILED(dst_y_tiled, src_y_tiled);
640 
641    brw_emit_mi_flush(brw);
642 
643    return true;
644 }
645 
646 bool
intelEmitImmediateColorExpandBlit(struct brw_context * brw,GLuint cpp,GLubyte * src_bits,GLuint src_size,GLuint fg_color,GLshort dst_pitch,struct brw_bo * dst_buffer,GLuint dst_offset,enum isl_tiling dst_tiling,GLshort x,GLshort y,GLshort w,GLshort h,GLenum logic_op)647 intelEmitImmediateColorExpandBlit(struct brw_context *brw,
648 				  GLuint cpp,
649 				  GLubyte *src_bits, GLuint src_size,
650 				  GLuint fg_color,
651 				  GLshort dst_pitch,
652 				  struct brw_bo *dst_buffer,
653 				  GLuint dst_offset,
654 				  enum isl_tiling dst_tiling,
655 				  GLshort x, GLshort y,
656 				  GLshort w, GLshort h,
657 				  GLenum logic_op)
658 {
659    const struct gen_device_info *devinfo = &brw->screen->devinfo;
660    int dwords = ALIGN(src_size, 8) / 4;
661    uint32_t opcode, br13, blit_cmd;
662 
663    if (dst_tiling != ISL_TILING_LINEAR) {
664       if (dst_offset & 4095)
665 	 return false;
666       if (dst_tiling == ISL_TILING_Y0)
667 	 return false;
668    }
669 
670    assert((logic_op >= GL_CLEAR) && (logic_op <= (GL_CLEAR + 0x0f)));
671    assert(dst_pitch > 0);
672 
673    if (w < 0 || h < 0)
674       return true;
675 
676    DBG("%s dst:buf(%p)/%d+%d %d,%d sz:%dx%d, %d bytes %d dwords\n",
677        __func__,
678        dst_buffer, dst_pitch, dst_offset, x, y, w, h, src_size, dwords);
679 
680    unsigned xy_setup_blt_length = devinfo->gen >= 8 ? 10 : 8;
681    intel_batchbuffer_require_space(brw, (xy_setup_blt_length * 4) +
682                                         (3 * 4) + dwords * 4, BLT_RING);
683 
684    opcode = XY_SETUP_BLT_CMD;
685    if (cpp == 4)
686       opcode |= XY_BLT_WRITE_ALPHA | XY_BLT_WRITE_RGB;
687    if (dst_tiling != ISL_TILING_LINEAR) {
688       opcode |= XY_DST_TILED;
689       dst_pitch /= 4;
690    }
691 
692    br13 = dst_pitch | (translate_raster_op(logic_op) << 16) | (1 << 29);
693    br13 |= br13_for_cpp(cpp);
694 
695    blit_cmd = XY_TEXT_IMMEDIATE_BLIT_CMD | XY_TEXT_BYTE_PACKED; /* packing? */
696    if (dst_tiling != ISL_TILING_LINEAR)
697       blit_cmd |= XY_DST_TILED;
698 
699    BEGIN_BATCH_BLT(xy_setup_blt_length + 3);
700    OUT_BATCH(opcode | (xy_setup_blt_length - 2));
701    OUT_BATCH(br13);
702    OUT_BATCH((0 << 16) | 0); /* clip x1, y1 */
703    OUT_BATCH((100 << 16) | 100); /* clip x2, y2 */
704    if (devinfo->gen >= 8) {
705       OUT_RELOC64(dst_buffer, RELOC_WRITE, dst_offset);
706    } else {
707       OUT_RELOC(dst_buffer, RELOC_WRITE, dst_offset);
708    }
709    OUT_BATCH(0); /* bg */
710    OUT_BATCH(fg_color); /* fg */
711    OUT_BATCH(0); /* pattern base addr */
712    if (devinfo->gen >= 8)
713       OUT_BATCH(0);
714 
715    OUT_BATCH(blit_cmd | ((3 - 2) + dwords));
716    OUT_BATCH(SET_FIELD(y, BLT_Y) | SET_FIELD(x, BLT_X));
717    OUT_BATCH(SET_FIELD(y + h, BLT_Y) | SET_FIELD(x + w, BLT_X));
718    ADVANCE_BATCH();
719 
720    intel_batchbuffer_data(brw, src_bits, dwords * 4, BLT_RING);
721 
722    brw_emit_mi_flush(brw);
723 
724    return true;
725 }
726 
727 /* We don't have a memmove-type blit like some other hardware, so we'll do a
728  * rectangular blit covering a large space, then emit 1-scanline blit at the
729  * end to cover the last if we need.
730  */
731 void
intel_emit_linear_blit(struct brw_context * brw,struct brw_bo * dst_bo,unsigned int dst_offset,struct brw_bo * src_bo,unsigned int src_offset,unsigned int size)732 intel_emit_linear_blit(struct brw_context *brw,
733 		       struct brw_bo *dst_bo,
734 		       unsigned int dst_offset,
735 		       struct brw_bo *src_bo,
736 		       unsigned int src_offset,
737 		       unsigned int size)
738 {
739    struct gl_context *ctx = &brw->ctx;
740    GLuint pitch, height;
741    int16_t src_x, dst_x;
742    bool ok;
743 
744    do {
745       /* The pitch given to the GPU must be DWORD aligned, and
746        * we want width to match pitch. Max width is (1 << 15 - 1),
747        * rounding that down to the nearest DWORD is 1 << 15 - 4
748        */
749       pitch = ROUND_DOWN_TO(MIN2(size, (1 << 15) - 64), 4);
750       height = (size < pitch || pitch == 0) ? 1 : size / pitch;
751 
752       src_x = src_offset % 64;
753       dst_x = dst_offset % 64;
754       pitch = ALIGN(MIN2(size, (1 << 15) - 64), 4);
755       assert(src_x + pitch < 1 << 15);
756       assert(dst_x + pitch < 1 << 15);
757 
758       ok = intelEmitCopyBlit(brw, 1,
759                              pitch, src_bo, src_offset - src_x,
760                              ISL_TILING_LINEAR,
761                              pitch, dst_bo, dst_offset - dst_x,
762                              ISL_TILING_LINEAR,
763                              src_x, 0, /* src x/y */
764                              dst_x, 0, /* dst x/y */
765                              MIN2(size, pitch), height, /* w, h */
766                              GL_COPY);
767       if (!ok) {
768          _mesa_problem(ctx, "Failed to linear blit %dx%d\n",
769                        MIN2(size, pitch), height);
770          return;
771       }
772 
773       pitch *= height;
774       if (size <= pitch)
775          return;
776 
777       src_offset += pitch;
778       dst_offset += pitch;
779       size -= pitch;
780    } while (1);
781 }
782 
783 /**
784  * Used to initialize the alpha value of an ARGB8888 miptree after copying
785  * into it from an XRGB8888 source.
786  *
787  * This is very common with glCopyTexImage2D().  Note that the coordinates are
788  * relative to the start of the miptree, not relative to a slice within the
789  * miptree.
790  */
791 static void
intel_miptree_set_alpha_to_one(struct brw_context * brw,struct intel_mipmap_tree * mt,int x,int y,int width,int height)792 intel_miptree_set_alpha_to_one(struct brw_context *brw,
793                               struct intel_mipmap_tree *mt,
794                               int x, int y, int width, int height)
795 {
796    const struct gen_device_info *devinfo = &brw->screen->devinfo;
797    uint32_t BR13, CMD;
798    int pitch, cpp;
799 
800    pitch = mt->surf.row_pitch;
801    cpp = mt->cpp;
802 
803    DBG("%s dst:buf(%p)/%d %d,%d sz:%dx%d\n",
804        __func__, mt->bo, pitch, x, y, width, height);
805 
806    /* Note: Currently only handles 8 bit alpha channel. Extension to < 8 Bit
807     * alpha channel would be likely possible via ROP code 0xfa instead of 0xf0
808     * and writing a suitable bit-mask instead of 0xffffffff.
809     */
810    BR13 = br13_for_cpp(cpp) | 0xf0 << 16;
811    CMD = XY_COLOR_BLT_CMD;
812    CMD |= XY_BLT_WRITE_ALPHA;
813 
814    if (mt->surf.tiling != ISL_TILING_LINEAR) {
815       CMD |= XY_DST_TILED;
816       pitch /= 4;
817    }
818    BR13 |= pitch;
819 
820    /* do space check before going any further */
821    if (!brw_batch_has_aperture_space(brw, mt->bo->size))
822       intel_batchbuffer_flush(brw);
823 
824    unsigned length = devinfo->gen >= 8 ? 7 : 6;
825    const bool dst_y_tiled = mt->surf.tiling == ISL_TILING_Y0;
826 
827    /* We need to split the blit into chunks that each fit within the blitter's
828     * restrictions.  We can't use a chunk size of 32768 because we need to
829     * ensure that src_tile_x + chunk_size fits.  We choose 16384 because it's
830     * a nice round power of two, big enough that performance won't suffer, and
831     * small enough to guarantee everything fits.
832     */
833    const uint32_t max_chunk_size = 16384;
834 
835    for (uint32_t chunk_x = 0; chunk_x < width; chunk_x += max_chunk_size) {
836       for (uint32_t chunk_y = 0; chunk_y < height; chunk_y += max_chunk_size) {
837          const uint32_t chunk_w = MIN2(max_chunk_size, width - chunk_x);
838          const uint32_t chunk_h = MIN2(max_chunk_size, height - chunk_y);
839 
840          uint32_t offset, tile_x, tile_y;
841          get_blit_intratile_offset_el(brw, mt,
842                                       x + chunk_x, y + chunk_y,
843                                       &offset, &tile_x, &tile_y);
844 
845          BEGIN_BATCH_BLT_TILED(length, dst_y_tiled, false);
846          OUT_BATCH(CMD | (length - 2));
847          OUT_BATCH(BR13);
848          OUT_BATCH(SET_FIELD(y + chunk_y, BLT_Y) |
849                    SET_FIELD(x + chunk_x, BLT_X));
850          OUT_BATCH(SET_FIELD(y + chunk_y + chunk_h, BLT_Y) |
851                    SET_FIELD(x + chunk_x + chunk_w, BLT_X));
852          if (devinfo->gen >= 8) {
853             OUT_RELOC64(mt->bo, RELOC_WRITE, mt->offset + offset);
854          } else {
855             OUT_RELOC(mt->bo, RELOC_WRITE, mt->offset + offset);
856          }
857          OUT_BATCH(0xffffffff); /* white, but only alpha gets written */
858          ADVANCE_BATCH_TILED(dst_y_tiled, false);
859       }
860    }
861 
862    brw_emit_mi_flush(brw);
863 }
864