Details | Last modification | View Log | RSS feed
Rev | Author | Line No. | Line |
---|---|---|---|
5564 | serge | 1 | /* |
2 | * Copyright © 2012 Intel Corporation |
||
3 | * |
||
4 | * Permission is hereby granted, free of charge, to any person obtaining a |
||
5 | * copy of this software and associated documentation files (the "Software"), |
||
6 | * to deal in the Software without restriction, including without limitation |
||
7 | * the rights to use, copy, modify, merge, publish, distribute, sublicense, |
||
8 | * and/or sell copies of the Software, and to permit persons to whom the |
||
9 | * Software is furnished to do so, subject to the following conditions: |
||
10 | * |
||
11 | * The above copyright notice and this permission notice (including the next |
||
12 | * paragraph) shall be included in all copies or substantial portions of the |
||
13 | * Software. |
||
14 | * |
||
15 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
||
16 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
||
17 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL |
||
18 | * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
||
19 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING |
||
20 | * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS |
||
21 | * IN THE SOFTWARE. |
||
22 | */ |
||
23 | |||
24 | #include "main/teximage.h" |
||
25 | #include "main/fbobject.h" |
||
26 | #include "main/renderbuffer.h" |
||
27 | |||
28 | #include "intel_fbo.h" |
||
29 | |||
30 | #include "brw_blorp.h" |
||
31 | #include "brw_context.h" |
||
32 | #include "brw_blorp_blit_eu.h" |
||
33 | #include "brw_state.h" |
||
34 | #include "brw_meta_util.h" |
||
35 | |||
36 | #define FILE_DEBUG_FLAG DEBUG_BLORP |
||
37 | |||
38 | static struct intel_mipmap_tree * |
||
39 | find_miptree(GLbitfield buffer_bit, struct intel_renderbuffer *irb) |
||
40 | { |
||
41 | struct intel_mipmap_tree *mt = irb->mt; |
||
42 | if (buffer_bit == GL_STENCIL_BUFFER_BIT && mt->stencil_mt) |
||
43 | mt = mt->stencil_mt; |
||
44 | return mt; |
||
45 | } |
||
46 | |||
47 | |||
48 | /** |
||
49 | * Note: if the src (or dst) is a 2D multisample array texture on Gen7+ using |
||
50 | * INTEL_MSAA_LAYOUT_UMS or INTEL_MSAA_LAYOUT_CMS, src_layer (dst_layer) is |
||
51 | * the physical layer holding sample 0. So, for example, if |
||
52 | * src_mt->num_samples == 4, then logical layer n corresponds to src_layer == |
||
53 | * 4*n. |
||
54 | */ |
||
55 | void |
||
56 | brw_blorp_blit_miptrees(struct brw_context *brw, |
||
57 | struct intel_mipmap_tree *src_mt, |
||
58 | unsigned src_level, unsigned src_layer, |
||
59 | mesa_format src_format, |
||
60 | struct intel_mipmap_tree *dst_mt, |
||
61 | unsigned dst_level, unsigned dst_layer, |
||
62 | mesa_format dst_format, |
||
63 | float src_x0, float src_y0, |
||
64 | float src_x1, float src_y1, |
||
65 | float dst_x0, float dst_y0, |
||
66 | float dst_x1, float dst_y1, |
||
67 | GLenum filter, bool mirror_x, bool mirror_y) |
||
68 | { |
||
69 | /* Get ready to blit. This includes depth resolving the src and dst |
||
70 | * buffers if necessary. Note: it's not necessary to do a color resolve on |
||
71 | * the destination buffer because we use the standard render path to render |
||
72 | * to destination color buffers, and the standard render path is |
||
73 | * fast-color-aware. |
||
74 | */ |
||
75 | intel_miptree_resolve_color(brw, src_mt); |
||
76 | intel_miptree_slice_resolve_depth(brw, src_mt, src_level, src_layer); |
||
77 | intel_miptree_slice_resolve_depth(brw, dst_mt, dst_level, dst_layer); |
||
78 | |||
79 | DBG("%s from %dx %s mt %p %d %d (%f,%f) (%f,%f)" |
||
80 | "to %dx %s mt %p %d %d (%f,%f) (%f,%f) (flip %d,%d)\n", |
||
81 | __func__, |
||
82 | src_mt->num_samples, _mesa_get_format_name(src_mt->format), src_mt, |
||
83 | src_level, src_layer, src_x0, src_y0, src_x1, src_y1, |
||
84 | dst_mt->num_samples, _mesa_get_format_name(dst_mt->format), dst_mt, |
||
85 | dst_level, dst_layer, dst_x0, dst_y0, dst_x1, dst_y1, |
||
86 | mirror_x, mirror_y); |
||
87 | |||
88 | brw_blorp_blit_params params(brw, |
||
89 | src_mt, src_level, src_layer, src_format, |
||
90 | dst_mt, dst_level, dst_layer, dst_format, |
||
91 | src_x0, src_y0, |
||
92 | src_x1, src_y1, |
||
93 | dst_x0, dst_y0, |
||
94 | dst_x1, dst_y1, |
||
95 | filter, mirror_x, mirror_y); |
||
96 | brw_blorp_exec(brw, ¶ms); |
||
97 | |||
98 | intel_miptree_slice_set_needs_hiz_resolve(dst_mt, dst_level, dst_layer); |
||
99 | } |
||
100 | |||
101 | static void |
||
102 | do_blorp_blit(struct brw_context *brw, GLbitfield buffer_bit, |
||
103 | struct intel_renderbuffer *src_irb, mesa_format src_format, |
||
104 | struct intel_renderbuffer *dst_irb, mesa_format dst_format, |
||
105 | GLfloat srcX0, GLfloat srcY0, GLfloat srcX1, GLfloat srcY1, |
||
106 | GLfloat dstX0, GLfloat dstY0, GLfloat dstX1, GLfloat dstY1, |
||
107 | GLenum filter, bool mirror_x, bool mirror_y) |
||
108 | { |
||
109 | /* Find source/dst miptrees */ |
||
110 | struct intel_mipmap_tree *src_mt = find_miptree(buffer_bit, src_irb); |
||
111 | struct intel_mipmap_tree *dst_mt = find_miptree(buffer_bit, dst_irb); |
||
112 | |||
113 | /* Do the blit */ |
||
114 | brw_blorp_blit_miptrees(brw, |
||
115 | src_mt, src_irb->mt_level, src_irb->mt_layer, |
||
116 | src_format, |
||
117 | dst_mt, dst_irb->mt_level, dst_irb->mt_layer, |
||
118 | dst_format, |
||
119 | srcX0, srcY0, srcX1, srcY1, |
||
120 | dstX0, dstY0, dstX1, dstY1, |
||
121 | filter, mirror_x, mirror_y); |
||
122 | |||
123 | dst_irb->need_downsample = true; |
||
124 | } |
||
125 | |||
126 | static bool |
||
127 | try_blorp_blit(struct brw_context *brw, |
||
128 | const struct gl_framebuffer *read_fb, |
||
129 | const struct gl_framebuffer *draw_fb, |
||
130 | GLfloat srcX0, GLfloat srcY0, GLfloat srcX1, GLfloat srcY1, |
||
131 | GLfloat dstX0, GLfloat dstY0, GLfloat dstX1, GLfloat dstY1, |
||
132 | GLenum filter, GLbitfield buffer_bit) |
||
133 | { |
||
134 | struct gl_context *ctx = &brw->ctx; |
||
135 | |||
136 | /* Sync up the state of window system buffers. We need to do this before |
||
137 | * we go looking for the buffers. |
||
138 | */ |
||
139 | intel_prepare_render(brw); |
||
140 | |||
141 | bool mirror_x, mirror_y; |
||
142 | if (brw_meta_mirror_clip_and_scissor(ctx, read_fb, draw_fb, |
||
143 | &srcX0, &srcY0, &srcX1, &srcY1, |
||
144 | &dstX0, &dstY0, &dstX1, &dstY1, |
||
145 | &mirror_x, &mirror_y)) |
||
146 | return true; |
||
147 | |||
148 | /* Find buffers */ |
||
149 | struct intel_renderbuffer *src_irb; |
||
150 | struct intel_renderbuffer *dst_irb; |
||
151 | struct intel_mipmap_tree *src_mt; |
||
152 | struct intel_mipmap_tree *dst_mt; |
||
153 | switch (buffer_bit) { |
||
154 | case GL_COLOR_BUFFER_BIT: |
||
155 | src_irb = intel_renderbuffer(read_fb->_ColorReadBuffer); |
||
156 | for (unsigned i = 0; i < draw_fb->_NumColorDrawBuffers; ++i) { |
||
157 | dst_irb = intel_renderbuffer(draw_fb->_ColorDrawBuffers[i]); |
||
158 | if (dst_irb) |
||
159 | do_blorp_blit(brw, buffer_bit, |
||
160 | src_irb, src_irb->Base.Base.Format, |
||
161 | dst_irb, dst_irb->Base.Base.Format, |
||
162 | srcX0, srcY0, srcX1, srcY1, |
||
163 | dstX0, dstY0, dstX1, dstY1, |
||
164 | filter, mirror_x, mirror_y); |
||
165 | } |
||
166 | break; |
||
167 | case GL_DEPTH_BUFFER_BIT: |
||
168 | src_irb = |
||
169 | intel_renderbuffer(read_fb->Attachment[BUFFER_DEPTH].Renderbuffer); |
||
170 | dst_irb = |
||
171 | intel_renderbuffer(draw_fb->Attachment[BUFFER_DEPTH].Renderbuffer); |
||
172 | src_mt = find_miptree(buffer_bit, src_irb); |
||
173 | dst_mt = find_miptree(buffer_bit, dst_irb); |
||
174 | |||
175 | /* We can't handle format conversions between Z24 and other formats |
||
176 | * since we have to lie about the surface format. See the comments in |
||
177 | * brw_blorp_surface_info::set(). |
||
178 | */ |
||
179 | if ((src_mt->format == MESA_FORMAT_Z24_UNORM_X8_UINT) != |
||
180 | (dst_mt->format == MESA_FORMAT_Z24_UNORM_X8_UINT)) |
||
181 | return false; |
||
182 | |||
183 | do_blorp_blit(brw, buffer_bit, src_irb, MESA_FORMAT_NONE, |
||
184 | dst_irb, MESA_FORMAT_NONE, srcX0, srcY0, |
||
185 | srcX1, srcY1, dstX0, dstY0, dstX1, dstY1, |
||
186 | filter, mirror_x, mirror_y); |
||
187 | break; |
||
188 | case GL_STENCIL_BUFFER_BIT: |
||
189 | src_irb = |
||
190 | intel_renderbuffer(read_fb->Attachment[BUFFER_STENCIL].Renderbuffer); |
||
191 | dst_irb = |
||
192 | intel_renderbuffer(draw_fb->Attachment[BUFFER_STENCIL].Renderbuffer); |
||
193 | do_blorp_blit(brw, buffer_bit, src_irb, MESA_FORMAT_NONE, |
||
194 | dst_irb, MESA_FORMAT_NONE, srcX0, srcY0, |
||
195 | srcX1, srcY1, dstX0, dstY0, dstX1, dstY1, |
||
196 | filter, mirror_x, mirror_y); |
||
197 | break; |
||
198 | default: |
||
199 | unreachable("not reached"); |
||
200 | } |
||
201 | |||
202 | return true; |
||
203 | } |
||
204 | |||
205 | bool |
||
206 | brw_blorp_copytexsubimage(struct brw_context *brw, |
||
207 | struct gl_renderbuffer *src_rb, |
||
208 | struct gl_texture_image *dst_image, |
||
209 | int slice, |
||
210 | int srcX0, int srcY0, |
||
211 | int dstX0, int dstY0, |
||
212 | int width, int height) |
||
213 | { |
||
214 | struct gl_context *ctx = &brw->ctx; |
||
215 | struct intel_renderbuffer *src_irb = intel_renderbuffer(src_rb); |
||
216 | struct intel_texture_image *intel_image = intel_texture_image(dst_image); |
||
217 | |||
218 | /* Sync up the state of window system buffers. We need to do this before |
||
219 | * we go looking at the src renderbuffer's miptree. |
||
220 | */ |
||
221 | intel_prepare_render(brw); |
||
222 | |||
223 | struct intel_mipmap_tree *src_mt = src_irb->mt; |
||
224 | struct intel_mipmap_tree *dst_mt = intel_image->mt; |
||
225 | |||
226 | /* BLORP is only supported for Gen6-7. */ |
||
227 | if (brw->gen < 6 || brw->gen > 7) |
||
228 | return false; |
||
229 | |||
230 | if (_mesa_get_format_base_format(src_rb->Format) != |
||
231 | _mesa_get_format_base_format(dst_image->TexFormat)) { |
||
232 | return false; |
||
233 | } |
||
234 | |||
235 | /* We can't handle format conversions between Z24 and other formats since |
||
236 | * we have to lie about the surface format. See the comments in |
||
237 | * brw_blorp_surface_info::set(). |
||
238 | */ |
||
239 | if ((src_mt->format == MESA_FORMAT_Z24_UNORM_X8_UINT) != |
||
240 | (dst_mt->format == MESA_FORMAT_Z24_UNORM_X8_UINT)) { |
||
241 | return false; |
||
242 | } |
||
243 | |||
244 | if (!brw->format_supported_as_render_target[dst_image->TexFormat]) |
||
245 | return false; |
||
246 | |||
247 | /* Source clipping shouldn't be necessary, since copytexsubimage (in |
||
248 | * src/mesa/main/teximage.c) calls _mesa_clip_copytexsubimage() which |
||
249 | * takes care of it. |
||
250 | * |
||
251 | * Destination clipping shouldn't be necessary since the restrictions on |
||
252 | * glCopyTexSubImage prevent the user from specifying a destination rectangle |
||
253 | * that falls outside the bounds of the destination texture. |
||
254 | * See error_check_subtexture_dimensions(). |
||
255 | */ |
||
256 | |||
257 | int srcY1 = srcY0 + height; |
||
258 | int srcX1 = srcX0 + width; |
||
259 | int dstX1 = dstX0 + width; |
||
260 | int dstY1 = dstY0 + height; |
||
261 | |||
262 | /* Account for the fact that in the system framebuffer, the origin is at |
||
263 | * the lower left. |
||
264 | */ |
||
265 | bool mirror_y = false; |
||
266 | if (_mesa_is_winsys_fbo(ctx->ReadBuffer)) { |
||
267 | GLint tmp = src_rb->Height - srcY0; |
||
268 | srcY0 = src_rb->Height - srcY1; |
||
269 | srcY1 = tmp; |
||
270 | mirror_y = true; |
||
271 | } |
||
272 | |||
273 | /* Account for face selection and texture view MinLayer */ |
||
274 | int dst_slice = slice + dst_image->TexObject->MinLayer + dst_image->Face; |
||
275 | int dst_level = dst_image->Level + dst_image->TexObject->MinLevel; |
||
276 | |||
277 | brw_blorp_blit_miptrees(brw, |
||
278 | src_mt, src_irb->mt_level, src_irb->mt_layer, |
||
279 | src_rb->Format, |
||
280 | dst_mt, dst_level, dst_slice, |
||
281 | dst_image->TexFormat, |
||
282 | srcX0, srcY0, srcX1, srcY1, |
||
283 | dstX0, dstY0, dstX1, dstY1, |
||
284 | GL_NEAREST, false, mirror_y); |
||
285 | |||
286 | /* If we're copying to a packed depth stencil texture and the source |
||
287 | * framebuffer has separate stencil, we need to also copy the stencil data |
||
288 | * over. |
||
289 | */ |
||
290 | src_rb = ctx->ReadBuffer->Attachment[BUFFER_STENCIL].Renderbuffer; |
||
291 | if (_mesa_get_format_bits(dst_image->TexFormat, GL_STENCIL_BITS) > 0 && |
||
292 | src_rb != NULL) { |
||
293 | src_irb = intel_renderbuffer(src_rb); |
||
294 | src_mt = src_irb->mt; |
||
295 | |||
296 | if (src_mt->stencil_mt) |
||
297 | src_mt = src_mt->stencil_mt; |
||
298 | if (dst_mt->stencil_mt) |
||
299 | dst_mt = dst_mt->stencil_mt; |
||
300 | |||
301 | if (src_mt != dst_mt) { |
||
302 | brw_blorp_blit_miptrees(brw, |
||
303 | src_mt, src_irb->mt_level, src_irb->mt_layer, |
||
304 | src_mt->format, |
||
305 | dst_mt, dst_level, dst_slice, |
||
306 | dst_mt->format, |
||
307 | srcX0, srcY0, srcX1, srcY1, |
||
308 | dstX0, dstY0, dstX1, dstY1, |
||
309 | GL_NEAREST, false, mirror_y); |
||
310 | } |
||
311 | } |
||
312 | |||
313 | return true; |
||
314 | } |
||
315 | |||
316 | |||
317 | GLbitfield |
||
318 | brw_blorp_framebuffer(struct brw_context *brw, |
||
319 | struct gl_framebuffer *readFb, |
||
320 | struct gl_framebuffer *drawFb, |
||
321 | GLint srcX0, GLint srcY0, GLint srcX1, GLint srcY1, |
||
322 | GLint dstX0, GLint dstY0, GLint dstX1, GLint dstY1, |
||
323 | GLbitfield mask, GLenum filter) |
||
324 | { |
||
325 | /* BLORP is not supported before Gen6. */ |
||
326 | if (brw->gen < 6 || brw->gen >= 8) |
||
327 | return mask; |
||
328 | |||
329 | static GLbitfield buffer_bits[] = { |
||
330 | GL_COLOR_BUFFER_BIT, |
||
331 | GL_DEPTH_BUFFER_BIT, |
||
332 | GL_STENCIL_BUFFER_BIT, |
||
333 | }; |
||
334 | |||
335 | for (unsigned int i = 0; i < ARRAY_SIZE(buffer_bits); ++i) { |
||
336 | if ((mask & buffer_bits[i]) && |
||
337 | try_blorp_blit(brw, readFb, drawFb, |
||
338 | srcX0, srcY0, srcX1, srcY1, |
||
339 | dstX0, dstY0, dstX1, dstY1, |
||
340 | filter, buffer_bits[i])) { |
||
341 | mask &= ~buffer_bits[i]; |
||
342 | } |
||
343 | } |
||
344 | |||
345 | return mask; |
||
346 | } |
||
347 | |||
348 | |||
349 | /** |
||
350 | * Enum to specify the order of arguments in a sampler message |
||
351 | */ |
||
352 | enum sampler_message_arg |
||
353 | { |
||
354 | SAMPLER_MESSAGE_ARG_U_FLOAT, |
||
355 | SAMPLER_MESSAGE_ARG_V_FLOAT, |
||
356 | SAMPLER_MESSAGE_ARG_U_INT, |
||
357 | SAMPLER_MESSAGE_ARG_V_INT, |
||
358 | SAMPLER_MESSAGE_ARG_SI_INT, |
||
359 | SAMPLER_MESSAGE_ARG_MCS_INT, |
||
360 | SAMPLER_MESSAGE_ARG_ZERO_INT, |
||
361 | }; |
||
362 | |||
363 | /** |
||
364 | * Generator for WM programs used in BLORP blits. |
||
365 | * |
||
366 | * The bulk of the work done by the WM program is to wrap and unwrap the |
||
367 | * coordinate transformations used by the hardware to store surfaces in |
||
368 | * memory. The hardware transforms a pixel location (X, Y, S) (where S is the |
||
369 | * sample index for a multisampled surface) to a memory offset by the |
||
370 | * following formulas: |
||
371 | * |
||
372 | * offset = tile(tiling_format, encode_msaa(num_samples, layout, X, Y, S)) |
||
373 | * (X, Y, S) = decode_msaa(num_samples, layout, detile(tiling_format, offset)) |
||
374 | * |
||
375 | * For a single-sampled surface, or for a multisampled surface using |
||
376 | * INTEL_MSAA_LAYOUT_UMS, encode_msaa() and decode_msaa are the identity |
||
377 | * function: |
||
378 | * |
||
379 | * encode_msaa(1, NONE, X, Y, 0) = (X, Y, 0) |
||
380 | * decode_msaa(1, NONE, X, Y, 0) = (X, Y, 0) |
||
381 | * encode_msaa(n, UMS, X, Y, S) = (X, Y, S) |
||
382 | * decode_msaa(n, UMS, X, Y, S) = (X, Y, S) |
||
383 | * |
||
384 | * For a 4x multisampled surface using INTEL_MSAA_LAYOUT_IMS, encode_msaa() |
||
385 | * embeds the sample number into bit 1 of the X and Y coordinates: |
||
386 | * |
||
387 | * encode_msaa(4, IMS, X, Y, S) = (X', Y', 0) |
||
388 | * where X' = (X & ~0b1) << 1 | (S & 0b1) << 1 | (X & 0b1) |
||
389 | * Y' = (Y & ~0b1 ) << 1 | (S & 0b10) | (Y & 0b1) |
||
390 | * decode_msaa(4, IMS, X, Y, 0) = (X', Y', S) |
||
391 | * where X' = (X & ~0b11) >> 1 | (X & 0b1) |
||
392 | * Y' = (Y & ~0b11) >> 1 | (Y & 0b1) |
||
393 | * S = (Y & 0b10) | (X & 0b10) >> 1 |
||
394 | * |
||
395 | * For an 8x multisampled surface using INTEL_MSAA_LAYOUT_IMS, encode_msaa() |
||
396 | * embeds the sample number into bits 1 and 2 of the X coordinate and bit 1 of |
||
397 | * the Y coordinate: |
||
398 | * |
||
399 | * encode_msaa(8, IMS, X, Y, S) = (X', Y', 0) |
||
400 | * where X' = (X & ~0b1) << 2 | (S & 0b100) | (S & 0b1) << 1 | (X & 0b1) |
||
401 | * Y' = (Y & ~0b1) << 1 | (S & 0b10) | (Y & 0b1) |
||
402 | * decode_msaa(8, IMS, X, Y, 0) = (X', Y', S) |
||
403 | * where X' = (X & ~0b111) >> 2 | (X & 0b1) |
||
404 | * Y' = (Y & ~0b11) >> 1 | (Y & 0b1) |
||
405 | * S = (X & 0b100) | (Y & 0b10) | (X & 0b10) >> 1 |
||
406 | * |
||
407 | * For X tiling, tile() combines together the low-order bits of the X and Y |
||
408 | * coordinates in the pattern 0byyyxxxxxxxxx, creating 4k tiles that are 512 |
||
409 | * bytes wide and 8 rows high: |
||
410 | * |
||
411 | * tile(x_tiled, X, Y, S) = A |
||
412 | * where A = tile_num << 12 | offset |
||
413 | * tile_num = (Y' >> 3) * tile_pitch + (X' >> 9) |
||
414 | * offset = (Y' & 0b111) << 9 |
||
415 | * | (X & 0b111111111) |
||
416 | * X' = X * cpp |
||
417 | * Y' = Y + S * qpitch |
||
418 | * detile(x_tiled, A) = (X, Y, S) |
||
419 | * where X = X' / cpp |
||
420 | * Y = Y' % qpitch |
||
421 | * S = Y' / qpitch |
||
422 | * Y' = (tile_num / tile_pitch) << 3 |
||
423 | * | (A & 0b111000000000) >> 9 |
||
424 | * X' = (tile_num % tile_pitch) << 9 |
||
425 | * | (A & 0b111111111) |
||
426 | * |
||
427 | * (In all tiling formulas, cpp is the number of bytes occupied by a single |
||
428 | * sample ("chars per pixel"), tile_pitch is the number of 4k tiles required |
||
429 | * to fill the width of the surface, and qpitch is the spacing (in rows) |
||
430 | * between array slices). |
||
431 | * |
||
432 | * For Y tiling, tile() combines together the low-order bits of the X and Y |
||
433 | * coordinates in the pattern 0bxxxyyyyyxxxx, creating 4k tiles that are 128 |
||
434 | * bytes wide and 32 rows high: |
||
435 | * |
||
436 | * tile(y_tiled, X, Y, S) = A |
||
437 | * where A = tile_num << 12 | offset |
||
438 | * tile_num = (Y' >> 5) * tile_pitch + (X' >> 7) |
||
439 | * offset = (X' & 0b1110000) << 5 |
||
440 | * | (Y' & 0b11111) << 4 |
||
441 | * | (X' & 0b1111) |
||
442 | * X' = X * cpp |
||
443 | * Y' = Y + S * qpitch |
||
444 | * detile(y_tiled, A) = (X, Y, S) |
||
445 | * where X = X' / cpp |
||
446 | * Y = Y' % qpitch |
||
447 | * S = Y' / qpitch |
||
448 | * Y' = (tile_num / tile_pitch) << 5 |
||
449 | * | (A & 0b111110000) >> 4 |
||
450 | * X' = (tile_num % tile_pitch) << 7 |
||
451 | * | (A & 0b111000000000) >> 5 |
||
452 | * | (A & 0b1111) |
||
453 | * |
||
454 | * For W tiling, tile() combines together the low-order bits of the X and Y |
||
455 | * coordinates in the pattern 0bxxxyyyyxyxyx, creating 4k tiles that are 64 |
||
456 | * bytes wide and 64 rows high (note that W tiling is only used for stencil |
||
457 | * buffers, which always have cpp = 1 and S=0): |
||
458 | * |
||
459 | * tile(w_tiled, X, Y, S) = A |
||
460 | * where A = tile_num << 12 | offset |
||
461 | * tile_num = (Y' >> 6) * tile_pitch + (X' >> 6) |
||
462 | * offset = (X' & 0b111000) << 6 |
||
463 | * | (Y' & 0b111100) << 3 |
||
464 | * | (X' & 0b100) << 2 |
||
465 | * | (Y' & 0b10) << 2 |
||
466 | * | (X' & 0b10) << 1 |
||
467 | * | (Y' & 0b1) << 1 |
||
468 | * | (X' & 0b1) |
||
469 | * X' = X * cpp = X |
||
470 | * Y' = Y + S * qpitch |
||
471 | * detile(w_tiled, A) = (X, Y, S) |
||
472 | * where X = X' / cpp = X' |
||
473 | * Y = Y' % qpitch = Y' |
||
474 | * S = Y / qpitch = 0 |
||
475 | * Y' = (tile_num / tile_pitch) << 6 |
||
476 | * | (A & 0b111100000) >> 3 |
||
477 | * | (A & 0b1000) >> 2 |
||
478 | * | (A & 0b10) >> 1 |
||
479 | * X' = (tile_num % tile_pitch) << 6 |
||
480 | * | (A & 0b111000000000) >> 6 |
||
481 | * | (A & 0b10000) >> 2 |
||
482 | * | (A & 0b100) >> 1 |
||
483 | * | (A & 0b1) |
||
484 | * |
||
485 | * Finally, for a non-tiled surface, tile() simply combines together the X and |
||
486 | * Y coordinates in the natural way: |
||
487 | * |
||
488 | * tile(untiled, X, Y, S) = A |
||
489 | * where A = Y * pitch + X' |
||
490 | * X' = X * cpp |
||
491 | * Y' = Y + S * qpitch |
||
492 | * detile(untiled, A) = (X, Y, S) |
||
493 | * where X = X' / cpp |
||
494 | * Y = Y' % qpitch |
||
495 | * S = Y' / qpitch |
||
496 | * X' = A % pitch |
||
497 | * Y' = A / pitch |
||
498 | * |
||
499 | * (In these formulas, pitch is the number of bytes occupied by a single row |
||
500 | * of samples). |
||
501 | */ |
||
502 | class brw_blorp_blit_program : public brw_blorp_eu_emitter |
||
503 | { |
||
504 | public: |
||
505 | brw_blorp_blit_program(struct brw_context *brw, |
||
506 | const brw_blorp_blit_prog_key *key, bool debug_flag); |
||
507 | |||
508 | const GLuint *compile(struct brw_context *brw, GLuint *program_size); |
||
509 | |||
510 | brw_blorp_prog_data prog_data; |
||
511 | |||
512 | private: |
||
513 | void alloc_regs(); |
||
514 | void alloc_push_const_regs(int base_reg); |
||
515 | void compute_frag_coords(); |
||
516 | void translate_tiling(bool old_tiled_w, bool new_tiled_w); |
||
517 | void encode_msaa(unsigned num_samples, intel_msaa_layout layout); |
||
518 | void decode_msaa(unsigned num_samples, intel_msaa_layout layout); |
||
519 | void translate_dst_to_src(); |
||
520 | void clamp_tex_coords(struct brw_reg regX, struct brw_reg regY, |
||
521 | struct brw_reg clampX0, struct brw_reg clampY0, |
||
522 | struct brw_reg clampX1, struct brw_reg clampY1); |
||
523 | void single_to_blend(); |
||
524 | void manual_blend_average(unsigned num_samples); |
||
525 | void manual_blend_bilinear(unsigned num_samples); |
||
526 | void sample(struct brw_reg dst); |
||
527 | void texel_fetch(struct brw_reg dst); |
||
528 | void mcs_fetch(); |
||
529 | void texture_lookup(struct brw_reg dst, enum opcode op, |
||
530 | const sampler_message_arg *args, int num_args); |
||
531 | void render_target_write(); |
||
532 | |||
533 | /** |
||
534 | * Base-2 logarithm of the maximum number of samples that can be blended. |
||
535 | */ |
||
536 | static const unsigned LOG2_MAX_BLEND_SAMPLES = 3; |
||
537 | |||
538 | struct brw_context *brw; |
||
539 | const brw_blorp_blit_prog_key *key; |
||
540 | |||
541 | /* Thread dispatch header */ |
||
542 | struct brw_reg R0; |
||
543 | |||
544 | /* Pixel X/Y coordinates (always in R1). */ |
||
545 | struct brw_reg R1; |
||
546 | |||
547 | /* Push constants */ |
||
548 | struct brw_reg dst_x0; |
||
549 | struct brw_reg dst_x1; |
||
550 | struct brw_reg dst_y0; |
||
551 | struct brw_reg dst_y1; |
||
552 | /* Top right coordinates of the rectangular grid used for scaled blitting */ |
||
553 | struct brw_reg rect_grid_x1; |
||
554 | struct brw_reg rect_grid_y1; |
||
555 | struct { |
||
556 | struct brw_reg multiplier; |
||
557 | struct brw_reg offset; |
||
558 | } x_transform, y_transform; |
||
559 | |||
560 | /* Data read from texture (4 vec16's per array element) */ |
||
561 | struct brw_reg texture_data[LOG2_MAX_BLEND_SAMPLES + 1]; |
||
562 | |||
563 | /* Auxiliary storage for the contents of the MCS surface. |
||
564 | * |
||
565 | * Since the sampler always returns 8 registers worth of data, this is 8 |
||
566 | * registers wide, even though we only use the first 2 registers of it. |
||
567 | */ |
||
568 | struct brw_reg mcs_data; |
||
569 | |||
570 | /* X coordinates. We have two of them so that we can perform coordinate |
||
571 | * transformations easily. |
||
572 | */ |
||
573 | struct brw_reg x_coords[2]; |
||
574 | |||
575 | /* Y coordinates. We have two of them so that we can perform coordinate |
||
576 | * transformations easily. |
||
577 | */ |
||
578 | struct brw_reg y_coords[2]; |
||
579 | |||
580 | /* X, Y coordinates of the pixel from which we need to fetch the specific |
||
581 | * sample. These are used for multisample scaled blitting. |
||
582 | */ |
||
583 | struct brw_reg x_sample_coords; |
||
584 | struct brw_reg y_sample_coords; |
||
585 | |||
586 | /* Fractional parts of the x and y coordinates, used as bilinear interpolation coefficients */ |
||
587 | struct brw_reg x_frac; |
||
588 | struct brw_reg y_frac; |
||
589 | |||
590 | /* Which element of x_coords and y_coords is currently in use. |
||
591 | */ |
||
592 | int xy_coord_index; |
||
593 | |||
594 | /* True if, at the point in the program currently being compiled, the |
||
595 | * sample index is known to be zero. |
||
596 | */ |
||
597 | bool s_is_zero; |
||
598 | |||
599 | /* Register storing the sample index when s_is_zero is false. */ |
||
600 | struct brw_reg sample_index; |
||
601 | |||
602 | /* Temporaries */ |
||
603 | struct brw_reg t1; |
||
604 | struct brw_reg t2; |
||
605 | |||
606 | /* MRF used for sampling and render target writes */ |
||
607 | GLuint base_mrf; |
||
608 | }; |
||
609 | |||
610 | brw_blorp_blit_program::brw_blorp_blit_program( |
||
611 | struct brw_context *brw, |
||
612 | const brw_blorp_blit_prog_key *key, |
||
613 | bool debug_flag) |
||
614 | : brw_blorp_eu_emitter(brw, debug_flag), |
||
615 | brw(brw), |
||
616 | key(key) |
||
617 | { |
||
618 | } |
||
619 | |||
620 | const GLuint * |
||
621 | brw_blorp_blit_program::compile(struct brw_context *brw, |
||
622 | GLuint *program_size) |
||
623 | { |
||
624 | /* Sanity checks */ |
||
625 | if (key->dst_tiled_w && key->rt_samples > 0) { |
||
626 | /* If the destination image is W tiled and multisampled, then the thread |
||
627 | * must be dispatched once per sample, not once per pixel. This is |
||
628 | * necessary because after conversion between W and Y tiling, there's no |
||
629 | * guarantee that all samples corresponding to a single pixel will still |
||
630 | * be together. |
||
631 | */ |
||
632 | assert(key->persample_msaa_dispatch); |
||
633 | } |
||
634 | |||
635 | if (key->blend) { |
||
636 | /* We are blending, which means we won't have an opportunity to |
||
637 | * translate the tiling and sample count for the texture surface. So |
||
638 | * the surface state for the texture must be configured with the correct |
||
639 | * tiling and sample count. |
||
640 | */ |
||
641 | assert(!key->src_tiled_w); |
||
642 | assert(key->tex_samples == key->src_samples); |
||
643 | assert(key->tex_layout == key->src_layout); |
||
644 | assert(key->tex_samples > 0); |
||
645 | } |
||
646 | |||
647 | if (key->persample_msaa_dispatch) { |
||
648 | /* It only makes sense to do persample dispatch if the render target is |
||
649 | * configured as multisampled. |
||
650 | */ |
||
651 | assert(key->rt_samples > 0); |
||
652 | } |
||
653 | |||
654 | /* Make sure layout is consistent with sample count */ |
||
655 | assert((key->tex_layout == INTEL_MSAA_LAYOUT_NONE) == |
||
656 | (key->tex_samples == 0)); |
||
657 | assert((key->rt_layout == INTEL_MSAA_LAYOUT_NONE) == |
||
658 | (key->rt_samples == 0)); |
||
659 | assert((key->src_layout == INTEL_MSAA_LAYOUT_NONE) == |
||
660 | (key->src_samples == 0)); |
||
661 | assert((key->dst_layout == INTEL_MSAA_LAYOUT_NONE) == |
||
662 | (key->dst_samples == 0)); |
||
663 | |||
664 | /* Set up prog_data */ |
||
665 | memset(&prog_data, 0, sizeof(prog_data)); |
||
666 | prog_data.persample_msaa_dispatch = key->persample_msaa_dispatch; |
||
667 | |||
668 | alloc_regs(); |
||
669 | compute_frag_coords(); |
||
670 | |||
671 | /* Render target and texture hardware don't support W tiling. */ |
||
672 | const bool rt_tiled_w = false; |
||
673 | const bool tex_tiled_w = false; |
||
674 | |||
675 | /* The address that data will be written to is determined by the |
||
676 | * coordinates supplied to the WM thread and the tiling and sample count of |
||
677 | * the render target, according to the formula: |
||
678 | * |
||
679 | * (X, Y, S) = decode_msaa(rt_samples, detile(rt_tiling, offset)) |
||
680 | * |
||
681 | * If the actual tiling and sample count of the destination surface are not |
||
682 | * the same as the configuration of the render target, then these |
||
683 | * coordinates are wrong and we have to adjust them to compensate for the |
||
684 | * difference. |
||
685 | */ |
||
686 | if (rt_tiled_w != key->dst_tiled_w || |
||
687 | key->rt_samples != key->dst_samples || |
||
688 | key->rt_layout != key->dst_layout) { |
||
689 | encode_msaa(key->rt_samples, key->rt_layout); |
||
690 | /* Now (X, Y, S) = detile(rt_tiling, offset) */ |
||
691 | translate_tiling(rt_tiled_w, key->dst_tiled_w); |
||
692 | /* Now (X, Y, S) = detile(dst_tiling, offset) */ |
||
693 | decode_msaa(key->dst_samples, key->dst_layout); |
||
694 | } |
||
695 | |||
696 | /* Now (X, Y, S) = decode_msaa(dst_samples, detile(dst_tiling, offset)). |
||
697 | * |
||
698 | * That is: X, Y and S now contain the true coordinates and sample index of |
||
699 | * the data that the WM thread should output. |
||
700 | * |
||
701 | * If we need to kill pixels that are outside the destination rectangle, |
||
702 | * now is the time to do it. |
||
703 | */ |
||
704 | |||
705 | if (key->use_kill) |
||
706 | emit_kill_if_outside_rect(x_coords[xy_coord_index], |
||
707 | y_coords[xy_coord_index], |
||
708 | dst_x0, dst_x1, dst_y0, dst_y1); |
||
709 | |||
710 | /* Next, apply a translation to obtain coordinates in the source image. */ |
||
711 | translate_dst_to_src(); |
||
712 | |||
713 | /* If the source image is not multisampled, then we want to fetch sample |
||
714 | * number 0, because that's the only sample there is. |
||
715 | */ |
||
716 | if (key->src_samples == 0) |
||
717 | s_is_zero = true; |
||
718 | |||
719 | /* X, Y, and S are now the coordinates of the pixel in the source image |
||
720 | * that we want to texture from. Exception: if we are blending, then S is |
||
721 | * irrelevant, because we are going to fetch all samples. |
||
722 | */ |
||
723 | if (key->blend && !key->blit_scaled) { |
||
724 | if (brw->gen == 6) { |
||
725 | /* Gen6 hardware an automatically blend using the SAMPLE message */ |
||
726 | single_to_blend(); |
||
727 | sample(texture_data[0]); |
||
728 | } else { |
||
729 | /* Gen7+ hardware doesn't automaticaly blend. */ |
||
730 | manual_blend_average(key->src_samples); |
||
731 | } |
||
732 | } else if(key->blend && key->blit_scaled) { |
||
733 | manual_blend_bilinear(key->src_samples); |
||
734 | } else { |
||
735 | /* We aren't blending, which means we just want to fetch a single sample |
||
736 | * from the source surface. The address that we want to fetch from is |
||
737 | * related to the X, Y and S values according to the formula: |
||
738 | * |
||
739 | * (X, Y, S) = decode_msaa(src_samples, detile(src_tiling, offset)). |
||
740 | * |
||
741 | * If the actual tiling and sample count of the source surface are not |
||
742 | * the same as the configuration of the texture, then we need to adjust |
||
743 | * the coordinates to compensate for the difference. |
||
744 | */ |
||
745 | if ((tex_tiled_w != key->src_tiled_w || |
||
746 | key->tex_samples != key->src_samples || |
||
747 | key->tex_layout != key->src_layout) && |
||
748 | !key->bilinear_filter) { |
||
749 | encode_msaa(key->src_samples, key->src_layout); |
||
750 | /* Now (X, Y, S) = detile(src_tiling, offset) */ |
||
751 | translate_tiling(key->src_tiled_w, tex_tiled_w); |
||
752 | /* Now (X, Y, S) = detile(tex_tiling, offset) */ |
||
753 | decode_msaa(key->tex_samples, key->tex_layout); |
||
754 | } |
||
755 | |||
756 | if (key->bilinear_filter) { |
||
757 | sample(texture_data[0]); |
||
758 | } |
||
759 | else { |
||
760 | /* Now (X, Y, S) = decode_msaa(tex_samples, detile(tex_tiling, offset)). |
||
761 | * |
||
762 | * In other words: X, Y, and S now contain values which, when passed to |
||
763 | * the texturing unit, will cause data to be read from the correct |
||
764 | * memory location. So we can fetch the texel now. |
||
765 | */ |
||
766 | if (key->tex_layout == INTEL_MSAA_LAYOUT_CMS) |
||
767 | mcs_fetch(); |
||
768 | texel_fetch(texture_data[0]); |
||
769 | } |
||
770 | } |
||
771 | |||
772 | /* Finally, write the fetched (or blended) value to the render target and |
||
773 | * terminate the thread. |
||
774 | */ |
||
775 | render_target_write(); |
||
776 | |||
777 | return get_program(program_size); |
||
778 | } |
||
779 | |||
780 | void |
||
781 | brw_blorp_blit_program::alloc_push_const_regs(int base_reg) |
||
782 | { |
||
783 | #define CONST_LOC(name) offsetof(brw_blorp_wm_push_constants, name) |
||
784 | #define ALLOC_REG(name, type) \ |
||
785 | this->name = \ |
||
786 | retype(brw_vec1_reg(BRW_GENERAL_REGISTER_FILE, \ |
||
787 | base_reg + CONST_LOC(name) / 32, \ |
||
788 | (CONST_LOC(name) % 32) / 4), type) |
||
789 | |||
790 | ALLOC_REG(dst_x0, BRW_REGISTER_TYPE_UD); |
||
791 | ALLOC_REG(dst_x1, BRW_REGISTER_TYPE_UD); |
||
792 | ALLOC_REG(dst_y0, BRW_REGISTER_TYPE_UD); |
||
793 | ALLOC_REG(dst_y1, BRW_REGISTER_TYPE_UD); |
||
794 | ALLOC_REG(rect_grid_x1, BRW_REGISTER_TYPE_F); |
||
795 | ALLOC_REG(rect_grid_y1, BRW_REGISTER_TYPE_F); |
||
796 | ALLOC_REG(x_transform.multiplier, BRW_REGISTER_TYPE_F); |
||
797 | ALLOC_REG(x_transform.offset, BRW_REGISTER_TYPE_F); |
||
798 | ALLOC_REG(y_transform.multiplier, BRW_REGISTER_TYPE_F); |
||
799 | ALLOC_REG(y_transform.offset, BRW_REGISTER_TYPE_F); |
||
800 | #undef CONST_LOC |
||
801 | #undef ALLOC_REG |
||
802 | } |
||
803 | |||
804 | void |
||
805 | brw_blorp_blit_program::alloc_regs() |
||
806 | { |
||
807 | int reg = 0; |
||
808 | this->R0 = retype(brw_vec8_grf(reg++, 0), BRW_REGISTER_TYPE_UW); |
||
809 | this->R1 = retype(brw_vec8_grf(reg++, 0), BRW_REGISTER_TYPE_UW); |
||
810 | prog_data.first_curbe_grf = reg; |
||
811 | alloc_push_const_regs(reg); |
||
812 | reg += BRW_BLORP_NUM_PUSH_CONST_REGS; |
||
813 | for (unsigned i = 0; i < ARRAY_SIZE(texture_data); ++i) { |
||
814 | this->texture_data[i] = |
||
815 | retype(vec16(brw_vec8_grf(reg, 0)), key->texture_data_type); |
||
816 | reg += 8; |
||
817 | } |
||
818 | this->mcs_data = |
||
819 | retype(brw_vec8_grf(reg, 0), BRW_REGISTER_TYPE_UD); reg += 8; |
||
820 | |||
821 | for (int i = 0; i < 2; ++i) { |
||
822 | this->x_coords[i] |
||
823 | = retype(brw_vec8_grf(reg, 0), BRW_REGISTER_TYPE_UD); |
||
824 | reg += 2; |
||
825 | this->y_coords[i] |
||
826 | = retype(brw_vec8_grf(reg, 0), BRW_REGISTER_TYPE_UD); |
||
827 | reg += 2; |
||
828 | } |
||
829 | |||
830 | if (key->blit_scaled && key->blend) { |
||
831 | this->x_sample_coords = brw_vec8_grf(reg, 0); |
||
832 | reg += 2; |
||
833 | this->y_sample_coords = brw_vec8_grf(reg, 0); |
||
834 | reg += 2; |
||
835 | this->x_frac = brw_vec8_grf(reg, 0); |
||
836 | reg += 2; |
||
837 | this->y_frac = brw_vec8_grf(reg, 0); |
||
838 | reg += 2; |
||
839 | } |
||
840 | |||
841 | this->xy_coord_index = 0; |
||
842 | this->sample_index |
||
843 | = retype(brw_vec8_grf(reg, 0), BRW_REGISTER_TYPE_UD); |
||
844 | reg += 2; |
||
845 | this->t1 = retype(brw_vec8_grf(reg, 0), BRW_REGISTER_TYPE_UD); |
||
846 | reg += 2; |
||
847 | this->t2 = retype(brw_vec8_grf(reg, 0), BRW_REGISTER_TYPE_UD); |
||
848 | reg += 2; |
||
849 | |||
850 | /* Make sure we didn't run out of registers */ |
||
851 | assert(reg <= GEN7_MRF_HACK_START); |
||
852 | |||
853 | int mrf = 2; |
||
854 | this->base_mrf = mrf; |
||
855 | } |
||
856 | |||
857 | /* In the code that follows, X and Y can be used to quickly refer to the |
||
858 | * active elements of x_coords and y_coords, and Xp and Yp ("X prime" and "Y |
||
859 | * prime") to the inactive elements. |
||
860 | * |
||
861 | * S can be used to quickly refer to sample_index. |
||
862 | */ |
||
863 | #define X x_coords[xy_coord_index] |
||
864 | #define Y y_coords[xy_coord_index] |
||
865 | #define Xp x_coords[!xy_coord_index] |
||
866 | #define Yp y_coords[!xy_coord_index] |
||
867 | #define S sample_index |
||
868 | |||
869 | /* Quickly swap the roles of (X, Y) and (Xp, Yp). Saves us from having to do |
||
870 | * MOVs to transfor (Xp, Yp) to (X, Y) after a coordinate transformation. |
||
871 | */ |
||
872 | #define SWAP_XY_AND_XPYP() xy_coord_index = !xy_coord_index; |
||
873 | |||
874 | /** |
||
875 | * Emit code to compute the X and Y coordinates of the pixels being rendered |
||
876 | * by this WM invocation. |
||
877 | * |
||
878 | * Assuming the render target is set up for Y tiling, these (X, Y) values are |
||
879 | * related to the address offset where outputs will be written by the formula: |
||
880 | * |
||
881 | * (X, Y, S) = decode_msaa(detile(offset)). |
||
882 | * |
||
883 | * (See brw_blorp_blit_program). |
||
884 | */ |
||
885 | void |
||
886 | brw_blorp_blit_program::compute_frag_coords() |
||
887 | { |
||
888 | /* R1.2[15:0] = X coordinate of upper left pixel of subspan 0 (pixel 0) |
||
889 | * R1.3[15:0] = X coordinate of upper left pixel of subspan 1 (pixel 4) |
||
890 | * R1.4[15:0] = X coordinate of upper left pixel of subspan 2 (pixel 8) |
||
891 | * R1.5[15:0] = X coordinate of upper left pixel of subspan 3 (pixel 12) |
||
892 | * |
||
893 | * Pixels within a subspan are laid out in this arrangement: |
||
894 | * 0 1 |
||
895 | * 2 3 |
||
896 | * |
||
897 | * So, to compute the coordinates of each pixel, we need to read every 2nd |
||
898 | * 16-bit value (vstride=2) from R1, starting at the 4th 16-bit value |
||
899 | * (suboffset=4), and duplicate each value 4 times (hstride=0, width=4). |
||
900 | * In other words, the data we want to access is R1.4<2;4,0>UW. |
||
901 | * |
||
902 | * Then, we need to add the repeating sequence (0, 1, 0, 1, ...) to the |
||
903 | * result, since pixels n+1 and n+3 are in the right half of the subspan. |
||
904 | */ |
||
905 | emit_add(vec16(retype(X, BRW_REGISTER_TYPE_UW)), |
||
906 | stride(suboffset(R1, 4), 2, 4, 0), brw_imm_v(0x10101010)); |
||
907 | |||
908 | /* Similarly, Y coordinates for subspans come from R1.2[31:16] through |
||
909 | * R1.5[31:16], so to get pixel Y coordinates we need to start at the 5th |
||
910 | * 16-bit value instead of the 4th (R1.5<2;4,0>UW instead of |
||
911 | * R1.4<2;4,0>UW). |
||
912 | * |
||
913 | * And we need to add the repeating sequence (0, 0, 1, 1, ...), since |
||
914 | * pixels n+2 and n+3 are in the bottom half of the subspan. |
||
915 | */ |
||
916 | emit_add(vec16(retype(Y, BRW_REGISTER_TYPE_UW)), |
||
917 | stride(suboffset(R1, 5), 2, 4, 0), brw_imm_v(0x11001100)); |
||
918 | |||
919 | /* Move the coordinates to UD registers. */ |
||
920 | emit_mov(vec16(Xp), retype(X, BRW_REGISTER_TYPE_UW)); |
||
921 | emit_mov(vec16(Yp), retype(Y, BRW_REGISTER_TYPE_UW)); |
||
922 | SWAP_XY_AND_XPYP(); |
||
923 | |||
924 | if (key->persample_msaa_dispatch) { |
||
925 | switch (key->rt_samples) { |
||
926 | case 4: { |
||
927 | /* The WM will be run in MSDISPMODE_PERSAMPLE with num_samples == 4. |
||
928 | * Therefore, subspan 0 will represent sample 0, subspan 1 will |
||
929 | * represent sample 1, and so on. |
||
930 | * |
||
931 | * So we need to populate S with the sequence (0, 0, 0, 0, 1, 1, 1, |
||
932 | * 1, 2, 2, 2, 2, 3, 3, 3, 3). The easiest way to do this is to |
||
933 | * populate a temporary variable with the sequence (0, 1, 2, 3), and |
||
934 | * then copy from it using vstride=1, width=4, hstride=0. |
||
935 | */ |
||
936 | struct brw_reg t1_uw1 = retype(t1, BRW_REGISTER_TYPE_UW); |
||
937 | emit_mov(vec16(t1_uw1), brw_imm_v(0x3210)); |
||
938 | /* Move to UD sample_index register. */ |
||
939 | emit_mov_8(S, stride(t1_uw1, 1, 4, 0)); |
||
940 | emit_mov_8(offset(S, 1), suboffset(stride(t1_uw1, 1, 4, 0), 2)); |
||
941 | break; |
||
942 | } |
||
943 | case 8: { |
||
944 | /* The WM will be run in MSDISPMODE_PERSAMPLE with num_samples == 8. |
||
945 | * Therefore, subspan 0 will represent sample N (where N is 0 or 4), |
||
946 | * subspan 1 will represent sample 1, and so on. We can find the |
||
947 | * value of N by looking at R0.0 bits 7:6 ("Starting Sample Pair |
||
948 | * Index") and multiplying by two (since samples are always delivered |
||
949 | * in pairs). That is, we compute 2*((R0.0 & 0xc0) >> 6) == (R0.0 & |
||
950 | * 0xc0) >> 5. |
||
951 | * |
||
952 | * Then we need to add N to the sequence (0, 0, 0, 0, 1, 1, 1, 1, 2, |
||
953 | * 2, 2, 2, 3, 3, 3, 3), which we compute by populating a temporary |
||
954 | * variable with the sequence (0, 1, 2, 3), and then reading from it |
||
955 | * using vstride=1, width=4, hstride=0. |
||
956 | */ |
||
957 | struct brw_reg t1_ud1 = vec1(retype(t1, BRW_REGISTER_TYPE_UD)); |
||
958 | struct brw_reg t2_uw1 = retype(t2, BRW_REGISTER_TYPE_UW); |
||
959 | struct brw_reg r0_ud1 = vec1(retype(R0, BRW_REGISTER_TYPE_UD)); |
||
960 | emit_and(t1_ud1, r0_ud1, brw_imm_ud(0xc0)); |
||
961 | emit_shr(t1_ud1, t1_ud1, brw_imm_ud(5)); |
||
962 | emit_mov(vec16(t2_uw1), brw_imm_v(0x3210)); |
||
963 | emit_add(vec16(S), retype(t1_ud1, BRW_REGISTER_TYPE_UW), |
||
964 | stride(t2_uw1, 1, 4, 0)); |
||
965 | emit_add_8(offset(S, 1), |
||
966 | retype(t1_ud1, BRW_REGISTER_TYPE_UW), |
||
967 | suboffset(stride(t2_uw1, 1, 4, 0), 2)); |
||
968 | break; |
||
969 | } |
||
970 | default: |
||
971 | unreachable("Unrecognized sample count in " |
||
972 | "brw_blorp_blit_program::compute_frag_coords()"); |
||
973 | } |
||
974 | s_is_zero = false; |
||
975 | } else { |
||
976 | /* Either the destination surface is single-sampled, or the WM will be |
||
977 | * run in MSDISPMODE_PERPIXEL (which causes a single fragment dispatch |
||
978 | * per pixel). In either case, it's not meaningful to compute a sample |
||
979 | * value. Just set it to 0. |
||
980 | */ |
||
981 | s_is_zero = true; |
||
982 | } |
||
983 | } |
||
984 | |||
985 | /** |
||
986 | * Emit code to compensate for the difference between Y and W tiling. |
||
987 | * |
||
988 | * This code modifies the X and Y coordinates according to the formula: |
||
989 | * |
||
990 | * (X', Y', S') = detile(new_tiling, tile(old_tiling, X, Y, S)) |
||
991 | * |
||
992 | * (See brw_blorp_blit_program). |
||
993 | * |
||
994 | * It can only translate between W and Y tiling, so new_tiling and old_tiling |
||
995 | * are booleans where true represents W tiling and false represents Y tiling. |
||
996 | */ |
||
997 | void |
||
998 | brw_blorp_blit_program::translate_tiling(bool old_tiled_w, bool new_tiled_w) |
||
999 | { |
||
1000 | if (old_tiled_w == new_tiled_w) |
||
1001 | return; |
||
1002 | |||
1003 | /* In the code that follows, we can safely assume that S = 0, because W |
||
1004 | * tiling formats always use IMS layout. |
||
1005 | */ |
||
1006 | assert(s_is_zero); |
||
1007 | |||
1008 | if (new_tiled_w) { |
||
1009 | /* Given X and Y coordinates that describe an address using Y tiling, |
||
1010 | * translate to the X and Y coordinates that describe the same address |
||
1011 | * using W tiling. |
||
1012 | * |
||
1013 | * If we break down the low order bits of X and Y, using a |
||
1014 | * single letter to represent each low-order bit: |
||
1015 | * |
||
1016 | * X = A << 7 | 0bBCDEFGH |
||
1017 | * Y = J << 5 | 0bKLMNP (1) |
||
1018 | * |
||
1019 | * Then we can apply the Y tiling formula to see the memory offset being |
||
1020 | * addressed: |
||
1021 | * |
||
1022 | * offset = (J * tile_pitch + A) << 12 | 0bBCDKLMNPEFGH (2) |
||
1023 | * |
||
1024 | * If we apply the W detiling formula to this memory location, that the |
||
1025 | * corresponding X' and Y' coordinates are: |
||
1026 | * |
||
1027 | * X' = A << 6 | 0bBCDPFH (3) |
||
1028 | * Y' = J << 6 | 0bKLMNEG |
||
1029 | * |
||
1030 | * Combining (1) and (3), we see that to transform (X, Y) to (X', Y'), |
||
1031 | * we need to make the following computation: |
||
1032 | * |
||
1033 | * X' = (X & ~0b1011) >> 1 | (Y & 0b1) << 2 | X & 0b1 (4) |
||
1034 | * Y' = (Y & ~0b1) << 1 | (X & 0b1000) >> 2 | (X & 0b10) >> 1 |
||
1035 | */ |
||
1036 | emit_and(t1, X, brw_imm_uw(0xfff4)); /* X & ~0b1011 */ |
||
1037 | emit_shr(t1, t1, brw_imm_uw(1)); /* (X & ~0b1011) >> 1 */ |
||
1038 | emit_and(t2, Y, brw_imm_uw(1)); /* Y & 0b1 */ |
||
1039 | emit_shl(t2, t2, brw_imm_uw(2)); /* (Y & 0b1) << 2 */ |
||
1040 | emit_or(t1, t1, t2); /* (X & ~0b1011) >> 1 | (Y & 0b1) << 2 */ |
||
1041 | emit_and(t2, X, brw_imm_uw(1)); /* X & 0b1 */ |
||
1042 | emit_or(Xp, t1, t2); |
||
1043 | emit_and(t1, Y, brw_imm_uw(0xfffe)); /* Y & ~0b1 */ |
||
1044 | emit_shl(t1, t1, brw_imm_uw(1)); /* (Y & ~0b1) << 1 */ |
||
1045 | emit_and(t2, X, brw_imm_uw(8)); /* X & 0b1000 */ |
||
1046 | emit_shr(t2, t2, brw_imm_uw(2)); /* (X & 0b1000) >> 2 */ |
||
1047 | emit_or(t1, t1, t2); /* (Y & ~0b1) << 1 | (X & 0b1000) >> 2 */ |
||
1048 | emit_and(t2, X, brw_imm_uw(2)); /* X & 0b10 */ |
||
1049 | emit_shr(t2, t2, brw_imm_uw(1)); /* (X & 0b10) >> 1 */ |
||
1050 | emit_or(Yp, t1, t2); |
||
1051 | SWAP_XY_AND_XPYP(); |
||
1052 | } else { |
||
1053 | /* Applying the same logic as above, but in reverse, we obtain the |
||
1054 | * formulas: |
||
1055 | * |
||
1056 | * X' = (X & ~0b101) << 1 | (Y & 0b10) << 2 | (Y & 0b1) << 1 | X & 0b1 |
||
1057 | * Y' = (Y & ~0b11) >> 1 | (X & 0b100) >> 2 |
||
1058 | */ |
||
1059 | emit_and(t1, X, brw_imm_uw(0xfffa)); /* X & ~0b101 */ |
||
1060 | emit_shl(t1, t1, brw_imm_uw(1)); /* (X & ~0b101) << 1 */ |
||
1061 | emit_and(t2, Y, brw_imm_uw(2)); /* Y & 0b10 */ |
||
1062 | emit_shl(t2, t2, brw_imm_uw(2)); /* (Y & 0b10) << 2 */ |
||
1063 | emit_or(t1, t1, t2); /* (X & ~0b101) << 1 | (Y & 0b10) << 2 */ |
||
1064 | emit_and(t2, Y, brw_imm_uw(1)); /* Y & 0b1 */ |
||
1065 | emit_shl(t2, t2, brw_imm_uw(1)); /* (Y & 0b1) << 1 */ |
||
1066 | emit_or(t1, t1, t2); /* (X & ~0b101) << 1 | (Y & 0b10) << 2 |
||
1067 | | (Y & 0b1) << 1 */ |
||
1068 | emit_and(t2, X, brw_imm_uw(1)); /* X & 0b1 */ |
||
1069 | emit_or(Xp, t1, t2); |
||
1070 | emit_and(t1, Y, brw_imm_uw(0xfffc)); /* Y & ~0b11 */ |
||
1071 | emit_shr(t1, t1, brw_imm_uw(1)); /* (Y & ~0b11) >> 1 */ |
||
1072 | emit_and(t2, X, brw_imm_uw(4)); /* X & 0b100 */ |
||
1073 | emit_shr(t2, t2, brw_imm_uw(2)); /* (X & 0b100) >> 2 */ |
||
1074 | emit_or(Yp, t1, t2); |
||
1075 | SWAP_XY_AND_XPYP(); |
||
1076 | } |
||
1077 | } |
||
1078 | |||
1079 | /** |
||
1080 | * Emit code to compensate for the difference between MSAA and non-MSAA |
||
1081 | * surfaces. |
||
1082 | * |
||
1083 | * This code modifies the X and Y coordinates according to the formula: |
||
1084 | * |
||
1085 | * (X', Y', S') = encode_msaa(num_samples, IMS, X, Y, S) |
||
1086 | * |
||
1087 | * (See brw_blorp_blit_program). |
||
1088 | */ |
||
1089 | void |
||
1090 | brw_blorp_blit_program::encode_msaa(unsigned num_samples, |
||
1091 | intel_msaa_layout layout) |
||
1092 | { |
||
1093 | switch (layout) { |
||
1094 | case INTEL_MSAA_LAYOUT_NONE: |
||
1095 | /* No translation necessary, and S should already be zero. */ |
||
1096 | assert(s_is_zero); |
||
1097 | break; |
||
1098 | case INTEL_MSAA_LAYOUT_CMS: |
||
1099 | /* We can't compensate for compressed layout since at this point in the |
||
1100 | * program we haven't read from the MCS buffer. |
||
1101 | */ |
||
1102 | unreachable("Bad layout in encode_msaa"); |
||
1103 | case INTEL_MSAA_LAYOUT_UMS: |
||
1104 | /* No translation necessary. */ |
||
1105 | break; |
||
1106 | case INTEL_MSAA_LAYOUT_IMS: |
||
1107 | switch (num_samples) { |
||
1108 | case 4: |
||
1109 | /* encode_msaa(4, IMS, X, Y, S) = (X', Y', 0) |
||
1110 | * where X' = (X & ~0b1) << 1 | (S & 0b1) << 1 | (X & 0b1) |
||
1111 | * Y' = (Y & ~0b1) << 1 | (S & 0b10) | (Y & 0b1) |
||
1112 | */ |
||
1113 | emit_and(t1, X, brw_imm_uw(0xfffe)); /* X & ~0b1 */ |
||
1114 | if (!s_is_zero) { |
||
1115 | emit_and(t2, S, brw_imm_uw(1)); /* S & 0b1 */ |
||
1116 | emit_or(t1, t1, t2); /* (X & ~0b1) | (S & 0b1) */ |
||
1117 | } |
||
1118 | emit_shl(t1, t1, brw_imm_uw(1)); /* (X & ~0b1) << 1 |
||
1119 | | (S & 0b1) << 1 */ |
||
1120 | emit_and(t2, X, brw_imm_uw(1)); /* X & 0b1 */ |
||
1121 | emit_or(Xp, t1, t2); |
||
1122 | emit_and(t1, Y, brw_imm_uw(0xfffe)); /* Y & ~0b1 */ |
||
1123 | emit_shl(t1, t1, brw_imm_uw(1)); /* (Y & ~0b1) << 1 */ |
||
1124 | if (!s_is_zero) { |
||
1125 | emit_and(t2, S, brw_imm_uw(2)); /* S & 0b10 */ |
||
1126 | emit_or(t1, t1, t2); /* (Y & ~0b1) << 1 | (S & 0b10) */ |
||
1127 | } |
||
1128 | emit_and(t2, Y, brw_imm_uw(1)); /* Y & 0b1 */ |
||
1129 | emit_or(Yp, t1, t2); |
||
1130 | break; |
||
1131 | case 8: |
||
1132 | /* encode_msaa(8, IMS, X, Y, S) = (X', Y', 0) |
||
1133 | * where X' = (X & ~0b1) << 2 | (S & 0b100) | (S & 0b1) << 1 |
||
1134 | * | (X & 0b1) |
||
1135 | * Y' = (Y & ~0b1) << 1 | (S & 0b10) | (Y & 0b1) |
||
1136 | */ |
||
1137 | emit_and(t1, X, brw_imm_uw(0xfffe)); /* X & ~0b1 */ |
||
1138 | emit_shl(t1, t1, brw_imm_uw(2)); /* (X & ~0b1) << 2 */ |
||
1139 | if (!s_is_zero) { |
||
1140 | emit_and(t2, S, brw_imm_uw(4)); /* S & 0b100 */ |
||
1141 | emit_or(t1, t1, t2); /* (X & ~0b1) << 2 | (S & 0b100) */ |
||
1142 | emit_and(t2, S, brw_imm_uw(1)); /* S & 0b1 */ |
||
1143 | emit_shl(t2, t2, brw_imm_uw(1)); /* (S & 0b1) << 1 */ |
||
1144 | emit_or(t1, t1, t2); /* (X & ~0b1) << 2 | (S & 0b100) |
||
1145 | | (S & 0b1) << 1 */ |
||
1146 | } |
||
1147 | emit_and(t2, X, brw_imm_uw(1)); /* X & 0b1 */ |
||
1148 | emit_or(Xp, t1, t2); |
||
1149 | emit_and(t1, Y, brw_imm_uw(0xfffe)); /* Y & ~0b1 */ |
||
1150 | emit_shl(t1, t1, brw_imm_uw(1)); /* (Y & ~0b1) << 1 */ |
||
1151 | if (!s_is_zero) { |
||
1152 | emit_and(t2, S, brw_imm_uw(2)); /* S & 0b10 */ |
||
1153 | emit_or(t1, t1, t2); /* (Y & ~0b1) << 1 | (S & 0b10) */ |
||
1154 | } |
||
1155 | emit_and(t2, Y, brw_imm_uw(1)); /* Y & 0b1 */ |
||
1156 | emit_or(Yp, t1, t2); |
||
1157 | break; |
||
1158 | } |
||
1159 | SWAP_XY_AND_XPYP(); |
||
1160 | s_is_zero = true; |
||
1161 | break; |
||
1162 | } |
||
1163 | } |
||
1164 | |||
1165 | /** |
||
1166 | * Emit code to compensate for the difference between MSAA and non-MSAA |
||
1167 | * surfaces. |
||
1168 | * |
||
1169 | * This code modifies the X and Y coordinates according to the formula: |
||
1170 | * |
||
1171 | * (X', Y', S) = decode_msaa(num_samples, IMS, X, Y, S) |
||
1172 | * |
||
1173 | * (See brw_blorp_blit_program). |
||
1174 | */ |
||
1175 | void |
||
1176 | brw_blorp_blit_program::decode_msaa(unsigned num_samples, |
||
1177 | intel_msaa_layout layout) |
||
1178 | { |
||
1179 | switch (layout) { |
||
1180 | case INTEL_MSAA_LAYOUT_NONE: |
||
1181 | /* No translation necessary, and S should already be zero. */ |
||
1182 | assert(s_is_zero); |
||
1183 | break; |
||
1184 | case INTEL_MSAA_LAYOUT_CMS: |
||
1185 | /* We can't compensate for compressed layout since at this point in the |
||
1186 | * program we don't have access to the MCS buffer. |
||
1187 | */ |
||
1188 | unreachable("Bad layout in encode_msaa"); |
||
1189 | case INTEL_MSAA_LAYOUT_UMS: |
||
1190 | /* No translation necessary. */ |
||
1191 | break; |
||
1192 | case INTEL_MSAA_LAYOUT_IMS: |
||
1193 | assert(s_is_zero); |
||
1194 | switch (num_samples) { |
||
1195 | case 4: |
||
1196 | /* decode_msaa(4, IMS, X, Y, 0) = (X', Y', S) |
||
1197 | * where X' = (X & ~0b11) >> 1 | (X & 0b1) |
||
1198 | * Y' = (Y & ~0b11) >> 1 | (Y & 0b1) |
||
1199 | * S = (Y & 0b10) | (X & 0b10) >> 1 |
||
1200 | */ |
||
1201 | emit_and(t1, X, brw_imm_uw(0xfffc)); /* X & ~0b11 */ |
||
1202 | emit_shr(t1, t1, brw_imm_uw(1)); /* (X & ~0b11) >> 1 */ |
||
1203 | emit_and(t2, X, brw_imm_uw(1)); /* X & 0b1 */ |
||
1204 | emit_or(Xp, t1, t2); |
||
1205 | emit_and(t1, Y, brw_imm_uw(0xfffc)); /* Y & ~0b11 */ |
||
1206 | emit_shr(t1, t1, brw_imm_uw(1)); /* (Y & ~0b11) >> 1 */ |
||
1207 | emit_and(t2, Y, brw_imm_uw(1)); /* Y & 0b1 */ |
||
1208 | emit_or(Yp, t1, t2); |
||
1209 | emit_and(t1, Y, brw_imm_uw(2)); /* Y & 0b10 */ |
||
1210 | emit_and(t2, X, brw_imm_uw(2)); /* X & 0b10 */ |
||
1211 | emit_shr(t2, t2, brw_imm_uw(1)); /* (X & 0b10) >> 1 */ |
||
1212 | emit_or(S, t1, t2); |
||
1213 | break; |
||
1214 | case 8: |
||
1215 | /* decode_msaa(8, IMS, X, Y, 0) = (X', Y', S) |
||
1216 | * where X' = (X & ~0b111) >> 2 | (X & 0b1) |
||
1217 | * Y' = (Y & ~0b11) >> 1 | (Y & 0b1) |
||
1218 | * S = (X & 0b100) | (Y & 0b10) | (X & 0b10) >> 1 |
||
1219 | */ |
||
1220 | emit_and(t1, X, brw_imm_uw(0xfff8)); /* X & ~0b111 */ |
||
1221 | emit_shr(t1, t1, brw_imm_uw(2)); /* (X & ~0b111) >> 2 */ |
||
1222 | emit_and(t2, X, brw_imm_uw(1)); /* X & 0b1 */ |
||
1223 | emit_or(Xp, t1, t2); |
||
1224 | emit_and(t1, Y, brw_imm_uw(0xfffc)); /* Y & ~0b11 */ |
||
1225 | emit_shr(t1, t1, brw_imm_uw(1)); /* (Y & ~0b11) >> 1 */ |
||
1226 | emit_and(t2, Y, brw_imm_uw(1)); /* Y & 0b1 */ |
||
1227 | emit_or(Yp, t1, t2); |
||
1228 | emit_and(t1, X, brw_imm_uw(4)); /* X & 0b100 */ |
||
1229 | emit_and(t2, Y, brw_imm_uw(2)); /* Y & 0b10 */ |
||
1230 | emit_or(t1, t1, t2); /* (X & 0b100) | (Y & 0b10) */ |
||
1231 | emit_and(t2, X, brw_imm_uw(2)); /* X & 0b10 */ |
||
1232 | emit_shr(t2, t2, brw_imm_uw(1)); /* (X & 0b10) >> 1 */ |
||
1233 | emit_or(S, t1, t2); |
||
1234 | break; |
||
1235 | } |
||
1236 | s_is_zero = false; |
||
1237 | SWAP_XY_AND_XPYP(); |
||
1238 | break; |
||
1239 | } |
||
1240 | } |
||
1241 | |||
1242 | /** |
||
1243 | * Emit code to translate from destination (X, Y) coordinates to source (X, Y) |
||
1244 | * coordinates. |
||
1245 | */ |
||
1246 | void |
||
1247 | brw_blorp_blit_program::translate_dst_to_src() |
||
1248 | { |
||
1249 | struct brw_reg X_f = retype(X, BRW_REGISTER_TYPE_F); |
||
1250 | struct brw_reg Y_f = retype(Y, BRW_REGISTER_TYPE_F); |
||
1251 | struct brw_reg Xp_f = retype(Xp, BRW_REGISTER_TYPE_F); |
||
1252 | struct brw_reg Yp_f = retype(Yp, BRW_REGISTER_TYPE_F); |
||
1253 | |||
1254 | /* Move the UD coordinates to float registers. */ |
||
1255 | emit_mov(Xp_f, X); |
||
1256 | emit_mov(Yp_f, Y); |
||
1257 | /* Scale and offset */ |
||
1258 | emit_mad(X_f, x_transform.offset, Xp_f, x_transform.multiplier); |
||
1259 | emit_mad(Y_f, y_transform.offset, Yp_f, y_transform.multiplier); |
||
1260 | if (key->blit_scaled && key->blend) { |
||
1261 | /* Translate coordinates to lay out the samples in a rectangular grid |
||
1262 | * roughly corresponding to sample locations. |
||
1263 | */ |
||
1264 | emit_mul(X_f, X_f, brw_imm_f(key->x_scale)); |
||
1265 | emit_mul(Y_f, Y_f, brw_imm_f(key->y_scale)); |
||
1266 | /* Adjust coordinates so that integers represent pixel centers rather |
||
1267 | * than pixel edges. |
||
1268 | */ |
||
1269 | emit_add(X_f, X_f, brw_imm_f(-0.5)); |
||
1270 | emit_add(Y_f, Y_f, brw_imm_f(-0.5)); |
||
1271 | |||
1272 | /* Clamp the X, Y texture coordinates to properly handle the sampling of |
||
1273 | * texels on texture edges. |
||
1274 | */ |
||
1275 | clamp_tex_coords(X_f, Y_f, |
||
1276 | brw_imm_f(0.0), brw_imm_f(0.0), |
||
1277 | rect_grid_x1, rect_grid_y1); |
||
1278 | |||
1279 | /* Store the fractional parts to be used as bilinear interpolation |
||
1280 | * coefficients. |
||
1281 | */ |
||
1282 | emit_frc(x_frac, X_f); |
||
1283 | emit_frc(y_frac, Y_f); |
||
1284 | |||
1285 | /* Round the float coordinates down to nearest integer */ |
||
1286 | emit_rndd(Xp_f, X_f); |
||
1287 | emit_rndd(Yp_f, Y_f); |
||
1288 | emit_mul(X_f, Xp_f, brw_imm_f(1 / key->x_scale)); |
||
1289 | emit_mul(Y_f, Yp_f, brw_imm_f(1 / key->y_scale)); |
||
1290 | SWAP_XY_AND_XPYP(); |
||
1291 | } else if (!key->bilinear_filter) { |
||
1292 | /* Round the float coordinates down to nearest integer by moving to |
||
1293 | * UD registers. |
||
1294 | */ |
||
1295 | emit_mov(Xp, X_f); |
||
1296 | emit_mov(Yp, Y_f); |
||
1297 | SWAP_XY_AND_XPYP(); |
||
1298 | } |
||
1299 | } |
||
1300 | |||
1301 | void |
||
1302 | brw_blorp_blit_program::clamp_tex_coords(struct brw_reg regX, |
||
1303 | struct brw_reg regY, |
||
1304 | struct brw_reg clampX0, |
||
1305 | struct brw_reg clampY0, |
||
1306 | struct brw_reg clampX1, |
||
1307 | struct brw_reg clampY1) |
||
1308 | { |
||
1309 | emit_max(regX, regX, clampX0); |
||
1310 | emit_max(regY, regY, clampY0); |
||
1311 | emit_min(regX, regX, clampX1); |
||
1312 | emit_min(regY, regY, clampY1); |
||
1313 | } |
||
1314 | |||
1315 | /** |
||
1316 | * Emit code to transform the X and Y coordinates as needed for blending |
||
1317 | * together the different samples in an MSAA texture. |
||
1318 | */ |
||
1319 | void |
||
1320 | brw_blorp_blit_program::single_to_blend() |
||
1321 | { |
||
1322 | /* When looking up samples in an MSAA texture using the SAMPLE message, |
||
1323 | * Gen6 requires the texture coordinates to be odd integers (so that they |
||
1324 | * correspond to the center of a 2x2 block representing the four samples |
||
1325 | * that maxe up a pixel). So we need to multiply our X and Y coordinates |
||
1326 | * each by 2 and then add 1. |
||
1327 | */ |
||
1328 | emit_shl(t1, X, brw_imm_w(1)); |
||
1329 | emit_shl(t2, Y, brw_imm_w(1)); |
||
1330 | emit_add(Xp, t1, brw_imm_w(1)); |
||
1331 | emit_add(Yp, t2, brw_imm_w(1)); |
||
1332 | SWAP_XY_AND_XPYP(); |
||
1333 | } |
||
1334 | |||
1335 | |||
1336 | /** |
||
1337 | * Count the number of trailing 1 bits in the given value. For example: |
||
1338 | * |
||
1339 | * count_trailing_one_bits(0) == 0 |
||
1340 | * count_trailing_one_bits(7) == 3 |
||
1341 | * count_trailing_one_bits(11) == 2 |
||
1342 | */ |
||
1343 | inline int count_trailing_one_bits(unsigned value) |
||
1344 | { |
||
1345 | #ifdef HAVE___BUILTIN_CTZ |
||
1346 | return __builtin_ctz(~value); |
||
1347 | #else |
||
1348 | return _mesa_bitcount(value & ~(value + 1)); |
||
1349 | #endif |
||
1350 | } |
||
1351 | |||
1352 | |||
1353 | void |
||
1354 | brw_blorp_blit_program::manual_blend_average(unsigned num_samples) |
||
1355 | { |
||
1356 | if (key->tex_layout == INTEL_MSAA_LAYOUT_CMS) |
||
1357 | mcs_fetch(); |
||
1358 | |||
1359 | /* We add together samples using a binary tree structure, e.g. for 4x MSAA: |
||
1360 | * |
||
1361 | * result = ((sample[0] + sample[1]) + (sample[2] + sample[3])) / 4 |
||
1362 | * |
||
1363 | * This ensures that when all samples have the same value, no numerical |
||
1364 | * precision is lost, since each addition operation always adds two equal |
||
1365 | * values, and summing two equal floating point values does not lose |
||
1366 | * precision. |
||
1367 | * |
||
1368 | * We perform this computation by treating the texture_data array as a |
||
1369 | * stack and performing the following operations: |
||
1370 | * |
||
1371 | * - push sample 0 onto stack |
||
1372 | * - push sample 1 onto stack |
||
1373 | * - add top two stack entries |
||
1374 | * - push sample 2 onto stack |
||
1375 | * - push sample 3 onto stack |
||
1376 | * - add top two stack entries |
||
1377 | * - add top two stack entries |
||
1378 | * - divide top stack entry by 4 |
||
1379 | * |
||
1380 | * Note that after pushing sample i onto the stack, the number of add |
||
1381 | * operations we do is equal to the number of trailing 1 bits in i. This |
||
1382 | * works provided the total number of samples is a power of two, which it |
||
1383 | * always is for i965. |
||
1384 | * |
||
1385 | * For integer formats, we replace the add operations with average |
||
1386 | * operations and skip the final division. |
||
1387 | */ |
||
1388 | unsigned stack_depth = 0; |
||
1389 | for (unsigned i = 0; i < num_samples; ++i) { |
||
1390 | assert(stack_depth == _mesa_bitcount(i)); /* Loop invariant */ |
||
1391 | |||
1392 | /* Push sample i onto the stack */ |
||
1393 | assert(stack_depth < ARRAY_SIZE(texture_data)); |
||
1394 | if (i == 0) { |
||
1395 | s_is_zero = true; |
||
1396 | } else { |
||
1397 | s_is_zero = false; |
||
1398 | emit_mov(vec16(S), brw_imm_ud(i)); |
||
1399 | } |
||
1400 | texel_fetch(texture_data[stack_depth++]); |
||
1401 | |||
1402 | if (i == 0 && key->tex_layout == INTEL_MSAA_LAYOUT_CMS) { |
||
1403 | /* The Ivy Bridge PRM, Vol4 Part1 p27 (Multisample Control Surface) |
||
1404 | * suggests an optimization: |
||
1405 | * |
||
1406 | * "A simple optimization with probable large return in |
||
1407 | * performance is to compare the MCS value to zero (indicating |
||
1408 | * all samples are on sample slice 0), and sample only from |
||
1409 | * sample slice 0 using ld2dss if MCS is zero." |
||
1410 | * |
||
1411 | * Note that in the case where the MCS value is zero, sampling from |
||
1412 | * sample slice 0 using ld2dss and sampling from sample 0 using |
||
1413 | * ld2dms are equivalent (since all samples are on sample slice 0). |
||
1414 | * Since we have already sampled from sample 0, all we need to do is |
||
1415 | * skip the remaining fetches and averaging if MCS is zero. |
||
1416 | */ |
||
1417 | emit_cmp_if(BRW_CONDITIONAL_NZ, mcs_data, brw_imm_ud(0)); |
||
1418 | } |
||
1419 | |||
1420 | /* Do count_trailing_one_bits(i) times */ |
||
1421 | for (int j = count_trailing_one_bits(i); j-- > 0; ) { |
||
1422 | assert(stack_depth >= 2); |
||
1423 | --stack_depth; |
||
1424 | |||
1425 | /* TODO: should use a smaller loop bound for non_RGBA formats */ |
||
1426 | for (int k = 0; k < 4; ++k) { |
||
1427 | emit_combine(key->texture_data_type == BRW_REGISTER_TYPE_F ? |
||
1428 | BRW_OPCODE_ADD : BRW_OPCODE_AVG, |
||
1429 | offset(texture_data[stack_depth - 1], 2*k), |
||
1430 | offset(vec8(texture_data[stack_depth - 1]), 2*k), |
||
1431 | offset(vec8(texture_data[stack_depth]), 2*k)); |
||
1432 | } |
||
1433 | } |
||
1434 | } |
||
1435 | |||
1436 | /* We should have just 1 sample on the stack now. */ |
||
1437 | assert(stack_depth == 1); |
||
1438 | |||
1439 | if (key->texture_data_type == BRW_REGISTER_TYPE_F) { |
||
1440 | /* Scale the result down by a factor of num_samples */ |
||
1441 | /* TODO: should use a smaller loop bound for non-RGBA formats */ |
||
1442 | for (int j = 0; j < 4; ++j) { |
||
1443 | emit_mul(offset(texture_data[0], 2*j), |
||
1444 | offset(vec8(texture_data[0]), 2*j), |
||
1445 | brw_imm_f(1.0/num_samples)); |
||
1446 | } |
||
1447 | } |
||
1448 | |||
1449 | if (key->tex_layout == INTEL_MSAA_LAYOUT_CMS) |
||
1450 | emit_endif(); |
||
1451 | } |
||
1452 | |||
1453 | void |
||
1454 | brw_blorp_blit_program::manual_blend_bilinear(unsigned num_samples) |
||
1455 | { |
||
1456 | /* We do this computation by performing the following operations: |
||
1457 | * |
||
1458 | * In case of 4x, 8x MSAA: |
||
1459 | * - Compute the pixel coordinates and sample numbers (a, b, c, d) |
||
1460 | * which are later used for interpolation |
||
1461 | * - linearly interpolate samples a and b in X |
||
1462 | * - linearly interpolate samples c and d in X |
||
1463 | * - linearly interpolate the results of last two operations in Y |
||
1464 | * |
||
1465 | * result = lrp(lrp(a + b) + lrp(c + d)) |
||
1466 | */ |
||
1467 | struct brw_reg Xp_f = retype(Xp, BRW_REGISTER_TYPE_F); |
||
1468 | struct brw_reg Yp_f = retype(Yp, BRW_REGISTER_TYPE_F); |
||
1469 | struct brw_reg t1_f = retype(t1, BRW_REGISTER_TYPE_F); |
||
1470 | struct brw_reg t2_f = retype(t2, BRW_REGISTER_TYPE_F); |
||
1471 | |||
1472 | for (unsigned i = 0; i < 4; ++i) { |
||
1473 | assert(i < ARRAY_SIZE(texture_data)); |
||
1474 | s_is_zero = false; |
||
1475 | |||
1476 | /* Compute pixel coordinates */ |
||
1477 | emit_add(vec16(x_sample_coords), Xp_f, |
||
1478 | brw_imm_f((float)(i & 0x1) * (1.0 / key->x_scale))); |
||
1479 | emit_add(vec16(y_sample_coords), Yp_f, |
||
1480 | brw_imm_f((float)((i >> 1) & 0x1) * (1.0 / key->y_scale))); |
||
1481 | emit_mov(vec16(X), x_sample_coords); |
||
1482 | emit_mov(vec16(Y), y_sample_coords); |
||
1483 | |||
1484 | /* The MCS value we fetch has to match up with the pixel that we're |
||
1485 | * sampling from. Since we sample from different pixels in each |
||
1486 | * iteration of this "for" loop, the call to mcs_fetch() should be |
||
1487 | * here inside the loop after computing the pixel coordinates. |
||
1488 | */ |
||
1489 | if (key->tex_layout == INTEL_MSAA_LAYOUT_CMS) |
||
1490 | mcs_fetch(); |
||
1491 | |||
1492 | /* Compute sample index and map the sample index to a sample number. |
||
1493 | * Sample index layout shows the numbering of slots in a rectangular |
||
1494 | * grid of samples with in a pixel. Sample number layout shows the |
||
1495 | * rectangular grid of samples roughly corresponding to the real sample |
||
1496 | * locations with in a pixel. |
||
1497 | * In case of 4x MSAA, layout of sample indices matches the layout of |
||
1498 | * sample numbers: |
||
1499 | * --------- |
||
1500 | * | 0 | 1 | |
||
1501 | * --------- |
||
1502 | * | 2 | 3 | |
||
1503 | * --------- |
||
1504 | * |
||
1505 | * In case of 8x MSAA the two layouts don't match. |
||
1506 | * sample index layout : --------- sample number layout : --------- |
||
1507 | * | 0 | 1 | | 5 | 2 | |
||
1508 | * --------- --------- |
||
1509 | * | 2 | 3 | | 4 | 6 | |
||
1510 | * --------- --------- |
||
1511 | * | 4 | 5 | | 0 | 3 | |
||
1512 | * --------- --------- |
||
1513 | * | 6 | 7 | | 7 | 1 | |
||
1514 | * --------- --------- |
||
1515 | */ |
||
1516 | emit_frc(vec16(t1_f), x_sample_coords); |
||
1517 | emit_frc(vec16(t2_f), y_sample_coords); |
||
1518 | emit_mul(vec16(t1_f), t1_f, brw_imm_f(key->x_scale)); |
||
1519 | emit_mul(vec16(t2_f), t2_f, brw_imm_f(key->x_scale * key->y_scale)); |
||
1520 | emit_add(vec16(t1_f), t1_f, t2_f); |
||
1521 | emit_mov(vec16(S), t1_f); |
||
1522 | |||
1523 | if (num_samples == 8) { |
||
1524 | /* Map the sample index to a sample number */ |
||
1525 | emit_cmp_if(BRW_CONDITIONAL_L, S, brw_imm_d(4)); |
||
1526 | { |
||
1527 | emit_mov(vec16(t2), brw_imm_d(5)); |
||
1528 | emit_if_eq_mov(S, 1, vec16(t2), 2); |
||
1529 | emit_if_eq_mov(S, 2, vec16(t2), 4); |
||
1530 | emit_if_eq_mov(S, 3, vec16(t2), 6); |
||
1531 | } |
||
1532 | emit_else(); |
||
1533 | { |
||
1534 | emit_mov(vec16(t2), brw_imm_d(0)); |
||
1535 | emit_if_eq_mov(S, 5, vec16(t2), 3); |
||
1536 | emit_if_eq_mov(S, 6, vec16(t2), 7); |
||
1537 | emit_if_eq_mov(S, 7, vec16(t2), 1); |
||
1538 | } |
||
1539 | emit_endif(); |
||
1540 | emit_mov(vec16(S), t2); |
||
1541 | } |
||
1542 | texel_fetch(texture_data[i]); |
||
1543 | } |
||
1544 | |||
1545 | #define SAMPLE(x, y) offset(texture_data[x], y) |
||
1546 | for (int index = 3; index > 0; ) { |
||
1547 | /* Since we're doing SIMD16, 4 color channels fits in to 8 registers. |
||
1548 | * Counter value of 8 in 'for' loop below is used to interpolate all |
||
1549 | * the color components. |
||
1550 | */ |
||
1551 | for (int k = 0; k < 8; k += 2) |
||
1552 | emit_lrp(vec8(SAMPLE(index - 1, k)), |
||
1553 | x_frac, |
||
1554 | vec8(SAMPLE(index, k)), |
||
1555 | vec8(SAMPLE(index - 1, k))); |
||
1556 | index -= 2; |
||
1557 | } |
||
1558 | for (int k = 0; k < 8; k += 2) |
||
1559 | emit_lrp(vec8(SAMPLE(0, k)), |
||
1560 | y_frac, |
||
1561 | vec8(SAMPLE(2, k)), |
||
1562 | vec8(SAMPLE(0, k))); |
||
1563 | #undef SAMPLE |
||
1564 | } |
||
1565 | |||
1566 | /** |
||
1567 | * Emit code to look up a value in the texture using the SAMPLE message (which |
||
1568 | * does blending of MSAA surfaces). |
||
1569 | */ |
||
1570 | void |
||
1571 | brw_blorp_blit_program::sample(struct brw_reg dst) |
||
1572 | { |
||
1573 | static const sampler_message_arg args[2] = { |
||
1574 | SAMPLER_MESSAGE_ARG_U_FLOAT, |
||
1575 | SAMPLER_MESSAGE_ARG_V_FLOAT |
||
1576 | }; |
||
1577 | |||
1578 | texture_lookup(dst, SHADER_OPCODE_TEX, args, ARRAY_SIZE(args)); |
||
1579 | } |
||
1580 | |||
1581 | /** |
||
1582 | * Emit code to look up a value in the texture using the SAMPLE_LD message |
||
1583 | * (which does a simple texel fetch). |
||
1584 | */ |
||
1585 | void |
||
1586 | brw_blorp_blit_program::texel_fetch(struct brw_reg dst) |
||
1587 | { |
||
1588 | static const sampler_message_arg gen6_args[5] = { |
||
1589 | SAMPLER_MESSAGE_ARG_U_INT, |
||
1590 | SAMPLER_MESSAGE_ARG_V_INT, |
||
1591 | SAMPLER_MESSAGE_ARG_ZERO_INT, /* R */ |
||
1592 | SAMPLER_MESSAGE_ARG_ZERO_INT, /* LOD */ |
||
1593 | SAMPLER_MESSAGE_ARG_SI_INT |
||
1594 | }; |
||
1595 | static const sampler_message_arg gen7_ld_args[3] = { |
||
1596 | SAMPLER_MESSAGE_ARG_U_INT, |
||
1597 | SAMPLER_MESSAGE_ARG_ZERO_INT, /* LOD */ |
||
1598 | SAMPLER_MESSAGE_ARG_V_INT |
||
1599 | }; |
||
1600 | static const sampler_message_arg gen7_ld2dss_args[3] = { |
||
1601 | SAMPLER_MESSAGE_ARG_SI_INT, |
||
1602 | SAMPLER_MESSAGE_ARG_U_INT, |
||
1603 | SAMPLER_MESSAGE_ARG_V_INT |
||
1604 | }; |
||
1605 | static const sampler_message_arg gen7_ld2dms_args[4] = { |
||
1606 | SAMPLER_MESSAGE_ARG_SI_INT, |
||
1607 | SAMPLER_MESSAGE_ARG_MCS_INT, |
||
1608 | SAMPLER_MESSAGE_ARG_U_INT, |
||
1609 | SAMPLER_MESSAGE_ARG_V_INT |
||
1610 | }; |
||
1611 | |||
1612 | switch (brw->gen) { |
||
1613 | case 6: |
||
1614 | texture_lookup(dst, SHADER_OPCODE_TXF, gen6_args, s_is_zero ? 2 : 5); |
||
1615 | break; |
||
1616 | case 7: |
||
1617 | switch (key->tex_layout) { |
||
1618 | case INTEL_MSAA_LAYOUT_IMS: |
||
1619 | /* From the Ivy Bridge PRM, Vol4 Part1 p72 (Multisampled Surface Storage |
||
1620 | * Format): |
||
1621 | * |
||
1622 | * If this field is MSFMT_DEPTH_STENCIL |
||
1623 | * [a.k.a. INTEL_MSAA_LAYOUT_IMS], the only sampling engine |
||
1624 | * messages allowed are "ld2dms", "resinfo", and "sampleinfo". |
||
1625 | * |
||
1626 | * So fall through to emit the same message as we use for |
||
1627 | * INTEL_MSAA_LAYOUT_CMS. |
||
1628 | */ |
||
1629 | case INTEL_MSAA_LAYOUT_CMS: |
||
1630 | texture_lookup(dst, SHADER_OPCODE_TXF_CMS, |
||
1631 | gen7_ld2dms_args, ARRAY_SIZE(gen7_ld2dms_args)); |
||
1632 | break; |
||
1633 | case INTEL_MSAA_LAYOUT_UMS: |
||
1634 | texture_lookup(dst, SHADER_OPCODE_TXF_UMS, |
||
1635 | gen7_ld2dss_args, ARRAY_SIZE(gen7_ld2dss_args)); |
||
1636 | break; |
||
1637 | case INTEL_MSAA_LAYOUT_NONE: |
||
1638 | assert(s_is_zero); |
||
1639 | texture_lookup(dst, SHADER_OPCODE_TXF, gen7_ld_args, |
||
1640 | ARRAY_SIZE(gen7_ld_args)); |
||
1641 | break; |
||
1642 | } |
||
1643 | break; |
||
1644 | default: |
||
1645 | unreachable("Should not get here."); |
||
1646 | }; |
||
1647 | } |
||
1648 | |||
1649 | void |
||
1650 | brw_blorp_blit_program::mcs_fetch() |
||
1651 | { |
||
1652 | static const sampler_message_arg gen7_ld_mcs_args[2] = { |
||
1653 | SAMPLER_MESSAGE_ARG_U_INT, |
||
1654 | SAMPLER_MESSAGE_ARG_V_INT |
||
1655 | }; |
||
1656 | texture_lookup(vec16(mcs_data), SHADER_OPCODE_TXF_MCS, |
||
1657 | gen7_ld_mcs_args, ARRAY_SIZE(gen7_ld_mcs_args)); |
||
1658 | } |
||
1659 | |||
1660 | void |
||
1661 | brw_blorp_blit_program::texture_lookup(struct brw_reg dst, |
||
1662 | enum opcode op, |
||
1663 | const sampler_message_arg *args, |
||
1664 | int num_args) |
||
1665 | { |
||
1666 | struct brw_reg mrf = |
||
1667 | retype(vec16(brw_message_reg(base_mrf)), BRW_REGISTER_TYPE_UD); |
||
1668 | for (int arg = 0; arg < num_args; ++arg) { |
||
1669 | switch (args[arg]) { |
||
1670 | case SAMPLER_MESSAGE_ARG_U_FLOAT: |
||
1671 | if (key->bilinear_filter) |
||
1672 | emit_mov(retype(mrf, BRW_REGISTER_TYPE_F), |
||
1673 | retype(X, BRW_REGISTER_TYPE_F)); |
||
1674 | else |
||
1675 | emit_mov(retype(mrf, BRW_REGISTER_TYPE_F), X); |
||
1676 | break; |
||
1677 | case SAMPLER_MESSAGE_ARG_V_FLOAT: |
||
1678 | if (key->bilinear_filter) |
||
1679 | emit_mov(retype(mrf, BRW_REGISTER_TYPE_F), |
||
1680 | retype(Y, BRW_REGISTER_TYPE_F)); |
||
1681 | else |
||
1682 | emit_mov(retype(mrf, BRW_REGISTER_TYPE_F), Y); |
||
1683 | break; |
||
1684 | case SAMPLER_MESSAGE_ARG_U_INT: |
||
1685 | emit_mov(mrf, X); |
||
1686 | break; |
||
1687 | case SAMPLER_MESSAGE_ARG_V_INT: |
||
1688 | emit_mov(mrf, Y); |
||
1689 | break; |
||
1690 | case SAMPLER_MESSAGE_ARG_SI_INT: |
||
1691 | /* Note: on Gen7, this code may be reached with s_is_zero==true |
||
1692 | * because in Gen7's ld2dss message, the sample index is the first |
||
1693 | * argument. When this happens, we need to move a 0 into the |
||
1694 | * appropriate message register. |
||
1695 | */ |
||
1696 | if (s_is_zero) |
||
1697 | emit_mov(mrf, brw_imm_ud(0)); |
||
1698 | else |
||
1699 | emit_mov(mrf, S); |
||
1700 | break; |
||
1701 | case SAMPLER_MESSAGE_ARG_MCS_INT: |
||
1702 | switch (key->tex_layout) { |
||
1703 | case INTEL_MSAA_LAYOUT_CMS: |
||
1704 | emit_mov(mrf, mcs_data); |
||
1705 | break; |
||
1706 | case INTEL_MSAA_LAYOUT_IMS: |
||
1707 | /* When sampling from an IMS surface, MCS data is not relevant, |
||
1708 | * and the hardware ignores it. So don't bother populating it. |
||
1709 | */ |
||
1710 | break; |
||
1711 | default: |
||
1712 | /* We shouldn't be trying to send MCS data with any other |
||
1713 | * layouts. |
||
1714 | */ |
||
1715 | assert (!"Unsupported layout for MCS data"); |
||
1716 | break; |
||
1717 | } |
||
1718 | break; |
||
1719 | case SAMPLER_MESSAGE_ARG_ZERO_INT: |
||
1720 | emit_mov(mrf, brw_imm_ud(0)); |
||
1721 | break; |
||
1722 | } |
||
1723 | mrf.nr += 2; |
||
1724 | } |
||
1725 | |||
1726 | emit_texture_lookup(retype(dst, BRW_REGISTER_TYPE_UW) /* dest */, |
||
1727 | op, |
||
1728 | base_mrf, |
||
1729 | mrf.nr - base_mrf /* msg_length */); |
||
1730 | } |
||
1731 | |||
1732 | #undef X |
||
1733 | #undef Y |
||
1734 | #undef U |
||
1735 | #undef V |
||
1736 | #undef S |
||
1737 | #undef SWAP_XY_AND_XPYP |
||
1738 | |||
1739 | void |
||
1740 | brw_blorp_blit_program::render_target_write() |
||
1741 | { |
||
1742 | struct brw_reg mrf_rt_write = |
||
1743 | retype(vec16(brw_message_reg(base_mrf)), key->texture_data_type); |
||
1744 | int mrf_offset = 0; |
||
1745 | |||
1746 | /* If we may have killed pixels, then we need to send R0 and R1 in a header |
||
1747 | * so that the render target knows which pixels we killed. |
||
1748 | */ |
||
1749 | bool use_header = key->use_kill; |
||
1750 | if (use_header) { |
||
1751 | /* Copy R0/1 to MRF */ |
||
1752 | emit_mov(retype(mrf_rt_write, BRW_REGISTER_TYPE_UD), |
||
1753 | retype(R0, BRW_REGISTER_TYPE_UD)); |
||
1754 | mrf_offset += 2; |
||
1755 | } |
||
1756 | |||
1757 | /* Copy texture data to MRFs */ |
||
1758 | for (int i = 0; i < 4; ++i) { |
||
1759 | /* E.g. mov(16) m2.0<1>:f r2.0<8;8,1>:f { Align1, H1 } */ |
||
1760 | emit_mov(offset(mrf_rt_write, mrf_offset), |
||
1761 | offset(vec8(texture_data[0]), 2*i)); |
||
1762 | mrf_offset += 2; |
||
1763 | } |
||
1764 | |||
1765 | /* Now write to the render target and terminate the thread */ |
||
1766 | emit_render_target_write( |
||
1767 | mrf_rt_write, |
||
1768 | base_mrf, |
||
1769 | mrf_offset /* msg_length. TODO: Should be smaller for non-RGBA formats. */, |
||
1770 | use_header); |
||
1771 | } |
||
1772 | |||
1773 | |||
1774 | void |
||
1775 | brw_blorp_coord_transform_params::setup(GLfloat src0, GLfloat src1, |
||
1776 | GLfloat dst0, GLfloat dst1, |
||
1777 | bool mirror) |
||
1778 | { |
||
1779 | float scale = (src1 - src0) / (dst1 - dst0); |
||
1780 | if (!mirror) { |
||
1781 | /* When not mirroring a coordinate (say, X), we need: |
||
1782 | * src_x - src_x0 = (dst_x - dst_x0 + 0.5) * scale |
||
1783 | * Therefore: |
||
1784 | * src_x = src_x0 + (dst_x - dst_x0 + 0.5) * scale |
||
1785 | * |
||
1786 | * blorp program uses "round toward zero" to convert the |
||
1787 | * transformed floating point coordinates to integer coordinates, |
||
1788 | * whereas the behaviour we actually want is "round to nearest", |
||
1789 | * so 0.5 provides the necessary correction. |
||
1790 | */ |
||
1791 | multiplier = scale; |
||
1792 | offset = src0 + (-dst0 + 0.5) * scale; |
||
1793 | } else { |
||
1794 | /* When mirroring X we need: |
||
1795 | * src_x - src_x0 = dst_x1 - dst_x - 0.5 |
||
1796 | * Therefore: |
||
1797 | * src_x = src_x0 + (dst_x1 -dst_x - 0.5) * scale |
||
1798 | */ |
||
1799 | multiplier = -scale; |
||
1800 | offset = src0 + (dst1 - 0.5) * scale; |
||
1801 | } |
||
1802 | } |
||
1803 | |||
1804 | |||
1805 | /** |
||
1806 | * Determine which MSAA layout the GPU pipeline should be configured for, |
||
1807 | * based on the chip generation, the number of samples, and the true layout of |
||
1808 | * the image in memory. |
||
1809 | */ |
||
1810 | inline intel_msaa_layout |
||
1811 | compute_msaa_layout_for_pipeline(struct brw_context *brw, unsigned num_samples, |
||
1812 | intel_msaa_layout true_layout) |
||
1813 | { |
||
1814 | if (num_samples <= 1) { |
||
1815 | /* When configuring the GPU for non-MSAA, we can still accommodate IMS |
||
1816 | * format buffers, by transforming coordinates appropriately. |
||
1817 | */ |
||
1818 | assert(true_layout == INTEL_MSAA_LAYOUT_NONE || |
||
1819 | true_layout == INTEL_MSAA_LAYOUT_IMS); |
||
1820 | return INTEL_MSAA_LAYOUT_NONE; |
||
1821 | } else { |
||
1822 | assert(true_layout != INTEL_MSAA_LAYOUT_NONE); |
||
1823 | } |
||
1824 | |||
1825 | /* Prior to Gen7, all MSAA surfaces use IMS layout. */ |
||
1826 | if (brw->gen == 6) { |
||
1827 | assert(true_layout == INTEL_MSAA_LAYOUT_IMS); |
||
1828 | } |
||
1829 | |||
1830 | return true_layout; |
||
1831 | } |
||
1832 | |||
1833 | |||
1834 | brw_blorp_blit_params::brw_blorp_blit_params(struct brw_context *brw, |
||
1835 | struct intel_mipmap_tree *src_mt, |
||
1836 | unsigned src_level, unsigned src_layer, |
||
1837 | mesa_format src_format, |
||
1838 | struct intel_mipmap_tree *dst_mt, |
||
1839 | unsigned dst_level, unsigned dst_layer, |
||
1840 | mesa_format dst_format, |
||
1841 | GLfloat src_x0, GLfloat src_y0, |
||
1842 | GLfloat src_x1, GLfloat src_y1, |
||
1843 | GLfloat dst_x0, GLfloat dst_y0, |
||
1844 | GLfloat dst_x1, GLfloat dst_y1, |
||
1845 | GLenum filter, |
||
1846 | bool mirror_x, bool mirror_y) |
||
1847 | { |
||
1848 | src.set(brw, src_mt, src_level, src_layer, src_format, false); |
||
1849 | dst.set(brw, dst_mt, dst_level, dst_layer, dst_format, true); |
||
1850 | |||
1851 | /* Even though we do multisample resolves at the time of the blit, OpenGL |
||
1852 | * specification defines them as if they happen at the time of rendering, |
||
1853 | * which means that the type of averaging we do during the resolve should |
||
1854 | * only depend on the source format; the destination format should be |
||
1855 | * ignored. But, specification doesn't seem to be strict about it. |
||
1856 | * |
||
1857 | * It has been observed that mulitisample resolves produce slightly better |
||
1858 | * looking images when averaging is done using destination format. NVIDIA's |
||
1859 | * proprietary OpenGL driver also follow this approach. So, we choose to |
||
1860 | * follow it in our driver. |
||
1861 | * |
||
1862 | * When multisampling, if the source and destination formats are equal |
||
1863 | * (aside from the color space), we choose to blit in sRGB space to get |
||
1864 | * this higher quality image. |
||
1865 | */ |
||
1866 | if (src.num_samples > 1 && |
||
1867 | _mesa_get_format_color_encoding(dst_mt->format) == GL_SRGB && |
||
1868 | _mesa_get_srgb_format_linear(src_mt->format) == |
||
1869 | _mesa_get_srgb_format_linear(dst_mt->format)) { |
||
1870 | assert(brw->format_supported_as_render_target[dst_mt->format]); |
||
1871 | dst.brw_surfaceformat = brw->render_target_format[dst_mt->format]; |
||
1872 | src.brw_surfaceformat = brw_format_for_mesa_format(dst_mt->format); |
||
1873 | } |
||
1874 | |||
1875 | /* When doing a multisample resolve of a GL_LUMINANCE32F or GL_INTENSITY32F |
||
1876 | * texture, the above code configures the source format for L32_FLOAT or |
||
1877 | * I32_FLOAT, and the destination format for R32_FLOAT. On Sandy Bridge, |
||
1878 | * the SAMPLE message appears to handle multisampled L32_FLOAT and |
||
1879 | * I32_FLOAT textures incorrectly, resulting in blocky artifacts. So work |
||
1880 | * around the problem by using a source format of R32_FLOAT. This |
||
1881 | * shouldn't affect rendering correctness, since the destination format is |
||
1882 | * R32_FLOAT, so only the contents of the red channel matters. |
||
1883 | */ |
||
1884 | if (brw->gen == 6 && src.num_samples > 1 && dst.num_samples <= 1 && |
||
1885 | src_mt->format == dst_mt->format && |
||
1886 | dst.brw_surfaceformat == BRW_SURFACEFORMAT_R32_FLOAT) { |
||
1887 | src.brw_surfaceformat = dst.brw_surfaceformat; |
||
1888 | } |
||
1889 | |||
1890 | use_wm_prog = true; |
||
1891 | memset(&wm_prog_key, 0, sizeof(wm_prog_key)); |
||
1892 | |||
1893 | /* texture_data_type indicates the register type that should be used to |
||
1894 | * manipulate texture data. |
||
1895 | */ |
||
1896 | switch (_mesa_get_format_datatype(src_mt->format)) { |
||
1897 | case GL_UNSIGNED_NORMALIZED: |
||
1898 | case GL_SIGNED_NORMALIZED: |
||
1899 | case GL_FLOAT: |
||
1900 | wm_prog_key.texture_data_type = BRW_REGISTER_TYPE_F; |
||
1901 | break; |
||
1902 | case GL_UNSIGNED_INT: |
||
1903 | if (src_mt->format == MESA_FORMAT_S_UINT8) { |
||
1904 | /* We process stencil as though it's an unsigned normalized color */ |
||
1905 | wm_prog_key.texture_data_type = BRW_REGISTER_TYPE_F; |
||
1906 | } else { |
||
1907 | wm_prog_key.texture_data_type = BRW_REGISTER_TYPE_UD; |
||
1908 | } |
||
1909 | break; |
||
1910 | case GL_INT: |
||
1911 | wm_prog_key.texture_data_type = BRW_REGISTER_TYPE_D; |
||
1912 | break; |
||
1913 | default: |
||
1914 | unreachable("Unrecognized blorp format"); |
||
1915 | } |
||
1916 | |||
1917 | if (brw->gen > 6) { |
||
1918 | /* Gen7's rendering hardware only supports the IMS layout for depth and |
||
1919 | * stencil render targets. Blorp always maps its destination surface as |
||
1920 | * a color render target (even if it's actually a depth or stencil |
||
1921 | * buffer). So if the destination is IMS, we'll have to map it as a |
||
1922 | * single-sampled texture and interleave the samples ourselves. |
||
1923 | */ |
||
1924 | if (dst_mt->msaa_layout == INTEL_MSAA_LAYOUT_IMS) |
||
1925 | dst.num_samples = 0; |
||
1926 | } |
||
1927 | |||
1928 | if (dst.map_stencil_as_y_tiled && dst.num_samples > 1) { |
||
1929 | /* If the destination surface is a W-tiled multisampled stencil buffer |
||
1930 | * that we're mapping as Y tiled, then we need to arrange for the WM |
||
1931 | * program to run once per sample rather than once per pixel, because |
||
1932 | * the memory layout of related samples doesn't match between W and Y |
||
1933 | * tiling. |
||
1934 | */ |
||
1935 | wm_prog_key.persample_msaa_dispatch = true; |
||
1936 | } |
||
1937 | |||
1938 | if (src.num_samples > 0 && dst.num_samples > 1) { |
||
1939 | /* We are blitting from a multisample buffer to a multisample buffer, so |
||
1940 | * we must preserve samples within a pixel. This means we have to |
||
1941 | * arrange for the WM program to run once per sample rather than once |
||
1942 | * per pixel. |
||
1943 | */ |
||
1944 | wm_prog_key.persample_msaa_dispatch = true; |
||
1945 | } |
||
1946 | |||
1947 | /* Scaled blitting or not. */ |
||
1948 | wm_prog_key.blit_scaled = |
||
1949 | ((dst_x1 - dst_x0) == (src_x1 - src_x0) && |
||
1950 | (dst_y1 - dst_y0) == (src_y1 - src_y0)) ? false : true; |
||
1951 | |||
1952 | /* Scaling factors used for bilinear filtering in multisample scaled |
||
1953 | * blits. |
||
1954 | */ |
||
1955 | wm_prog_key.x_scale = 2.0; |
||
1956 | wm_prog_key.y_scale = src_mt->num_samples / 2.0; |
||
1957 | |||
1958 | if (filter == GL_LINEAR && src.num_samples <= 1 && dst.num_samples <= 1) |
||
1959 | wm_prog_key.bilinear_filter = true; |
||
1960 | |||
1961 | GLenum base_format = _mesa_get_format_base_format(src_mt->format); |
||
1962 | if (base_format != GL_DEPTH_COMPONENT && /* TODO: what about depth/stencil? */ |
||
1963 | base_format != GL_STENCIL_INDEX && |
||
1964 | src_mt->num_samples > 1 && dst_mt->num_samples <= 1) { |
||
1965 | /* We are downsampling a color buffer, so blend. */ |
||
1966 | wm_prog_key.blend = true; |
||
1967 | } |
||
1968 | |||
1969 | /* src_samples and dst_samples are the true sample counts */ |
||
1970 | wm_prog_key.src_samples = src_mt->num_samples; |
||
1971 | wm_prog_key.dst_samples = dst_mt->num_samples; |
||
1972 | |||
1973 | /* tex_samples and rt_samples are the sample counts that are set up in |
||
1974 | * SURFACE_STATE. |
||
1975 | */ |
||
1976 | wm_prog_key.tex_samples = src.num_samples; |
||
1977 | wm_prog_key.rt_samples = dst.num_samples; |
||
1978 | |||
1979 | /* tex_layout and rt_layout indicate the MSAA layout the GPU pipeline will |
||
1980 | * use to access the source and destination surfaces. |
||
1981 | */ |
||
1982 | wm_prog_key.tex_layout = |
||
1983 | compute_msaa_layout_for_pipeline(brw, src.num_samples, src.msaa_layout); |
||
1984 | wm_prog_key.rt_layout = |
||
1985 | compute_msaa_layout_for_pipeline(brw, dst.num_samples, dst.msaa_layout); |
||
1986 | |||
1987 | /* src_layout and dst_layout indicate the true MSAA layout used by src and |
||
1988 | * dst. |
||
1989 | */ |
||
1990 | wm_prog_key.src_layout = src_mt->msaa_layout; |
||
1991 | wm_prog_key.dst_layout = dst_mt->msaa_layout; |
||
1992 | |||
1993 | wm_prog_key.src_tiled_w = src.map_stencil_as_y_tiled; |
||
1994 | wm_prog_key.dst_tiled_w = dst.map_stencil_as_y_tiled; |
||
1995 | /* Round floating point values to nearest integer to avoid "off by one texel" |
||
1996 | * kind of errors when blitting. |
||
1997 | */ |
||
1998 | x0 = wm_push_consts.dst_x0 = roundf(dst_x0); |
||
1999 | y0 = wm_push_consts.dst_y0 = roundf(dst_y0); |
||
2000 | x1 = wm_push_consts.dst_x1 = roundf(dst_x1); |
||
2001 | y1 = wm_push_consts.dst_y1 = roundf(dst_y1); |
||
2002 | wm_push_consts.rect_grid_x1 = (minify(src_mt->logical_width0, src_level) * |
||
2003 | wm_prog_key.x_scale - 1.0); |
||
2004 | wm_push_consts.rect_grid_y1 = (minify(src_mt->logical_height0, src_level) * |
||
2005 | wm_prog_key.y_scale - 1.0); |
||
2006 | |||
2007 | wm_push_consts.x_transform.setup(src_x0, src_x1, dst_x0, dst_x1, mirror_x); |
||
2008 | wm_push_consts.y_transform.setup(src_y0, src_y1, dst_y0, dst_y1, mirror_y); |
||
2009 | |||
2010 | if (dst.num_samples <= 1 && dst_mt->num_samples > 1) { |
||
2011 | /* We must expand the rectangle we send through the rendering pipeline, |
||
2012 | * to account for the fact that we are mapping the destination region as |
||
2013 | * single-sampled when it is in fact multisampled. We must also align |
||
2014 | * it to a multiple of the multisampling pattern, because the |
||
2015 | * differences between multisampled and single-sampled surface formats |
||
2016 | * will mean that pixels are scrambled within the multisampling pattern. |
||
2017 | * TODO: what if this makes the coordinates too large? |
||
2018 | * |
||
2019 | * Note: this only works if the destination surface uses the IMS layout. |
||
2020 | * If it's UMS, then we have no choice but to set up the rendering |
||
2021 | * pipeline as multisampled. |
||
2022 | */ |
||
2023 | assert(dst_mt->msaa_layout == INTEL_MSAA_LAYOUT_IMS); |
||
2024 | switch (dst_mt->num_samples) { |
||
2025 | case 4: |
||
2026 | x0 = ROUND_DOWN_TO(x0 * 2, 4); |
||
2027 | y0 = ROUND_DOWN_TO(y0 * 2, 4); |
||
2028 | x1 = ALIGN(x1 * 2, 4); |
||
2029 | y1 = ALIGN(y1 * 2, 4); |
||
2030 | break; |
||
2031 | case 8: |
||
2032 | x0 = ROUND_DOWN_TO(x0 * 4, 8); |
||
2033 | y0 = ROUND_DOWN_TO(y0 * 2, 4); |
||
2034 | x1 = ALIGN(x1 * 4, 8); |
||
2035 | y1 = ALIGN(y1 * 2, 4); |
||
2036 | break; |
||
2037 | default: |
||
2038 | unreachable("Unrecognized sample count in brw_blorp_blit_params ctor"); |
||
2039 | } |
||
2040 | wm_prog_key.use_kill = true; |
||
2041 | } |
||
2042 | |||
2043 | if (dst.map_stencil_as_y_tiled) { |
||
2044 | /* We must modify the rectangle we send through the rendering pipeline |
||
2045 | * (and the size and x/y offset of the destination surface), to account |
||
2046 | * for the fact that we are mapping it as Y-tiled when it is in fact |
||
2047 | * W-tiled. |
||
2048 | * |
||
2049 | * Both Y tiling and W tiling can be understood as organizations of |
||
2050 | * 32-byte sub-tiles; within each 32-byte sub-tile, the layout of pixels |
||
2051 | * is different, but the layout of the 32-byte sub-tiles within the 4k |
||
2052 | * tile is the same (8 sub-tiles across by 16 sub-tiles down, in |
||
2053 | * column-major order). In Y tiling, the sub-tiles are 16 bytes wide |
||
2054 | * and 2 rows high; in W tiling, they are 8 bytes wide and 4 rows high. |
||
2055 | * |
||
2056 | * Therefore, to account for the layout differences within the 32-byte |
||
2057 | * sub-tiles, we must expand the rectangle so the X coordinates of its |
||
2058 | * edges are multiples of 8 (the W sub-tile width), and its Y |
||
2059 | * coordinates of its edges are multiples of 4 (the W sub-tile height). |
||
2060 | * Then we need to scale the X and Y coordinates of the rectangle to |
||
2061 | * account for the differences in aspect ratio between the Y and W |
||
2062 | * sub-tiles. We need to modify the layer width and height similarly. |
||
2063 | * |
||
2064 | * A correction needs to be applied when MSAA is in use: since |
||
2065 | * INTEL_MSAA_LAYOUT_IMS uses an interleaving pattern whose height is 4, |
||
2066 | * we need to align the Y coordinates to multiples of 8, so that when |
||
2067 | * they are divided by two they are still multiples of 4. |
||
2068 | * |
||
2069 | * Note: Since the x/y offset of the surface will be applied using the |
||
2070 | * SURFACE_STATE command packet, it will be invisible to the swizzling |
||
2071 | * code in the shader; therefore it needs to be in a multiple of the |
||
2072 | * 32-byte sub-tile size. Fortunately it is, since the sub-tile is 8 |
||
2073 | * pixels wide and 4 pixels high (when viewed as a W-tiled stencil |
||
2074 | * buffer), and the miplevel alignment used for stencil buffers is 8 |
||
2075 | * pixels horizontally and either 4 or 8 pixels vertically (see |
||
2076 | * intel_horizontal_texture_alignment_unit() and |
||
2077 | * intel_vertical_texture_alignment_unit()). |
||
2078 | * |
||
2079 | * Note: Also, since the SURFACE_STATE command packet can only apply |
||
2080 | * offsets that are multiples of 4 pixels horizontally and 2 pixels |
||
2081 | * vertically, it is important that the offsets will be multiples of |
||
2082 | * these sizes after they are converted into Y-tiled coordinates. |
||
2083 | * Fortunately they will be, since we know from above that the offsets |
||
2084 | * are a multiple of the 32-byte sub-tile size, and in Y-tiled |
||
2085 | * coordinates the sub-tile is 16 pixels wide and 2 pixels high. |
||
2086 | * |
||
2087 | * TODO: what if this makes the coordinates (or the texture size) too |
||
2088 | * large? |
||
2089 | */ |
||
2090 | const unsigned x_align = 8, y_align = dst.num_samples != 0 ? 8 : 4; |
||
2091 | x0 = ROUND_DOWN_TO(x0, x_align) * 2; |
||
2092 | y0 = ROUND_DOWN_TO(y0, y_align) / 2; |
||
2093 | x1 = ALIGN(x1, x_align) * 2; |
||
2094 | y1 = ALIGN(y1, y_align) / 2; |
||
2095 | dst.width = ALIGN(dst.width, x_align) * 2; |
||
2096 | dst.height = ALIGN(dst.height, y_align) / 2; |
||
2097 | dst.x_offset *= 2; |
||
2098 | dst.y_offset /= 2; |
||
2099 | wm_prog_key.use_kill = true; |
||
2100 | } |
||
2101 | |||
2102 | if (src.map_stencil_as_y_tiled) { |
||
2103 | /* We must modify the size and x/y offset of the source surface to |
||
2104 | * account for the fact that we are mapping it as Y-tiled when it is in |
||
2105 | * fact W tiled. |
||
2106 | * |
||
2107 | * See the comments above concerning x/y offset alignment for the |
||
2108 | * destination surface. |
||
2109 | * |
||
2110 | * TODO: what if this makes the texture size too large? |
||
2111 | */ |
||
2112 | const unsigned x_align = 8, y_align = src.num_samples != 0 ? 8 : 4; |
||
2113 | src.width = ALIGN(src.width, x_align) * 2; |
||
2114 | src.height = ALIGN(src.height, y_align) / 2; |
||
2115 | src.x_offset *= 2; |
||
2116 | src.y_offset /= 2; |
||
2117 | } |
||
2118 | } |
||
2119 | |||
2120 | uint32_t |
||
2121 | brw_blorp_blit_params::get_wm_prog(struct brw_context *brw, |
||
2122 | brw_blorp_prog_data **prog_data) const |
||
2123 | { |
||
2124 | uint32_t prog_offset = 0; |
||
2125 | if (!brw_search_cache(&brw->cache, BRW_CACHE_BLORP_BLIT_PROG, |
||
2126 | &this->wm_prog_key, sizeof(this->wm_prog_key), |
||
2127 | &prog_offset, prog_data)) { |
||
2128 | brw_blorp_blit_program prog(brw, &this->wm_prog_key, |
||
2129 | INTEL_DEBUG & DEBUG_BLORP); |
||
2130 | GLuint program_size; |
||
2131 | const GLuint *program = prog.compile(brw, &program_size); |
||
2132 | brw_upload_cache(&brw->cache, BRW_CACHE_BLORP_BLIT_PROG, |
||
2133 | &this->wm_prog_key, sizeof(this->wm_prog_key), |
||
2134 | program, program_size, |
||
2135 | &prog.prog_data, sizeof(prog.prog_data), |
||
2136 | &prog_offset, prog_data); |
||
2137 | } |
||
2138 | return prog_offset; |
||
2139 | }=>=>=>=>=>=>8;8,1>1>>>>>>>>>>>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>2;4,0>2;4,0>2;4,0>=>>>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>><>>>>> |