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 
27 #include "glsl/ralloc.h"
28 
29 #include "intel_fbo.h"
30 
31 #include "brw_blorp.h"
32 #include "brw_context.h"
33 #include "brw_eu.h"
34 #include "brw_state.h"
35 
36 
37 /**
38  * Helper function for handling mirror image blits.
39  *
40  * If coord0 > coord1, swap them and invert the "mirror" boolean.
41  */
42 static inline void
fixup_mirroring(bool & mirror,GLint & coord0,GLint & coord1)43 fixup_mirroring(bool &mirror, GLint &coord0, GLint &coord1)
44 {
45    if (coord0 > coord1) {
46       mirror = !mirror;
47       GLint tmp = coord0;
48       coord0 = coord1;
49       coord1 = tmp;
50    }
51 }
52 
53 
54 /**
55  * Adjust {src,dst}_x{0,1} to account for clipping and scissoring of
56  * destination coordinates.
57  *
58  * Return true if there is still blitting to do, false if all pixels got
59  * rejected by the clip and/or scissor.
60  *
61  * For clarity, the nomenclature of this function assumes we are clipping and
62  * scissoring the X coordinate; the exact same logic applies for Y
63  * coordinates.
64  *
65  * Note: this function may also be used to account for clipping of source
66  * coordinates, by swapping the roles of src and dst.
67  */
68 static inline bool
clip_or_scissor(bool mirror,GLint & src_x0,GLint & src_x1,GLint & dst_x0,GLint & dst_x1,GLint fb_xmin,GLint fb_xmax)69 clip_or_scissor(bool mirror, GLint &src_x0, GLint &src_x1, GLint &dst_x0,
70                 GLint &dst_x1, GLint fb_xmin, GLint fb_xmax)
71 {
72    /* If we are going to scissor everything away, stop. */
73    if (!(fb_xmin < fb_xmax &&
74          dst_x0 < fb_xmax &&
75          fb_xmin < dst_x1 &&
76          dst_x0 < dst_x1)) {
77       return false;
78    }
79 
80    /* Clip the destination rectangle, and keep track of how many pixels we
81     * clipped off of the left and right sides of it.
82     */
83    GLint pixels_clipped_left = 0;
84    GLint pixels_clipped_right = 0;
85    if (dst_x0 < fb_xmin) {
86       pixels_clipped_left = fb_xmin - dst_x0;
87       dst_x0 = fb_xmin;
88    }
89    if (fb_xmax < dst_x1) {
90       pixels_clipped_right = dst_x1 - fb_xmax;
91       dst_x1 = fb_xmax;
92    }
93 
94    /* If we are mirrored, then before applying pixels_clipped_{left,right} to
95     * the source coordinates, we need to flip them to account for the
96     * mirroring.
97     */
98    if (mirror) {
99       GLint tmp = pixels_clipped_left;
100       pixels_clipped_left = pixels_clipped_right;
101       pixels_clipped_right = tmp;
102    }
103 
104    /* Adjust the source rectangle to remove the pixels corresponding to those
105     * that were clipped/scissored out of the destination rectangle.
106     */
107    src_x0 += pixels_clipped_left;
108    src_x1 -= pixels_clipped_right;
109 
110    return true;
111 }
112 
113 
114 static struct intel_mipmap_tree *
find_miptree(GLbitfield buffer_bit,struct intel_renderbuffer * irb)115 find_miptree(GLbitfield buffer_bit, struct intel_renderbuffer *irb)
116 {
117    struct intel_mipmap_tree *mt = irb->mt;
118    if (buffer_bit == GL_STENCIL_BUFFER_BIT && mt->stencil_mt)
119       mt = mt->stencil_mt;
120    return mt;
121 }
122 
123 void
brw_blorp_blit_miptrees(struct intel_context * intel,struct intel_mipmap_tree * src_mt,unsigned src_level,unsigned src_layer,struct intel_mipmap_tree * dst_mt,unsigned dst_level,unsigned dst_layer,int src_x0,int src_y0,int dst_x0,int dst_y0,int dst_x1,int dst_y1,bool mirror_x,bool mirror_y)124 brw_blorp_blit_miptrees(struct intel_context *intel,
125                         struct intel_mipmap_tree *src_mt,
126                         unsigned src_level, unsigned src_layer,
127                         struct intel_mipmap_tree *dst_mt,
128                         unsigned dst_level, unsigned dst_layer,
129                         int src_x0, int src_y0,
130                         int dst_x0, int dst_y0,
131                         int dst_x1, int dst_y1,
132                         bool mirror_x, bool mirror_y)
133 {
134    brw_blorp_blit_params params(brw_context(&intel->ctx),
135                                 src_mt, src_level, src_layer,
136                                 dst_mt, dst_level, dst_layer,
137                                 src_x0, src_y0,
138                                 dst_x0, dst_y0,
139                                 dst_x1, dst_y1,
140                                 mirror_x, mirror_y);
141    brw_blorp_exec(intel, &params);
142 }
143 
144 static void
do_blorp_blit(struct intel_context * intel,GLbitfield buffer_bit,struct intel_renderbuffer * src_irb,struct intel_renderbuffer * dst_irb,GLint srcX0,GLint srcY0,GLint dstX0,GLint dstY0,GLint dstX1,GLint dstY1,bool mirror_x,bool mirror_y)145 do_blorp_blit(struct intel_context *intel, GLbitfield buffer_bit,
146               struct intel_renderbuffer *src_irb,
147               struct intel_renderbuffer *dst_irb,
148               GLint srcX0, GLint srcY0,
149               GLint dstX0, GLint dstY0, GLint dstX1, GLint dstY1,
150               bool mirror_x, bool mirror_y)
151 {
152    /* Find source/dst miptrees */
153    struct intel_mipmap_tree *src_mt = find_miptree(buffer_bit, src_irb);
154    struct intel_mipmap_tree *dst_mt = find_miptree(buffer_bit, dst_irb);
155 
156    /* Get ready to blit.  This includes depth resolving the src and dst
157     * buffers if necessary.
158     */
159    intel_renderbuffer_resolve_depth(intel, src_irb);
160    intel_renderbuffer_resolve_depth(intel, dst_irb);
161 
162    /* Do the blit */
163    brw_blorp_blit_miptrees(intel,
164                            src_mt, src_irb->mt_level, src_irb->mt_layer,
165                            dst_mt, dst_irb->mt_level, dst_irb->mt_layer,
166                            srcX0, srcY0, dstX0, dstY0, dstX1, dstY1,
167                            mirror_x, mirror_y);
168 
169    intel_renderbuffer_set_needs_hiz_resolve(dst_irb);
170    intel_renderbuffer_set_needs_downsample(dst_irb);
171 }
172 
173 
174 static bool
formats_match(GLbitfield buffer_bit,struct intel_renderbuffer * src_irb,struct intel_renderbuffer * dst_irb)175 formats_match(GLbitfield buffer_bit, struct intel_renderbuffer *src_irb,
176               struct intel_renderbuffer *dst_irb)
177 {
178    /* Note: don't just check gl_renderbuffer::Format, because in some cases
179     * multiple gl_formats resolve to the same native type in the miptree (for
180     * example MESA_FORMAT_X8_Z24 and MESA_FORMAT_S8_Z24), and we can blit
181     * between those formats.
182     */
183    gl_format src_format = find_miptree(buffer_bit, src_irb)->format;
184    gl_format dst_format = find_miptree(buffer_bit, dst_irb)->format;
185 
186    return _mesa_get_srgb_format_linear(src_format) ==
187           _mesa_get_srgb_format_linear(dst_format);
188 }
189 
190 
191 static bool
try_blorp_blit(struct intel_context * intel,GLint srcX0,GLint srcY0,GLint srcX1,GLint srcY1,GLint dstX0,GLint dstY0,GLint dstX1,GLint dstY1,GLenum filter,GLbitfield buffer_bit)192 try_blorp_blit(struct intel_context *intel,
193                GLint srcX0, GLint srcY0, GLint srcX1, GLint srcY1,
194                GLint dstX0, GLint dstY0, GLint dstX1, GLint dstY1,
195                GLenum filter, GLbitfield buffer_bit)
196 {
197    struct gl_context *ctx = &intel->ctx;
198 
199    /* Sync up the state of window system buffers.  We need to do this before
200     * we go looking for the buffers.
201     */
202    intel_prepare_render(intel);
203 
204    const struct gl_framebuffer *read_fb = ctx->ReadBuffer;
205    const struct gl_framebuffer *draw_fb = ctx->DrawBuffer;
206 
207    /* Detect if the blit needs to be mirrored */
208    bool mirror_x = false, mirror_y = false;
209    fixup_mirroring(mirror_x, srcX0, srcX1);
210    fixup_mirroring(mirror_x, dstX0, dstX1);
211    fixup_mirroring(mirror_y, srcY0, srcY1);
212    fixup_mirroring(mirror_y, dstY0, dstY1);
213 
214    /* Make sure width and height match */
215    if (srcX1 - srcX0 != dstX1 - dstX0) return false;
216    if (srcY1 - srcY0 != dstY1 - dstY0) return false;
217 
218    /* If the destination rectangle needs to be clipped or scissored, do so.
219     */
220    if (!(clip_or_scissor(mirror_x, srcX0, srcX1, dstX0, dstX1,
221                          draw_fb->_Xmin, draw_fb->_Xmax) &&
222          clip_or_scissor(mirror_y, srcY0, srcY1, dstY0, dstY1,
223                          draw_fb->_Ymin, draw_fb->_Ymax))) {
224       /* Everything got clipped/scissored away, so the blit was successful. */
225       return true;
226    }
227 
228    /* If the source rectangle needs to be clipped or scissored, do so. */
229    if (!(clip_or_scissor(mirror_x, dstX0, dstX1, srcX0, srcX1,
230                          0, read_fb->Width) &&
231          clip_or_scissor(mirror_y, dstY0, dstY1, srcY0, srcY1,
232                          0, read_fb->Height))) {
233       /* Everything got clipped/scissored away, so the blit was successful. */
234       return true;
235    }
236 
237    /* Account for the fact that in the system framebuffer, the origin is at
238     * the lower left.
239     */
240    if (_mesa_is_winsys_fbo(read_fb)) {
241       GLint tmp = read_fb->Height - srcY0;
242       srcY0 = read_fb->Height - srcY1;
243       srcY1 = tmp;
244       mirror_y = !mirror_y;
245    }
246    if (_mesa_is_winsys_fbo(draw_fb)) {
247       GLint tmp = draw_fb->Height - dstY0;
248       dstY0 = draw_fb->Height - dstY1;
249       dstY1 = tmp;
250       mirror_y = !mirror_y;
251    }
252 
253    /* Find buffers */
254    struct intel_renderbuffer *src_irb;
255    struct intel_renderbuffer *dst_irb;
256    switch (buffer_bit) {
257    case GL_COLOR_BUFFER_BIT:
258       src_irb = intel_renderbuffer(read_fb->_ColorReadBuffer);
259       for (unsigned i = 0; i < ctx->DrawBuffer->_NumColorDrawBuffers; ++i) {
260          dst_irb = intel_renderbuffer(ctx->DrawBuffer->_ColorDrawBuffers[i]);
261          if (dst_irb && !formats_match(buffer_bit, src_irb, dst_irb))
262             return false;
263       }
264       for (unsigned i = 0; i < ctx->DrawBuffer->_NumColorDrawBuffers; ++i) {
265          dst_irb = intel_renderbuffer(ctx->DrawBuffer->_ColorDrawBuffers[i]);
266          do_blorp_blit(intel, buffer_bit, src_irb, dst_irb, srcX0, srcY0,
267                        dstX0, dstY0, dstX1, dstY1, mirror_x, mirror_y);
268       }
269       break;
270    case GL_DEPTH_BUFFER_BIT:
271       src_irb =
272          intel_renderbuffer(read_fb->Attachment[BUFFER_DEPTH].Renderbuffer);
273       dst_irb =
274          intel_renderbuffer(draw_fb->Attachment[BUFFER_DEPTH].Renderbuffer);
275       if (!formats_match(buffer_bit, src_irb, dst_irb))
276          return false;
277       do_blorp_blit(intel, buffer_bit, src_irb, dst_irb, srcX0, srcY0,
278                     dstX0, dstY0, dstX1, dstY1, mirror_x, mirror_y);
279       break;
280    case GL_STENCIL_BUFFER_BIT:
281       src_irb =
282          intel_renderbuffer(read_fb->Attachment[BUFFER_STENCIL].Renderbuffer);
283       dst_irb =
284          intel_renderbuffer(draw_fb->Attachment[BUFFER_STENCIL].Renderbuffer);
285       if (!formats_match(buffer_bit, src_irb, dst_irb))
286          return false;
287       do_blorp_blit(intel, buffer_bit, src_irb, dst_irb, srcX0, srcY0,
288                     dstX0, dstY0, dstX1, dstY1, mirror_x, mirror_y);
289       break;
290    default:
291       assert(false);
292    }
293 
294    return true;
295 }
296 
297 GLbitfield
brw_blorp_framebuffer(struct intel_context * intel,GLint srcX0,GLint srcY0,GLint srcX1,GLint srcY1,GLint dstX0,GLint dstY0,GLint dstX1,GLint dstY1,GLbitfield mask,GLenum filter)298 brw_blorp_framebuffer(struct intel_context *intel,
299                       GLint srcX0, GLint srcY0, GLint srcX1, GLint srcY1,
300                       GLint dstX0, GLint dstY0, GLint dstX1, GLint dstY1,
301                       GLbitfield mask, GLenum filter)
302 {
303    /* BLORP is not supported before Gen6. */
304    if (intel->gen < 6)
305       return mask;
306 
307    static GLbitfield buffer_bits[] = {
308       GL_COLOR_BUFFER_BIT,
309       GL_DEPTH_BUFFER_BIT,
310       GL_STENCIL_BUFFER_BIT,
311    };
312 
313    for (unsigned int i = 0; i < ARRAY_SIZE(buffer_bits); ++i) {
314       if ((mask & buffer_bits[i]) &&
315        try_blorp_blit(intel,
316                       srcX0, srcY0, srcX1, srcY1,
317                       dstX0, dstY0, dstX1, dstY1,
318                       filter, buffer_bits[i])) {
319          mask &= ~buffer_bits[i];
320       }
321    }
322 
323    return mask;
324 }
325 
326 
327 /**
328  * Enum to specify the order of arguments in a sampler message
329  */
330 enum sampler_message_arg
331 {
332    SAMPLER_MESSAGE_ARG_U_FLOAT,
333    SAMPLER_MESSAGE_ARG_V_FLOAT,
334    SAMPLER_MESSAGE_ARG_U_INT,
335    SAMPLER_MESSAGE_ARG_V_INT,
336    SAMPLER_MESSAGE_ARG_SI_INT,
337    SAMPLER_MESSAGE_ARG_MCS_INT,
338    SAMPLER_MESSAGE_ARG_ZERO_INT,
339 };
340 
341 /**
342  * Generator for WM programs used in BLORP blits.
343  *
344  * The bulk of the work done by the WM program is to wrap and unwrap the
345  * coordinate transformations used by the hardware to store surfaces in
346  * memory.  The hardware transforms a pixel location (X, Y, S) (where S is the
347  * sample index for a multisampled surface) to a memory offset by the
348  * following formulas:
349  *
350  *   offset = tile(tiling_format, encode_msaa(num_samples, layout, X, Y, S))
351  *   (X, Y, S) = decode_msaa(num_samples, layout, detile(tiling_format, offset))
352  *
353  * For a single-sampled surface, or for a multisampled surface using
354  * INTEL_MSAA_LAYOUT_UMS, encode_msaa() and decode_msaa are the identity
355  * function:
356  *
357  *   encode_msaa(1, NONE, X, Y, 0) = (X, Y, 0)
358  *   decode_msaa(1, NONE, X, Y, 0) = (X, Y, 0)
359  *   encode_msaa(n, UMS, X, Y, S) = (X, Y, S)
360  *   decode_msaa(n, UMS, X, Y, S) = (X, Y, S)
361  *
362  * For a 4x multisampled surface using INTEL_MSAA_LAYOUT_IMS, encode_msaa()
363  * embeds the sample number into bit 1 of the X and Y coordinates:
364  *
365  *   encode_msaa(4, IMS, X, Y, S) = (X', Y', 0)
366  *     where X' = (X & ~0b1) << 1 | (S & 0b1) << 1 | (X & 0b1)
367  *           Y' = (Y & ~0b1 ) << 1 | (S & 0b10) | (Y & 0b1)
368  *   decode_msaa(4, IMS, X, Y, 0) = (X', Y', S)
369  *     where X' = (X & ~0b11) >> 1 | (X & 0b1)
370  *           Y' = (Y & ~0b11) >> 1 | (Y & 0b1)
371  *           S = (Y & 0b10) | (X & 0b10) >> 1
372  *
373  * For an 8x multisampled surface using INTEL_MSAA_LAYOUT_IMS, encode_msaa()
374  * embeds the sample number into bits 1 and 2 of the X coordinate and bit 1 of
375  * the Y coordinate:
376  *
377  *   encode_msaa(8, IMS, X, Y, S) = (X', Y', 0)
378  *     where X' = (X & ~0b1) << 2 | (S & 0b100) | (S & 0b1) << 1 | (X & 0b1)
379  *           Y' = (Y & ~0b1) << 1 | (S & 0b10) | (Y & 0b1)
380  *   decode_msaa(8, IMS, X, Y, 0) = (X', Y', S)
381  *     where X' = (X & ~0b111) >> 2 | (X & 0b1)
382  *           Y' = (Y & ~0b11) >> 1 | (Y & 0b1)
383  *           S = (X & 0b100) | (Y & 0b10) | (X & 0b10) >> 1
384  *
385  * For X tiling, tile() combines together the low-order bits of the X and Y
386  * coordinates in the pattern 0byyyxxxxxxxxx, creating 4k tiles that are 512
387  * bytes wide and 8 rows high:
388  *
389  *   tile(x_tiled, X, Y, S) = A
390  *     where A = tile_num << 12 | offset
391  *           tile_num = (Y' >> 3) * tile_pitch + (X' >> 9)
392  *           offset = (Y' & 0b111) << 9
393  *                    | (X & 0b111111111)
394  *           X' = X * cpp
395  *           Y' = Y + S * qpitch
396  *   detile(x_tiled, A) = (X, Y, S)
397  *     where X = X' / cpp
398  *           Y = Y' % qpitch
399  *           S = Y' / qpitch
400  *           Y' = (tile_num / tile_pitch) << 3
401  *                | (A & 0b111000000000) >> 9
402  *           X' = (tile_num % tile_pitch) << 9
403  *                | (A & 0b111111111)
404  *
405  * (In all tiling formulas, cpp is the number of bytes occupied by a single
406  * sample ("chars per pixel"), tile_pitch is the number of 4k tiles required
407  * to fill the width of the surface, and qpitch is the spacing (in rows)
408  * between array slices).
409  *
410  * For Y tiling, tile() combines together the low-order bits of the X and Y
411  * coordinates in the pattern 0bxxxyyyyyxxxx, creating 4k tiles that are 128
412  * bytes wide and 32 rows high:
413  *
414  *   tile(y_tiled, X, Y, S) = A
415  *     where A = tile_num << 12 | offset
416  *           tile_num = (Y' >> 5) * tile_pitch + (X' >> 7)
417  *           offset = (X' & 0b1110000) << 5
418  *                    | (Y' & 0b11111) << 4
419  *                    | (X' & 0b1111)
420  *           X' = X * cpp
421  *           Y' = Y + S * qpitch
422  *   detile(y_tiled, A) = (X, Y, S)
423  *     where X = X' / cpp
424  *           Y = Y' % qpitch
425  *           S = Y' / qpitch
426  *           Y' = (tile_num / tile_pitch) << 5
427  *                | (A & 0b111110000) >> 4
428  *           X' = (tile_num % tile_pitch) << 7
429  *                | (A & 0b111000000000) >> 5
430  *                | (A & 0b1111)
431  *
432  * For W tiling, tile() combines together the low-order bits of the X and Y
433  * coordinates in the pattern 0bxxxyyyyxyxyx, creating 4k tiles that are 64
434  * bytes wide and 64 rows high (note that W tiling is only used for stencil
435  * buffers, which always have cpp = 1 and S=0):
436  *
437  *   tile(w_tiled, X, Y, S) = A
438  *     where A = tile_num << 12 | offset
439  *           tile_num = (Y' >> 6) * tile_pitch + (X' >> 6)
440  *           offset = (X' & 0b111000) << 6
441  *                    | (Y' & 0b111100) << 3
442  *                    | (X' & 0b100) << 2
443  *                    | (Y' & 0b10) << 2
444  *                    | (X' & 0b10) << 1
445  *                    | (Y' & 0b1) << 1
446  *                    | (X' & 0b1)
447  *           X' = X * cpp = X
448  *           Y' = Y + S * qpitch
449  *   detile(w_tiled, A) = (X, Y, S)
450  *     where X = X' / cpp = X'
451  *           Y = Y' % qpitch = Y'
452  *           S = Y / qpitch = 0
453  *           Y' = (tile_num / tile_pitch) << 6
454  *                | (A & 0b111100000) >> 3
455  *                | (A & 0b1000) >> 2
456  *                | (A & 0b10) >> 1
457  *           X' = (tile_num % tile_pitch) << 6
458  *                | (A & 0b111000000000) >> 6
459  *                | (A & 0b10000) >> 2
460  *                | (A & 0b100) >> 1
461  *                | (A & 0b1)
462  *
463  * Finally, for a non-tiled surface, tile() simply combines together the X and
464  * Y coordinates in the natural way:
465  *
466  *   tile(untiled, X, Y, S) = A
467  *     where A = Y * pitch + X'
468  *           X' = X * cpp
469  *           Y' = Y + S * qpitch
470  *   detile(untiled, A) = (X, Y, S)
471  *     where X = X' / cpp
472  *           Y = Y' % qpitch
473  *           S = Y' / qpitch
474  *           X' = A % pitch
475  *           Y' = A / pitch
476  *
477  * (In these formulas, pitch is the number of bytes occupied by a single row
478  * of samples).
479  */
480 class brw_blorp_blit_program
481 {
482 public:
483    brw_blorp_blit_program(struct brw_context *brw,
484                           const brw_blorp_blit_prog_key *key);
485    ~brw_blorp_blit_program();
486 
487    const GLuint *compile(struct brw_context *brw, GLuint *program_size);
488 
489    brw_blorp_prog_data prog_data;
490 
491 private:
492    void alloc_regs();
493    void alloc_push_const_regs(int base_reg);
494    void compute_frag_coords();
495    void translate_tiling(bool old_tiled_w, bool new_tiled_w);
496    void encode_msaa(unsigned num_samples, intel_msaa_layout layout);
497    void decode_msaa(unsigned num_samples, intel_msaa_layout layout);
498    void kill_if_outside_dst_rect();
499    void translate_dst_to_src();
500    void single_to_blend();
501    void manual_blend(unsigned num_samples);
502    void sample(struct brw_reg dst);
503    void texel_fetch(struct brw_reg dst);
504    void mcs_fetch();
505    void expand_to_32_bits(struct brw_reg src, struct brw_reg dst);
506    void texture_lookup(struct brw_reg dst, GLuint msg_type,
507                        const sampler_message_arg *args, int num_args);
508    void render_target_write();
509 
510    /**
511     * Base-2 logarithm of the maximum number of samples that can be blended.
512     */
513    static const unsigned LOG2_MAX_BLEND_SAMPLES = 3;
514 
515    void *mem_ctx;
516    struct brw_context *brw;
517    const brw_blorp_blit_prog_key *key;
518    struct brw_compile func;
519 
520    /* Thread dispatch header */
521    struct brw_reg R0;
522 
523    /* Pixel X/Y coordinates (always in R1). */
524    struct brw_reg R1;
525 
526    /* Push constants */
527    struct brw_reg dst_x0;
528    struct brw_reg dst_x1;
529    struct brw_reg dst_y0;
530    struct brw_reg dst_y1;
531    struct {
532       struct brw_reg multiplier;
533       struct brw_reg offset;
534    } x_transform, y_transform;
535 
536    /* Data read from texture (4 vec16's per array element) */
537    struct brw_reg texture_data[LOG2_MAX_BLEND_SAMPLES + 1];
538 
539    /* Auxiliary storage for the contents of the MCS surface.
540     *
541     * Since the sampler always returns 8 registers worth of data, this is 8
542     * registers wide, even though we only use the first 2 registers of it.
543     */
544    struct brw_reg mcs_data;
545 
546    /* X coordinates.  We have two of them so that we can perform coordinate
547     * transformations easily.
548     */
549    struct brw_reg x_coords[2];
550 
551    /* Y coordinates.  We have two of them so that we can perform coordinate
552     * transformations easily.
553     */
554    struct brw_reg y_coords[2];
555 
556    /* Which element of x_coords and y_coords is currently in use.
557     */
558    int xy_coord_index;
559 
560    /* True if, at the point in the program currently being compiled, the
561     * sample index is known to be zero.
562     */
563    bool s_is_zero;
564 
565    /* Register storing the sample index when s_is_zero is false. */
566    struct brw_reg sample_index;
567 
568    /* Temporaries */
569    struct brw_reg t1;
570    struct brw_reg t2;
571 
572    /* MRF used for sampling and render target writes */
573    GLuint base_mrf;
574 };
575 
brw_blorp_blit_program(struct brw_context * brw,const brw_blorp_blit_prog_key * key)576 brw_blorp_blit_program::brw_blorp_blit_program(
577       struct brw_context *brw,
578       const brw_blorp_blit_prog_key *key)
579    : mem_ctx(ralloc_context(NULL)),
580      brw(brw),
581      key(key)
582 {
583    brw_init_compile(brw, &func, mem_ctx);
584 }
585 
~brw_blorp_blit_program()586 brw_blorp_blit_program::~brw_blorp_blit_program()
587 {
588    ralloc_free(mem_ctx);
589 }
590 
591 const GLuint *
compile(struct brw_context * brw,GLuint * program_size)592 brw_blorp_blit_program::compile(struct brw_context *brw,
593                                 GLuint *program_size)
594 {
595    /* Sanity checks */
596    if (key->dst_tiled_w && key->rt_samples > 0) {
597       /* If the destination image is W tiled and multisampled, then the thread
598        * must be dispatched once per sample, not once per pixel.  This is
599        * necessary because after conversion between W and Y tiling, there's no
600        * guarantee that all samples corresponding to a single pixel will still
601        * be together.
602        */
603       assert(key->persample_msaa_dispatch);
604    }
605 
606    if (key->blend) {
607       /* We are blending, which means we won't have an opportunity to
608        * translate the tiling and sample count for the texture surface.  So
609        * the surface state for the texture must be configured with the correct
610        * tiling and sample count.
611        */
612       assert(!key->src_tiled_w);
613       assert(key->tex_samples == key->src_samples);
614       assert(key->tex_layout == key->src_layout);
615       assert(key->tex_samples > 0);
616    }
617 
618    if (key->persample_msaa_dispatch) {
619       /* It only makes sense to do persample dispatch if the render target is
620        * configured as multisampled.
621        */
622       assert(key->rt_samples > 0);
623    }
624 
625    /* Make sure layout is consistent with sample count */
626    assert((key->tex_layout == INTEL_MSAA_LAYOUT_NONE) ==
627           (key->tex_samples == 0));
628    assert((key->rt_layout == INTEL_MSAA_LAYOUT_NONE) ==
629           (key->rt_samples == 0));
630    assert((key->src_layout == INTEL_MSAA_LAYOUT_NONE) ==
631           (key->src_samples == 0));
632    assert((key->dst_layout == INTEL_MSAA_LAYOUT_NONE) ==
633           (key->dst_samples == 0));
634 
635    /* Set up prog_data */
636    memset(&prog_data, 0, sizeof(prog_data));
637    prog_data.persample_msaa_dispatch = key->persample_msaa_dispatch;
638 
639    brw_set_compression_control(&func, BRW_COMPRESSION_NONE);
640 
641    alloc_regs();
642    compute_frag_coords();
643 
644    /* Render target and texture hardware don't support W tiling. */
645    const bool rt_tiled_w = false;
646    const bool tex_tiled_w = false;
647 
648    /* The address that data will be written to is determined by the
649     * coordinates supplied to the WM thread and the tiling and sample count of
650     * the render target, according to the formula:
651     *
652     * (X, Y, S) = decode_msaa(rt_samples, detile(rt_tiling, offset))
653     *
654     * If the actual tiling and sample count of the destination surface are not
655     * the same as the configuration of the render target, then these
656     * coordinates are wrong and we have to adjust them to compensate for the
657     * difference.
658     */
659    if (rt_tiled_w != key->dst_tiled_w ||
660        key->rt_samples != key->dst_samples ||
661        key->rt_layout != key->dst_layout) {
662       encode_msaa(key->rt_samples, key->rt_layout);
663       /* Now (X, Y, S) = detile(rt_tiling, offset) */
664       translate_tiling(rt_tiled_w, key->dst_tiled_w);
665       /* Now (X, Y, S) = detile(dst_tiling, offset) */
666       decode_msaa(key->dst_samples, key->dst_layout);
667    }
668 
669    /* Now (X, Y, S) = decode_msaa(dst_samples, detile(dst_tiling, offset)).
670     *
671     * That is: X, Y and S now contain the true coordinates and sample index of
672     * the data that the WM thread should output.
673     *
674     * If we need to kill pixels that are outside the destination rectangle,
675     * now is the time to do it.
676     */
677 
678    if (key->use_kill)
679       kill_if_outside_dst_rect();
680 
681    /* Next, apply a translation to obtain coordinates in the source image. */
682    translate_dst_to_src();
683 
684    /* If the source image is not multisampled, then we want to fetch sample
685     * number 0, because that's the only sample there is.
686     */
687    if (key->src_samples == 0)
688       s_is_zero = true;
689 
690    /* X, Y, and S are now the coordinates of the pixel in the source image
691     * that we want to texture from.  Exception: if we are blending, then S is
692     * irrelevant, because we are going to fetch all samples.
693     */
694    if (key->blend) {
695       if (brw->intel.gen == 6) {
696          /* Gen6 hardware an automatically blend using the SAMPLE message */
697          single_to_blend();
698          sample(texture_data[0]);
699       } else {
700          /* Gen7+ hardware doesn't automaticaly blend. */
701          manual_blend(key->src_samples);
702       }
703    } else {
704       /* We aren't blending, which means we just want to fetch a single sample
705        * from the source surface.  The address that we want to fetch from is
706        * related to the X, Y and S values according to the formula:
707        *
708        * (X, Y, S) = decode_msaa(src_samples, detile(src_tiling, offset)).
709        *
710        * If the actual tiling and sample count of the source surface are not
711        * the same as the configuration of the texture, then we need to adjust
712        * the coordinates to compensate for the difference.
713        */
714       if (tex_tiled_w != key->src_tiled_w ||
715           key->tex_samples != key->src_samples ||
716           key->tex_layout != key->src_layout) {
717          encode_msaa(key->src_samples, key->src_layout);
718          /* Now (X, Y, S) = detile(src_tiling, offset) */
719          translate_tiling(key->src_tiled_w, tex_tiled_w);
720          /* Now (X, Y, S) = detile(tex_tiling, offset) */
721          decode_msaa(key->tex_samples, key->tex_layout);
722       }
723 
724       /* Now (X, Y, S) = decode_msaa(tex_samples, detile(tex_tiling, offset)).
725        *
726        * In other words: X, Y, and S now contain values which, when passed to
727        * the texturing unit, will cause data to be read from the correct
728        * memory location.  So we can fetch the texel now.
729        */
730       if (key->tex_layout == INTEL_MSAA_LAYOUT_CMS)
731          mcs_fetch();
732       texel_fetch(texture_data[0]);
733    }
734 
735    /* Finally, write the fetched (or blended) value to the render target and
736     * terminate the thread.
737     */
738    render_target_write();
739    return brw_get_program(&func, program_size);
740 }
741 
742 void
alloc_push_const_regs(int base_reg)743 brw_blorp_blit_program::alloc_push_const_regs(int base_reg)
744 {
745 #define CONST_LOC(name) offsetof(brw_blorp_wm_push_constants, name)
746 #define ALLOC_REG(name) \
747    this->name = \
748       brw_uw1_reg(BRW_GENERAL_REGISTER_FILE, base_reg, CONST_LOC(name) / 2)
749 
750    ALLOC_REG(dst_x0);
751    ALLOC_REG(dst_x1);
752    ALLOC_REG(dst_y0);
753    ALLOC_REG(dst_y1);
754    ALLOC_REG(x_transform.multiplier);
755    ALLOC_REG(x_transform.offset);
756    ALLOC_REG(y_transform.multiplier);
757    ALLOC_REG(y_transform.offset);
758 #undef CONST_LOC
759 #undef ALLOC_REG
760 }
761 
762 void
alloc_regs()763 brw_blorp_blit_program::alloc_regs()
764 {
765    int reg = 0;
766    this->R0 = retype(brw_vec8_grf(reg++, 0), BRW_REGISTER_TYPE_UW);
767    this->R1 = retype(brw_vec8_grf(reg++, 0), BRW_REGISTER_TYPE_UW);
768    prog_data.first_curbe_grf = reg;
769    alloc_push_const_regs(reg);
770    reg += BRW_BLORP_NUM_PUSH_CONST_REGS;
771    for (unsigned i = 0; i < ARRAY_SIZE(texture_data); ++i) {
772       this->texture_data[i] =
773          retype(vec16(brw_vec8_grf(reg, 0)), key->texture_data_type);
774       reg += 8;
775    }
776    this->mcs_data =
777       retype(brw_vec8_grf(reg, 0), BRW_REGISTER_TYPE_UD); reg += 8;
778    for (int i = 0; i < 2; ++i) {
779       this->x_coords[i]
780          = vec16(retype(brw_vec8_grf(reg++, 0), BRW_REGISTER_TYPE_UW));
781       this->y_coords[i]
782          = vec16(retype(brw_vec8_grf(reg++, 0), BRW_REGISTER_TYPE_UW));
783    }
784    this->xy_coord_index = 0;
785    this->sample_index
786       = vec16(retype(brw_vec8_grf(reg++, 0), BRW_REGISTER_TYPE_UW));
787    this->t1 = vec16(retype(brw_vec8_grf(reg++, 0), BRW_REGISTER_TYPE_UW));
788    this->t2 = vec16(retype(brw_vec8_grf(reg++, 0), BRW_REGISTER_TYPE_UW));
789 
790    /* Make sure we didn't run out of registers */
791    assert(reg <= GEN7_MRF_HACK_START);
792 
793    int mrf = 2;
794    this->base_mrf = mrf;
795 }
796 
797 /* In the code that follows, X and Y can be used to quickly refer to the
798  * active elements of x_coords and y_coords, and Xp and Yp ("X prime" and "Y
799  * prime") to the inactive elements.
800  *
801  * S can be used to quickly refer to sample_index.
802  */
803 #define X x_coords[xy_coord_index]
804 #define Y y_coords[xy_coord_index]
805 #define Xp x_coords[!xy_coord_index]
806 #define Yp y_coords[!xy_coord_index]
807 #define S sample_index
808 
809 /* Quickly swap the roles of (X, Y) and (Xp, Yp).  Saves us from having to do
810  * MOVs to transfor (Xp, Yp) to (X, Y) after a coordinate transformation.
811  */
812 #define SWAP_XY_AND_XPYP() xy_coord_index = !xy_coord_index;
813 
814 /**
815  * Emit code to compute the X and Y coordinates of the pixels being rendered
816  * by this WM invocation.
817  *
818  * Assuming the render target is set up for Y tiling, these (X, Y) values are
819  * related to the address offset where outputs will be written by the formula:
820  *
821  *   (X, Y, S) = decode_msaa(detile(offset)).
822  *
823  * (See brw_blorp_blit_program).
824  */
825 void
compute_frag_coords()826 brw_blorp_blit_program::compute_frag_coords()
827 {
828    /* R1.2[15:0] = X coordinate of upper left pixel of subspan 0 (pixel 0)
829     * R1.3[15:0] = X coordinate of upper left pixel of subspan 1 (pixel 4)
830     * R1.4[15:0] = X coordinate of upper left pixel of subspan 2 (pixel 8)
831     * R1.5[15:0] = X coordinate of upper left pixel of subspan 3 (pixel 12)
832     *
833     * Pixels within a subspan are laid out in this arrangement:
834     * 0 1
835     * 2 3
836     *
837     * So, to compute the coordinates of each pixel, we need to read every 2nd
838     * 16-bit value (vstride=2) from R1, starting at the 4th 16-bit value
839     * (suboffset=4), and duplicate each value 4 times (hstride=0, width=4).
840     * In other words, the data we want to access is R1.4<2;4,0>UW.
841     *
842     * Then, we need to add the repeating sequence (0, 1, 0, 1, ...) to the
843     * result, since pixels n+1 and n+3 are in the right half of the subspan.
844     */
845    brw_ADD(&func, X, stride(suboffset(R1, 4), 2, 4, 0), brw_imm_v(0x10101010));
846 
847    /* Similarly, Y coordinates for subspans come from R1.2[31:16] through
848     * R1.5[31:16], so to get pixel Y coordinates we need to start at the 5th
849     * 16-bit value instead of the 4th (R1.5<2;4,0>UW instead of
850     * R1.4<2;4,0>UW).
851     *
852     * And we need to add the repeating sequence (0, 0, 1, 1, ...), since
853     * pixels n+2 and n+3 are in the bottom half of the subspan.
854     */
855    brw_ADD(&func, Y, stride(suboffset(R1, 5), 2, 4, 0), brw_imm_v(0x11001100));
856 
857    if (key->persample_msaa_dispatch) {
858       switch (key->rt_samples) {
859       case 4:
860          /* The WM will be run in MSDISPMODE_PERSAMPLE with num_samples == 4.
861           * Therefore, subspan 0 will represent sample 0, subspan 1 will
862           * represent sample 1, and so on.
863           *
864           * So we need to populate S with the sequence (0, 0, 0, 0, 1, 1, 1,
865           * 1, 2, 2, 2, 2, 3, 3, 3, 3).  The easiest way to do this is to
866           * populate a temporary variable with the sequence (0, 1, 2, 3), and
867           * then copy from it using vstride=1, width=4, hstride=0.
868           */
869          brw_MOV(&func, t1, brw_imm_v(0x3210));
870          brw_MOV(&func, S, stride(t1, 1, 4, 0));
871          break;
872       case 8: {
873          /* The WM will be run in MSDISPMODE_PERSAMPLE with num_samples == 8.
874           * Therefore, subspan 0 will represent sample N (where N is 0 or 4),
875           * subspan 1 will represent sample 1, and so on.  We can find the
876           * value of N by looking at R0.0 bits 7:6 ("Starting Sample Pair
877           * Index") and multiplying by two (since samples are always delivered
878           * in pairs).  That is, we compute 2*((R0.0 & 0xc0) >> 6) == (R0.0 &
879           * 0xc0) >> 5.
880           *
881           * Then we need to add N to the sequence (0, 0, 0, 0, 1, 1, 1, 1, 2,
882           * 2, 2, 2, 3, 3, 3, 3), which we compute by populating a temporary
883           * variable with the sequence (0, 1, 2, 3), and then reading from it
884           * using vstride=1, width=4, hstride=0.
885           */
886          struct brw_reg t1_ud1 = vec1(retype(t1, BRW_REGISTER_TYPE_UD));
887          struct brw_reg r0_ud1 = vec1(retype(R0, BRW_REGISTER_TYPE_UD));
888          brw_AND(&func, t1_ud1, r0_ud1, brw_imm_ud(0xc0));
889          brw_SHR(&func, t1_ud1, t1_ud1, brw_imm_ud(5));
890          brw_MOV(&func, t2, brw_imm_v(0x3210));
891          brw_ADD(&func, S, retype(t1_ud1, BRW_REGISTER_TYPE_UW),
892                  stride(t2, 1, 4, 0));
893          break;
894       }
895       default:
896          assert(!"Unrecognized sample count in "
897                 "brw_blorp_blit_program::compute_frag_coords()");
898          break;
899       }
900       s_is_zero = false;
901    } else {
902       /* Either the destination surface is single-sampled, or the WM will be
903        * run in MSDISPMODE_PERPIXEL (which causes a single fragment dispatch
904        * per pixel).  In either case, it's not meaningful to compute a sample
905        * value.  Just set it to 0.
906        */
907       s_is_zero = true;
908    }
909 }
910 
911 /**
912  * Emit code to compensate for the difference between Y and W tiling.
913  *
914  * This code modifies the X and Y coordinates according to the formula:
915  *
916  *   (X', Y', S') = detile(new_tiling, tile(old_tiling, X, Y, S))
917  *
918  * (See brw_blorp_blit_program).
919  *
920  * It can only translate between W and Y tiling, so new_tiling and old_tiling
921  * are booleans where true represents W tiling and false represents Y tiling.
922  */
923 void
translate_tiling(bool old_tiled_w,bool new_tiled_w)924 brw_blorp_blit_program::translate_tiling(bool old_tiled_w, bool new_tiled_w)
925 {
926    if (old_tiled_w == new_tiled_w)
927       return;
928 
929    /* In the code that follows, we can safely assume that S = 0, because W
930     * tiling formats always use IMS layout.
931     */
932    assert(s_is_zero);
933 
934    if (new_tiled_w) {
935       /* Given X and Y coordinates that describe an address using Y tiling,
936        * translate to the X and Y coordinates that describe the same address
937        * using W tiling.
938        *
939        * If we break down the low order bits of X and Y, using a
940        * single letter to represent each low-order bit:
941        *
942        *   X = A << 7 | 0bBCDEFGH
943        *   Y = J << 5 | 0bKLMNP                                       (1)
944        *
945        * Then we can apply the Y tiling formula to see the memory offset being
946        * addressed:
947        *
948        *   offset = (J * tile_pitch + A) << 12 | 0bBCDKLMNPEFGH       (2)
949        *
950        * If we apply the W detiling formula to this memory location, that the
951        * corresponding X' and Y' coordinates are:
952        *
953        *   X' = A << 6 | 0bBCDPFH                                     (3)
954        *   Y' = J << 6 | 0bKLMNEG
955        *
956        * Combining (1) and (3), we see that to transform (X, Y) to (X', Y'),
957        * we need to make the following computation:
958        *
959        *   X' = (X & ~0b1011) >> 1 | (Y & 0b1) << 2 | X & 0b1         (4)
960        *   Y' = (Y & ~0b1) << 1 | (X & 0b1000) >> 2 | (X & 0b10) >> 1
961        */
962       brw_AND(&func, t1, X, brw_imm_uw(0xfff4)); /* X & ~0b1011 */
963       brw_SHR(&func, t1, t1, brw_imm_uw(1)); /* (X & ~0b1011) >> 1 */
964       brw_AND(&func, t2, Y, brw_imm_uw(1)); /* Y & 0b1 */
965       brw_SHL(&func, t2, t2, brw_imm_uw(2)); /* (Y & 0b1) << 2 */
966       brw_OR(&func, t1, t1, t2); /* (X & ~0b1011) >> 1 | (Y & 0b1) << 2 */
967       brw_AND(&func, t2, X, brw_imm_uw(1)); /* X & 0b1 */
968       brw_OR(&func, Xp, t1, t2);
969       brw_AND(&func, t1, Y, brw_imm_uw(0xfffe)); /* Y & ~0b1 */
970       brw_SHL(&func, t1, t1, brw_imm_uw(1)); /* (Y & ~0b1) << 1 */
971       brw_AND(&func, t2, X, brw_imm_uw(8)); /* X & 0b1000 */
972       brw_SHR(&func, t2, t2, brw_imm_uw(2)); /* (X & 0b1000) >> 2 */
973       brw_OR(&func, t1, t1, t2); /* (Y & ~0b1) << 1 | (X & 0b1000) >> 2 */
974       brw_AND(&func, t2, X, brw_imm_uw(2)); /* X & 0b10 */
975       brw_SHR(&func, t2, t2, brw_imm_uw(1)); /* (X & 0b10) >> 1 */
976       brw_OR(&func, Yp, t1, t2);
977       SWAP_XY_AND_XPYP();
978    } else {
979       /* Applying the same logic as above, but in reverse, we obtain the
980        * formulas:
981        *
982        * X' = (X & ~0b101) << 1 | (Y & 0b10) << 2 | (Y & 0b1) << 1 | X & 0b1
983        * Y' = (Y & ~0b11) >> 1 | (X & 0b100) >> 2
984        */
985       brw_AND(&func, t1, X, brw_imm_uw(0xfffa)); /* X & ~0b101 */
986       brw_SHL(&func, t1, t1, brw_imm_uw(1)); /* (X & ~0b101) << 1 */
987       brw_AND(&func, t2, Y, brw_imm_uw(2)); /* Y & 0b10 */
988       brw_SHL(&func, t2, t2, brw_imm_uw(2)); /* (Y & 0b10) << 2 */
989       brw_OR(&func, t1, t1, t2); /* (X & ~0b101) << 1 | (Y & 0b10) << 2 */
990       brw_AND(&func, t2, Y, brw_imm_uw(1)); /* Y & 0b1 */
991       brw_SHL(&func, t2, t2, brw_imm_uw(1)); /* (Y & 0b1) << 1 */
992       brw_OR(&func, t1, t1, t2); /* (X & ~0b101) << 1 | (Y & 0b10) << 2
993                                     | (Y & 0b1) << 1 */
994       brw_AND(&func, t2, X, brw_imm_uw(1)); /* X & 0b1 */
995       brw_OR(&func, Xp, t1, t2);
996       brw_AND(&func, t1, Y, brw_imm_uw(0xfffc)); /* Y & ~0b11 */
997       brw_SHR(&func, t1, t1, brw_imm_uw(1)); /* (Y & ~0b11) >> 1 */
998       brw_AND(&func, t2, X, brw_imm_uw(4)); /* X & 0b100 */
999       brw_SHR(&func, t2, t2, brw_imm_uw(2)); /* (X & 0b100) >> 2 */
1000       brw_OR(&func, Yp, t1, t2);
1001       SWAP_XY_AND_XPYP();
1002    }
1003 }
1004 
1005 /**
1006  * Emit code to compensate for the difference between MSAA and non-MSAA
1007  * surfaces.
1008  *
1009  * This code modifies the X and Y coordinates according to the formula:
1010  *
1011  *   (X', Y', S') = encode_msaa(num_samples, IMS, X, Y, S)
1012  *
1013  * (See brw_blorp_blit_program).
1014  */
1015 void
encode_msaa(unsigned num_samples,intel_msaa_layout layout)1016 brw_blorp_blit_program::encode_msaa(unsigned num_samples,
1017                                     intel_msaa_layout layout)
1018 {
1019    switch (layout) {
1020    case INTEL_MSAA_LAYOUT_NONE:
1021       /* No translation necessary, and S should already be zero. */
1022       assert(s_is_zero);
1023       break;
1024    case INTEL_MSAA_LAYOUT_CMS:
1025       /* We can't compensate for compressed layout since at this point in the
1026        * program we haven't read from the MCS buffer.
1027        */
1028       assert(!"Bad layout in encode_msaa");
1029       break;
1030    case INTEL_MSAA_LAYOUT_UMS:
1031       /* No translation necessary. */
1032       break;
1033    case INTEL_MSAA_LAYOUT_IMS:
1034       switch (num_samples) {
1035       case 4:
1036          /* encode_msaa(4, IMS, X, Y, S) = (X', Y', 0)
1037           *   where X' = (X & ~0b1) << 1 | (S & 0b1) << 1 | (X & 0b1)
1038           *         Y' = (Y & ~0b1) << 1 | (S & 0b10) | (Y & 0b1)
1039           */
1040          brw_AND(&func, t1, X, brw_imm_uw(0xfffe)); /* X & ~0b1 */
1041          if (!s_is_zero) {
1042             brw_AND(&func, t2, S, brw_imm_uw(1)); /* S & 0b1 */
1043             brw_OR(&func, t1, t1, t2); /* (X & ~0b1) | (S & 0b1) */
1044          }
1045          brw_SHL(&func, t1, t1, brw_imm_uw(1)); /* (X & ~0b1) << 1
1046                                                    | (S & 0b1) << 1 */
1047          brw_AND(&func, t2, X, brw_imm_uw(1)); /* X & 0b1 */
1048          brw_OR(&func, Xp, t1, t2);
1049          brw_AND(&func, t1, Y, brw_imm_uw(0xfffe)); /* Y & ~0b1 */
1050          brw_SHL(&func, t1, t1, brw_imm_uw(1)); /* (Y & ~0b1) << 1 */
1051          if (!s_is_zero) {
1052             brw_AND(&func, t2, S, brw_imm_uw(2)); /* S & 0b10 */
1053             brw_OR(&func, t1, t1, t2); /* (Y & ~0b1) << 1 | (S & 0b10) */
1054          }
1055          brw_AND(&func, t2, Y, brw_imm_uw(1)); /* Y & 0b1 */
1056          brw_OR(&func, Yp, t1, t2);
1057          break;
1058       case 8:
1059          /* encode_msaa(8, IMS, X, Y, S) = (X', Y', 0)
1060           *   where X' = (X & ~0b1) << 2 | (S & 0b100) | (S & 0b1) << 1
1061           *              | (X & 0b1)
1062           *         Y' = (Y & ~0b1) << 1 | (S & 0b10) | (Y & 0b1)
1063           */
1064          brw_AND(&func, t1, X, brw_imm_uw(0xfffe)); /* X & ~0b1 */
1065          brw_SHL(&func, t1, t1, brw_imm_uw(2)); /* (X & ~0b1) << 2 */
1066          if (!s_is_zero) {
1067             brw_AND(&func, t2, S, brw_imm_uw(4)); /* S & 0b100 */
1068             brw_OR(&func, t1, t1, t2); /* (X & ~0b1) << 2 | (S & 0b100) */
1069             brw_AND(&func, t2, S, brw_imm_uw(1)); /* S & 0b1 */
1070             brw_SHL(&func, t2, t2, brw_imm_uw(1)); /* (S & 0b1) << 1 */
1071             brw_OR(&func, t1, t1, t2); /* (X & ~0b1) << 2 | (S & 0b100)
1072                                           | (S & 0b1) << 1 */
1073          }
1074          brw_AND(&func, t2, X, brw_imm_uw(1)); /* X & 0b1 */
1075          brw_OR(&func, Xp, t1, t2);
1076          brw_AND(&func, t1, Y, brw_imm_uw(0xfffe)); /* Y & ~0b1 */
1077          brw_SHL(&func, t1, t1, brw_imm_uw(1)); /* (Y & ~0b1) << 1 */
1078          if (!s_is_zero) {
1079             brw_AND(&func, t2, S, brw_imm_uw(2)); /* S & 0b10 */
1080             brw_OR(&func, t1, t1, t2); /* (Y & ~0b1) << 1 | (S & 0b10) */
1081          }
1082          brw_AND(&func, t2, Y, brw_imm_uw(1)); /* Y & 0b1 */
1083          brw_OR(&func, Yp, t1, t2);
1084          break;
1085       }
1086       SWAP_XY_AND_XPYP();
1087       s_is_zero = true;
1088       break;
1089    }
1090 }
1091 
1092 /**
1093  * Emit code to compensate for the difference between MSAA and non-MSAA
1094  * surfaces.
1095  *
1096  * This code modifies the X and Y coordinates according to the formula:
1097  *
1098  *   (X', Y', S) = decode_msaa(num_samples, IMS, X, Y, S)
1099  *
1100  * (See brw_blorp_blit_program).
1101  */
1102 void
decode_msaa(unsigned num_samples,intel_msaa_layout layout)1103 brw_blorp_blit_program::decode_msaa(unsigned num_samples,
1104                                     intel_msaa_layout layout)
1105 {
1106    switch (layout) {
1107    case INTEL_MSAA_LAYOUT_NONE:
1108       /* No translation necessary, and S should already be zero. */
1109       assert(s_is_zero);
1110       break;
1111    case INTEL_MSAA_LAYOUT_CMS:
1112       /* We can't compensate for compressed layout since at this point in the
1113        * program we don't have access to the MCS buffer.
1114        */
1115       assert(!"Bad layout in encode_msaa");
1116       break;
1117    case INTEL_MSAA_LAYOUT_UMS:
1118       /* No translation necessary. */
1119       break;
1120    case INTEL_MSAA_LAYOUT_IMS:
1121       assert(s_is_zero);
1122       switch (num_samples) {
1123       case 4:
1124          /* decode_msaa(4, IMS, X, Y, 0) = (X', Y', S)
1125           *   where X' = (X & ~0b11) >> 1 | (X & 0b1)
1126           *         Y' = (Y & ~0b11) >> 1 | (Y & 0b1)
1127           *         S = (Y & 0b10) | (X & 0b10) >> 1
1128           */
1129          brw_AND(&func, t1, X, brw_imm_uw(0xfffc)); /* X & ~0b11 */
1130          brw_SHR(&func, t1, t1, brw_imm_uw(1)); /* (X & ~0b11) >> 1 */
1131          brw_AND(&func, t2, X, brw_imm_uw(1)); /* X & 0b1 */
1132          brw_OR(&func, Xp, t1, t2);
1133          brw_AND(&func, t1, Y, brw_imm_uw(0xfffc)); /* Y & ~0b11 */
1134          brw_SHR(&func, t1, t1, brw_imm_uw(1)); /* (Y & ~0b11) >> 1 */
1135          brw_AND(&func, t2, Y, brw_imm_uw(1)); /* Y & 0b1 */
1136          brw_OR(&func, Yp, t1, t2);
1137          brw_AND(&func, t1, Y, brw_imm_uw(2)); /* Y & 0b10 */
1138          brw_AND(&func, t2, X, brw_imm_uw(2)); /* X & 0b10 */
1139          brw_SHR(&func, t2, t2, brw_imm_uw(1)); /* (X & 0b10) >> 1 */
1140          brw_OR(&func, S, t1, t2);
1141          break;
1142       case 8:
1143          /* decode_msaa(8, IMS, X, Y, 0) = (X', Y', S)
1144           *   where X' = (X & ~0b111) >> 2 | (X & 0b1)
1145           *         Y' = (Y & ~0b11) >> 1 | (Y & 0b1)
1146           *         S = (X & 0b100) | (Y & 0b10) | (X & 0b10) >> 1
1147           */
1148          brw_AND(&func, t1, X, brw_imm_uw(0xfff8)); /* X & ~0b111 */
1149          brw_SHR(&func, t1, t1, brw_imm_uw(2)); /* (X & ~0b111) >> 2 */
1150          brw_AND(&func, t2, X, brw_imm_uw(1)); /* X & 0b1 */
1151          brw_OR(&func, Xp, t1, t2);
1152          brw_AND(&func, t1, Y, brw_imm_uw(0xfffc)); /* Y & ~0b11 */
1153          brw_SHR(&func, t1, t1, brw_imm_uw(1)); /* (Y & ~0b11) >> 1 */
1154          brw_AND(&func, t2, Y, brw_imm_uw(1)); /* Y & 0b1 */
1155          brw_OR(&func, Yp, t1, t2);
1156          brw_AND(&func, t1, X, brw_imm_uw(4)); /* X & 0b100 */
1157          brw_AND(&func, t2, Y, brw_imm_uw(2)); /* Y & 0b10 */
1158          brw_OR(&func, t1, t1, t2); /* (X & 0b100) | (Y & 0b10) */
1159          brw_AND(&func, t2, X, brw_imm_uw(2)); /* X & 0b10 */
1160          brw_SHR(&func, t2, t2, brw_imm_uw(1)); /* (X & 0b10) >> 1 */
1161          brw_OR(&func, S, t1, t2);
1162          break;
1163       }
1164       s_is_zero = false;
1165       SWAP_XY_AND_XPYP();
1166       break;
1167    }
1168 }
1169 
1170 /**
1171  * Emit code that kills pixels whose X and Y coordinates are outside the
1172  * boundary of the rectangle defined by the push constants (dst_x0, dst_y0,
1173  * dst_x1, dst_y1).
1174  */
1175 void
kill_if_outside_dst_rect()1176 brw_blorp_blit_program::kill_if_outside_dst_rect()
1177 {
1178    struct brw_reg f0 = brw_flag_reg();
1179    struct brw_reg g1 = retype(brw_vec1_grf(1, 7), BRW_REGISTER_TYPE_UW);
1180    struct brw_reg null16 = vec16(retype(brw_null_reg(), BRW_REGISTER_TYPE_UW));
1181 
1182    brw_CMP(&func, null16, BRW_CONDITIONAL_GE, X, dst_x0);
1183    brw_CMP(&func, null16, BRW_CONDITIONAL_GE, Y, dst_y0);
1184    brw_CMP(&func, null16, BRW_CONDITIONAL_L, X, dst_x1);
1185    brw_CMP(&func, null16, BRW_CONDITIONAL_L, Y, dst_y1);
1186 
1187    brw_set_predicate_control(&func, BRW_PREDICATE_NONE);
1188    brw_push_insn_state(&func);
1189    brw_set_mask_control(&func, BRW_MASK_DISABLE);
1190    brw_AND(&func, g1, f0, g1);
1191    brw_pop_insn_state(&func);
1192 }
1193 
1194 /**
1195  * Emit code to translate from destination (X, Y) coordinates to source (X, Y)
1196  * coordinates.
1197  */
1198 void
translate_dst_to_src()1199 brw_blorp_blit_program::translate_dst_to_src()
1200 {
1201    brw_MUL(&func, Xp, X, x_transform.multiplier);
1202    brw_MUL(&func, Yp, Y, y_transform.multiplier);
1203    brw_ADD(&func, Xp, Xp, x_transform.offset);
1204    brw_ADD(&func, Yp, Yp, y_transform.offset);
1205    SWAP_XY_AND_XPYP();
1206 }
1207 
1208 /**
1209  * Emit code to transform the X and Y coordinates as needed for blending
1210  * together the different samples in an MSAA texture.
1211  */
1212 void
single_to_blend()1213 brw_blorp_blit_program::single_to_blend()
1214 {
1215    /* When looking up samples in an MSAA texture using the SAMPLE message,
1216     * Gen6 requires the texture coordinates to be odd integers (so that they
1217     * correspond to the center of a 2x2 block representing the four samples
1218     * that maxe up a pixel).  So we need to multiply our X and Y coordinates
1219     * each by 2 and then add 1.
1220     */
1221    brw_SHL(&func, t1, X, brw_imm_w(1));
1222    brw_SHL(&func, t2, Y, brw_imm_w(1));
1223    brw_ADD(&func, Xp, t1, brw_imm_w(1));
1224    brw_ADD(&func, Yp, t2, brw_imm_w(1));
1225    SWAP_XY_AND_XPYP();
1226 }
1227 
1228 
1229 /**
1230  * Count the number of trailing 1 bits in the given value.  For example:
1231  *
1232  * count_trailing_one_bits(0) == 0
1233  * count_trailing_one_bits(7) == 3
1234  * count_trailing_one_bits(11) == 2
1235  */
count_trailing_one_bits(unsigned value)1236 inline int count_trailing_one_bits(unsigned value)
1237 {
1238 #if defined(__GNUC__) && ((__GNUC__ * 100 + __GNUC_MINOR__) >= 304) /* gcc 3.4 or later */
1239    return __builtin_ctz(~value);
1240 #else
1241    return _mesa_bitcount(value & ~(value + 1));
1242 #endif
1243 }
1244 
1245 
1246 void
manual_blend(unsigned num_samples)1247 brw_blorp_blit_program::manual_blend(unsigned num_samples)
1248 {
1249    if (key->tex_layout == INTEL_MSAA_LAYOUT_CMS)
1250       mcs_fetch();
1251 
1252    /* We add together samples using a binary tree structure, e.g. for 4x MSAA:
1253     *
1254     *   result = ((sample[0] + sample[1]) + (sample[2] + sample[3])) / 4
1255     *
1256     * This ensures that when all samples have the same value, no numerical
1257     * precision is lost, since each addition operation always adds two equal
1258     * values, and summing two equal floating point values does not lose
1259     * precision.
1260     *
1261     * We perform this computation by treating the texture_data array as a
1262     * stack and performing the following operations:
1263     *
1264     * - push sample 0 onto stack
1265     * - push sample 1 onto stack
1266     * - add top two stack entries
1267     * - push sample 2 onto stack
1268     * - push sample 3 onto stack
1269     * - add top two stack entries
1270     * - add top two stack entries
1271     * - divide top stack entry by 4
1272     *
1273     * Note that after pushing sample i onto the stack, the number of add
1274     * operations we do is equal to the number of trailing 1 bits in i.  This
1275     * works provided the total number of samples is a power of two, which it
1276     * always is for i965.
1277     *
1278     * For integer formats, we replace the add operations with average
1279     * operations and skip the final division.
1280     */
1281    typedef struct brw_instruction *(*brw_op2_ptr)(struct brw_compile *,
1282                                                   struct brw_reg,
1283                                                   struct brw_reg,
1284                                                   struct brw_reg);
1285    brw_op2_ptr combine_op =
1286       key->texture_data_type == BRW_REGISTER_TYPE_F ? brw_ADD : brw_AVG;
1287    unsigned stack_depth = 0;
1288    for (unsigned i = 0; i < num_samples; ++i) {
1289       assert(stack_depth == _mesa_bitcount(i)); /* Loop invariant */
1290 
1291       /* Push sample i onto the stack */
1292       assert(stack_depth < ARRAY_SIZE(texture_data));
1293       if (i == 0) {
1294          s_is_zero = true;
1295       } else {
1296          s_is_zero = false;
1297          brw_MOV(&func, S, brw_imm_uw(i));
1298       }
1299       texel_fetch(texture_data[stack_depth++]);
1300 
1301       if (i == 0 && key->tex_layout == INTEL_MSAA_LAYOUT_CMS) {
1302          /* The Ivy Bridge PRM, Vol4 Part1 p27 (Multisample Control Surface)
1303           * suggests an optimization:
1304           *
1305           *     "A simple optimization with probable large return in
1306           *     performance is to compare the MCS value to zero (indicating
1307           *     all samples are on sample slice 0), and sample only from
1308           *     sample slice 0 using ld2dss if MCS is zero."
1309           *
1310           * Note that in the case where the MCS value is zero, sampling from
1311           * sample slice 0 using ld2dss and sampling from sample 0 using
1312           * ld2dms are equivalent (since all samples are on sample slice 0).
1313           * Since we have already sampled from sample 0, all we need to do is
1314           * skip the remaining fetches and averaging if MCS is zero.
1315           */
1316          brw_CMP(&func, vec16(brw_null_reg()), BRW_CONDITIONAL_NZ,
1317                  mcs_data, brw_imm_ud(0));
1318          brw_IF(&func, BRW_EXECUTE_16);
1319       }
1320 
1321       /* Do count_trailing_one_bits(i) times */
1322       for (int j = count_trailing_one_bits(i); j-- > 0; ) {
1323          assert(stack_depth >= 2);
1324          --stack_depth;
1325 
1326          /* TODO: should use a smaller loop bound for non_RGBA formats */
1327          for (int k = 0; k < 4; ++k) {
1328             combine_op(&func, offset(texture_data[stack_depth - 1], 2*k),
1329                        offset(vec8(texture_data[stack_depth - 1]), 2*k),
1330                        offset(vec8(texture_data[stack_depth]), 2*k));
1331          }
1332       }
1333    }
1334 
1335    /* We should have just 1 sample on the stack now. */
1336    assert(stack_depth == 1);
1337 
1338    if (key->texture_data_type == BRW_REGISTER_TYPE_F) {
1339       /* Scale the result down by a factor of num_samples */
1340       /* TODO: should use a smaller loop bound for non-RGBA formats */
1341       for (int j = 0; j < 4; ++j) {
1342          brw_MUL(&func, offset(texture_data[0], 2*j),
1343                  offset(vec8(texture_data[0]), 2*j),
1344                  brw_imm_f(1.0/num_samples));
1345       }
1346    }
1347 
1348    if (key->tex_layout == INTEL_MSAA_LAYOUT_CMS)
1349       brw_ENDIF(&func);
1350 }
1351 
1352 /**
1353  * Emit code to look up a value in the texture using the SAMPLE message (which
1354  * does blending of MSAA surfaces).
1355  */
1356 void
sample(struct brw_reg dst)1357 brw_blorp_blit_program::sample(struct brw_reg dst)
1358 {
1359    static const sampler_message_arg args[2] = {
1360       SAMPLER_MESSAGE_ARG_U_FLOAT,
1361       SAMPLER_MESSAGE_ARG_V_FLOAT
1362    };
1363 
1364    texture_lookup(dst, GEN5_SAMPLER_MESSAGE_SAMPLE, args, ARRAY_SIZE(args));
1365 }
1366 
1367 /**
1368  * Emit code to look up a value in the texture using the SAMPLE_LD message
1369  * (which does a simple texel fetch).
1370  */
1371 void
texel_fetch(struct brw_reg dst)1372 brw_blorp_blit_program::texel_fetch(struct brw_reg dst)
1373 {
1374    static const sampler_message_arg gen6_args[5] = {
1375       SAMPLER_MESSAGE_ARG_U_INT,
1376       SAMPLER_MESSAGE_ARG_V_INT,
1377       SAMPLER_MESSAGE_ARG_ZERO_INT, /* R */
1378       SAMPLER_MESSAGE_ARG_ZERO_INT, /* LOD */
1379       SAMPLER_MESSAGE_ARG_SI_INT
1380    };
1381    static const sampler_message_arg gen7_ld_args[3] = {
1382       SAMPLER_MESSAGE_ARG_U_INT,
1383       SAMPLER_MESSAGE_ARG_ZERO_INT, /* LOD */
1384       SAMPLER_MESSAGE_ARG_V_INT
1385    };
1386    static const sampler_message_arg gen7_ld2dss_args[3] = {
1387       SAMPLER_MESSAGE_ARG_SI_INT,
1388       SAMPLER_MESSAGE_ARG_U_INT,
1389       SAMPLER_MESSAGE_ARG_V_INT
1390    };
1391    static const sampler_message_arg gen7_ld2dms_args[4] = {
1392       SAMPLER_MESSAGE_ARG_SI_INT,
1393       SAMPLER_MESSAGE_ARG_MCS_INT,
1394       SAMPLER_MESSAGE_ARG_U_INT,
1395       SAMPLER_MESSAGE_ARG_V_INT
1396    };
1397 
1398    switch (brw->intel.gen) {
1399    case 6:
1400       texture_lookup(dst, GEN5_SAMPLER_MESSAGE_SAMPLE_LD, gen6_args,
1401                      s_is_zero ? 2 : 5);
1402       break;
1403    case 7:
1404       switch (key->tex_layout) {
1405       case INTEL_MSAA_LAYOUT_IMS:
1406          /* From the Ivy Bridge PRM, Vol4 Part1 p72 (Multisampled Surface Storage
1407           * Format):
1408           *
1409           *     If this field is MSFMT_DEPTH_STENCIL
1410           *     [a.k.a. INTEL_MSAA_LAYOUT_IMS], the only sampling engine
1411           *     messages allowed are "ld2dms", "resinfo", and "sampleinfo".
1412           *
1413           * So fall through to emit the same message as we use for
1414           * INTEL_MSAA_LAYOUT_CMS.
1415           */
1416       case INTEL_MSAA_LAYOUT_CMS:
1417          texture_lookup(dst, GEN7_SAMPLER_MESSAGE_SAMPLE_LD2DMS,
1418                         gen7_ld2dms_args, ARRAY_SIZE(gen7_ld2dms_args));
1419          break;
1420       case INTEL_MSAA_LAYOUT_UMS:
1421          texture_lookup(dst, GEN7_SAMPLER_MESSAGE_SAMPLE_LD2DSS,
1422                         gen7_ld2dss_args, ARRAY_SIZE(gen7_ld2dss_args));
1423          break;
1424       case INTEL_MSAA_LAYOUT_NONE:
1425          assert(s_is_zero);
1426          texture_lookup(dst, GEN5_SAMPLER_MESSAGE_SAMPLE_LD, gen7_ld_args,
1427                         ARRAY_SIZE(gen7_ld_args));
1428          break;
1429       }
1430       break;
1431    default:
1432       assert(!"Should not get here.");
1433       break;
1434    };
1435 }
1436 
1437 void
mcs_fetch()1438 brw_blorp_blit_program::mcs_fetch()
1439 {
1440    static const sampler_message_arg gen7_ld_mcs_args[2] = {
1441       SAMPLER_MESSAGE_ARG_U_INT,
1442       SAMPLER_MESSAGE_ARG_V_INT
1443    };
1444    texture_lookup(vec16(mcs_data), GEN7_SAMPLER_MESSAGE_SAMPLE_LD_MCS,
1445                   gen7_ld_mcs_args, ARRAY_SIZE(gen7_ld_mcs_args));
1446 }
1447 
1448 void
expand_to_32_bits(struct brw_reg src,struct brw_reg dst)1449 brw_blorp_blit_program::expand_to_32_bits(struct brw_reg src,
1450                                           struct brw_reg dst)
1451 {
1452    brw_MOV(&func, vec8(dst), vec8(src));
1453    brw_set_compression_control(&func, BRW_COMPRESSION_2NDHALF);
1454    brw_MOV(&func, offset(vec8(dst), 1), suboffset(vec8(src), 8));
1455    brw_set_compression_control(&func, BRW_COMPRESSION_NONE);
1456 }
1457 
1458 void
texture_lookup(struct brw_reg dst,GLuint msg_type,const sampler_message_arg * args,int num_args)1459 brw_blorp_blit_program::texture_lookup(struct brw_reg dst,
1460                                        GLuint msg_type,
1461                                        const sampler_message_arg *args,
1462                                        int num_args)
1463 {
1464    struct brw_reg mrf =
1465       retype(vec16(brw_message_reg(base_mrf)), BRW_REGISTER_TYPE_UD);
1466    for (int arg = 0; arg < num_args; ++arg) {
1467       switch (args[arg]) {
1468       case SAMPLER_MESSAGE_ARG_U_FLOAT:
1469          expand_to_32_bits(X, retype(mrf, BRW_REGISTER_TYPE_F));
1470          break;
1471       case SAMPLER_MESSAGE_ARG_V_FLOAT:
1472          expand_to_32_bits(Y, retype(mrf, BRW_REGISTER_TYPE_F));
1473          break;
1474       case SAMPLER_MESSAGE_ARG_U_INT:
1475          expand_to_32_bits(X, mrf);
1476          break;
1477       case SAMPLER_MESSAGE_ARG_V_INT:
1478          expand_to_32_bits(Y, mrf);
1479          break;
1480       case SAMPLER_MESSAGE_ARG_SI_INT:
1481          /* Note: on Gen7, this code may be reached with s_is_zero==true
1482           * because in Gen7's ld2dss message, the sample index is the first
1483           * argument.  When this happens, we need to move a 0 into the
1484           * appropriate message register.
1485           */
1486          if (s_is_zero)
1487             brw_MOV(&func, mrf, brw_imm_ud(0));
1488          else
1489             expand_to_32_bits(S, mrf);
1490          break;
1491       case SAMPLER_MESSAGE_ARG_MCS_INT:
1492          switch (key->tex_layout) {
1493          case INTEL_MSAA_LAYOUT_CMS:
1494             brw_MOV(&func, mrf, mcs_data);
1495             break;
1496          case INTEL_MSAA_LAYOUT_IMS:
1497             /* When sampling from an IMS surface, MCS data is not relevant,
1498              * and the hardware ignores it.  So don't bother populating it.
1499              */
1500             break;
1501          default:
1502             /* We shouldn't be trying to send MCS data with any other
1503              * layouts.
1504              */
1505             assert (!"Unsupported layout for MCS data");
1506             break;
1507          }
1508          break;
1509       case SAMPLER_MESSAGE_ARG_ZERO_INT:
1510          brw_MOV(&func, mrf, brw_imm_ud(0));
1511          break;
1512       }
1513       mrf.nr += 2;
1514    }
1515 
1516    brw_SAMPLE(&func,
1517               retype(dst, BRW_REGISTER_TYPE_UW) /* dest */,
1518               base_mrf /* msg_reg_nr */,
1519               brw_message_reg(base_mrf) /* src0 */,
1520               BRW_BLORP_TEXTURE_BINDING_TABLE_INDEX,
1521               0 /* sampler */,
1522               WRITEMASK_XYZW,
1523               msg_type,
1524               8 /* response_length.  TODO: should be smaller for non-RGBA formats? */,
1525               mrf.nr - base_mrf /* msg_length */,
1526               0 /* header_present */,
1527               BRW_SAMPLER_SIMD_MODE_SIMD16,
1528               BRW_SAMPLER_RETURN_FORMAT_FLOAT32);
1529 }
1530 
1531 #undef X
1532 #undef Y
1533 #undef U
1534 #undef V
1535 #undef S
1536 #undef SWAP_XY_AND_XPYP
1537 
1538 void
render_target_write()1539 brw_blorp_blit_program::render_target_write()
1540 {
1541    struct brw_reg mrf_rt_write =
1542       retype(vec16(brw_message_reg(base_mrf)), key->texture_data_type);
1543    int mrf_offset = 0;
1544 
1545    /* If we may have killed pixels, then we need to send R0 and R1 in a header
1546     * so that the render target knows which pixels we killed.
1547     */
1548    bool use_header = key->use_kill;
1549    if (use_header) {
1550       /* Copy R0/1 to MRF */
1551       brw_MOV(&func, retype(mrf_rt_write, BRW_REGISTER_TYPE_UD),
1552               retype(R0, BRW_REGISTER_TYPE_UD));
1553       mrf_offset += 2;
1554    }
1555 
1556    /* Copy texture data to MRFs */
1557    for (int i = 0; i < 4; ++i) {
1558       /* E.g. mov(16) m2.0<1>:f r2.0<8;8,1>:f { Align1, H1 } */
1559       brw_MOV(&func, offset(mrf_rt_write, mrf_offset),
1560               offset(vec8(texture_data[0]), 2*i));
1561       mrf_offset += 2;
1562    }
1563 
1564    /* Now write to the render target and terminate the thread */
1565    brw_fb_WRITE(&func,
1566                 16 /* dispatch_width */,
1567                 base_mrf /* msg_reg_nr */,
1568                 mrf_rt_write /* src0 */,
1569                 BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE,
1570                 BRW_BLORP_RENDERBUFFER_BINDING_TABLE_INDEX,
1571                 mrf_offset /* msg_length.  TODO: Should be smaller for non-RGBA formats. */,
1572                 0 /* response_length */,
1573                 true /* eot */,
1574                 use_header);
1575 }
1576 
1577 
1578 void
setup(GLuint src0,GLuint dst0,GLuint dst1,bool mirror)1579 brw_blorp_coord_transform_params::setup(GLuint src0, GLuint dst0, GLuint dst1,
1580                                         bool mirror)
1581 {
1582    if (!mirror) {
1583       /* When not mirroring a coordinate (say, X), we need:
1584        *   x' - src_x0 = x - dst_x0
1585        * Therefore:
1586        *   x' = 1*x + (src_x0 - dst_x0)
1587        */
1588       multiplier = 1;
1589       offset = src0 - dst0;
1590    } else {
1591       /* When mirroring X we need:
1592        *   x' - src_x0 = dst_x1 - x - 1
1593        * Therefore:
1594        *   x' = -1*x + (src_x0 + dst_x1 - 1)
1595        */
1596       multiplier = -1;
1597       offset = src0 + dst1 - 1;
1598    }
1599 }
1600 
1601 
1602 /**
1603  * Determine which MSAA layout the GPU pipeline should be configured for,
1604  * based on the chip generation, the number of samples, and the true layout of
1605  * the image in memory.
1606  */
1607 inline intel_msaa_layout
compute_msaa_layout_for_pipeline(struct brw_context * brw,unsigned num_samples,intel_msaa_layout true_layout)1608 compute_msaa_layout_for_pipeline(struct brw_context *brw, unsigned num_samples,
1609                                  intel_msaa_layout true_layout)
1610 {
1611    if (num_samples <= 1) {
1612       /* When configuring the GPU for non-MSAA, we can still accommodate IMS
1613        * format buffers, by transforming coordinates appropriately.
1614        */
1615       assert(true_layout == INTEL_MSAA_LAYOUT_NONE ||
1616              true_layout == INTEL_MSAA_LAYOUT_IMS);
1617       return INTEL_MSAA_LAYOUT_NONE;
1618    } else {
1619       assert(true_layout != INTEL_MSAA_LAYOUT_NONE);
1620    }
1621 
1622    /* Prior to Gen7, all MSAA surfaces use IMS layout. */
1623    if (brw->intel.gen == 6) {
1624       assert(true_layout == INTEL_MSAA_LAYOUT_IMS);
1625    }
1626 
1627    return true_layout;
1628 }
1629 
1630 
brw_blorp_blit_params(struct brw_context * brw,struct intel_mipmap_tree * src_mt,unsigned src_level,unsigned src_layer,struct intel_mipmap_tree * dst_mt,unsigned dst_level,unsigned dst_layer,GLuint src_x0,GLuint src_y0,GLuint dst_x0,GLuint dst_y0,GLuint dst_x1,GLuint dst_y1,bool mirror_x,bool mirror_y)1631 brw_blorp_blit_params::brw_blorp_blit_params(struct brw_context *brw,
1632                                              struct intel_mipmap_tree *src_mt,
1633                                              unsigned src_level, unsigned src_layer,
1634                                              struct intel_mipmap_tree *dst_mt,
1635                                              unsigned dst_level, unsigned dst_layer,
1636                                              GLuint src_x0, GLuint src_y0,
1637                                              GLuint dst_x0, GLuint dst_y0,
1638                                              GLuint dst_x1, GLuint dst_y1,
1639                                              bool mirror_x, bool mirror_y)
1640 {
1641    src.set(brw, src_mt, src_level, src_layer);
1642    dst.set(brw, dst_mt, dst_level, dst_layer);
1643 
1644    /* If we are blitting from sRGB to linear or vice versa, we still want the
1645     * blit to be a direct copy, so we need source and destination to use the
1646     * same format.  However, we want the destination sRGB/linear state to be
1647     * correct (so that sRGB blending is used when doing an MSAA resolve to an
1648     * sRGB surface, and linear blending is used when doing an MSAA resolve to
1649     * a linear surface).  Since blorp blits don't support any format
1650     * conversion (except between sRGB and linear), we can accomplish this by
1651     * simply setting up the source to use the same format as the destination.
1652     */
1653    assert(_mesa_get_srgb_format_linear(src_mt->format) ==
1654           _mesa_get_srgb_format_linear(dst_mt->format));
1655    src.brw_surfaceformat = dst.brw_surfaceformat;
1656 
1657    use_wm_prog = true;
1658    memset(&wm_prog_key, 0, sizeof(wm_prog_key));
1659 
1660    /* texture_data_type indicates the register type that should be used to
1661     * manipulate texture data.
1662     */
1663    switch (_mesa_get_format_datatype(src_mt->format)) {
1664    case GL_UNSIGNED_NORMALIZED:
1665    case GL_SIGNED_NORMALIZED:
1666    case GL_FLOAT:
1667       wm_prog_key.texture_data_type = BRW_REGISTER_TYPE_F;
1668       break;
1669    case GL_UNSIGNED_INT:
1670       if (src_mt->format == MESA_FORMAT_S8) {
1671          /* We process stencil as though it's an unsigned normalized color */
1672          wm_prog_key.texture_data_type = BRW_REGISTER_TYPE_F;
1673       } else {
1674          wm_prog_key.texture_data_type = BRW_REGISTER_TYPE_UD;
1675       }
1676       break;
1677    case GL_INT:
1678       wm_prog_key.texture_data_type = BRW_REGISTER_TYPE_D;
1679       break;
1680    default:
1681       assert(!"Unrecognized blorp format");
1682       break;
1683    }
1684 
1685    if (brw->intel.gen > 6) {
1686       /* Gen7's rendering hardware only supports the IMS layout for depth and
1687        * stencil render targets.  Blorp always maps its destination surface as
1688        * a color render target (even if it's actually a depth or stencil
1689        * buffer).  So if the destination is IMS, we'll have to map it as a
1690        * single-sampled texture and interleave the samples ourselves.
1691        */
1692       if (dst_mt->msaa_layout == INTEL_MSAA_LAYOUT_IMS)
1693          dst.num_samples = 0;
1694    }
1695 
1696    if (dst.map_stencil_as_y_tiled && dst.num_samples > 1) {
1697       /* If the destination surface is a W-tiled multisampled stencil buffer
1698        * that we're mapping as Y tiled, then we need to arrange for the WM
1699        * program to run once per sample rather than once per pixel, because
1700        * the memory layout of related samples doesn't match between W and Y
1701        * tiling.
1702        */
1703       wm_prog_key.persample_msaa_dispatch = true;
1704    }
1705 
1706    if (src.num_samples > 0 && dst.num_samples > 1) {
1707       /* We are blitting from a multisample buffer to a multisample buffer, so
1708        * we must preserve samples within a pixel.  This means we have to
1709        * arrange for the WM program to run once per sample rather than once
1710        * per pixel.
1711        */
1712       wm_prog_key.persample_msaa_dispatch = true;
1713    }
1714 
1715    /* The render path must be configured to use the same number of samples as
1716     * the destination buffer.
1717     */
1718    num_samples = dst.num_samples;
1719 
1720    GLenum base_format = _mesa_get_format_base_format(src_mt->format);
1721    if (base_format != GL_DEPTH_COMPONENT && /* TODO: what about depth/stencil? */
1722        base_format != GL_STENCIL_INDEX &&
1723        src_mt->num_samples > 1 && dst_mt->num_samples <= 1) {
1724       /* We are downsampling a color buffer, so blend. */
1725       wm_prog_key.blend = true;
1726    }
1727 
1728    /* src_samples and dst_samples are the true sample counts */
1729    wm_prog_key.src_samples = src_mt->num_samples;
1730    wm_prog_key.dst_samples = dst_mt->num_samples;
1731 
1732    /* tex_samples and rt_samples are the sample counts that are set up in
1733     * SURFACE_STATE.
1734     */
1735    wm_prog_key.tex_samples = src.num_samples;
1736    wm_prog_key.rt_samples  = dst.num_samples;
1737 
1738    /* tex_layout and rt_layout indicate the MSAA layout the GPU pipeline will
1739     * use to access the source and destination surfaces.
1740     */
1741    wm_prog_key.tex_layout =
1742       compute_msaa_layout_for_pipeline(brw, src.num_samples, src.msaa_layout);
1743    wm_prog_key.rt_layout =
1744       compute_msaa_layout_for_pipeline(brw, dst.num_samples, dst.msaa_layout);
1745 
1746    /* src_layout and dst_layout indicate the true MSAA layout used by src and
1747     * dst.
1748     */
1749    wm_prog_key.src_layout = src_mt->msaa_layout;
1750    wm_prog_key.dst_layout = dst_mt->msaa_layout;
1751 
1752    wm_prog_key.src_tiled_w = src.map_stencil_as_y_tiled;
1753    wm_prog_key.dst_tiled_w = dst.map_stencil_as_y_tiled;
1754    x0 = wm_push_consts.dst_x0 = dst_x0;
1755    y0 = wm_push_consts.dst_y0 = dst_y0;
1756    x1 = wm_push_consts.dst_x1 = dst_x1;
1757    y1 = wm_push_consts.dst_y1 = dst_y1;
1758    wm_push_consts.x_transform.setup(src_x0, dst_x0, dst_x1, mirror_x);
1759    wm_push_consts.y_transform.setup(src_y0, dst_y0, dst_y1, mirror_y);
1760 
1761    if (dst.num_samples <= 1 && dst_mt->num_samples > 1) {
1762       /* We must expand the rectangle we send through the rendering pipeline,
1763        * to account for the fact that we are mapping the destination region as
1764        * single-sampled when it is in fact multisampled.  We must also align
1765        * it to a multiple of the multisampling pattern, because the
1766        * differences between multisampled and single-sampled surface formats
1767        * will mean that pixels are scrambled within the multisampling pattern.
1768        * TODO: what if this makes the coordinates too large?
1769        *
1770        * Note: this only works if the destination surface uses the IMS layout.
1771        * If it's UMS, then we have no choice but to set up the rendering
1772        * pipeline as multisampled.
1773        */
1774       assert(dst_mt->msaa_layout == INTEL_MSAA_LAYOUT_IMS);
1775       switch (dst_mt->num_samples) {
1776       case 4:
1777          x0 = ROUND_DOWN_TO(x0 * 2, 4);
1778          y0 = ROUND_DOWN_TO(y0 * 2, 4);
1779          x1 = ALIGN(x1 * 2, 4);
1780          y1 = ALIGN(y1 * 2, 4);
1781          break;
1782       case 8:
1783          x0 = ROUND_DOWN_TO(x0 * 4, 8);
1784          y0 = ROUND_DOWN_TO(y0 * 2, 4);
1785          x1 = ALIGN(x1 * 4, 8);
1786          y1 = ALIGN(y1 * 2, 4);
1787          break;
1788       default:
1789          assert(!"Unrecognized sample count in brw_blorp_blit_params ctor");
1790          break;
1791       }
1792       wm_prog_key.use_kill = true;
1793    }
1794 
1795    if (dst.map_stencil_as_y_tiled) {
1796       /* We must modify the rectangle we send through the rendering pipeline
1797        * (and the size and x/y offset of the destination surface), to account
1798        * for the fact that we are mapping it as Y-tiled when it is in fact
1799        * W-tiled.
1800        *
1801        * Both Y tiling and W tiling can be understood as organizations of
1802        * 32-byte sub-tiles; within each 32-byte sub-tile, the layout of pixels
1803        * is different, but the layout of the 32-byte sub-tiles within the 4k
1804        * tile is the same (8 sub-tiles across by 16 sub-tiles down, in
1805        * column-major order).  In Y tiling, the sub-tiles are 16 bytes wide
1806        * and 2 rows high; in W tiling, they are 8 bytes wide and 4 rows high.
1807        *
1808        * Therefore, to account for the layout differences within the 32-byte
1809        * sub-tiles, we must expand the rectangle so the X coordinates of its
1810        * edges are multiples of 8 (the W sub-tile width), and its Y
1811        * coordinates of its edges are multiples of 4 (the W sub-tile height).
1812        * Then we need to scale the X and Y coordinates of the rectangle to
1813        * account for the differences in aspect ratio between the Y and W
1814        * sub-tiles.  We need to modify the layer width and height similarly.
1815        *
1816        * A correction needs to be applied when MSAA is in use: since
1817        * INTEL_MSAA_LAYOUT_IMS uses an interleaving pattern whose height is 4,
1818        * we need to align the Y coordinates to multiples of 8, so that when
1819        * they are divided by two they are still multiples of 4.
1820        *
1821        * Note: Since the x/y offset of the surface will be applied using the
1822        * SURFACE_STATE command packet, it will be invisible to the swizzling
1823        * code in the shader; therefore it needs to be in a multiple of the
1824        * 32-byte sub-tile size.  Fortunately it is, since the sub-tile is 8
1825        * pixels wide and 4 pixels high (when viewed as a W-tiled stencil
1826        * buffer), and the miplevel alignment used for stencil buffers is 8
1827        * pixels horizontally and either 4 or 8 pixels vertically (see
1828        * intel_horizontal_texture_alignment_unit() and
1829        * intel_vertical_texture_alignment_unit()).
1830        *
1831        * Note: Also, since the SURFACE_STATE command packet can only apply
1832        * offsets that are multiples of 4 pixels horizontally and 2 pixels
1833        * vertically, it is important that the offsets will be multiples of
1834        * these sizes after they are converted into Y-tiled coordinates.
1835        * Fortunately they will be, since we know from above that the offsets
1836        * are a multiple of the 32-byte sub-tile size, and in Y-tiled
1837        * coordinates the sub-tile is 16 pixels wide and 2 pixels high.
1838        *
1839        * TODO: what if this makes the coordinates (or the texture size) too
1840        * large?
1841        */
1842       const unsigned x_align = 8, y_align = dst.num_samples != 0 ? 8 : 4;
1843       x0 = ROUND_DOWN_TO(x0, x_align) * 2;
1844       y0 = ROUND_DOWN_TO(y0, y_align) / 2;
1845       x1 = ALIGN(x1, x_align) * 2;
1846       y1 = ALIGN(y1, y_align) / 2;
1847       dst.width = ALIGN(dst.width, x_align) * 2;
1848       dst.height = ALIGN(dst.height, y_align) / 2;
1849       dst.x_offset *= 2;
1850       dst.y_offset /= 2;
1851       wm_prog_key.use_kill = true;
1852    }
1853 
1854    if (src.map_stencil_as_y_tiled) {
1855       /* We must modify the size and x/y offset of the source surface to
1856        * account for the fact that we are mapping it as Y-tiled when it is in
1857        * fact W tiled.
1858        *
1859        * See the comments above concerning x/y offset alignment for the
1860        * destination surface.
1861        *
1862        * TODO: what if this makes the texture size too large?
1863        */
1864       const unsigned x_align = 8, y_align = src.num_samples != 0 ? 8 : 4;
1865       src.width = ALIGN(src.width, x_align) * 2;
1866       src.height = ALIGN(src.height, y_align) / 2;
1867       src.x_offset *= 2;
1868       src.y_offset /= 2;
1869    }
1870 }
1871 
1872 uint32_t
get_wm_prog(struct brw_context * brw,brw_blorp_prog_data ** prog_data) const1873 brw_blorp_blit_params::get_wm_prog(struct brw_context *brw,
1874                                    brw_blorp_prog_data **prog_data) const
1875 {
1876    uint32_t prog_offset;
1877    if (!brw_search_cache(&brw->cache, BRW_BLORP_BLIT_PROG,
1878                          &this->wm_prog_key, sizeof(this->wm_prog_key),
1879                          &prog_offset, prog_data)) {
1880       brw_blorp_blit_program prog(brw, &this->wm_prog_key);
1881       GLuint program_size;
1882       const GLuint *program = prog.compile(brw, &program_size);
1883       brw_upload_cache(&brw->cache, BRW_BLORP_BLIT_PROG,
1884                        &this->wm_prog_key, sizeof(this->wm_prog_key),
1885                        program, program_size,
1886                        &prog.prog_data, sizeof(prog.prog_data),
1887                        &prog_offset, prog_data);
1888    }
1889    return prog_offset;
1890 }
1891