Go to most recent revision | Details | Last modification | View Log | RSS feed
Rev | Author | Line No. | Line |
---|---|---|---|
5564 | serge | 1 | /* |
2 | * Copyright © 2011 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 |
||
21 | * DEALINGS IN THE SOFTWARE. |
||
22 | */ |
||
23 | |||
24 | /** |
||
25 | * \file lower_varyings_to_packed.cpp |
||
26 | * |
||
27 | * This lowering pass generates GLSL code that manually packs varyings into |
||
28 | * vec4 slots, for the benefit of back-ends that don't support packed varyings |
||
29 | * natively. |
||
30 | * |
||
31 | * For example, the following shader: |
||
32 | * |
||
33 | * out mat3x2 foo; // location=4, location_frac=0 |
||
34 | * out vec3 bar[2]; // location=5, location_frac=2 |
||
35 | * |
||
36 | * main() |
||
37 | * { |
||
38 | * ... |
||
39 | * } |
||
40 | * |
||
41 | * Is rewritten to: |
||
42 | * |
||
43 | * mat3x2 foo; |
||
44 | * vec3 bar[2]; |
||
45 | * out vec4 packed4; // location=4, location_frac=0 |
||
46 | * out vec4 packed5; // location=5, location_frac=0 |
||
47 | * out vec4 packed6; // location=6, location_frac=0 |
||
48 | * |
||
49 | * main() |
||
50 | * { |
||
51 | * ... |
||
52 | * packed4.xy = foo[0]; |
||
53 | * packed4.zw = foo[1]; |
||
54 | * packed5.xy = foo[2]; |
||
55 | * packed5.zw = bar[0].xy; |
||
56 | * packed6.x = bar[0].z; |
||
57 | * packed6.yzw = bar[1]; |
||
58 | * } |
||
59 | * |
||
60 | * This lowering pass properly handles "double parking" of a varying vector |
||
61 | * across two varying slots. For example, in the code above, two of the |
||
62 | * components of bar[0] are stored in packed5, and the remaining component is |
||
63 | * stored in packed6. |
||
64 | * |
||
65 | * Note that in theory, the extra instructions may cause some loss of |
||
66 | * performance. However, hopefully in most cases the performance loss will |
||
67 | * either be absorbed by a later optimization pass, or it will be offset by |
||
68 | * memory bandwidth savings (because fewer varyings are used). |
||
69 | * |
||
70 | * This lowering pass also packs flat floats, ints, and uints together, by |
||
71 | * using ivec4 as the base type of flat "varyings", and using appropriate |
||
72 | * casts to convert floats and uints into ints. |
||
73 | * |
||
74 | * This lowering pass also handles varyings whose type is a struct or an array |
||
75 | * of struct. Structs are packed in order and with no gaps, so there may be a |
||
76 | * performance penalty due to structure elements being double-parked. |
||
77 | * |
||
78 | * Lowering of geometry shader inputs is slightly more complex, since geometry |
||
79 | * inputs are always arrays, so we need to lower arrays to arrays. For |
||
80 | * example, the following input: |
||
81 | * |
||
82 | * in struct Foo { |
||
83 | * float f; |
||
84 | * vec3 v; |
||
85 | * vec2 a[2]; |
||
86 | * } arr[3]; // location=4, location_frac=0 |
||
87 | * |
||
88 | * Would get lowered like this if it occurred in a fragment shader: |
||
89 | * |
||
90 | * struct Foo { |
||
91 | * float f; |
||
92 | * vec3 v; |
||
93 | * vec2 a[2]; |
||
94 | * } arr[3]; |
||
95 | * in vec4 packed4; // location=4, location_frac=0 |
||
96 | * in vec4 packed5; // location=5, location_frac=0 |
||
97 | * in vec4 packed6; // location=6, location_frac=0 |
||
98 | * in vec4 packed7; // location=7, location_frac=0 |
||
99 | * in vec4 packed8; // location=8, location_frac=0 |
||
100 | * in vec4 packed9; // location=9, location_frac=0 |
||
101 | * |
||
102 | * main() |
||
103 | * { |
||
104 | * arr[0].f = packed4.x; |
||
105 | * arr[0].v = packed4.yzw; |
||
106 | * arr[0].a[0] = packed5.xy; |
||
107 | * arr[0].a[1] = packed5.zw; |
||
108 | * arr[1].f = packed6.x; |
||
109 | * arr[1].v = packed6.yzw; |
||
110 | * arr[1].a[0] = packed7.xy; |
||
111 | * arr[1].a[1] = packed7.zw; |
||
112 | * arr[2].f = packed8.x; |
||
113 | * arr[2].v = packed8.yzw; |
||
114 | * arr[2].a[0] = packed9.xy; |
||
115 | * arr[2].a[1] = packed9.zw; |
||
116 | * ... |
||
117 | * } |
||
118 | * |
||
119 | * But it would get lowered like this if it occurred in a geometry shader: |
||
120 | * |
||
121 | * struct Foo { |
||
122 | * float f; |
||
123 | * vec3 v; |
||
124 | * vec2 a[2]; |
||
125 | * } arr[3]; |
||
126 | * in vec4 packed4[3]; // location=4, location_frac=0 |
||
127 | * in vec4 packed5[3]; // location=5, location_frac=0 |
||
128 | * |
||
129 | * main() |
||
130 | * { |
||
131 | * arr[0].f = packed4[0].x; |
||
132 | * arr[0].v = packed4[0].yzw; |
||
133 | * arr[0].a[0] = packed5[0].xy; |
||
134 | * arr[0].a[1] = packed5[0].zw; |
||
135 | * arr[1].f = packed4[1].x; |
||
136 | * arr[1].v = packed4[1].yzw; |
||
137 | * arr[1].a[0] = packed5[1].xy; |
||
138 | * arr[1].a[1] = packed5[1].zw; |
||
139 | * arr[2].f = packed4[2].x; |
||
140 | * arr[2].v = packed4[2].yzw; |
||
141 | * arr[2].a[0] = packed5[2].xy; |
||
142 | * arr[2].a[1] = packed5[2].zw; |
||
143 | * ... |
||
144 | * } |
||
145 | */ |
||
146 | |||
147 | #include "glsl_symbol_table.h" |
||
148 | #include "ir.h" |
||
149 | #include "ir_builder.h" |
||
150 | #include "ir_optimization.h" |
||
151 | #include "program/prog_instruction.h" |
||
152 | |||
153 | using namespace ir_builder; |
||
154 | |||
155 | namespace { |
||
156 | |||
157 | /** |
||
158 | * Visitor that performs varying packing. For each varying declared in the |
||
159 | * shader, this visitor determines whether it needs to be packed. If so, it |
||
160 | * demotes it to an ordinary global, creates new packed varyings, and |
||
161 | * generates assignments to convert between the original varying and the |
||
162 | * packed varying. |
||
163 | */ |
||
164 | class lower_packed_varyings_visitor |
||
165 | { |
||
166 | public: |
||
167 | lower_packed_varyings_visitor(void *mem_ctx, unsigned locations_used, |
||
168 | ir_variable_mode mode, |
||
169 | unsigned gs_input_vertices, |
||
170 | exec_list *out_instructions, |
||
171 | exec_list *out_variables); |
||
172 | |||
173 | void run(exec_list *instructions); |
||
174 | |||
175 | private: |
||
176 | void bitwise_assign_pack(ir_rvalue *lhs, ir_rvalue *rhs); |
||
177 | void bitwise_assign_unpack(ir_rvalue *lhs, ir_rvalue *rhs); |
||
178 | unsigned lower_rvalue(ir_rvalue *rvalue, unsigned fine_location, |
||
179 | ir_variable *unpacked_var, const char *name, |
||
180 | bool gs_input_toplevel, unsigned vertex_index); |
||
181 | unsigned lower_arraylike(ir_rvalue *rvalue, unsigned array_size, |
||
182 | unsigned fine_location, |
||
183 | ir_variable *unpacked_var, const char *name, |
||
184 | bool gs_input_toplevel, unsigned vertex_index); |
||
185 | ir_dereference *get_packed_varying_deref(unsigned location, |
||
186 | ir_variable *unpacked_var, |
||
187 | const char *name, |
||
188 | unsigned vertex_index); |
||
189 | bool needs_lowering(ir_variable *var); |
||
190 | |||
191 | /** |
||
192 | * Memory context used to allocate new instructions for the shader. |
||
193 | */ |
||
194 | void * const mem_ctx; |
||
195 | |||
196 | /** |
||
197 | * Number of generic varying slots which are used by this shader. This is |
||
198 | * used to allocate temporary intermediate data structures. If any varying |
||
199 | * used by this shader has a location greater than or equal to |
||
200 | * VARYING_SLOT_VAR0 + locations_used, an assertion will fire. |
||
201 | */ |
||
202 | const unsigned locations_used; |
||
203 | |||
204 | /** |
||
205 | * Array of pointers to the packed varyings that have been created for each |
||
206 | * generic varying slot. NULL entries in this array indicate varying slots |
||
207 | * for which a packed varying has not been created yet. |
||
208 | */ |
||
209 | ir_variable **packed_varyings; |
||
210 | |||
211 | /** |
||
212 | * Type of varying which is being lowered in this pass (either |
||
213 | * ir_var_shader_in or ir_var_shader_out). |
||
214 | */ |
||
215 | const ir_variable_mode mode; |
||
216 | |||
217 | /** |
||
218 | * If we are currently lowering geometry shader inputs, the number of input |
||
219 | * vertices the geometry shader accepts. Otherwise zero. |
||
220 | */ |
||
221 | const unsigned gs_input_vertices; |
||
222 | |||
223 | /** |
||
224 | * Exec list into which the visitor should insert the packing instructions. |
||
225 | * Caller provides this list; it should insert the instructions into the |
||
226 | * appropriate place in the shader once the visitor has finished running. |
||
227 | */ |
||
228 | exec_list *out_instructions; |
||
229 | |||
230 | /** |
||
231 | * Exec list into which the visitor should insert any new variables. |
||
232 | */ |
||
233 | exec_list *out_variables; |
||
234 | }; |
||
235 | |||
236 | } /* anonymous namespace */ |
||
237 | |||
238 | lower_packed_varyings_visitor::lower_packed_varyings_visitor( |
||
239 | void *mem_ctx, unsigned locations_used, ir_variable_mode mode, |
||
240 | unsigned gs_input_vertices, exec_list *out_instructions, |
||
241 | exec_list *out_variables) |
||
242 | : mem_ctx(mem_ctx), |
||
243 | locations_used(locations_used), |
||
244 | packed_varyings((ir_variable **) |
||
245 | rzalloc_array_size(mem_ctx, sizeof(*packed_varyings), |
||
246 | locations_used)), |
||
247 | mode(mode), |
||
248 | gs_input_vertices(gs_input_vertices), |
||
249 | out_instructions(out_instructions), |
||
250 | out_variables(out_variables) |
||
251 | { |
||
252 | } |
||
253 | |||
254 | void |
||
255 | lower_packed_varyings_visitor::run(exec_list *instructions) |
||
256 | { |
||
257 | foreach_in_list(ir_instruction, node, instructions) { |
||
258 | ir_variable *var = node->as_variable(); |
||
259 | if (var == NULL) |
||
260 | continue; |
||
261 | |||
262 | if (var->data.mode != this->mode || |
||
263 | var->data.location < VARYING_SLOT_VAR0 || |
||
264 | !this->needs_lowering(var)) |
||
265 | continue; |
||
266 | |||
267 | /* This lowering pass is only capable of packing floats and ints |
||
268 | * together when their interpolation mode is "flat". Therefore, to be |
||
269 | * safe, caller should ensure that integral varyings always use flat |
||
270 | * interpolation, even when this is not required by GLSL. |
||
271 | */ |
||
272 | assert(var->data.interpolation == INTERP_QUALIFIER_FLAT || |
||
273 | !var->type->contains_integer()); |
||
274 | |||
275 | /* Change the old varying into an ordinary global. */ |
||
276 | assert(var->data.mode != ir_var_temporary); |
||
277 | var->data.mode = ir_var_auto; |
||
278 | |||
279 | /* Create a reference to the old varying. */ |
||
280 | ir_dereference_variable *deref |
||
281 | = new(this->mem_ctx) ir_dereference_variable(var); |
||
282 | |||
283 | /* Recursively pack or unpack it. */ |
||
284 | this->lower_rvalue(deref, var->data.location * 4 + var->data.location_frac, var, |
||
285 | var->name, this->gs_input_vertices != 0, 0); |
||
286 | } |
||
287 | } |
||
288 | |||
289 | #define SWIZZLE_ZWZW MAKE_SWIZZLE4(SWIZZLE_Z, SWIZZLE_W, SWIZZLE_Z, SWIZZLE_W) |
||
290 | |||
291 | /** |
||
292 | * Make an ir_assignment from \c rhs to \c lhs, performing appropriate |
||
293 | * bitcasts if necessary to match up types. |
||
294 | * |
||
295 | * This function is called when packing varyings. |
||
296 | */ |
||
297 | void |
||
298 | lower_packed_varyings_visitor::bitwise_assign_pack(ir_rvalue *lhs, |
||
299 | ir_rvalue *rhs) |
||
300 | { |
||
301 | if (lhs->type->base_type != rhs->type->base_type) { |
||
302 | /* Since we only mix types in flat varyings, and we always store flat |
||
303 | * varyings as type ivec4, we need only produce conversions from (uint |
||
304 | * or float) to int. |
||
305 | */ |
||
306 | assert(lhs->type->base_type == GLSL_TYPE_INT); |
||
307 | switch (rhs->type->base_type) { |
||
308 | case GLSL_TYPE_UINT: |
||
309 | rhs = new(this->mem_ctx) |
||
310 | ir_expression(ir_unop_u2i, lhs->type, rhs); |
||
311 | break; |
||
312 | case GLSL_TYPE_FLOAT: |
||
313 | rhs = new(this->mem_ctx) |
||
314 | ir_expression(ir_unop_bitcast_f2i, lhs->type, rhs); |
||
315 | break; |
||
316 | case GLSL_TYPE_DOUBLE: |
||
317 | assert(rhs->type->vector_elements <= 2); |
||
318 | if (rhs->type->vector_elements == 2) { |
||
319 | ir_variable *t = new(mem_ctx) ir_variable(lhs->type, "pack", ir_var_temporary); |
||
320 | |||
321 | assert(lhs->type->vector_elements == 4); |
||
322 | this->out_variables->push_tail(t); |
||
323 | this->out_instructions->push_tail( |
||
324 | assign(t, u2i(expr(ir_unop_unpack_double_2x32, swizzle_x(rhs->clone(mem_ctx, NULL)))), 0x3)); |
||
325 | this->out_instructions->push_tail( |
||
326 | assign(t, u2i(expr(ir_unop_unpack_double_2x32, swizzle_y(rhs))), 0xc)); |
||
327 | rhs = deref(t).val; |
||
328 | } else { |
||
329 | rhs = u2i(expr(ir_unop_unpack_double_2x32, rhs)); |
||
330 | } |
||
331 | break; |
||
332 | default: |
||
333 | assert(!"Unexpected type conversion while lowering varyings"); |
||
334 | break; |
||
335 | } |
||
336 | } |
||
337 | this->out_instructions->push_tail(new (this->mem_ctx) ir_assignment(lhs, rhs)); |
||
338 | } |
||
339 | |||
340 | |||
341 | /** |
||
342 | * Make an ir_assignment from \c rhs to \c lhs, performing appropriate |
||
343 | * bitcasts if necessary to match up types. |
||
344 | * |
||
345 | * This function is called when unpacking varyings. |
||
346 | */ |
||
347 | void |
||
348 | lower_packed_varyings_visitor::bitwise_assign_unpack(ir_rvalue *lhs, |
||
349 | ir_rvalue *rhs) |
||
350 | { |
||
351 | if (lhs->type->base_type != rhs->type->base_type) { |
||
352 | /* Since we only mix types in flat varyings, and we always store flat |
||
353 | * varyings as type ivec4, we need only produce conversions from int to |
||
354 | * (uint or float). |
||
355 | */ |
||
356 | assert(rhs->type->base_type == GLSL_TYPE_INT); |
||
357 | switch (lhs->type->base_type) { |
||
358 | case GLSL_TYPE_UINT: |
||
359 | rhs = new(this->mem_ctx) |
||
360 | ir_expression(ir_unop_i2u, lhs->type, rhs); |
||
361 | break; |
||
362 | case GLSL_TYPE_FLOAT: |
||
363 | rhs = new(this->mem_ctx) |
||
364 | ir_expression(ir_unop_bitcast_i2f, lhs->type, rhs); |
||
365 | break; |
||
366 | case GLSL_TYPE_DOUBLE: |
||
367 | assert(lhs->type->vector_elements <= 2); |
||
368 | if (lhs->type->vector_elements == 2) { |
||
369 | ir_variable *t = new(mem_ctx) ir_variable(lhs->type, "unpack", ir_var_temporary); |
||
370 | assert(rhs->type->vector_elements == 4); |
||
371 | this->out_variables->push_tail(t); |
||
372 | this->out_instructions->push_tail( |
||
373 | assign(t, expr(ir_unop_pack_double_2x32, i2u(swizzle_xy(rhs->clone(mem_ctx, NULL)))), 0x1)); |
||
374 | this->out_instructions->push_tail( |
||
375 | assign(t, expr(ir_unop_pack_double_2x32, i2u(swizzle(rhs->clone(mem_ctx, NULL), SWIZZLE_ZWZW, 2))), 0x2)); |
||
376 | rhs = deref(t).val; |
||
377 | } else { |
||
378 | rhs = expr(ir_unop_pack_double_2x32, i2u(rhs)); |
||
379 | } |
||
380 | break; |
||
381 | default: |
||
382 | assert(!"Unexpected type conversion while lowering varyings"); |
||
383 | break; |
||
384 | } |
||
385 | } |
||
386 | this->out_instructions->push_tail(new(this->mem_ctx) ir_assignment(lhs, rhs)); |
||
387 | } |
||
388 | |||
389 | |||
390 | /** |
||
391 | * Recursively pack or unpack the given varying (or portion of a varying) by |
||
392 | * traversing all of its constituent vectors. |
||
393 | * |
||
394 | * \param fine_location is the location where the first constituent vector |
||
395 | * should be packed--the word "fine" indicates that this location is expressed |
||
396 | * in multiples of a float, rather than multiples of a vec4 as is used |
||
397 | * elsewhere in Mesa. |
||
398 | * |
||
399 | * \param gs_input_toplevel should be set to true if we are lowering geometry |
||
400 | * shader inputs, and we are currently lowering the whole input variable |
||
401 | * (i.e. we are lowering the array whose index selects the vertex). |
||
402 | * |
||
403 | * \param vertex_index: if we are lowering geometry shader inputs, and the |
||
404 | * level of the array that we are currently lowering is *not* the top level, |
||
405 | * then this indicates which vertex we are currently lowering. Otherwise it |
||
406 | * is ignored. |
||
407 | * |
||
408 | * \return the location where the next constituent vector (after this one) |
||
409 | * should be packed. |
||
410 | */ |
||
411 | unsigned |
||
412 | lower_packed_varyings_visitor::lower_rvalue(ir_rvalue *rvalue, |
||
413 | unsigned fine_location, |
||
414 | ir_variable *unpacked_var, |
||
415 | const char *name, |
||
416 | bool gs_input_toplevel, |
||
417 | unsigned vertex_index) |
||
418 | { |
||
419 | unsigned dmul = rvalue->type->is_double() ? 2 : 1; |
||
420 | /* When gs_input_toplevel is set, we should be looking at a geometry shader |
||
421 | * input array. |
||
422 | */ |
||
423 | assert(!gs_input_toplevel || rvalue->type->is_array()); |
||
424 | |||
425 | if (rvalue->type->is_record()) { |
||
426 | for (unsigned i = 0; i < rvalue->type->length; i++) { |
||
427 | if (i != 0) |
||
428 | rvalue = rvalue->clone(this->mem_ctx, NULL); |
||
429 | const char *field_name = rvalue->type->fields.structure[i].name; |
||
430 | ir_dereference_record *dereference_record = new(this->mem_ctx) |
||
431 | ir_dereference_record(rvalue, field_name); |
||
432 | char *deref_name |
||
433 | = ralloc_asprintf(this->mem_ctx, "%s.%s", name, field_name); |
||
434 | fine_location = this->lower_rvalue(dereference_record, fine_location, |
||
435 | unpacked_var, deref_name, false, |
||
436 | vertex_index); |
||
437 | } |
||
438 | return fine_location; |
||
439 | } else if (rvalue->type->is_array()) { |
||
440 | /* Arrays are packed/unpacked by considering each array element in |
||
441 | * sequence. |
||
442 | */ |
||
443 | return this->lower_arraylike(rvalue, rvalue->type->array_size(), |
||
444 | fine_location, unpacked_var, name, |
||
445 | gs_input_toplevel, vertex_index); |
||
446 | } else if (rvalue->type->is_matrix()) { |
||
447 | /* Matrices are packed/unpacked by considering each column vector in |
||
448 | * sequence. |
||
449 | */ |
||
450 | return this->lower_arraylike(rvalue, rvalue->type->matrix_columns, |
||
451 | fine_location, unpacked_var, name, |
||
452 | false, vertex_index); |
||
453 | } else if (rvalue->type->vector_elements * dmul + |
||
454 | fine_location % 4 > 4) { |
||
455 | /* This vector is going to be "double parked" across two varying slots, |
||
456 | * so handle it as two separate assignments. For doubles, a dvec3/dvec4 |
||
457 | * can end up being spread over 3 slots. However the second splitting |
||
458 | * will happen later, here we just always want to split into 2. |
||
459 | */ |
||
460 | unsigned left_components, right_components; |
||
461 | unsigned left_swizzle_values[4] = { 0, 0, 0, 0 }; |
||
462 | unsigned right_swizzle_values[4] = { 0, 0, 0, 0 }; |
||
463 | char left_swizzle_name[4] = { 0, 0, 0, 0 }; |
||
464 | char right_swizzle_name[4] = { 0, 0, 0, 0 }; |
||
465 | |||
466 | left_components = 4 - fine_location % 4; |
||
467 | if (rvalue->type->is_double()) { |
||
468 | /* We might actually end up with 0 left components! */ |
||
469 | left_components /= 2; |
||
470 | } |
||
471 | right_components = rvalue->type->vector_elements - left_components; |
||
472 | |||
473 | for (unsigned i = 0; i < left_components; i++) { |
||
474 | left_swizzle_values[i] = i; |
||
475 | left_swizzle_name[i] = "xyzw"[i]; |
||
476 | } |
||
477 | for (unsigned i = 0; i < right_components; i++) { |
||
478 | right_swizzle_values[i] = i + left_components; |
||
479 | right_swizzle_name[i] = "xyzw"[i + left_components]; |
||
480 | } |
||
481 | ir_swizzle *left_swizzle = new(this->mem_ctx) |
||
482 | ir_swizzle(rvalue, left_swizzle_values, left_components); |
||
483 | ir_swizzle *right_swizzle = new(this->mem_ctx) |
||
484 | ir_swizzle(rvalue->clone(this->mem_ctx, NULL), right_swizzle_values, |
||
485 | right_components); |
||
486 | char *left_name |
||
487 | = ralloc_asprintf(this->mem_ctx, "%s.%s", name, left_swizzle_name); |
||
488 | char *right_name |
||
489 | = ralloc_asprintf(this->mem_ctx, "%s.%s", name, right_swizzle_name); |
||
490 | if (left_components) |
||
491 | fine_location = this->lower_rvalue(left_swizzle, fine_location, |
||
492 | unpacked_var, left_name, false, |
||
493 | vertex_index); |
||
494 | else |
||
495 | /* Top up the fine location to the next slot */ |
||
496 | fine_location++; |
||
497 | return this->lower_rvalue(right_swizzle, fine_location, unpacked_var, |
||
498 | right_name, false, vertex_index); |
||
499 | } else { |
||
500 | /* No special handling is necessary; pack the rvalue into the |
||
501 | * varying. |
||
502 | */ |
||
503 | unsigned swizzle_values[4] = { 0, 0, 0, 0 }; |
||
504 | unsigned components = rvalue->type->vector_elements * dmul; |
||
505 | unsigned location = fine_location / 4; |
||
506 | unsigned location_frac = fine_location % 4; |
||
507 | for (unsigned i = 0; i < components; ++i) |
||
508 | swizzle_values[i] = i + location_frac; |
||
509 | ir_dereference *packed_deref = |
||
510 | this->get_packed_varying_deref(location, unpacked_var, name, |
||
511 | vertex_index); |
||
512 | ir_swizzle *swizzle = new(this->mem_ctx) |
||
513 | ir_swizzle(packed_deref, swizzle_values, components); |
||
514 | if (this->mode == ir_var_shader_out) { |
||
515 | this->bitwise_assign_pack(swizzle, rvalue); |
||
516 | } else { |
||
517 | this->bitwise_assign_unpack(rvalue, swizzle); |
||
518 | } |
||
519 | return fine_location + components; |
||
520 | } |
||
521 | } |
||
522 | |||
523 | /** |
||
524 | * Recursively pack or unpack a varying for which we need to iterate over its |
||
525 | * constituent elements, accessing each one using an ir_dereference_array. |
||
526 | * This takes care of both arrays and matrices, since ir_dereference_array |
||
527 | * treats a matrix like an array of its column vectors. |
||
528 | * |
||
529 | * \param gs_input_toplevel should be set to true if we are lowering geometry |
||
530 | * shader inputs, and we are currently lowering the whole input variable |
||
531 | * (i.e. we are lowering the array whose index selects the vertex). |
||
532 | * |
||
533 | * \param vertex_index: if we are lowering geometry shader inputs, and the |
||
534 | * level of the array that we are currently lowering is *not* the top level, |
||
535 | * then this indicates which vertex we are currently lowering. Otherwise it |
||
536 | * is ignored. |
||
537 | */ |
||
538 | unsigned |
||
539 | lower_packed_varyings_visitor::lower_arraylike(ir_rvalue *rvalue, |
||
540 | unsigned array_size, |
||
541 | unsigned fine_location, |
||
542 | ir_variable *unpacked_var, |
||
543 | const char *name, |
||
544 | bool gs_input_toplevel, |
||
545 | unsigned vertex_index) |
||
546 | { |
||
547 | for (unsigned i = 0; i < array_size; i++) { |
||
548 | if (i != 0) |
||
549 | rvalue = rvalue->clone(this->mem_ctx, NULL); |
||
550 | ir_constant *constant = new(this->mem_ctx) ir_constant(i); |
||
551 | ir_dereference_array *dereference_array = new(this->mem_ctx) |
||
552 | ir_dereference_array(rvalue, constant); |
||
553 | if (gs_input_toplevel) { |
||
554 | /* Geometry shader inputs are a special case. Instead of storing |
||
555 | * each element of the array at a different location, all elements |
||
556 | * are at the same location, but with a different vertex index. |
||
557 | */ |
||
558 | (void) this->lower_rvalue(dereference_array, fine_location, |
||
559 | unpacked_var, name, false, i); |
||
560 | } else { |
||
561 | char *subscripted_name |
||
562 | = ralloc_asprintf(this->mem_ctx, "%s[%d]", name, i); |
||
563 | fine_location = |
||
564 | this->lower_rvalue(dereference_array, fine_location, |
||
565 | unpacked_var, subscripted_name, |
||
566 | false, vertex_index); |
||
567 | } |
||
568 | } |
||
569 | return fine_location; |
||
570 | } |
||
571 | |||
572 | /** |
||
573 | * Retrieve the packed varying corresponding to the given varying location. |
||
574 | * If no packed varying has been created for the given varying location yet, |
||
575 | * create it and add it to the shader before returning it. |
||
576 | * |
||
577 | * The newly created varying inherits its interpolation parameters from \c |
||
578 | * unpacked_var. Its base type is ivec4 if we are lowering a flat varying, |
||
579 | * vec4 otherwise. |
||
580 | * |
||
581 | * \param vertex_index: if we are lowering geometry shader inputs, then this |
||
582 | * indicates which vertex we are currently lowering. Otherwise it is ignored. |
||
583 | */ |
||
584 | ir_dereference * |
||
585 | lower_packed_varyings_visitor::get_packed_varying_deref( |
||
586 | unsigned location, ir_variable *unpacked_var, const char *name, |
||
587 | unsigned vertex_index) |
||
588 | { |
||
589 | unsigned slot = location - VARYING_SLOT_VAR0; |
||
590 | assert(slot < locations_used); |
||
591 | if (this->packed_varyings[slot] == NULL) { |
||
592 | char *packed_name = ralloc_asprintf(this->mem_ctx, "packed:%s", name); |
||
593 | const glsl_type *packed_type; |
||
594 | if (unpacked_var->data.interpolation == INTERP_QUALIFIER_FLAT) |
||
595 | packed_type = glsl_type::ivec4_type; |
||
596 | else |
||
597 | packed_type = glsl_type::vec4_type; |
||
598 | if (this->gs_input_vertices != 0) { |
||
599 | packed_type = |
||
600 | glsl_type::get_array_instance(packed_type, |
||
601 | this->gs_input_vertices); |
||
602 | } |
||
603 | ir_variable *packed_var = new(this->mem_ctx) |
||
604 | ir_variable(packed_type, packed_name, this->mode); |
||
605 | if (this->gs_input_vertices != 0) { |
||
606 | /* Prevent update_array_sizes() from messing with the size of the |
||
607 | * array. |
||
608 | */ |
||
609 | packed_var->data.max_array_access = this->gs_input_vertices - 1; |
||
610 | } |
||
611 | packed_var->data.centroid = unpacked_var->data.centroid; |
||
612 | packed_var->data.sample = unpacked_var->data.sample; |
||
613 | packed_var->data.interpolation = unpacked_var->data.interpolation; |
||
614 | packed_var->data.location = location; |
||
615 | unpacked_var->insert_before(packed_var); |
||
616 | this->packed_varyings[slot] = packed_var; |
||
617 | } else { |
||
618 | /* For geometry shader inputs, only update the packed variable name the |
||
619 | * first time we visit each component. |
||
620 | */ |
||
621 | if (this->gs_input_vertices == 0 || vertex_index == 0) { |
||
622 | ralloc_asprintf_append((char **) &this->packed_varyings[slot]->name, |
||
623 | ",%s", name); |
||
624 | } |
||
625 | } |
||
626 | |||
627 | ir_dereference *deref = new(this->mem_ctx) |
||
628 | ir_dereference_variable(this->packed_varyings[slot]); |
||
629 | if (this->gs_input_vertices != 0) { |
||
630 | /* When lowering GS inputs, the packed variable is an array, so we need |
||
631 | * to dereference it using vertex_index. |
||
632 | */ |
||
633 | ir_constant *constant = new(this->mem_ctx) ir_constant(vertex_index); |
||
634 | deref = new(this->mem_ctx) ir_dereference_array(deref, constant); |
||
635 | } |
||
636 | return deref; |
||
637 | } |
||
638 | |||
639 | bool |
||
640 | lower_packed_varyings_visitor::needs_lowering(ir_variable *var) |
||
641 | { |
||
642 | /* Things composed of vec4's and varyings with explicitly assigned |
||
643 | * locations don't need lowering. Everything else does. |
||
644 | */ |
||
645 | if (var->data.explicit_location) |
||
646 | return false; |
||
647 | |||
648 | const glsl_type *type = var->type->without_array(); |
||
649 | if (type->vector_elements == 4 && !type->is_double()) |
||
650 | return false; |
||
651 | return true; |
||
652 | } |
||
653 | |||
654 | |||
655 | /** |
||
656 | * Visitor that splices varying packing code before every use of EmitVertex() |
||
657 | * in a geometry shader. |
||
658 | */ |
||
659 | class lower_packed_varyings_gs_splicer : public ir_hierarchical_visitor |
||
660 | { |
||
661 | public: |
||
662 | explicit lower_packed_varyings_gs_splicer(void *mem_ctx, |
||
663 | const exec_list *instructions); |
||
664 | |||
665 | virtual ir_visitor_status visit_leave(ir_emit_vertex *ev); |
||
666 | |||
667 | private: |
||
668 | /** |
||
669 | * Memory context used to allocate new instructions for the shader. |
||
670 | */ |
||
671 | void * const mem_ctx; |
||
672 | |||
673 | /** |
||
674 | * Instructions that should be spliced into place before each EmitVertex() |
||
675 | * call. |
||
676 | */ |
||
677 | const exec_list *instructions; |
||
678 | }; |
||
679 | |||
680 | |||
681 | lower_packed_varyings_gs_splicer::lower_packed_varyings_gs_splicer( |
||
682 | void *mem_ctx, const exec_list *instructions) |
||
683 | : mem_ctx(mem_ctx), instructions(instructions) |
||
684 | { |
||
685 | } |
||
686 | |||
687 | |||
688 | ir_visitor_status |
||
689 | lower_packed_varyings_gs_splicer::visit_leave(ir_emit_vertex *ev) |
||
690 | { |
||
691 | foreach_in_list(ir_instruction, ir, this->instructions) { |
||
692 | ev->insert_before(ir->clone(this->mem_ctx, NULL)); |
||
693 | } |
||
694 | return visit_continue; |
||
695 | } |
||
696 | |||
697 | |||
698 | void |
||
699 | lower_packed_varyings(void *mem_ctx, unsigned locations_used, |
||
700 | ir_variable_mode mode, unsigned gs_input_vertices, |
||
701 | gl_shader *shader) |
||
702 | { |
||
703 | exec_list *instructions = shader->ir; |
||
704 | ir_function *main_func = shader->symbols->get_function("main"); |
||
705 | exec_list void_parameters; |
||
706 | ir_function_signature *main_func_sig |
||
707 | = main_func->matching_signature(NULL, &void_parameters, false); |
||
708 | exec_list new_instructions, new_variables; |
||
709 | lower_packed_varyings_visitor visitor(mem_ctx, locations_used, mode, |
||
710 | gs_input_vertices, |
||
711 | &new_instructions, |
||
712 | &new_variables); |
||
713 | visitor.run(instructions); |
||
714 | if (mode == ir_var_shader_out) { |
||
715 | if (shader->Stage == MESA_SHADER_GEOMETRY) { |
||
716 | /* For geometry shaders, outputs need to be lowered before each call |
||
717 | * to EmitVertex() |
||
718 | */ |
||
719 | lower_packed_varyings_gs_splicer splicer(mem_ctx, &new_instructions); |
||
720 | |||
721 | /* Add all the variables in first. */ |
||
722 | main_func_sig->body.head->insert_before(&new_variables); |
||
723 | |||
724 | /* Now update all the EmitVertex instances */ |
||
725 | splicer.run(instructions); |
||
726 | } else { |
||
727 | /* For other shader types, outputs need to be lowered at the end of |
||
728 | * main() |
||
729 | */ |
||
730 | main_func_sig->body.append_list(&new_variables); |
||
731 | main_func_sig->body.append_list(&new_instructions); |
||
732 | } |
||
733 | } else { |
||
734 | /* Shader inputs need to be lowered at the beginning of main() */ |
||
735 | main_func_sig->body.head->insert_before(&new_instructions); |
||
736 | main_func_sig->body.head->insert_before(&new_variables); |
||
737 | } |
||
738 | }>>>>>>=>=>> |