Merge branch 'master' into gallium-texture-transfer
[mesa.git] / src / mesa / drivers / dri / i965 / brw_wm_pass2.c
1 /*
2 Copyright (C) Intel Corp. 2006. All Rights Reserved.
3 Intel funded Tungsten Graphics (http://www.tungstengraphics.com) to
4 develop this 3D driver.
5
6 Permission is hereby granted, free of charge, to any person obtaining
7 a copy of this software and associated documentation files (the
8 "Software"), to deal in the Software without restriction, including
9 without limitation the rights to use, copy, modify, merge, publish,
10 distribute, sublicense, and/or sell copies of the Software, and to
11 permit persons to whom the Software is furnished to do so, subject to
12 the following conditions:
13
14 The above copyright notice and this permission notice (including the
15 next paragraph) shall be included in all copies or substantial
16 portions of the Software.
17
18 THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
19 EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
21 IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
22 LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
23 OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
24 WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
25
26 **********************************************************************/
27 /*
28 * Authors:
29 * Keith Whitwell <keith@tungstengraphics.com>
30 */
31
32
33 #include "brw_context.h"
34 #include "brw_wm.h"
35
36
37 /* Use these to force spilling so that that functionality can be
38 * tested with known-good examples rather than having to construct new
39 * tests.
40 */
41 #define TEST_PAYLOAD_SPILLS 0
42 #define TEST_DST_SPILLS 0
43
44 static void spill_value(struct brw_wm_compile *c,
45 struct brw_wm_value *value);
46
47 static void prealloc_reg(struct brw_wm_compile *c,
48 struct brw_wm_value *value,
49 GLuint reg)
50 {
51 if (value->lastuse) {
52 /* Set nextuse to zero, it will be corrected by
53 * update_register_usage().
54 */
55 c->pass2_grf[reg].value = value;
56 c->pass2_grf[reg].nextuse = 0;
57
58 value->resident = &c->pass2_grf[reg];
59 value->hw_reg = brw_vec8_grf(reg*2, 0);
60
61 if (TEST_PAYLOAD_SPILLS)
62 spill_value(c, value);
63 }
64 }
65
66
67 /* Initialize all the register values. Do the initial setup
68 * calculations for interpolants.
69 */
70 static void init_registers( struct brw_wm_compile *c )
71 {
72 struct brw_context *brw = c->func.brw;
73 GLuint inputs = (brw->vs.prog_data->outputs_written & DO_SETUP_BITS);
74 GLuint nr_interp_regs = 0;
75 GLuint i = 0;
76 GLuint j;
77
78 for (j = 0; j < c->grf_limit; j++)
79 c->pass2_grf[j].nextuse = BRW_WM_MAX_INSN;
80
81 for (j = 0; j < c->key.nr_depth_regs; j++)
82 prealloc_reg(c, &c->payload.depth[j], i++);
83
84 for (j = 0; j < c->nr_creg; j++)
85 prealloc_reg(c, &c->creg[j], i++);
86
87 for (j = 0; j < FRAG_ATTRIB_MAX; j++) {
88 if (inputs & (1<<j)) {
89 /* index for vs output and ps input are not the same
90 in shader varying */
91 GLuint index;
92 if (j > FRAG_ATTRIB_VAR0)
93 index = j - (VERT_RESULT_VAR0 - FRAG_ATTRIB_VAR0);
94 else
95 index = j;
96 nr_interp_regs++;
97 prealloc_reg(c, &c->payload.input_interp[index], i++);
98 }
99 }
100
101 assert(nr_interp_regs >= 1);
102
103 c->prog_data.first_curbe_grf = c->key.nr_depth_regs * 2;
104 c->prog_data.urb_read_length = nr_interp_regs * 2;
105 c->prog_data.curb_read_length = c->nr_creg * 2;
106
107 c->max_wm_grf = i * 2;
108 }
109
110
111 /* Update the nextuse value for each register in our file.
112 */
113 static void update_register_usage(struct brw_wm_compile *c,
114 GLuint thisinsn)
115 {
116 GLuint i;
117
118 for (i = 1; i < c->grf_limit; i++) {
119 struct brw_wm_grf *grf = &c->pass2_grf[i];
120
121 /* Only search those which can change:
122 */
123 if (grf->nextuse < thisinsn) {
124 const struct brw_wm_ref *ref = grf->value->lastuse;
125
126 /* Has last use of value been passed?
127 */
128 if (ref->insn < thisinsn) {
129 grf->value->resident = 0;
130 grf->value = 0;
131 grf->nextuse = BRW_WM_MAX_INSN;
132 }
133 else {
134 /* Else loop through chain to update:
135 */
136 while (ref->prevuse && ref->prevuse->insn >= thisinsn)
137 ref = ref->prevuse;
138
139 grf->nextuse = ref->insn;
140 }
141 }
142 }
143 }
144
145
146 static void spill_value(struct brw_wm_compile *c,
147 struct brw_wm_value *value)
148 {
149 /* Allocate a spill slot. Note that allocations start from 0x40 -
150 * the first slot is reserved to mean "undef" in brw_wm_emit.c
151 */
152 if (!value->spill_slot) {
153 c->last_scratch += 0x40;
154 value->spill_slot = c->last_scratch;
155 }
156
157 /* The spill will be done in brw_wm_emit.c immediately after the
158 * value is calculated, so we can just take this reg without any
159 * further work.
160 */
161 value->resident->value = NULL;
162 value->resident->nextuse = BRW_WM_MAX_INSN;
163 value->resident = NULL;
164 }
165
166
167
168 /* Search for contiguous region with the most distant nearest
169 * member. Free regs count as very distant.
170 *
171 * TODO: implement spill-to-reg so that we can rearrange discontigous
172 * free regs and then spill the oldest non-free regs in sequence.
173 * This would mean inserting instructions in this pass.
174 */
175 static GLuint search_contiguous_regs(struct brw_wm_compile *c,
176 GLuint nr,
177 GLuint thisinsn)
178 {
179 struct brw_wm_grf *grf = c->pass2_grf;
180 GLuint furthest = 0;
181 GLuint reg = 0;
182 GLuint i, j;
183
184 /* Start search at 1: r0 is special and can't be used or spilled.
185 */
186 for (i = 1; i < c->grf_limit && furthest < BRW_WM_MAX_INSN; i++) {
187 GLuint group_nextuse = BRW_WM_MAX_INSN;
188
189 for (j = 0; j < nr; j++) {
190 if (grf[i+j].nextuse < group_nextuse)
191 group_nextuse = grf[i+j].nextuse;
192 }
193
194 if (group_nextuse > furthest) {
195 furthest = group_nextuse;
196 reg = i;
197 }
198 }
199
200 assert(furthest != thisinsn);
201
202 /* Any non-empty regs will need to be spilled:
203 */
204 for (j = 0; j < nr; j++)
205 if (grf[reg+j].value)
206 spill_value(c, grf[reg+j].value);
207
208 return reg;
209 }
210
211
212 static void alloc_contiguous_dest(struct brw_wm_compile *c,
213 struct brw_wm_value *dst[],
214 GLuint nr,
215 GLuint thisinsn)
216 {
217 GLuint reg = search_contiguous_regs(c, nr, thisinsn);
218 GLuint i;
219
220 for (i = 0; i < nr; i++) {
221 if (!dst[i]) {
222 /* Need to grab a dummy value in TEX case. Don't introduce
223 * it into the tracking scheme.
224 */
225 dst[i] = &c->vreg[c->nr_vreg++];
226 }
227 else {
228 assert(!dst[i]->resident);
229 assert(c->pass2_grf[reg+i].nextuse != thisinsn);
230
231 c->pass2_grf[reg+i].value = dst[i];
232 c->pass2_grf[reg+i].nextuse = thisinsn;
233
234 dst[i]->resident = &c->pass2_grf[reg+i];
235 }
236
237 dst[i]->hw_reg = brw_vec8_grf((reg+i)*2, 0);
238 }
239
240 if ((reg+nr)*2 > c->max_wm_grf)
241 c->max_wm_grf = (reg+nr) * 2;
242 }
243
244
245 static void load_args(struct brw_wm_compile *c,
246 struct brw_wm_instruction *inst)
247 {
248 GLuint thisinsn = inst - c->instruction;
249 GLuint i,j;
250
251 for (i = 0; i < 3; i++) {
252 for (j = 0; j < 4; j++) {
253 struct brw_wm_ref *ref = inst->src[i][j];
254
255 if (ref) {
256 if (!ref->value->resident) {
257 /* Need to bring the value in from scratch space. The code for
258 * this will be done in brw_wm_emit.c, here we just do the
259 * register allocation and mark the ref as requiring a fill.
260 */
261 GLuint reg = search_contiguous_regs(c, 1, thisinsn);
262
263 c->pass2_grf[reg].value = ref->value;
264 c->pass2_grf[reg].nextuse = thisinsn;
265
266 ref->value->resident = &c->pass2_grf[reg];
267
268 /* Note that a fill is required:
269 */
270 ref->unspill_reg = reg*2;
271 }
272
273 /* Adjust the hw_reg to point at the value's current location:
274 */
275 assert(ref->value == ref->value->resident->value);
276 ref->hw_reg.nr += (ref->value->resident - c->pass2_grf) * 2;
277 }
278 }
279 }
280 }
281
282
283
284 /* Step 3: Work forwards once again. Perform register allocations,
285 * taking into account instructions like TEX which require contiguous
286 * result registers. Where necessary spill registers to scratch space
287 * and reload later.
288 */
289 void brw_wm_pass2( struct brw_wm_compile *c )
290 {
291 GLuint insn;
292 GLuint i;
293
294 init_registers(c);
295
296 for (insn = 0; insn < c->nr_insns; insn++) {
297 struct brw_wm_instruction *inst = &c->instruction[insn];
298
299 /* Update registers' nextuse values:
300 */
301 update_register_usage(c, insn);
302
303 /* May need to unspill some args.
304 */
305 load_args(c, inst);
306
307 /* Allocate registers to hold results:
308 */
309 switch (inst->opcode) {
310 case OPCODE_TEX:
311 case OPCODE_TXB:
312 case OPCODE_TXP:
313 alloc_contiguous_dest(c, inst->dst, 4, insn);
314 break;
315
316 default:
317 for (i = 0; i < 4; i++) {
318 if (inst->writemask & (1<<i)) {
319 assert(inst->dst[i]);
320 alloc_contiguous_dest(c, &inst->dst[i], 1, insn);
321 }
322 }
323 break;
324 }
325
326 if (TEST_DST_SPILLS && inst->opcode != WM_PIXELXY) {
327 for (i = 0; i < 4; i++)
328 if (inst->dst[i])
329 spill_value(c, inst->dst[i]);
330 }
331 }
332
333 if (INTEL_DEBUG & DEBUG_WM) {
334 brw_wm_print_program(c, "pass2");
335 }
336
337 c->state = PASS2_DONE;
338
339 if (INTEL_DEBUG & DEBUG_WM) {
340 brw_wm_print_program(c, "pass2/done");
341 }
342 }