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 
prealloc_reg(struct brw_wm_compile * c,struct brw_wm_value * value,GLuint reg)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  */
init_registers(struct brw_wm_compile * c)70 static void init_registers( struct brw_wm_compile *c )
71 {
72    struct brw_context *brw = c->func.brw;
73    struct intel_context *intel = &brw->intel;
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->nr_payload_regs + 1) / 2; 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    if (intel->gen >= 6) {
88       for (unsigned int j = 0; j < FRAG_ATTRIB_MAX; j++) {
89 	 if (brw->fragment_program->Base.InputsRead & BITFIELD64_BIT(j)) {
90 	    nr_interp_regs++;
91 	    prealloc_reg(c, &c->payload.input_interp[j], i++);
92 	 }
93       }
94    } else {
95       for (j = 0; j < VERT_RESULT_MAX; j++) {
96          /* Point size is packed into the header, not as a general attribute */
97          if (j == VERT_RESULT_PSIZ)
98             continue;
99 
100 	 if (c->key.vp_outputs_written & BITFIELD64_BIT(j)) {
101 	    int fp_index = _mesa_vert_result_to_frag_attrib(j);
102 
103 	    nr_interp_regs++;
104 
105 	    /* The back color slot is skipped when the front color is
106 	     * also written to.  In addition, some slots can be
107 	     * written in the vertex shader and not read in the
108 	     * fragment shader.  So the register number must always be
109 	     * incremented, mapped or not.
110 	     */
111 	    if (fp_index >= 0)
112 	       prealloc_reg(c, &c->payload.input_interp[fp_index], i);
113             i++;
114 	 }
115       }
116       assert(nr_interp_regs >= 1);
117    }
118 
119 
120    c->prog_data.first_curbe_grf = ALIGN(c->nr_payload_regs, 2);
121    c->prog_data.urb_read_length = nr_interp_regs * 2;
122    c->prog_data.curb_read_length = c->nr_creg * 2;
123 
124    c->max_wm_grf = i * 2;
125 }
126 
127 
128 /* Update the nextuse value for each register in our file.
129  */
update_register_usage(struct brw_wm_compile * c,GLuint thisinsn)130 static void update_register_usage(struct brw_wm_compile *c,
131 				  GLuint thisinsn)
132 {
133    GLuint i;
134 
135    for (i = 1; i < c->grf_limit; i++) {
136       struct brw_wm_grf *grf = &c->pass2_grf[i];
137 
138       /* Only search those which can change:
139        */
140       if (grf->nextuse < thisinsn) {
141 	 const struct brw_wm_ref *ref = grf->value->lastuse;
142 
143 	 /* Has last use of value been passed?
144 	  */
145 	 if (ref->insn < thisinsn) {
146 	    grf->value->resident = 0;
147 	    grf->value = 0;
148 	    grf->nextuse = BRW_WM_MAX_INSN;
149 	 }
150 	 else {
151 	    /* Else loop through chain to update:
152 	     */
153 	    while (ref->prevuse && ref->prevuse->insn >= thisinsn)
154 	       ref = ref->prevuse;
155 
156 	    grf->nextuse = ref->insn;
157 	 }
158       }
159    }
160 }
161 
162 
spill_value(struct brw_wm_compile * c,struct brw_wm_value * value)163 static void spill_value(struct brw_wm_compile *c,
164 			struct brw_wm_value *value)
165 {
166    /* Allocate a spill slot.  Note that allocations start from 0x40 -
167     * the first slot is reserved to mean "undef" in brw_wm_emit.c
168     */
169    if (!value->spill_slot) {
170       c->last_scratch += 0x40;
171       value->spill_slot = c->last_scratch;
172    }
173 
174    /* The spill will be done in brw_wm_emit.c immediately after the
175     * value is calculated, so we can just take this reg without any
176     * further work.
177     */
178    value->resident->value = NULL;
179    value->resident->nextuse = BRW_WM_MAX_INSN;
180    value->resident = NULL;
181 }
182 
183 
184 
185 /* Search for contiguous region with the most distant nearest
186  * member.  Free regs count as very distant.
187  *
188  * TODO: implement spill-to-reg so that we can rearrange discontigous
189  * free regs and then spill the oldest non-free regs in sequence.
190  * This would mean inserting instructions in this pass.
191  */
search_contiguous_regs(struct brw_wm_compile * c,GLuint nr,GLuint thisinsn)192 static GLuint search_contiguous_regs(struct brw_wm_compile *c,
193 				     GLuint nr,
194 				     GLuint thisinsn)
195 {
196    struct brw_wm_grf *grf = c->pass2_grf;
197    GLuint furthest = 0;
198    GLuint reg = 0;
199    GLuint i, j;
200 
201    /* Start search at 1: r0 is special and can't be used or spilled.
202     */
203    for (i = 1; i < c->grf_limit && furthest < BRW_WM_MAX_INSN; i++) {
204       GLuint group_nextuse = BRW_WM_MAX_INSN;
205 
206       for (j = 0; j < nr; j++) {
207 	 if (grf[i+j].nextuse < group_nextuse)
208 	    group_nextuse = grf[i+j].nextuse;
209       }
210 
211       if (group_nextuse > furthest) {
212 	 furthest = group_nextuse;
213 	 reg = i;
214       }
215    }
216 
217    assert(furthest != thisinsn);
218 
219    /* Any non-empty regs will need to be spilled:
220     */
221    for (j = 0; j < nr; j++)
222       if (grf[reg+j].value)
223 	 spill_value(c, grf[reg+j].value);
224 
225    return reg;
226 }
227 
228 
alloc_contiguous_dest(struct brw_wm_compile * c,struct brw_wm_value * dst[],GLuint nr,GLuint thisinsn)229 static void alloc_contiguous_dest(struct brw_wm_compile *c,
230 				  struct brw_wm_value *dst[],
231 				  GLuint nr,
232 				  GLuint thisinsn)
233 {
234    GLuint reg = search_contiguous_regs(c, nr, thisinsn);
235    GLuint i;
236 
237    for (i = 0; i < nr; i++) {
238       if (!dst[i]) {
239 	 /* Need to grab a dummy value in TEX case.  Don't introduce
240 	  * it into the tracking scheme.
241 	  */
242 	 dst[i] = &c->vreg[c->nr_vreg++];
243       }
244       else {
245 	 assert(!dst[i]->resident);
246 	 assert(c->pass2_grf[reg+i].nextuse != thisinsn);
247 
248 	 c->pass2_grf[reg+i].value = dst[i];
249 	 c->pass2_grf[reg+i].nextuse = thisinsn;
250 
251 	 dst[i]->resident = &c->pass2_grf[reg+i];
252       }
253 
254       dst[i]->hw_reg = brw_vec8_grf((reg+i)*2, 0);
255    }
256 
257    if ((reg+nr)*2 > c->max_wm_grf)
258       c->max_wm_grf = (reg+nr) * 2;
259 }
260 
261 
load_args(struct brw_wm_compile * c,struct brw_wm_instruction * inst)262 static void load_args(struct brw_wm_compile *c,
263 		      struct brw_wm_instruction *inst)
264 {
265    GLuint thisinsn = inst - c->instruction;
266    GLuint i,j;
267 
268    for (i = 0; i < 3; i++) {
269       for (j = 0; j < 4; j++) {
270 	 struct brw_wm_ref *ref = inst->src[i][j];
271 
272 	 if (ref) {
273 	    if (!ref->value->resident) {
274 	       /* Need to bring the value in from scratch space.  The code for
275 		* this will be done in brw_wm_emit.c, here we just do the
276 		* register allocation and mark the ref as requiring a fill.
277 		*/
278 	       GLuint reg = search_contiguous_regs(c, 1, thisinsn);
279 
280 	       c->pass2_grf[reg].value = ref->value;
281 	       c->pass2_grf[reg].nextuse = thisinsn;
282 
283 	       ref->value->resident = &c->pass2_grf[reg];
284 
285 	       /* Note that a fill is required:
286 		*/
287 	       ref->unspill_reg = reg*2;
288 	    }
289 
290 	    /* Adjust the hw_reg to point at the value's current location:
291 	     */
292 	    assert(ref->value == ref->value->resident->value);
293 	    ref->hw_reg.nr += (ref->value->resident - c->pass2_grf) * 2;
294 	 }
295       }
296    }
297 }
298 
299 
300 
301 /* Step 3: Work forwards once again.  Perform register allocations,
302  * taking into account instructions like TEX which require contiguous
303  * result registers.  Where necessary spill registers to scratch space
304  * and reload later.
305  */
brw_wm_pass2(struct brw_wm_compile * c)306 void brw_wm_pass2( struct brw_wm_compile *c )
307 {
308    GLuint insn;
309    GLuint i;
310 
311    init_registers(c);
312 
313    for (insn = 0; insn < c->nr_insns; insn++) {
314       struct brw_wm_instruction *inst = &c->instruction[insn];
315 
316       /* Update registers' nextuse values:
317        */
318       update_register_usage(c, insn);
319 
320       /* May need to unspill some args.
321        */
322       load_args(c, inst);
323 
324       /* Allocate registers to hold results:
325        */
326       switch (inst->opcode) {
327       case OPCODE_TEX:
328       case OPCODE_TXB:
329       case OPCODE_TXP:
330 	 alloc_contiguous_dest(c, inst->dst, 4, insn);
331 	 break;
332 
333       default:
334 	 for (i = 0; i < 4; i++) {
335 	    if (inst->writemask & (1<<i)) {
336 	       assert(inst->dst[i]);
337 	       alloc_contiguous_dest(c, &inst->dst[i], 1, insn);
338 	    }
339 	 }
340 	 break;
341       }
342 
343       if (TEST_DST_SPILLS && inst->opcode != WM_PIXELXY) {
344 	 for (i = 0; i < 4; i++)
345 	    if (inst->dst[i])
346 	       spill_value(c, inst->dst[i]);
347       }
348    }
349 
350    if (unlikely(INTEL_DEBUG & DEBUG_WM)) {
351       brw_wm_print_program(c, "pass2");
352    }
353 
354    c->state = PASS2_DONE;
355 
356    if (unlikely(INTEL_DEBUG & DEBUG_WM)) {
357        brw_wm_print_program(c, "pass2/done");
358    }
359 }
360