i915_fpc_emit.c revision 2bc5e0e97ba7b6c32f6ff90cb90448173d74b89b 
1/**************************************************************************
2 *
3 * Copyright 2003 Tungsten Graphics, Inc., Cedar Park, Texas.
4 * All Rights Reserved.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * 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, sub license, 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 portions
16 * of the Software.
17 *
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
25 *
26 **************************************************************************/
27
28#include "i915_reg.h"
29#include "i915_context.h"
30#include "i915_fpc.h"
31#include "util/u_math.h"
32
33
34#define A0_DEST( reg ) (((reg)&UREG_TYPE_NR_MASK)>>UREG_A0_DEST_SHIFT_LEFT)
35#define D0_DEST( reg ) (((reg)&UREG_TYPE_NR_MASK)>>UREG_A0_DEST_SHIFT_LEFT)
36#define T0_DEST( reg ) (((reg)&UREG_TYPE_NR_MASK)>>UREG_A0_DEST_SHIFT_LEFT)
37#define A0_SRC0( reg ) (((reg)&UREG_MASK)>>UREG_A0_SRC0_SHIFT_LEFT)
38#define A1_SRC0( reg ) (((reg)&UREG_MASK)<<UREG_A1_SRC0_SHIFT_RIGHT)
39#define A1_SRC1( reg ) (((reg)&UREG_MASK)>>UREG_A1_SRC1_SHIFT_LEFT)
40#define A2_SRC1( reg ) (((reg)&UREG_MASK)<<UREG_A2_SRC1_SHIFT_RIGHT)
41#define A2_SRC2( reg ) (((reg)&UREG_MASK)>>UREG_A2_SRC2_SHIFT_LEFT)
42
43/* These are special, and don't have swizzle/negate bits.
44 */
45#define T0_SAMPLER( reg )     (GET_UREG_NR(reg)<<T0_SAMPLER_NR_SHIFT)
46#define T1_ADDRESS_REG( reg ) ((GET_UREG_NR(reg)<<T1_ADDRESS_REG_NR_SHIFT) | \
47			       (GET_UREG_TYPE(reg)<<T1_ADDRESS_REG_TYPE_SHIFT))
48
49
50/* Macros for translating UREG's into the various register fields used
51 * by the I915 programmable unit.
52 */
53#define UREG_A0_DEST_SHIFT_LEFT  (UREG_TYPE_SHIFT - A0_DEST_TYPE_SHIFT)
54#define UREG_A0_SRC0_SHIFT_LEFT  (UREG_TYPE_SHIFT - A0_SRC0_TYPE_SHIFT)
55#define UREG_A1_SRC0_SHIFT_RIGHT (A1_SRC0_CHANNEL_W_SHIFT - UREG_CHANNEL_W_SHIFT)
56#define UREG_A1_SRC1_SHIFT_LEFT  (UREG_TYPE_SHIFT - A1_SRC1_TYPE_SHIFT)
57#define UREG_A2_SRC1_SHIFT_RIGHT (A2_SRC1_CHANNEL_W_SHIFT - UREG_CHANNEL_W_SHIFT)
58#define UREG_A2_SRC2_SHIFT_LEFT  (UREG_TYPE_SHIFT - A2_SRC2_TYPE_SHIFT)
59
60#define UREG_MASK         0xffffff00
61#define UREG_TYPE_NR_MASK ((REG_TYPE_MASK << UREG_TYPE_SHIFT) | \
62  			   (REG_NR_MASK << UREG_NR_SHIFT))
63
64
65uint
66i915_get_temp(struct i915_fp_compile *p)
67{
68   int bit = ffs(~p->temp_flag);
69   if (!bit) {
70      i915_program_error(p, "i915_get_temp: out of temporaries");
71      return 0;
72   }
73
74   p->temp_flag |= 1 << (bit - 1);
75   return bit - 1;
76}
77
78
79static void
80i915_release_temp(struct i915_fp_compile *p, int reg)
81{
82   p->temp_flag &= ~(1 << reg);
83}
84
85
86/**
87 * Get unpreserved temporary, a temp whose value is not preserved between
88 * PS program phases.
89 */
90uint
91i915_get_utemp(struct i915_fp_compile * p)
92{
93   int bit = ffs(~p->utemp_flag);
94   if (!bit) {
95      i915_program_error(p, "i915_get_utemp: out of temporaries");
96      return 0;
97   }
98
99   p->utemp_flag |= 1 << (bit - 1);
100   return UREG(REG_TYPE_U, (bit - 1));
101}
102
103void
104i915_release_utemps(struct i915_fp_compile *p)
105{
106   p->utemp_flag = ~0x7;
107}
108
109
110uint
111i915_emit_decl(struct i915_fp_compile *p,
112               uint type, uint nr, uint d0_flags)
113{
114   uint reg = UREG(type, nr);
115
116   if (type == REG_TYPE_T) {
117      if (p->decl_t & (1 << nr))
118         return reg;
119
120      p->decl_t |= (1 << nr);
121   }
122   else if (type == REG_TYPE_S) {
123      if (p->decl_s & (1 << nr))
124         return reg;
125
126      p->decl_s |= (1 << nr);
127   }
128   else
129      return reg;
130
131   if (p->decl< p->declarations + I915_PROGRAM_SIZE) {
132      *(p->decl++) = (D0_DCL | D0_DEST(reg) | d0_flags);
133      *(p->decl++) = D1_MBZ;
134      *(p->decl++) = D2_MBZ;
135   }
136   else
137      i915_program_error(p, "Out of declarations");
138
139   p->nr_decl_insn++;
140   return reg;
141}
142
143uint
144i915_emit_arith(struct i915_fp_compile * p,
145                uint op,
146                uint dest,
147                uint mask,
148                uint saturate, uint src0, uint src1, uint src2)
149{
150   uint c[3];
151   uint nr_const = 0;
152
153   assert(GET_UREG_TYPE(dest) != REG_TYPE_CONST);
154   dest = UREG(GET_UREG_TYPE(dest), GET_UREG_NR(dest));
155   assert(dest);
156
157   if (GET_UREG_TYPE(src0) == REG_TYPE_CONST)
158      c[nr_const++] = 0;
159   if (GET_UREG_TYPE(src1) == REG_TYPE_CONST)
160      c[nr_const++] = 1;
161   if (GET_UREG_TYPE(src2) == REG_TYPE_CONST)
162      c[nr_const++] = 2;
163
164   /* Recursively call this function to MOV additional const values
165    * into temporary registers.  Use utemp registers for this -
166    * currently shouldn't be possible to run out, but keep an eye on
167    * this.
168    */
169   if (nr_const > 1) {
170      uint s[3], first, i, old_utemp_flag;
171
172      s[0] = src0;
173      s[1] = src1;
174      s[2] = src2;
175      old_utemp_flag = p->utemp_flag;
176
177      first = GET_UREG_NR(s[c[0]]);
178      for (i = 1; i < nr_const; i++) {
179         if (GET_UREG_NR(s[c[i]]) != first) {
180            uint tmp = i915_get_utemp(p);
181
182            i915_emit_arith(p, A0_MOV, tmp, A0_DEST_CHANNEL_ALL, 0,
183                            s[c[i]], 0, 0);
184            s[c[i]] = tmp;
185         }
186      }
187
188      src0 = s[0];
189      src1 = s[1];
190      src2 = s[2];
191      p->utemp_flag = old_utemp_flag;   /* restore */
192   }
193
194   if (p->csr< p->program + I915_PROGRAM_SIZE) {
195      *(p->csr++) = (op | A0_DEST(dest) | mask | saturate | A0_SRC0(src0));
196      *(p->csr++) = (A1_SRC0(src0) | A1_SRC1(src1));
197      *(p->csr++) = (A2_SRC1(src1) | A2_SRC2(src2));
198   }
199   else
200      i915_program_error(p, "Out of instructions");
201
202   if (GET_UREG_TYPE(dest) == REG_TYPE_R)
203      p->register_phases[GET_UREG_NR(dest)] = p->nr_tex_indirect;
204
205   p->nr_alu_insn++;
206   return dest;
207}
208
209
210/**
211 * Emit a texture load or texkill instruction.
212 * \param dest  the dest i915 register
213 * \param destmask  the dest register writemask
214 * \param sampler  the i915 sampler register
215 * \param coord  the i915 source texcoord operand
216 * \param opcode  the instruction opcode
217 */
218uint i915_emit_texld( struct i915_fp_compile *p,
219			uint dest,
220			uint destmask,
221			uint sampler,
222			uint coord,
223			uint opcode )
224{
225   const uint k = UREG(GET_UREG_TYPE(coord), GET_UREG_NR(coord));
226   int temp = -1;
227
228   if (coord != k) {
229      /* texcoord is swizzled or negated.  Need to allocate a new temporary
230       * register (a utemp / unpreserved temp) won't do.
231       */
232      uint tempReg;
233
234      temp = i915_get_temp(p);           /* get temp reg index */
235      tempReg = UREG(REG_TYPE_R, temp);  /* make i915 register */
236
237      i915_emit_arith( p, A0_MOV,
238                       tempReg, A0_DEST_CHANNEL_ALL, /* dest reg, writemask */
239                       0,                            /* saturate */
240                       coord, 0, 0 );                /* src0, src1, src2 */
241
242      /* new src texcoord is tempReg */
243      coord = tempReg;
244   }
245
246   /* Don't worry about saturate as we only support
247    */
248   if (destmask != A0_DEST_CHANNEL_ALL) {
249      /* if not writing to XYZW... */
250      uint tmp = i915_get_utemp(p);
251      i915_emit_texld( p, tmp, A0_DEST_CHANNEL_ALL, sampler, coord, opcode );
252      i915_emit_arith( p, A0_MOV, dest, destmask, 0, tmp, 0, 0 );
253      /* XXX release utemp here? */
254   }
255   else {
256      assert(GET_UREG_TYPE(dest) != REG_TYPE_CONST);
257      assert(dest == UREG(GET_UREG_TYPE(dest), GET_UREG_NR(dest)));
258
259      /* Output register being oC or oD defines a phase boundary */
260      if (GET_UREG_TYPE(dest) == REG_TYPE_OC ||
261          GET_UREG_TYPE(dest) == REG_TYPE_OD)
262         p->nr_tex_indirect++;
263
264      /* Reading from an r# register whose contents depend on output of the
265       * current phase defines a phase boundary.
266       */
267      if (GET_UREG_TYPE(coord) == REG_TYPE_R &&
268          p->register_phases[GET_UREG_NR(coord)] == p->nr_tex_indirect)
269         p->nr_tex_indirect++;
270
271      if (p->csr< p->program + I915_PROGRAM_SIZE) {
272         *(p->csr++) = (opcode |
273                        T0_DEST( dest ) |
274                        T0_SAMPLER( sampler ));
275
276         *(p->csr++) = T1_ADDRESS_REG( coord );
277         *(p->csr++) = T2_MBZ;
278      }
279      else
280         i915_program_error(p, "Out of instructions");
281
282      if (GET_UREG_TYPE(dest) == REG_TYPE_R)
283         p->register_phases[GET_UREG_NR(dest)] = p->nr_tex_indirect;
284
285      p->nr_tex_insn++;
286   }
287
288   if (temp >= 0)
289      i915_release_temp(p, temp);
290
291   return dest;
292}
293
294
295uint
296i915_emit_const1f(struct i915_fp_compile * p, float c0)
297{
298   struct i915_fragment_shader *ifs = p->shader;
299   unsigned reg, idx;
300
301   if (c0 == 0.0)
302      return swizzle(UREG(REG_TYPE_R, 0), ZERO, ZERO, ZERO, ZERO);
303   if (c0 == 1.0)
304      return swizzle(UREG(REG_TYPE_R, 0), ONE, ONE, ONE, ONE);
305
306   for (reg = 0; reg < I915_MAX_CONSTANT; reg++) {
307      if (ifs->constant_flags[reg] == I915_CONSTFLAG_USER)
308         continue;
309      for (idx = 0; idx < 4; idx++) {
310         if (!(ifs->constant_flags[reg] & (1 << idx)) ||
311             ifs->constants[reg][idx] == c0) {
312            ifs->constants[reg][idx] = c0;
313            ifs->constant_flags[reg] |= 1 << idx;
314            if (reg + 1 > ifs->num_constants)
315               ifs->num_constants = reg + 1;
316            return swizzle(UREG(REG_TYPE_CONST, reg), idx, ZERO, ZERO, ONE);
317         }
318      }
319   }
320
321   i915_program_error(p, "i915_emit_const1f: out of constants");
322   return 0;
323}
324
325uint
326i915_emit_const2f(struct i915_fp_compile * p, float c0, float c1)
327{
328   struct i915_fragment_shader *ifs = p->shader;
329   unsigned reg, idx;
330
331   if (c0 == 0.0)
332      return swizzle(i915_emit_const1f(p, c1), ZERO, X, Z, W);
333   if (c0 == 1.0)
334      return swizzle(i915_emit_const1f(p, c1), ONE, X, Z, W);
335
336   if (c1 == 0.0)
337      return swizzle(i915_emit_const1f(p, c0), X, ZERO, Z, W);
338   if (c1 == 1.0)
339      return swizzle(i915_emit_const1f(p, c0), X, ONE, Z, W);
340
341   // XXX emit swizzle here for 0, 1, -1 and any combination thereof
342   // we can use swizzle + neg for that
343   for (reg = 0; reg < I915_MAX_CONSTANT; reg++) {
344      if (ifs->constant_flags[reg] == 0xf ||
345          ifs->constant_flags[reg] == I915_CONSTFLAG_USER)
346         continue;
347      for (idx = 0; idx < 3; idx++) {
348         if (!(ifs->constant_flags[reg] & (3 << idx))) {
349            ifs->constants[reg][idx + 0] = c0;
350            ifs->constants[reg][idx + 1] = c1;
351            ifs->constant_flags[reg] |= 3 << idx;
352            if (reg + 1 > ifs->num_constants)
353               ifs->num_constants = reg + 1;
354            return swizzle(UREG(REG_TYPE_CONST, reg), idx, idx + 1, ZERO, ONE);
355         }
356      }
357   }
358
359   i915_program_error(p, "i915_emit_const2f: out of constants");
360   return 0;
361}
362
363uint
364i915_emit_const4f(struct i915_fp_compile * p,
365                  float c0, float c1, float c2, float c3)
366{
367   struct i915_fragment_shader *ifs = p->shader;
368   unsigned reg;
369
370   // XXX emit swizzle here for 0, 1, -1 and any combination thereof
371   // we can use swizzle + neg for that
372   for (reg = 0; reg < I915_MAX_CONSTANT; reg++) {
373      if (ifs->constant_flags[reg] == 0xf &&
374          ifs->constants[reg][0] == c0 &&
375          ifs->constants[reg][1] == c1 &&
376          ifs->constants[reg][2] == c2 &&
377          ifs->constants[reg][3] == c3) {
378         return UREG(REG_TYPE_CONST, reg);
379      }
380      else if (ifs->constant_flags[reg] == 0) {
381
382         ifs->constants[reg][0] = c0;
383         ifs->constants[reg][1] = c1;
384         ifs->constants[reg][2] = c2;
385         ifs->constants[reg][3] = c3;
386         ifs->constant_flags[reg] = 0xf;
387         if (reg + 1 > ifs->num_constants)
388            ifs->num_constants = reg + 1;
389         return UREG(REG_TYPE_CONST, reg);
390      }
391   }
392
393   i915_program_error(p, "i915_emit_const4f: out of constants");
394   return 0;
395}
396
397
398uint
399i915_emit_const4fv(struct i915_fp_compile * p, const float * c)
400{
401   return i915_emit_const4f(p, c[0], c[1], c[2], c[3]);
402}
403