1 /*
2 * Copyright 2010 Corbin Simpson
3 * Copyright 2010 Marek Olšák <maraeo@gmail.com>
4 * SPDX-License-Identifier: MIT
5 */
6
7 #include "radeon_program_tex.h"
8
9 #include "radeon_compiler_util.h"
10
11 /* Series of transformations to be done on textures. */
12
13 static void
scale_texcoords(struct r300_fragment_program_compiler * compiler,struct rc_instruction * inst,unsigned state_constant)14 scale_texcoords(struct r300_fragment_program_compiler *compiler, struct rc_instruction *inst,
15 unsigned state_constant)
16 {
17 struct rc_instruction *inst_mov;
18
19 unsigned temp = rc_find_free_temporary(&compiler->Base);
20
21 inst_mov = rc_insert_new_instruction(&compiler->Base, inst->Prev);
22
23 inst_mov->U.I.Opcode = RC_OPCODE_MUL;
24 inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
25 inst_mov->U.I.DstReg.Index = temp;
26 inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
27 inst_mov->U.I.SrcReg[1].File = RC_FILE_CONSTANT;
28 inst_mov->U.I.SrcReg[1].Index = rc_constants_add_state(&compiler->Base.Program.Constants,
29 state_constant, inst->U.I.TexSrcUnit);
30
31 reset_srcreg(&inst->U.I.SrcReg[0]);
32 inst->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
33 inst->U.I.SrcReg[0].Index = temp;
34 }
35
36 static void
projective_divide(struct r300_fragment_program_compiler * compiler,struct rc_instruction * inst)37 projective_divide(struct r300_fragment_program_compiler *compiler, struct rc_instruction *inst)
38 {
39 struct rc_instruction *inst_mul, *inst_rcp;
40
41 /* Make sure there is no temp reusing, some later passes depend on the SSA-like form. */
42 unsigned temp_rcp = rc_find_free_temporary(&compiler->Base);
43 unsigned temp_mul = rc_find_free_temporary(&compiler->Base);
44
45 inst_rcp = rc_insert_new_instruction(&compiler->Base, inst->Prev);
46 inst_rcp->U.I.Opcode = RC_OPCODE_RCP;
47 inst_rcp->U.I.DstReg.File = RC_FILE_TEMPORARY;
48 inst_rcp->U.I.DstReg.Index = temp_rcp;
49 inst_rcp->U.I.DstReg.WriteMask = RC_MASK_W;
50 inst_rcp->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
51 /* Because the input can be arbitrarily swizzled,
52 * read the component mapped to W. */
53 inst_rcp->U.I.SrcReg[0].Swizzle = RC_MAKE_SWIZZLE_SMEAR(GET_SWZ(inst->U.I.SrcReg[0].Swizzle, 3));
54
55 inst_mul = rc_insert_new_instruction(&compiler->Base, inst->Prev);
56 inst_mul->U.I.Opcode = RC_OPCODE_MUL;
57 inst_mul->U.I.DstReg.File = RC_FILE_TEMPORARY;
58 inst_mul->U.I.DstReg.Index = temp_mul;
59 inst_mul->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
60 inst_mul->U.I.SrcReg[1].File = RC_FILE_TEMPORARY;
61 inst_mul->U.I.SrcReg[1].Index = temp_rcp;
62 inst_mul->U.I.SrcReg[1].Swizzle = RC_SWIZZLE_WWWW;
63
64 reset_srcreg(&inst->U.I.SrcReg[0]);
65 inst->U.I.Opcode = RC_OPCODE_TEX;
66 inst->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
67 inst->U.I.SrcReg[0].Index = temp_mul;
68 }
69
70 /**
71 * Transform TEX, TXP, TXB, and KIL instructions in the following ways:
72 * - implement texture compare (shadow extensions)
73 * - extract non-native source / destination operands
74 * - premultiply texture coordinates for RECT
75 * - extract operand swizzles
76 * - introduce a temporary register when write masks are needed
77 */
78 int
radeonTransformTEX(struct radeon_compiler * c,struct rc_instruction * inst,void * data)79 radeonTransformTEX(struct radeon_compiler *c, struct rc_instruction *inst, void *data)
80 {
81 struct r300_fragment_program_compiler *compiler = (struct r300_fragment_program_compiler *)data;
82 rc_wrap_mode wrapmode = compiler->state.unit[inst->U.I.TexSrcUnit].wrap_mode;
83 int is_rect = inst->U.I.TexSrcTarget == RC_TEXTURE_RECT;
84
85 if (inst->U.I.Opcode != RC_OPCODE_TEX && inst->U.I.Opcode != RC_OPCODE_TXB &&
86 inst->U.I.Opcode != RC_OPCODE_TXP && inst->U.I.Opcode != RC_OPCODE_TXD &&
87 inst->U.I.Opcode != RC_OPCODE_TXL && inst->U.I.Opcode != RC_OPCODE_KIL)
88 return 0;
89
90 /* R300 cannot sample from rectangles and the wrap mode fallback needs
91 * normalized coordinates anyway. */
92 if (inst->U.I.Opcode != RC_OPCODE_KIL && is_rect && (!c->is_r500 || wrapmode != RC_WRAP_NONE)) {
93 scale_texcoords(compiler, inst, RC_STATE_R300_TEXRECT_FACTOR);
94 inst->U.I.TexSrcTarget = RC_TEXTURE_2D;
95 }
96
97 /* Divide by W if needed. */
98 if (inst->U.I.Opcode == RC_OPCODE_TXP &&
99 (wrapmode == RC_WRAP_REPEAT || wrapmode == RC_WRAP_MIRRORED_REPEAT ||
100 compiler->state.unit[inst->U.I.TexSrcUnit].clamp_and_scale_before_fetch)) {
101 projective_divide(compiler, inst);
102 }
103
104 /* Texture wrap modes don't work on NPOT textures.
105 *
106 * Non-wrapped/clamped texcoords with NPOT are free in HW. Repeat and
107 * mirroring are not. If we need to repeat, we do:
108 *
109 * MUL temp, texcoord, <scaling factor constant>
110 * FRC temp, temp ; Discard integer portion of coords
111 *
112 * This gives us coords in [0, 1].
113 *
114 * Mirroring is trickier. We're going to start out like repeat:
115 *
116 * MUL temp, texcoord, <scaling factor constant> ; De-mirror across axes
117 * MUL temp, temp, 0.5 ; Pattern repeats in [0, 2]
118 * ; so scale to [0, 1]
119 * FRC temp, temp ; Make the pattern repeat
120 * MAD temp, temp, 2, -1 ; Move the pattern to [-1, 1]
121 * ADD temp, 1, -abs(temp) ; Now comes a neat trick: use abs to mirror the pattern.
122 * ; The pattern is backwards, so reverse it (1-x).
123 *
124 * This gives us coords in [0, 1].
125 *
126 * ~ C & M. ;)
127 */
128 if (inst->U.I.Opcode != RC_OPCODE_KIL && wrapmode != RC_WRAP_NONE) {
129 struct rc_instruction *inst_mov;
130 unsigned temp = rc_find_free_temporary(c);
131
132 if (wrapmode == RC_WRAP_REPEAT) {
133 /* Both instructions will be paired up. */
134 struct rc_instruction *inst_frc = rc_insert_new_instruction(c, inst->Prev);
135
136 inst_frc->U.I.Opcode = RC_OPCODE_FRC;
137 inst_frc->U.I.DstReg.File = RC_FILE_TEMPORARY;
138 inst_frc->U.I.DstReg.Index = temp;
139 inst_frc->U.I.DstReg.WriteMask = RC_MASK_XYZ;
140 inst_frc->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
141 } else if (wrapmode == RC_WRAP_MIRRORED_REPEAT) {
142 /*
143 * Function:
144 * f(v) = 1 - abs(frac(v * 0.5) * 2 - 1)
145 *
146 * Code:
147 * MUL temp, src0, 0.5
148 * FRC temp, temp
149 * MAD temp, temp, 2, -1
150 * ADD temp, 1, -abs(temp)
151 */
152
153 struct rc_instruction *inst_mul, *inst_frc, *inst_mad, *inst_add;
154 unsigned two, two_swizzle;
155
156 inst_mul = rc_insert_new_instruction(c, inst->Prev);
157 unsigned temp_mul = rc_find_free_temporary(c);
158
159 inst_mul->U.I.Opcode = RC_OPCODE_MUL;
160 inst_mul->U.I.DstReg.File = RC_FILE_TEMPORARY;
161 inst_mul->U.I.DstReg.Index = temp_mul;
162 inst_mul->U.I.DstReg.WriteMask = RC_MASK_XYZ;
163 inst_mul->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
164 inst_mul->U.I.SrcReg[1].Swizzle = RC_SWIZZLE_HHHH;
165
166 inst_frc = rc_insert_new_instruction(c, inst->Prev);
167 unsigned temp_frc = rc_find_free_temporary(c);
168
169 inst_frc->U.I.Opcode = RC_OPCODE_FRC;
170 inst_frc->U.I.DstReg.File = RC_FILE_TEMPORARY;
171 inst_frc->U.I.DstReg.Index = temp_frc;
172 inst_frc->U.I.DstReg.WriteMask = RC_MASK_XYZ;
173 inst_frc->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
174 inst_frc->U.I.SrcReg[0].Index = temp_mul;
175 inst_frc->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_XYZ0;
176
177 two = rc_constants_add_immediate_scalar(&c->Program.Constants, 2, &two_swizzle);
178 inst_mad = rc_insert_new_instruction(c, inst->Prev);
179 unsigned temp_mad = rc_find_free_temporary(c);
180
181 inst_mad->U.I.Opcode = RC_OPCODE_MAD;
182 inst_mad->U.I.DstReg.File = RC_FILE_TEMPORARY;
183 inst_mad->U.I.DstReg.Index = temp_mad;
184 inst_mad->U.I.DstReg.WriteMask = RC_MASK_XYZ;
185 inst_mad->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
186 inst_mad->U.I.SrcReg[0].Index = temp_frc;
187 inst_mad->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_XYZ0;
188 inst_mad->U.I.SrcReg[1].File = RC_FILE_CONSTANT;
189 inst_mad->U.I.SrcReg[1].Index = two;
190 inst_mad->U.I.SrcReg[1].Swizzle = two_swizzle;
191 inst_mad->U.I.SrcReg[2].Swizzle = RC_SWIZZLE_1111;
192 inst_mad->U.I.SrcReg[2].Negate = RC_MASK_XYZ;
193
194 inst_add = rc_insert_new_instruction(c, inst->Prev);
195
196 inst_add->U.I.Opcode = RC_OPCODE_ADD;
197 inst_add->U.I.DstReg.File = RC_FILE_TEMPORARY;
198 inst_add->U.I.DstReg.Index = temp;
199 inst_add->U.I.DstReg.WriteMask = RC_MASK_XYZ;
200 inst_add->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_1111;
201 inst_add->U.I.SrcReg[1].File = RC_FILE_TEMPORARY;
202 inst_add->U.I.SrcReg[1].Index = temp_mad;
203 inst_add->U.I.SrcReg[1].Swizzle = RC_SWIZZLE_XYZ0;
204 inst_add->U.I.SrcReg[1].Abs = 1;
205 inst_add->U.I.SrcReg[1].Negate = RC_MASK_XYZ;
206 } else if (wrapmode == RC_WRAP_MIRRORED_CLAMP) {
207 /*
208 * Mirrored clamp modes are bloody simple, we just use abs
209 * to mirror [0, 1] into [-1, 0]. This works for
210 * all modes i.e. CLAMP, CLAMP_TO_EDGE, and CLAMP_TO_BORDER.
211 */
212 struct rc_instruction *inst_mov;
213
214 inst_mov = rc_insert_new_instruction(c, inst->Prev);
215
216 inst_mov->U.I.Opcode = RC_OPCODE_MOV;
217 inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
218 inst_mov->U.I.DstReg.Index = temp;
219 inst_mov->U.I.DstReg.WriteMask = RC_MASK_XYZ;
220 inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
221 inst_mov->U.I.SrcReg[0].Abs = 1;
222 }
223
224 /* Preserve W for TXP/TXB. */
225 inst_mov = rc_insert_new_instruction(c, inst->Prev);
226
227 inst_mov->U.I.Opcode = RC_OPCODE_MOV;
228 inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
229 inst_mov->U.I.DstReg.Index = temp;
230 inst_mov->U.I.DstReg.WriteMask = RC_MASK_W;
231 inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
232
233 reset_srcreg(&inst->U.I.SrcReg[0]);
234 inst->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
235 inst->U.I.SrcReg[0].Index = temp;
236 }
237
238 /* NPOT -> POT conversion for 3D textures. */
239 if (inst->U.I.Opcode != RC_OPCODE_KIL &&
240 compiler->state.unit[inst->U.I.TexSrcUnit].clamp_and_scale_before_fetch) {
241 struct rc_instruction *inst_mov;
242 unsigned temp = rc_find_free_temporary(c);
243
244 /* Saturate XYZ. */
245 inst_mov = rc_insert_new_instruction(c, inst->Prev);
246 inst_mov->U.I.Opcode = RC_OPCODE_MOV;
247 inst_mov->U.I.SaturateMode = RC_SATURATE_ZERO_ONE;
248 inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
249 inst_mov->U.I.DstReg.Index = temp;
250 inst_mov->U.I.DstReg.WriteMask = RC_MASK_XYZ;
251 inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
252
253 /* Copy W. */
254 inst_mov = rc_insert_new_instruction(c, inst->Prev);
255 inst_mov->U.I.Opcode = RC_OPCODE_MOV;
256 inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
257 inst_mov->U.I.DstReg.Index = temp;
258 inst_mov->U.I.DstReg.WriteMask = RC_MASK_W;
259 inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
260
261 reset_srcreg(&inst->U.I.SrcReg[0]);
262 inst->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
263 inst->U.I.SrcReg[0].Index = temp;
264
265 scale_texcoords(compiler, inst, RC_STATE_R300_TEXSCALE_FACTOR);
266 }
267
268 /* Cannot write texture to output registers or with saturate (all chips),
269 * or with masks (non-r500). */
270 if (inst->U.I.Opcode != RC_OPCODE_KIL) {
271 /* We should not be getting saturates on TEX, but assert just to be sure. */
272 assert(!inst->U.I.SaturateMode);
273
274 if (inst->U.I.DstReg.File != RC_FILE_TEMPORARY || inst->U.I.SaturateMode ||
275 (!c->is_r500 && inst->U.I.DstReg.WriteMask != RC_MASK_XYZW)) {
276 struct rc_instruction *inst_mov = rc_insert_new_instruction(c, inst);
277
278 inst_mov->U.I.Opcode = RC_OPCODE_MOV;
279 inst_mov->U.I.SaturateMode = inst->U.I.SaturateMode;
280 inst_mov->U.I.DstReg = inst->U.I.DstReg;
281 inst_mov->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
282 inst_mov->U.I.SrcReg[0].Index = rc_find_free_temporary(c);
283
284 inst->U.I.DstReg.File = RC_FILE_TEMPORARY;
285 inst->U.I.DstReg.Index = inst_mov->U.I.SrcReg[0].Index;
286 inst->U.I.DstReg.WriteMask = RC_MASK_XYZW;
287 }
288 }
289
290 return 1;
291 }
292