1 /**************************************************************************
2 *
3 * Copyright 2007 VMware, Inc.
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 VMWARE 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 /*
29 * Authors:
30 * Keith Whitwell <keithw@vmware.com>
31 */
32
33 #include "util/u_memory.h"
34 #include "util/u_format.h"
35 #include "util/u_half.h"
36 #include "util/u_math.h"
37 #include "pipe/p_state.h"
38 #include "translate.h"
39
40
41 #define DRAW_DBG 0
42
43 typedef void (*fetch_func)(void *dst,
44 const uint8_t *src,
45 unsigned i, unsigned j);
46 typedef void (*emit_func)(const void *attrib, void *ptr);
47
48
49
50 struct translate_generic {
51 struct translate translate;
52
53 struct {
54 enum translate_element_type type;
55
56 fetch_func fetch;
57 unsigned buffer;
58 unsigned input_offset;
59 unsigned instance_divisor;
60
61 emit_func emit;
62 unsigned output_offset;
63
64 const uint8_t *input_ptr;
65 unsigned input_stride;
66 unsigned max_index;
67
68 /* this value is set to -1 if this is a normal element with
69 * output_format != input_format: in this case, u_format is used
70 * to do a full conversion
71 *
72 * this value is set to the format size in bytes if
73 * output_format == input_format or for 32-bit instance ids:
74 * in this case, memcpy is used to copy this amount of bytes
75 */
76 int copy_size;
77
78 } attrib[TRANSLATE_MAX_ATTRIBS];
79
80 unsigned nr_attrib;
81 };
82
83
84 static struct translate_generic *
translate_generic(struct translate * translate)85 translate_generic(struct translate *translate)
86 {
87 return (struct translate_generic *)translate;
88 }
89
90
91 /**
92 * Fetch a dword[4] vertex attribute from memory, doing format/type
93 * conversion as needed.
94 *
95 * This is probably needed/dupliocated elsewhere, eg format
96 * conversion, texture sampling etc.
97 */
98 #define ATTRIB(NAME, SZ, SRCTYPE, DSTTYPE, TO) \
99 static void \
100 emit_##NAME(const void *attrib, void *ptr) \
101 { \
102 unsigned i; \
103 SRCTYPE *in = (SRCTYPE *)attrib; \
104 DSTTYPE *out = (DSTTYPE *)ptr; \
105 \
106 for (i = 0; i < SZ; i++) { \
107 out[i] = TO(in[i]); \
108 } \
109 }
110
111
112 #define TO_64_FLOAT(x) ((double) x)
113 #define TO_32_FLOAT(x) (x)
114 #define TO_16_FLOAT(x) util_float_to_half(x)
115
116 #define TO_8_USCALED(x) ((unsigned char) x)
117 #define TO_16_USCALED(x) ((unsigned short) x)
118 #define TO_32_USCALED(x) ((unsigned int) x)
119
120 #define TO_8_SSCALED(x) ((char) x)
121 #define TO_16_SSCALED(x) ((short) x)
122 #define TO_32_SSCALED(x) ((int) x)
123
124 #define TO_8_UNORM(x) ((unsigned char) (x * 255.0f))
125 #define TO_16_UNORM(x) ((unsigned short) (x * 65535.0f))
126 #define TO_32_UNORM(x) ((unsigned int) (x * 4294967295.0f))
127
128 #define TO_8_SNORM(x) ((char) (x * 127.0f))
129 #define TO_16_SNORM(x) ((short) (x * 32767.0f))
130 #define TO_32_SNORM(x) ((int) (x * 2147483647.0f))
131
132 #define TO_32_FIXED(x) ((int) (x * 65536.0f))
133
134 #define TO_INT(x) (x)
135
136
137 ATTRIB(R64G64B64A64_FLOAT, 4, float, double, TO_64_FLOAT)
138 ATTRIB(R64G64B64_FLOAT, 3, float, double, TO_64_FLOAT)
139 ATTRIB(R64G64_FLOAT, 2, float, double, TO_64_FLOAT)
140 ATTRIB(R64_FLOAT, 1, float, double, TO_64_FLOAT)
141
142 ATTRIB(R32G32B32A32_FLOAT, 4, float, float, TO_32_FLOAT)
143 ATTRIB(R32G32B32_FLOAT, 3, float, float, TO_32_FLOAT)
144 ATTRIB(R32G32_FLOAT, 2, float, float, TO_32_FLOAT)
145 ATTRIB(R32_FLOAT, 1, float, float, TO_32_FLOAT)
146
147 ATTRIB(R16G16B16A16_FLOAT, 4, float, ushort, TO_16_FLOAT)
148 ATTRIB(R16G16B16_FLOAT, 3, float, ushort, TO_16_FLOAT)
149 ATTRIB(R16G16_FLOAT, 2, float, ushort, TO_16_FLOAT)
150 ATTRIB(R16_FLOAT, 1, float, ushort, TO_16_FLOAT)
151
152 ATTRIB(R32G32B32A32_USCALED, 4, float, unsigned, TO_32_USCALED)
153 ATTRIB(R32G32B32_USCALED, 3, float, unsigned, TO_32_USCALED)
154 ATTRIB(R32G32_USCALED, 2, float, unsigned, TO_32_USCALED)
155 ATTRIB(R32_USCALED, 1, float, unsigned, TO_32_USCALED)
156
157 ATTRIB(R32G32B32A32_SSCALED, 4, float, int, TO_32_SSCALED)
158 ATTRIB(R32G32B32_SSCALED, 3, float, int, TO_32_SSCALED)
159 ATTRIB(R32G32_SSCALED, 2, float, int, TO_32_SSCALED)
160 ATTRIB(R32_SSCALED, 1, float, int, TO_32_SSCALED)
161
162 ATTRIB(R32G32B32A32_UNORM, 4, float, unsigned, TO_32_UNORM)
163 ATTRIB(R32G32B32_UNORM, 3, float, unsigned, TO_32_UNORM)
164 ATTRIB(R32G32_UNORM, 2, float, unsigned, TO_32_UNORM)
165 ATTRIB(R32_UNORM, 1, float, unsigned, TO_32_UNORM)
166
167 ATTRIB(R32G32B32A32_SNORM, 4, float, int, TO_32_SNORM)
168 ATTRIB(R32G32B32_SNORM, 3, float, int, TO_32_SNORM)
169 ATTRIB(R32G32_SNORM, 2, float, int, TO_32_SNORM)
170 ATTRIB(R32_SNORM, 1, float, int, TO_32_SNORM)
171
172 ATTRIB(R16G16B16A16_USCALED, 4, float, ushort, TO_16_USCALED)
173 ATTRIB(R16G16B16_USCALED, 3, float, ushort, TO_16_USCALED)
174 ATTRIB(R16G16_USCALED, 2, float, ushort, TO_16_USCALED)
175 ATTRIB(R16_USCALED, 1, float, ushort, TO_16_USCALED)
176
177 ATTRIB(R16G16B16A16_SSCALED, 4, float, short, TO_16_SSCALED)
178 ATTRIB(R16G16B16_SSCALED, 3, float, short, TO_16_SSCALED)
179 ATTRIB(R16G16_SSCALED, 2, float, short, TO_16_SSCALED)
180 ATTRIB(R16_SSCALED, 1, float, short, TO_16_SSCALED)
181
182 ATTRIB(R16G16B16A16_UNORM, 4, float, ushort, TO_16_UNORM)
183 ATTRIB(R16G16B16_UNORM, 3, float, ushort, TO_16_UNORM)
184 ATTRIB(R16G16_UNORM, 2, float, ushort, TO_16_UNORM)
185 ATTRIB(R16_UNORM, 1, float, ushort, TO_16_UNORM)
186
187 ATTRIB(R16G16B16A16_SNORM, 4, float, short, TO_16_SNORM)
188 ATTRIB(R16G16B16_SNORM, 3, float, short, TO_16_SNORM)
189 ATTRIB(R16G16_SNORM, 2, float, short, TO_16_SNORM)
190 ATTRIB(R16_SNORM, 1, float, short, TO_16_SNORM)
191
192 ATTRIB(R8G8B8A8_USCALED, 4, float, ubyte, TO_8_USCALED)
193 ATTRIB(R8G8B8_USCALED, 3, float, ubyte, TO_8_USCALED)
194 ATTRIB(R8G8_USCALED, 2, float, ubyte, TO_8_USCALED)
195 ATTRIB(R8_USCALED, 1, float, ubyte, TO_8_USCALED)
196
197 ATTRIB(R8G8B8A8_SSCALED, 4, float, char, TO_8_SSCALED)
198 ATTRIB(R8G8B8_SSCALED, 3, float, char, TO_8_SSCALED)
199 ATTRIB(R8G8_SSCALED, 2, float, char, TO_8_SSCALED)
200 ATTRIB(R8_SSCALED, 1, float, char, TO_8_SSCALED)
201
202 ATTRIB(R8G8B8A8_UNORM, 4, float, ubyte, TO_8_UNORM)
203 ATTRIB(R8G8B8_UNORM, 3, float, ubyte, TO_8_UNORM)
204 ATTRIB(R8G8_UNORM, 2, float, ubyte, TO_8_UNORM)
205 ATTRIB(R8_UNORM, 1, float, ubyte, TO_8_UNORM)
206
207 ATTRIB(R8G8B8A8_SNORM, 4, float, char, TO_8_SNORM)
208 ATTRIB(R8G8B8_SNORM, 3, float, char, TO_8_SNORM)
209 ATTRIB(R8G8_SNORM, 2, float, char, TO_8_SNORM)
210 ATTRIB(R8_SNORM, 1, float, char, TO_8_SNORM)
211
212 ATTRIB(R32G32B32A32_UINT, 4, uint32_t, unsigned, TO_INT)
213 ATTRIB(R32G32B32_UINT, 3, uint32_t, unsigned, TO_INT)
214 ATTRIB(R32G32_UINT, 2, uint32_t, unsigned, TO_INT)
215 ATTRIB(R32_UINT, 1, uint32_t, unsigned, TO_INT)
216
217 ATTRIB(R16G16B16A16_UINT, 4, uint32_t, ushort, TO_INT)
218 ATTRIB(R16G16B16_UINT, 3, uint32_t, ushort, TO_INT)
219 ATTRIB(R16G16_UINT, 2, uint32_t, ushort, TO_INT)
220 ATTRIB(R16_UINT, 1, uint32_t, ushort, TO_INT)
221
222 ATTRIB(R8G8B8A8_UINT, 4, uint32_t, ubyte, TO_INT)
223 ATTRIB(R8G8B8_UINT, 3, uint32_t, ubyte, TO_INT)
224 ATTRIB(R8G8_UINT, 2, uint32_t, ubyte, TO_INT)
225 ATTRIB(R8_UINT, 1, uint32_t, ubyte, TO_INT)
226
227 ATTRIB(R32G32B32A32_SINT, 4, int32_t, int, TO_INT)
228 ATTRIB(R32G32B32_SINT, 3, int32_t, int, TO_INT)
229 ATTRIB(R32G32_SINT, 2, int32_t, int, TO_INT)
230 ATTRIB(R32_SINT, 1, int32_t, int, TO_INT)
231
232 ATTRIB(R16G16B16A16_SINT, 4, int32_t, short, TO_INT)
233 ATTRIB(R16G16B16_SINT, 3, int32_t, short, TO_INT)
234 ATTRIB(R16G16_SINT, 2, int32_t, short, TO_INT)
235 ATTRIB(R16_SINT, 1, int32_t, short, TO_INT)
236
237 ATTRIB(R8G8B8A8_SINT, 4, int32_t, char, TO_INT)
238 ATTRIB(R8G8B8_SINT, 3, int32_t, char, TO_INT)
239 ATTRIB(R8G8_SINT, 2, int32_t, char, TO_INT)
240 ATTRIB(R8_SINT, 1, int32_t, char, TO_INT)
241
242 static void
emit_A8R8G8B8_UNORM(const void * attrib,void * ptr)243 emit_A8R8G8B8_UNORM(const void *attrib, void *ptr)
244 {
245 float *in = (float *)attrib;
246 ubyte *out = (ubyte *)ptr;
247 out[0] = TO_8_UNORM(in[3]);
248 out[1] = TO_8_UNORM(in[0]);
249 out[2] = TO_8_UNORM(in[1]);
250 out[3] = TO_8_UNORM(in[2]);
251 }
252
253 static void
emit_B8G8R8A8_UNORM(const void * attrib,void * ptr)254 emit_B8G8R8A8_UNORM(const void *attrib, void *ptr)
255 {
256 float *in = (float *)attrib;
257 ubyte *out = (ubyte *)ptr;
258 out[2] = TO_8_UNORM(in[0]);
259 out[1] = TO_8_UNORM(in[1]);
260 out[0] = TO_8_UNORM(in[2]);
261 out[3] = TO_8_UNORM(in[3]);
262 }
263
264 static void
emit_B10G10R10A2_UNORM(const void * attrib,void * ptr)265 emit_B10G10R10A2_UNORM(const void *attrib, void *ptr)
266 {
267 float *src = (float *)ptr;
268 uint32_t value = 0;
269 value |= ((uint32_t)(CLAMP(src[2], 0, 1) * 0x3ff)) & 0x3ff;
270 value |= (((uint32_t)(CLAMP(src[1], 0, 1) * 0x3ff)) & 0x3ff) << 10;
271 value |= (((uint32_t)(CLAMP(src[0], 0, 1) * 0x3ff)) & 0x3ff) << 20;
272 value |= ((uint32_t)(CLAMP(src[3], 0, 1) * 0x3)) << 30;
273 *(uint32_t *)attrib = util_le32_to_cpu(value);
274 }
275
276 static void
emit_B10G10R10A2_USCALED(const void * attrib,void * ptr)277 emit_B10G10R10A2_USCALED(const void *attrib, void *ptr)
278 {
279 float *src = (float *)ptr;
280 uint32_t value = 0;
281 value |= ((uint32_t)CLAMP(src[2], 0, 1023)) & 0x3ff;
282 value |= (((uint32_t)CLAMP(src[1], 0, 1023)) & 0x3ff) << 10;
283 value |= (((uint32_t)CLAMP(src[0], 0, 1023)) & 0x3ff) << 20;
284 value |= ((uint32_t)CLAMP(src[3], 0, 3)) << 30;
285 *(uint32_t *)attrib = util_le32_to_cpu(value);
286 }
287
288 static void
emit_B10G10R10A2_SNORM(const void * attrib,void * ptr)289 emit_B10G10R10A2_SNORM(const void *attrib, void *ptr)
290 {
291 float *src = (float *)ptr;
292 uint32_t value = 0;
293 value |= (uint32_t)(((uint32_t)(CLAMP(src[2], -1, 1) * 0x1ff)) & 0x3ff) ;
294 value |= (uint32_t)((((uint32_t)(CLAMP(src[1], -1, 1) * 0x1ff)) & 0x3ff) << 10) ;
295 value |= (uint32_t)((((uint32_t)(CLAMP(src[0], -1, 1) * 0x1ff)) & 0x3ff) << 20) ;
296 value |= (uint32_t)(((uint32_t)(CLAMP(src[3], -1, 1) * 0x1)) << 30) ;
297 *(uint32_t *)attrib = util_le32_to_cpu(value);
298 }
299
300 static void
emit_B10G10R10A2_SSCALED(const void * attrib,void * ptr)301 emit_B10G10R10A2_SSCALED(const void *attrib, void *ptr)
302 {
303 float *src = (float *)ptr;
304 uint32_t value = 0;
305 value |= (uint32_t)(((uint32_t)CLAMP(src[2], -512, 511)) & 0x3ff) ;
306 value |= (uint32_t)((((uint32_t)CLAMP(src[1], -512, 511)) & 0x3ff) << 10) ;
307 value |= (uint32_t)((((uint32_t)CLAMP(src[0], -512, 511)) & 0x3ff) << 20) ;
308 value |= (uint32_t)(((uint32_t)CLAMP(src[3], -2, 1)) << 30) ;
309 *(uint32_t *)attrib = util_le32_to_cpu(value);
310 }
311
312 static void
emit_R10G10B10A2_UNORM(const void * attrib,void * ptr)313 emit_R10G10B10A2_UNORM(const void *attrib, void *ptr)
314 {
315 float *src = (float *)ptr;
316 uint32_t value = 0;
317 value |= ((uint32_t)(CLAMP(src[0], 0, 1) * 0x3ff)) & 0x3ff;
318 value |= (((uint32_t)(CLAMP(src[1], 0, 1) * 0x3ff)) & 0x3ff) << 10;
319 value |= (((uint32_t)(CLAMP(src[2], 0, 1) * 0x3ff)) & 0x3ff) << 20;
320 value |= ((uint32_t)(CLAMP(src[3], 0, 1) * 0x3)) << 30;
321 *(uint32_t *)attrib = util_le32_to_cpu(value);
322 }
323
324 static void
emit_R10G10B10A2_USCALED(const void * attrib,void * ptr)325 emit_R10G10B10A2_USCALED(const void *attrib, void *ptr)
326 {
327 float *src = (float *)ptr;
328 uint32_t value = 0;
329 value |= ((uint32_t)CLAMP(src[0], 0, 1023)) & 0x3ff;
330 value |= (((uint32_t)CLAMP(src[1], 0, 1023)) & 0x3ff) << 10;
331 value |= (((uint32_t)CLAMP(src[2], 0, 1023)) & 0x3ff) << 20;
332 value |= ((uint32_t)CLAMP(src[3], 0, 3)) << 30;
333 *(uint32_t *)attrib = util_le32_to_cpu(value);
334 }
335
336 static void
emit_R10G10B10A2_SNORM(const void * attrib,void * ptr)337 emit_R10G10B10A2_SNORM(const void *attrib, void *ptr)
338 {
339 float *src = (float *)ptr;
340 uint32_t value = 0;
341 value |= (uint32_t)(((uint32_t)(CLAMP(src[0], -1, 1) * 0x1ff)) & 0x3ff) ;
342 value |= (uint32_t)((((uint32_t)(CLAMP(src[1], -1, 1) * 0x1ff)) & 0x3ff) << 10) ;
343 value |= (uint32_t)((((uint32_t)(CLAMP(src[2], -1, 1) * 0x1ff)) & 0x3ff) << 20) ;
344 value |= (uint32_t)(((uint32_t)(CLAMP(src[3], -1, 1) * 0x1)) << 30) ;
345 *(uint32_t *)attrib = util_le32_to_cpu(value);
346 }
347
348 static void
emit_R10G10B10A2_SSCALED(const void * attrib,void * ptr)349 emit_R10G10B10A2_SSCALED(const void *attrib, void *ptr)
350 {
351 float *src = (float *)ptr;
352 uint32_t value = 0;
353 value |= (uint32_t)(((uint32_t)CLAMP(src[0], -512, 511)) & 0x3ff) ;
354 value |= (uint32_t)((((uint32_t)CLAMP(src[1], -512, 511)) & 0x3ff) << 10) ;
355 value |= (uint32_t)((((uint32_t)CLAMP(src[2], -512, 511)) & 0x3ff) << 20) ;
356 value |= (uint32_t)(((uint32_t)CLAMP(src[3], -2, 1)) << 30) ;
357 *(uint32_t *)attrib = util_le32_to_cpu(value);
358 }
359
360 static void
emit_NULL(const void * attrib,void * ptr)361 emit_NULL(const void *attrib, void *ptr)
362 {
363 /* do nothing is the only sensible option */
364 }
365
366 static emit_func
get_emit_func(enum pipe_format format)367 get_emit_func(enum pipe_format format)
368 {
369 switch (format) {
370 case PIPE_FORMAT_R64_FLOAT:
371 return &emit_R64_FLOAT;
372 case PIPE_FORMAT_R64G64_FLOAT:
373 return &emit_R64G64_FLOAT;
374 case PIPE_FORMAT_R64G64B64_FLOAT:
375 return &emit_R64G64B64_FLOAT;
376 case PIPE_FORMAT_R64G64B64A64_FLOAT:
377 return &emit_R64G64B64A64_FLOAT;
378
379 case PIPE_FORMAT_R32_FLOAT:
380 return &emit_R32_FLOAT;
381 case PIPE_FORMAT_R32G32_FLOAT:
382 return &emit_R32G32_FLOAT;
383 case PIPE_FORMAT_R32G32B32_FLOAT:
384 return &emit_R32G32B32_FLOAT;
385 case PIPE_FORMAT_R32G32B32A32_FLOAT:
386 return &emit_R32G32B32A32_FLOAT;
387
388 case PIPE_FORMAT_R16_FLOAT:
389 return &emit_R16_FLOAT;
390 case PIPE_FORMAT_R16G16_FLOAT:
391 return &emit_R16G16_FLOAT;
392 case PIPE_FORMAT_R16G16B16_FLOAT:
393 return &emit_R16G16B16_FLOAT;
394 case PIPE_FORMAT_R16G16B16A16_FLOAT:
395 return &emit_R16G16B16A16_FLOAT;
396
397 case PIPE_FORMAT_R32_UNORM:
398 return &emit_R32_UNORM;
399 case PIPE_FORMAT_R32G32_UNORM:
400 return &emit_R32G32_UNORM;
401 case PIPE_FORMAT_R32G32B32_UNORM:
402 return &emit_R32G32B32_UNORM;
403 case PIPE_FORMAT_R32G32B32A32_UNORM:
404 return &emit_R32G32B32A32_UNORM;
405
406 case PIPE_FORMAT_R32_USCALED:
407 return &emit_R32_USCALED;
408 case PIPE_FORMAT_R32G32_USCALED:
409 return &emit_R32G32_USCALED;
410 case PIPE_FORMAT_R32G32B32_USCALED:
411 return &emit_R32G32B32_USCALED;
412 case PIPE_FORMAT_R32G32B32A32_USCALED:
413 return &emit_R32G32B32A32_USCALED;
414
415 case PIPE_FORMAT_R32_SNORM:
416 return &emit_R32_SNORM;
417 case PIPE_FORMAT_R32G32_SNORM:
418 return &emit_R32G32_SNORM;
419 case PIPE_FORMAT_R32G32B32_SNORM:
420 return &emit_R32G32B32_SNORM;
421 case PIPE_FORMAT_R32G32B32A32_SNORM:
422 return &emit_R32G32B32A32_SNORM;
423
424 case PIPE_FORMAT_R32_SSCALED:
425 return &emit_R32_SSCALED;
426 case PIPE_FORMAT_R32G32_SSCALED:
427 return &emit_R32G32_SSCALED;
428 case PIPE_FORMAT_R32G32B32_SSCALED:
429 return &emit_R32G32B32_SSCALED;
430 case PIPE_FORMAT_R32G32B32A32_SSCALED:
431 return &emit_R32G32B32A32_SSCALED;
432
433 case PIPE_FORMAT_R16_UNORM:
434 return &emit_R16_UNORM;
435 case PIPE_FORMAT_R16G16_UNORM:
436 return &emit_R16G16_UNORM;
437 case PIPE_FORMAT_R16G16B16_UNORM:
438 return &emit_R16G16B16_UNORM;
439 case PIPE_FORMAT_R16G16B16A16_UNORM:
440 return &emit_R16G16B16A16_UNORM;
441
442 case PIPE_FORMAT_R16_USCALED:
443 return &emit_R16_USCALED;
444 case PIPE_FORMAT_R16G16_USCALED:
445 return &emit_R16G16_USCALED;
446 case PIPE_FORMAT_R16G16B16_USCALED:
447 return &emit_R16G16B16_USCALED;
448 case PIPE_FORMAT_R16G16B16A16_USCALED:
449 return &emit_R16G16B16A16_USCALED;
450
451 case PIPE_FORMAT_R16_SNORM:
452 return &emit_R16_SNORM;
453 case PIPE_FORMAT_R16G16_SNORM:
454 return &emit_R16G16_SNORM;
455 case PIPE_FORMAT_R16G16B16_SNORM:
456 return &emit_R16G16B16_SNORM;
457 case PIPE_FORMAT_R16G16B16A16_SNORM:
458 return &emit_R16G16B16A16_SNORM;
459
460 case PIPE_FORMAT_R16_SSCALED:
461 return &emit_R16_SSCALED;
462 case PIPE_FORMAT_R16G16_SSCALED:
463 return &emit_R16G16_SSCALED;
464 case PIPE_FORMAT_R16G16B16_SSCALED:
465 return &emit_R16G16B16_SSCALED;
466 case PIPE_FORMAT_R16G16B16A16_SSCALED:
467 return &emit_R16G16B16A16_SSCALED;
468
469 case PIPE_FORMAT_R8_UNORM:
470 return &emit_R8_UNORM;
471 case PIPE_FORMAT_R8G8_UNORM:
472 return &emit_R8G8_UNORM;
473 case PIPE_FORMAT_R8G8B8_UNORM:
474 return &emit_R8G8B8_UNORM;
475 case PIPE_FORMAT_R8G8B8A8_UNORM:
476 return &emit_R8G8B8A8_UNORM;
477
478 case PIPE_FORMAT_R8_USCALED:
479 return &emit_R8_USCALED;
480 case PIPE_FORMAT_R8G8_USCALED:
481 return &emit_R8G8_USCALED;
482 case PIPE_FORMAT_R8G8B8_USCALED:
483 return &emit_R8G8B8_USCALED;
484 case PIPE_FORMAT_R8G8B8A8_USCALED:
485 return &emit_R8G8B8A8_USCALED;
486
487 case PIPE_FORMAT_R8_SNORM:
488 return &emit_R8_SNORM;
489 case PIPE_FORMAT_R8G8_SNORM:
490 return &emit_R8G8_SNORM;
491 case PIPE_FORMAT_R8G8B8_SNORM:
492 return &emit_R8G8B8_SNORM;
493 case PIPE_FORMAT_R8G8B8A8_SNORM:
494 return &emit_R8G8B8A8_SNORM;
495
496 case PIPE_FORMAT_R8_SSCALED:
497 return &emit_R8_SSCALED;
498 case PIPE_FORMAT_R8G8_SSCALED:
499 return &emit_R8G8_SSCALED;
500 case PIPE_FORMAT_R8G8B8_SSCALED:
501 return &emit_R8G8B8_SSCALED;
502 case PIPE_FORMAT_R8G8B8A8_SSCALED:
503 return &emit_R8G8B8A8_SSCALED;
504
505 case PIPE_FORMAT_B8G8R8A8_UNORM:
506 return &emit_B8G8R8A8_UNORM;
507
508 case PIPE_FORMAT_A8R8G8B8_UNORM:
509 return &emit_A8R8G8B8_UNORM;
510
511 case PIPE_FORMAT_R32_UINT:
512 return &emit_R32_UINT;
513 case PIPE_FORMAT_R32G32_UINT:
514 return &emit_R32G32_UINT;
515 case PIPE_FORMAT_R32G32B32_UINT:
516 return &emit_R32G32B32_UINT;
517 case PIPE_FORMAT_R32G32B32A32_UINT:
518 return &emit_R32G32B32A32_UINT;
519
520 case PIPE_FORMAT_R16_UINT:
521 return &emit_R16_UINT;
522 case PIPE_FORMAT_R16G16_UINT:
523 return &emit_R16G16_UINT;
524 case PIPE_FORMAT_R16G16B16_UINT:
525 return &emit_R16G16B16_UINT;
526 case PIPE_FORMAT_R16G16B16A16_UINT:
527 return &emit_R16G16B16A16_UINT;
528
529 case PIPE_FORMAT_R8_UINT:
530 return &emit_R8_UINT;
531 case PIPE_FORMAT_R8G8_UINT:
532 return &emit_R8G8_UINT;
533 case PIPE_FORMAT_R8G8B8_UINT:
534 return &emit_R8G8B8_UINT;
535 case PIPE_FORMAT_R8G8B8A8_UINT:
536 return &emit_R8G8B8A8_UINT;
537
538 case PIPE_FORMAT_R32_SINT:
539 return &emit_R32_SINT;
540 case PIPE_FORMAT_R32G32_SINT:
541 return &emit_R32G32_SINT;
542 case PIPE_FORMAT_R32G32B32_SINT:
543 return &emit_R32G32B32_SINT;
544 case PIPE_FORMAT_R32G32B32A32_SINT:
545 return &emit_R32G32B32A32_SINT;
546
547 case PIPE_FORMAT_R16_SINT:
548 return &emit_R16_SINT;
549 case PIPE_FORMAT_R16G16_SINT:
550 return &emit_R16G16_SINT;
551 case PIPE_FORMAT_R16G16B16_SINT:
552 return &emit_R16G16B16_SINT;
553 case PIPE_FORMAT_R16G16B16A16_SINT:
554 return &emit_R16G16B16A16_SINT;
555
556 case PIPE_FORMAT_R8_SINT:
557 return &emit_R8_SINT;
558 case PIPE_FORMAT_R8G8_SINT:
559 return &emit_R8G8_SINT;
560 case PIPE_FORMAT_R8G8B8_SINT:
561 return &emit_R8G8B8_SINT;
562 case PIPE_FORMAT_R8G8B8A8_SINT:
563 return &emit_R8G8B8A8_SINT;
564
565 case PIPE_FORMAT_B10G10R10A2_UNORM:
566 return &emit_B10G10R10A2_UNORM;
567 case PIPE_FORMAT_B10G10R10A2_USCALED:
568 return &emit_B10G10R10A2_USCALED;
569 case PIPE_FORMAT_B10G10R10A2_SNORM:
570 return &emit_B10G10R10A2_SNORM;
571 case PIPE_FORMAT_B10G10R10A2_SSCALED:
572 return &emit_B10G10R10A2_SSCALED;
573
574 case PIPE_FORMAT_R10G10B10A2_UNORM:
575 return &emit_R10G10B10A2_UNORM;
576 case PIPE_FORMAT_R10G10B10A2_USCALED:
577 return &emit_R10G10B10A2_USCALED;
578 case PIPE_FORMAT_R10G10B10A2_SNORM:
579 return &emit_R10G10B10A2_SNORM;
580 case PIPE_FORMAT_R10G10B10A2_SSCALED:
581 return &emit_R10G10B10A2_SSCALED;
582
583 default:
584 assert(0);
585 return &emit_NULL;
586 }
587 }
588
589 static ALWAYS_INLINE void PIPE_CDECL
generic_run_one(struct translate_generic * tg,unsigned elt,unsigned start_instance,unsigned instance_id,void * vert)590 generic_run_one(struct translate_generic *tg,
591 unsigned elt,
592 unsigned start_instance,
593 unsigned instance_id,
594 void *vert)
595 {
596 unsigned nr_attrs = tg->nr_attrib;
597 unsigned attr;
598
599 for (attr = 0; attr < nr_attrs; attr++) {
600 float data[4];
601 uint8_t *dst = (uint8_t *)vert + tg->attrib[attr].output_offset;
602
603 if (tg->attrib[attr].type == TRANSLATE_ELEMENT_NORMAL) {
604 const uint8_t *src;
605 unsigned index;
606 int copy_size;
607
608 if (tg->attrib[attr].instance_divisor) {
609 index = start_instance;
610 index += (instance_id / tg->attrib[attr].instance_divisor);
611 /* XXX we need to clamp the index here too, but to a
612 * per-array max value, not the draw->pt.max_index value
613 * that's being given to us via translate->set_buffer().
614 */
615 }
616 else {
617 index = elt;
618 /* clamp to avoid going out of bounds */
619 index = MIN2(index, tg->attrib[attr].max_index);
620 }
621
622 src = tg->attrib[attr].input_ptr +
623 (ptrdiff_t)tg->attrib[attr].input_stride * index;
624
625 copy_size = tg->attrib[attr].copy_size;
626 if (likely(copy_size >= 0)) {
627 memcpy(dst, src, copy_size);
628 } else {
629 tg->attrib[attr].fetch(data, src, 0, 0);
630
631 if (0)
632 debug_printf("Fetch linear attr %d from %p stride %d index %d: "
633 " %f, %f, %f, %f \n",
634 attr,
635 tg->attrib[attr].input_ptr,
636 tg->attrib[attr].input_stride,
637 index,
638 data[0], data[1],data[2], data[3]);
639
640 tg->attrib[attr].emit(data, dst);
641 }
642 } else {
643 if (likely(tg->attrib[attr].copy_size >= 0)) {
644 memcpy(data, &instance_id, 4);
645 } else {
646 data[0] = (float)instance_id;
647 tg->attrib[attr].emit(data, dst);
648 }
649 }
650 }
651 }
652
653 /**
654 * Fetch vertex attributes for 'count' vertices.
655 */
656 static void PIPE_CDECL
generic_run_elts(struct translate * translate,const unsigned * elts,unsigned count,unsigned start_instance,unsigned instance_id,void * output_buffer)657 generic_run_elts(struct translate *translate,
658 const unsigned *elts,
659 unsigned count,
660 unsigned start_instance,
661 unsigned instance_id,
662 void *output_buffer)
663 {
664 struct translate_generic *tg = translate_generic(translate);
665 char *vert = output_buffer;
666 unsigned i;
667
668 for (i = 0; i < count; i++) {
669 generic_run_one(tg, *elts++, start_instance, instance_id, vert);
670 vert += tg->translate.key.output_stride;
671 }
672 }
673
674 static void PIPE_CDECL
generic_run_elts16(struct translate * translate,const uint16_t * elts,unsigned count,unsigned start_instance,unsigned instance_id,void * output_buffer)675 generic_run_elts16(struct translate *translate,
676 const uint16_t *elts,
677 unsigned count,
678 unsigned start_instance,
679 unsigned instance_id,
680 void *output_buffer)
681 {
682 struct translate_generic *tg = translate_generic(translate);
683 char *vert = output_buffer;
684 unsigned i;
685
686 for (i = 0; i < count; i++) {
687 generic_run_one(tg, *elts++, start_instance, instance_id, vert);
688 vert += tg->translate.key.output_stride;
689 }
690 }
691
692 static void PIPE_CDECL
generic_run_elts8(struct translate * translate,const uint8_t * elts,unsigned count,unsigned start_instance,unsigned instance_id,void * output_buffer)693 generic_run_elts8(struct translate *translate,
694 const uint8_t *elts,
695 unsigned count,
696 unsigned start_instance,
697 unsigned instance_id,
698 void *output_buffer)
699 {
700 struct translate_generic *tg = translate_generic(translate);
701 char *vert = output_buffer;
702 unsigned i;
703
704 for (i = 0; i < count; i++) {
705 generic_run_one(tg, *elts++, start_instance, instance_id, vert);
706 vert += tg->translate.key.output_stride;
707 }
708 }
709
710 static void PIPE_CDECL
generic_run(struct translate * translate,unsigned start,unsigned count,unsigned start_instance,unsigned instance_id,void * output_buffer)711 generic_run(struct translate *translate,
712 unsigned start,
713 unsigned count,
714 unsigned start_instance,
715 unsigned instance_id,
716 void *output_buffer)
717 {
718 struct translate_generic *tg = translate_generic(translate);
719 char *vert = output_buffer;
720 unsigned i;
721
722 for (i = 0; i < count; i++) {
723 generic_run_one(tg, start + i, start_instance, instance_id, vert);
724 vert += tg->translate.key.output_stride;
725 }
726 }
727
728
729
730 static void
generic_set_buffer(struct translate * translate,unsigned buf,const void * ptr,unsigned stride,unsigned max_index)731 generic_set_buffer(struct translate *translate,
732 unsigned buf,
733 const void *ptr,
734 unsigned stride,
735 unsigned max_index)
736 {
737 struct translate_generic *tg = translate_generic(translate);
738 unsigned i;
739
740 for (i = 0; i < tg->nr_attrib; i++) {
741 if (tg->attrib[i].buffer == buf) {
742 tg->attrib[i].input_ptr = ((const uint8_t *)ptr +
743 tg->attrib[i].input_offset);
744 tg->attrib[i].input_stride = stride;
745 tg->attrib[i].max_index = max_index;
746 }
747 }
748 }
749
750
751 static void
generic_release(struct translate * translate)752 generic_release(struct translate *translate)
753 {
754 /* Refcount?
755 */
756 FREE(translate);
757 }
758
759 static boolean
is_legal_int_format_combo(const struct util_format_description * src,const struct util_format_description * dst)760 is_legal_int_format_combo(const struct util_format_description *src,
761 const struct util_format_description *dst)
762 {
763 unsigned i;
764 unsigned nr = MIN2(src->nr_channels, dst->nr_channels);
765
766 for (i = 0; i < nr; i++) {
767 /* The signs must match. */
768 if (src->channel[i].type != dst->channel[i].type) {
769 return FALSE;
770 }
771
772 /* Integers must not lose precision at any point in the pipeline. */
773 if (src->channel[i].size > dst->channel[i].size) {
774 return FALSE;
775 }
776 }
777 return TRUE;
778 }
779
780 struct translate *
translate_generic_create(const struct translate_key * key)781 translate_generic_create(const struct translate_key *key)
782 {
783 struct translate_generic *tg = CALLOC_STRUCT(translate_generic);
784 unsigned i;
785
786 if (!tg)
787 return NULL;
788
789 assert(key->nr_elements <= TRANSLATE_MAX_ATTRIBS);
790
791 tg->translate.key = *key;
792 tg->translate.release = generic_release;
793 tg->translate.set_buffer = generic_set_buffer;
794 tg->translate.run_elts = generic_run_elts;
795 tg->translate.run_elts16 = generic_run_elts16;
796 tg->translate.run_elts8 = generic_run_elts8;
797 tg->translate.run = generic_run;
798
799 for (i = 0; i < key->nr_elements; i++) {
800 const struct util_format_description *format_desc =
801 util_format_description(key->element[i].input_format);
802
803 assert(format_desc);
804
805 tg->attrib[i].type = key->element[i].type;
806
807 if (format_desc->channel[0].pure_integer) {
808 const struct util_format_description *out_format_desc =
809 util_format_description(key->element[i].output_format);
810
811 if (!is_legal_int_format_combo(format_desc, out_format_desc)) {
812 FREE(tg);
813 return NULL;
814 }
815
816 if (format_desc->channel[0].type == UTIL_FORMAT_TYPE_SIGNED) {
817 assert(format_desc->fetch_rgba_sint);
818 tg->attrib[i].fetch = (fetch_func)format_desc->fetch_rgba_sint;
819 } else {
820 assert(format_desc->fetch_rgba_uint);
821 tg->attrib[i].fetch = (fetch_func)format_desc->fetch_rgba_uint;
822 }
823 } else {
824 assert(format_desc->fetch_rgba_float);
825 tg->attrib[i].fetch = (fetch_func)format_desc->fetch_rgba_float;
826 }
827
828 tg->attrib[i].buffer = key->element[i].input_buffer;
829 tg->attrib[i].input_offset = key->element[i].input_offset;
830 tg->attrib[i].instance_divisor = key->element[i].instance_divisor;
831
832 tg->attrib[i].output_offset = key->element[i].output_offset;
833
834 tg->attrib[i].copy_size = -1;
835 if (tg->attrib[i].type == TRANSLATE_ELEMENT_INSTANCE_ID) {
836 if (key->element[i].output_format == PIPE_FORMAT_R32_USCALED
837 || key->element[i].output_format == PIPE_FORMAT_R32_SSCALED)
838 tg->attrib[i].copy_size = 4;
839 } else {
840 if (key->element[i].input_format == key->element[i].output_format
841 && format_desc->block.width == 1
842 && format_desc->block.height == 1
843 && !(format_desc->block.bits & 7))
844 tg->attrib[i].copy_size = format_desc->block.bits >> 3;
845 }
846
847 if (tg->attrib[i].copy_size < 0)
848 tg->attrib[i].emit = get_emit_func(key->element[i].output_format);
849 else
850 tg->attrib[i].emit = NULL;
851 }
852
853 tg->nr_attrib = key->nr_elements;
854
855 return &tg->translate;
856 }
857
858 boolean
translate_generic_is_output_format_supported(enum pipe_format format)859 translate_generic_is_output_format_supported(enum pipe_format format)
860 {
861 switch(format) {
862 case PIPE_FORMAT_R64G64B64A64_FLOAT: return TRUE;
863 case PIPE_FORMAT_R64G64B64_FLOAT: return TRUE;
864 case PIPE_FORMAT_R64G64_FLOAT: return TRUE;
865 case PIPE_FORMAT_R64_FLOAT: return TRUE;
866
867 case PIPE_FORMAT_R32G32B32A32_FLOAT: return TRUE;
868 case PIPE_FORMAT_R32G32B32_FLOAT: return TRUE;
869 case PIPE_FORMAT_R32G32_FLOAT: return TRUE;
870 case PIPE_FORMAT_R32_FLOAT: return TRUE;
871
872 case PIPE_FORMAT_R16G16B16A16_FLOAT: return TRUE;
873 case PIPE_FORMAT_R16G16B16_FLOAT: return TRUE;
874 case PIPE_FORMAT_R16G16_FLOAT: return TRUE;
875 case PIPE_FORMAT_R16_FLOAT: return TRUE;
876
877 case PIPE_FORMAT_R32G32B32A32_USCALED: return TRUE;
878 case PIPE_FORMAT_R32G32B32_USCALED: return TRUE;
879 case PIPE_FORMAT_R32G32_USCALED: return TRUE;
880 case PIPE_FORMAT_R32_USCALED: return TRUE;
881
882 case PIPE_FORMAT_R32G32B32A32_SSCALED: return TRUE;
883 case PIPE_FORMAT_R32G32B32_SSCALED: return TRUE;
884 case PIPE_FORMAT_R32G32_SSCALED: return TRUE;
885 case PIPE_FORMAT_R32_SSCALED: return TRUE;
886
887 case PIPE_FORMAT_R32G32B32A32_UNORM: return TRUE;
888 case PIPE_FORMAT_R32G32B32_UNORM: return TRUE;
889 case PIPE_FORMAT_R32G32_UNORM: return TRUE;
890 case PIPE_FORMAT_R32_UNORM: return TRUE;
891
892 case PIPE_FORMAT_R32G32B32A32_SNORM: return TRUE;
893 case PIPE_FORMAT_R32G32B32_SNORM: return TRUE;
894 case PIPE_FORMAT_R32G32_SNORM: return TRUE;
895 case PIPE_FORMAT_R32_SNORM: return TRUE;
896
897 case PIPE_FORMAT_R16G16B16A16_USCALED: return TRUE;
898 case PIPE_FORMAT_R16G16B16_USCALED: return TRUE;
899 case PIPE_FORMAT_R16G16_USCALED: return TRUE;
900 case PIPE_FORMAT_R16_USCALED: return TRUE;
901
902 case PIPE_FORMAT_R16G16B16A16_SSCALED: return TRUE;
903 case PIPE_FORMAT_R16G16B16_SSCALED: return TRUE;
904 case PIPE_FORMAT_R16G16_SSCALED: return TRUE;
905 case PIPE_FORMAT_R16_SSCALED: return TRUE;
906
907 case PIPE_FORMAT_R16G16B16A16_UNORM: return TRUE;
908 case PIPE_FORMAT_R16G16B16_UNORM: return TRUE;
909 case PIPE_FORMAT_R16G16_UNORM: return TRUE;
910 case PIPE_FORMAT_R16_UNORM: return TRUE;
911
912 case PIPE_FORMAT_R16G16B16A16_SNORM: return TRUE;
913 case PIPE_FORMAT_R16G16B16_SNORM: return TRUE;
914 case PIPE_FORMAT_R16G16_SNORM: return TRUE;
915 case PIPE_FORMAT_R16_SNORM: return TRUE;
916
917 case PIPE_FORMAT_R8G8B8A8_USCALED: return TRUE;
918 case PIPE_FORMAT_R8G8B8_USCALED: return TRUE;
919 case PIPE_FORMAT_R8G8_USCALED: return TRUE;
920 case PIPE_FORMAT_R8_USCALED: return TRUE;
921
922 case PIPE_FORMAT_R8G8B8A8_SSCALED: return TRUE;
923 case PIPE_FORMAT_R8G8B8_SSCALED: return TRUE;
924 case PIPE_FORMAT_R8G8_SSCALED: return TRUE;
925 case PIPE_FORMAT_R8_SSCALED: return TRUE;
926
927 case PIPE_FORMAT_R8G8B8A8_UNORM: return TRUE;
928 case PIPE_FORMAT_R8G8B8_UNORM: return TRUE;
929 case PIPE_FORMAT_R8G8_UNORM: return TRUE;
930 case PIPE_FORMAT_R8_UNORM: return TRUE;
931
932 case PIPE_FORMAT_R8G8B8A8_SNORM: return TRUE;
933 case PIPE_FORMAT_R8G8B8_SNORM: return TRUE;
934 case PIPE_FORMAT_R8G8_SNORM: return TRUE;
935 case PIPE_FORMAT_R8_SNORM: return TRUE;
936
937 case PIPE_FORMAT_A8R8G8B8_UNORM: return TRUE;
938 case PIPE_FORMAT_B8G8R8A8_UNORM: return TRUE;
939
940 case PIPE_FORMAT_R32G32B32A32_UINT: return TRUE;
941 case PIPE_FORMAT_R32G32B32_UINT: return TRUE;
942 case PIPE_FORMAT_R32G32_UINT: return TRUE;
943 case PIPE_FORMAT_R32_UINT: return TRUE;
944
945 case PIPE_FORMAT_R16G16B16A16_UINT: return TRUE;
946 case PIPE_FORMAT_R16G16B16_UINT: return TRUE;
947 case PIPE_FORMAT_R16G16_UINT: return TRUE;
948 case PIPE_FORMAT_R16_UINT: return TRUE;
949
950 case PIPE_FORMAT_R8G8B8A8_UINT: return TRUE;
951 case PIPE_FORMAT_R8G8B8_UINT: return TRUE;
952 case PIPE_FORMAT_R8G8_UINT: return TRUE;
953 case PIPE_FORMAT_R8_UINT: return TRUE;
954
955 case PIPE_FORMAT_R32G32B32A32_SINT: return TRUE;
956 case PIPE_FORMAT_R32G32B32_SINT: return TRUE;
957 case PIPE_FORMAT_R32G32_SINT: return TRUE;
958 case PIPE_FORMAT_R32_SINT: return TRUE;
959
960 case PIPE_FORMAT_R16G16B16A16_SINT: return TRUE;
961 case PIPE_FORMAT_R16G16B16_SINT: return TRUE;
962 case PIPE_FORMAT_R16G16_SINT: return TRUE;
963 case PIPE_FORMAT_R16_SINT: return TRUE;
964
965 case PIPE_FORMAT_R8G8B8A8_SINT: return TRUE;
966 case PIPE_FORMAT_R8G8B8_SINT: return TRUE;
967 case PIPE_FORMAT_R8G8_SINT: return TRUE;
968 case PIPE_FORMAT_R8_SINT: return TRUE;
969
970 case PIPE_FORMAT_B10G10R10A2_UNORM: return TRUE;
971 case PIPE_FORMAT_B10G10R10A2_USCALED: return TRUE;
972 case PIPE_FORMAT_B10G10R10A2_SNORM: return TRUE;
973 case PIPE_FORMAT_B10G10R10A2_SSCALED: return TRUE;
974
975 case PIPE_FORMAT_R10G10B10A2_UNORM: return TRUE;
976 case PIPE_FORMAT_R10G10B10A2_USCALED: return TRUE;
977 case PIPE_FORMAT_R10G10B10A2_SNORM: return TRUE;
978 case PIPE_FORMAT_R10G10B10A2_SSCALED: return TRUE;
979
980 default: return FALSE;
981 }
982 }
983