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