1 /*
2 * Copyright © 2018 Red Hat
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 *
23 * Authors:
24 * Rob Clark (robdclark@gmail.com)
25 */
26
27 #include "math.h"
28 #include "nir/nir_builtin_builder.h"
29
30 #include "vtn_private.h"
31 #include "OpenCL.std.h"
32
33 typedef nir_ssa_def *(*nir_handler)(struct vtn_builder *b,
34 enum OpenCLstd_Entrypoints opcode,
35 unsigned num_srcs, nir_ssa_def **srcs,
36 const struct glsl_type *dest_type);
37
38 static void
handle_instr(struct vtn_builder * b,enum OpenCLstd_Entrypoints opcode,const uint32_t * w,unsigned count,nir_handler handler)39 handle_instr(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode,
40 const uint32_t *w, unsigned count, nir_handler handler)
41 {
42 const struct glsl_type *dest_type = vtn_get_type(b, w[1])->type;
43
44 unsigned num_srcs = count - 5;
45 nir_ssa_def *srcs[3] = { NULL };
46 vtn_assert(num_srcs <= ARRAY_SIZE(srcs));
47 for (unsigned i = 0; i < num_srcs; i++) {
48 srcs[i] = vtn_get_nir_ssa(b, w[i + 5]);
49 }
50
51 nir_ssa_def *result = handler(b, opcode, num_srcs, srcs, dest_type);
52 if (result) {
53 vtn_push_nir_ssa(b, w[2], result);
54 } else {
55 vtn_assert(dest_type == glsl_void_type());
56 }
57 }
58
59 static nir_op
nir_alu_op_for_opencl_opcode(struct vtn_builder * b,enum OpenCLstd_Entrypoints opcode)60 nir_alu_op_for_opencl_opcode(struct vtn_builder *b,
61 enum OpenCLstd_Entrypoints opcode)
62 {
63 switch (opcode) {
64 case OpenCLstd_Fabs: return nir_op_fabs;
65 case OpenCLstd_SAbs: return nir_op_iabs;
66 case OpenCLstd_SAdd_sat: return nir_op_iadd_sat;
67 case OpenCLstd_UAdd_sat: return nir_op_uadd_sat;
68 case OpenCLstd_Ceil: return nir_op_fceil;
69 case OpenCLstd_Cos: return nir_op_fcos;
70 case OpenCLstd_Exp2: return nir_op_fexp2;
71 case OpenCLstd_Log2: return nir_op_flog2;
72 case OpenCLstd_Floor: return nir_op_ffloor;
73 case OpenCLstd_SHadd: return nir_op_ihadd;
74 case OpenCLstd_UHadd: return nir_op_uhadd;
75 case OpenCLstd_Fma: return nir_op_ffma;
76 case OpenCLstd_Fmax: return nir_op_fmax;
77 case OpenCLstd_SMax: return nir_op_imax;
78 case OpenCLstd_UMax: return nir_op_umax;
79 case OpenCLstd_Fmin: return nir_op_fmin;
80 case OpenCLstd_SMin: return nir_op_imin;
81 case OpenCLstd_UMin: return nir_op_umin;
82 case OpenCLstd_Fmod: return nir_op_fmod;
83 case OpenCLstd_Mix: return nir_op_flrp;
84 case OpenCLstd_Native_cos: return nir_op_fcos;
85 case OpenCLstd_Native_divide: return nir_op_fdiv;
86 case OpenCLstd_Native_exp2: return nir_op_fexp2;
87 case OpenCLstd_Native_log2: return nir_op_flog2;
88 case OpenCLstd_Native_powr: return nir_op_fpow;
89 case OpenCLstd_Native_recip: return nir_op_frcp;
90 case OpenCLstd_Native_rsqrt: return nir_op_frsq;
91 case OpenCLstd_Native_sin: return nir_op_fsin;
92 case OpenCLstd_Native_sqrt: return nir_op_fsqrt;
93 case OpenCLstd_SMul_hi: return nir_op_imul_high;
94 case OpenCLstd_UMul_hi: return nir_op_umul_high;
95 case OpenCLstd_Popcount: return nir_op_bit_count;
96 case OpenCLstd_Pow: return nir_op_fpow;
97 case OpenCLstd_Remainder: return nir_op_frem;
98 case OpenCLstd_SRhadd: return nir_op_irhadd;
99 case OpenCLstd_URhadd: return nir_op_urhadd;
100 case OpenCLstd_Rsqrt: return nir_op_frsq;
101 case OpenCLstd_Sign: return nir_op_fsign;
102 case OpenCLstd_Sin: return nir_op_fsin;
103 case OpenCLstd_Sqrt: return nir_op_fsqrt;
104 case OpenCLstd_SSub_sat: return nir_op_isub_sat;
105 case OpenCLstd_USub_sat: return nir_op_usub_sat;
106 case OpenCLstd_Trunc: return nir_op_ftrunc;
107 case OpenCLstd_Rint: return nir_op_fround_even;
108 /* uhm... */
109 case OpenCLstd_UAbs: return nir_op_mov;
110 default:
111 vtn_fail("No NIR equivalent");
112 }
113 }
114
115 static nir_ssa_def *
handle_alu(struct vtn_builder * b,enum OpenCLstd_Entrypoints opcode,unsigned num_srcs,nir_ssa_def ** srcs,const struct glsl_type * dest_type)116 handle_alu(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode,
117 unsigned num_srcs, nir_ssa_def **srcs,
118 const struct glsl_type *dest_type)
119 {
120 return nir_build_alu(&b->nb, nir_alu_op_for_opencl_opcode(b, opcode),
121 srcs[0], srcs[1], srcs[2], NULL);
122 }
123
124 static nir_ssa_def *
handle_special(struct vtn_builder * b,enum OpenCLstd_Entrypoints opcode,unsigned num_srcs,nir_ssa_def ** srcs,const struct glsl_type * dest_type)125 handle_special(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode,
126 unsigned num_srcs, nir_ssa_def **srcs,
127 const struct glsl_type *dest_type)
128 {
129 nir_builder *nb = &b->nb;
130
131 switch (opcode) {
132 case OpenCLstd_SAbs_diff:
133 return nir_iabs_diff(nb, srcs[0], srcs[1]);
134 case OpenCLstd_UAbs_diff:
135 return nir_uabs_diff(nb, srcs[0], srcs[1]);
136 case OpenCLstd_Bitselect:
137 return nir_bitselect(nb, srcs[0], srcs[1], srcs[2]);
138 case OpenCLstd_SMad_hi:
139 return nir_imad_hi(nb, srcs[0], srcs[1], srcs[2]);
140 case OpenCLstd_UMad_hi:
141 return nir_umad_hi(nb, srcs[0], srcs[1], srcs[2]);
142 case OpenCLstd_SMul24:
143 return nir_imul24(nb, srcs[0], srcs[1]);
144 case OpenCLstd_UMul24:
145 return nir_umul24(nb, srcs[0], srcs[1]);
146 case OpenCLstd_SMad24:
147 return nir_imad24(nb, srcs[0], srcs[1], srcs[2]);
148 case OpenCLstd_UMad24:
149 return nir_umad24(nb, srcs[0], srcs[1], srcs[2]);
150 case OpenCLstd_FClamp:
151 return nir_fclamp(nb, srcs[0], srcs[1], srcs[2]);
152 case OpenCLstd_SClamp:
153 return nir_iclamp(nb, srcs[0], srcs[1], srcs[2]);
154 case OpenCLstd_UClamp:
155 return nir_uclamp(nb, srcs[0], srcs[1], srcs[2]);
156 case OpenCLstd_Copysign:
157 return nir_copysign(nb, srcs[0], srcs[1]);
158 case OpenCLstd_Cross:
159 if (glsl_get_components(dest_type) == 4)
160 return nir_cross4(nb, srcs[0], srcs[1]);
161 return nir_cross3(nb, srcs[0], srcs[1]);
162 case OpenCLstd_Degrees:
163 return nir_degrees(nb, srcs[0]);
164 case OpenCLstd_Fdim:
165 return nir_fdim(nb, srcs[0], srcs[1]);
166 case OpenCLstd_Distance:
167 return nir_distance(nb, srcs[0], srcs[1]);
168 case OpenCLstd_Fast_distance:
169 return nir_fast_distance(nb, srcs[0], srcs[1]);
170 case OpenCLstd_Fast_length:
171 return nir_fast_length(nb, srcs[0]);
172 case OpenCLstd_Fast_normalize:
173 return nir_fast_normalize(nb, srcs[0]);
174 case OpenCLstd_Length:
175 return nir_length(nb, srcs[0]);
176 case OpenCLstd_Mad:
177 return nir_fmad(nb, srcs[0], srcs[1], srcs[2]);
178 case OpenCLstd_Maxmag:
179 return nir_maxmag(nb, srcs[0], srcs[1]);
180 case OpenCLstd_Minmag:
181 return nir_minmag(nb, srcs[0], srcs[1]);
182 case OpenCLstd_Nan:
183 return nir_nan(nb, srcs[0]);
184 case OpenCLstd_Nextafter:
185 return nir_nextafter(nb, srcs[0], srcs[1]);
186 case OpenCLstd_Normalize:
187 return nir_normalize(nb, srcs[0]);
188 case OpenCLstd_Radians:
189 return nir_radians(nb, srcs[0]);
190 case OpenCLstd_Rotate:
191 return nir_rotate(nb, srcs[0], srcs[1]);
192 case OpenCLstd_Smoothstep:
193 return nir_smoothstep(nb, srcs[0], srcs[1], srcs[2]);
194 case OpenCLstd_Clz:
195 return nir_clz_u(nb, srcs[0]);
196 case OpenCLstd_Select:
197 return nir_select(nb, srcs[0], srcs[1], srcs[2]);
198 case OpenCLstd_Step:
199 return nir_sge(nb, srcs[1], srcs[0]);
200 case OpenCLstd_S_Upsample:
201 case OpenCLstd_U_Upsample:
202 return nir_upsample(nb, srcs[0], srcs[1]);
203 case OpenCLstd_Native_exp:
204 return nir_fexp(nb, srcs[0]);
205 case OpenCLstd_Native_exp10:
206 return nir_fexp2(nb, nir_fmul_imm(nb, srcs[0], log(10) / log(2)));
207 case OpenCLstd_Native_log:
208 return nir_flog(nb, srcs[0]);
209 case OpenCLstd_Native_log10:
210 return nir_fmul_imm(nb, nir_flog2(nb, srcs[0]), log(2) / log(10));
211 case OpenCLstd_Native_tan:
212 return nir_ftan(nb, srcs[0]);
213 default:
214 vtn_fail("No NIR equivalent");
215 return NULL;
216 }
217 }
218
219 static void
_handle_v_load_store(struct vtn_builder * b,enum OpenCLstd_Entrypoints opcode,const uint32_t * w,unsigned count,bool load)220 _handle_v_load_store(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode,
221 const uint32_t *w, unsigned count, bool load)
222 {
223 struct vtn_type *type;
224 if (load)
225 type = vtn_get_type(b, w[1]);
226 else
227 type = vtn_get_value_type(b, w[5]);
228 unsigned a = load ? 0 : 1;
229
230 const struct glsl_type *dest_type = type->type;
231 unsigned components = glsl_get_vector_elements(dest_type);
232
233 nir_ssa_def *offset = vtn_get_nir_ssa(b, w[5 + a]);
234 struct vtn_value *p = vtn_value(b, w[6 + a], vtn_value_type_pointer);
235
236 struct vtn_ssa_value *comps[NIR_MAX_VEC_COMPONENTS];
237 nir_ssa_def *ncomps[NIR_MAX_VEC_COMPONENTS];
238
239 nir_ssa_def *moffset = nir_imul_imm(&b->nb, offset, components);
240 nir_deref_instr *deref = vtn_pointer_to_deref(b, p->pointer);
241
242 for (int i = 0; i < components; i++) {
243 nir_ssa_def *coffset = nir_iadd_imm(&b->nb, moffset, i);
244 nir_deref_instr *arr_deref = nir_build_deref_ptr_as_array(&b->nb, deref, coffset);
245
246 if (load) {
247 comps[i] = vtn_local_load(b, arr_deref, p->type->access);
248 ncomps[i] = comps[i]->def;
249 } else {
250 struct vtn_ssa_value *ssa = vtn_create_ssa_value(b, glsl_scalar_type(glsl_get_base_type(dest_type)));
251 struct vtn_ssa_value *val = vtn_ssa_value(b, w[5]);
252 ssa->def = nir_channel(&b->nb, val->def, i);
253 vtn_local_store(b, ssa, arr_deref, p->type->access);
254 }
255 }
256 if (load) {
257 vtn_push_nir_ssa(b, w[2], nir_vec(&b->nb, ncomps, components));
258 }
259 }
260
261 static void
vtn_handle_opencl_vload(struct vtn_builder * b,enum OpenCLstd_Entrypoints opcode,const uint32_t * w,unsigned count)262 vtn_handle_opencl_vload(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode,
263 const uint32_t *w, unsigned count)
264 {
265 _handle_v_load_store(b, opcode, w, count, true);
266 }
267
268 static void
vtn_handle_opencl_vstore(struct vtn_builder * b,enum OpenCLstd_Entrypoints opcode,const uint32_t * w,unsigned count)269 vtn_handle_opencl_vstore(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode,
270 const uint32_t *w, unsigned count)
271 {
272 _handle_v_load_store(b, opcode, w, count, false);
273 }
274
275 static nir_ssa_def *
handle_printf(struct vtn_builder * b,enum OpenCLstd_Entrypoints opcode,unsigned num_srcs,nir_ssa_def ** srcs,const struct glsl_type * dest_type)276 handle_printf(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode,
277 unsigned num_srcs, nir_ssa_def **srcs,
278 const struct glsl_type *dest_type)
279 {
280 /* hahah, yeah, right.. */
281 return nir_imm_int(&b->nb, -1);
282 }
283
284 static nir_ssa_def *
handle_shuffle(struct vtn_builder * b,enum OpenCLstd_Entrypoints opcode,unsigned num_srcs,nir_ssa_def ** srcs,const struct glsl_type * dest_type)285 handle_shuffle(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode, unsigned num_srcs,
286 nir_ssa_def **srcs, const struct glsl_type *dest_type)
287 {
288 struct nir_ssa_def *input = srcs[0];
289 struct nir_ssa_def *mask = srcs[1];
290
291 unsigned out_elems = glsl_get_vector_elements(dest_type);
292 nir_ssa_def *outres[NIR_MAX_VEC_COMPONENTS];
293 unsigned in_elems = input->num_components;
294 if (mask->bit_size != 32)
295 mask = nir_u2u32(&b->nb, mask);
296 mask = nir_iand(&b->nb, mask, nir_imm_intN_t(&b->nb, in_elems - 1, mask->bit_size));
297 for (unsigned i = 0; i < out_elems; i++)
298 outres[i] = nir_vector_extract(&b->nb, input, nir_channel(&b->nb, mask, i));
299
300 return nir_vec(&b->nb, outres, out_elems);
301 }
302
303 static nir_ssa_def *
handle_shuffle2(struct vtn_builder * b,enum OpenCLstd_Entrypoints opcode,unsigned num_srcs,nir_ssa_def ** srcs,const struct glsl_type * dest_type)304 handle_shuffle2(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode, unsigned num_srcs,
305 nir_ssa_def **srcs, const struct glsl_type *dest_type)
306 {
307 struct nir_ssa_def *input0 = srcs[0];
308 struct nir_ssa_def *input1 = srcs[1];
309 struct nir_ssa_def *mask = srcs[2];
310
311 unsigned out_elems = glsl_get_vector_elements(dest_type);
312 nir_ssa_def *outres[NIR_MAX_VEC_COMPONENTS];
313 unsigned in_elems = input0->num_components;
314 unsigned total_mask = 2 * in_elems - 1;
315 unsigned half_mask = in_elems - 1;
316 if (mask->bit_size != 32)
317 mask = nir_u2u32(&b->nb, mask);
318 mask = nir_iand(&b->nb, mask, nir_imm_intN_t(&b->nb, total_mask, mask->bit_size));
319 for (unsigned i = 0; i < out_elems; i++) {
320 nir_ssa_def *this_mask = nir_channel(&b->nb, mask, i);
321 nir_ssa_def *vmask = nir_iand(&b->nb, this_mask, nir_imm_intN_t(&b->nb, half_mask, mask->bit_size));
322 nir_ssa_def *val0 = nir_vector_extract(&b->nb, input0, vmask);
323 nir_ssa_def *val1 = nir_vector_extract(&b->nb, input1, vmask);
324 nir_ssa_def *sel = nir_ilt(&b->nb, this_mask, nir_imm_intN_t(&b->nb, in_elems, mask->bit_size));
325 outres[i] = nir_bcsel(&b->nb, sel, val0, val1);
326 }
327 return nir_vec(&b->nb, outres, out_elems);
328 }
329
330 bool
vtn_handle_opencl_instruction(struct vtn_builder * b,SpvOp ext_opcode,const uint32_t * w,unsigned count)331 vtn_handle_opencl_instruction(struct vtn_builder *b, SpvOp ext_opcode,
332 const uint32_t *w, unsigned count)
333 {
334 enum OpenCLstd_Entrypoints cl_opcode = (enum OpenCLstd_Entrypoints) ext_opcode;
335
336 switch (cl_opcode) {
337 case OpenCLstd_Fabs:
338 case OpenCLstd_SAbs:
339 case OpenCLstd_UAbs:
340 case OpenCLstd_SAdd_sat:
341 case OpenCLstd_UAdd_sat:
342 case OpenCLstd_Ceil:
343 case OpenCLstd_Cos:
344 case OpenCLstd_Exp2:
345 case OpenCLstd_Log2:
346 case OpenCLstd_Floor:
347 case OpenCLstd_Fma:
348 case OpenCLstd_Fmax:
349 case OpenCLstd_SHadd:
350 case OpenCLstd_UHadd:
351 case OpenCLstd_SMax:
352 case OpenCLstd_UMax:
353 case OpenCLstd_Fmin:
354 case OpenCLstd_SMin:
355 case OpenCLstd_UMin:
356 case OpenCLstd_Mix:
357 case OpenCLstd_Native_cos:
358 case OpenCLstd_Native_divide:
359 case OpenCLstd_Native_exp2:
360 case OpenCLstd_Native_log2:
361 case OpenCLstd_Native_powr:
362 case OpenCLstd_Native_recip:
363 case OpenCLstd_Native_rsqrt:
364 case OpenCLstd_Native_sin:
365 case OpenCLstd_Native_sqrt:
366 case OpenCLstd_Fmod:
367 case OpenCLstd_SMul_hi:
368 case OpenCLstd_UMul_hi:
369 case OpenCLstd_Popcount:
370 case OpenCLstd_Pow:
371 case OpenCLstd_Remainder:
372 case OpenCLstd_SRhadd:
373 case OpenCLstd_URhadd:
374 case OpenCLstd_Rsqrt:
375 case OpenCLstd_Sign:
376 case OpenCLstd_Sin:
377 case OpenCLstd_Sqrt:
378 case OpenCLstd_SSub_sat:
379 case OpenCLstd_USub_sat:
380 case OpenCLstd_Trunc:
381 case OpenCLstd_Rint:
382 handle_instr(b, cl_opcode, w, count, handle_alu);
383 return true;
384 case OpenCLstd_SAbs_diff:
385 case OpenCLstd_UAbs_diff:
386 case OpenCLstd_SMad_hi:
387 case OpenCLstd_UMad_hi:
388 case OpenCLstd_SMad24:
389 case OpenCLstd_UMad24:
390 case OpenCLstd_SMul24:
391 case OpenCLstd_UMul24:
392 case OpenCLstd_Bitselect:
393 case OpenCLstd_FClamp:
394 case OpenCLstd_SClamp:
395 case OpenCLstd_UClamp:
396 case OpenCLstd_Copysign:
397 case OpenCLstd_Cross:
398 case OpenCLstd_Degrees:
399 case OpenCLstd_Fdim:
400 case OpenCLstd_Distance:
401 case OpenCLstd_Fast_distance:
402 case OpenCLstd_Fast_length:
403 case OpenCLstd_Fast_normalize:
404 case OpenCLstd_Length:
405 case OpenCLstd_Mad:
406 case OpenCLstd_Maxmag:
407 case OpenCLstd_Minmag:
408 case OpenCLstd_Nan:
409 case OpenCLstd_Nextafter:
410 case OpenCLstd_Normalize:
411 case OpenCLstd_Radians:
412 case OpenCLstd_Rotate:
413 case OpenCLstd_Select:
414 case OpenCLstd_Step:
415 case OpenCLstd_Smoothstep:
416 case OpenCLstd_S_Upsample:
417 case OpenCLstd_U_Upsample:
418 case OpenCLstd_Clz:
419 case OpenCLstd_Native_exp:
420 case OpenCLstd_Native_exp10:
421 case OpenCLstd_Native_log:
422 case OpenCLstd_Native_log10:
423 case OpenCLstd_Native_tan:
424 handle_instr(b, cl_opcode, w, count, handle_special);
425 return true;
426 case OpenCLstd_Vloadn:
427 vtn_handle_opencl_vload(b, cl_opcode, w, count);
428 return true;
429 case OpenCLstd_Vstoren:
430 vtn_handle_opencl_vstore(b, cl_opcode, w, count);
431 return true;
432 case OpenCLstd_Shuffle:
433 handle_instr(b, cl_opcode, w, count, handle_shuffle);
434 return true;
435 case OpenCLstd_Shuffle2:
436 handle_instr(b, cl_opcode, w, count, handle_shuffle2);
437 return true;
438 case OpenCLstd_Printf:
439 handle_instr(b, cl_opcode, w, count, handle_printf);
440 return true;
441 case OpenCLstd_Prefetch:
442 /* TODO maybe add a nir instruction for this? */
443 return true;
444 default:
445 vtn_fail("unhandled opencl opc: %u\n", ext_opcode);
446 return false;
447 }
448 }
449