vtn/opencl: native rsqrt support
[mesa.git] / src / compiler / spirv / vtn_opencl.c
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
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 =
43 vtn_value(b, w[1], vtn_value_type_type)->type->type;
44
45 unsigned num_srcs = count - 5;
46 nir_ssa_def *srcs[3] = { NULL };
47 vtn_assert(num_srcs <= ARRAY_SIZE(srcs));
48 for (unsigned i = 0; i < num_srcs; i++) {
49 srcs[i] = vtn_ssa_value(b, w[i + 5])->def;
50 }
51
52 nir_ssa_def *result = handler(b, opcode, num_srcs, srcs, dest_type);
53 if (result) {
54 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
55 val->ssa = vtn_create_ssa_value(b, dest_type);
56 val->ssa->def = result;
57 } else {
58 vtn_assert(dest_type == glsl_void_type());
59 }
60 }
61
62 static nir_op
63 nir_alu_op_for_opencl_opcode(struct vtn_builder *b,
64 enum OpenCLstd_Entrypoints opcode)
65 {
66 switch (opcode) {
67 case OpenCLstd_Fabs: return nir_op_fabs;
68 case OpenCLstd_SAbs: return nir_op_iabs;
69 case OpenCLstd_SAdd_sat: return nir_op_iadd_sat;
70 case OpenCLstd_UAdd_sat: return nir_op_uadd_sat;
71 case OpenCLstd_Ceil: return nir_op_fceil;
72 case OpenCLstd_Cos: return nir_op_fcos;
73 case OpenCLstd_Exp2: return nir_op_fexp2;
74 case OpenCLstd_Log2: return nir_op_flog2;
75 case OpenCLstd_Floor: return nir_op_ffloor;
76 case OpenCLstd_SHadd: return nir_op_ihadd;
77 case OpenCLstd_UHadd: return nir_op_uhadd;
78 case OpenCLstd_Fma: return nir_op_ffma;
79 case OpenCLstd_Fmax: return nir_op_fmax;
80 case OpenCLstd_SMax: return nir_op_imax;
81 case OpenCLstd_UMax: return nir_op_umax;
82 case OpenCLstd_Fmin: return nir_op_fmin;
83 case OpenCLstd_SMin: return nir_op_imin;
84 case OpenCLstd_UMin: return nir_op_umin;
85 case OpenCLstd_Fmod: return nir_op_fmod;
86 case OpenCLstd_Mix: return nir_op_flrp;
87 case OpenCLstd_Native_cos: return nir_op_fcos;
88 case OpenCLstd_Native_divide: return nir_op_fdiv;
89 case OpenCLstd_Native_exp2: return nir_op_fexp2;
90 case OpenCLstd_Native_log2: return nir_op_flog2;
91 case OpenCLstd_Native_powr: return nir_op_fpow;
92 case OpenCLstd_Native_recip: return nir_op_frcp;
93 case OpenCLstd_Native_rsqrt: return nir_op_frsq;
94 case OpenCLstd_Native_sin: return nir_op_fsin;
95 case OpenCLstd_SMul_hi: return nir_op_imul_high;
96 case OpenCLstd_UMul_hi: return nir_op_umul_high;
97 case OpenCLstd_Popcount: return nir_op_bit_count;
98 case OpenCLstd_Pow: return nir_op_fpow;
99 case OpenCLstd_Remainder: return nir_op_frem;
100 case OpenCLstd_SRhadd: return nir_op_irhadd;
101 case OpenCLstd_URhadd: return nir_op_urhadd;
102 case OpenCLstd_Rsqrt: return nir_op_frsq;
103 case OpenCLstd_Sign: return nir_op_fsign;
104 case OpenCLstd_Sin: return nir_op_fsin;
105 case OpenCLstd_Sqrt: return nir_op_fsqrt;
106 case OpenCLstd_SSub_sat: return nir_op_isub_sat;
107 case OpenCLstd_USub_sat: return nir_op_usub_sat;
108 case OpenCLstd_Trunc: return nir_op_ftrunc;
109 case OpenCLstd_Rint: return nir_op_fround_even;
110 /* uhm... */
111 case OpenCLstd_UAbs: return nir_op_mov;
112 default:
113 vtn_fail("No NIR equivalent");
114 }
115 }
116
117 static nir_ssa_def *
118 handle_alu(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode,
119 unsigned num_srcs, nir_ssa_def **srcs,
120 const struct glsl_type *dest_type)
121 {
122 return nir_build_alu(&b->nb, nir_alu_op_for_opencl_opcode(b, opcode),
123 srcs[0], srcs[1], srcs[2], NULL);
124 }
125
126 static nir_ssa_def *
127 handle_special(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode,
128 unsigned num_srcs, nir_ssa_def **srcs,
129 const struct glsl_type *dest_type)
130 {
131 nir_builder *nb = &b->nb;
132
133 switch (opcode) {
134 case OpenCLstd_SAbs_diff:
135 return nir_iabs_diff(nb, srcs[0], srcs[1]);
136 case OpenCLstd_UAbs_diff:
137 return nir_uabs_diff(nb, srcs[0], srcs[1]);
138 case OpenCLstd_Bitselect:
139 return nir_bitselect(nb, srcs[0], srcs[1], srcs[2]);
140 case OpenCLstd_SMad_hi:
141 return nir_imad_hi(nb, srcs[0], srcs[1], srcs[2]);
142 case OpenCLstd_UMad_hi:
143 return nir_umad_hi(nb, srcs[0], srcs[1], srcs[2]);
144 case OpenCLstd_SMul24:
145 return nir_imul24(nb, srcs[0], srcs[1]);
146 case OpenCLstd_UMul24:
147 return nir_umul24(nb, srcs[0], srcs[1]);
148 case OpenCLstd_SMad24:
149 return nir_imad24(nb, srcs[0], srcs[1], srcs[2]);
150 case OpenCLstd_UMad24:
151 return nir_umad24(nb, srcs[0], srcs[1], srcs[2]);
152 case OpenCLstd_FClamp:
153 return nir_fclamp(nb, srcs[0], srcs[1], srcs[2]);
154 case OpenCLstd_SClamp:
155 return nir_iclamp(nb, srcs[0], srcs[1], srcs[2]);
156 case OpenCLstd_UClamp:
157 return nir_uclamp(nb, srcs[0], srcs[1], srcs[2]);
158 case OpenCLstd_Copysign:
159 return nir_copysign(nb, srcs[0], srcs[1]);
160 case OpenCLstd_Cross:
161 if (glsl_get_components(dest_type) == 4)
162 return nir_cross4(nb, srcs[0], srcs[1]);
163 return nir_cross3(nb, srcs[0], srcs[1]);
164 case OpenCLstd_Degrees:
165 return nir_degrees(nb, srcs[0]);
166 case OpenCLstd_Fdim:
167 return nir_fdim(nb, srcs[0], srcs[1]);
168 case OpenCLstd_Distance:
169 return nir_distance(nb, srcs[0], srcs[1]);
170 case OpenCLstd_Fast_distance:
171 return nir_fast_distance(nb, srcs[0], srcs[1]);
172 case OpenCLstd_Fast_length:
173 return nir_fast_length(nb, srcs[0]);
174 case OpenCLstd_Fast_normalize:
175 return nir_fast_normalize(nb, srcs[0]);
176 case OpenCLstd_Length:
177 return nir_length(nb, srcs[0]);
178 case OpenCLstd_Mad:
179 return nir_fmad(nb, srcs[0], srcs[1], srcs[2]);
180 case OpenCLstd_Maxmag:
181 return nir_maxmag(nb, srcs[0], srcs[1]);
182 case OpenCLstd_Minmag:
183 return nir_minmag(nb, srcs[0], srcs[1]);
184 case OpenCLstd_Nan:
185 return nir_nan(nb, srcs[0]);
186 case OpenCLstd_Nextafter:
187 return nir_nextafter(nb, srcs[0], srcs[1]);
188 case OpenCLstd_Normalize:
189 return nir_normalize(nb, srcs[0]);
190 case OpenCLstd_Radians:
191 return nir_radians(nb, srcs[0]);
192 case OpenCLstd_Rotate:
193 return nir_rotate(nb, srcs[0], srcs[1]);
194 case OpenCLstd_Smoothstep:
195 return nir_smoothstep(nb, srcs[0], srcs[1], srcs[2]);
196 case OpenCLstd_Clz:
197 return nir_clz_u(nb, srcs[0]);
198 case OpenCLstd_Select:
199 return nir_select(nb, srcs[0], srcs[1], srcs[2]);
200 case OpenCLstd_Step:
201 return nir_sge(nb, srcs[1], srcs[0]);
202 case OpenCLstd_S_Upsample:
203 case OpenCLstd_U_Upsample:
204 return nir_upsample(nb, srcs[0], srcs[1]);
205 case OpenCLstd_Native_exp:
206 return nir_fexp(nb, srcs[0]);
207 case OpenCLstd_Native_exp10:
208 return nir_fexp2(nb, nir_fmul_imm(nb, srcs[0], log(10) / log(2)));
209 case OpenCLstd_Native_log:
210 return nir_flog(nb, srcs[0]);
211 case OpenCLstd_Native_log10:
212 return nir_fmul_imm(nb, nir_flog2(nb, srcs[0]), log(2) / log(10));
213 case OpenCLstd_Native_tan:
214 return nir_ftan(nb, srcs[0]);
215 default:
216 vtn_fail("No NIR equivalent");
217 return NULL;
218 }
219 }
220
221 static void
222 _handle_v_load_store(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode,
223 const uint32_t *w, unsigned count, bool load)
224 {
225 struct vtn_type *type;
226 if (load)
227 type = vtn_value(b, w[1], vtn_value_type_type)->type;
228 else
229 type = vtn_untyped_value(b, w[5])->type;
230 unsigned a = load ? 0 : 1;
231
232 const struct glsl_type *dest_type = type->type;
233 unsigned components = glsl_get_vector_elements(dest_type);
234
235 nir_ssa_def *offset = vtn_ssa_value(b, w[5 + a])->def;
236 struct vtn_value *p = vtn_value(b, w[6 + a], vtn_value_type_pointer);
237
238 struct vtn_ssa_value *comps[NIR_MAX_VEC_COMPONENTS];
239 nir_ssa_def *ncomps[NIR_MAX_VEC_COMPONENTS];
240
241 nir_ssa_def *moffset = nir_imul_imm(&b->nb, offset, components);
242 nir_deref_instr *deref = vtn_pointer_to_deref(b, p->pointer);
243
244 for (int i = 0; i < components; i++) {
245 nir_ssa_def *coffset = nir_iadd_imm(&b->nb, moffset, i);
246 nir_deref_instr *arr_deref = nir_build_deref_ptr_as_array(&b->nb, deref, coffset);
247
248 if (load) {
249 comps[i] = vtn_local_load(b, arr_deref, p->type->access);
250 ncomps[i] = comps[i]->def;
251 } else {
252 struct vtn_ssa_value *ssa = vtn_create_ssa_value(b, glsl_scalar_type(glsl_get_base_type(dest_type)));
253 struct vtn_ssa_value *val = vtn_ssa_value(b, w[5]);
254 ssa->def = nir_channel(&b->nb, val->def, i);
255 vtn_local_store(b, ssa, arr_deref, p->type->access);
256 }
257 }
258 if (load) {
259 struct vtn_ssa_value *ssa = vtn_create_ssa_value(b, dest_type);
260 ssa->def = nir_vec(&b->nb, ncomps, components);
261 vtn_push_ssa(b, w[2], type, ssa);
262 }
263 }
264
265 static void
266 vtn_handle_opencl_vload(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode,
267 const uint32_t *w, unsigned count)
268 {
269 _handle_v_load_store(b, opcode, w, count, true);
270 }
271
272 static void
273 vtn_handle_opencl_vstore(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode,
274 const uint32_t *w, unsigned count)
275 {
276 _handle_v_load_store(b, opcode, w, count, false);
277 }
278
279 static nir_ssa_def *
280 handle_printf(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode,
281 unsigned num_srcs, nir_ssa_def **srcs,
282 const struct glsl_type *dest_type)
283 {
284 /* hahah, yeah, right.. */
285 return nir_imm_int(&b->nb, -1);
286 }
287
288 static nir_ssa_def *
289 handle_shuffle(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode, unsigned num_srcs,
290 nir_ssa_def **srcs, const struct glsl_type *dest_type)
291 {
292 struct nir_ssa_def *input = srcs[0];
293 struct nir_ssa_def *mask = srcs[1];
294
295 unsigned out_elems = glsl_get_vector_elements(dest_type);
296 nir_ssa_def *outres[NIR_MAX_VEC_COMPONENTS];
297 unsigned in_elems = input->num_components;
298 if (mask->bit_size != 32)
299 mask = nir_u2u32(&b->nb, mask);
300 mask = nir_iand(&b->nb, mask, nir_imm_intN_t(&b->nb, in_elems - 1, mask->bit_size));
301 for (unsigned i = 0; i < out_elems; i++)
302 outres[i] = nir_vector_extract(&b->nb, input, nir_channel(&b->nb, mask, i));
303
304 return nir_vec(&b->nb, outres, out_elems);
305 }
306
307 static nir_ssa_def *
308 handle_shuffle2(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode, unsigned num_srcs,
309 nir_ssa_def **srcs, const struct glsl_type *dest_type)
310 {
311 struct nir_ssa_def *input0 = srcs[0];
312 struct nir_ssa_def *input1 = srcs[1];
313 struct nir_ssa_def *mask = srcs[2];
314
315 unsigned out_elems = glsl_get_vector_elements(dest_type);
316 nir_ssa_def *outres[NIR_MAX_VEC_COMPONENTS];
317 unsigned in_elems = input0->num_components;
318 unsigned total_mask = 2 * in_elems - 1;
319 unsigned half_mask = in_elems - 1;
320 if (mask->bit_size != 32)
321 mask = nir_u2u32(&b->nb, mask);
322 mask = nir_iand(&b->nb, mask, nir_imm_intN_t(&b->nb, total_mask, mask->bit_size));
323 for (unsigned i = 0; i < out_elems; i++) {
324 nir_ssa_def *this_mask = nir_channel(&b->nb, mask, i);
325 nir_ssa_def *vmask = nir_iand(&b->nb, this_mask, nir_imm_intN_t(&b->nb, half_mask, mask->bit_size));
326 nir_ssa_def *val0 = nir_vector_extract(&b->nb, input0, vmask);
327 nir_ssa_def *val1 = nir_vector_extract(&b->nb, input1, vmask);
328 nir_ssa_def *sel = nir_ilt(&b->nb, this_mask, nir_imm_intN_t(&b->nb, in_elems, mask->bit_size));
329 outres[i] = nir_bcsel(&b->nb, sel, val0, val1);
330 }
331 return nir_vec(&b->nb, outres, out_elems);
332 }
333
334 bool
335 vtn_handle_opencl_instruction(struct vtn_builder *b, SpvOp ext_opcode,
336 const uint32_t *w, unsigned count)
337 {
338 enum OpenCLstd_Entrypoints cl_opcode = (enum OpenCLstd_Entrypoints) ext_opcode;
339
340 switch (cl_opcode) {
341 case OpenCLstd_Fabs:
342 case OpenCLstd_SAbs:
343 case OpenCLstd_UAbs:
344 case OpenCLstd_SAdd_sat:
345 case OpenCLstd_UAdd_sat:
346 case OpenCLstd_Ceil:
347 case OpenCLstd_Cos:
348 case OpenCLstd_Exp2:
349 case OpenCLstd_Log2:
350 case OpenCLstd_Floor:
351 case OpenCLstd_Fma:
352 case OpenCLstd_Fmax:
353 case OpenCLstd_SHadd:
354 case OpenCLstd_UHadd:
355 case OpenCLstd_SMax:
356 case OpenCLstd_UMax:
357 case OpenCLstd_Fmin:
358 case OpenCLstd_SMin:
359 case OpenCLstd_UMin:
360 case OpenCLstd_Mix:
361 case OpenCLstd_Native_cos:
362 case OpenCLstd_Native_divide:
363 case OpenCLstd_Native_exp2:
364 case OpenCLstd_Native_log2:
365 case OpenCLstd_Native_powr:
366 case OpenCLstd_Native_recip:
367 case OpenCLstd_Native_rsqrt:
368 case OpenCLstd_Native_sin:
369 case OpenCLstd_Fmod:
370 case OpenCLstd_SMul_hi:
371 case OpenCLstd_UMul_hi:
372 case OpenCLstd_Popcount:
373 case OpenCLstd_Pow:
374 case OpenCLstd_Remainder:
375 case OpenCLstd_SRhadd:
376 case OpenCLstd_URhadd:
377 case OpenCLstd_Rsqrt:
378 case OpenCLstd_Sign:
379 case OpenCLstd_Sin:
380 case OpenCLstd_Sqrt:
381 case OpenCLstd_SSub_sat:
382 case OpenCLstd_USub_sat:
383 case OpenCLstd_Trunc:
384 case OpenCLstd_Rint:
385 handle_instr(b, cl_opcode, w, count, handle_alu);
386 return true;
387 case OpenCLstd_SAbs_diff:
388 case OpenCLstd_UAbs_diff:
389 case OpenCLstd_SMad_hi:
390 case OpenCLstd_UMad_hi:
391 case OpenCLstd_SMad24:
392 case OpenCLstd_UMad24:
393 case OpenCLstd_SMul24:
394 case OpenCLstd_UMul24:
395 case OpenCLstd_Bitselect:
396 case OpenCLstd_FClamp:
397 case OpenCLstd_SClamp:
398 case OpenCLstd_UClamp:
399 case OpenCLstd_Copysign:
400 case OpenCLstd_Cross:
401 case OpenCLstd_Degrees:
402 case OpenCLstd_Fdim:
403 case OpenCLstd_Distance:
404 case OpenCLstd_Fast_distance:
405 case OpenCLstd_Fast_length:
406 case OpenCLstd_Fast_normalize:
407 case OpenCLstd_Length:
408 case OpenCLstd_Mad:
409 case OpenCLstd_Maxmag:
410 case OpenCLstd_Minmag:
411 case OpenCLstd_Nan:
412 case OpenCLstd_Nextafter:
413 case OpenCLstd_Normalize:
414 case OpenCLstd_Radians:
415 case OpenCLstd_Rotate:
416 case OpenCLstd_Select:
417 case OpenCLstd_Step:
418 case OpenCLstd_Smoothstep:
419 case OpenCLstd_S_Upsample:
420 case OpenCLstd_U_Upsample:
421 case OpenCLstd_Clz:
422 case OpenCLstd_Native_exp:
423 case OpenCLstd_Native_exp10:
424 case OpenCLstd_Native_log:
425 case OpenCLstd_Native_log10:
426 case OpenCLstd_Native_tan:
427 handle_instr(b, cl_opcode, w, count, handle_special);
428 return true;
429 case OpenCLstd_Vloadn:
430 vtn_handle_opencl_vload(b, cl_opcode, w, count);
431 return true;
432 case OpenCLstd_Vstoren:
433 vtn_handle_opencl_vstore(b, cl_opcode, w, count);
434 return true;
435 case OpenCLstd_Shuffle:
436 handle_instr(b, cl_opcode, w, count, handle_shuffle);
437 return true;
438 case OpenCLstd_Shuffle2:
439 handle_instr(b, cl_opcode, w, count, handle_shuffle2);
440 return true;
441 case OpenCLstd_Printf:
442 handle_instr(b, cl_opcode, w, count, handle_printf);
443 return true;
444 case OpenCLstd_Prefetch:
445 /* TODO maybe add a nir instruction for this? */
446 return true;
447 default:
448 vtn_fail("unhandled opencl opc: %u\n", ext_opcode);
449 return false;
450 }
451 }