vtn/opencl: add shuffle/shuffle 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_SMul_hi: return nir_op_imul_high;
88 case OpenCLstd_UMul_hi: return nir_op_umul_high;
89 case OpenCLstd_Popcount: return nir_op_bit_count;
90 case OpenCLstd_Pow: return nir_op_fpow;
91 case OpenCLstd_Remainder: return nir_op_frem;
92 case OpenCLstd_SRhadd: return nir_op_irhadd;
93 case OpenCLstd_URhadd: return nir_op_urhadd;
94 case OpenCLstd_Rsqrt: return nir_op_frsq;
95 case OpenCLstd_Sign: return nir_op_fsign;
96 case OpenCLstd_Sin: return nir_op_fsin;
97 case OpenCLstd_Sqrt: return nir_op_fsqrt;
98 case OpenCLstd_SSub_sat: return nir_op_isub_sat;
99 case OpenCLstd_USub_sat: return nir_op_usub_sat;
100 case OpenCLstd_Trunc: return nir_op_ftrunc;
101 /* uhm... */
102 case OpenCLstd_UAbs: return nir_op_mov;
103 default:
104 vtn_fail("No NIR equivalent");
105 }
106 }
107
108 static nir_ssa_def *
109 handle_alu(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode,
110 unsigned num_srcs, nir_ssa_def **srcs,
111 const struct glsl_type *dest_type)
112 {
113 return nir_build_alu(&b->nb, nir_alu_op_for_opencl_opcode(b, opcode),
114 srcs[0], srcs[1], srcs[2], NULL);
115 }
116
117 static nir_ssa_def *
118 handle_special(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 nir_builder *nb = &b->nb;
123
124 switch (opcode) {
125 case OpenCLstd_SAbs_diff:
126 return nir_iabs_diff(nb, srcs[0], srcs[1]);
127 case OpenCLstd_UAbs_diff:
128 return nir_uabs_diff(nb, srcs[0], srcs[1]);
129 case OpenCLstd_Bitselect:
130 return nir_bitselect(nb, srcs[0], srcs[1], srcs[2]);
131 case OpenCLstd_SMad_hi:
132 return nir_imad_hi(nb, srcs[0], srcs[1], srcs[2]);
133 case OpenCLstd_UMad_hi:
134 return nir_umad_hi(nb, srcs[0], srcs[1], srcs[2]);
135 case OpenCLstd_SMul24:
136 return nir_imul24(nb, srcs[0], srcs[1]);
137 case OpenCLstd_UMul24:
138 return nir_umul24(nb, srcs[0], srcs[1]);
139 case OpenCLstd_SMad24:
140 return nir_imad24(nb, srcs[0], srcs[1], srcs[2]);
141 case OpenCLstd_UMad24:
142 return nir_umad24(nb, srcs[0], srcs[1], srcs[2]);
143 case OpenCLstd_FClamp:
144 return nir_fclamp(nb, srcs[0], srcs[1], srcs[2]);
145 case OpenCLstd_SClamp:
146 return nir_iclamp(nb, srcs[0], srcs[1], srcs[2]);
147 case OpenCLstd_UClamp:
148 return nir_uclamp(nb, srcs[0], srcs[1], srcs[2]);
149 case OpenCLstd_Copysign:
150 return nir_copysign(nb, srcs[0], srcs[1]);
151 case OpenCLstd_Cross:
152 if (glsl_get_components(dest_type) == 4)
153 return nir_cross4(nb, srcs[0], srcs[1]);
154 return nir_cross3(nb, srcs[0], srcs[1]);
155 case OpenCLstd_Degrees:
156 return nir_degrees(nb, srcs[0]);
157 case OpenCLstd_Fdim:
158 return nir_fdim(nb, srcs[0], srcs[1]);
159 case OpenCLstd_Distance:
160 return nir_distance(nb, srcs[0], srcs[1]);
161 case OpenCLstd_Fast_distance:
162 return nir_fast_distance(nb, srcs[0], srcs[1]);
163 case OpenCLstd_Fast_length:
164 return nir_fast_length(nb, srcs[0]);
165 case OpenCLstd_Fast_normalize:
166 return nir_fast_normalize(nb, srcs[0]);
167 case OpenCLstd_Length:
168 return nir_length(nb, srcs[0]);
169 case OpenCLstd_Mad:
170 return nir_fmad(nb, srcs[0], srcs[1], srcs[2]);
171 case OpenCLstd_Maxmag:
172 return nir_maxmag(nb, srcs[0], srcs[1]);
173 case OpenCLstd_Minmag:
174 return nir_minmag(nb, srcs[0], srcs[1]);
175 case OpenCLstd_Nan:
176 return nir_nan(nb, srcs[0]);
177 case OpenCLstd_Nextafter:
178 return nir_nextafter(nb, srcs[0], srcs[1]);
179 case OpenCLstd_Normalize:
180 return nir_normalize(nb, srcs[0]);
181 case OpenCLstd_Radians:
182 return nir_radians(nb, srcs[0]);
183 case OpenCLstd_Rotate:
184 return nir_rotate(nb, srcs[0], srcs[1]);
185 case OpenCLstd_Smoothstep:
186 return nir_smoothstep(nb, srcs[0], srcs[1], srcs[2]);
187 case OpenCLstd_Clz:
188 return nir_clz_u(nb, srcs[0]);
189 case OpenCLstd_Select:
190 return nir_select(nb, srcs[0], srcs[1], srcs[2]);
191 case OpenCLstd_Step:
192 return nir_sge(nb, srcs[1], srcs[0]);
193 case OpenCLstd_S_Upsample:
194 case OpenCLstd_U_Upsample:
195 return nir_upsample(nb, srcs[0], srcs[1]);
196 default:
197 vtn_fail("No NIR equivalent");
198 return NULL;
199 }
200 }
201
202 static void
203 _handle_v_load_store(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode,
204 const uint32_t *w, unsigned count, bool load)
205 {
206 struct vtn_type *type;
207 if (load)
208 type = vtn_value(b, w[1], vtn_value_type_type)->type;
209 else
210 type = vtn_untyped_value(b, w[5])->type;
211 unsigned a = load ? 0 : 1;
212
213 const struct glsl_type *dest_type = type->type;
214 unsigned components = glsl_get_vector_elements(dest_type);
215
216 nir_ssa_def *offset = vtn_ssa_value(b, w[5 + a])->def;
217 struct vtn_value *p = vtn_value(b, w[6 + a], vtn_value_type_pointer);
218
219 struct vtn_ssa_value *comps[NIR_MAX_VEC_COMPONENTS];
220 nir_ssa_def *ncomps[NIR_MAX_VEC_COMPONENTS];
221
222 nir_ssa_def *moffset = nir_imul_imm(&b->nb, offset, components);
223 nir_deref_instr *deref = vtn_pointer_to_deref(b, p->pointer);
224
225 for (int i = 0; i < components; i++) {
226 nir_ssa_def *coffset = nir_iadd_imm(&b->nb, moffset, i);
227 nir_deref_instr *arr_deref = nir_build_deref_ptr_as_array(&b->nb, deref, coffset);
228
229 if (load) {
230 comps[i] = vtn_local_load(b, arr_deref, p->type->access);
231 ncomps[i] = comps[i]->def;
232 } else {
233 struct vtn_ssa_value *ssa = vtn_create_ssa_value(b, glsl_scalar_type(glsl_get_base_type(dest_type)));
234 struct vtn_ssa_value *val = vtn_ssa_value(b, w[5]);
235 ssa->def = vtn_vector_extract(b, val->def, i);
236 vtn_local_store(b, ssa, arr_deref, p->type->access);
237 }
238 }
239 if (load) {
240 struct vtn_ssa_value *ssa = vtn_create_ssa_value(b, dest_type);
241 ssa->def = nir_vec(&b->nb, ncomps, components);
242 vtn_push_ssa(b, w[2], type, ssa);
243 }
244 }
245
246 static void
247 vtn_handle_opencl_vload(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode,
248 const uint32_t *w, unsigned count)
249 {
250 _handle_v_load_store(b, opcode, w, count, true);
251 }
252
253 static void
254 vtn_handle_opencl_vstore(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode,
255 const uint32_t *w, unsigned count)
256 {
257 _handle_v_load_store(b, opcode, w, count, false);
258 }
259
260 static nir_ssa_def *
261 handle_printf(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode,
262 unsigned num_srcs, nir_ssa_def **srcs,
263 const struct glsl_type *dest_type)
264 {
265 /* hahah, yeah, right.. */
266 return nir_imm_int(&b->nb, -1);
267 }
268
269 static nir_ssa_def *
270 handle_shuffle(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode, unsigned num_srcs,
271 nir_ssa_def **srcs, const struct glsl_type *dest_type)
272 {
273 struct nir_ssa_def *input = srcs[0];
274 struct nir_ssa_def *mask = srcs[1];
275
276 unsigned out_elems = glsl_get_vector_elements(dest_type);
277 nir_ssa_def *outres[NIR_MAX_VEC_COMPONENTS];
278 unsigned in_elems = input->num_components;
279 if (mask->bit_size != 32)
280 mask = nir_u2u32(&b->nb, mask);
281 mask = nir_iand(&b->nb, mask, nir_imm_intN_t(&b->nb, in_elems - 1, mask->bit_size));
282 for (unsigned i = 0; i < out_elems; i++)
283 outres[i] = nir_vector_extract(&b->nb, input, nir_channel(&b->nb, mask, i));
284
285 return nir_vec(&b->nb, outres, out_elems);
286 }
287
288 static nir_ssa_def *
289 handle_shuffle2(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 *input0 = srcs[0];
293 struct nir_ssa_def *input1 = srcs[1];
294 struct nir_ssa_def *mask = srcs[2];
295
296 unsigned out_elems = glsl_get_vector_elements(dest_type);
297 nir_ssa_def *outres[NIR_MAX_VEC_COMPONENTS];
298 unsigned in_elems = input0->num_components;
299 unsigned total_mask = 2 * in_elems - 1;
300 unsigned half_mask = in_elems - 1;
301 if (mask->bit_size != 32)
302 mask = nir_u2u32(&b->nb, mask);
303 mask = nir_iand(&b->nb, mask, nir_imm_intN_t(&b->nb, total_mask, mask->bit_size));
304 for (unsigned i = 0; i < out_elems; i++) {
305 nir_ssa_def *this_mask = nir_channel(&b->nb, mask, i);
306 nir_ssa_def *vmask = nir_iand(&b->nb, this_mask, nir_imm_intN_t(&b->nb, half_mask, mask->bit_size));
307 nir_ssa_def *val0 = nir_vector_extract(&b->nb, input0, vmask);
308 nir_ssa_def *val1 = nir_vector_extract(&b->nb, input1, vmask);
309 nir_ssa_def *sel = nir_ilt(&b->nb, this_mask, nir_imm_intN_t(&b->nb, in_elems, mask->bit_size));
310 outres[i] = nir_bcsel(&b->nb, sel, val0, val1);
311 }
312 return nir_vec(&b->nb, outres, out_elems);
313 }
314
315 bool
316 vtn_handle_opencl_instruction(struct vtn_builder *b, SpvOp ext_opcode,
317 const uint32_t *w, unsigned count)
318 {
319 switch ((enum OpenCLstd_Entrypoints)ext_opcode) {
320 case OpenCLstd_Fabs:
321 case OpenCLstd_SAbs:
322 case OpenCLstd_UAbs:
323 case OpenCLstd_SAdd_sat:
324 case OpenCLstd_UAdd_sat:
325 case OpenCLstd_Ceil:
326 case OpenCLstd_Cos:
327 case OpenCLstd_Exp2:
328 case OpenCLstd_Log2:
329 case OpenCLstd_Floor:
330 case OpenCLstd_Fma:
331 case OpenCLstd_Fmax:
332 case OpenCLstd_SHadd:
333 case OpenCLstd_UHadd:
334 case OpenCLstd_SMax:
335 case OpenCLstd_UMax:
336 case OpenCLstd_Fmin:
337 case OpenCLstd_SMin:
338 case OpenCLstd_UMin:
339 case OpenCLstd_Mix:
340 case OpenCLstd_Fmod:
341 case OpenCLstd_SMul_hi:
342 case OpenCLstd_UMul_hi:
343 case OpenCLstd_Popcount:
344 case OpenCLstd_Pow:
345 case OpenCLstd_Remainder:
346 case OpenCLstd_SRhadd:
347 case OpenCLstd_URhadd:
348 case OpenCLstd_Rsqrt:
349 case OpenCLstd_Sign:
350 case OpenCLstd_Sin:
351 case OpenCLstd_Sqrt:
352 case OpenCLstd_SSub_sat:
353 case OpenCLstd_USub_sat:
354 case OpenCLstd_Trunc:
355 handle_instr(b, ext_opcode, w, count, handle_alu);
356 return true;
357 case OpenCLstd_SAbs_diff:
358 case OpenCLstd_UAbs_diff:
359 case OpenCLstd_SMad_hi:
360 case OpenCLstd_UMad_hi:
361 case OpenCLstd_SMad24:
362 case OpenCLstd_UMad24:
363 case OpenCLstd_SMul24:
364 case OpenCLstd_UMul24:
365 case OpenCLstd_Bitselect:
366 case OpenCLstd_FClamp:
367 case OpenCLstd_SClamp:
368 case OpenCLstd_UClamp:
369 case OpenCLstd_Copysign:
370 case OpenCLstd_Cross:
371 case OpenCLstd_Degrees:
372 case OpenCLstd_Fdim:
373 case OpenCLstd_Distance:
374 case OpenCLstd_Fast_distance:
375 case OpenCLstd_Fast_length:
376 case OpenCLstd_Fast_normalize:
377 case OpenCLstd_Length:
378 case OpenCLstd_Mad:
379 case OpenCLstd_Maxmag:
380 case OpenCLstd_Minmag:
381 case OpenCLstd_Nan:
382 case OpenCLstd_Nextafter:
383 case OpenCLstd_Normalize:
384 case OpenCLstd_Radians:
385 case OpenCLstd_Rotate:
386 case OpenCLstd_Select:
387 case OpenCLstd_Step:
388 case OpenCLstd_Smoothstep:
389 case OpenCLstd_S_Upsample:
390 case OpenCLstd_U_Upsample:
391 handle_instr(b, ext_opcode, w, count, handle_special);
392 return true;
393 case OpenCLstd_Vloadn:
394 vtn_handle_opencl_vload(b, ext_opcode, w, count);
395 return true;
396 case OpenCLstd_Vstoren:
397 vtn_handle_opencl_vstore(b, ext_opcode, w, count);
398 return true;
399 case OpenCLstd_Shuffle:
400 handle_instr(b, ext_opcode, w, count, handle_shuffle);
401 return true;
402 case OpenCLstd_Shuffle2:
403 handle_instr(b, ext_opcode, w, count, handle_shuffle2);
404 return true;
405 case OpenCLstd_Printf:
406 handle_instr(b, ext_opcode, w, count, handle_printf);
407 return true;
408 case OpenCLstd_Prefetch:
409 /* TODO maybe add a nir instruction for this? */
410 return true;
411 default:
412 vtn_fail("unhandled opencl opc: %u\n", ext_opcode);
413 return false;
414 }
415 }