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_exp2: return nir_op_fexp2;
  88    case OpenCLstd_Native_log2: return nir_op_flog2;
  89    case OpenCLstd_SMul_hi: return nir_op_imul_high;
  90    case OpenCLstd_UMul_hi: return nir_op_umul_high;
  91    case OpenCLstd_Popcount: return nir_op_bit_count;
  92    case OpenCLstd_Pow: return nir_op_fpow;
  93    case OpenCLstd_Remainder: return nir_op_frem;
  94    case OpenCLstd_SRhadd: return nir_op_irhadd;
  95    case OpenCLstd_URhadd: return nir_op_urhadd;
  96    case OpenCLstd_Rsqrt: return nir_op_frsq;
  97    case OpenCLstd_Sign: return nir_op_fsign;
  98    case OpenCLstd_Sin: return nir_op_fsin;
  99    case OpenCLstd_Sqrt: return nir_op_fsqrt;
 100    case OpenCLstd_SSub_sat: return nir_op_isub_sat;
 101    case OpenCLstd_USub_sat: return nir_op_usub_sat;
 102    case OpenCLstd_Trunc: return nir_op_ftrunc;
 103    /* uhm... */
 104    case OpenCLstd_UAbs: return nir_op_mov;
 105    default:
 106       vtn_fail("No NIR equivalent");
 107    }
 108 }
 109
 110 static nir_ssa_def *
 111 handle_alu(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode,
 112            unsigned num_srcs, nir_ssa_def **srcs,
 113            const struct glsl_type *dest_type)
 114 {
 115    return nir_build_alu(&b->nb, nir_alu_op_for_opencl_opcode(b, opcode),
 116                         srcs[0], srcs[1], srcs[2], NULL);
 117 }
 118
 119 static nir_ssa_def *
 120 handle_special(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode,
 121                unsigned num_srcs, nir_ssa_def **srcs,
 122                const struct glsl_type *dest_type)
 123 {
 124    nir_builder *nb = &b->nb;
 125
 126    switch (opcode) {
 127    case OpenCLstd_SAbs_diff:
 128       return nir_iabs_diff(nb, srcs[0], srcs[1]);
 129    case OpenCLstd_UAbs_diff:
 130       return nir_uabs_diff(nb, srcs[0], srcs[1]);
 131    case OpenCLstd_Bitselect:
 132       return nir_bitselect(nb, srcs[0], srcs[1], srcs[2]);
 133    case OpenCLstd_SMad_hi:
 134       return nir_imad_hi(nb, srcs[0], srcs[1], srcs[2]);
 135    case OpenCLstd_UMad_hi:
 136       return nir_umad_hi(nb, srcs[0], srcs[1], srcs[2]);
 137    case OpenCLstd_SMul24:
 138       return nir_imul24(nb, srcs[0], srcs[1]);
 139    case OpenCLstd_UMul24:
 140       return nir_umul24(nb, srcs[0], srcs[1]);
 141    case OpenCLstd_SMad24:
 142       return nir_imad24(nb, srcs[0], srcs[1], srcs[2]);
 143    case OpenCLstd_UMad24:
 144       return nir_umad24(nb, srcs[0], srcs[1], srcs[2]);
 145    case OpenCLstd_FClamp:
 146       return nir_fclamp(nb, srcs[0], srcs[1], srcs[2]);
 147    case OpenCLstd_SClamp:
 148       return nir_iclamp(nb, srcs[0], srcs[1], srcs[2]);
 149    case OpenCLstd_UClamp:
 150       return nir_uclamp(nb, srcs[0], srcs[1], srcs[2]);
 151    case OpenCLstd_Copysign:
 152       return nir_copysign(nb, srcs[0], srcs[1]);
 153    case OpenCLstd_Cross:
 154       if (glsl_get_components(dest_type) == 4)
 155          return nir_cross4(nb, srcs[0], srcs[1]);
 156       return nir_cross3(nb, srcs[0], srcs[1]);
 157    case OpenCLstd_Degrees:
 158       return nir_degrees(nb, srcs[0]);
 159    case OpenCLstd_Fdim:
 160       return nir_fdim(nb, srcs[0], srcs[1]);
 161    case OpenCLstd_Distance:
 162       return nir_distance(nb, srcs[0], srcs[1]);
 163    case OpenCLstd_Fast_distance:
 164       return nir_fast_distance(nb, srcs[0], srcs[1]);
 165    case OpenCLstd_Fast_length:
 166       return nir_fast_length(nb, srcs[0]);
 167    case OpenCLstd_Fast_normalize:
 168       return nir_fast_normalize(nb, srcs[0]);
 169    case OpenCLstd_Length:
 170       return nir_length(nb, srcs[0]);
 171    case OpenCLstd_Mad:
 172       return nir_fmad(nb, srcs[0], srcs[1], srcs[2]);
 173    case OpenCLstd_Maxmag:
 174       return nir_maxmag(nb, srcs[0], srcs[1]);
 175    case OpenCLstd_Minmag:
 176       return nir_minmag(nb, srcs[0], srcs[1]);
 177    case OpenCLstd_Nan:
 178       return nir_nan(nb, srcs[0]);
 179    case OpenCLstd_Nextafter:
 180       return nir_nextafter(nb, srcs[0], srcs[1]);
 181    case OpenCLstd_Normalize:
 182       return nir_normalize(nb, srcs[0]);
 183    case OpenCLstd_Radians:
 184       return nir_radians(nb, srcs[0]);
 185    case OpenCLstd_Rotate:
 186       return nir_rotate(nb, srcs[0], srcs[1]);
 187    case OpenCLstd_Smoothstep:
 188       return nir_smoothstep(nb, srcs[0], srcs[1], srcs[2]);
 189    case OpenCLstd_Clz:
 190       return nir_clz_u(nb, srcs[0]);
 191    case OpenCLstd_Select:
 192       return nir_select(nb, srcs[0], srcs[1], srcs[2]);
 193    case OpenCLstd_Step:
 194       return nir_sge(nb, srcs[1], srcs[0]);
 195    case OpenCLstd_S_Upsample:
 196    case OpenCLstd_U_Upsample:
 197       return nir_upsample(nb, srcs[0], srcs[1]);
 198    case OpenCLstd_Native_exp:
 199       return nir_fexp(nb, srcs[0]);
 200    case OpenCLstd_Native_exp10:
 201       return nir_fexp2(nb, nir_fmul_imm(nb, srcs[0], log(10) / log(2)));
 202    case OpenCLstd_Native_log:
 203       return nir_flog(nb, srcs[0]);
 204    case OpenCLstd_Native_log10:
 205       return nir_fmul_imm(nb, nir_flog2(nb, srcs[0]), log(2) / log(10));
 206    default:
 207       vtn_fail("No NIR equivalent");
 208       return NULL;
 209    }
 210 }
 211
 212 static void
 213 _handle_v_load_store(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode,
 214                      const uint32_t *w, unsigned count, bool load)
 215 {
 216    struct vtn_type *type;
 217    if (load)
 218       type = vtn_value(b, w[1], vtn_value_type_type)->type;
 219    else
 220       type = vtn_untyped_value(b, w[5])->type;
 221    unsigned a = load ? 0 : 1;
 222
 223    const struct glsl_type *dest_type = type->type;
 224    unsigned components = glsl_get_vector_elements(dest_type);
 225
 226    nir_ssa_def *offset = vtn_ssa_value(b, w[5 + a])->def;
 227    struct vtn_value *p = vtn_value(b, w[6 + a], vtn_value_type_pointer);
 228
 229    struct vtn_ssa_value *comps[NIR_MAX_VEC_COMPONENTS];
 230    nir_ssa_def *ncomps[NIR_MAX_VEC_COMPONENTS];
 231
 232    nir_ssa_def *moffset = nir_imul_imm(&b->nb, offset, components);
 233    nir_deref_instr *deref = vtn_pointer_to_deref(b, p->pointer);
 234
 235    for (int i = 0; i < components; i++) {
 236       nir_ssa_def *coffset = nir_iadd_imm(&b->nb, moffset, i);
 237       nir_deref_instr *arr_deref = nir_build_deref_ptr_as_array(&b->nb, deref, coffset);
 238
 239       if (load) {
 240          comps[i] = vtn_local_load(b, arr_deref, p->type->access);
 241          ncomps[i] = comps[i]->def;
 242       } else {
 243          struct vtn_ssa_value *ssa = vtn_create_ssa_value(b, glsl_scalar_type(glsl_get_base_type(dest_type)));
 244          struct vtn_ssa_value *val = vtn_ssa_value(b, w[5]);
 245          ssa->def = vtn_vector_extract(b, val->def, i);
 246          vtn_local_store(b, ssa, arr_deref, p->type->access);
 247       }
 248    }
 249    if (load) {
 250       struct vtn_ssa_value *ssa = vtn_create_ssa_value(b, dest_type);
 251       ssa->def = nir_vec(&b->nb, ncomps, components);
 252       vtn_push_ssa(b, w[2], type, ssa);
 253    }
 254 }
 255
 256 static void
 257 vtn_handle_opencl_vload(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode,
 258                         const uint32_t *w, unsigned count)
 259 {
 260    _handle_v_load_store(b, opcode, w, count, true);
 261 }
 262
 263 static void
 264 vtn_handle_opencl_vstore(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode,
 265                          const uint32_t *w, unsigned count)
 266 {
 267    _handle_v_load_store(b, opcode, w, count, false);
 268 }
 269
 270 static nir_ssa_def *
 271 handle_printf(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode,
 272               unsigned num_srcs, nir_ssa_def **srcs,
 273               const struct glsl_type *dest_type)
 274 {
 275    /* hahah, yeah, right.. */
 276    return nir_imm_int(&b->nb, -1);
 277 }
 278
 279 static nir_ssa_def *
 280 handle_shuffle(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode, unsigned num_srcs,
 281                nir_ssa_def **srcs, const struct glsl_type *dest_type)
 282 {
 283    struct nir_ssa_def *input = srcs[0];
 284    struct nir_ssa_def *mask = srcs[1];
 285
 286    unsigned out_elems = glsl_get_vector_elements(dest_type);
 287    nir_ssa_def *outres[NIR_MAX_VEC_COMPONENTS];
 288    unsigned in_elems = input->num_components;
 289    if (mask->bit_size != 32)
 290       mask = nir_u2u32(&b->nb, mask);
 291    mask = nir_iand(&b->nb, mask, nir_imm_intN_t(&b->nb, in_elems - 1, mask->bit_size));
 292    for (unsigned i = 0; i < out_elems; i++)
 293       outres[i] = nir_vector_extract(&b->nb, input, nir_channel(&b->nb, mask, i));
 294
 295    return nir_vec(&b->nb, outres, out_elems);
 296 }
 297
 298 static nir_ssa_def *
 299 handle_shuffle2(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode, unsigned num_srcs,
 300                 nir_ssa_def **srcs, const struct glsl_type *dest_type)
 301 {
 302    struct nir_ssa_def *input0 = srcs[0];
 303    struct nir_ssa_def *input1 = srcs[1];
 304    struct nir_ssa_def *mask = srcs[2];
 305
 306    unsigned out_elems = glsl_get_vector_elements(dest_type);
 307    nir_ssa_def *outres[NIR_MAX_VEC_COMPONENTS];
 308    unsigned in_elems = input0->num_components;
 309    unsigned total_mask = 2 * in_elems - 1;
 310    unsigned half_mask = in_elems - 1;
 311    if (mask->bit_size != 32)
 312       mask = nir_u2u32(&b->nb, mask);
 313    mask = nir_iand(&b->nb, mask, nir_imm_intN_t(&b->nb, total_mask, mask->bit_size));
 314    for (unsigned i = 0; i < out_elems; i++) {
 315       nir_ssa_def *this_mask = nir_channel(&b->nb, mask, i);
 316       nir_ssa_def *vmask = nir_iand(&b->nb, this_mask, nir_imm_intN_t(&b->nb, half_mask, mask->bit_size));
 317       nir_ssa_def *val0 = nir_vector_extract(&b->nb, input0, vmask);
 318       nir_ssa_def *val1 = nir_vector_extract(&b->nb, input1, vmask);
 319       nir_ssa_def *sel = nir_ilt(&b->nb, this_mask, nir_imm_intN_t(&b->nb, in_elems, mask->bit_size));
 320       outres[i] = nir_bcsel(&b->nb, sel, val0, val1);
 321    }
 322    return nir_vec(&b->nb, outres, out_elems);
 323 }
 324
 325 bool
 326 vtn_handle_opencl_instruction(struct vtn_builder *b, SpvOp ext_opcode,
 327                               const uint32_t *w, unsigned count)
 328 {
 329    enum OpenCLstd_Entrypoints cl_opcode = (enum OpenCLstd_Entrypoints) ext_opcode;
 330
 331    switch (cl_opcode) {
 332    case OpenCLstd_Fabs:
 333    case OpenCLstd_SAbs:
 334    case OpenCLstd_UAbs:
 335    case OpenCLstd_SAdd_sat:
 336    case OpenCLstd_UAdd_sat:
 337    case OpenCLstd_Ceil:
 338    case OpenCLstd_Cos:
 339    case OpenCLstd_Exp2:
 340    case OpenCLstd_Log2:
 341    case OpenCLstd_Floor:
 342    case OpenCLstd_Fma:
 343    case OpenCLstd_Fmax:
 344    case OpenCLstd_SHadd:
 345    case OpenCLstd_UHadd:
 346    case OpenCLstd_SMax:
 347    case OpenCLstd_UMax:
 348    case OpenCLstd_Fmin:
 349    case OpenCLstd_SMin:
 350    case OpenCLstd_UMin:
 351    case OpenCLstd_Mix:
 352    case OpenCLstd_Native_exp2:
 353    case OpenCLstd_Native_log2:
 354    case OpenCLstd_Fmod:
 355    case OpenCLstd_SMul_hi:
 356    case OpenCLstd_UMul_hi:
 357    case OpenCLstd_Popcount:
 358    case OpenCLstd_Pow:
 359    case OpenCLstd_Remainder:
 360    case OpenCLstd_SRhadd:
 361    case OpenCLstd_URhadd:
 362    case OpenCLstd_Rsqrt:
 363    case OpenCLstd_Sign:
 364    case OpenCLstd_Sin:
 365    case OpenCLstd_Sqrt:
 366    case OpenCLstd_SSub_sat:
 367    case OpenCLstd_USub_sat:
 368    case OpenCLstd_Trunc:
 369       handle_instr(b, cl_opcode, w, count, handle_alu);
 370       return true;
 371    case OpenCLstd_SAbs_diff:
 372    case OpenCLstd_UAbs_diff:
 373    case OpenCLstd_SMad_hi:
 374    case OpenCLstd_UMad_hi:
 375    case OpenCLstd_SMad24:
 376    case OpenCLstd_UMad24:
 377    case OpenCLstd_SMul24:
 378    case OpenCLstd_UMul24:
 379    case OpenCLstd_Bitselect:
 380    case OpenCLstd_FClamp:
 381    case OpenCLstd_SClamp:
 382    case OpenCLstd_UClamp:
 383    case OpenCLstd_Copysign:
 384    case OpenCLstd_Cross:
 385    case OpenCLstd_Degrees:
 386    case OpenCLstd_Fdim:
 387    case OpenCLstd_Distance:
 388    case OpenCLstd_Fast_distance:
 389    case OpenCLstd_Fast_length:
 390    case OpenCLstd_Fast_normalize:
 391    case OpenCLstd_Length:
 392    case OpenCLstd_Mad:
 393    case OpenCLstd_Maxmag:
 394    case OpenCLstd_Minmag:
 395    case OpenCLstd_Nan:
 396    case OpenCLstd_Nextafter:
 397    case OpenCLstd_Normalize:
 398    case OpenCLstd_Radians:
 399    case OpenCLstd_Rotate:
 400    case OpenCLstd_Select:
 401    case OpenCLstd_Step:
 402    case OpenCLstd_Smoothstep:
 403    case OpenCLstd_S_Upsample:
 404    case OpenCLstd_U_Upsample:
 405    case OpenCLstd_Clz:
 406    case OpenCLstd_Native_exp:
 407    case OpenCLstd_Native_exp10:
 408    case OpenCLstd_Native_log:
 409    case OpenCLstd_Native_log10:
 410       handle_instr(b, cl_opcode, w, count, handle_special);
 411       return true;
 412    case OpenCLstd_Vloadn:
 413       vtn_handle_opencl_vload(b, cl_opcode, w, count);
 414       return true;
 415    case OpenCLstd_Vstoren:
 416       vtn_handle_opencl_vstore(b, cl_opcode, w, count);
 417       return true;
 418    case OpenCLstd_Shuffle:
 419       handle_instr(b, cl_opcode, w, count, handle_shuffle);
 420       return true;
 421    case OpenCLstd_Shuffle2:
 422       handle_instr(b, cl_opcode, w, count, handle_shuffle2);
 423       return true;
 424    case OpenCLstd_Printf:
 425       handle_instr(b, cl_opcode, w, count, handle_printf);
 426       return true;
 427    case OpenCLstd_Prefetch:
 428       /* TODO maybe add a nir instruction for this? */
 429       return true;
 430    default:
 431       vtn_fail("unhandled opencl opc: %u\n", ext_opcode);
 432       return false;
 433    }
 434 }