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