src/gallium/state_trackers/clover/core/kernel.cpp

   1 //
   2 // Copyright 2012 Francisco Jerez
   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 shall be included in
  12 // all copies or substantial portions of the Software.
  13 //
  14 // THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15 // IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16 // FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17 // THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18 // OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19 // ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20 // OTHER DEALINGS IN THE SOFTWARE.
  21 //
  22
  23 #include "core/kernel.hpp"
  24 #include "core/resource.hpp"
  25 #include "util/u_math.h"
  26 #include "pipe/p_context.h"
  27
  28 using namespace clover;
  29
  30 kernel::kernel(program &prog, const std::string &name,
  31                const std::vector<module::argument> &margs) :
  32    prog(prog), _name(name), exec(*this) {
  33    for (auto &marg : margs) {
  34       if (marg.type == module::argument::scalar)
  35          _args.emplace_back(new scalar_argument(marg.size));
  36       else if (marg.type == module::argument::global)
  37          _args.emplace_back(new global_argument);
  38       else if (marg.type == module::argument::local)
  39          _args.emplace_back(new local_argument);
  40       else if (marg.type == module::argument::constant)
  41          _args.emplace_back(new constant_argument);
  42       else if (marg.type == module::argument::image2d_rd ||
  43                marg.type == module::argument::image3d_rd)
  44          _args.emplace_back(new image_rd_argument);
  45       else if (marg.type == module::argument::image2d_wr ||
  46                marg.type == module::argument::image3d_wr)
  47          _args.emplace_back(new image_wr_argument);
  48       else if (marg.type == module::argument::sampler)
  49          _args.emplace_back(new sampler_argument);
  50       else
  51          throw error(CL_INVALID_KERNEL_DEFINITION);
  52    }
  53 }
  54
  55 template<typename V>
  56 static inline std::vector<uint>
  57 pad_vector(command_queue &q, const V &v, uint x) {
  58    std::vector<uint> w { v.begin(), v.end() };
  59    w.resize(q.dev.max_block_size().size(), x);
  60    return w;
  61 }
  62
  63 void
  64 kernel::launch(command_queue &q,
  65                const std::vector<size_t> &grid_offset,
  66                const std::vector<size_t> &grid_size,
  67                const std::vector<size_t> &block_size) {
  68    const auto m = prog.binary(q.dev);
  69    const auto reduced_grid_size =
  70       map(divides(), grid_size, block_size);
  71    void *st = exec.bind(&q);
  72
  73    // The handles are created during exec_context::bind(), so we need make
  74    // sure to call exec_context::bind() before retrieving them.
  75    std::vector<uint32_t *> g_handles = map([&](size_t h) {
  76          return (uint32_t *)&exec.input[h];
  77       }, exec.g_handles);
  78
  79    q.pipe->bind_compute_state(q.pipe, st);
  80    q.pipe->bind_sampler_states(q.pipe, PIPE_SHADER_COMPUTE,
  81                                0, exec.samplers.size(),
  82                                exec.samplers.data());
  83
  84    q.pipe->set_sampler_views(q.pipe, PIPE_SHADER_COMPUTE, 0,
  85                              exec.sviews.size(), exec.sviews.data());
  86    q.pipe->set_compute_resources(q.pipe, 0, exec.resources.size(),
  87                                  exec.resources.data());
  88    q.pipe->set_global_binding(q.pipe, 0, exec.g_buffers.size(),
  89                               exec.g_buffers.data(), g_handles.data());
  90
  91    q.pipe->launch_grid(q.pipe,
  92                        pad_vector(q, block_size, 1).data(),
  93                        pad_vector(q, reduced_grid_size, 1).data(),
  94                        find(name_equals(_name), m.syms).offset,
  95                        exec.input.data());
  96
  97    q.pipe->set_global_binding(q.pipe, 0, exec.g_buffers.size(), NULL, NULL);
  98    q.pipe->set_compute_resources(q.pipe, 0, exec.resources.size(), NULL);
  99    q.pipe->set_sampler_views(q.pipe, PIPE_SHADER_COMPUTE, 0,
 100                              exec.sviews.size(), NULL);
 101    q.pipe->bind_sampler_states(q.pipe, PIPE_SHADER_COMPUTE, 0,
 102                                exec.samplers.size(), NULL);
 103    exec.unbind();
 104 }
 105
 106 size_t
 107 kernel::mem_local() const {
 108    size_t sz = 0;
 109
 110    for (auto &arg : args()) {
 111       if (dynamic_cast<local_argument *>(&arg))
 112          sz += arg.storage();
 113    }
 114
 115    return sz;
 116 }
 117
 118 size_t
 119 kernel::mem_private() const {
 120    return 0;
 121 }
 122
 123 const std::string &
 124 kernel::name() const {
 125    return _name;
 126 }
 127
 128 std::vector<size_t>
 129 kernel::block_size() const {
 130    return { 0, 0, 0 };
 131 }
 132
 133 kernel::argument_range
 134 kernel::args() {
 135    return map(derefs(), _args);
 136 }
 137
 138 kernel::const_argument_range
 139 kernel::args() const {
 140    return map(derefs(), _args);
 141 }
 142
 143 const module &
 144 kernel::module(const command_queue &q) const {
 145    return prog.binary(q.dev);
 146 }
 147
 148 kernel::exec_context::exec_context(kernel &kern) :
 149    kern(kern), q(NULL), mem_local(0), st(NULL), cs() {
 150 }
 151
 152 kernel::exec_context::~exec_context() {
 153    if (st)
 154       q->pipe->delete_compute_state(q->pipe, st);
 155 }
 156
 157 void *
 158 kernel::exec_context::bind(command_queue *_q) {
 159    std::swap(q, _q);
 160
 161    // Bind kernel arguments.
 162    auto &m = kern.prog.binary(q->dev);
 163    auto margs = find(name_equals(kern.name()), m.syms).args;
 164    auto msec = find(type_equals(module::section::text), m.secs);
 165
 166    for_each([=](kernel::argument &karg, const module::argument &marg) {
 167                karg.bind(*this, marg);
 168             }, kern.args(), margs);
 169
 170    // Create a new compute state if anything changed.
 171    if (!st || q != _q ||
 172        cs.req_local_mem != mem_local ||
 173        cs.req_input_mem != input.size()) {
 174       if (st)
 175          _q->pipe->delete_compute_state(_q->pipe, st);
 176
 177       cs.prog = msec.data.begin();
 178       cs.req_local_mem = mem_local;
 179       cs.req_input_mem = input.size();
 180       st = q->pipe->create_compute_state(q->pipe, &cs);
 181    }
 182
 183    return st;
 184 }
 185
 186 void
 187 kernel::exec_context::unbind() {
 188    for (auto &arg : kern.args())
 189       arg.unbind(*this);
 190
 191    input.clear();
 192    samplers.clear();
 193    sviews.clear();
 194    resources.clear();
 195    g_buffers.clear();
 196    g_handles.clear();
 197    mem_local = 0;
 198 }
 199
 200 namespace {
 201    template<typename T>
 202    std::vector<uint8_t>
 203    bytes(const T& x) {
 204       return { (uint8_t *)&x, (uint8_t *)&x + sizeof(x) };
 205    }
 206
 207    ///
 208    /// Transform buffer \a v from the native byte order into the byte
 209    /// order specified by \a e.
 210    ///
 211    template<typename T>
 212    void
 213    byteswap(T &v, pipe_endian e) {
 214       if (PIPE_ENDIAN_NATIVE != e)
 215          std::reverse(v.begin(), v.end());
 216    }
 217
 218    ///
 219    /// Pad buffer \a v to the next multiple of \a n.
 220    ///
 221    template<typename T>
 222    void
 223    align(T &v, size_t n) {
 224       v.resize(util_align_npot(v.size(), n));
 225    }
 226
 227    bool
 228    msb(const std::vector<uint8_t> &s) {
 229       if (PIPE_ENDIAN_NATIVE == PIPE_ENDIAN_LITTLE)
 230          return s.back() & 0x80;
 231       else
 232          return s.front() & 0x80;
 233    }
 234
 235    ///
 236    /// Resize buffer \a v to size \a n using sign or zero extension
 237    /// according to \a ext.
 238    ///
 239    template<typename T>
 240    void
 241    extend(T &v, enum module::argument::ext_type ext, size_t n) {
 242       const size_t m = std::min(v.size(), n);
 243       const bool sign_ext = (ext == module::argument::sign_ext);
 244       const uint8_t fill = (sign_ext && msb(v) ? ~0 : 0);
 245       T w(n, fill);
 246
 247       if (PIPE_ENDIAN_NATIVE == PIPE_ENDIAN_LITTLE)
 248          std::copy_n(v.begin(), m, w.begin());
 249       else
 250          std::copy_n(v.end() - m, m, w.end() - m);
 251
 252       std::swap(v, w);
 253    }
 254
 255    ///
 256    /// Append buffer \a w to \a v.
 257    ///
 258    template<typename T>
 259    void
 260    insert(T &v, const T &w) {
 261       v.insert(v.end(), w.begin(), w.end());
 262    }
 263
 264    ///
 265    /// Append \a n elements to the end of buffer \a v.
 266    ///
 267    template<typename T>
 268    size_t
 269    allocate(T &v, size_t n) {
 270       size_t pos = v.size();
 271       v.resize(pos + n);
 272       return pos;
 273    }
 274 }
 275
 276 kernel::argument::argument() : _set(false) {
 277 }
 278
 279 bool
 280 kernel::argument::set() const {
 281    return _set;
 282 }
 283
 284 size_t
 285 kernel::argument::storage() const {
 286    return 0;
 287 }
 288
 289 kernel::scalar_argument::scalar_argument(size_t size) : size(size) {
 290 }
 291
 292 void
 293 kernel::scalar_argument::set(size_t size, const void *value) {
 294    if (size != this->size)
 295       throw error(CL_INVALID_ARG_SIZE);
 296
 297    v = { (uint8_t *)value, (uint8_t *)value + size };
 298    _set = true;
 299 }
 300
 301 void
 302 kernel::scalar_argument::bind(exec_context &ctx,
 303                               const module::argument &marg) {
 304    auto w = v;
 305
 306    extend(w, marg.ext_type, marg.target_size);
 307    byteswap(w, ctx.q->dev.endianness());
 308    align(ctx.input, marg.target_align);
 309    insert(ctx.input, w);
 310 }
 311
 312 void
 313 kernel::scalar_argument::unbind(exec_context &ctx) {
 314 }
 315
 316 void
 317 kernel::global_argument::set(size_t size, const void *value) {
 318    if (size != sizeof(cl_mem))
 319       throw error(CL_INVALID_ARG_SIZE);
 320
 321    buf = &obj<buffer>(*(cl_mem *)value);
 322    _set = true;
 323 }
 324
 325 void
 326 kernel::global_argument::bind(exec_context &ctx,
 327                               const module::argument &marg) {
 328    align(ctx.input, marg.target_align);
 329    ctx.g_handles.push_back(allocate(ctx.input, marg.target_size));
 330    ctx.g_buffers.push_back(buf->resource(*ctx.q).pipe);
 331 }
 332
 333 void
 334 kernel::global_argument::unbind(exec_context &ctx) {
 335 }
 336
 337 size_t
 338 kernel::local_argument::storage() const {
 339    return _storage;
 340 }
 341
 342 void
 343 kernel::local_argument::set(size_t size, const void *value) {
 344    if (value)
 345       throw error(CL_INVALID_ARG_VALUE);
 346
 347    _storage = size;
 348    _set = true;
 349 }
 350
 351 void
 352 kernel::local_argument::bind(exec_context &ctx,
 353                              const module::argument &marg) {
 354    auto v = bytes(ctx.mem_local);
 355
 356    extend(v, module::argument::zero_ext, marg.target_size);
 357    byteswap(v, ctx.q->dev.endianness());
 358    align(ctx.input, marg.target_align);
 359    insert(ctx.input, v);
 360
 361    ctx.mem_local += _storage;
 362 }
 363
 364 void
 365 kernel::local_argument::unbind(exec_context &ctx) {
 366 }
 367
 368 void
 369 kernel::constant_argument::set(size_t size, const void *value) {
 370    if (size != sizeof(cl_mem))
 371       throw error(CL_INVALID_ARG_SIZE);
 372
 373    buf = &obj<buffer>(*(cl_mem *)value);
 374    _set = true;
 375 }
 376
 377 void
 378 kernel::constant_argument::bind(exec_context &ctx,
 379                                 const module::argument &marg) {
 380    auto v = bytes(ctx.resources.size() << 24);
 381
 382    extend(v, module::argument::zero_ext, marg.target_size);
 383    byteswap(v, ctx.q->dev.endianness());
 384    align(ctx.input, marg.target_align);
 385    insert(ctx.input, v);
 386
 387    st = buf->resource(*ctx.q).bind_surface(*ctx.q, false);
 388    ctx.resources.push_back(st);
 389 }
 390
 391 void
 392 kernel::constant_argument::unbind(exec_context &ctx) {
 393    buf->resource(*ctx.q).unbind_surface(*ctx.q, st);
 394 }
 395
 396 void
 397 kernel::image_rd_argument::set(size_t size, const void *value) {
 398    if (size != sizeof(cl_mem))
 399       throw error(CL_INVALID_ARG_SIZE);
 400
 401    img = &obj<image>(*(cl_mem *)value);
 402    _set = true;
 403 }
 404
 405 void
 406 kernel::image_rd_argument::bind(exec_context &ctx,
 407                                 const module::argument &marg) {
 408    auto v = bytes(ctx.sviews.size());
 409
 410    extend(v, module::argument::zero_ext, marg.target_size);
 411    byteswap(v, ctx.q->dev.endianness());
 412    align(ctx.input, marg.target_align);
 413    insert(ctx.input, v);
 414
 415    st = img->resource(*ctx.q).bind_sampler_view(*ctx.q);
 416    ctx.sviews.push_back(st);
 417 }
 418
 419 void
 420 kernel::image_rd_argument::unbind(exec_context &ctx) {
 421    img->resource(*ctx.q).unbind_sampler_view(*ctx.q, st);
 422 }
 423
 424 void
 425 kernel::image_wr_argument::set(size_t size, const void *value) {
 426    if (size != sizeof(cl_mem))
 427       throw error(CL_INVALID_ARG_SIZE);
 428
 429    img = &obj<image>(*(cl_mem *)value);
 430    _set = true;
 431 }
 432
 433 void
 434 kernel::image_wr_argument::bind(exec_context &ctx,
 435                                 const module::argument &marg) {
 436    auto v = bytes(ctx.resources.size());
 437
 438    extend(v, module::argument::zero_ext, marg.target_size);
 439    byteswap(v, ctx.q->dev.endianness());
 440    align(ctx.input, marg.target_align);
 441    insert(ctx.input, v);
 442
 443    st = img->resource(*ctx.q).bind_surface(*ctx.q, true);
 444    ctx.resources.push_back(st);
 445 }
 446
 447 void
 448 kernel::image_wr_argument::unbind(exec_context &ctx) {
 449    img->resource(*ctx.q).unbind_surface(*ctx.q, st);
 450 }
 451
 452 void
 453 kernel::sampler_argument::set(size_t size, const void *value) {
 454    if (size != sizeof(cl_sampler))
 455       throw error(CL_INVALID_ARG_SIZE);
 456
 457    s = &obj(*(cl_sampler *)value);
 458    _set = true;
 459 }
 460
 461 void
 462 kernel::sampler_argument::bind(exec_context &ctx,
 463                                const module::argument &marg) {
 464    st = s->bind(*ctx.q);
 465    ctx.samplers.push_back(st);
 466 }
 467
 468 void
 469 kernel::sampler_argument::unbind(exec_context &ctx) {
 470    s->unbind(*ctx.q, st);
 471 }