src/gallium/auxiliary/util/u_helpers.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2012 Marek Olšák <maraeo@gmail.com>
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sub license, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  19  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  21  * IN NO EVENT SHALL THE AUTHORS AND/OR THEIR SUPPLIERS BE LIABLE FOR
  22  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  23  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  24  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  **************************************************************************/
  27
  28 #include "util/u_cpu_detect.h"
  29 #include "util/u_helpers.h"
  30 #include "util/u_inlines.h"
  31 #include "util/u_upload_mgr.h"
  32 #include "util/u_thread.h"
  33 #include "util/os_time.h"
  34 #include <inttypes.h>
  35
  36 /**
  37  * This function is used to copy an array of pipe_vertex_buffer structures,
  38  * while properly referencing the pipe_vertex_buffer::buffer member.
  39  *
  40  * enabled_buffers is updated such that the bits corresponding to the indices
  41  * of disabled buffers are set to 0 and the enabled ones are set to 1.
  42  *
  43  * \sa util_copy_framebuffer_state
  44  */
  45 void util_set_vertex_buffers_mask(struct pipe_vertex_buffer *dst,
  46                                   uint32_t *enabled_buffers,
  47                                   const struct pipe_vertex_buffer *src,
  48                                   unsigned start_slot, unsigned count)
  49 {
  50    unsigned i;
  51    uint32_t bitmask = 0;
  52
  53    dst += start_slot;
  54
  55    *enabled_buffers &= ~u_bit_consecutive(start_slot, count);
  56
  57    if (src) {
  58       for (i = 0; i < count; i++) {
  59          if (src[i].buffer.resource)
  60             bitmask |= 1 << i;
  61
  62          pipe_vertex_buffer_unreference(&dst[i]);
  63
  64          if (!src[i].is_user_buffer)
  65             pipe_resource_reference(&dst[i].buffer.resource, src[i].buffer.resource);
  66       }
  67
  68       /* Copy over the other members of pipe_vertex_buffer. */
  69       memcpy(dst, src, count * sizeof(struct pipe_vertex_buffer));
  70
  71       *enabled_buffers |= bitmask << start_slot;
  72    }
  73    else {
  74       /* Unreference the buffers. */
  75       for (i = 0; i < count; i++)
  76          pipe_vertex_buffer_unreference(&dst[i]);
  77    }
  78 }
  79
  80 /**
  81  * Same as util_set_vertex_buffers_mask, but it only returns the number
  82  * of bound buffers.
  83  */
  84 void util_set_vertex_buffers_count(struct pipe_vertex_buffer *dst,
  85                                    unsigned *dst_count,
  86                                    const struct pipe_vertex_buffer *src,
  87                                    unsigned start_slot, unsigned count)
  88 {
  89    unsigned i;
  90    uint32_t enabled_buffers = 0;
  91
  92    for (i = 0; i < *dst_count; i++) {
  93       if (dst[i].buffer.resource)
  94          enabled_buffers |= (1ull << i);
  95    }
  96
  97    util_set_vertex_buffers_mask(dst, &enabled_buffers, src, start_slot,
  98                                 count);
  99
 100    *dst_count = util_last_bit(enabled_buffers);
 101 }
 102
 103 /**
 104  * This function is used to copy an array of pipe_shader_buffer structures,
 105  * while properly referencing the pipe_shader_buffer::buffer member.
 106  *
 107  * \sa util_set_vertex_buffer_mask
 108  */
 109 void util_set_shader_buffers_mask(struct pipe_shader_buffer *dst,
 110                                   uint32_t *enabled_buffers,
 111                                   const struct pipe_shader_buffer *src,
 112                                   unsigned start_slot, unsigned count)
 113 {
 114    unsigned i;
 115
 116    dst += start_slot;
 117
 118    if (src) {
 119       for (i = 0; i < count; i++) {
 120          pipe_resource_reference(&dst[i].buffer, src[i].buffer);
 121
 122          if (src[i].buffer)
 123             *enabled_buffers |= (1ull << (start_slot + i));
 124          else
 125             *enabled_buffers &= ~(1ull << (start_slot + i));
 126       }
 127
 128       /* Copy over the other members of pipe_shader_buffer. */
 129       memcpy(dst, src, count * sizeof(struct pipe_shader_buffer));
 130    }
 131    else {
 132       /* Unreference the buffers. */
 133       for (i = 0; i < count; i++)
 134          pipe_resource_reference(&dst[i].buffer, NULL);
 135
 136       *enabled_buffers &= ~(((1ull << count) - 1) << start_slot);
 137    }
 138 }
 139
 140 /**
 141  * Given a user index buffer, save the structure to "saved", and upload it.
 142  */
 143 bool
 144 util_upload_index_buffer(struct pipe_context *pipe,
 145                          const struct pipe_draw_info *info,
 146                          struct pipe_resource **out_buffer,
 147                          unsigned *out_offset, unsigned alignment)
 148 {
 149    unsigned start_offset = info->start * info->index_size;
 150
 151    u_upload_data(pipe->stream_uploader, start_offset,
 152                  info->count * info->index_size, alignment,
 153                  (char*)info->index.user + start_offset,
 154                  out_offset, out_buffer);
 155    u_upload_unmap(pipe->stream_uploader);
 156    *out_offset -= start_offset;
 157    return *out_buffer != NULL;
 158 }
 159
 160 /**
 161  * Called by MakeCurrent. Used to notify the driver that the application
 162  * thread may have been changed.
 163  *
 164  * The function pins the current thread and driver threads to a group of
 165  * CPU cores that share the same L3 cache. This is needed for good multi-
 166  * threading performance on AMD Zen CPUs.
 167  *
 168  * \param upper_thread  thread in gallium frontends that also needs to be
 169  *                      pinned.
 170  */
 171 void
 172 util_pin_driver_threads_to_random_L3(struct pipe_context *ctx,
 173                                      thrd_t *upper_thread)
 174 {
 175    /* If pinning has no effect, don't do anything. */
 176    if (util_cpu_caps.nr_cpus == util_cpu_caps.cores_per_L3)
 177       return;
 178
 179    unsigned num_L3_caches = util_cpu_caps.nr_cpus /
 180                             util_cpu_caps.cores_per_L3;
 181
 182    /* Get a semi-random number. */
 183    int64_t t = os_time_get_nano();
 184    unsigned cache = (t ^ (t >> 8) ^ (t >> 16)) % num_L3_caches;
 185
 186    /* Tell the driver to pin its threads to the selected L3 cache. */
 187    if (ctx->set_context_param) {
 188       ctx->set_context_param(ctx, PIPE_CONTEXT_PARAM_PIN_THREADS_TO_L3_CACHE,
 189                              cache);
 190    }
 191
 192    /* Do the same for the upper level thread if there is any (e.g. glthread) */
 193    if (upper_thread)
 194       util_pin_thread_to_L3(*upper_thread, cache, util_cpu_caps.cores_per_L3);
 195
 196    /* Optionally pin the application thread to the same L3 to get maximum
 197     * performance with glthread on AMD Zen. (this function is only called
 198     * with glthread) This is used to estimate and remove the overhead of
 199     * Infinity Fabric between L3 caches.
 200     */
 201 #if defined(HAVE_PTHREAD)
 202    if (debug_get_bool_option("pin_app_thread", false))
 203       util_pin_thread_to_L3(pthread_self(), cache, util_cpu_caps.cores_per_L3);
 204 #endif
 205 }
 206
 207 /* This is a helper for hardware bring-up. Don't remove. */
 208 struct pipe_query *
 209 util_begin_pipestat_query(struct pipe_context *ctx)
 210 {
 211    struct pipe_query *q =
 212       ctx->create_query(ctx, PIPE_QUERY_PIPELINE_STATISTICS, 0);
 213    if (!q)
 214       return NULL;
 215
 216    ctx->begin_query(ctx, q);
 217    return q;
 218 }
 219
 220 /* This is a helper for hardware bring-up. Don't remove. */
 221 void
 222 util_end_pipestat_query(struct pipe_context *ctx, struct pipe_query *q,
 223                         FILE *f)
 224 {
 225    static unsigned counter;
 226    struct pipe_query_data_pipeline_statistics stats;
 227
 228    ctx->end_query(ctx, q);
 229    ctx->get_query_result(ctx, q, true, (void*)&stats);
 230    ctx->destroy_query(ctx, q);
 231
 232    fprintf(f,
 233            "Draw call %u:\n"
 234            "    ia_vertices    = %"PRIu64"\n"
 235            "    ia_primitives  = %"PRIu64"\n"
 236            "    vs_invocations = %"PRIu64"\n"
 237            "    gs_invocations = %"PRIu64"\n"
 238            "    gs_primitives  = %"PRIu64"\n"
 239            "    c_invocations  = %"PRIu64"\n"
 240            "    c_primitives   = %"PRIu64"\n"
 241            "    ps_invocations = %"PRIu64"\n"
 242            "    hs_invocations = %"PRIu64"\n"
 243            "    ds_invocations = %"PRIu64"\n"
 244            "    cs_invocations = %"PRIu64"\n",
 245            (unsigned)p_atomic_inc_return(&counter),
 246            stats.ia_vertices,
 247            stats.ia_primitives,
 248            stats.vs_invocations,
 249            stats.gs_invocations,
 250            stats.gs_primitives,
 251            stats.c_invocations,
 252            stats.c_primitives,
 253            stats.ps_invocations,
 254            stats.hs_invocations,
 255            stats.ds_invocations,
 256            stats.cs_invocations);
 257 }
 258
 259 /* This is a helper for hardware bring-up. Don't remove. */
 260 void
 261 util_wait_for_idle(struct pipe_context *ctx)
 262 {
 263    struct pipe_fence_handle *fence = NULL;
 264
 265    ctx->flush(ctx, &fence, 0);
 266    ctx->screen->fence_finish(ctx->screen, NULL, fence, PIPE_TIMEOUT_INFINITE);
 267 }
 268
 269 void
 270 util_throttle_init(struct util_throttle *t, uint64_t max_mem_usage)
 271 {
 272    t->max_mem_usage = max_mem_usage;
 273 }
 274
 275 void
 276 util_throttle_deinit(struct pipe_screen *screen, struct util_throttle *t)
 277 {
 278    for (unsigned i = 0; i < ARRAY_SIZE(t->ring); i++)
 279       screen->fence_reference(screen, &t->ring[i].fence, NULL);
 280 }
 281
 282 static uint64_t
 283 util_get_throttle_total_memory_usage(struct util_throttle *t)
 284 {
 285    uint64_t total_usage = 0;
 286
 287    for (unsigned i = 0; i < ARRAY_SIZE(t->ring); i++)
 288       total_usage += t->ring[i].mem_usage;
 289    return total_usage;
 290 }
 291
 292 static void util_dump_throttle_ring(struct util_throttle *t)
 293 {
 294    printf("Throttle:\n");
 295    for (unsigned i = 0; i < ARRAY_SIZE(t->ring); i++) {
 296       printf("  ring[%u]: fence = %s, mem_usage = %"PRIu64"%s%s\n",
 297              i, t->ring[i].fence ? "yes" : " no",
 298              t->ring[i].mem_usage,
 299              t->flush_index == i ? " [flush]" : "",
 300              t->wait_index == i ? " [wait]" : "");
 301    }
 302 }
 303
 304 /**
 305  * Notify util_throttle that the next operation allocates memory.
 306  * util_throttle tracks memory usage and waits for fences until its tracked
 307  * memory usage decreases.
 308  *
 309  * Example:
 310  *   util_throttle_memory_usage(..., w*h*d*Bpp);
 311  *   TexSubImage(..., w, h, d, ...);
 312  *
 313  * This means that TexSubImage can't allocate more memory its maximum limit
 314  * set during initialization.
 315  */
 316 void
 317 util_throttle_memory_usage(struct pipe_context *pipe,
 318                            struct util_throttle *t, uint64_t memory_size)
 319 {
 320    (void)util_dump_throttle_ring; /* silence warning */
 321
 322    if (!t->max_mem_usage)
 323       return;
 324
 325    struct pipe_screen *screen = pipe->screen;
 326    struct pipe_fence_handle **fence = NULL;
 327    unsigned ring_size = ARRAY_SIZE(t->ring);
 328    uint64_t total = util_get_throttle_total_memory_usage(t);
 329
 330    /* If there is not enough memory, walk the list of fences and find
 331     * the latest one that we need to wait for.
 332     */
 333    while (t->wait_index != t->flush_index &&
 334           total && total + memory_size > t->max_mem_usage) {
 335       assert(t->ring[t->wait_index].fence);
 336
 337       /* Release an older fence if we need to wait for a newer one. */
 338       if (fence)
 339          screen->fence_reference(screen, fence, NULL);
 340
 341       fence = &t->ring[t->wait_index].fence;
 342       t->ring[t->wait_index].mem_usage = 0;
 343       t->wait_index = (t->wait_index + 1) % ring_size;
 344
 345       total = util_get_throttle_total_memory_usage(t);
 346    }
 347
 348    /* Wait for the fence to decrease memory usage. */
 349    if (fence) {
 350       screen->fence_finish(screen, pipe, *fence, PIPE_TIMEOUT_INFINITE);
 351       screen->fence_reference(screen, fence, NULL);
 352    }
 353
 354    /* Flush and get a fence if we've exhausted memory usage for the current
 355     * slot.
 356     */
 357    if (t->ring[t->flush_index].mem_usage &&
 358        t->ring[t->flush_index].mem_usage + memory_size >
 359        t->max_mem_usage / (ring_size / 2)) {
 360       struct pipe_fence_handle **fence =
 361          &t->ring[t->flush_index].fence;
 362
 363       /* Expect that the current flush slot doesn't have a fence yet. */
 364       assert(!*fence);
 365
 366       pipe->flush(pipe, fence, PIPE_FLUSH_ASYNC);
 367       t->flush_index = (t->flush_index + 1) % ring_size;
 368
 369       /* Vacate the next slot if it's occupied. This should be rare. */
 370       if (t->flush_index == t->wait_index) {
 371          struct pipe_fence_handle **fence =
 372             &t->ring[t->wait_index].fence;
 373
 374          t->ring[t->wait_index].mem_usage = 0;
 375          t->wait_index = (t->wait_index + 1) % ring_size;
 376
 377          assert(*fence);
 378          screen->fence_finish(screen, pipe, *fence, PIPE_TIMEOUT_INFINITE);
 379          screen->fence_reference(screen, fence, NULL);
 380       }
 381
 382       assert(!t->ring[t->flush_index].mem_usage);
 383       assert(!t->ring[t->flush_index].fence);
 384    }
 385
 386    t->ring[t->flush_index].mem_usage += memory_size;
 387 }