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    if (src) {
  56       for (i = 0; i < count; i++) {
  57          if (src[i].buffer.resource)
  58             bitmask |= 1 << i;
  59
  60          pipe_vertex_buffer_unreference(&dst[i]);
  61
  62          if (!src[i].is_user_buffer)
  63             pipe_resource_reference(&dst[i].buffer.resource, src[i].buffer.resource);
  64       }
  65
  66       /* Copy over the other members of pipe_vertex_buffer. */
  67       memcpy(dst, src, count * sizeof(struct pipe_vertex_buffer));
  68
  69       *enabled_buffers &= ~(((1ull << count) - 1) << start_slot);
  70       *enabled_buffers |= bitmask << start_slot;
  71    }
  72    else {
  73       /* Unreference the buffers. */
  74       for (i = 0; i < count; i++)
  75          pipe_vertex_buffer_unreference(&dst[i]);
  76
  77       *enabled_buffers &= ~(((1ull << count) - 1) << start_slot);
  78    }
  79 }
  80
  81 /**
  82  * Same as util_set_vertex_buffers_mask, but it only returns the number
  83  * of bound buffers.
  84  */
  85 void util_set_vertex_buffers_count(struct pipe_vertex_buffer *dst,
  86                                    unsigned *dst_count,
  87                                    const struct pipe_vertex_buffer *src,
  88                                    unsigned start_slot, unsigned count)
  89 {
  90    unsigned i;
  91    uint32_t enabled_buffers = 0;
  92
  93    for (i = 0; i < *dst_count; i++) {
  94       if (dst[i].buffer.resource)
  95          enabled_buffers |= (1ull << i);
  96    }
  97
  98    util_set_vertex_buffers_mask(dst, &enabled_buffers, src, start_slot,
  99                                 count);
 100
 101    *dst_count = util_last_bit(enabled_buffers);
 102 }
 103
 104 /**
 105  * Given a user index buffer, save the structure to "saved", and upload it.
 106  */
 107 bool
 108 util_upload_index_buffer(struct pipe_context *pipe,
 109                          const struct pipe_draw_info *info,
 110                          struct pipe_resource **out_buffer,
 111                          unsigned *out_offset)
 112 {
 113    unsigned start_offset = info->start * info->index_size;
 114
 115    u_upload_data(pipe->stream_uploader, start_offset,
 116                  info->count * info->index_size, 4,
 117                  (char*)info->index.user + start_offset,
 118                  out_offset, out_buffer);
 119    u_upload_unmap(pipe->stream_uploader);
 120    *out_offset -= start_offset;
 121    return *out_buffer != NULL;
 122 }
 123
 124 /**
 125  * Called by MakeCurrent. Used to notify the driver that the application
 126  * thread may have been changed.
 127  *
 128  * The function pins the current thread and driver threads to a group of
 129  * CPU cores that share the same L3 cache. This is needed for good multi-
 130  * threading performance on AMD Zen CPUs.
 131  *
 132  * \param upper_thread  thread in the state tracker that also needs to be
 133  *                      pinned.
 134  */
 135 void
 136 util_pin_driver_threads_to_random_L3(struct pipe_context *ctx,
 137                                      thrd_t *upper_thread)
 138 {
 139    /* If pinning has no effect, don't do anything. */
 140    if (util_cpu_caps.nr_cpus == util_cpu_caps.cores_per_L3)
 141       return;
 142
 143    unsigned num_L3_caches = util_cpu_caps.nr_cpus /
 144                             util_cpu_caps.cores_per_L3;
 145
 146    /* Get a semi-random number. */
 147    int64_t t = os_time_get_nano();
 148    unsigned cache = (t ^ (t >> 8) ^ (t >> 16)) % num_L3_caches;
 149
 150    /* Tell the driver to pin its threads to the selected L3 cache. */
 151    if (ctx->set_context_param) {
 152       ctx->set_context_param(ctx, PIPE_CONTEXT_PARAM_PIN_THREADS_TO_L3_CACHE,
 153                              cache);
 154    }
 155
 156    /* Do the same for the upper level thread if there is any (e.g. glthread) */
 157    if (upper_thread)
 158       util_pin_thread_to_L3(*upper_thread, cache, util_cpu_caps.cores_per_L3);
 159 }
 160
 161 /* This is a helper for hardware bring-up. Don't remove. */
 162 struct pipe_query *
 163 util_begin_pipestat_query(struct pipe_context *ctx)
 164 {
 165    struct pipe_query *q =
 166       ctx->create_query(ctx, PIPE_QUERY_PIPELINE_STATISTICS, 0);
 167    if (!q)
 168       return NULL;
 169
 170    ctx->begin_query(ctx, q);
 171    return q;
 172 }
 173
 174 /* This is a helper for hardware bring-up. Don't remove. */
 175 void
 176 util_end_pipestat_query(struct pipe_context *ctx, struct pipe_query *q,
 177                         FILE *f)
 178 {
 179    static unsigned counter;
 180    struct pipe_query_data_pipeline_statistics stats;
 181
 182    ctx->end_query(ctx, q);
 183    ctx->get_query_result(ctx, q, true, (void*)&stats);
 184    ctx->destroy_query(ctx, q);
 185
 186    fprintf(f,
 187            "Draw call %u:\n"
 188            "    ia_vertices    = %"PRIu64"\n"
 189            "    ia_primitives  = %"PRIu64"\n"
 190            "    vs_invocations = %"PRIu64"\n"
 191            "    gs_invocations = %"PRIu64"\n"
 192            "    gs_primitives  = %"PRIu64"\n"
 193            "    c_invocations  = %"PRIu64"\n"
 194            "    c_primitives   = %"PRIu64"\n"
 195            "    ps_invocations = %"PRIu64"\n"
 196            "    hs_invocations = %"PRIu64"\n"
 197            "    ds_invocations = %"PRIu64"\n"
 198            "    cs_invocations = %"PRIu64"\n",
 199            p_atomic_inc_return(&counter),
 200            stats.ia_vertices,
 201            stats.ia_primitives,
 202            stats.vs_invocations,
 203            stats.gs_invocations,
 204            stats.gs_primitives,
 205            stats.c_invocations,
 206            stats.c_primitives,
 207            stats.ps_invocations,
 208            stats.hs_invocations,
 209            stats.ds_invocations,
 210            stats.cs_invocations);
 211 }
 212
 213 /* This is a helper for hardware bring-up. Don't remove. */
 214 void
 215 util_wait_for_idle(struct pipe_context *ctx)
 216 {
 217    struct pipe_fence_handle *fence = NULL;
 218
 219    ctx->flush(ctx, &fence, 0);
 220    ctx->screen->fence_finish(ctx->screen, NULL, fence, PIPE_TIMEOUT_INFINITE);
 221 }
 222
 223 void
 224 util_throttle_init(struct util_throttle *t, uint64_t max_mem_usage)
 225 {
 226    t->max_mem_usage = max_mem_usage;
 227 }
 228
 229 void
 230 util_throttle_deinit(struct pipe_screen *screen, struct util_throttle *t)
 231 {
 232    for (unsigned i = 0; i < ARRAY_SIZE(t->ring); i++)
 233       screen->fence_reference(screen, &t->ring[i].fence, NULL);
 234 }
 235
 236 static uint64_t
 237 util_get_throttle_total_memory_usage(struct util_throttle *t)
 238 {
 239    uint64_t total_usage = 0;
 240
 241    for (unsigned i = 0; i < ARRAY_SIZE(t->ring); i++)
 242       total_usage += t->ring[i].mem_usage;
 243    return total_usage;
 244 }
 245
 246 static void util_dump_throttle_ring(struct util_throttle *t)
 247 {
 248    printf("Throttle:\n");
 249    for (unsigned i = 0; i < ARRAY_SIZE(t->ring); i++) {
 250       printf("  ring[%u]: fence = %s, mem_usage = %"PRIu64"%s%s\n",
 251              i, t->ring[i].fence ? "yes" : " no",
 252              t->ring[i].mem_usage,
 253              t->flush_index == i ? " [flush]" : "",
 254              t->wait_index == i ? " [wait]" : "");
 255    }
 256 }
 257
 258 /**
 259  * Notify util_throttle that the next operation allocates memory.
 260  * util_throttle tracks memory usage and waits for fences until its tracked
 261  * memory usage decreases.
 262  *
 263  * Example:
 264  *   util_throttle_memory_usage(..., w*h*d*Bpp);
 265  *   TexSubImage(..., w, h, d, ...);
 266  *
 267  * This means that TexSubImage can't allocate more memory its maximum limit
 268  * set during initialization.
 269  */
 270 void
 271 util_throttle_memory_usage(struct pipe_context *pipe,
 272                            struct util_throttle *t, uint64_t memory_size)
 273 {
 274    (void)util_dump_throttle_ring; /* silence warning */
 275
 276    if (!t->max_mem_usage)
 277       return;
 278
 279    struct pipe_screen *screen = pipe->screen;
 280    struct pipe_fence_handle **fence = NULL;
 281    unsigned ring_size = ARRAY_SIZE(t->ring);
 282    uint64_t total = util_get_throttle_total_memory_usage(t);
 283
 284    /* If there is not enough memory, walk the list of fences and find
 285     * the latest one that we need to wait for.
 286     */
 287    while (t->wait_index != t->flush_index &&
 288           total && total + memory_size > t->max_mem_usage) {
 289       assert(t->ring[t->wait_index].fence);
 290
 291       /* Release an older fence if we need to wait for a newer one. */
 292       if (fence)
 293          screen->fence_reference(screen, fence, NULL);
 294
 295       fence = &t->ring[t->wait_index].fence;
 296       t->ring[t->wait_index].mem_usage = 0;
 297       t->wait_index = (t->wait_index + 1) % ring_size;
 298
 299       total = util_get_throttle_total_memory_usage(t);
 300    }
 301
 302    /* Wait for the fence to decrease memory usage. */
 303    if (fence) {
 304       screen->fence_finish(screen, pipe, *fence, PIPE_TIMEOUT_INFINITE);
 305       screen->fence_reference(screen, fence, NULL);
 306    }
 307
 308    /* Flush and get a fence if we've exhausted memory usage for the current
 309     * slot.
 310     */
 311    if (t->ring[t->flush_index].mem_usage &&
 312        t->ring[t->flush_index].mem_usage + memory_size >
 313        t->max_mem_usage / (ring_size / 2)) {
 314       struct pipe_fence_handle **fence =
 315          &t->ring[t->flush_index].fence;
 316
 317       /* Expect that the current flush slot doesn't have a fence yet. */
 318       assert(!*fence);
 319
 320       pipe->flush(pipe, fence, PIPE_FLUSH_ASYNC);
 321       t->flush_index = (t->flush_index + 1) % ring_size;
 322
 323       /* Vacate the next slot if it's occupied. This should be rare. */
 324       if (t->flush_index == t->wait_index) {
 325          struct pipe_fence_handle **fence =
 326             &t->ring[t->wait_index].fence;
 327
 328          t->ring[t->wait_index].mem_usage = 0;
 329          t->wait_index = (t->wait_index + 1) % ring_size;
 330
 331          assert(*fence);
 332          screen->fence_finish(screen, pipe, *fence, PIPE_TIMEOUT_INFINITE);
 333          screen->fence_reference(screen, fence, NULL);
 334       }
 335
 336       assert(!t->ring[t->flush_index].mem_usage);
 337       assert(!t->ring[t->flush_index].fence);
 338    }
 339
 340    t->ring[t->flush_index].mem_usage += memory_size;
 341 }