src/mesa/drivers/dri/i965/brw_compute.c

   1 /*
   2  * Copyright © 2014 Intel Corporation
   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
  21  * DEALINGS IN THE SOFTWARE.
  22  */
  23
  24 #include <sys/errno.h>
  25
  26 #include "main/condrender.h"
  27 #include "main/mtypes.h"
  28 #include "main/state.h"
  29 #include "brw_context.h"
  30 #include "brw_draw.h"
  31 #include "brw_state.h"
  32 #include "intel_batchbuffer.h"
  33 #include "intel_buffer_objects.h"
  34 #include "brw_defines.h"
  35
  36
  37 static void
  38 prepare_indirect_gpgpu_walker(struct brw_context *brw)
  39 {
  40    GLintptr indirect_offset = brw->compute.num_work_groups_offset;
  41    drm_intel_bo *bo = brw->compute.num_work_groups_bo;
  42
  43    brw_load_register_mem(brw, GEN7_GPGPU_DISPATCHDIMX, bo,
  44                          I915_GEM_DOMAIN_VERTEX, 0,
  45                          indirect_offset + 0);
  46    brw_load_register_mem(brw, GEN7_GPGPU_DISPATCHDIMY, bo,
  47                          I915_GEM_DOMAIN_VERTEX, 0,
  48                          indirect_offset + 4);
  49    brw_load_register_mem(brw, GEN7_GPGPU_DISPATCHDIMZ, bo,
  50                          I915_GEM_DOMAIN_VERTEX, 0,
  51                          indirect_offset + 8);
  52
  53    if (brw->gen > 7)
  54       return;
  55
  56    /* Clear upper 32-bits of SRC0 and all 64-bits of SRC1 */
  57    BEGIN_BATCH(7);
  58    OUT_BATCH(MI_LOAD_REGISTER_IMM | (7 - 2));
  59    OUT_BATCH(MI_PREDICATE_SRC0 + 4);
  60    OUT_BATCH(0u);
  61    OUT_BATCH(MI_PREDICATE_SRC1 + 0);
  62    OUT_BATCH(0u);
  63    OUT_BATCH(MI_PREDICATE_SRC1 + 4);
  64    OUT_BATCH(0u);
  65    ADVANCE_BATCH();
  66
  67    /* Load compute_dispatch_indirect_x_size into SRC0 */
  68    brw_load_register_mem(brw, MI_PREDICATE_SRC0, bo,
  69                          I915_GEM_DOMAIN_INSTRUCTION, 0,
  70                          indirect_offset + 0);
  71
  72    /* predicate = (compute_dispatch_indirect_x_size == 0); */
  73    BEGIN_BATCH(1);
  74    OUT_BATCH(GEN7_MI_PREDICATE |
  75              MI_PREDICATE_LOADOP_LOAD |
  76              MI_PREDICATE_COMBINEOP_SET |
  77              MI_PREDICATE_COMPAREOP_SRCS_EQUAL);
  78    ADVANCE_BATCH();
  79
  80    /* Load compute_dispatch_indirect_y_size into SRC0 */
  81    brw_load_register_mem(brw, MI_PREDICATE_SRC0, bo,
  82                          I915_GEM_DOMAIN_INSTRUCTION, 0,
  83                          indirect_offset + 4);
  84
  85    /* predicate |= (compute_dispatch_indirect_y_size == 0); */
  86    BEGIN_BATCH(1);
  87    OUT_BATCH(GEN7_MI_PREDICATE |
  88              MI_PREDICATE_LOADOP_LOAD |
  89              MI_PREDICATE_COMBINEOP_OR |
  90              MI_PREDICATE_COMPAREOP_SRCS_EQUAL);
  91    ADVANCE_BATCH();
  92
  93    /* Load compute_dispatch_indirect_z_size into SRC0 */
  94    brw_load_register_mem(brw, MI_PREDICATE_SRC0, bo,
  95                          I915_GEM_DOMAIN_INSTRUCTION, 0,
  96                          indirect_offset + 8);
  97
  98    /* predicate |= (compute_dispatch_indirect_z_size == 0); */
  99    BEGIN_BATCH(1);
 100    OUT_BATCH(GEN7_MI_PREDICATE |
 101              MI_PREDICATE_LOADOP_LOAD |
 102              MI_PREDICATE_COMBINEOP_OR |
 103              MI_PREDICATE_COMPAREOP_SRCS_EQUAL);
 104    ADVANCE_BATCH();
 105
 106    /* predicate = !predicate; */
 107    BEGIN_BATCH(1);
 108    OUT_BATCH(GEN7_MI_PREDICATE |
 109              MI_PREDICATE_LOADOP_LOADINV |
 110              MI_PREDICATE_COMBINEOP_OR |
 111              MI_PREDICATE_COMPAREOP_FALSE);
 112    ADVANCE_BATCH();
 113 }
 114
 115 static void
 116 brw_emit_gpgpu_walker(struct brw_context *brw)
 117 {
 118    const struct brw_cs_prog_data *prog_data = brw->cs.prog_data;
 119
 120    const GLuint *num_groups = brw->compute.num_work_groups;
 121    uint32_t indirect_flag;
 122
 123    if (brw->compute.num_work_groups_bo == NULL) {
 124       indirect_flag = 0;
 125    } else {
 126       indirect_flag =
 127          GEN7_GPGPU_INDIRECT_PARAMETER_ENABLE |
 128          (brw->gen == 7 ? GEN7_GPGPU_PREDICATE_ENABLE : 0);
 129       prepare_indirect_gpgpu_walker(brw);
 130    }
 131
 132    const unsigned simd_size = prog_data->simd_size;
 133    unsigned group_size = prog_data->local_size[0] *
 134       prog_data->local_size[1] * prog_data->local_size[2];
 135    unsigned thread_width_max =
 136       (group_size + simd_size - 1) / simd_size;
 137
 138    uint32_t right_mask = 0xffffffffu >> (32 - simd_size);
 139    const unsigned right_non_aligned = group_size & (simd_size - 1);
 140    if (right_non_aligned != 0)
 141       right_mask >>= (simd_size - right_non_aligned);
 142
 143    uint32_t dwords = brw->gen < 8 ? 11 : 15;
 144    BEGIN_BATCH(dwords);
 145    OUT_BATCH(GPGPU_WALKER << 16 | (dwords - 2) | indirect_flag);
 146    OUT_BATCH(0);
 147    if (brw->gen >= 8) {
 148       OUT_BATCH(0);                     /* Indirect Data Length */
 149       OUT_BATCH(0);                     /* Indirect Data Start Address */
 150    }
 151    assert(thread_width_max <= brw->max_cs_threads);
 152    OUT_BATCH(SET_FIELD(simd_size / 16, GPGPU_WALKER_SIMD_SIZE) |
 153              SET_FIELD(thread_width_max - 1, GPGPU_WALKER_THREAD_WIDTH_MAX));
 154    OUT_BATCH(0);                        /* Thread Group ID Starting X */
 155    if (brw->gen >= 8)
 156       OUT_BATCH(0);                     /* MBZ */
 157    OUT_BATCH(num_groups[0]);            /* Thread Group ID X Dimension */
 158    OUT_BATCH(0);                        /* Thread Group ID Starting Y */
 159    if (brw->gen >= 8)
 160       OUT_BATCH(0);                     /* MBZ */
 161    OUT_BATCH(num_groups[1]);            /* Thread Group ID Y Dimension */
 162    OUT_BATCH(0);                        /* Thread Group ID Starting/Resume Z */
 163    OUT_BATCH(num_groups[2]);            /* Thread Group ID Z Dimension */
 164    OUT_BATCH(right_mask);               /* Right Execution Mask */
 165    OUT_BATCH(0xffffffff);               /* Bottom Execution Mask */
 166    ADVANCE_BATCH();
 167
 168    BEGIN_BATCH(2);
 169    OUT_BATCH(MEDIA_STATE_FLUSH << 16 | (2 - 2));
 170    OUT_BATCH(0);
 171    ADVANCE_BATCH();
 172 }
 173
 174
 175 static void
 176 brw_dispatch_compute_common(struct gl_context *ctx)
 177 {
 178    struct brw_context *brw = brw_context(ctx);
 179    int estimated_buffer_space_needed;
 180    bool fail_next = false;
 181
 182    if (!_mesa_check_conditional_render(ctx))
 183       return;
 184
 185    if (ctx->NewState)
 186       _mesa_update_state(ctx);
 187
 188    brw_validate_textures(brw);
 189
 190    const int sampler_state_size = 16; /* 16 bytes */
 191    estimated_buffer_space_needed = 512; /* batchbuffer commands */
 192    estimated_buffer_space_needed += (BRW_MAX_TEX_UNIT *
 193                                      (sampler_state_size +
 194                                       sizeof(struct gen5_sampler_default_color)));
 195    estimated_buffer_space_needed += 1024; /* push constants */
 196    estimated_buffer_space_needed += 512; /* misc. pad */
 197
 198    /* Flush the batch if it's approaching full, so that we don't wrap while
 199     * we've got validated state that needs to be in the same batch as the
 200     * primitives.
 201     */
 202    intel_batchbuffer_require_space(brw, estimated_buffer_space_needed,
 203                                    RENDER_RING);
 204    intel_batchbuffer_save_state(brw);
 205
 206  retry:
 207    brw->no_batch_wrap = true;
 208    brw_upload_compute_state(brw);
 209
 210    brw_emit_gpgpu_walker(brw);
 211
 212    brw->no_batch_wrap = false;
 213
 214    if (dri_bufmgr_check_aperture_space(&brw->batch.bo, 1)) {
 215       if (!fail_next) {
 216          intel_batchbuffer_reset_to_saved(brw);
 217          intel_batchbuffer_flush(brw);
 218          fail_next = true;
 219          goto retry;
 220       } else {
 221          if (intel_batchbuffer_flush(brw) == -ENOSPC) {
 222             static bool warned = false;
 223
 224             if (!warned) {
 225                fprintf(stderr, "i965: Single compute shader dispatch "
 226                        "exceeded available aperture space\n");
 227                warned = true;
 228             }
 229          }
 230       }
 231    }
 232
 233    /* Now that we know we haven't run out of aperture space, we can safely
 234     * reset the dirty bits.
 235     */
 236    brw_compute_state_finished(brw);
 237
 238    if (brw->always_flush_batch)
 239       intel_batchbuffer_flush(brw);
 240
 241    brw_state_cache_check_size(brw);
 242
 243    /* Note: since compute shaders can't write to framebuffers, there's no need
 244     * to call brw_postdraw_set_buffers_need_resolve().
 245     */
 246 }
 247
 248 static void
 249 brw_dispatch_compute(struct gl_context *ctx, const GLuint *num_groups) {
 250    struct brw_context *brw = brw_context(ctx);
 251
 252    brw->compute.num_work_groups_bo = NULL;
 253    brw->compute.num_work_groups = num_groups;
 254    ctx->NewDriverState |= BRW_NEW_CS_WORK_GROUPS;
 255
 256    brw_dispatch_compute_common(ctx);
 257 }
 258
 259 static void
 260 brw_dispatch_compute_indirect(struct gl_context *ctx, GLintptr indirect)
 261 {
 262    struct brw_context *brw = brw_context(ctx);
 263    static const GLuint indirect_group_counts[3] = { 0, 0, 0 };
 264    struct gl_buffer_object *indirect_buffer = ctx->DispatchIndirectBuffer;
 265    drm_intel_bo *bo =
 266       intel_bufferobj_buffer(brw,
 267                              intel_buffer_object(indirect_buffer),
 268                              indirect, 3 * sizeof(GLuint));
 269
 270    brw->compute.num_work_groups_bo = bo;
 271    brw->compute.num_work_groups_offset = indirect;
 272    brw->compute.num_work_groups = indirect_group_counts;
 273    ctx->NewDriverState |= BRW_NEW_CS_WORK_GROUPS;
 274
 275    brw_dispatch_compute_common(ctx);
 276 }
 277
 278 void
 279 brw_init_compute_functions(struct dd_function_table *functions)
 280 {
 281    functions->DispatchCompute = brw_dispatch_compute;
 282    functions->DispatchComputeIndirect = brw_dispatch_compute_indirect;
 283 }