src/mesa/drivers/dri/i965/gen7_cs_state.c

   1 /*
   2  * Copyright © 2015 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 "util/ralloc.h"
  25 #include "brw_context.h"
  26 #include "brw_cs.h"
  27 #include "brw_eu.h"
  28 #include "brw_wm.h"
  29 #include "brw_shader.h"
  30 #include "intel_mipmap_tree.h"
  31 #include "intel_batchbuffer.h"
  32 #include "brw_state.h"
  33
  34 static unsigned
  35 get_cs_thread_count(const struct brw_cs_prog_data *cs_prog_data)
  36 {
  37    const unsigned simd_size = cs_prog_data->simd_size;
  38    unsigned group_size = cs_prog_data->local_size[0] *
  39       cs_prog_data->local_size[1] * cs_prog_data->local_size[2];
  40
  41    return (group_size + simd_size - 1) / simd_size;
  42 }
  43
  44
  45 static void
  46 brw_upload_cs_state(struct brw_context *brw)
  47 {
  48    if (!brw->cs.prog_data)
  49       return;
  50
  51    uint32_t offset;
  52    uint32_t *desc = (uint32_t*) brw_state_batch(brw, AUB_TRACE_SURFACE_STATE,
  53                                                 8 * 4, 64, &offset);
  54    struct gl_program *prog = (struct gl_program *) brw->compute_program;
  55    struct brw_stage_state *stage_state = &brw->cs.base;
  56    struct brw_cs_prog_data *cs_prog_data = brw->cs.prog_data;
  57    struct brw_stage_prog_data *prog_data = &cs_prog_data->base;
  58
  59    if (INTEL_DEBUG & DEBUG_SHADER_TIME) {
  60       brw->vtbl.emit_buffer_surface_state(
  61          brw, &stage_state->surf_offset[
  62                  prog_data->binding_table.shader_time_start],
  63          brw->shader_time.bo, 0, BRW_SURFACEFORMAT_RAW,
  64          brw->shader_time.bo->size, 1, true);
  65    }
  66
  67    uint32_t *bind = (uint32_t*) brw_state_batch(brw, AUB_TRACE_BINDING_TABLE,
  68                                             prog_data->binding_table.size_bytes,
  69                                             32, &stage_state->bind_bo_offset);
  70
  71    unsigned local_id_dwords = 0;
  72
  73    if (prog->SystemValuesRead & SYSTEM_BIT_LOCAL_INVOCATION_ID) {
  74       local_id_dwords =
  75          brw_cs_prog_local_id_payload_dwords(prog, cs_prog_data->simd_size);
  76    }
  77
  78    unsigned push_constant_data_size =
  79       (prog_data->nr_params + local_id_dwords) * sizeof(gl_constant_value);
  80    unsigned reg_aligned_constant_size = ALIGN(push_constant_data_size, 32);
  81    unsigned push_constant_regs = reg_aligned_constant_size / 32;
  82    unsigned threads = get_cs_thread_count(cs_prog_data);
  83
  84    uint32_t dwords = brw->gen < 8 ? 8 : 9;
  85    BEGIN_BATCH(dwords);
  86    OUT_BATCH(MEDIA_VFE_STATE << 16 | (dwords - 2));
  87
  88    if (prog_data->total_scratch) {
  89       if (brw->gen >= 8)
  90          OUT_RELOC64(stage_state->scratch_bo,
  91                      I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER,
  92                      ffs(prog_data->total_scratch) - 11);
  93       else
  94          OUT_RELOC(stage_state->scratch_bo,
  95                    I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER,
  96                    ffs(prog_data->total_scratch) - 11);
  97    } else {
  98       OUT_BATCH(0);
  99       if (brw->gen >= 8)
 100          OUT_BATCH(0);
 101    }
 102
 103    const uint32_t vfe_num_urb_entries = brw->gen >= 8 ? 2 : 0;
 104    const uint32_t vfe_gpgpu_mode =
 105       brw->gen == 7 ? SET_FIELD(1, GEN7_MEDIA_VFE_STATE_GPGPU_MODE) : 0;
 106    OUT_BATCH(SET_FIELD(brw->max_cs_threads - 1, MEDIA_VFE_STATE_MAX_THREADS) |
 107              SET_FIELD(vfe_num_urb_entries, MEDIA_VFE_STATE_URB_ENTRIES) |
 108              SET_FIELD(1, MEDIA_VFE_STATE_RESET_GTW_TIMER) |
 109              SET_FIELD(1, MEDIA_VFE_STATE_BYPASS_GTW) |
 110              vfe_gpgpu_mode);
 111
 112    OUT_BATCH(0);
 113    const uint32_t vfe_urb_allocation = brw->gen >= 8 ? 2 : 0;
 114
 115    /* We are uploading duplicated copies of push constant uniforms for each
 116     * thread. Although the local id data needs to vary per thread, it won't
 117     * change for other uniform data. Unfortunately this duplication is
 118     * required for gen7. As of Haswell, this duplication can be avoided, but
 119     * this older mechanism with duplicated data continues to work.
 120     *
 121     * FINISHME: As of Haswell, we could make use of the
 122     * INTERFACE_DESCRIPTOR_DATA "Cross-Thread Constant Data Read Length" field
 123     * to only store one copy of uniform data.
 124     *
 125     * FINISHME: Broadwell adds a new alternative "Indirect Payload Storage"
 126     * which is described in the GPGPU_WALKER command and in the Broadwell PRM
 127     * Volume 7: 3D Media GPGPU, under Media GPGPU Pipeline => Mode of
 128     * Operations => GPGPU Mode => Indirect Payload Storage.
 129     *
 130     * Note: The constant data is built in brw_upload_cs_push_constants below.
 131     */
 132    const uint32_t vfe_curbe_allocation = push_constant_regs * threads;
 133    OUT_BATCH(SET_FIELD(vfe_urb_allocation, MEDIA_VFE_STATE_URB_ALLOC) |
 134              SET_FIELD(vfe_curbe_allocation, MEDIA_VFE_STATE_CURBE_ALLOC));
 135    OUT_BATCH(0);
 136    OUT_BATCH(0);
 137    OUT_BATCH(0);
 138    ADVANCE_BATCH();
 139
 140    if (reg_aligned_constant_size > 0) {
 141       BEGIN_BATCH(4);
 142       OUT_BATCH(MEDIA_CURBE_LOAD << 16 | (4 - 2));
 143       OUT_BATCH(0);
 144       OUT_BATCH(reg_aligned_constant_size * threads);
 145       OUT_BATCH(stage_state->push_const_offset);
 146       ADVANCE_BATCH();
 147    }
 148
 149    /* BRW_NEW_SURFACES and BRW_NEW_*_CONSTBUF */
 150    memcpy(bind, stage_state->surf_offset,
 151           prog_data->binding_table.size_bytes);
 152
 153    memset(desc, 0, 8 * 4);
 154
 155    int dw = 0;
 156    desc[dw++] = brw->cs.base.prog_offset;
 157    if (brw->gen >= 8)
 158       desc[dw++] = 0; /* Kernel Start Pointer High */
 159    desc[dw++] = 0;
 160    desc[dw++] = stage_state->sampler_offset |
 161       ((stage_state->sampler_count + 3) / 4);
 162    desc[dw++] = stage_state->bind_bo_offset;
 163    desc[dw++] = SET_FIELD(push_constant_regs, MEDIA_CURBE_READ_LENGTH);
 164    const uint32_t media_threads =
 165       brw->gen >= 8 ?
 166       SET_FIELD(threads, GEN8_MEDIA_GPGPU_THREAD_COUNT) :
 167       SET_FIELD(threads, MEDIA_GPGPU_THREAD_COUNT);
 168    assert(threads <= brw->max_cs_threads);
 169    desc[dw++] =
 170       SET_FIELD(cs_prog_data->uses_barrier, MEDIA_BARRIER_ENABLE) |
 171       media_threads;
 172
 173    BEGIN_BATCH(4);
 174    OUT_BATCH(MEDIA_INTERFACE_DESCRIPTOR_LOAD << 16 | (4 - 2));
 175    OUT_BATCH(0);
 176    OUT_BATCH(8 * 4);
 177    OUT_BATCH(offset);
 178    ADVANCE_BATCH();
 179 }
 180
 181 const struct brw_tracked_state brw_cs_state = {
 182    .dirty = {
 183       .mesa = _NEW_PROGRAM_CONSTANTS,
 184       .brw = BRW_NEW_BATCH |
 185              BRW_NEW_CS_PROG_DATA |
 186              BRW_NEW_PUSH_CONSTANT_ALLOCATION |
 187              BRW_NEW_SURFACES,
 188    },
 189    .emit = brw_upload_cs_state
 190 };
 191
 192
 193 /**
 194  * We are building the local ID push constant data using the simplest possible
 195  * method. We simply push the local IDs directly as they should appear in the
 196  * registers for the uvec3 gl_LocalInvocationID variable.
 197  *
 198  * Therefore, for SIMD8, we use 3 full registers, and for SIMD16 we use 6
 199  * registers worth of push constant space.
 200  *
 201  * Note: Any updates to brw_cs_prog_local_id_payload_dwords,
 202  * fill_local_id_payload or fs_visitor::emit_cs_local_invocation_id_setup need
 203  * to coordinated.
 204  *
 205  * FINISHME: There are a few easy optimizations to consider.
 206  *
 207  * 1. If gl_WorkGroupSize x, y or z is 1, we can just use zero, and there is
 208  *    no need for using push constant space for that dimension.
 209  *
 210  * 2. Since GL_MAX_COMPUTE_WORK_GROUP_SIZE is currently 1024 or less, we can
 211  *    easily use 16-bit words rather than 32-bit dwords in the push constant
 212  *    data.
 213  *
 214  * 3. If gl_WorkGroupSize x, y or z is small, then we can use bytes for
 215  *    conveying the data, and thereby reduce push constant usage.
 216  *
 217  */
 218 unsigned
 219 brw_cs_prog_local_id_payload_dwords(const struct gl_program *prog,
 220                                     unsigned dispatch_width)
 221 {
 222    return 3 * dispatch_width;
 223 }
 224
 225
 226 static void
 227 fill_local_id_payload(const struct brw_cs_prog_data *cs_prog_data,
 228                       void *buffer, unsigned *x, unsigned *y, unsigned *z)
 229 {
 230    uint32_t *param = (uint32_t *)buffer;
 231    for (unsigned i = 0; i < cs_prog_data->simd_size; i++) {
 232       param[0 * cs_prog_data->simd_size + i] = *x;
 233       param[1 * cs_prog_data->simd_size + i] = *y;
 234       param[2 * cs_prog_data->simd_size + i] = *z;
 235
 236       (*x)++;
 237       if (*x == cs_prog_data->local_size[0]) {
 238          *x = 0;
 239          (*y)++;
 240          if (*y == cs_prog_data->local_size[1]) {
 241             *y = 0;
 242             (*z)++;
 243             if (*z == cs_prog_data->local_size[2])
 244                *z = 0;
 245          }
 246       }
 247    }
 248 }
 249
 250
 251 /**
 252  * Creates a region containing the push constants for the CS on gen7+.
 253  *
 254  * Push constants are constant values (such as GLSL uniforms) that are
 255  * pre-loaded into a shader stage's register space at thread spawn time.
 256  *
 257  * For other stages, see brw_curbe.c:brw_upload_constant_buffer for the
 258  * equivalent gen4/5 code and gen6_vs_state.c:gen6_upload_push_constants for
 259  * gen6+.
 260  */
 261 static void
 262 brw_upload_cs_push_constants(struct brw_context *brw,
 263                              const struct gl_program *prog,
 264                              const struct brw_cs_prog_data *cs_prog_data,
 265                              struct brw_stage_state *stage_state,
 266                              enum aub_state_struct_type type)
 267 {
 268    struct gl_context *ctx = &brw->ctx;
 269    const struct brw_stage_prog_data *prog_data =
 270       (struct brw_stage_prog_data*) cs_prog_data;
 271    unsigned local_id_dwords = 0;
 272
 273    if (prog->SystemValuesRead & SYSTEM_BIT_LOCAL_INVOCATION_ID) {
 274       local_id_dwords =
 275          brw_cs_prog_local_id_payload_dwords(prog, cs_prog_data->simd_size);
 276    }
 277
 278    /* Updates the ParamaterValues[i] pointers for all parameters of the
 279     * basic type of PROGRAM_STATE_VAR.
 280     */
 281    /* XXX: Should this happen somewhere before to get our state flag set? */
 282    _mesa_load_state_parameters(ctx, prog->Parameters);
 283
 284    if (prog_data->nr_params == 0 && local_id_dwords == 0) {
 285       stage_state->push_const_size = 0;
 286    } else {
 287       gl_constant_value *param;
 288       unsigned i, t;
 289
 290       const unsigned push_constant_data_size =
 291          (local_id_dwords + prog_data->nr_params) * sizeof(gl_constant_value);
 292       const unsigned reg_aligned_constant_size = ALIGN(push_constant_data_size, 32);
 293       const unsigned param_aligned_count =
 294          reg_aligned_constant_size / sizeof(*param);
 295
 296       unsigned threads = get_cs_thread_count(cs_prog_data);
 297
 298       param = (gl_constant_value*)
 299          brw_state_batch(brw, type,
 300                          reg_aligned_constant_size * threads,
 301                          32, &stage_state->push_const_offset);
 302       assert(param);
 303
 304       STATIC_ASSERT(sizeof(gl_constant_value) == sizeof(float));
 305
 306       /* _NEW_PROGRAM_CONSTANTS */
 307       unsigned x = 0, y = 0, z = 0;
 308       for (t = 0; t < threads; t++) {
 309          gl_constant_value *next_param = &param[t * param_aligned_count];
 310          if (local_id_dwords > 0) {
 311             fill_local_id_payload(cs_prog_data, (void*)next_param, &x, &y, &z);
 312             next_param += local_id_dwords;
 313          }
 314          for (i = 0; i < prog_data->nr_params; i++) {
 315             next_param[i] = *prog_data->param[i];
 316          }
 317       }
 318
 319       stage_state->push_const_size = ALIGN(prog_data->nr_params, 8) / 8;
 320    }
 321 }
 322
 323
 324 static void
 325 gen7_upload_cs_push_constants(struct brw_context *brw)
 326 {
 327    struct brw_stage_state *stage_state = &brw->cs.base;
 328
 329    /* BRW_NEW_COMPUTE_PROGRAM */
 330    const struct brw_compute_program *cp =
 331       (struct brw_compute_program *) brw->compute_program;
 332
 333    if (cp) {
 334       /* CACHE_NEW_CS_PROG */
 335       struct brw_cs_prog_data *cs_prog_data = brw->cs.prog_data;
 336
 337       brw_upload_cs_push_constants(brw, &cp->program.Base, cs_prog_data,
 338                                    stage_state, AUB_TRACE_WM_CONSTANTS);
 339    }
 340 }
 341
 342 const struct brw_tracked_state gen7_cs_push_constants = {
 343    .dirty = {
 344       .mesa = _NEW_PROGRAM_CONSTANTS,
 345       .brw = BRW_NEW_BATCH |
 346              BRW_NEW_COMPUTE_PROGRAM |
 347              BRW_NEW_PUSH_CONSTANT_ALLOCATION,
 348    },
 349    .emit = gen7_upload_cs_push_constants,
 350 };