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 #include "program/prog_statevars.h"
  34 #include "glsl/ir_uniform.h"
  35
  36 static unsigned
  37 get_cs_thread_count(const struct brw_cs_prog_data *cs_prog_data)
  38 {
  39    const unsigned simd_size = cs_prog_data->simd_size;
  40    unsigned group_size = cs_prog_data->local_size[0] *
  41       cs_prog_data->local_size[1] * cs_prog_data->local_size[2];
  42
  43    return (group_size + simd_size - 1) / simd_size;
  44 }
  45
  46
  47 static void
  48 brw_upload_cs_state(struct brw_context *brw)
  49 {
  50    if (!brw->cs.prog_data)
  51       return;
  52
  53    uint32_t offset;
  54    uint32_t *desc = (uint32_t*) brw_state_batch(brw, AUB_TRACE_SURFACE_STATE,
  55                                                 8 * 4, 64, &offset);
  56    struct gl_program *prog = (struct gl_program *) brw->compute_program;
  57    struct brw_stage_state *stage_state = &brw->cs.base;
  58    struct brw_cs_prog_data *cs_prog_data = brw->cs.prog_data;
  59    struct brw_stage_prog_data *prog_data = &cs_prog_data->base;
  60
  61    if (INTEL_DEBUG & DEBUG_SHADER_TIME) {
  62       brw->vtbl.emit_buffer_surface_state(
  63          brw, &stage_state->surf_offset[
  64                  prog_data->binding_table.shader_time_start],
  65          brw->shader_time.bo, 0, BRW_SURFACEFORMAT_RAW,
  66          brw->shader_time.bo->size, 1, true);
  67    }
  68
  69    uint32_t *bind = (uint32_t*) brw_state_batch(brw, AUB_TRACE_BINDING_TABLE,
  70                                             prog_data->binding_table.size_bytes,
  71                                             32, &stage_state->bind_bo_offset);
  72
  73    unsigned local_id_dwords = 0;
  74
  75    if (prog->SystemValuesRead & SYSTEM_BIT_LOCAL_INVOCATION_ID)
  76       local_id_dwords = cs_prog_data->local_invocation_id_regs * 8;
  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
 170    assert(prog_data->total_shared <= 64 * 1024);
 171    uint32_t slm_size = 0;
 172    if (prog_data->total_shared > 0) {
 173       /* slm_size is in 4k increments, but must be a power of 2. */
 174       slm_size = 4 * 1024;
 175       while (slm_size < prog_data->total_shared)
 176          slm_size <<= 1;
 177       slm_size /= 4 * 1024;
 178    }
 179
 180    desc[dw++] =
 181       SET_FIELD(cs_prog_data->uses_barrier, MEDIA_BARRIER_ENABLE) |
 182       SET_FIELD(slm_size, MEDIA_SHARED_LOCAL_MEMORY_SIZE) |
 183       media_threads;
 184
 185    BEGIN_BATCH(4);
 186    OUT_BATCH(MEDIA_INTERFACE_DESCRIPTOR_LOAD << 16 | (4 - 2));
 187    OUT_BATCH(0);
 188    OUT_BATCH(8 * 4);
 189    OUT_BATCH(offset);
 190    ADVANCE_BATCH();
 191 }
 192
 193 const struct brw_tracked_state brw_cs_state = {
 194    .dirty = {
 195       .mesa = _NEW_PROGRAM_CONSTANTS,
 196       .brw = BRW_NEW_BATCH |
 197              BRW_NEW_CS_PROG_DATA |
 198              BRW_NEW_PUSH_CONSTANT_ALLOCATION |
 199              BRW_NEW_SURFACES,
 200    },
 201    .emit = brw_upload_cs_state
 202 };
 203
 204
 205 /**
 206  * Creates a region containing the push constants for the CS on gen7+.
 207  *
 208  * Push constants are constant values (such as GLSL uniforms) that are
 209  * pre-loaded into a shader stage's register space at thread spawn time.
 210  *
 211  * For other stages, see brw_curbe.c:brw_upload_constant_buffer for the
 212  * equivalent gen4/5 code and gen6_vs_state.c:gen6_upload_push_constants for
 213  * gen6+.
 214  */
 215 static void
 216 brw_upload_cs_push_constants(struct brw_context *brw,
 217                              const struct gl_program *prog,
 218                              const struct brw_cs_prog_data *cs_prog_data,
 219                              struct brw_stage_state *stage_state,
 220                              enum aub_state_struct_type type)
 221 {
 222    struct gl_context *ctx = &brw->ctx;
 223    const struct brw_stage_prog_data *prog_data =
 224       (struct brw_stage_prog_data*) cs_prog_data;
 225    unsigned local_id_dwords = 0;
 226
 227    if (prog->SystemValuesRead & SYSTEM_BIT_LOCAL_INVOCATION_ID)
 228       local_id_dwords = cs_prog_data->local_invocation_id_regs * 8;
 229
 230    /* Updates the ParamaterValues[i] pointers for all parameters of the
 231     * basic type of PROGRAM_STATE_VAR.
 232     */
 233    /* XXX: Should this happen somewhere before to get our state flag set? */
 234    _mesa_load_state_parameters(ctx, prog->Parameters);
 235
 236    if (prog_data->nr_params == 0 && local_id_dwords == 0) {
 237       stage_state->push_const_size = 0;
 238    } else {
 239       gl_constant_value *param;
 240       unsigned i, t;
 241
 242       const unsigned push_constant_data_size =
 243          (local_id_dwords + prog_data->nr_params) * sizeof(gl_constant_value);
 244       const unsigned reg_aligned_constant_size = ALIGN(push_constant_data_size, 32);
 245       const unsigned param_aligned_count =
 246          reg_aligned_constant_size / sizeof(*param);
 247
 248       unsigned threads = get_cs_thread_count(cs_prog_data);
 249
 250       param = (gl_constant_value*)
 251          brw_state_batch(brw, type,
 252                          reg_aligned_constant_size * threads,
 253                          32, &stage_state->push_const_offset);
 254       assert(param);
 255
 256       STATIC_ASSERT(sizeof(gl_constant_value) == sizeof(float));
 257
 258       brw_cs_fill_local_id_payload(cs_prog_data, param, threads,
 259                                    reg_aligned_constant_size);
 260
 261       /* _NEW_PROGRAM_CONSTANTS */
 262       for (t = 0; t < threads; t++) {
 263          gl_constant_value *next_param =
 264             &param[t * param_aligned_count + local_id_dwords];
 265          for (i = 0; i < prog_data->nr_params; i++) {
 266             next_param[i] = *prog_data->param[i];
 267          }
 268       }
 269
 270       stage_state->push_const_size = ALIGN(prog_data->nr_params, 8) / 8;
 271    }
 272 }
 273
 274
 275 static void
 276 gen7_upload_cs_push_constants(struct brw_context *brw)
 277 {
 278    struct brw_stage_state *stage_state = &brw->cs.base;
 279
 280    /* BRW_NEW_COMPUTE_PROGRAM */
 281    const struct brw_compute_program *cp =
 282       (struct brw_compute_program *) brw->compute_program;
 283
 284    if (cp) {
 285       /* CACHE_NEW_CS_PROG */
 286       struct brw_cs_prog_data *cs_prog_data = brw->cs.prog_data;
 287
 288       brw_upload_cs_push_constants(brw, &cp->program.Base, cs_prog_data,
 289                                    stage_state, AUB_TRACE_WM_CONSTANTS);
 290    }
 291 }
 292
 293 const struct brw_tracked_state gen7_cs_push_constants = {
 294    .dirty = {
 295       .mesa = _NEW_PROGRAM_CONSTANTS,
 296       .brw = BRW_NEW_BATCH |
 297              BRW_NEW_COMPUTE_PROGRAM |
 298              BRW_NEW_PUSH_CONSTANT_ALLOCATION,
 299    },
 300    .emit = gen7_upload_cs_push_constants,
 301 };
 302
 303 /**
 304  * Creates a new CS constant buffer reflecting the current CS program's
 305  * constants, if needed by the CS program.
 306  */
 307 static void
 308 brw_upload_cs_pull_constants(struct brw_context *brw)
 309 {
 310    struct brw_stage_state *stage_state = &brw->cs.base;
 311
 312    /* BRW_NEW_COMPUTE_PROGRAM */
 313    struct brw_compute_program *cp =
 314       (struct brw_compute_program *) brw->compute_program;
 315
 316    /* BRW_NEW_CS_PROG_DATA */
 317    const struct brw_stage_prog_data *prog_data = &brw->cs.prog_data->base;
 318
 319    /* _NEW_PROGRAM_CONSTANTS */
 320    brw_upload_pull_constants(brw, BRW_NEW_SURFACES, &cp->program.Base,
 321                              stage_state, prog_data);
 322 }
 323
 324 const struct brw_tracked_state brw_cs_pull_constants = {
 325    .dirty = {
 326       .mesa = _NEW_PROGRAM_CONSTANTS,
 327       .brw = BRW_NEW_BATCH |
 328              BRW_NEW_COMPUTE_PROGRAM |
 329              BRW_NEW_CS_PROG_DATA,
 330    },
 331    .emit = brw_upload_cs_pull_constants,
 332 };