src/gallium/drivers/r600/sfn/sfn_shader_geometry.cpp

   1 /* -*- mesa-c++  -*-
   2  *
   3  * Copyright (c) 2018 Collabora LTD
   4  *
   5  * Author: Gert Wollny <gert.wollny@collabora.com>
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a
   8  * copy of this software and associated documentation files (the "Software"),
   9  * to deal in the Software without restriction, including without limitation
  10  * on the rights to use, copy, modify, merge, publish, distribute, sub
  11  * license, and/or sell copies of the Software, and to permit persons to whom
  12  * the Software is furnished to do so, subject to the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the next
  15  * paragraph) shall be included in all copies or substantial portions of the
  16  * Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  19  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  20  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  21  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
  22  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  23  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  24  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  */
  26
  27 #include "tgsi/tgsi_from_mesa.h"
  28 #include "sfn_shader_geometry.h"
  29 #include "sfn_instruction_misc.h"
  30 #include "sfn_instruction_fetch.h"
  31
  32 namespace r600 {
  33
  34 GeometryShaderFromNir::GeometryShaderFromNir(r600_pipe_shader *sh,
  35                                              r600_pipe_shader_selector &sel,
  36                                              const r600_shader_key &key):
  37    ShaderFromNirProcessor (PIPE_SHADER_GEOMETRY, sel, sh->shader,
  38                            sh->scratch_space_needed),
  39    m_pipe_shader(sh),
  40    m_so_info(&sel.so),
  41    m_first_vertex_emitted(false),
  42    m_offset(0),
  43    m_next_input_ring_offset(0),
  44    m_key(key),
  45    m_num_clip_dist(0),
  46    m_cur_ring_output(0),
  47    m_gs_tri_strip_adj_fix(false)
  48 {
  49    sh_info().atomic_base = key.gs.first_atomic_counter;
  50 }
  51
  52 bool GeometryShaderFromNir::do_emit_load_deref(UNUSED const nir_variable *in_var, UNUSED nir_intrinsic_instr* instr)
  53 {
  54    return false;
  55 }
  56
  57 bool GeometryShaderFromNir::do_emit_store_deref(const nir_variable *out_var, nir_intrinsic_instr* instr)
  58 {
  59    uint32_t write_mask =  (1 << instr->num_components) - 1;
  60    GPRVector::Swizzle swz = swizzle_from_mask(instr->num_components);
  61    std::unique_ptr<GPRVector> vec(vec_from_nir_with_fetch_constant(instr->src[1], write_mask, swz));
  62
  63    GPRVector out_value  = *vec;
  64
  65    sh_info().output[out_var->data.driver_location].write_mask =
  66          (1 << instr->num_components) - 1;
  67
  68    auto ir = new MemRingOutIntruction(cf_mem_ring, mem_write_ind, out_value,
  69                                       4 * out_var->data.driver_location,
  70                                       4, m_export_base);
  71    emit_instruction(ir);
  72
  73    return true;
  74 }
  75
  76 bool GeometryShaderFromNir::scan_sysvalue_access(UNUSED nir_instr *instr)
  77 {
  78    return true;
  79 }
  80
  81 bool GeometryShaderFromNir::do_process_inputs(nir_variable *input)
  82 {
  83
  84    if (input->data.location == VARYING_SLOT_POS ||
  85        input->data.location == VARYING_SLOT_PSIZ ||
  86        input->data.location == VARYING_SLOT_CLIP_DIST0 ||
  87        input->data.location == VARYING_SLOT_CLIP_DIST1 ||
  88        (input->data.location >= VARYING_SLOT_VAR0 &&
  89        input->data.location <= VARYING_SLOT_VAR31) ||
  90        (input->data.location >= VARYING_SLOT_TEX0 &&
  91        input->data.location <= VARYING_SLOT_TEX7)) {
  92
  93       r600_shader_io& io = sh_info().input[input->data.driver_location];
  94       tgsi_get_gl_varying_semantic(static_cast<gl_varying_slot>( input->data.location),
  95                                    true, &io.name, &io.sid);
  96       io.ring_offset = 16 * input->data.driver_location;
  97       ++sh_info().ninput;
  98       m_next_input_ring_offset += 16;
  99       return true;
 100    }
 101
 102    return false;
 103 }
 104
 105 bool GeometryShaderFromNir::do_process_outputs(nir_variable *output)
 106 {
 107    if (output->data.location == VARYING_SLOT_COL0 ||
 108        output->data.location == VARYING_SLOT_COL1 ||
 109        (output->data.location >= VARYING_SLOT_VAR0 &&
 110        output->data.location <= VARYING_SLOT_VAR31) ||
 111        (output->data.location >= VARYING_SLOT_TEX0 &&
 112        output->data.location <= VARYING_SLOT_TEX7) ||
 113        output->data.location == VARYING_SLOT_BFC0 ||
 114        output->data.location == VARYING_SLOT_BFC1 ||
 115        output->data.location == VARYING_SLOT_CLIP_VERTEX ||
 116        output->data.location == VARYING_SLOT_CLIP_DIST0 ||
 117        output->data.location == VARYING_SLOT_CLIP_DIST1 ||
 118        output->data.location == VARYING_SLOT_PRIMITIVE_ID ||
 119        output->data.location == VARYING_SLOT_POS ||
 120        output->data.location == VARYING_SLOT_PSIZ ||
 121        output->data.location == VARYING_SLOT_LAYER ||
 122        output->data.location == VARYING_SLOT_VIEWPORT ||
 123        output->data.location == VARYING_SLOT_FOGC) {
 124       r600_shader_io& io = sh_info().output[output->data.driver_location];
 125
 126       tgsi_get_gl_varying_semantic(static_cast<gl_varying_slot>( output->data.location),
 127                                    true, &io.name, &io.sid);
 128       evaluate_spi_sid(io);
 129       ++sh_info().noutput;
 130
 131       if (output->data.location == VARYING_SLOT_CLIP_DIST0 ||
 132           output->data.location == VARYING_SLOT_CLIP_DIST1) {
 133          m_num_clip_dist += 4;
 134       }
 135       return true;
 136    }
 137    return false;
 138 }
 139
 140
 141 bool GeometryShaderFromNir::allocate_reserved_registers()
 142 {
 143    const int sel[6] = {0, 0 ,0, 1, 1, 1};
 144    const int chan[6] = {0, 1 ,3, 0, 1, 2};
 145
 146    increment_reserved_registers();
 147    increment_reserved_registers();
 148
 149    /* Reserve registers used by the shaders (should check how many
 150     * components are actually used */
 151    for (int i = 0; i < 6; ++i) {
 152       auto reg = new GPRValue(sel[i], chan[i]);
 153       reg->set_as_input();
 154       m_per_vertex_offsets[i].reset(reg);
 155       inject_register(sel[i], chan[i], m_per_vertex_offsets[i], false);
 156    }
 157    auto reg = new GPRValue(0, 2);
 158    reg->set_as_input();
 159    m_primitive_id.reset(reg);
 160    inject_register(0, 2, m_primitive_id, false);
 161
 162    reg = new GPRValue(1, 3);
 163    reg->set_as_input();
 164    m_invocation_id.reset(reg);
 165    inject_register(1, 3, m_invocation_id, false);
 166
 167    m_export_base = get_temp_register();
 168    emit_instruction(new AluInstruction(op1_mov, m_export_base, Value::zero, {alu_write, alu_last_instr}));
 169
 170    sh_info().ring_item_sizes[0] = m_next_input_ring_offset;
 171
 172    if (m_key.gs.tri_strip_adj_fix)
 173       emit_adj_fix();
 174
 175    return true;
 176 }
 177
 178 void GeometryShaderFromNir::emit_adj_fix()
 179 {
 180    PValue adjhelp0(new  GPRValue(m_export_base->sel(), 1));
 181    emit_instruction(op2_and_int, adjhelp0, {m_primitive_id, Value::one_i}, {alu_write, alu_last_instr});
 182
 183    int help2 = allocate_temp_register();
 184    int reg_indices[6];
 185    int reg_chanels[6] = {0, 1, 2, 3, 2, 3};
 186
 187    int rotate_indices[6] = {4, 5, 0, 1, 2, 3};
 188
 189    reg_indices[0] = reg_indices[1] = reg_indices[2] = reg_indices[3] = help2;
 190    reg_indices[4] = reg_indices[5] = m_export_base->sel();
 191
 192    std::array<PValue, 6> adjhelp;
 193
 194    AluInstruction *ir = nullptr;
 195    for (int i = 0; i < 6; i++) {
 196       adjhelp[i].reset(new GPRValue(reg_indices[i], reg_chanels[i]));
 197       ir = new AluInstruction(op3_cnde_int, adjhelp[i],
 198                              {adjhelp0, m_per_vertex_offsets[i],
 199                               m_per_vertex_offsets[rotate_indices[i]]},
 200                              {alu_write});
 201       if (i == 3)
 202          ir->set_flag(alu_last_instr);
 203       emit_instruction(ir);
 204    }
 205    ir->set_flag(alu_last_instr);
 206
 207    for (int i = 0; i < 6; i++)
 208       m_per_vertex_offsets[i] = adjhelp[i];
 209 }
 210
 211 bool GeometryShaderFromNir::emit_deref_instruction_override(nir_deref_instr* instr)
 212 {
 213    if (instr->deref_type  == nir_deref_type_array) {
 214       auto var = get_deref_location(instr->parent);
 215       ArrayDeref ad = {var, &instr->arr.index};
 216       assert(instr->dest.is_ssa);
 217       m_in_array_deref[instr->dest.ssa.index] = ad;
 218
 219       /* Problem: nir_intrinsice_load_deref tries to lookup the
 220        * variable, and will not find it, need to override that too */
 221       return true;
 222    }
 223    return false;
 224 }
 225
 226 bool GeometryShaderFromNir::emit_intrinsic_instruction_override(nir_intrinsic_instr* instr)
 227 {
 228    switch (instr->intrinsic) {
 229    case nir_intrinsic_load_deref: {
 230       auto& src = instr->src[0];
 231       assert(src.is_ssa);
 232       auto array = m_in_array_deref.find(src.ssa->index);
 233       if (array != m_in_array_deref.end())
 234          return emit_load_from_array(instr, array->second);
 235    } break;
 236    case nir_intrinsic_emit_vertex:
 237       return emit_vertex(instr, false);
 238    case nir_intrinsic_end_primitive:
 239       return emit_vertex(instr, true);
 240    case nir_intrinsic_load_primitive_id:
 241       return load_preloaded_value(instr->dest, 0, m_primitive_id);
 242    case nir_intrinsic_load_invocation_id:
 243       return load_preloaded_value(instr->dest, 0, m_invocation_id);
 244    default:
 245       ;
 246    }
 247    return false;
 248 }
 249
 250 bool GeometryShaderFromNir::emit_vertex(nir_intrinsic_instr* instr, bool cut)
 251 {
 252    int stream = nir_intrinsic_stream_id(instr);
 253    assert(stream < 4);
 254
 255    emit_instruction(new EmitVertex(stream, cut));
 256
 257    if (!cut)
 258       emit_instruction(new AluInstruction(op2_add_int, m_export_base, m_export_base,
 259                                           PValue(new LiteralValue(sh_info().noutput)),
 260                                           {alu_write, alu_last_instr}));
 261
 262    return true;
 263 }
 264
 265 bool GeometryShaderFromNir::emit_load_from_array(nir_intrinsic_instr* instr,
 266                                                  const ArrayDeref& array_deref)
 267 {
 268    auto dest = vec_from_nir(instr->dest, instr->num_components);
 269
 270    const nir_load_const_instr* literal_index = nullptr;
 271
 272    if (array_deref.index->is_ssa)
 273       literal_index = get_literal_constant(array_deref.index->ssa->index);
 274
 275    if (!literal_index) {
 276       sfn_log << SfnLog::err << "GS: Indirect input addressing not (yet) supported\n";
 277       return false;
 278    }
 279    assert(literal_index->value[0].u32 < 6);
 280    PValue addr = m_per_vertex_offsets[literal_index->value[0].u32];
 281
 282    auto fetch = new FetchInstruction(vc_fetch, no_index_offset, dest, addr,
 283                                      16 * array_deref.var->data.driver_location,
 284                                      R600_GS_RING_CONST_BUFFER, PValue(), bim_none, true);
 285    emit_instruction(fetch);
 286    return true;
 287 }
 288
 289 void GeometryShaderFromNir::do_finalize()
 290 {
 291    if (m_num_clip_dist) {
 292       sh_info().cc_dist_mask = (1 << m_num_clip_dist) - 1;
 293       sh_info().clip_dist_write = (1 << m_num_clip_dist) - 1;
 294    }
 295 }
 296
 297 }