xref: /aosp_15_r20/external/mesa3d/src/gallium/drivers/etnaviv/etnaviv_compiler_nir_ra.c (revision 6104692788411f58d303aa86923a9ff6ecaded22)
1 /*
2  * Copyright (c) 2019 Zodiac Inflight Innovations
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, sub license,
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
12  * next paragraph) shall be included in all copies or substantial portions
13  * of the 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 NON-INFRINGEMENT. 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  * Authors:
24  *    Jonathan Marek <[email protected]>
25  */
26 
27 #include "etnaviv_compiler_nir.h"
28 #include "util/register_allocate.h"
29 
30 /* use "r63.z" for depth reg, it will wrap around to r0.z by reg_get_base
31  * (fs registers are offset by 1 to avoid reserving r0)
32  */
33 #define REG_FRAG_DEPTH ((ETNA_MAX_TEMPS - 1) * NUM_REG_TYPES + REG_TYPE_VIRT_SCALAR_Z)
34 
35 /* precomputed by register_allocate  */
36 static unsigned int *q_values[] = {
37    (unsigned int[]) {1, 2, 3, 4, 2, 2, 3, },
38    (unsigned int[]) {3, 5, 6, 6, 5, 5, 6, },
39    (unsigned int[]) {3, 4, 4, 4, 4, 4, 4, },
40    (unsigned int[]) {1, 1, 1, 1, 1, 1, 1, },
41    (unsigned int[]) {1, 2, 2, 2, 1, 2, 2, },
42    (unsigned int[]) {2, 3, 3, 3, 2, 3, 3, },
43    (unsigned int[]) {2, 2, 2, 2, 2, 2, 2, },
44 };
45 
reg_get_class(int virt_reg)46 static inline int reg_get_class(int virt_reg)
47 {
48    switch (reg_get_type(virt_reg)) {
49    case REG_TYPE_VEC4:
50       return REG_CLASS_VEC4;
51    case REG_TYPE_VIRT_VEC3_XYZ:
52    case REG_TYPE_VIRT_VEC3_XYW:
53    case REG_TYPE_VIRT_VEC3_XZW:
54    case REG_TYPE_VIRT_VEC3_YZW:
55       return REG_CLASS_VIRT_VEC3;
56    case REG_TYPE_VIRT_VEC2_XY:
57    case REG_TYPE_VIRT_VEC2_XZ:
58    case REG_TYPE_VIRT_VEC2_XW:
59    case REG_TYPE_VIRT_VEC2_YZ:
60    case REG_TYPE_VIRT_VEC2_YW:
61    case REG_TYPE_VIRT_VEC2_ZW:
62       return REG_CLASS_VIRT_VEC2;
63    case REG_TYPE_VIRT_SCALAR_X:
64    case REG_TYPE_VIRT_SCALAR_Y:
65    case REG_TYPE_VIRT_SCALAR_Z:
66    case REG_TYPE_VIRT_SCALAR_W:
67       return REG_CLASS_VIRT_SCALAR;
68    case REG_TYPE_VIRT_VEC2T_XY:
69    case REG_TYPE_VIRT_VEC2T_ZW:
70       return REG_CLASS_VIRT_VEC2T;
71    case REG_TYPE_VIRT_VEC2C_XY:
72    case REG_TYPE_VIRT_VEC2C_YZ:
73    case REG_TYPE_VIRT_VEC2C_ZW:
74       return REG_CLASS_VIRT_VEC2C;
75    case REG_TYPE_VIRT_VEC3C_XYZ:
76    case REG_TYPE_VIRT_VEC3C_YZW:
77       return REG_CLASS_VIRT_VEC3C;
78    }
79 
80    assert(false);
81    return 0;
82 }
83 
84 struct ra_regs *
etna_ra_setup(void * mem_ctx)85 etna_ra_setup(void *mem_ctx)
86 {
87    struct ra_regs *regs = ra_alloc_reg_set(mem_ctx, ETNA_MAX_TEMPS *
88                   NUM_REG_TYPES, false);
89 
90    /* classes always be created from index 0, so equal to the class enum
91     * which represents a register with (c+1) components
92     */
93    struct ra_class *classes[NUM_REG_CLASSES];
94    for (int c = 0; c < NUM_REG_CLASSES; c++)
95       classes[c] = ra_alloc_reg_class(regs);
96    /* add each register of each class */
97    for (int r = 0; r < NUM_REG_TYPES * ETNA_MAX_TEMPS; r++)
98       ra_class_add_reg(classes[reg_get_class(r)], r);
99    /* set conflicts */
100    for (int r = 0; r < ETNA_MAX_TEMPS; r++) {
101       for (int i = 0; i < NUM_REG_TYPES; i++) {
102          for (int j = 0; j < i; j++) {
103             if (reg_writemask[i] & reg_writemask[j]) {
104                ra_add_reg_conflict(regs, NUM_REG_TYPES * r + i,
105                                          NUM_REG_TYPES * r + j);
106             }
107          }
108       }
109    }
110    ra_set_finalize(regs, q_values);
111 
112    return regs;
113 }
114 
115 void
etna_ra_assign(struct etna_compile * c,nir_shader * shader)116 etna_ra_assign(struct etna_compile *c, nir_shader *shader)
117 {
118    struct etna_compiler *compiler = c->variant->shader->compiler;
119    struct ra_regs *regs = compiler->regs;
120 
121    nir_function_impl *impl = nir_shader_get_entrypoint(shader);
122 
123    /* liveness and interference */
124 
125    nir_index_blocks(impl);
126    nir_index_ssa_defs(impl);
127 
128    nir_foreach_function_impl(impl, shader) {
129       nir_foreach_block(block, impl) {
130          nir_foreach_instr(instr, block) {
131             instr->pass_flags &= ~PASS_FLAGS_IS_DEAD_MASK;
132          }
133       }
134    }
135 
136    /* this gives an approximation/upper limit on how many nodes are needed
137     * (some ssa values do not represent an allocated register)
138     */
139    unsigned max_nodes = impl->ssa_alloc;
140    unsigned *live_map = ralloc_array(NULL, unsigned, max_nodes);
141    memset(live_map, 0xff, sizeof(unsigned) * max_nodes);
142    struct live_def *defs = rzalloc_array(NULL, struct live_def, max_nodes);
143 
144    unsigned num_nodes = etna_live_defs(impl, defs, live_map);
145    struct ra_graph *g = ra_alloc_interference_graph(regs, num_nodes);
146 
147    /* set classes from num_components */
148    for (unsigned i = 0; i < num_nodes; i++) {
149       nir_instr *instr = defs[i].instr;
150       nir_def *def = defs[i].def;
151       unsigned comp = def->num_components - 1;
152 
153       if (instr->type == nir_instr_type_alu &&
154           c->specs->has_new_transcendentals) {
155          switch (nir_instr_as_alu(instr)->op) {
156          case nir_op_fdiv:
157          case nir_op_flog2:
158          case nir_op_fsin:
159          case nir_op_fcos:
160             comp = REG_CLASS_VIRT_VEC2T;
161             break;
162          default:
163             break;
164          }
165       }
166 
167       if (instr->type == nir_instr_type_intrinsic) {
168          nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
169          /* can't have dst swizzle or sparse writemask on UBO loads */
170          if (intr->intrinsic == nir_intrinsic_load_ubo) {
171             assert(def == &intr->def);
172             if (def->num_components == 2)
173                comp = REG_CLASS_VIRT_VEC2C;
174             if (def->num_components == 3)
175                comp = REG_CLASS_VIRT_VEC3C;
176          }
177       }
178 
179       ra_set_node_class(g, i, ra_get_class_from_index(regs, comp));
180    }
181 
182    nir_foreach_block(block, impl) {
183       nir_foreach_instr(instr, block) {
184          if (instr->type != nir_instr_type_intrinsic)
185             continue;
186 
187          nir_def *def = def_for_instr(instr);
188          nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
189          unsigned reg;
190 
191          switch (intr->intrinsic) {
192          case nir_intrinsic_store_deref: {
193             /* don't want outputs to be swizzled
194              * TODO: better would be to set the type to X/XY/XYZ/XYZW
195              * TODO: what if fragcoord.z is read after writing fragdepth?
196              */
197             nir_deref_instr *deref = nir_src_as_deref(intr->src[0]);
198             unsigned index = live_map[src_index(impl, &intr->src[1])];
199 
200             if (shader->info.stage == MESA_SHADER_FRAGMENT &&
201                 deref->var->data.location == FRAG_RESULT_DEPTH) {
202                ra_set_node_reg(g, index, REG_FRAG_DEPTH);
203             } else {
204                ra_set_node_class(g, index, ra_get_class_from_index(regs, REG_CLASS_VEC4));
205             }
206          } continue;
207          case nir_intrinsic_load_input:
208             reg = nir_intrinsic_base(intr) * NUM_REG_TYPES + (unsigned[]) {
209                REG_TYPE_VIRT_SCALAR_X,
210                REG_TYPE_VIRT_VEC2_XY,
211                REG_TYPE_VIRT_VEC3_XYZ,
212                REG_TYPE_VEC4,
213             }[def->num_components - 1];
214             break;
215          case nir_intrinsic_load_instance_id:
216             reg = c->variant->infile.num_reg * NUM_REG_TYPES + REG_TYPE_VIRT_SCALAR_Y;
217             break;
218          case nir_intrinsic_load_vertex_id:
219             reg = c->variant->infile.num_reg * NUM_REG_TYPES + REG_TYPE_VIRT_SCALAR_X;
220             break;
221          default:
222             continue;
223          }
224 
225          ra_set_node_reg(g, live_map[def_index(impl, def)], reg);
226       }
227    }
228 
229    /* add interference for intersecting live ranges */
230    for (unsigned i = 0; i < num_nodes; i++) {
231       assert(defs[i].live_start < defs[i].live_end);
232       for (unsigned j = 0; j < i; j++) {
233          if (defs[i].live_start >= defs[j].live_end || defs[j].live_start >= defs[i].live_end)
234             continue;
235          ra_add_node_interference(g, i, j);
236       }
237    }
238 
239    ralloc_free(defs);
240 
241    /* Allocate registers */
242    ASSERTED bool ok = ra_allocate(g);
243    assert(ok);
244 
245    c->g = g;
246    c->live_map = live_map;
247    c->num_nodes = num_nodes;
248 }
249 
250 unsigned
etna_ra_finish(struct etna_compile * c)251 etna_ra_finish(struct etna_compile *c)
252 {
253    /* TODO: better way to get number of registers used? */
254    unsigned j = 0;
255    for (unsigned i = 0; i < c->num_nodes; i++) {
256       j = MAX2(j, reg_get_base(c, ra_get_node_reg(c->g, i)) + 1);
257    }
258 
259    ralloc_free(c->g);
260    ralloc_free(c->live_map);
261 
262    return j;
263 }
264