• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright 2020 Advanced Micro Devices, Inc.
3  * All Rights Reserved.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * on the rights to use, copy, modify, merge, publish, distribute, sub
9  * license, and/or sell copies of the Software, and to permit persons to whom
10  * the Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice (including the next
13  * paragraph) shall be included in all copies or substantial portions of the
14  * Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22  * USE OR OTHER DEALINGS IN THE SOFTWARE.
23  */
24 
25 #include "si_build_pm4.h"
26 #include "ac_debug.h"
27 #include "ac_shadowed_regs.h"
28 #include "util/u_memory.h"
29 
si_build_load_reg(struct si_screen * sscreen,struct si_pm4_state * pm4,enum ac_reg_range_type type,struct si_resource * shadow_regs)30 static void si_build_load_reg(struct si_screen *sscreen, struct si_pm4_state *pm4,
31                               enum ac_reg_range_type type,
32                               struct si_resource *shadow_regs)
33 {
34    uint64_t gpu_address = shadow_regs->gpu_address;
35    unsigned packet, num_ranges, offset;
36    const struct ac_reg_range *ranges;
37 
38    ac_get_reg_ranges(sscreen->info.chip_class, sscreen->info.family,
39                      type, &num_ranges, &ranges);
40 
41    switch (type) {
42    case SI_REG_RANGE_UCONFIG:
43       gpu_address += SI_SHADOWED_UCONFIG_REG_OFFSET;
44       offset = CIK_UCONFIG_REG_OFFSET;
45       packet = PKT3_LOAD_UCONFIG_REG;
46       break;
47    case SI_REG_RANGE_CONTEXT:
48       gpu_address += SI_SHADOWED_CONTEXT_REG_OFFSET;
49       offset = SI_CONTEXT_REG_OFFSET;
50       packet = PKT3_LOAD_CONTEXT_REG;
51       break;
52    default:
53       gpu_address += SI_SHADOWED_SH_REG_OFFSET;
54       offset = SI_SH_REG_OFFSET;
55       packet = PKT3_LOAD_SH_REG;
56       break;
57    }
58 
59    si_pm4_cmd_add(pm4, PKT3(packet, 1 + num_ranges * 2, 0));
60    si_pm4_cmd_add(pm4, gpu_address);
61    si_pm4_cmd_add(pm4, gpu_address >> 32);
62    for (unsigned i = 0; i < num_ranges; i++) {
63       si_pm4_cmd_add(pm4, (ranges[i].offset - offset) / 4);
64       si_pm4_cmd_add(pm4, ranges[i].size / 4);
65    }
66 }
67 
68 static struct si_pm4_state *
si_create_shadowing_ib_preamble(struct si_context * sctx)69 si_create_shadowing_ib_preamble(struct si_context *sctx)
70 {
71    struct si_pm4_state *pm4 = CALLOC_STRUCT(si_pm4_state);
72 
73    if (sctx->chip_class == GFX10) {
74       /* SQ_NON_EVENT must be emitted before GE_PC_ALLOC is written. */
75       si_pm4_cmd_add(pm4, PKT3(PKT3_EVENT_WRITE, 0, 0));
76       si_pm4_cmd_add(pm4, EVENT_TYPE(V_028A90_SQ_NON_EVENT) | EVENT_INDEX(0));
77    }
78 
79    if (sctx->screen->dpbb_allowed) {
80       si_pm4_cmd_add(pm4, PKT3(PKT3_EVENT_WRITE, 0, 0));
81       si_pm4_cmd_add(pm4, EVENT_TYPE(V_028A90_BREAK_BATCH) | EVENT_INDEX(0));
82    }
83 
84    /* Wait for idle, because we'll update VGT ring pointers. */
85    si_pm4_cmd_add(pm4, PKT3(PKT3_EVENT_WRITE, 0, 0));
86    si_pm4_cmd_add(pm4, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH) | EVENT_INDEX(4));
87 
88    /* VGT_FLUSH is required even if VGT is idle. It resets VGT pointers. */
89    si_pm4_cmd_add(pm4, PKT3(PKT3_EVENT_WRITE, 0, 0));
90    si_pm4_cmd_add(pm4, EVENT_TYPE(V_028A90_VGT_FLUSH) | EVENT_INDEX(0));
91 
92    if (sctx->chip_class >= GFX10) {
93       unsigned gcr_cntl = S_586_GL2_INV(1) | S_586_GL2_WB(1) |
94                           S_586_GLM_INV(1) | S_586_GLM_WB(1) |
95                           S_586_GL1_INV(1) | S_586_GLV_INV(1) |
96                           S_586_GLK_INV(1) | S_586_GLI_INV(V_586_GLI_ALL);
97 
98       si_pm4_cmd_add(pm4, PKT3(PKT3_ACQUIRE_MEM, 6, 0));
99       si_pm4_cmd_add(pm4, 0);		/* CP_COHER_CNTL */
100       si_pm4_cmd_add(pm4, 0xffffffff);  /* CP_COHER_SIZE */
101       si_pm4_cmd_add(pm4, 0xffffff);	/* CP_COHER_SIZE_HI */
102       si_pm4_cmd_add(pm4, 0);		/* CP_COHER_BASE */
103       si_pm4_cmd_add(pm4, 0);		/* CP_COHER_BASE_HI */
104       si_pm4_cmd_add(pm4, 0x0000000A);  /* POLL_INTERVAL */
105       si_pm4_cmd_add(pm4, gcr_cntl);	/* GCR_CNTL */
106    } else if (sctx->chip_class == GFX9) {
107       unsigned cp_coher_cntl = S_0301F0_SH_ICACHE_ACTION_ENA(1) |
108                                S_0301F0_SH_KCACHE_ACTION_ENA(1) |
109                                S_0301F0_TC_ACTION_ENA(1) |
110                                S_0301F0_TCL1_ACTION_ENA(1) |
111                                S_0301F0_TC_WB_ACTION_ENA(1);
112 
113       si_pm4_cmd_add(pm4, PKT3(PKT3_ACQUIRE_MEM, 5, 0));
114       si_pm4_cmd_add(pm4, cp_coher_cntl); /* CP_COHER_CNTL */
115       si_pm4_cmd_add(pm4, 0xffffffff);    /* CP_COHER_SIZE */
116       si_pm4_cmd_add(pm4, 0xffffff);	  /* CP_COHER_SIZE_HI */
117       si_pm4_cmd_add(pm4, 0);		  /* CP_COHER_BASE */
118       si_pm4_cmd_add(pm4, 0);		  /* CP_COHER_BASE_HI */
119       si_pm4_cmd_add(pm4, 0x0000000A);    /* POLL_INTERVAL */
120    } else {
121       unreachable("invalid chip");
122    }
123 
124    si_pm4_cmd_add(pm4, PKT3(PKT3_PFP_SYNC_ME, 0, 0));
125    si_pm4_cmd_add(pm4, 0);
126 
127    si_pm4_cmd_add(pm4, PKT3(PKT3_CONTEXT_CONTROL, 1, 0));
128    si_pm4_cmd_add(pm4,
129                   CC0_UPDATE_LOAD_ENABLES(1) |
130                   CC0_LOAD_PER_CONTEXT_STATE(1) |
131                   CC0_LOAD_CS_SH_REGS(1) |
132                   CC0_LOAD_GFX_SH_REGS(1) |
133                   CC0_LOAD_GLOBAL_UCONFIG(1));
134    si_pm4_cmd_add(pm4,
135                   CC1_UPDATE_SHADOW_ENABLES(1) |
136                   CC1_SHADOW_PER_CONTEXT_STATE(1) |
137                   CC1_SHADOW_CS_SH_REGS(1) |
138                   CC1_SHADOW_GFX_SH_REGS(1) |
139                   CC1_SHADOW_GLOBAL_UCONFIG(1));
140 
141    for (unsigned i = 0; i < SI_NUM_SHADOWED_REG_RANGES; i++)
142       si_build_load_reg(sctx->screen, pm4, i, sctx->shadowed_regs);
143 
144    return pm4;
145 }
146 
si_init_cp_reg_shadowing(struct si_context * sctx)147 void si_init_cp_reg_shadowing(struct si_context *sctx)
148 {
149    if (sctx->screen->info.mid_command_buffer_preemption_enabled ||
150        sctx->screen->debug_flags & DBG(SHADOW_REGS)) {
151       sctx->shadowed_regs =
152             si_aligned_buffer_create(sctx->b.screen,
153                                      SI_RESOURCE_FLAG_UNMAPPABLE | SI_RESOURCE_FLAG_DRIVER_INTERNAL,
154                                      PIPE_USAGE_DEFAULT,
155                                      SI_SHADOWED_REG_BUFFER_SIZE,
156                                      4096);
157       if (!sctx->shadowed_regs)
158          fprintf(stderr, "radeonsi: cannot create a shadowed_regs buffer\n");
159    }
160 
161    si_init_cs_preamble_state(sctx, sctx->shadowed_regs != NULL);
162 
163    if (sctx->shadowed_regs) {
164       /* We need to clear the shadowed reg buffer. */
165       si_cp_dma_clear_buffer(sctx, sctx->gfx_cs, &sctx->shadowed_regs->b.b,
166                              0, sctx->shadowed_regs->bo_size, 0, 0, SI_COHERENCY_CP,
167                              L2_BYPASS);
168 
169       /* Create the shadowing preamble. */
170       struct si_pm4_state *shadowing_preamble =
171             si_create_shadowing_ib_preamble(sctx);
172 
173       /* Initialize shadowed registers as follows. */
174       radeon_add_to_buffer_list(sctx, sctx->gfx_cs, sctx->shadowed_regs,
175                                 RADEON_USAGE_READWRITE, RADEON_PRIO_DESCRIPTORS);
176       si_pm4_emit(sctx, shadowing_preamble);
177       ac_emulate_clear_state(&sctx->screen->info, sctx->gfx_cs,
178                              radeon_set_context_reg_seq_array);
179       si_pm4_emit(sctx, sctx->cs_preamble_state);
180 
181       /* The register values are shadowed, so we won't need to set them again. */
182       si_pm4_free_state(sctx, sctx->cs_preamble_state, ~0);
183       sctx->cs_preamble_state = NULL;
184 
185       si_set_tracked_regs_to_clear_state(sctx);
186 
187       /* Setup preemption. The shadowing preamble will be executed as a preamble IB,
188        * which will load register values from memory on a context switch.
189        */
190       sctx->ws->cs_setup_preemption(sctx->gfx_cs, shadowing_preamble->pm4,
191                                     shadowing_preamble->ndw);
192       si_pm4_free_state(sctx, shadowing_preamble, ~0);
193    }
194 }
195