1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
|
#include "util/u_memory.h"
#include "../radeon/r600_cs.h"
#include "radeonsi_pipe.h"
#include "radeonsi_shader.h"
#include "radeon_llvm_util.h"
#define MAX_GLOBAL_BUFFERS 20
struct si_pipe_compute {
struct r600_context *ctx;
unsigned local_size;
unsigned private_size;
unsigned input_size;
unsigned num_kernels;
struct si_pipe_shader *kernels;
unsigned num_user_sgprs;
struct pipe_resource *global_buffers[MAX_GLOBAL_BUFFERS];
};
static void *radeonsi_create_compute_state(
struct pipe_context *ctx,
const struct pipe_compute_state *cso)
{
struct r600_context *rctx = (struct r600_context *)ctx;
struct si_pipe_compute *program =
CALLOC_STRUCT(si_pipe_compute);
const struct pipe_llvm_program_header *header;
const unsigned char *code;
unsigned i;
header = cso->prog;
code = cso->prog + sizeof(struct pipe_llvm_program_header);
program->ctx = rctx;
program->local_size = cso->req_local_mem;
program->private_size = cso->req_private_mem;
program->input_size = cso->req_input_mem;
program->num_kernels = radeon_llvm_get_num_kernels(code,
header->num_bytes);
program->kernels = CALLOC(sizeof(struct si_pipe_shader),
program->num_kernels);
for (i = 0; i < program->num_kernels; i++) {
LLVMModuleRef mod = radeon_llvm_get_kernel_module(i, code,
header->num_bytes);
si_compile_llvm(rctx, &program->kernels[i], mod);
}
return program;
}
static void radeonsi_bind_compute_state(struct pipe_context *ctx, void *state)
{
struct r600_context *rctx = (struct r600_context*)ctx;
rctx->cs_shader_state.program = (struct si_pipe_compute*)state;
}
static void radeonsi_set_global_binding(
struct pipe_context *ctx, unsigned first, unsigned n,
struct pipe_resource **resources,
uint32_t **handles)
{
unsigned i;
struct r600_context *rctx = (struct r600_context*)ctx;
struct si_pipe_compute *program = rctx->cs_shader_state.program;
if (!resources) {
for (i = first; i < first + n; i++) {
program->global_buffers[i] = NULL;
}
return;
}
for (i = first; i < first + n; i++) {
uint64_t va;
program->global_buffers[i] = resources[i];
va = r600_resource_va(ctx->screen, resources[i]);
memcpy(handles[i], &va, sizeof(va));
}
}
static void radeonsi_launch_grid(
struct pipe_context *ctx,
const uint *block_layout, const uint *grid_layout,
uint32_t pc, const void *input)
{
struct r600_context *rctx = (struct r600_context*)ctx;
struct si_pipe_compute *program = rctx->cs_shader_state.program;
struct si_pm4_state *pm4 = CALLOC_STRUCT(si_pm4_state);
struct r600_resource *kernel_args_buffer = NULL;
unsigned kernel_args_size;
unsigned num_work_size_bytes = 36;
uint32_t kernel_args_offset = 0;
uint32_t *kernel_args;
uint64_t kernel_args_va;
uint64_t shader_va;
unsigned arg_user_sgpr_count = 2;
unsigned i;
struct si_pipe_shader *shader = &program->kernels[pc];
pm4->compute_pkt = true;
si_cmd_context_control(pm4);
si_pm4_cmd_begin(pm4, PKT3_EVENT_WRITE);
si_pm4_cmd_add(pm4, EVENT_TYPE(EVENT_TYPE_CACHE_FLUSH) |
EVENT_INDEX(0x7) |
EVENT_WRITE_INV_L2);
si_pm4_cmd_end(pm4, false);
si_pm4_inval_texture_cache(pm4);
si_pm4_inval_shader_cache(pm4);
si_cmd_surface_sync(pm4, pm4->cp_coher_cntl);
/* Upload the kernel arguments */
/* The extra num_work_size_bytes are for work group / work item size information */
kernel_args_size = program->input_size + num_work_size_bytes;
kernel_args = MALLOC(kernel_args_size);
for (i = 0; i < 3; i++) {
kernel_args[i] = grid_layout[i];
kernel_args[i + 3] = grid_layout[i] * block_layout[i];
kernel_args[i + 6] = block_layout[i];
}
memcpy(kernel_args + (num_work_size_bytes / 4), input, program->input_size);
r600_upload_const_buffer(rctx, &kernel_args_buffer, (uint8_t*)kernel_args,
kernel_args_size, &kernel_args_offset);
kernel_args_va = r600_resource_va(ctx->screen,
(struct pipe_resource*)kernel_args_buffer);
kernel_args_va += kernel_args_offset;
si_pm4_add_bo(pm4, kernel_args_buffer, RADEON_USAGE_READ);
si_pm4_set_reg(pm4, R_00B900_COMPUTE_USER_DATA_0, kernel_args_va);
si_pm4_set_reg(pm4, R_00B900_COMPUTE_USER_DATA_0 + 4, S_008F04_BASE_ADDRESS_HI (kernel_args_va >> 32) | S_008F04_STRIDE(0));
si_pm4_set_reg(pm4, R_00B810_COMPUTE_START_X, 0);
si_pm4_set_reg(pm4, R_00B814_COMPUTE_START_Y, 0);
si_pm4_set_reg(pm4, R_00B818_COMPUTE_START_Z, 0);
si_pm4_set_reg(pm4, R_00B81C_COMPUTE_NUM_THREAD_X,
S_00B81C_NUM_THREAD_FULL(block_layout[0]));
si_pm4_set_reg(pm4, R_00B820_COMPUTE_NUM_THREAD_Y,
S_00B820_NUM_THREAD_FULL(block_layout[1]));
si_pm4_set_reg(pm4, R_00B824_COMPUTE_NUM_THREAD_Z,
S_00B824_NUM_THREAD_FULL(block_layout[2]));
/* Global buffers */
for (i = 0; i < MAX_GLOBAL_BUFFERS; i++) {
struct r600_resource *buffer =
(struct r600_resource*)program->global_buffers[i];
if (!buffer) {
continue;
}
si_pm4_add_bo(pm4, buffer, RADEON_USAGE_READWRITE);
}
/* XXX: This should be:
* (number of compute units) * 4 * (waves per simd) - 1 */
si_pm4_set_reg(pm4, R_00B82C_COMPUTE_MAX_WAVE_ID, 0x190 /* Default value */);
shader_va = r600_resource_va(ctx->screen, (void *)shader->bo);
si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ);
si_pm4_set_reg(pm4, R_00B830_COMPUTE_PGM_LO, (shader_va >> 8) & 0xffffffff);
si_pm4_set_reg(pm4, R_00B834_COMPUTE_PGM_HI, shader_va >> 40);
si_pm4_set_reg(pm4, R_00B848_COMPUTE_PGM_RSRC1,
/* We always use at least 3 VGPRS, these come from
* TIDIG_COMP_CNT.
* XXX: The compiler should account for this.
*/
S_00B848_VGPRS((MAX2(3, shader->num_vgprs) - 1) / 4)
/* We always use at least 4 + arg_user_sgpr_count. The 4 extra
* sgprs are from TGID_X_EN, TGID_Y_EN, TGID_Z_EN, TG_SIZE_EN
* XXX: The compiler should account for this.
*/
| S_00B848_SGPRS(((MAX2(4 + arg_user_sgpr_count,
shader->num_sgprs)) - 1) / 8))
;
si_pm4_set_reg(pm4, R_00B84C_COMPUTE_PGM_RSRC2,
S_00B84C_SCRATCH_EN(0)
| S_00B84C_USER_SGPR(arg_user_sgpr_count)
| S_00B84C_TGID_X_EN(1)
| S_00B84C_TGID_Y_EN(1)
| S_00B84C_TGID_Z_EN(1)
| S_00B84C_TG_SIZE_EN(1)
| S_00B84C_TIDIG_COMP_CNT(2)
| S_00B84C_LDS_SIZE(shader->lds_size)
| S_00B84C_EXCP_EN(0))
;
si_pm4_set_reg(pm4, R_00B854_COMPUTE_RESOURCE_LIMITS, 0);
si_pm4_set_reg(pm4, R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE0,
S_00B858_SH0_CU_EN(0xffff /* Default value */)
| S_00B858_SH1_CU_EN(0xffff /* Default value */))
;
si_pm4_set_reg(pm4, R_00B85C_COMPUTE_STATIC_THREAD_MGMT_SE1,
S_00B85C_SH0_CU_EN(0xffff /* Default value */)
| S_00B85C_SH1_CU_EN(0xffff /* Default value */))
;
si_pm4_cmd_begin(pm4, PKT3_DISPATCH_DIRECT);
si_pm4_cmd_add(pm4, grid_layout[0]); /* Thread groups DIM_X */
si_pm4_cmd_add(pm4, grid_layout[1]); /* Thread groups DIM_Y */
si_pm4_cmd_add(pm4, grid_layout[2]); /* Thread gropus DIM_Z */
si_pm4_cmd_add(pm4, 1); /* DISPATCH_INITIATOR */
si_pm4_cmd_end(pm4, false);
si_pm4_cmd_begin(pm4, PKT3_EVENT_WRITE);
si_pm4_cmd_add(pm4, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH | EVENT_INDEX(0x4)));
si_pm4_cmd_end(pm4, false);
si_pm4_inval_texture_cache(pm4);
si_pm4_inval_shader_cache(pm4);
si_cmd_surface_sync(pm4, pm4->cp_coher_cntl);
si_pm4_emit(rctx, pm4);
#if 0
fprintf(stderr, "cdw: %i\n", rctx->cs->cdw);
for (i = 0; i < rctx->cs->cdw; i++) {
fprintf(stderr, "%4i : 0x%08X\n", i, rctx->cs->buf[i]);
}
#endif
FREE(pm4);
FREE(kernel_args);
}
static void si_delete_compute_state(struct pipe_context *ctx, void* state){}
static void si_set_compute_resources(struct pipe_context * ctx_,
unsigned start, unsigned count,
struct pipe_surface ** surfaces) { }
void si_init_compute_functions(struct r600_context *rctx)
{
rctx->b.b.create_compute_state = radeonsi_create_compute_state;
rctx->b.b.delete_compute_state = si_delete_compute_state;
rctx->b.b.bind_compute_state = radeonsi_bind_compute_state;
/* ctx->context.create_sampler_view = evergreen_compute_create_sampler_view; */
rctx->b.b.set_compute_resources = si_set_compute_resources;
rctx->b.b.set_global_binding = radeonsi_set_global_binding;
rctx->b.b.launch_grid = radeonsi_launch_grid;
}
|
