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
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
|
#include "r600_llvm.h"
#include "gallivm/lp_bld_const.h"
#include "gallivm/lp_bld_intr.h"
#include "gallivm/lp_bld_gather.h"
#include "tgsi/tgsi_parse.h"
#include "util/u_double_list.h"
#include "r600.h"
#include "r600_asm.h"
#include "r600_opcodes.h"
#include "r600_shader.h"
#include "radeon_llvm.h"
#include "radeon_llvm_emit.h"
#include <stdio.h>
static LLVMValueRef llvm_fetch_const(
struct lp_build_tgsi_context * bld_base,
const struct tgsi_full_src_register *reg,
enum tgsi_opcode_type type,
unsigned swizzle)
{
LLVMValueRef cval = lp_build_intrinsic_unary(bld_base->base.gallivm->builder,
"llvm.AMDGPU.load.const", bld_base->base.elem_type,
lp_build_const_int32(bld_base->base.gallivm,
radeon_llvm_reg_index_soa(reg->Register.Index, swizzle)));
return bitcast(bld_base, type, cval);
}
static void llvm_load_input(
struct radeon_llvm_context * ctx,
unsigned input_index,
const struct tgsi_full_declaration *decl)
{
unsigned chan;
for (chan = 0; chan < 4; chan++) {
unsigned soa_index = radeon_llvm_reg_index_soa(input_index,
chan);
/* The * 4 is assuming that we are in soa mode. */
LLVMValueRef reg = lp_build_const_int32(
ctx->soa.bld_base.base.gallivm,
soa_index + (ctx->reserved_reg_count * 4));
ctx->inputs[soa_index] = lp_build_intrinsic_unary(
ctx->soa.bld_base.base.gallivm->builder,
"llvm.R600.load.input",
ctx->soa.bld_base.base.elem_type, reg);
}
}
static void llvm_emit_prologue(struct lp_build_tgsi_context * bld_base)
{
struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
struct lp_build_context * base = &bld_base->base;
unsigned i;
/* Reserve special input registers */
for (i = 0; i < ctx->reserved_reg_count; i++) {
unsigned chan;
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
LLVMValueRef reg;
LLVMValueRef reg_index = lp_build_const_int32(
base->gallivm,
radeon_llvm_reg_index_soa(i, chan));
reg = lp_build_intrinsic_unary(base->gallivm->builder,
"llvm.AMDGPU.reserve.reg",
base->elem_type, reg_index);
lp_build_intrinsic_unary(base->gallivm->builder,
"llvm.AMDGPU.export.reg",
LLVMVoidTypeInContext(base->gallivm->context),
reg);
}
}
}
static void llvm_emit_epilogue(struct lp_build_tgsi_context * bld_base)
{
struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
struct lp_build_context * base = &bld_base->base;
unsigned i;
/* Add the necessary export instructions */
for (i = 0; i < ctx->output_reg_count; i++) {
unsigned chan;
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
LLVMValueRef output;
LLVMValueRef store_output;
unsigned adjusted_reg_idx = i +
ctx->reserved_reg_count;
LLVMValueRef reg_index = lp_build_const_int32(
base->gallivm,
radeon_llvm_reg_index_soa(adjusted_reg_idx, chan));
output = LLVMBuildLoad(base->gallivm->builder,
ctx->soa.outputs[i][chan], "");
store_output = lp_build_intrinsic_binary(
base->gallivm->builder,
"llvm.AMDGPU.store.output",
base->elem_type,
output, reg_index);
lp_build_intrinsic_unary(base->gallivm->builder,
"llvm.AMDGPU.export.reg",
LLVMVoidTypeInContext(base->gallivm->context),
store_output);
}
}
}
static void llvm_emit_tex(
const struct lp_build_tgsi_action * action,
struct lp_build_tgsi_context * bld_base,
struct lp_build_emit_data * emit_data)
{
struct gallivm_state * gallivm = bld_base->base.gallivm;
LLVMValueRef args[3];
args[0] = emit_data->args[0];
args[1] = lp_build_const_int32(gallivm,
emit_data->inst->Src[1].Register.Index);
args[2] = lp_build_const_int32(gallivm,
emit_data->inst->Texture.Texture);
emit_data->output[0] = lp_build_intrinsic(gallivm->builder,
action->intr_name,
emit_data->dst_type, args, 3);
}
static void dp_fetch_args(
struct lp_build_tgsi_context * bld_base,
struct lp_build_emit_data * emit_data)
{
struct lp_build_context * base = &bld_base->base;
unsigned chan;
LLVMValueRef elements[2][4];
unsigned opcode = emit_data->inst->Instruction.Opcode;
unsigned dp_components = (opcode == TGSI_OPCODE_DP2 ? 2 :
(opcode == TGSI_OPCODE_DP3 ? 3 : 4));
for (chan = 0 ; chan < dp_components; chan++) {
elements[0][chan] = lp_build_emit_fetch(bld_base,
emit_data->inst, 0, chan);
elements[1][chan] = lp_build_emit_fetch(bld_base,
emit_data->inst, 1, chan);
}
for ( ; chan < 4; chan++) {
elements[0][chan] = base->zero;
elements[1][chan] = base->zero;
}
/* Fix up for DPH */
if (opcode == TGSI_OPCODE_DPH) {
elements[0][TGSI_CHAN_W] = base->one;
}
emit_data->args[0] = lp_build_gather_values(bld_base->base.gallivm,
elements[0], 4);
emit_data->args[1] = lp_build_gather_values(bld_base->base.gallivm,
elements[1], 4);
emit_data->arg_count = 2;
emit_data->dst_type = base->elem_type;
}
static struct lp_build_tgsi_action dot_action = {
.fetch_args = dp_fetch_args,
.emit = lp_build_tgsi_intrinsic,
.intr_name = "llvm.AMDGPU.dp4"
};
static void txp_fetch_args(
struct lp_build_tgsi_context * bld_base,
struct lp_build_emit_data * emit_data)
{
LLVMValueRef src_w;
unsigned chan;
LLVMValueRef coords[4];
emit_data->dst_type = LLVMVectorType(bld_base->base.elem_type, 4);
src_w = lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_W);
for (chan = 0; chan < 3; chan++ ) {
LLVMValueRef arg = lp_build_emit_fetch(bld_base,
emit_data->inst, 0, chan);
coords[chan] = lp_build_emit_llvm_binary(bld_base,
TGSI_OPCODE_DIV, arg, src_w);
}
coords[3] = bld_base->base.one;
emit_data->args[0] = lp_build_gather_values(bld_base->base.gallivm,
coords, 4);
emit_data->arg_count = 1;
}
LLVMModuleRef r600_tgsi_llvm(
struct radeon_llvm_context * ctx,
const struct tgsi_token * tokens)
{
struct tgsi_shader_info shader_info;
struct lp_build_tgsi_context * bld_base = &ctx->soa.bld_base;
radeon_llvm_context_init(ctx);
tgsi_scan_shader(tokens, &shader_info);
bld_base->info = &shader_info;
bld_base->userdata = ctx;
bld_base->emit_fetch_funcs[TGSI_FILE_CONSTANT] = llvm_fetch_const;
bld_base->emit_prologue = llvm_emit_prologue;
bld_base->emit_epilogue = llvm_emit_epilogue;
ctx->userdata = ctx;
ctx->load_input = llvm_load_input;
bld_base->op_actions[TGSI_OPCODE_DP2] = dot_action;
bld_base->op_actions[TGSI_OPCODE_DP3] = dot_action;
bld_base->op_actions[TGSI_OPCODE_DP4] = dot_action;
bld_base->op_actions[TGSI_OPCODE_DPH] = dot_action;
bld_base->op_actions[TGSI_OPCODE_TEX].emit = llvm_emit_tex;
bld_base->op_actions[TGSI_OPCODE_TXB].emit = llvm_emit_tex;
bld_base->op_actions[TGSI_OPCODE_TXD].emit = llvm_emit_tex;
bld_base->op_actions[TGSI_OPCODE_TXL].emit = llvm_emit_tex;
bld_base->op_actions[TGSI_OPCODE_TXP].fetch_args = txp_fetch_args;
bld_base->op_actions[TGSI_OPCODE_TXP].emit = llvm_emit_tex;
lp_build_tgsi_llvm(bld_base, tokens);
radeon_llvm_finalize_module(ctx);
return ctx->gallivm.module;
}
const char * r600_llvm_gpu_string(enum radeon_family family)
{
const char * gpu_family;
switch (family) {
case CHIP_R600:
case CHIP_RV610:
case CHIP_RV630:
case CHIP_RV620:
case CHIP_RV635:
case CHIP_RS780:
case CHIP_RS880:
case CHIP_RV710:
gpu_family = "rv710";
break;
case CHIP_RV730:
gpu_family = "rv730";
break;
case CHIP_RV670:
case CHIP_RV740:
case CHIP_RV770:
gpu_family = "rv770";
break;
case CHIP_PALM:
case CHIP_CEDAR:
gpu_family = "cedar";
break;
case CHIP_SUMO:
case CHIP_SUMO2:
case CHIP_REDWOOD:
gpu_family = "redwood";
break;
case CHIP_JUNIPER:
gpu_family = "juniper";
break;
case CHIP_HEMLOCK:
case CHIP_CYPRESS:
gpu_family = "cypress";
break;
case CHIP_BARTS:
gpu_family = "barts";
break;
case CHIP_TURKS:
gpu_family = "turks";
break;
case CHIP_CAICOS:
gpu_family = "caicos";
break;
case CHIP_CAYMAN:
case CHIP_ARUBA:
gpu_family = "cayman";
break;
default:
gpu_family = "";
fprintf(stderr, "Chip not supported by r600 llvm "
"backend, please file a bug at bugs.freedesktop.org\n");
break;
}
return gpu_family;
}
unsigned r600_llvm_compile(
LLVMModuleRef mod,
unsigned char ** inst_bytes,
unsigned * inst_byte_count,
enum radeon_family family,
unsigned dump)
{
const char * gpu_family = r600_llvm_gpu_string(family);
return radeon_llvm_compile(mod, inst_bytes, inst_byte_count,
gpu_family, dump);
}
|
