3 #include "gallivm/lp_bld_const.h"
4 #include "gallivm/lp_bld_intr.h"
5 #include "gallivm/lp_bld_gather.h"
6 #include "tgsi/tgsi_parse.h"
7 #include "util/u_double_list.h"
8 #include "util/u_memory.h"
12 #include "r600_opcodes.h"
13 #include "r600_shader.h"
14 #include "radeon_llvm.h"
15 #include "radeon_llvm_emit.h"
19 #if defined R600_USE_LLVM || defined HAVE_OPENCL
21 static LLVMValueRef
llvm_fetch_const(
22 struct lp_build_tgsi_context
* bld_base
,
23 const struct tgsi_full_src_register
*reg
,
24 enum tgsi_opcode_type type
,
27 LLVMValueRef idx
= lp_build_const_int32(bld_base
->base
.gallivm
,
28 radeon_llvm_reg_index_soa(reg
->Register
.Index
, swizzle
));
29 LLVMValueRef cval
= build_intrinsic(bld_base
->base
.gallivm
->builder
,
30 "llvm.AMDGPU.load.const", bld_base
->base
.elem_type
,
31 &idx
, 1, LLVMReadNoneAttribute
);
33 return bitcast(bld_base
, type
, cval
);
36 static void llvm_load_system_value(
37 struct radeon_llvm_context
* ctx
,
39 const struct tgsi_full_declaration
*decl
)
43 switch (decl
->Semantic
.Name
) {
44 case TGSI_SEMANTIC_INSTANCEID
: chan
= 3; break;
45 case TGSI_SEMANTIC_VERTEXID
: chan
= 0; break;
46 default: assert(!"unknown system value");
49 LLVMValueRef reg
= lp_build_const_int32(
50 ctx
->soa
.bld_base
.base
.gallivm
, chan
);
51 ctx
->system_values
[index
] = build_intrinsic(
52 ctx
->soa
.bld_base
.base
.gallivm
->builder
,
53 "llvm.R600.load.input",
54 ctx
->soa
.bld_base
.base
.elem_type
, ®
, 1,
55 LLVMReadNoneAttribute
);
58 static LLVMValueRef
llvm_fetch_system_value(
59 struct lp_build_tgsi_context
* bld_base
,
60 const struct tgsi_full_src_register
*reg
,
61 enum tgsi_opcode_type type
,
64 struct radeon_llvm_context
* ctx
= radeon_llvm_context(bld_base
);
65 LLVMValueRef cval
= ctx
->system_values
[reg
->Register
.Index
];
66 return bitcast(bld_base
, type
, cval
);
69 static void llvm_load_input(
70 struct radeon_llvm_context
* ctx
,
72 const struct tgsi_full_declaration
*decl
)
76 for (chan
= 0; chan
< 4; chan
++) {
77 unsigned soa_index
= radeon_llvm_reg_index_soa(input_index
,
80 /* The * 4 is assuming that we are in soa mode. */
81 LLVMValueRef reg
= lp_build_const_int32(
82 ctx
->soa
.bld_base
.base
.gallivm
,
83 soa_index
+ (ctx
->reserved_reg_count
* 4));
84 ctx
->inputs
[soa_index
] = build_intrinsic(
85 ctx
->soa
.bld_base
.base
.gallivm
->builder
,
86 "llvm.R600.load.input",
87 ctx
->soa
.bld_base
.base
.elem_type
, ®
, 1,
88 LLVMReadNoneAttribute
);
92 static void llvm_emit_prologue(struct lp_build_tgsi_context
* bld_base
)
94 struct radeon_llvm_context
* ctx
= radeon_llvm_context(bld_base
);
95 struct lp_build_context
* base
= &bld_base
->base
;
98 /* Reserve special input registers */
99 for (i
= 0; i
< ctx
->reserved_reg_count
; i
++) {
101 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
102 LLVMValueRef reg_index
= lp_build_const_int32(
104 radeon_llvm_reg_index_soa(i
, chan
));
105 lp_build_intrinsic_unary(base
->gallivm
->builder
,
106 "llvm.AMDGPU.reserve.reg",
107 LLVMVoidTypeInContext(base
->gallivm
->context
),
113 static void llvm_emit_epilogue(struct lp_build_tgsi_context
* bld_base
)
115 struct radeon_llvm_context
* ctx
= radeon_llvm_context(bld_base
);
116 struct lp_build_context
* base
= &bld_base
->base
;
119 /* Add the necessary export instructions */
120 for (i
= 0; i
< ctx
->output_reg_count
; i
++) {
122 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
124 unsigned adjusted_reg_idx
= i
+
125 ctx
->reserved_reg_count
;
126 LLVMValueRef reg_index
= lp_build_const_int32(
128 radeon_llvm_reg_index_soa(adjusted_reg_idx
, chan
));
130 output
= LLVMBuildLoad(base
->gallivm
->builder
,
131 ctx
->soa
.outputs
[i
][chan
], "");
133 lp_build_intrinsic_binary(
134 base
->gallivm
->builder
,
135 "llvm.AMDGPU.store.output",
136 LLVMVoidTypeInContext(base
->gallivm
->context
),
142 static void llvm_emit_tex(
143 const struct lp_build_tgsi_action
* action
,
144 struct lp_build_tgsi_context
* bld_base
,
145 struct lp_build_emit_data
* emit_data
)
147 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
148 LLVMValueRef args
[6];
149 unsigned c
, sampler_src
;
151 assert(emit_data
->arg_count
+ 2 <= Elements(args
));
153 for (c
= 0; c
< emit_data
->arg_count
; ++c
)
154 args
[c
] = emit_data
->args
[c
];
156 sampler_src
= emit_data
->inst
->Instruction
.NumSrcRegs
-1;
158 args
[c
++] = lp_build_const_int32(gallivm
,
159 emit_data
->inst
->Src
[sampler_src
].Register
.Index
);
160 args
[c
++] = lp_build_const_int32(gallivm
,
161 emit_data
->inst
->Texture
.Texture
);
163 emit_data
->output
[0] = build_intrinsic(gallivm
->builder
,
165 emit_data
->dst_type
, args
, c
, LLVMReadNoneAttribute
);
168 static void dp_fetch_args(
169 struct lp_build_tgsi_context
* bld_base
,
170 struct lp_build_emit_data
* emit_data
)
172 struct lp_build_context
* base
= &bld_base
->base
;
174 LLVMValueRef elements
[2][4];
175 unsigned opcode
= emit_data
->inst
->Instruction
.Opcode
;
176 unsigned dp_components
= (opcode
== TGSI_OPCODE_DP2
? 2 :
177 (opcode
== TGSI_OPCODE_DP3
? 3 : 4));
178 for (chan
= 0 ; chan
< dp_components
; chan
++) {
179 elements
[0][chan
] = lp_build_emit_fetch(bld_base
,
180 emit_data
->inst
, 0, chan
);
181 elements
[1][chan
] = lp_build_emit_fetch(bld_base
,
182 emit_data
->inst
, 1, chan
);
185 for ( ; chan
< 4; chan
++) {
186 elements
[0][chan
] = base
->zero
;
187 elements
[1][chan
] = base
->zero
;
191 if (opcode
== TGSI_OPCODE_DPH
) {
192 elements
[0][TGSI_CHAN_W
] = base
->one
;
195 emit_data
->args
[0] = lp_build_gather_values(bld_base
->base
.gallivm
,
197 emit_data
->args
[1] = lp_build_gather_values(bld_base
->base
.gallivm
,
199 emit_data
->arg_count
= 2;
201 emit_data
->dst_type
= base
->elem_type
;
204 static struct lp_build_tgsi_action dot_action
= {
205 .fetch_args
= dp_fetch_args
,
206 .emit
= build_tgsi_intrinsic_nomem
,
207 .intr_name
= "llvm.AMDGPU.dp4"
212 LLVMModuleRef
r600_tgsi_llvm(
213 struct radeon_llvm_context
* ctx
,
214 const struct tgsi_token
* tokens
)
216 struct tgsi_shader_info shader_info
;
217 struct lp_build_tgsi_context
* bld_base
= &ctx
->soa
.bld_base
;
218 radeon_llvm_context_init(ctx
);
219 tgsi_scan_shader(tokens
, &shader_info
);
221 bld_base
->info
= &shader_info
;
222 bld_base
->userdata
= ctx
;
223 bld_base
->emit_fetch_funcs
[TGSI_FILE_CONSTANT
] = llvm_fetch_const
;
224 bld_base
->emit_fetch_funcs
[TGSI_FILE_SYSTEM_VALUE
] = llvm_fetch_system_value
;
225 bld_base
->emit_prologue
= llvm_emit_prologue
;
226 bld_base
->emit_epilogue
= llvm_emit_epilogue
;
228 ctx
->load_input
= llvm_load_input
;
229 ctx
->load_system_value
= llvm_load_system_value
;
231 bld_base
->op_actions
[TGSI_OPCODE_DP2
] = dot_action
;
232 bld_base
->op_actions
[TGSI_OPCODE_DP3
] = dot_action
;
233 bld_base
->op_actions
[TGSI_OPCODE_DP4
] = dot_action
;
234 bld_base
->op_actions
[TGSI_OPCODE_DPH
] = dot_action
;
235 bld_base
->op_actions
[TGSI_OPCODE_DDX
].emit
= llvm_emit_tex
;
236 bld_base
->op_actions
[TGSI_OPCODE_DDY
].emit
= llvm_emit_tex
;
237 bld_base
->op_actions
[TGSI_OPCODE_TEX
].emit
= llvm_emit_tex
;
238 bld_base
->op_actions
[TGSI_OPCODE_TXB
].emit
= llvm_emit_tex
;
239 bld_base
->op_actions
[TGSI_OPCODE_TXD
].emit
= llvm_emit_tex
;
240 bld_base
->op_actions
[TGSI_OPCODE_TXL
].emit
= llvm_emit_tex
;
241 bld_base
->op_actions
[TGSI_OPCODE_TXF
].emit
= llvm_emit_tex
;
242 bld_base
->op_actions
[TGSI_OPCODE_TXQ
].emit
= llvm_emit_tex
;
243 bld_base
->op_actions
[TGSI_OPCODE_TXP
].emit
= llvm_emit_tex
;
245 lp_build_tgsi_llvm(bld_base
, tokens
);
247 radeon_llvm_finalize_module(ctx
);
249 return ctx
->gallivm
.module
;
252 const char * r600_llvm_gpu_string(enum radeon_family family
)
254 const char * gpu_family
;
265 gpu_family
= "rv710";
268 gpu_family
= "rv730";
273 gpu_family
= "rv770";
277 gpu_family
= "cedar";
282 gpu_family
= "redwood";
285 gpu_family
= "juniper";
289 gpu_family
= "cypress";
292 gpu_family
= "barts";
295 gpu_family
= "turks";
298 gpu_family
= "caicos";
302 gpu_family
= "cayman";
306 fprintf(stderr
, "Chip not supported by r600 llvm "
307 "backend, please file a bug at bugs.freedesktop.org\n");
313 unsigned r600_llvm_compile(
315 unsigned char ** inst_bytes
,
316 unsigned * inst_byte_count
,
317 enum radeon_family family
,
320 const char * gpu_family
= r600_llvm_gpu_string(family
);
321 return radeon_llvm_compile(mod
, inst_bytes
, inst_byte_count
,