3 * Copyright 2012 Advanced Micro Devices, Inc.
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:
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
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.
25 * Tom Stellard <thomas.stellard@amd.com>
26 * Michel Dänzer <michel.daenzer@amd.com>
27 * Christian König <christian.koenig@amd.com>
30 #include "gallivm/lp_bld_tgsi_action.h"
31 #include "gallivm/lp_bld_const.h"
32 #include "gallivm/lp_bld_gather.h"
33 #include "gallivm/lp_bld_intr.h"
34 #include "gallivm/lp_bld_logic.h"
35 #include "gallivm/lp_bld_tgsi.h"
36 #include "radeon_llvm.h"
37 #include "radeon_llvm_emit.h"
38 #include "tgsi/tgsi_info.h"
39 #include "tgsi/tgsi_parse.h"
40 #include "tgsi/tgsi_scan.h"
41 #include "tgsi/tgsi_dump.h"
43 #include "radeonsi_pipe.h"
44 #include "radeonsi_shader.h"
52 struct si_shader_context
54 struct radeon_llvm_context radeon_bld
;
55 struct r600_context
*rctx
;
56 struct tgsi_parse_context parse
;
57 struct tgsi_token
* tokens
;
58 struct si_pipe_shader
*shader
;
59 struct si_shader_key key
;
60 unsigned type
; /* TGSI_PROCESSOR_* specifies the type of shader. */
61 unsigned ninput_emitted
;
62 /* struct list_head inputs; */
63 /* unsigned * input_mappings *//* From TGSI to SI hw */
64 /* struct tgsi_shader_info info;*/
67 static struct si_shader_context
* si_shader_context(
68 struct lp_build_tgsi_context
* bld_base
)
70 return (struct si_shader_context
*)bld_base
;
74 #define PERSPECTIVE_BASE 0
77 #define SAMPLE_OFFSET 0
78 #define CENTER_OFFSET 2
79 #define CENTROID_OFSET 4
81 #define USE_SGPR_MAX_SUFFIX_LEN 5
82 #define CONST_ADDR_SPACE 2
83 #define USER_SGPR_ADDR_SPACE 8
94 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad
96 * @param offset The offset parameter specifies the number of
97 * elements to offset, not the number of bytes or dwords. An element is the
98 * the type pointed to by the base_ptr parameter (e.g. int is the element of
101 * When LLVM lowers the load instruction, it will convert the element offset
102 * into a dword offset automatically.
105 static LLVMValueRef
build_indexed_load(
106 struct gallivm_state
* gallivm
,
107 LLVMValueRef base_ptr
,
110 LLVMValueRef computed_ptr
= LLVMBuildGEP(
111 gallivm
->builder
, base_ptr
, &offset
, 1, "");
113 return LLVMBuildLoad(gallivm
->builder
, computed_ptr
, "");
117 * Load a value stored in one of the user SGPRs
119 * @param sgpr This is the sgpr to load the value from. If you need to load a
120 * value that is stored in consecutive SGPR registers (e.g. a 64-bit pointer),
121 * then you should pass the index of the first SGPR that holds the value. For
122 * example, if you want to load a pointer that is stored in SGPRs 2 and 3, then
123 * use pass 2 for the sgpr parameter.
125 * The value of the sgpr parameter must also be aligned to the width of the type
126 * being loaded, so that the sgpr parameter is divisible by the dword width of the
127 * type. For example, if the value being loaded is two dwords wide, then the sgpr
128 * parameter must be divisible by two.
130 static LLVMValueRef
use_sgpr(
131 struct gallivm_state
* gallivm
,
135 LLVMValueRef sgpr_index
;
136 LLVMTypeRef ret_type
;
139 sgpr_index
= lp_build_const_int32(gallivm
, sgpr
);
142 case SGPR_CONST_PTR_F32
:
143 assert(sgpr
% 2 == 0);
144 ret_type
= LLVMFloatTypeInContext(gallivm
->context
);
145 ret_type
= LLVMPointerType(ret_type
, CONST_ADDR_SPACE
);
149 ret_type
= LLVMInt32TypeInContext(gallivm
->context
);
153 assert(sgpr
% 2 == 0);
154 ret_type
= LLVMInt64TypeInContext(gallivm
->context
);
157 case SGPR_CONST_PTR_V4I32
:
158 assert(sgpr
% 2 == 0);
159 ret_type
= LLVMInt32TypeInContext(gallivm
->context
);
160 ret_type
= LLVMVectorType(ret_type
, 4);
161 ret_type
= LLVMPointerType(ret_type
, CONST_ADDR_SPACE
);
164 case SGPR_CONST_PTR_V8I32
:
165 assert(sgpr
% 2 == 0);
166 ret_type
= LLVMInt32TypeInContext(gallivm
->context
);
167 ret_type
= LLVMVectorType(ret_type
, 8);
168 ret_type
= LLVMPointerType(ret_type
, CONST_ADDR_SPACE
);
172 assert(!"Unsupported SGPR type in use_sgpr()");
176 ret_type
= LLVMPointerType(ret_type
, USER_SGPR_ADDR_SPACE
);
177 ptr
= LLVMBuildIntToPtr(gallivm
->builder
, sgpr_index
, ret_type
, "");
178 return LLVMBuildLoad(gallivm
->builder
, ptr
, "");
181 static void declare_input_vs(
182 struct si_shader_context
* si_shader_ctx
,
183 unsigned input_index
,
184 const struct tgsi_full_declaration
*decl
)
186 LLVMValueRef t_list_ptr
;
187 LLVMValueRef t_offset
;
189 LLVMValueRef attribute_offset
;
190 LLVMValueRef buffer_index_reg
;
191 LLVMValueRef args
[3];
192 LLVMTypeRef vec4_type
;
194 struct lp_build_context
* uint
= &si_shader_ctx
->radeon_bld
.soa
.bld_base
.uint_bld
;
195 struct lp_build_context
* base
= &si_shader_ctx
->radeon_bld
.soa
.bld_base
.base
;
196 //struct pipe_vertex_element *velem = &rctx->vertex_elements->elements[input_index];
199 /* Load the T list */
200 t_list_ptr
= use_sgpr(base
->gallivm
, SGPR_CONST_PTR_V4I32
, SI_SGPR_VERTEX_BUFFER
);
202 t_offset
= lp_build_const_int32(base
->gallivm
, input_index
);
204 t_list
= build_indexed_load(base
->gallivm
, t_list_ptr
, t_offset
);
206 /* Build the attribute offset */
207 attribute_offset
= lp_build_const_int32(base
->gallivm
, 0);
209 /* Load the buffer index is always, which is always stored in VGPR0
210 * for Vertex Shaders */
211 buffer_index_reg
= build_intrinsic(base
->gallivm
->builder
,
212 "llvm.SI.vs.load.buffer.index", uint
->elem_type
, NULL
, 0,
213 LLVMReadNoneAttribute
);
215 vec4_type
= LLVMVectorType(base
->elem_type
, 4);
217 args
[1] = attribute_offset
;
218 args
[2] = buffer_index_reg
;
219 input
= lp_build_intrinsic(base
->gallivm
->builder
,
220 "llvm.SI.vs.load.input", vec4_type
, args
, 3);
222 /* Break up the vec4 into individual components */
223 for (chan
= 0; chan
< 4; chan
++) {
224 LLVMValueRef llvm_chan
= lp_build_const_int32(base
->gallivm
, chan
);
225 /* XXX: Use a helper function for this. There is one in
227 si_shader_ctx
->radeon_bld
.inputs
[radeon_llvm_reg_index_soa(input_index
, chan
)] =
228 LLVMBuildExtractElement(base
->gallivm
->builder
,
229 input
, llvm_chan
, "");
233 static void declare_input_fs(
234 struct si_shader_context
* si_shader_ctx
,
235 unsigned input_index
,
236 const struct tgsi_full_declaration
*decl
)
238 const char * intr_name
;
240 struct si_shader
*shader
= &si_shader_ctx
->shader
->shader
;
241 struct lp_build_context
* base
=
242 &si_shader_ctx
->radeon_bld
.soa
.bld_base
.base
;
243 struct gallivm_state
* gallivm
= base
->gallivm
;
244 LLVMTypeRef input_type
= LLVMFloatTypeInContext(gallivm
->context
);
247 * [15:0] NewPrimMask (Bit mask for each quad. It is set it the
248 * quad begins a new primitive. Bit 0 always needs
250 * [32:16] ParamOffset
253 LLVMValueRef params
= use_sgpr(base
->gallivm
, SGPR_I32
, SI_PS_NUM_USER_SGPR
);
254 LLVMValueRef attr_number
;
256 if (decl
->Semantic
.Name
== TGSI_SEMANTIC_POSITION
) {
257 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
258 LLVMValueRef args
[1];
260 radeon_llvm_reg_index_soa(input_index
, chan
);
261 args
[0] = lp_build_const_int32(gallivm
, chan
);
262 si_shader_ctx
->radeon_bld
.inputs
[soa_index
] =
263 build_intrinsic(base
->gallivm
->builder
,
264 "llvm.SI.fs.read.pos", input_type
,
265 args
, 1, LLVMReadNoneAttribute
);
270 if (decl
->Semantic
.Name
== TGSI_SEMANTIC_FACE
) {
271 LLVMValueRef face
, is_face_positive
;
273 face
= build_intrinsic(gallivm
->builder
,
274 "llvm.SI.fs.read.face",
276 NULL
, 0, LLVMReadNoneAttribute
);
277 is_face_positive
= LLVMBuildFCmp(gallivm
->builder
,
279 lp_build_const_float(gallivm
, 0.0f
),
282 si_shader_ctx
->radeon_bld
.inputs
[radeon_llvm_reg_index_soa(input_index
, 0)] =
283 LLVMBuildSelect(gallivm
->builder
,
285 lp_build_const_float(gallivm
, 1.0f
),
286 lp_build_const_float(gallivm
, 0.0f
),
288 si_shader_ctx
->radeon_bld
.inputs
[radeon_llvm_reg_index_soa(input_index
, 1)] =
289 si_shader_ctx
->radeon_bld
.inputs
[radeon_llvm_reg_index_soa(input_index
, 2)] =
290 lp_build_const_float(gallivm
, 0.0f
);
291 si_shader_ctx
->radeon_bld
.inputs
[radeon_llvm_reg_index_soa(input_index
, 3)] =
292 lp_build_const_float(gallivm
, 1.0f
);
297 shader
->input
[input_index
].param_offset
= shader
->ninterp
++;
298 attr_number
= lp_build_const_int32(gallivm
,
299 shader
->input
[input_index
].param_offset
);
301 /* XXX: Handle all possible interpolation modes */
302 switch (decl
->Interp
.Interpolate
) {
303 case TGSI_INTERPOLATE_COLOR
:
304 /* XXX: Flat shading hangs the GPU */
305 if (si_shader_ctx
->rctx
->queued
.named
.rasterizer
&&
306 si_shader_ctx
->rctx
->queued
.named
.rasterizer
->flatshade
) {
308 intr_name
= "llvm.SI.fs.interp.constant";
310 intr_name
= "llvm.SI.fs.interp.linear.center";
313 if (decl
->Interp
.Centroid
)
314 intr_name
= "llvm.SI.fs.interp.persp.centroid";
316 intr_name
= "llvm.SI.fs.interp.persp.center";
319 case TGSI_INTERPOLATE_CONSTANT
:
320 /* XXX: Flat shading hangs the GPU */
322 intr_name
= "llvm.SI.fs.interp.constant";
325 case TGSI_INTERPOLATE_LINEAR
:
326 if (decl
->Interp
.Centroid
)
327 intr_name
= "llvm.SI.fs.interp.linear.centroid";
329 intr_name
= "llvm.SI.fs.interp.linear.center";
331 case TGSI_INTERPOLATE_PERSPECTIVE
:
332 if (decl
->Interp
.Centroid
)
333 intr_name
= "llvm.SI.fs.interp.persp.centroid";
335 intr_name
= "llvm.SI.fs.interp.persp.center";
338 fprintf(stderr
, "Warning: Unhandled interpolation mode.\n");
342 if (!si_shader_ctx
->ninput_emitted
++) {
343 /* Enable whole quad mode */
344 lp_build_intrinsic(gallivm
->builder
,
346 LLVMVoidTypeInContext(gallivm
->context
),
350 /* XXX: Could there be more than TGSI_NUM_CHANNELS (4) ? */
351 if (decl
->Semantic
.Name
== TGSI_SEMANTIC_COLOR
&&
352 si_shader_ctx
->key
.color_two_side
) {
353 LLVMValueRef args
[3];
354 LLVMValueRef face
, is_face_positive
;
355 LLVMValueRef back_attr_number
=
356 lp_build_const_int32(gallivm
,
357 shader
->input
[input_index
].param_offset
+ 1);
359 face
= build_intrinsic(gallivm
->builder
,
360 "llvm.SI.fs.read.face",
362 NULL
, 0, LLVMReadNoneAttribute
);
363 is_face_positive
= LLVMBuildFCmp(gallivm
->builder
,
365 lp_build_const_float(gallivm
, 0.0f
),
369 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
370 LLVMValueRef llvm_chan
= lp_build_const_int32(gallivm
, chan
);
371 unsigned soa_index
= radeon_llvm_reg_index_soa(input_index
, chan
);
372 LLVMValueRef front
, back
;
375 args
[1] = attr_number
;
376 front
= build_intrinsic(base
->gallivm
->builder
, intr_name
,
377 input_type
, args
, 3, LLVMReadOnlyAttribute
);
379 args
[1] = back_attr_number
;
380 back
= build_intrinsic(base
->gallivm
->builder
, intr_name
,
381 input_type
, args
, 3, LLVMReadOnlyAttribute
);
383 si_shader_ctx
->radeon_bld
.inputs
[soa_index
] =
384 LLVMBuildSelect(gallivm
->builder
,
393 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
394 LLVMValueRef args
[3];
395 LLVMValueRef llvm_chan
= lp_build_const_int32(gallivm
, chan
);
396 unsigned soa_index
= radeon_llvm_reg_index_soa(input_index
, chan
);
398 args
[1] = attr_number
;
400 si_shader_ctx
->radeon_bld
.inputs
[soa_index
] =
401 build_intrinsic(base
->gallivm
->builder
, intr_name
,
402 input_type
, args
, 3, LLVMReadOnlyAttribute
);
407 static void declare_input(
408 struct radeon_llvm_context
* radeon_bld
,
409 unsigned input_index
,
410 const struct tgsi_full_declaration
*decl
)
412 struct si_shader_context
* si_shader_ctx
=
413 si_shader_context(&radeon_bld
->soa
.bld_base
);
414 if (si_shader_ctx
->type
== TGSI_PROCESSOR_VERTEX
) {
415 declare_input_vs(si_shader_ctx
, input_index
, decl
);
416 } else if (si_shader_ctx
->type
== TGSI_PROCESSOR_FRAGMENT
) {
417 declare_input_fs(si_shader_ctx
, input_index
, decl
);
419 fprintf(stderr
, "Warning: Unsupported shader type,\n");
423 static LLVMValueRef
fetch_constant(
424 struct lp_build_tgsi_context
* bld_base
,
425 const struct tgsi_full_src_register
*reg
,
426 enum tgsi_opcode_type type
,
429 struct lp_build_context
* base
= &bld_base
->base
;
432 LLVMValueRef const_ptr
;
436 /* currently not supported */
437 if (reg
->Register
.Indirect
) {
439 load
= lp_build_const_int32(base
->gallivm
, 0);
440 return bitcast(bld_base
, type
, load
);
443 const_ptr
= use_sgpr(base
->gallivm
, SGPR_CONST_PTR_F32
, SI_SGPR_CONST
);
445 /* XXX: This assumes that the constant buffer is not packed, so
446 * CONST[0].x will have an offset of 0 and CONST[1].x will have an
448 idx
= (reg
->Register
.Index
* 4) + swizzle
;
450 /* index loads above 255 are currently not supported */
455 offset
= lp_build_const_int32(base
->gallivm
, idx
);
457 load
= build_indexed_load(base
->gallivm
, const_ptr
, offset
);
458 return bitcast(bld_base
, type
, load
);
461 /* Initialize arguments for the shader export intrinsic */
462 static void si_llvm_init_export_args(struct lp_build_tgsi_context
*bld_base
,
463 struct tgsi_full_declaration
*d
,
468 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
469 struct lp_build_context
*uint
=
470 &si_shader_ctx
->radeon_bld
.soa
.bld_base
.uint_bld
;
471 struct lp_build_context
*base
= &bld_base
->base
;
472 unsigned compressed
= 0;
475 if (si_shader_ctx
->type
== TGSI_PROCESSOR_FRAGMENT
) {
476 int cbuf
= target
- V_008DFC_SQ_EXP_MRT
;
478 if (cbuf
>= 0 && cbuf
< 8) {
479 struct r600_context
*rctx
= si_shader_ctx
->rctx
;
480 compressed
= (si_shader_ctx
->key
.export_16bpc
>> cbuf
) & 0x1;
485 /* Pixel shader needs to pack output values before export */
486 for (chan
= 0; chan
< 2; chan
++ ) {
487 LLVMValueRef
*out_ptr
=
488 si_shader_ctx
->radeon_bld
.soa
.outputs
[index
];
489 args
[0] = LLVMBuildLoad(base
->gallivm
->builder
,
490 out_ptr
[2 * chan
], "");
491 args
[1] = LLVMBuildLoad(base
->gallivm
->builder
,
492 out_ptr
[2 * chan
+ 1], "");
494 build_intrinsic(base
->gallivm
->builder
,
496 LLVMInt32TypeInContext(base
->gallivm
->context
),
498 LLVMReadNoneAttribute
);
499 args
[chan
+ 7] = args
[chan
+ 5] =
500 LLVMBuildBitCast(base
->gallivm
->builder
,
502 LLVMFloatTypeInContext(base
->gallivm
->context
),
509 for (chan
= 0; chan
< 4; chan
++ ) {
510 LLVMValueRef out_ptr
=
511 si_shader_ctx
->radeon_bld
.soa
.outputs
[index
][chan
];
512 /* +5 because the first output value will be
513 * the 6th argument to the intrinsic. */
514 args
[chan
+ 5] = LLVMBuildLoad(base
->gallivm
->builder
,
518 /* Clear COMPR flag */
519 args
[4] = uint
->zero
;
522 /* XXX: This controls which components of the output
523 * registers actually get exported. (e.g bit 0 means export
524 * X component, bit 1 means export Y component, etc.) I'm
525 * hard coding this to 0xf for now. In the future, we might
526 * want to do something else. */
527 args
[0] = lp_build_const_int32(base
->gallivm
, 0xf);
529 /* Specify whether the EXEC mask represents the valid mask */
530 args
[1] = uint
->zero
;
532 /* Specify whether this is the last export */
533 args
[2] = uint
->zero
;
535 /* Specify the target we are exporting */
536 args
[3] = lp_build_const_int32(base
->gallivm
, target
);
538 /* XXX: We probably need to keep track of the output
539 * values, so we know what we are passing to the next
543 static void si_llvm_emit_prologue(struct lp_build_tgsi_context
*bld_base
)
545 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
546 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
547 lp_build_intrinsic_unary(gallivm
->builder
,
548 "llvm.AMDGPU.shader.type",
549 LLVMVoidTypeInContext(gallivm
->context
),
550 lp_build_const_int32(gallivm
, si_shader_ctx
->type
));
554 static void si_alpha_test(struct lp_build_tgsi_context
*bld_base
,
557 struct si_shader_context
*si_shader_ctx
= si_shader_context(bld_base
);
558 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
560 if (si_shader_ctx
->key
.alpha_func
!= PIPE_FUNC_NEVER
) {
561 LLVMValueRef out_ptr
= si_shader_ctx
->radeon_bld
.soa
.outputs
[index
][3];
562 LLVMValueRef alpha_pass
=
563 lp_build_cmp(&bld_base
->base
,
564 si_shader_ctx
->key
.alpha_func
,
565 LLVMBuildLoad(gallivm
->builder
, out_ptr
, ""),
566 lp_build_const_float(gallivm
, si_shader_ctx
->key
.alpha_ref
));
568 lp_build_select(&bld_base
->base
,
570 lp_build_const_float(gallivm
, 1.0f
),
571 lp_build_const_float(gallivm
, -1.0f
));
573 build_intrinsic(gallivm
->builder
,
575 LLVMVoidTypeInContext(gallivm
->context
),
578 build_intrinsic(gallivm
->builder
,
580 LLVMVoidTypeInContext(gallivm
->context
),
585 /* XXX: This is partially implemented for VS only at this point. It is not complete */
586 static void si_llvm_emit_epilogue(struct lp_build_tgsi_context
* bld_base
)
588 struct si_shader_context
* si_shader_ctx
= si_shader_context(bld_base
);
589 struct si_shader
* shader
= &si_shader_ctx
->shader
->shader
;
590 struct lp_build_context
* base
= &bld_base
->base
;
591 struct lp_build_context
* uint
=
592 &si_shader_ctx
->radeon_bld
.soa
.bld_base
.uint_bld
;
593 struct tgsi_parse_context
*parse
= &si_shader_ctx
->parse
;
594 LLVMValueRef args
[9];
595 LLVMValueRef last_args
[9] = { 0 };
596 unsigned color_count
= 0;
597 unsigned param_count
= 0;
598 int depth_index
= -1, stencil_index
= -1;
600 while (!tgsi_parse_end_of_tokens(parse
)) {
601 struct tgsi_full_declaration
*d
=
602 &parse
->FullToken
.FullDeclaration
;
607 tgsi_parse_token(parse
);
608 if (parse
->FullToken
.Token
.Type
!= TGSI_TOKEN_TYPE_DECLARATION
)
611 switch (d
->Declaration
.File
) {
612 case TGSI_FILE_INPUT
:
613 i
= shader
->ninput
++;
614 shader
->input
[i
].name
= d
->Semantic
.Name
;
615 shader
->input
[i
].sid
= d
->Semantic
.Index
;
616 shader
->input
[i
].interpolate
= d
->Interp
.Interpolate
;
617 shader
->input
[i
].centroid
= d
->Interp
.Centroid
;
620 case TGSI_FILE_OUTPUT
:
621 i
= shader
->noutput
++;
622 shader
->output
[i
].name
= d
->Semantic
.Name
;
623 shader
->output
[i
].sid
= d
->Semantic
.Index
;
624 shader
->output
[i
].interpolate
= d
->Interp
.Interpolate
;
631 for (index
= d
->Range
.First
; index
<= d
->Range
.Last
; index
++) {
632 /* Select the correct target */
633 switch(d
->Semantic
.Name
) {
634 case TGSI_SEMANTIC_PSIZE
:
635 target
= V_008DFC_SQ_EXP_POS
;
637 case TGSI_SEMANTIC_POSITION
:
638 if (si_shader_ctx
->type
== TGSI_PROCESSOR_VERTEX
) {
639 target
= V_008DFC_SQ_EXP_POS
;
645 case TGSI_SEMANTIC_STENCIL
:
646 stencil_index
= index
;
648 case TGSI_SEMANTIC_COLOR
:
649 if (si_shader_ctx
->type
== TGSI_PROCESSOR_VERTEX
) {
650 case TGSI_SEMANTIC_BCOLOR
:
651 target
= V_008DFC_SQ_EXP_PARAM
+ param_count
;
652 shader
->output
[i
].param_offset
= param_count
;
655 target
= V_008DFC_SQ_EXP_MRT
+ color_count
;
656 if (color_count
== 0 &&
657 si_shader_ctx
->key
.alpha_func
!= PIPE_FUNC_ALWAYS
)
658 si_alpha_test(bld_base
, index
);
663 case TGSI_SEMANTIC_FOG
:
664 case TGSI_SEMANTIC_GENERIC
:
665 target
= V_008DFC_SQ_EXP_PARAM
+ param_count
;
666 shader
->output
[i
].param_offset
= param_count
;
672 "Warning: SI unhandled output type:%d\n",
676 si_llvm_init_export_args(bld_base
, d
, index
, target
, args
);
678 if (si_shader_ctx
->type
== TGSI_PROCESSOR_VERTEX
?
679 (d
->Semantic
.Name
== TGSI_SEMANTIC_POSITION
) :
680 (d
->Semantic
.Name
== TGSI_SEMANTIC_COLOR
)) {
682 lp_build_intrinsic(base
->gallivm
->builder
,
684 LLVMVoidTypeInContext(base
->gallivm
->context
),
688 memcpy(last_args
, args
, sizeof(args
));
690 lp_build_intrinsic(base
->gallivm
->builder
,
692 LLVMVoidTypeInContext(base
->gallivm
->context
),
699 if (depth_index
>= 0 || stencil_index
>= 0) {
700 LLVMValueRef out_ptr
;
703 /* Specify the target we are exporting */
704 args
[3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_MRTZ
);
706 if (depth_index
>= 0) {
707 out_ptr
= si_shader_ctx
->radeon_bld
.soa
.outputs
[depth_index
][2];
708 args
[5] = LLVMBuildLoad(base
->gallivm
->builder
, out_ptr
, "");
711 if (stencil_index
< 0) {
718 if (stencil_index
>= 0) {
719 out_ptr
= si_shader_ctx
->radeon_bld
.soa
.outputs
[stencil_index
][1];
722 args
[6] = LLVMBuildLoad(base
->gallivm
->builder
, out_ptr
, "");
729 /* Specify which components to enable */
730 args
[0] = lp_build_const_int32(base
->gallivm
, mask
);
734 args
[4] = uint
->zero
;
737 lp_build_intrinsic(base
->gallivm
->builder
,
739 LLVMVoidTypeInContext(base
->gallivm
->context
),
742 memcpy(last_args
, args
, sizeof(args
));
746 assert(si_shader_ctx
->type
== TGSI_PROCESSOR_FRAGMENT
);
748 /* Specify which components to enable */
749 last_args
[0] = lp_build_const_int32(base
->gallivm
, 0x0);
751 /* Specify the target we are exporting */
752 last_args
[3] = lp_build_const_int32(base
->gallivm
, V_008DFC_SQ_EXP_MRT
);
754 /* Set COMPR flag to zero to export data as 32-bit */
755 last_args
[4] = uint
->zero
;
758 last_args
[5]= uint
->zero
;
759 last_args
[6]= uint
->zero
;
760 last_args
[7]= uint
->zero
;
761 last_args
[8]= uint
->zero
;
764 /* Specify whether the EXEC mask represents the valid mask */
765 last_args
[1] = lp_build_const_int32(base
->gallivm
,
766 si_shader_ctx
->type
== TGSI_PROCESSOR_FRAGMENT
);
768 /* Specify that this is the last export */
769 last_args
[2] = lp_build_const_int32(base
->gallivm
, 1);
771 lp_build_intrinsic(base
->gallivm
->builder
,
773 LLVMVoidTypeInContext(base
->gallivm
->context
),
776 /* XXX: Look up what this function does */
777 /* ctx->shader->output[i].spi_sid = r600_spi_sid(&ctx->shader->output[i]);*/
780 static void tex_fetch_args(
781 struct lp_build_tgsi_context
* bld_base
,
782 struct lp_build_emit_data
* emit_data
)
784 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
789 /* XXX: should be optimized using emit_data->inst->Dst[0].Register.WriteMask*/
790 emit_data
->args
[0] = lp_build_const_int32(bld_base
->base
.gallivm
, 0xf);
793 /* XXX: Not all sample instructions need 4 address arguments. */
794 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_TXP
) {
797 LLVMValueRef coords
[4];
799 emit_data
->dst_type
= LLVMVectorType(bld_base
->base
.elem_type
, 4);
800 src_w
= lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, TGSI_CHAN_W
);
802 for (chan
= 0; chan
< 3; chan
++ ) {
803 LLVMValueRef arg
= lp_build_emit_fetch(bld_base
,
804 emit_data
->inst
, 0, chan
);
805 coords
[chan
] = lp_build_emit_llvm_binary(bld_base
,
809 coords
[3] = bld_base
->base
.one
;
810 emit_data
->args
[1] = lp_build_gather_values(bld_base
->base
.gallivm
,
813 emit_data
->args
[1] = lp_build_emit_fetch(bld_base
, emit_data
->inst
,
816 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_TEX2
||
817 inst
->Instruction
.Opcode
== TGSI_OPCODE_TXB2
||
818 inst
->Instruction
.Opcode
== TGSI_OPCODE_TXL2
) {
819 /* These instructions have additional operand that should be packed
820 * into the cube coord vector by radeon_llvm_emit_prepare_cube_coords.
821 * That operand should be passed as a float value in the args array
822 * right after the coord vector. After packing it's not used anymore,
823 * that's why arg_count is not increased */
824 emit_data
->args
[2] = lp_build_emit_fetch(bld_base
, inst
, 1, 0);
827 if ((inst
->Texture
.Texture
== TGSI_TEXTURE_CUBE
||
828 inst
->Texture
.Texture
== TGSI_TEXTURE_SHADOWCUBE
) &&
829 inst
->Instruction
.Opcode
!= TGSI_OPCODE_TXQ
) {
830 radeon_llvm_emit_prepare_cube_coords(bld_base
, emit_data
, 1);
834 ptr
= use_sgpr(bld_base
->base
.gallivm
, SGPR_CONST_PTR_V8I32
, SI_SGPR_RESOURCE
);
835 offset
= lp_build_const_int32(bld_base
->base
.gallivm
,
836 emit_data
->inst
->Src
[1].Register
.Index
);
837 emit_data
->args
[2] = build_indexed_load(bld_base
->base
.gallivm
,
841 ptr
= use_sgpr(bld_base
->base
.gallivm
, SGPR_CONST_PTR_V4I32
, SI_SGPR_SAMPLER
);
842 offset
= lp_build_const_int32(bld_base
->base
.gallivm
,
843 emit_data
->inst
->Src
[1].Register
.Index
);
844 emit_data
->args
[3] = build_indexed_load(bld_base
->base
.gallivm
,
848 /* XXX: We might want to pass this information to the shader at some. */
849 /* emit_data->args[4] = lp_build_const_int32(bld_base->base.gallivm,
850 emit_data->inst->Texture.Texture);
853 emit_data
->arg_count
= 4;
854 /* XXX: To optimize, we could use a float or v2f32, if the last bits of
855 * the writemask are clear */
856 emit_data
->dst_type
= LLVMVectorType(
857 LLVMFloatTypeInContext(bld_base
->base
.gallivm
->context
),
861 static const struct lp_build_tgsi_action tex_action
= {
862 .fetch_args
= tex_fetch_args
,
863 .emit
= lp_build_tgsi_intrinsic
,
864 .intr_name
= "llvm.SI.sample"
867 static const struct lp_build_tgsi_action txb_action
= {
868 .fetch_args
= tex_fetch_args
,
869 .emit
= lp_build_tgsi_intrinsic
,
870 .intr_name
= "llvm.SI.sample.bias"
873 static const struct lp_build_tgsi_action txl_action
= {
874 .fetch_args
= tex_fetch_args
,
875 .emit
= lp_build_tgsi_intrinsic
,
876 .intr_name
= "llvm.SI.sample.lod"
880 int si_pipe_shader_create(
881 struct pipe_context
*ctx
,
882 struct si_pipe_shader
*shader
,
883 struct si_shader_key key
)
885 struct r600_context
*rctx
= (struct r600_context
*)ctx
;
886 struct si_pipe_shader_selector
*sel
= shader
->selector
;
887 struct si_shader_context si_shader_ctx
;
888 struct tgsi_shader_info shader_info
;
889 struct lp_build_tgsi_context
* bld_base
;
891 unsigned char * inst_bytes
;
892 unsigned inst_byte_count
;
897 dump
= debug_get_bool_option("RADEON_DUMP_SHADERS", FALSE
);
899 assert(shader
->shader
.noutput
== 0);
900 assert(shader
->shader
.ninterp
== 0);
901 assert(shader
->shader
.ninput
== 0);
903 memset(&si_shader_ctx
, 0, sizeof(si_shader_ctx
));
904 radeon_llvm_context_init(&si_shader_ctx
.radeon_bld
);
905 bld_base
= &si_shader_ctx
.radeon_bld
.soa
.bld_base
;
907 tgsi_scan_shader(sel
->tokens
, &shader_info
);
908 shader
->shader
.uses_kill
= shader_info
.uses_kill
;
909 bld_base
->info
= &shader_info
;
910 bld_base
->emit_fetch_funcs
[TGSI_FILE_CONSTANT
] = fetch_constant
;
911 bld_base
->emit_prologue
= si_llvm_emit_prologue
;
912 bld_base
->emit_epilogue
= si_llvm_emit_epilogue
;
914 bld_base
->op_actions
[TGSI_OPCODE_TEX
] = tex_action
;
915 bld_base
->op_actions
[TGSI_OPCODE_TXB
] = txb_action
;
916 bld_base
->op_actions
[TGSI_OPCODE_TXL
] = txl_action
;
917 bld_base
->op_actions
[TGSI_OPCODE_TXP
] = tex_action
;
919 si_shader_ctx
.radeon_bld
.load_input
= declare_input
;
920 si_shader_ctx
.tokens
= sel
->tokens
;
921 tgsi_parse_init(&si_shader_ctx
.parse
, si_shader_ctx
.tokens
);
922 si_shader_ctx
.shader
= shader
;
923 si_shader_ctx
.key
= key
;
924 si_shader_ctx
.type
= si_shader_ctx
.parse
.FullHeader
.Processor
.Processor
;
925 si_shader_ctx
.rctx
= rctx
;
927 shader
->shader
.nr_cbufs
= rctx
->framebuffer
.nr_cbufs
;
929 /* Dump TGSI code before doing TGSI->LLVM conversion in case the
930 * conversion fails. */
932 tgsi_dump(sel
->tokens
, 0);
935 if (!lp_build_tgsi_llvm(bld_base
, sel
->tokens
)) {
936 fprintf(stderr
, "Failed to translate shader from TGSI to LLVM\n");
940 radeon_llvm_finalize_module(&si_shader_ctx
.radeon_bld
);
942 mod
= bld_base
->base
.gallivm
->module
;
946 radeon_llvm_compile(mod
, &inst_bytes
, &inst_byte_count
, "SI", dump
);
948 fprintf(stderr
, "SI CODE:\n");
949 for (i
= 0; i
< inst_byte_count
; i
+=4 ) {
950 fprintf(stderr
, "%02x%02x%02x%02x\n", inst_bytes
[i
+ 3],
951 inst_bytes
[i
+ 2], inst_bytes
[i
+ 1],
956 shader
->num_sgprs
= util_le32_to_cpu(*(uint32_t*)inst_bytes
);
957 shader
->num_vgprs
= util_le32_to_cpu(*(uint32_t*)(inst_bytes
+ 4));
958 shader
->spi_ps_input_ena
= util_le32_to_cpu(*(uint32_t*)(inst_bytes
+ 8));
960 radeon_llvm_dispose(&si_shader_ctx
.radeon_bld
);
961 tgsi_parse_free(&si_shader_ctx
.parse
);
963 /* copy new shader */
964 si_resource_reference(&shader
->bo
, NULL
);
965 shader
->bo
= si_resource_create_custom(ctx
->screen
, PIPE_USAGE_IMMUTABLE
,
966 inst_byte_count
- 12);
967 if (shader
->bo
== NULL
) {
971 ptr
= (uint32_t*)rctx
->ws
->buffer_map(shader
->bo
->cs_buf
, rctx
->cs
, PIPE_TRANSFER_WRITE
);
972 if (0 /*R600_BIG_ENDIAN*/) {
973 for (i
= 0; i
< (inst_byte_count
-12)/4; ++i
) {
974 ptr
[i
] = util_bswap32(*(uint32_t*)(inst_bytes
+12 + i
*4));
977 memcpy(ptr
, inst_bytes
+ 12, inst_byte_count
- 12);
979 rctx
->ws
->buffer_unmap(shader
->bo
->cs_buf
);
986 void si_pipe_shader_destroy(struct pipe_context
*ctx
, struct si_pipe_shader
*shader
)
988 si_resource_reference(&shader
->bo
, NULL
);