2 * Copyright © 2016 Red Hat.
3 * Copyright © 2016 Bas Nieuwenhuizen
5 * based in part on anv driver which is:
6 * Copyright © 2015 Intel Corporation
8 * Permission is hereby granted, free of charge, to any person obtaining a
9 * copy of this software and associated documentation files (the "Software"),
10 * to deal in the Software without restriction, including without limitation
11 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12 * and/or sell copies of the Software, and to permit persons to whom the
13 * Software is furnished to do so, subject to the following conditions:
15 * The above copyright notice and this permission notice (including the next
16 * paragraph) shall be included in all copies or substantial portions of the
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
22 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
23 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
24 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
28 #include "radv_private.h"
29 #include "radv_shader.h"
30 #include "radv_shader_helper.h"
31 #include "radv_shader_args.h"
34 #include <llvm-c/Core.h>
35 #include <llvm-c/TargetMachine.h>
36 #include <llvm-c/Transforms/Scalar.h>
37 #include <llvm-c/Transforms/Utils.h>
40 #include "ac_binary.h"
41 #include "ac_llvm_util.h"
42 #include "ac_llvm_build.h"
43 #include "ac_shader_abi.h"
44 #include "ac_shader_util.h"
45 #include "ac_exp_param.h"
47 #define RADEON_LLVM_MAX_INPUTS (VARYING_SLOT_VAR31 + 1)
49 struct radv_shader_context
{
50 struct ac_llvm_context ac
;
51 const struct nir_shader
*shader
;
52 struct ac_shader_abi abi
;
53 const struct radv_shader_args
*args
;
55 gl_shader_stage stage
;
57 unsigned max_workgroup_size
;
58 LLVMContextRef context
;
59 LLVMValueRef main_function
;
61 LLVMValueRef descriptor_sets
[MAX_SETS
];
63 LLVMValueRef ring_offsets
;
65 LLVMValueRef rel_auto_id
;
67 LLVMValueRef gs_wave_id
;
68 LLVMValueRef gs_vtx_offset
[6];
70 LLVMValueRef esgs_ring
;
71 LLVMValueRef gsvs_ring
[4];
72 LLVMValueRef hs_ring_tess_offchip
;
73 LLVMValueRef hs_ring_tess_factor
;
75 LLVMValueRef inputs
[RADEON_LLVM_MAX_INPUTS
* 4];
79 LLVMValueRef gs_next_vertex
[4];
80 LLVMValueRef gs_curprim_verts
[4];
81 LLVMValueRef gs_generated_prims
[4];
82 LLVMValueRef gs_ngg_emit
;
83 LLVMValueRef gs_ngg_scratch
;
85 uint32_t tcs_num_inputs
;
86 uint32_t tcs_num_patches
;
88 LLVMValueRef vertexptr
; /* GFX10 only */
91 struct radv_shader_output_values
{
92 LLVMValueRef values
[4];
98 static inline struct radv_shader_context
*
99 radv_shader_context_from_abi(struct ac_shader_abi
*abi
)
101 struct radv_shader_context
*ctx
= NULL
;
102 return container_of(abi
, ctx
, abi
);
105 static LLVMValueRef
get_rel_patch_id(struct radv_shader_context
*ctx
)
107 switch (ctx
->stage
) {
108 case MESA_SHADER_TESS_CTRL
:
109 return ac_unpack_param(&ctx
->ac
,
110 ac_get_arg(&ctx
->ac
, ctx
->args
->ac
.tcs_rel_ids
),
112 case MESA_SHADER_TESS_EVAL
:
113 return ac_get_arg(&ctx
->ac
, ctx
->args
->tes_rel_patch_id
);
116 unreachable("Illegal stage");
121 get_tcs_num_patches(struct radv_shader_context
*ctx
)
123 unsigned num_tcs_input_cp
= ctx
->args
->options
->key
.tcs
.input_vertices
;
124 unsigned num_tcs_output_cp
= ctx
->shader
->info
.tess
.tcs_vertices_out
;
125 uint32_t input_vertex_size
= ctx
->tcs_num_inputs
* 16;
126 uint32_t input_patch_size
= ctx
->args
->options
->key
.tcs
.input_vertices
* input_vertex_size
;
127 uint32_t num_tcs_outputs
= util_last_bit64(ctx
->args
->shader_info
->tcs
.outputs_written
);
128 uint32_t num_tcs_patch_outputs
= util_last_bit64(ctx
->args
->shader_info
->tcs
.patch_outputs_written
);
129 uint32_t output_vertex_size
= num_tcs_outputs
* 16;
130 uint32_t pervertex_output_patch_size
= ctx
->shader
->info
.tess
.tcs_vertices_out
* output_vertex_size
;
131 uint32_t output_patch_size
= pervertex_output_patch_size
+ num_tcs_patch_outputs
* 16;
132 unsigned num_patches
;
133 unsigned hardware_lds_size
;
135 /* Ensure that we only need one wave per SIMD so we don't need to check
136 * resource usage. Also ensures that the number of tcs in and out
137 * vertices per threadgroup are at most 256.
139 num_patches
= 64 / MAX2(num_tcs_input_cp
, num_tcs_output_cp
) * 4;
140 /* Make sure that the data fits in LDS. This assumes the shaders only
141 * use LDS for the inputs and outputs.
143 hardware_lds_size
= 32768;
145 /* Looks like STONEY hangs if we use more than 32 KiB LDS in a single
146 * threadgroup, even though there is more than 32 KiB LDS.
148 * Test: dEQP-VK.tessellation.shader_input_output.barrier
150 if (ctx
->args
->options
->chip_class
>= GFX7
&& ctx
->args
->options
->family
!= CHIP_STONEY
)
151 hardware_lds_size
= 65536;
153 num_patches
= MIN2(num_patches
, hardware_lds_size
/ (input_patch_size
+ output_patch_size
));
154 /* Make sure the output data fits in the offchip buffer */
155 num_patches
= MIN2(num_patches
, (ctx
->args
->options
->tess_offchip_block_dw_size
* 4) / output_patch_size
);
156 /* Not necessary for correctness, but improves performance. The
157 * specific value is taken from the proprietary driver.
159 num_patches
= MIN2(num_patches
, 40);
161 /* GFX6 bug workaround - limit LS-HS threadgroups to only one wave. */
162 if (ctx
->args
->options
->chip_class
== GFX6
) {
163 unsigned one_wave
= 64 / MAX2(num_tcs_input_cp
, num_tcs_output_cp
);
164 num_patches
= MIN2(num_patches
, one_wave
);
170 calculate_tess_lds_size(struct radv_shader_context
*ctx
)
172 unsigned num_tcs_input_cp
= ctx
->args
->options
->key
.tcs
.input_vertices
;
173 unsigned num_tcs_output_cp
;
174 unsigned num_tcs_outputs
, num_tcs_patch_outputs
;
175 unsigned input_vertex_size
, output_vertex_size
;
176 unsigned input_patch_size
, output_patch_size
;
177 unsigned pervertex_output_patch_size
;
178 unsigned output_patch0_offset
;
179 unsigned num_patches
;
182 num_tcs_output_cp
= ctx
->shader
->info
.tess
.tcs_vertices_out
;
183 num_tcs_outputs
= util_last_bit64(ctx
->args
->shader_info
->tcs
.outputs_written
);
184 num_tcs_patch_outputs
= util_last_bit64(ctx
->args
->shader_info
->tcs
.patch_outputs_written
);
186 input_vertex_size
= ctx
->tcs_num_inputs
* 16;
187 output_vertex_size
= num_tcs_outputs
* 16;
189 input_patch_size
= num_tcs_input_cp
* input_vertex_size
;
191 pervertex_output_patch_size
= num_tcs_output_cp
* output_vertex_size
;
192 output_patch_size
= pervertex_output_patch_size
+ num_tcs_patch_outputs
* 16;
194 num_patches
= ctx
->tcs_num_patches
;
195 output_patch0_offset
= input_patch_size
* num_patches
;
197 lds_size
= output_patch0_offset
+ output_patch_size
* num_patches
;
201 /* Tessellation shaders pass outputs to the next shader using LDS.
203 * LS outputs = TCS inputs
204 * TCS outputs = TES inputs
207 * - TCS inputs for patch 0
208 * - TCS inputs for patch 1
209 * - TCS inputs for patch 2 = get_tcs_in_current_patch_offset (if RelPatchID==2)
211 * - TCS outputs for patch 0 = get_tcs_out_patch0_offset
212 * - Per-patch TCS outputs for patch 0 = get_tcs_out_patch0_patch_data_offset
213 * - TCS outputs for patch 1
214 * - Per-patch TCS outputs for patch 1
215 * - TCS outputs for patch 2 = get_tcs_out_current_patch_offset (if RelPatchID==2)
216 * - Per-patch TCS outputs for patch 2 = get_tcs_out_current_patch_data_offset (if RelPatchID==2)
219 * All three shaders VS(LS), TCS, TES share the same LDS space.
222 get_tcs_in_patch_stride(struct radv_shader_context
*ctx
)
224 assert(ctx
->stage
== MESA_SHADER_TESS_CTRL
);
225 uint32_t input_vertex_size
= ctx
->tcs_num_inputs
* 16;
226 uint32_t input_patch_size
= ctx
->args
->options
->key
.tcs
.input_vertices
* input_vertex_size
;
228 input_patch_size
/= 4;
229 return LLVMConstInt(ctx
->ac
.i32
, input_patch_size
, false);
233 get_tcs_out_patch_stride(struct radv_shader_context
*ctx
)
235 uint32_t num_tcs_outputs
= util_last_bit64(ctx
->args
->shader_info
->tcs
.outputs_written
);
236 uint32_t num_tcs_patch_outputs
= util_last_bit64(ctx
->args
->shader_info
->tcs
.patch_outputs_written
);
237 uint32_t output_vertex_size
= num_tcs_outputs
* 16;
238 uint32_t pervertex_output_patch_size
= ctx
->shader
->info
.tess
.tcs_vertices_out
* output_vertex_size
;
239 uint32_t output_patch_size
= pervertex_output_patch_size
+ num_tcs_patch_outputs
* 16;
240 output_patch_size
/= 4;
241 return LLVMConstInt(ctx
->ac
.i32
, output_patch_size
, false);
245 get_tcs_out_vertex_stride(struct radv_shader_context
*ctx
)
247 uint32_t num_tcs_outputs
= util_last_bit64(ctx
->args
->shader_info
->tcs
.outputs_written
);
248 uint32_t output_vertex_size
= num_tcs_outputs
* 16;
249 output_vertex_size
/= 4;
250 return LLVMConstInt(ctx
->ac
.i32
, output_vertex_size
, false);
254 get_tcs_out_patch0_offset(struct radv_shader_context
*ctx
)
256 assert (ctx
->stage
== MESA_SHADER_TESS_CTRL
);
257 uint32_t input_vertex_size
= ctx
->tcs_num_inputs
* 16;
258 uint32_t input_patch_size
= ctx
->args
->options
->key
.tcs
.input_vertices
* input_vertex_size
;
259 uint32_t output_patch0_offset
= input_patch_size
;
260 unsigned num_patches
= ctx
->tcs_num_patches
;
262 output_patch0_offset
*= num_patches
;
263 output_patch0_offset
/= 4;
264 return LLVMConstInt(ctx
->ac
.i32
, output_patch0_offset
, false);
268 get_tcs_out_patch0_patch_data_offset(struct radv_shader_context
*ctx
)
270 assert (ctx
->stage
== MESA_SHADER_TESS_CTRL
);
271 uint32_t input_vertex_size
= ctx
->tcs_num_inputs
* 16;
272 uint32_t input_patch_size
= ctx
->args
->options
->key
.tcs
.input_vertices
* input_vertex_size
;
273 uint32_t output_patch0_offset
= input_patch_size
;
275 uint32_t num_tcs_outputs
= util_last_bit64(ctx
->args
->shader_info
->tcs
.outputs_written
);
276 uint32_t output_vertex_size
= num_tcs_outputs
* 16;
277 uint32_t pervertex_output_patch_size
= ctx
->shader
->info
.tess
.tcs_vertices_out
* output_vertex_size
;
278 unsigned num_patches
= ctx
->tcs_num_patches
;
280 output_patch0_offset
*= num_patches
;
281 output_patch0_offset
+= pervertex_output_patch_size
;
282 output_patch0_offset
/= 4;
283 return LLVMConstInt(ctx
->ac
.i32
, output_patch0_offset
, false);
287 get_tcs_in_current_patch_offset(struct radv_shader_context
*ctx
)
289 LLVMValueRef patch_stride
= get_tcs_in_patch_stride(ctx
);
290 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
292 return LLVMBuildMul(ctx
->ac
.builder
, patch_stride
, rel_patch_id
, "");
296 get_tcs_out_current_patch_offset(struct radv_shader_context
*ctx
)
298 LLVMValueRef patch0_offset
= get_tcs_out_patch0_offset(ctx
);
299 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(ctx
);
300 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
302 return ac_build_imad(&ctx
->ac
, patch_stride
, rel_patch_id
,
307 get_tcs_out_current_patch_data_offset(struct radv_shader_context
*ctx
)
309 LLVMValueRef patch0_patch_data_offset
=
310 get_tcs_out_patch0_patch_data_offset(ctx
);
311 LLVMValueRef patch_stride
= get_tcs_out_patch_stride(ctx
);
312 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
314 return ac_build_imad(&ctx
->ac
, patch_stride
, rel_patch_id
,
315 patch0_patch_data_offset
);
319 create_llvm_function(struct ac_llvm_context
*ctx
, LLVMModuleRef module
,
320 LLVMBuilderRef builder
,
321 const struct ac_shader_args
*args
,
322 enum ac_llvm_calling_convention convention
,
323 unsigned max_workgroup_size
,
324 const struct radv_nir_compiler_options
*options
)
326 LLVMValueRef main_function
=
327 ac_build_main(args
, ctx
, convention
, "main", ctx
->voidt
, module
);
329 if (options
->address32_hi
) {
330 ac_llvm_add_target_dep_function_attr(main_function
,
331 "amdgpu-32bit-address-high-bits",
332 options
->address32_hi
);
335 ac_llvm_set_workgroup_size(main_function
, max_workgroup_size
);
337 return main_function
;
341 load_descriptor_sets(struct radv_shader_context
*ctx
)
343 uint32_t mask
= ctx
->args
->shader_info
->desc_set_used_mask
;
344 if (ctx
->args
->shader_info
->need_indirect_descriptor_sets
) {
345 LLVMValueRef desc_sets
=
346 ac_get_arg(&ctx
->ac
, ctx
->args
->descriptor_sets
[0]);
348 int i
= u_bit_scan(&mask
);
350 ctx
->descriptor_sets
[i
] =
351 ac_build_load_to_sgpr(&ctx
->ac
, desc_sets
,
352 LLVMConstInt(ctx
->ac
.i32
, i
, false));
357 int i
= u_bit_scan(&mask
);
359 ctx
->descriptor_sets
[i
] =
360 ac_get_arg(&ctx
->ac
, ctx
->args
->descriptor_sets
[i
]);
365 static enum ac_llvm_calling_convention
366 get_llvm_calling_convention(LLVMValueRef func
, gl_shader_stage stage
)
369 case MESA_SHADER_VERTEX
:
370 case MESA_SHADER_TESS_EVAL
:
371 return AC_LLVM_AMDGPU_VS
;
373 case MESA_SHADER_GEOMETRY
:
374 return AC_LLVM_AMDGPU_GS
;
376 case MESA_SHADER_TESS_CTRL
:
377 return AC_LLVM_AMDGPU_HS
;
379 case MESA_SHADER_FRAGMENT
:
380 return AC_LLVM_AMDGPU_PS
;
382 case MESA_SHADER_COMPUTE
:
383 return AC_LLVM_AMDGPU_CS
;
386 unreachable("Unhandle shader type");
390 /* Returns whether the stage is a stage that can be directly before the GS */
391 static bool is_pre_gs_stage(gl_shader_stage stage
)
393 return stage
== MESA_SHADER_VERTEX
|| stage
== MESA_SHADER_TESS_EVAL
;
396 static void create_function(struct radv_shader_context
*ctx
,
397 gl_shader_stage stage
,
398 bool has_previous_stage
)
400 if (ctx
->ac
.chip_class
>= GFX10
) {
401 if (is_pre_gs_stage(stage
) && ctx
->args
->options
->key
.vs_common_out
.as_ngg
) {
402 /* On GFX10, VS is merged into GS for NGG. */
403 stage
= MESA_SHADER_GEOMETRY
;
404 has_previous_stage
= true;
408 ctx
->main_function
= create_llvm_function(
409 &ctx
->ac
, ctx
->ac
.module
, ctx
->ac
.builder
, &ctx
->args
->ac
,
410 get_llvm_calling_convention(ctx
->main_function
, stage
),
411 ctx
->max_workgroup_size
,
414 ctx
->ring_offsets
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.implicit.buffer.ptr",
415 LLVMPointerType(ctx
->ac
.i8
, AC_ADDR_SPACE_CONST
),
416 NULL
, 0, AC_FUNC_ATTR_READNONE
);
417 ctx
->ring_offsets
= LLVMBuildBitCast(ctx
->ac
.builder
, ctx
->ring_offsets
,
418 ac_array_in_const_addr_space(ctx
->ac
.v4i32
), "");
420 load_descriptor_sets(ctx
);
422 if (stage
== MESA_SHADER_TESS_CTRL
||
423 (stage
== MESA_SHADER_VERTEX
&& ctx
->args
->options
->key
.vs_common_out
.as_ls
) ||
424 /* GFX9 has the ESGS ring buffer in LDS. */
425 (stage
== MESA_SHADER_GEOMETRY
&& has_previous_stage
)) {
426 ac_declare_lds_as_pointer(&ctx
->ac
);
433 radv_load_resource(struct ac_shader_abi
*abi
, LLVMValueRef index
,
434 unsigned desc_set
, unsigned binding
)
436 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
437 LLVMValueRef desc_ptr
= ctx
->descriptor_sets
[desc_set
];
438 struct radv_pipeline_layout
*pipeline_layout
= ctx
->args
->options
->layout
;
439 struct radv_descriptor_set_layout
*layout
= pipeline_layout
->set
[desc_set
].layout
;
440 unsigned base_offset
= layout
->binding
[binding
].offset
;
441 LLVMValueRef offset
, stride
;
443 if (layout
->binding
[binding
].type
== VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC
||
444 layout
->binding
[binding
].type
== VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC
) {
445 unsigned idx
= pipeline_layout
->set
[desc_set
].dynamic_offset_start
+
446 layout
->binding
[binding
].dynamic_offset_offset
;
447 desc_ptr
= ac_get_arg(&ctx
->ac
, ctx
->args
->ac
.push_constants
);
448 base_offset
= pipeline_layout
->push_constant_size
+ 16 * idx
;
449 stride
= LLVMConstInt(ctx
->ac
.i32
, 16, false);
451 stride
= LLVMConstInt(ctx
->ac
.i32
, layout
->binding
[binding
].size
, false);
453 offset
= LLVMConstInt(ctx
->ac
.i32
, base_offset
, false);
455 if (layout
->binding
[binding
].type
!= VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT
) {
456 offset
= ac_build_imad(&ctx
->ac
, index
, stride
, offset
);
459 desc_ptr
= LLVMBuildGEP(ctx
->ac
.builder
, desc_ptr
, &offset
, 1, "");
460 desc_ptr
= ac_cast_ptr(&ctx
->ac
, desc_ptr
, ctx
->ac
.v4i32
);
461 LLVMSetMetadata(desc_ptr
, ctx
->ac
.uniform_md_kind
, ctx
->ac
.empty_md
);
463 if (layout
->binding
[binding
].type
== VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT
) {
464 uint32_t desc_type
= S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
465 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
466 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
467 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
);
469 if (ctx
->ac
.chip_class
>= GFX10
) {
470 desc_type
|= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT
) |
471 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_RAW
) |
472 S_008F0C_RESOURCE_LEVEL(1);
474 desc_type
|= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
475 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
);
478 LLVMValueRef desc_components
[4] = {
479 LLVMBuildPtrToInt(ctx
->ac
.builder
, desc_ptr
, ctx
->ac
.intptr
, ""),
480 LLVMConstInt(ctx
->ac
.i32
, S_008F04_BASE_ADDRESS_HI(ctx
->args
->options
->address32_hi
), false),
481 /* High limit to support variable sizes. */
482 LLVMConstInt(ctx
->ac
.i32
, 0xffffffff, false),
483 LLVMConstInt(ctx
->ac
.i32
, desc_type
, false),
486 return ac_build_gather_values(&ctx
->ac
, desc_components
, 4);
493 /* The offchip buffer layout for TCS->TES is
495 * - attribute 0 of patch 0 vertex 0
496 * - attribute 0 of patch 0 vertex 1
497 * - attribute 0 of patch 0 vertex 2
499 * - attribute 0 of patch 1 vertex 0
500 * - attribute 0 of patch 1 vertex 1
502 * - attribute 1 of patch 0 vertex 0
503 * - attribute 1 of patch 0 vertex 1
505 * - per patch attribute 0 of patch 0
506 * - per patch attribute 0 of patch 1
509 * Note that every attribute has 4 components.
511 static LLVMValueRef
get_non_vertex_index_offset(struct radv_shader_context
*ctx
)
513 uint32_t num_patches
= ctx
->tcs_num_patches
;
514 uint32_t num_tcs_outputs
;
515 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
)
516 num_tcs_outputs
= util_last_bit64(ctx
->args
->shader_info
->tcs
.outputs_written
);
518 num_tcs_outputs
= ctx
->args
->options
->key
.tes
.tcs_num_outputs
;
520 uint32_t output_vertex_size
= num_tcs_outputs
* 16;
521 uint32_t pervertex_output_patch_size
= ctx
->shader
->info
.tess
.tcs_vertices_out
* output_vertex_size
;
523 return LLVMConstInt(ctx
->ac
.i32
, pervertex_output_patch_size
* num_patches
, false);
526 static LLVMValueRef
calc_param_stride(struct radv_shader_context
*ctx
,
527 LLVMValueRef vertex_index
)
529 LLVMValueRef param_stride
;
531 param_stride
= LLVMConstInt(ctx
->ac
.i32
, ctx
->shader
->info
.tess
.tcs_vertices_out
* ctx
->tcs_num_patches
, false);
533 param_stride
= LLVMConstInt(ctx
->ac
.i32
, ctx
->tcs_num_patches
, false);
537 static LLVMValueRef
get_tcs_tes_buffer_address(struct radv_shader_context
*ctx
,
538 LLVMValueRef vertex_index
,
539 LLVMValueRef param_index
)
541 LLVMValueRef base_addr
;
542 LLVMValueRef param_stride
, constant16
;
543 LLVMValueRef rel_patch_id
= get_rel_patch_id(ctx
);
544 LLVMValueRef vertices_per_patch
= LLVMConstInt(ctx
->ac
.i32
, ctx
->shader
->info
.tess
.tcs_vertices_out
, false);
545 constant16
= LLVMConstInt(ctx
->ac
.i32
, 16, false);
546 param_stride
= calc_param_stride(ctx
, vertex_index
);
548 base_addr
= ac_build_imad(&ctx
->ac
, rel_patch_id
,
549 vertices_per_patch
, vertex_index
);
551 base_addr
= rel_patch_id
;
554 base_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_addr
,
555 LLVMBuildMul(ctx
->ac
.builder
, param_index
,
556 param_stride
, ""), "");
558 base_addr
= LLVMBuildMul(ctx
->ac
.builder
, base_addr
, constant16
, "");
561 LLVMValueRef patch_data_offset
= get_non_vertex_index_offset(ctx
);
563 base_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_addr
,
564 patch_data_offset
, "");
569 static LLVMValueRef
get_tcs_tes_buffer_address_params(struct radv_shader_context
*ctx
,
571 unsigned const_index
,
573 LLVMValueRef vertex_index
,
574 LLVMValueRef indir_index
)
576 LLVMValueRef param_index
;
579 param_index
= LLVMBuildAdd(ctx
->ac
.builder
, LLVMConstInt(ctx
->ac
.i32
, param
, false),
582 if (const_index
&& !is_compact
)
583 param
+= const_index
;
584 param_index
= LLVMConstInt(ctx
->ac
.i32
, param
, false);
586 return get_tcs_tes_buffer_address(ctx
, vertex_index
, param_index
);
590 get_dw_address(struct radv_shader_context
*ctx
,
591 LLVMValueRef dw_addr
,
593 unsigned const_index
,
594 bool compact_const_index
,
595 LLVMValueRef vertex_index
,
597 LLVMValueRef indir_index
)
602 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
603 LLVMBuildMul(ctx
->ac
.builder
,
609 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
610 LLVMBuildMul(ctx
->ac
.builder
, indir_index
,
611 LLVMConstInt(ctx
->ac
.i32
, 4, false), ""), "");
612 else if (const_index
&& !compact_const_index
)
613 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
614 LLVMConstInt(ctx
->ac
.i32
, const_index
* 4, false), "");
616 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
617 LLVMConstInt(ctx
->ac
.i32
, param
* 4, false), "");
619 if (const_index
&& compact_const_index
)
620 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
621 LLVMConstInt(ctx
->ac
.i32
, const_index
, false), "");
626 load_tcs_varyings(struct ac_shader_abi
*abi
,
628 LLVMValueRef vertex_index
,
629 LLVMValueRef indir_index
,
630 unsigned const_index
,
632 unsigned driver_location
,
634 unsigned num_components
,
639 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
640 LLVMValueRef dw_addr
, stride
;
641 LLVMValueRef value
[4], result
;
642 unsigned param
= shader_io_get_unique_index(location
);
645 uint32_t input_vertex_size
= (ctx
->tcs_num_inputs
* 16) / 4;
646 stride
= LLVMConstInt(ctx
->ac
.i32
, input_vertex_size
, false);
647 dw_addr
= get_tcs_in_current_patch_offset(ctx
);
650 stride
= get_tcs_out_vertex_stride(ctx
);
651 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
653 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
658 dw_addr
= get_dw_address(ctx
, dw_addr
, param
, const_index
, is_compact
, vertex_index
, stride
,
661 for (unsigned i
= 0; i
< num_components
+ component
; i
++) {
662 value
[i
] = ac_lds_load(&ctx
->ac
, dw_addr
);
663 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
666 result
= ac_build_varying_gather_values(&ctx
->ac
, value
, num_components
, component
);
671 store_tcs_output(struct ac_shader_abi
*abi
,
672 const nir_variable
*var
,
673 LLVMValueRef vertex_index
,
674 LLVMValueRef param_index
,
675 unsigned const_index
,
679 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
680 const unsigned location
= var
->data
.location
;
681 unsigned component
= var
->data
.location_frac
;
682 const bool is_patch
= var
->data
.patch
;
683 const bool is_compact
= var
->data
.compact
;
684 LLVMValueRef dw_addr
;
685 LLVMValueRef stride
= NULL
;
686 LLVMValueRef buf_addr
= NULL
;
687 LLVMValueRef oc_lds
= ac_get_arg(&ctx
->ac
, ctx
->args
->oc_lds
);
689 bool store_lds
= true;
692 if (!(ctx
->shader
->info
.patch_outputs_read
& (1U << (location
- VARYING_SLOT_PATCH0
))))
695 if (!(ctx
->shader
->info
.outputs_read
& (1ULL << location
)))
699 param
= shader_io_get_unique_index(location
);
700 if ((location
== VARYING_SLOT_CLIP_DIST0
|| location
== VARYING_SLOT_CLIP_DIST1
) && is_compact
) {
701 const_index
+= component
;
704 if (const_index
>= 4) {
711 stride
= get_tcs_out_vertex_stride(ctx
);
712 dw_addr
= get_tcs_out_current_patch_offset(ctx
);
714 dw_addr
= get_tcs_out_current_patch_data_offset(ctx
);
717 dw_addr
= get_dw_address(ctx
, dw_addr
, param
, const_index
, is_compact
, vertex_index
, stride
,
719 buf_addr
= get_tcs_tes_buffer_address_params(ctx
, param
, const_index
, is_compact
,
720 vertex_index
, param_index
);
722 bool is_tess_factor
= false;
723 if (location
== VARYING_SLOT_TESS_LEVEL_INNER
||
724 location
== VARYING_SLOT_TESS_LEVEL_OUTER
)
725 is_tess_factor
= true;
727 unsigned base
= is_compact
? const_index
: 0;
728 for (unsigned chan
= 0; chan
< 8; chan
++) {
729 if (!(writemask
& (1 << chan
)))
731 LLVMValueRef value
= ac_llvm_extract_elem(&ctx
->ac
, src
, chan
- component
);
732 value
= ac_to_integer(&ctx
->ac
, value
);
733 value
= LLVMBuildZExtOrBitCast(ctx
->ac
.builder
, value
, ctx
->ac
.i32
, "");
735 if (store_lds
|| is_tess_factor
) {
736 LLVMValueRef dw_addr_chan
=
737 LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
738 LLVMConstInt(ctx
->ac
.i32
, chan
, false), "");
739 ac_lds_store(&ctx
->ac
, dw_addr_chan
, value
);
742 if (!is_tess_factor
&& writemask
!= 0xF)
743 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, value
, 1,
745 4 * (base
+ chan
), ac_glc
);
748 if (writemask
== 0xF) {
749 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, src
, 4,
756 load_tes_input(struct ac_shader_abi
*abi
,
758 LLVMValueRef vertex_index
,
759 LLVMValueRef param_index
,
760 unsigned const_index
,
762 unsigned driver_location
,
764 unsigned num_components
,
769 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
770 LLVMValueRef buf_addr
;
772 LLVMValueRef oc_lds
= ac_get_arg(&ctx
->ac
, ctx
->args
->oc_lds
);
773 unsigned param
= shader_io_get_unique_index(location
);
775 if ((location
== VARYING_SLOT_CLIP_DIST0
|| location
== VARYING_SLOT_CLIP_DIST1
) && is_compact
) {
776 const_index
+= component
;
778 if (const_index
>= 4) {
784 buf_addr
= get_tcs_tes_buffer_address_params(ctx
, param
, const_index
,
785 is_compact
, vertex_index
, param_index
);
787 LLVMValueRef comp_offset
= LLVMConstInt(ctx
->ac
.i32
, component
* 4, false);
788 buf_addr
= LLVMBuildAdd(ctx
->ac
.builder
, buf_addr
, comp_offset
, "");
790 result
= ac_build_buffer_load(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, num_components
, NULL
,
791 buf_addr
, oc_lds
, is_compact
? (4 * const_index
) : 0, ac_glc
, true, false);
792 result
= ac_trim_vector(&ctx
->ac
, result
, num_components
);
797 radv_emit_fetch_64bit(struct radv_shader_context
*ctx
,
798 LLVMTypeRef type
, LLVMValueRef a
, LLVMValueRef b
)
800 LLVMValueRef values
[2] = {
801 ac_to_integer(&ctx
->ac
, a
),
802 ac_to_integer(&ctx
->ac
, b
),
804 LLVMValueRef result
= ac_build_gather_values(&ctx
->ac
, values
, 2);
805 return LLVMBuildBitCast(ctx
->ac
.builder
, result
, type
, "");
809 load_gs_input(struct ac_shader_abi
*abi
,
811 unsigned driver_location
,
813 unsigned num_components
,
814 unsigned vertex_index
,
815 unsigned const_index
,
818 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
819 LLVMValueRef vtx_offset
;
820 unsigned param
, vtx_offset_param
;
821 LLVMValueRef value
[4], result
;
823 vtx_offset_param
= vertex_index
;
824 assert(vtx_offset_param
< 6);
825 vtx_offset
= LLVMBuildMul(ctx
->ac
.builder
, ctx
->gs_vtx_offset
[vtx_offset_param
],
826 LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
828 param
= shader_io_get_unique_index(location
);
830 for (unsigned i
= component
; i
< num_components
+ component
; i
++) {
831 if (ctx
->ac
.chip_class
>= GFX9
) {
832 LLVMValueRef dw_addr
= ctx
->gs_vtx_offset
[vtx_offset_param
];
833 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
834 LLVMConstInt(ctx
->ac
.i32
, param
* 4 + i
+ const_index
, 0), "");
835 value
[i
] = ac_lds_load(&ctx
->ac
, dw_addr
);
837 if (ac_get_type_size(type
) == 8) {
838 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
839 LLVMConstInt(ctx
->ac
.i32
, param
* 4 + i
+ const_index
+ 1, 0), "");
840 LLVMValueRef tmp
= ac_lds_load(&ctx
->ac
, dw_addr
);
842 value
[i
] = radv_emit_fetch_64bit(ctx
, type
, value
[i
], tmp
);
845 LLVMValueRef soffset
=
846 LLVMConstInt(ctx
->ac
.i32
,
847 (param
* 4 + i
+ const_index
) * 256,
850 value
[i
] = ac_build_buffer_load(&ctx
->ac
,
854 0, ac_glc
, true, false);
856 if (ac_get_type_size(type
) == 8) {
857 soffset
= LLVMConstInt(ctx
->ac
.i32
,
858 (param
* 4 + i
+ const_index
+ 1) * 256,
862 ac_build_buffer_load(&ctx
->ac
,
866 0, ac_glc
, true, false);
868 value
[i
] = radv_emit_fetch_64bit(ctx
, type
, value
[i
], tmp
);
872 if (ac_get_type_size(type
) == 2) {
873 value
[i
] = LLVMBuildBitCast(ctx
->ac
.builder
, value
[i
], ctx
->ac
.i32
, "");
874 value
[i
] = LLVMBuildTrunc(ctx
->ac
.builder
, value
[i
], ctx
->ac
.i16
, "");
876 value
[i
] = LLVMBuildBitCast(ctx
->ac
.builder
, value
[i
], type
, "");
878 result
= ac_build_varying_gather_values(&ctx
->ac
, value
, num_components
, component
);
879 result
= ac_to_integer(&ctx
->ac
, result
);
884 static void radv_emit_kill(struct ac_shader_abi
*abi
, LLVMValueRef visible
)
886 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
887 ac_build_kill_if_false(&ctx
->ac
, visible
);
891 radv_get_sample_pos_offset(uint32_t num_samples
)
893 uint32_t sample_pos_offset
= 0;
895 switch (num_samples
) {
897 sample_pos_offset
= 1;
900 sample_pos_offset
= 3;
903 sample_pos_offset
= 7;
908 return sample_pos_offset
;
911 static LLVMValueRef
load_sample_position(struct ac_shader_abi
*abi
,
912 LLVMValueRef sample_id
)
914 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
917 LLVMValueRef index
= LLVMConstInt(ctx
->ac
.i32
, RING_PS_SAMPLE_POSITIONS
, false);
918 LLVMValueRef ptr
= LLVMBuildGEP(ctx
->ac
.builder
, ctx
->ring_offsets
, &index
, 1, "");
920 ptr
= LLVMBuildBitCast(ctx
->ac
.builder
, ptr
,
921 ac_array_in_const_addr_space(ctx
->ac
.v2f32
), "");
923 uint32_t sample_pos_offset
=
924 radv_get_sample_pos_offset(ctx
->args
->options
->key
.fs
.num_samples
);
927 LLVMBuildAdd(ctx
->ac
.builder
, sample_id
,
928 LLVMConstInt(ctx
->ac
.i32
, sample_pos_offset
, false), "");
929 result
= ac_build_load_invariant(&ctx
->ac
, ptr
, sample_id
);
935 static LLVMValueRef
load_sample_mask_in(struct ac_shader_abi
*abi
)
937 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
938 uint8_t log2_ps_iter_samples
;
940 if (ctx
->args
->shader_info
->ps
.force_persample
) {
941 log2_ps_iter_samples
=
942 util_logbase2(ctx
->args
->options
->key
.fs
.num_samples
);
944 log2_ps_iter_samples
= ctx
->args
->options
->key
.fs
.log2_ps_iter_samples
;
947 /* The bit pattern matches that used by fixed function fragment
949 static const uint16_t ps_iter_masks
[] = {
950 0xffff, /* not used */
956 assert(log2_ps_iter_samples
< ARRAY_SIZE(ps_iter_masks
));
958 uint32_t ps_iter_mask
= ps_iter_masks
[log2_ps_iter_samples
];
960 LLVMValueRef result
, sample_id
;
961 sample_id
= ac_unpack_param(&ctx
->ac
, ac_get_arg(&ctx
->ac
, ctx
->args
->ac
.ancillary
), 8, 4);
962 sample_id
= LLVMBuildShl(ctx
->ac
.builder
, LLVMConstInt(ctx
->ac
.i32
, ps_iter_mask
, false), sample_id
, "");
963 result
= LLVMBuildAnd(ctx
->ac
.builder
, sample_id
,
964 ac_get_arg(&ctx
->ac
, ctx
->args
->ac
.sample_coverage
), "");
969 static void gfx10_ngg_gs_emit_vertex(struct radv_shader_context
*ctx
,
971 LLVMValueRef
*addrs
);
974 visit_emit_vertex(struct ac_shader_abi
*abi
, unsigned stream
, LLVMValueRef
*addrs
)
976 LLVMValueRef gs_next_vertex
;
977 LLVMValueRef can_emit
;
979 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
981 if (ctx
->args
->options
->key
.vs_common_out
.as_ngg
) {
982 gfx10_ngg_gs_emit_vertex(ctx
, stream
, addrs
);
986 /* Write vertex attribute values to GSVS ring */
987 gs_next_vertex
= LLVMBuildLoad(ctx
->ac
.builder
,
988 ctx
->gs_next_vertex
[stream
],
991 /* If this thread has already emitted the declared maximum number of
992 * vertices, don't emit any more: excessive vertex emissions are not
993 * supposed to have any effect.
995 can_emit
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntULT
, gs_next_vertex
,
996 LLVMConstInt(ctx
->ac
.i32
, ctx
->shader
->info
.gs
.vertices_out
, false), "");
998 bool use_kill
= !ctx
->args
->shader_info
->gs
.writes_memory
;
1000 ac_build_kill_if_false(&ctx
->ac
, can_emit
);
1002 ac_build_ifcc(&ctx
->ac
, can_emit
, 6505);
1004 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
1005 unsigned output_usage_mask
=
1006 ctx
->args
->shader_info
->gs
.output_usage_mask
[i
];
1007 uint8_t output_stream
=
1008 ctx
->args
->shader_info
->gs
.output_streams
[i
];
1009 LLVMValueRef
*out_ptr
= &addrs
[i
* 4];
1010 int length
= util_last_bit(output_usage_mask
);
1012 if (!(ctx
->output_mask
& (1ull << i
)) ||
1013 output_stream
!= stream
)
1016 for (unsigned j
= 0; j
< length
; j
++) {
1017 if (!(output_usage_mask
& (1 << j
)))
1020 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->ac
.builder
,
1022 LLVMValueRef voffset
=
1023 LLVMConstInt(ctx
->ac
.i32
, offset
*
1024 ctx
->shader
->info
.gs
.vertices_out
, false);
1028 voffset
= LLVMBuildAdd(ctx
->ac
.builder
, voffset
, gs_next_vertex
, "");
1029 voffset
= LLVMBuildMul(ctx
->ac
.builder
, voffset
, LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
1031 out_val
= ac_to_integer(&ctx
->ac
, out_val
);
1032 out_val
= LLVMBuildZExtOrBitCast(ctx
->ac
.builder
, out_val
, ctx
->ac
.i32
, "");
1034 ac_build_buffer_store_dword(&ctx
->ac
,
1035 ctx
->gsvs_ring
[stream
],
1038 ac_get_arg(&ctx
->ac
,
1039 ctx
->args
->gs2vs_offset
),
1040 0, ac_glc
| ac_slc
| ac_swizzled
);
1044 gs_next_vertex
= LLVMBuildAdd(ctx
->ac
.builder
, gs_next_vertex
,
1046 LLVMBuildStore(ctx
->ac
.builder
, gs_next_vertex
, ctx
->gs_next_vertex
[stream
]);
1048 ac_build_sendmsg(&ctx
->ac
,
1049 AC_SENDMSG_GS_OP_EMIT
| AC_SENDMSG_GS
| (stream
<< 8),
1053 ac_build_endif(&ctx
->ac
, 6505);
1057 visit_end_primitive(struct ac_shader_abi
*abi
, unsigned stream
)
1059 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1061 if (ctx
->args
->options
->key
.vs_common_out
.as_ngg
) {
1062 LLVMBuildStore(ctx
->ac
.builder
, ctx
->ac
.i32_0
, ctx
->gs_curprim_verts
[stream
]);
1066 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_CUT
| AC_SENDMSG_GS
| (stream
<< 8), ctx
->gs_wave_id
);
1070 load_tess_coord(struct ac_shader_abi
*abi
)
1072 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1074 LLVMValueRef coord
[4] = {
1075 ac_get_arg(&ctx
->ac
, ctx
->args
->tes_u
),
1076 ac_get_arg(&ctx
->ac
, ctx
->args
->tes_v
),
1081 if (ctx
->shader
->info
.tess
.primitive_mode
== GL_TRIANGLES
)
1082 coord
[2] = LLVMBuildFSub(ctx
->ac
.builder
, ctx
->ac
.f32_1
,
1083 LLVMBuildFAdd(ctx
->ac
.builder
, coord
[0], coord
[1], ""), "");
1085 return ac_build_gather_values(&ctx
->ac
, coord
, 3);
1089 load_patch_vertices_in(struct ac_shader_abi
*abi
)
1091 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1092 return LLVMConstInt(ctx
->ac
.i32
, ctx
->args
->options
->key
.tcs
.input_vertices
, false);
1096 static LLVMValueRef
radv_load_base_vertex(struct ac_shader_abi
*abi
)
1098 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1099 return ac_get_arg(&ctx
->ac
, ctx
->args
->ac
.base_vertex
);
1102 static LLVMValueRef
radv_load_ssbo(struct ac_shader_abi
*abi
,
1103 LLVMValueRef buffer_ptr
, bool write
)
1105 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1106 LLVMValueRef result
;
1108 LLVMSetMetadata(buffer_ptr
, ctx
->ac
.uniform_md_kind
, ctx
->ac
.empty_md
);
1110 result
= LLVMBuildLoad(ctx
->ac
.builder
, buffer_ptr
, "");
1111 LLVMSetMetadata(result
, ctx
->ac
.invariant_load_md_kind
, ctx
->ac
.empty_md
);
1116 static LLVMValueRef
radv_load_ubo(struct ac_shader_abi
*abi
, LLVMValueRef buffer_ptr
)
1118 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1119 LLVMValueRef result
;
1121 if (LLVMGetTypeKind(LLVMTypeOf(buffer_ptr
)) != LLVMPointerTypeKind
) {
1122 /* Do not load the descriptor for inlined uniform blocks. */
1126 LLVMSetMetadata(buffer_ptr
, ctx
->ac
.uniform_md_kind
, ctx
->ac
.empty_md
);
1128 result
= LLVMBuildLoad(ctx
->ac
.builder
, buffer_ptr
, "");
1129 LLVMSetMetadata(result
, ctx
->ac
.invariant_load_md_kind
, ctx
->ac
.empty_md
);
1134 static LLVMValueRef
radv_get_sampler_desc(struct ac_shader_abi
*abi
,
1135 unsigned descriptor_set
,
1136 unsigned base_index
,
1137 unsigned constant_index
,
1139 enum ac_descriptor_type desc_type
,
1140 bool image
, bool write
,
1143 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
1144 LLVMValueRef list
= ctx
->descriptor_sets
[descriptor_set
];
1145 struct radv_descriptor_set_layout
*layout
= ctx
->args
->options
->layout
->set
[descriptor_set
].layout
;
1146 struct radv_descriptor_set_binding_layout
*binding
= layout
->binding
+ base_index
;
1147 unsigned offset
= binding
->offset
;
1148 unsigned stride
= binding
->size
;
1150 LLVMBuilderRef builder
= ctx
->ac
.builder
;
1153 assert(base_index
< layout
->binding_count
);
1155 switch (desc_type
) {
1157 type
= ctx
->ac
.v8i32
;
1161 type
= ctx
->ac
.v8i32
;
1165 case AC_DESC_SAMPLER
:
1166 type
= ctx
->ac
.v4i32
;
1167 if (binding
->type
== VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER
) {
1168 offset
+= radv_combined_image_descriptor_sampler_offset(binding
);
1173 case AC_DESC_BUFFER
:
1174 type
= ctx
->ac
.v4i32
;
1177 case AC_DESC_PLANE_0
:
1178 case AC_DESC_PLANE_1
:
1179 case AC_DESC_PLANE_2
:
1180 type
= ctx
->ac
.v8i32
;
1182 offset
+= 32 * (desc_type
- AC_DESC_PLANE_0
);
1185 unreachable("invalid desc_type\n");
1188 offset
+= constant_index
* stride
;
1190 if (desc_type
== AC_DESC_SAMPLER
&& binding
->immutable_samplers_offset
&&
1191 (!index
|| binding
->immutable_samplers_equal
)) {
1192 if (binding
->immutable_samplers_equal
)
1195 const uint32_t *samplers
= radv_immutable_samplers(layout
, binding
);
1197 LLVMValueRef constants
[] = {
1198 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 0], 0),
1199 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 1], 0),
1200 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 2], 0),
1201 LLVMConstInt(ctx
->ac
.i32
, samplers
[constant_index
* 4 + 3], 0),
1203 return ac_build_gather_values(&ctx
->ac
, constants
, 4);
1206 assert(stride
% type_size
== 0);
1208 LLVMValueRef adjusted_index
= index
;
1209 if (!adjusted_index
)
1210 adjusted_index
= ctx
->ac
.i32_0
;
1212 adjusted_index
= LLVMBuildMul(builder
, adjusted_index
, LLVMConstInt(ctx
->ac
.i32
, stride
/ type_size
, 0), "");
1214 LLVMValueRef val_offset
= LLVMConstInt(ctx
->ac
.i32
, offset
, 0);
1215 list
= LLVMBuildGEP(builder
, list
, &val_offset
, 1, "");
1216 list
= LLVMBuildPointerCast(builder
, list
,
1217 ac_array_in_const32_addr_space(type
), "");
1219 LLVMValueRef descriptor
= ac_build_load_to_sgpr(&ctx
->ac
, list
, adjusted_index
);
1221 /* 3 plane formats always have same size and format for plane 1 & 2, so
1222 * use the tail from plane 1 so that we can store only the first 16 bytes
1223 * of the last plane. */
1224 if (desc_type
== AC_DESC_PLANE_2
) {
1225 LLVMValueRef descriptor2
= radv_get_sampler_desc(abi
, descriptor_set
, base_index
, constant_index
, index
, AC_DESC_PLANE_1
,image
, write
, bindless
);
1227 LLVMValueRef components
[8];
1228 for (unsigned i
= 0; i
< 4; ++i
)
1229 components
[i
] = ac_llvm_extract_elem(&ctx
->ac
, descriptor
, i
);
1231 for (unsigned i
= 4; i
< 8; ++i
)
1232 components
[i
] = ac_llvm_extract_elem(&ctx
->ac
, descriptor2
, i
);
1233 descriptor
= ac_build_gather_values(&ctx
->ac
, components
, 8);
1239 /* For 2_10_10_10 formats the alpha is handled as unsigned by pre-vega HW.
1240 * so we may need to fix it up. */
1242 adjust_vertex_fetch_alpha(struct radv_shader_context
*ctx
,
1243 unsigned adjustment
,
1246 if (adjustment
== RADV_ALPHA_ADJUST_NONE
)
1249 LLVMValueRef c30
= LLVMConstInt(ctx
->ac
.i32
, 30, 0);
1251 alpha
= LLVMBuildBitCast(ctx
->ac
.builder
, alpha
, ctx
->ac
.f32
, "");
1253 if (adjustment
== RADV_ALPHA_ADJUST_SSCALED
)
1254 alpha
= LLVMBuildFPToUI(ctx
->ac
.builder
, alpha
, ctx
->ac
.i32
, "");
1256 alpha
= ac_to_integer(&ctx
->ac
, alpha
);
1258 /* For the integer-like cases, do a natural sign extension.
1260 * For the SNORM case, the values are 0.0, 0.333, 0.666, 1.0
1261 * and happen to contain 0, 1, 2, 3 as the two LSBs of the
1264 alpha
= LLVMBuildShl(ctx
->ac
.builder
, alpha
,
1265 adjustment
== RADV_ALPHA_ADJUST_SNORM
?
1266 LLVMConstInt(ctx
->ac
.i32
, 7, 0) : c30
, "");
1267 alpha
= LLVMBuildAShr(ctx
->ac
.builder
, alpha
, c30
, "");
1269 /* Convert back to the right type. */
1270 if (adjustment
== RADV_ALPHA_ADJUST_SNORM
) {
1272 LLVMValueRef neg_one
= LLVMConstReal(ctx
->ac
.f32
, -1.0);
1273 alpha
= LLVMBuildSIToFP(ctx
->ac
.builder
, alpha
, ctx
->ac
.f32
, "");
1274 clamp
= LLVMBuildFCmp(ctx
->ac
.builder
, LLVMRealULT
, alpha
, neg_one
, "");
1275 alpha
= LLVMBuildSelect(ctx
->ac
.builder
, clamp
, neg_one
, alpha
, "");
1276 } else if (adjustment
== RADV_ALPHA_ADJUST_SSCALED
) {
1277 alpha
= LLVMBuildSIToFP(ctx
->ac
.builder
, alpha
, ctx
->ac
.f32
, "");
1280 return LLVMBuildBitCast(ctx
->ac
.builder
, alpha
, ctx
->ac
.i32
, "");
1283 static const struct vertex_format_info
{
1284 uint8_t vertex_byte_size
;
1285 uint8_t num_channels
;
1286 uint8_t chan_byte_size
;
1287 uint8_t chan_format
;
1288 } vertex_format_table
[] = {
1289 { 0, 4, 0, V_008F0C_BUF_DATA_FORMAT_INVALID
}, /* BUF_DATA_FORMAT_INVALID */
1290 { 1, 1, 1, V_008F0C_BUF_DATA_FORMAT_8
}, /* BUF_DATA_FORMAT_8 */
1291 { 2, 1, 2, V_008F0C_BUF_DATA_FORMAT_16
}, /* BUF_DATA_FORMAT_16 */
1292 { 2, 2, 1, V_008F0C_BUF_DATA_FORMAT_8
}, /* BUF_DATA_FORMAT_8_8 */
1293 { 4, 1, 4, V_008F0C_BUF_DATA_FORMAT_32
}, /* BUF_DATA_FORMAT_32 */
1294 { 4, 2, 2, V_008F0C_BUF_DATA_FORMAT_16
}, /* BUF_DATA_FORMAT_16_16 */
1295 { 4, 3, 0, V_008F0C_BUF_DATA_FORMAT_10_11_11
}, /* BUF_DATA_FORMAT_10_11_11 */
1296 { 4, 3, 0, V_008F0C_BUF_DATA_FORMAT_11_11_10
}, /* BUF_DATA_FORMAT_11_11_10 */
1297 { 4, 4, 0, V_008F0C_BUF_DATA_FORMAT_10_10_10_2
}, /* BUF_DATA_FORMAT_10_10_10_2 */
1298 { 4, 4, 0, V_008F0C_BUF_DATA_FORMAT_2_10_10_10
}, /* BUF_DATA_FORMAT_2_10_10_10 */
1299 { 4, 4, 1, V_008F0C_BUF_DATA_FORMAT_8
}, /* BUF_DATA_FORMAT_8_8_8_8 */
1300 { 8, 2, 4, V_008F0C_BUF_DATA_FORMAT_32
}, /* BUF_DATA_FORMAT_32_32 */
1301 { 8, 4, 2, V_008F0C_BUF_DATA_FORMAT_16
}, /* BUF_DATA_FORMAT_16_16_16_16 */
1302 { 12, 3, 4, V_008F0C_BUF_DATA_FORMAT_32
}, /* BUF_DATA_FORMAT_32_32_32 */
1303 { 16, 4, 4, V_008F0C_BUF_DATA_FORMAT_32
}, /* BUF_DATA_FORMAT_32_32_32_32 */
1307 radv_fixup_vertex_input_fetches(struct radv_shader_context
*ctx
,
1309 unsigned num_channels
,
1312 LLVMValueRef zero
= is_float
? ctx
->ac
.f32_0
: ctx
->ac
.i32_0
;
1313 LLVMValueRef one
= is_float
? ctx
->ac
.f32_1
: ctx
->ac
.i32_1
;
1314 LLVMValueRef chan
[4];
1316 if (LLVMGetTypeKind(LLVMTypeOf(value
)) == LLVMVectorTypeKind
) {
1317 unsigned vec_size
= LLVMGetVectorSize(LLVMTypeOf(value
));
1319 if (num_channels
== 4 && num_channels
== vec_size
)
1322 num_channels
= MIN2(num_channels
, vec_size
);
1324 for (unsigned i
= 0; i
< num_channels
; i
++)
1325 chan
[i
] = ac_llvm_extract_elem(&ctx
->ac
, value
, i
);
1327 assert(num_channels
== 1);
1331 for (unsigned i
= num_channels
; i
< 4; i
++) {
1332 chan
[i
] = i
== 3 ? one
: zero
;
1333 chan
[i
] = ac_to_integer(&ctx
->ac
, chan
[i
]);
1336 return ac_build_gather_values(&ctx
->ac
, chan
, 4);
1340 handle_vs_input_decl(struct radv_shader_context
*ctx
,
1341 struct nir_variable
*variable
)
1343 LLVMValueRef t_list_ptr
= ac_get_arg(&ctx
->ac
, ctx
->args
->vertex_buffers
);
1344 LLVMValueRef t_offset
;
1345 LLVMValueRef t_list
;
1347 LLVMValueRef buffer_index
;
1348 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, true);
1349 uint8_t input_usage_mask
=
1350 ctx
->args
->shader_info
->vs
.input_usage_mask
[variable
->data
.location
];
1351 unsigned num_input_channels
= util_last_bit(input_usage_mask
);
1353 variable
->data
.driver_location
= variable
->data
.location
* 4;
1355 enum glsl_base_type type
= glsl_get_base_type(variable
->type
);
1356 for (unsigned i
= 0; i
< attrib_count
; ++i
) {
1357 LLVMValueRef output
[4];
1358 unsigned attrib_index
= variable
->data
.location
+ i
- VERT_ATTRIB_GENERIC0
;
1359 unsigned attrib_format
= ctx
->args
->options
->key
.vs
.vertex_attribute_formats
[attrib_index
];
1360 unsigned data_format
= attrib_format
& 0x0f;
1361 unsigned num_format
= (attrib_format
>> 4) & 0x07;
1362 bool is_float
= num_format
!= V_008F0C_BUF_NUM_FORMAT_UINT
&&
1363 num_format
!= V_008F0C_BUF_NUM_FORMAT_SINT
;
1365 if (ctx
->args
->options
->key
.vs
.instance_rate_inputs
& (1u << attrib_index
)) {
1366 uint32_t divisor
= ctx
->args
->options
->key
.vs
.instance_rate_divisors
[attrib_index
];
1369 buffer_index
= ctx
->abi
.instance_id
;
1372 buffer_index
= LLVMBuildUDiv(ctx
->ac
.builder
, buffer_index
,
1373 LLVMConstInt(ctx
->ac
.i32
, divisor
, 0), "");
1376 buffer_index
= ctx
->ac
.i32_0
;
1379 buffer_index
= LLVMBuildAdd(ctx
->ac
.builder
,
1380 ac_get_arg(&ctx
->ac
,
1381 ctx
->args
->ac
.start_instance
),\
1384 buffer_index
= LLVMBuildAdd(ctx
->ac
.builder
,
1386 ac_get_arg(&ctx
->ac
,
1387 ctx
->args
->ac
.base_vertex
), "");
1390 assert(data_format
< ARRAY_SIZE(vertex_format_table
));
1391 const struct vertex_format_info
*vtx_info
= &vertex_format_table
[data_format
];
1393 /* Adjust the number of channels to load based on the vertex
1396 unsigned num_channels
= MIN2(num_input_channels
, vtx_info
->num_channels
);
1397 unsigned attrib_binding
= ctx
->args
->options
->key
.vs
.vertex_attribute_bindings
[attrib_index
];
1398 unsigned attrib_offset
= ctx
->args
->options
->key
.vs
.vertex_attribute_offsets
[attrib_index
];
1399 unsigned attrib_stride
= ctx
->args
->options
->key
.vs
.vertex_attribute_strides
[attrib_index
];
1401 if (ctx
->args
->options
->key
.vs
.post_shuffle
& (1 << attrib_index
)) {
1402 /* Always load, at least, 3 channels for formats that
1403 * need to be shuffled because X<->Z.
1405 num_channels
= MAX2(num_channels
, 3);
1408 t_offset
= LLVMConstInt(ctx
->ac
.i32
, attrib_binding
, false);
1409 t_list
= ac_build_load_to_sgpr(&ctx
->ac
, t_list_ptr
, t_offset
);
1411 /* Perform per-channel vertex fetch operations if unaligned
1412 * access are detected. Only GFX6 and GFX10 are affected.
1414 bool unaligned_vertex_fetches
= false;
1415 if ((ctx
->ac
.chip_class
== GFX6
|| ctx
->ac
.chip_class
== GFX10
) &&
1416 vtx_info
->chan_format
!= data_format
&&
1417 ((attrib_offset
% vtx_info
->vertex_byte_size
) ||
1418 (attrib_stride
% vtx_info
->vertex_byte_size
)))
1419 unaligned_vertex_fetches
= true;
1421 if (unaligned_vertex_fetches
) {
1422 unsigned chan_format
= vtx_info
->chan_format
;
1423 LLVMValueRef values
[4];
1425 assert(ctx
->ac
.chip_class
== GFX6
||
1426 ctx
->ac
.chip_class
== GFX10
);
1428 for (unsigned chan
= 0; chan
< num_channels
; chan
++) {
1429 unsigned chan_offset
= attrib_offset
+ chan
* vtx_info
->chan_byte_size
;
1430 LLVMValueRef chan_index
= buffer_index
;
1432 if (attrib_stride
!= 0 && chan_offset
> attrib_stride
) {
1433 LLVMValueRef buffer_offset
=
1434 LLVMConstInt(ctx
->ac
.i32
,
1435 chan_offset
/ attrib_stride
, false);
1437 chan_index
= LLVMBuildAdd(ctx
->ac
.builder
,
1441 chan_offset
= chan_offset
% attrib_stride
;
1444 values
[chan
] = ac_build_struct_tbuffer_load(&ctx
->ac
, t_list
,
1446 LLVMConstInt(ctx
->ac
.i32
, chan_offset
, false),
1447 ctx
->ac
.i32_0
, ctx
->ac
.i32_0
, 1,
1448 chan_format
, num_format
, 0, true);
1451 input
= ac_build_gather_values(&ctx
->ac
, values
, num_channels
);
1453 if (attrib_stride
!= 0 && attrib_offset
> attrib_stride
) {
1454 LLVMValueRef buffer_offset
=
1455 LLVMConstInt(ctx
->ac
.i32
,
1456 attrib_offset
/ attrib_stride
, false);
1458 buffer_index
= LLVMBuildAdd(ctx
->ac
.builder
,
1462 attrib_offset
= attrib_offset
% attrib_stride
;
1465 input
= ac_build_struct_tbuffer_load(&ctx
->ac
, t_list
,
1467 LLVMConstInt(ctx
->ac
.i32
, attrib_offset
, false),
1468 ctx
->ac
.i32_0
, ctx
->ac
.i32_0
,
1470 data_format
, num_format
, 0, true);
1473 if (ctx
->args
->options
->key
.vs
.post_shuffle
& (1 << attrib_index
)) {
1475 c
[0] = ac_llvm_extract_elem(&ctx
->ac
, input
, 2);
1476 c
[1] = ac_llvm_extract_elem(&ctx
->ac
, input
, 1);
1477 c
[2] = ac_llvm_extract_elem(&ctx
->ac
, input
, 0);
1478 c
[3] = ac_llvm_extract_elem(&ctx
->ac
, input
, 3);
1480 input
= ac_build_gather_values(&ctx
->ac
, c
, 4);
1483 input
= radv_fixup_vertex_input_fetches(ctx
, input
, num_channels
,
1486 for (unsigned chan
= 0; chan
< 4; chan
++) {
1487 LLVMValueRef llvm_chan
= LLVMConstInt(ctx
->ac
.i32
, chan
, false);
1488 output
[chan
] = LLVMBuildExtractElement(ctx
->ac
.builder
, input
, llvm_chan
, "");
1489 if (type
== GLSL_TYPE_FLOAT16
) {
1490 output
[chan
] = LLVMBuildBitCast(ctx
->ac
.builder
, output
[chan
], ctx
->ac
.f32
, "");
1491 output
[chan
] = LLVMBuildFPTrunc(ctx
->ac
.builder
, output
[chan
], ctx
->ac
.f16
, "");
1495 unsigned alpha_adjust
= (ctx
->args
->options
->key
.vs
.alpha_adjust
>> (attrib_index
* 2)) & 3;
1496 output
[3] = adjust_vertex_fetch_alpha(ctx
, alpha_adjust
, output
[3]);
1498 for (unsigned chan
= 0; chan
< 4; chan
++) {
1499 output
[chan
] = ac_to_integer(&ctx
->ac
, output
[chan
]);
1500 if (type
== GLSL_TYPE_UINT16
|| type
== GLSL_TYPE_INT16
)
1501 output
[chan
] = LLVMBuildTrunc(ctx
->ac
.builder
, output
[chan
], ctx
->ac
.i16
, "");
1503 ctx
->inputs
[ac_llvm_reg_index_soa(variable
->data
.location
+ i
, chan
)] = output
[chan
];
1509 handle_vs_inputs(struct radv_shader_context
*ctx
,
1510 struct nir_shader
*nir
) {
1511 nir_foreach_variable(variable
, &nir
->inputs
)
1512 handle_vs_input_decl(ctx
, variable
);
1516 prepare_interp_optimize(struct radv_shader_context
*ctx
,
1517 struct nir_shader
*nir
)
1519 bool uses_center
= false;
1520 bool uses_centroid
= false;
1521 nir_foreach_variable(variable
, &nir
->inputs
) {
1522 if (glsl_get_base_type(glsl_without_array(variable
->type
)) != GLSL_TYPE_FLOAT
||
1523 variable
->data
.sample
)
1526 if (variable
->data
.centroid
)
1527 uses_centroid
= true;
1532 ctx
->abi
.persp_centroid
= ac_get_arg(&ctx
->ac
, ctx
->args
->ac
.persp_centroid
);
1533 ctx
->abi
.linear_centroid
= ac_get_arg(&ctx
->ac
, ctx
->args
->ac
.linear_centroid
);
1535 if (uses_center
&& uses_centroid
) {
1536 LLVMValueRef sel
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntSLT
,
1537 ac_get_arg(&ctx
->ac
, ctx
->args
->ac
.prim_mask
),
1539 ctx
->abi
.persp_centroid
=
1540 LLVMBuildSelect(ctx
->ac
.builder
, sel
,
1541 ac_get_arg(&ctx
->ac
, ctx
->args
->ac
.persp_center
),
1542 ctx
->abi
.persp_centroid
, "");
1543 ctx
->abi
.linear_centroid
=
1544 LLVMBuildSelect(ctx
->ac
.builder
, sel
,
1545 ac_get_arg(&ctx
->ac
, ctx
->args
->ac
.linear_center
),
1546 ctx
->abi
.linear_centroid
, "");
1551 scan_shader_output_decl(struct radv_shader_context
*ctx
,
1552 struct nir_variable
*variable
,
1553 struct nir_shader
*shader
,
1554 gl_shader_stage stage
)
1556 int idx
= variable
->data
.location
+ variable
->data
.index
;
1557 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
, false);
1558 uint64_t mask_attribs
;
1560 variable
->data
.driver_location
= idx
* 4;
1562 /* tess ctrl has it's own load/store paths for outputs */
1563 if (stage
== MESA_SHADER_TESS_CTRL
)
1566 if (variable
->data
.compact
) {
1567 unsigned component_count
= variable
->data
.location_frac
+
1568 glsl_get_length(variable
->type
);
1569 attrib_count
= (component_count
+ 3) / 4;
1572 mask_attribs
= ((1ull << attrib_count
) - 1) << idx
;
1574 ctx
->output_mask
|= mask_attribs
;
1578 /* Initialize arguments for the shader export intrinsic */
1580 si_llvm_init_export_args(struct radv_shader_context
*ctx
,
1581 LLVMValueRef
*values
,
1582 unsigned enabled_channels
,
1584 struct ac_export_args
*args
)
1586 /* Specify the channels that are enabled. */
1587 args
->enabled_channels
= enabled_channels
;
1589 /* Specify whether the EXEC mask represents the valid mask */
1590 args
->valid_mask
= 0;
1592 /* Specify whether this is the last export */
1595 /* Specify the target we are exporting */
1596 args
->target
= target
;
1598 args
->compr
= false;
1599 args
->out
[0] = LLVMGetUndef(ctx
->ac
.f32
);
1600 args
->out
[1] = LLVMGetUndef(ctx
->ac
.f32
);
1601 args
->out
[2] = LLVMGetUndef(ctx
->ac
.f32
);
1602 args
->out
[3] = LLVMGetUndef(ctx
->ac
.f32
);
1607 bool is_16bit
= ac_get_type_size(LLVMTypeOf(values
[0])) == 2;
1608 if (ctx
->stage
== MESA_SHADER_FRAGMENT
) {
1609 unsigned index
= target
- V_008DFC_SQ_EXP_MRT
;
1610 unsigned col_format
= (ctx
->args
->options
->key
.fs
.col_format
>> (4 * index
)) & 0xf;
1611 bool is_int8
= (ctx
->args
->options
->key
.fs
.is_int8
>> index
) & 1;
1612 bool is_int10
= (ctx
->args
->options
->key
.fs
.is_int10
>> index
) & 1;
1615 LLVMValueRef (*packf
)(struct ac_llvm_context
*ctx
, LLVMValueRef args
[2]) = NULL
;
1616 LLVMValueRef (*packi
)(struct ac_llvm_context
*ctx
, LLVMValueRef args
[2],
1617 unsigned bits
, bool hi
) = NULL
;
1619 switch(col_format
) {
1620 case V_028714_SPI_SHADER_ZERO
:
1621 args
->enabled_channels
= 0; /* writemask */
1622 args
->target
= V_008DFC_SQ_EXP_NULL
;
1625 case V_028714_SPI_SHADER_32_R
:
1626 args
->enabled_channels
= 1;
1627 args
->out
[0] = values
[0];
1630 case V_028714_SPI_SHADER_32_GR
:
1631 args
->enabled_channels
= 0x3;
1632 args
->out
[0] = values
[0];
1633 args
->out
[1] = values
[1];
1636 case V_028714_SPI_SHADER_32_AR
:
1637 if (ctx
->ac
.chip_class
>= GFX10
) {
1638 args
->enabled_channels
= 0x3;
1639 args
->out
[0] = values
[0];
1640 args
->out
[1] = values
[3];
1642 args
->enabled_channels
= 0x9;
1643 args
->out
[0] = values
[0];
1644 args
->out
[3] = values
[3];
1648 case V_028714_SPI_SHADER_FP16_ABGR
:
1649 args
->enabled_channels
= 0x5;
1650 packf
= ac_build_cvt_pkrtz_f16
;
1652 for (unsigned chan
= 0; chan
< 4; chan
++)
1653 values
[chan
] = LLVMBuildFPExt(ctx
->ac
.builder
,
1659 case V_028714_SPI_SHADER_UNORM16_ABGR
:
1660 args
->enabled_channels
= 0x5;
1661 packf
= ac_build_cvt_pknorm_u16
;
1664 case V_028714_SPI_SHADER_SNORM16_ABGR
:
1665 args
->enabled_channels
= 0x5;
1666 packf
= ac_build_cvt_pknorm_i16
;
1669 case V_028714_SPI_SHADER_UINT16_ABGR
:
1670 args
->enabled_channels
= 0x5;
1671 packi
= ac_build_cvt_pk_u16
;
1673 for (unsigned chan
= 0; chan
< 4; chan
++)
1674 values
[chan
] = LLVMBuildZExt(ctx
->ac
.builder
,
1675 ac_to_integer(&ctx
->ac
, values
[chan
]),
1680 case V_028714_SPI_SHADER_SINT16_ABGR
:
1681 args
->enabled_channels
= 0x5;
1682 packi
= ac_build_cvt_pk_i16
;
1684 for (unsigned chan
= 0; chan
< 4; chan
++)
1685 values
[chan
] = LLVMBuildSExt(ctx
->ac
.builder
,
1686 ac_to_integer(&ctx
->ac
, values
[chan
]),
1692 case V_028714_SPI_SHADER_32_ABGR
:
1693 memcpy(&args
->out
[0], values
, sizeof(values
[0]) * 4);
1697 /* Pack f16 or norm_i16/u16. */
1699 for (chan
= 0; chan
< 2; chan
++) {
1700 LLVMValueRef pack_args
[2] = {
1702 values
[2 * chan
+ 1]
1704 LLVMValueRef packed
;
1706 packed
= packf(&ctx
->ac
, pack_args
);
1707 args
->out
[chan
] = ac_to_float(&ctx
->ac
, packed
);
1709 args
->compr
= 1; /* COMPR flag */
1714 for (chan
= 0; chan
< 2; chan
++) {
1715 LLVMValueRef pack_args
[2] = {
1716 ac_to_integer(&ctx
->ac
, values
[2 * chan
]),
1717 ac_to_integer(&ctx
->ac
, values
[2 * chan
+ 1])
1719 LLVMValueRef packed
;
1721 packed
= packi(&ctx
->ac
, pack_args
,
1722 is_int8
? 8 : is_int10
? 10 : 16,
1724 args
->out
[chan
] = ac_to_float(&ctx
->ac
, packed
);
1726 args
->compr
= 1; /* COMPR flag */
1732 for (unsigned chan
= 0; chan
< 4; chan
++) {
1733 values
[chan
] = LLVMBuildBitCast(ctx
->ac
.builder
, values
[chan
], ctx
->ac
.i16
, "");
1734 args
->out
[chan
] = LLVMBuildZExt(ctx
->ac
.builder
, values
[chan
], ctx
->ac
.i32
, "");
1737 memcpy(&args
->out
[0], values
, sizeof(values
[0]) * 4);
1739 for (unsigned i
= 0; i
< 4; ++i
)
1740 args
->out
[i
] = ac_to_float(&ctx
->ac
, args
->out
[i
]);
1744 radv_export_param(struct radv_shader_context
*ctx
, unsigned index
,
1745 LLVMValueRef
*values
, unsigned enabled_channels
)
1747 struct ac_export_args args
;
1749 si_llvm_init_export_args(ctx
, values
, enabled_channels
,
1750 V_008DFC_SQ_EXP_PARAM
+ index
, &args
);
1751 ac_build_export(&ctx
->ac
, &args
);
1755 radv_load_output(struct radv_shader_context
*ctx
, unsigned index
, unsigned chan
)
1757 LLVMValueRef output
= ctx
->abi
.outputs
[ac_llvm_reg_index_soa(index
, chan
)];
1758 return LLVMBuildLoad(ctx
->ac
.builder
, output
, "");
1762 radv_emit_stream_output(struct radv_shader_context
*ctx
,
1763 LLVMValueRef
const *so_buffers
,
1764 LLVMValueRef
const *so_write_offsets
,
1765 const struct radv_stream_output
*output
,
1766 struct radv_shader_output_values
*shader_out
)
1768 unsigned num_comps
= util_bitcount(output
->component_mask
);
1769 unsigned buf
= output
->buffer
;
1770 unsigned offset
= output
->offset
;
1772 LLVMValueRef out
[4];
1774 assert(num_comps
&& num_comps
<= 4);
1775 if (!num_comps
|| num_comps
> 4)
1778 /* Get the first component. */
1779 start
= ffs(output
->component_mask
) - 1;
1781 /* Load the output as int. */
1782 for (int i
= 0; i
< num_comps
; i
++) {
1783 out
[i
] = ac_to_integer(&ctx
->ac
, shader_out
->values
[start
+ i
]);
1786 /* Pack the output. */
1787 LLVMValueRef vdata
= NULL
;
1789 switch (num_comps
) {
1790 case 1: /* as i32 */
1793 case 2: /* as v2i32 */
1794 case 3: /* as v4i32 (aligned to 4) */
1795 out
[3] = LLVMGetUndef(ctx
->ac
.i32
);
1797 case 4: /* as v4i32 */
1798 vdata
= ac_build_gather_values(&ctx
->ac
, out
,
1799 !ac_has_vec3_support(ctx
->ac
.chip_class
, false) ?
1800 util_next_power_of_two(num_comps
) :
1805 ac_build_buffer_store_dword(&ctx
->ac
, so_buffers
[buf
],
1806 vdata
, num_comps
, so_write_offsets
[buf
],
1807 ctx
->ac
.i32_0
, offset
,
1812 radv_emit_streamout(struct radv_shader_context
*ctx
, unsigned stream
)
1816 /* Get bits [22:16], i.e. (so_param >> 16) & 127; */
1817 assert(ctx
->args
->streamout_config
.used
);
1818 LLVMValueRef so_vtx_count
=
1819 ac_build_bfe(&ctx
->ac
,
1820 ac_get_arg(&ctx
->ac
, ctx
->args
->streamout_config
),
1821 LLVMConstInt(ctx
->ac
.i32
, 16, false),
1822 LLVMConstInt(ctx
->ac
.i32
, 7, false), false);
1824 LLVMValueRef tid
= ac_get_thread_id(&ctx
->ac
);
1826 /* can_emit = tid < so_vtx_count; */
1827 LLVMValueRef can_emit
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntULT
,
1828 tid
, so_vtx_count
, "");
1830 /* Emit the streamout code conditionally. This actually avoids
1831 * out-of-bounds buffer access. The hw tells us via the SGPR
1832 * (so_vtx_count) which threads are allowed to emit streamout data.
1834 ac_build_ifcc(&ctx
->ac
, can_emit
, 6501);
1836 /* The buffer offset is computed as follows:
1837 * ByteOffset = streamout_offset[buffer_id]*4 +
1838 * (streamout_write_index + thread_id)*stride[buffer_id] +
1841 LLVMValueRef so_write_index
=
1842 ac_get_arg(&ctx
->ac
, ctx
->args
->streamout_write_idx
);
1844 /* Compute (streamout_write_index + thread_id). */
1846 LLVMBuildAdd(ctx
->ac
.builder
, so_write_index
, tid
, "");
1848 /* Load the descriptor and compute the write offset for each
1851 LLVMValueRef so_write_offset
[4] = {};
1852 LLVMValueRef so_buffers
[4] = {};
1853 LLVMValueRef buf_ptr
= ac_get_arg(&ctx
->ac
, ctx
->args
->streamout_buffers
);
1855 for (i
= 0; i
< 4; i
++) {
1856 uint16_t stride
= ctx
->args
->shader_info
->so
.strides
[i
];
1861 LLVMValueRef offset
=
1862 LLVMConstInt(ctx
->ac
.i32
, i
, false);
1864 so_buffers
[i
] = ac_build_load_to_sgpr(&ctx
->ac
,
1867 LLVMValueRef so_offset
=
1868 ac_get_arg(&ctx
->ac
, ctx
->args
->streamout_offset
[i
]);
1870 so_offset
= LLVMBuildMul(ctx
->ac
.builder
, so_offset
,
1871 LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
1873 so_write_offset
[i
] =
1874 ac_build_imad(&ctx
->ac
, so_write_index
,
1875 LLVMConstInt(ctx
->ac
.i32
,
1880 /* Write streamout data. */
1881 for (i
= 0; i
< ctx
->args
->shader_info
->so
.num_outputs
; i
++) {
1882 struct radv_shader_output_values shader_out
= {};
1883 struct radv_stream_output
*output
=
1884 &ctx
->args
->shader_info
->so
.outputs
[i
];
1886 if (stream
!= output
->stream
)
1889 for (int j
= 0; j
< 4; j
++) {
1890 shader_out
.values
[j
] =
1891 radv_load_output(ctx
, output
->location
, j
);
1894 radv_emit_stream_output(ctx
, so_buffers
,so_write_offset
,
1895 output
, &shader_out
);
1898 ac_build_endif(&ctx
->ac
, 6501);
1902 radv_build_param_exports(struct radv_shader_context
*ctx
,
1903 struct radv_shader_output_values
*outputs
,
1905 struct radv_vs_output_info
*outinfo
,
1906 bool export_clip_dists
)
1908 unsigned param_count
= 0;
1910 for (unsigned i
= 0; i
< noutput
; i
++) {
1911 unsigned slot_name
= outputs
[i
].slot_name
;
1912 unsigned usage_mask
= outputs
[i
].usage_mask
;
1914 if (slot_name
!= VARYING_SLOT_LAYER
&&
1915 slot_name
!= VARYING_SLOT_PRIMITIVE_ID
&&
1916 slot_name
!= VARYING_SLOT_CLIP_DIST0
&&
1917 slot_name
!= VARYING_SLOT_CLIP_DIST1
&&
1918 slot_name
< VARYING_SLOT_VAR0
)
1921 if ((slot_name
== VARYING_SLOT_CLIP_DIST0
||
1922 slot_name
== VARYING_SLOT_CLIP_DIST1
) && !export_clip_dists
)
1925 radv_export_param(ctx
, param_count
, outputs
[i
].values
, usage_mask
);
1927 assert(i
< ARRAY_SIZE(outinfo
->vs_output_param_offset
));
1928 outinfo
->vs_output_param_offset
[slot_name
] = param_count
++;
1931 outinfo
->param_exports
= param_count
;
1934 /* Generate export instructions for hardware VS shader stage or NGG GS stage
1935 * (position and parameter data only).
1938 radv_llvm_export_vs(struct radv_shader_context
*ctx
,
1939 struct radv_shader_output_values
*outputs
,
1941 struct radv_vs_output_info
*outinfo
,
1942 bool export_clip_dists
)
1944 LLVMValueRef psize_value
= NULL
, layer_value
= NULL
, viewport_value
= NULL
;
1945 struct ac_export_args pos_args
[4] = {};
1946 unsigned pos_idx
, index
;
1949 /* Build position exports */
1950 for (i
= 0; i
< noutput
; i
++) {
1951 switch (outputs
[i
].slot_name
) {
1952 case VARYING_SLOT_POS
:
1953 si_llvm_init_export_args(ctx
, outputs
[i
].values
, 0xf,
1954 V_008DFC_SQ_EXP_POS
, &pos_args
[0]);
1956 case VARYING_SLOT_PSIZ
:
1957 psize_value
= outputs
[i
].values
[0];
1959 case VARYING_SLOT_LAYER
:
1960 layer_value
= outputs
[i
].values
[0];
1962 case VARYING_SLOT_VIEWPORT
:
1963 viewport_value
= outputs
[i
].values
[0];
1965 case VARYING_SLOT_CLIP_DIST0
:
1966 case VARYING_SLOT_CLIP_DIST1
:
1967 index
= 2 + outputs
[i
].slot_index
;
1968 si_llvm_init_export_args(ctx
, outputs
[i
].values
, 0xf,
1969 V_008DFC_SQ_EXP_POS
+ index
,
1977 /* We need to add the position output manually if it's missing. */
1978 if (!pos_args
[0].out
[0]) {
1979 pos_args
[0].enabled_channels
= 0xf; /* writemask */
1980 pos_args
[0].valid_mask
= 0; /* EXEC mask */
1981 pos_args
[0].done
= 0; /* last export? */
1982 pos_args
[0].target
= V_008DFC_SQ_EXP_POS
;
1983 pos_args
[0].compr
= 0; /* COMPR flag */
1984 pos_args
[0].out
[0] = ctx
->ac
.f32_0
; /* X */
1985 pos_args
[0].out
[1] = ctx
->ac
.f32_0
; /* Y */
1986 pos_args
[0].out
[2] = ctx
->ac
.f32_0
; /* Z */
1987 pos_args
[0].out
[3] = ctx
->ac
.f32_1
; /* W */
1990 if (outinfo
->writes_pointsize
||
1991 outinfo
->writes_layer
||
1992 outinfo
->writes_viewport_index
) {
1993 pos_args
[1].enabled_channels
= ((outinfo
->writes_pointsize
== true ? 1 : 0) |
1994 (outinfo
->writes_layer
== true ? 4 : 0));
1995 pos_args
[1].valid_mask
= 0;
1996 pos_args
[1].done
= 0;
1997 pos_args
[1].target
= V_008DFC_SQ_EXP_POS
+ 1;
1998 pos_args
[1].compr
= 0;
1999 pos_args
[1].out
[0] = ctx
->ac
.f32_0
; /* X */
2000 pos_args
[1].out
[1] = ctx
->ac
.f32_0
; /* Y */
2001 pos_args
[1].out
[2] = ctx
->ac
.f32_0
; /* Z */
2002 pos_args
[1].out
[3] = ctx
->ac
.f32_0
; /* W */
2004 if (outinfo
->writes_pointsize
== true)
2005 pos_args
[1].out
[0] = psize_value
;
2006 if (outinfo
->writes_layer
== true)
2007 pos_args
[1].out
[2] = layer_value
;
2008 if (outinfo
->writes_viewport_index
== true) {
2009 if (ctx
->args
->options
->chip_class
>= GFX9
) {
2010 /* GFX9 has the layer in out.z[10:0] and the viewport
2011 * index in out.z[19:16].
2013 LLVMValueRef v
= viewport_value
;
2014 v
= ac_to_integer(&ctx
->ac
, v
);
2015 v
= LLVMBuildShl(ctx
->ac
.builder
, v
,
2016 LLVMConstInt(ctx
->ac
.i32
, 16, false),
2018 v
= LLVMBuildOr(ctx
->ac
.builder
, v
,
2019 ac_to_integer(&ctx
->ac
, pos_args
[1].out
[2]), "");
2021 pos_args
[1].out
[2] = ac_to_float(&ctx
->ac
, v
);
2022 pos_args
[1].enabled_channels
|= 1 << 2;
2024 pos_args
[1].out
[3] = viewport_value
;
2025 pos_args
[1].enabled_channels
|= 1 << 3;
2030 for (i
= 0; i
< 4; i
++) {
2031 if (pos_args
[i
].out
[0])
2032 outinfo
->pos_exports
++;
2035 /* Navi10-14 skip POS0 exports if EXEC=0 and DONE=0, causing a hang.
2036 * Setting valid_mask=1 prevents it and has no other effect.
2038 if (ctx
->ac
.family
== CHIP_NAVI10
||
2039 ctx
->ac
.family
== CHIP_NAVI12
||
2040 ctx
->ac
.family
== CHIP_NAVI14
)
2041 pos_args
[0].valid_mask
= 1;
2044 for (i
= 0; i
< 4; i
++) {
2045 if (!pos_args
[i
].out
[0])
2048 /* Specify the target we are exporting */
2049 pos_args
[i
].target
= V_008DFC_SQ_EXP_POS
+ pos_idx
++;
2051 if (pos_idx
== outinfo
->pos_exports
)
2052 /* Specify that this is the last export */
2053 pos_args
[i
].done
= 1;
2055 ac_build_export(&ctx
->ac
, &pos_args
[i
]);
2058 /* Build parameter exports */
2059 radv_build_param_exports(ctx
, outputs
, noutput
, outinfo
, export_clip_dists
);
2063 handle_vs_outputs_post(struct radv_shader_context
*ctx
,
2064 bool export_prim_id
,
2065 bool export_clip_dists
,
2066 struct radv_vs_output_info
*outinfo
)
2068 struct radv_shader_output_values
*outputs
;
2069 unsigned noutput
= 0;
2071 if (ctx
->args
->options
->key
.has_multiview_view_index
) {
2072 LLVMValueRef
* tmp_out
= &ctx
->abi
.outputs
[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER
, 0)];
2074 for(unsigned i
= 0; i
< 4; ++i
)
2075 ctx
->abi
.outputs
[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER
, i
)] =
2076 ac_build_alloca_undef(&ctx
->ac
, ctx
->ac
.f32
, "");
2079 LLVMValueRef view_index
= ac_get_arg(&ctx
->ac
, ctx
->args
->ac
.view_index
);
2080 LLVMBuildStore(ctx
->ac
.builder
, ac_to_float(&ctx
->ac
, view_index
), *tmp_out
);
2081 ctx
->output_mask
|= 1ull << VARYING_SLOT_LAYER
;
2084 memset(outinfo
->vs_output_param_offset
, AC_EXP_PARAM_UNDEFINED
,
2085 sizeof(outinfo
->vs_output_param_offset
));
2086 outinfo
->pos_exports
= 0;
2088 if (!ctx
->args
->options
->use_ngg_streamout
&&
2089 ctx
->args
->shader_info
->so
.num_outputs
&&
2090 !ctx
->args
->is_gs_copy_shader
) {
2091 /* The GS copy shader emission already emits streamout. */
2092 radv_emit_streamout(ctx
, 0);
2095 /* Allocate a temporary array for the output values. */
2096 unsigned num_outputs
= util_bitcount64(ctx
->output_mask
) + export_prim_id
;
2097 outputs
= malloc(num_outputs
* sizeof(outputs
[0]));
2099 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
2100 if (!(ctx
->output_mask
& (1ull << i
)))
2103 outputs
[noutput
].slot_name
= i
;
2104 outputs
[noutput
].slot_index
= i
== VARYING_SLOT_CLIP_DIST1
;
2106 if (ctx
->stage
== MESA_SHADER_VERTEX
&&
2107 !ctx
->args
->is_gs_copy_shader
) {
2108 outputs
[noutput
].usage_mask
=
2109 ctx
->args
->shader_info
->vs
.output_usage_mask
[i
];
2110 } else if (ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
2111 outputs
[noutput
].usage_mask
=
2112 ctx
->args
->shader_info
->tes
.output_usage_mask
[i
];
2114 assert(ctx
->args
->is_gs_copy_shader
);
2115 outputs
[noutput
].usage_mask
=
2116 ctx
->args
->shader_info
->gs
.output_usage_mask
[i
];
2119 for (unsigned j
= 0; j
< 4; j
++) {
2120 outputs
[noutput
].values
[j
] =
2121 ac_to_float(&ctx
->ac
, radv_load_output(ctx
, i
, j
));
2127 /* Export PrimitiveID. */
2128 if (export_prim_id
) {
2129 outputs
[noutput
].slot_name
= VARYING_SLOT_PRIMITIVE_ID
;
2130 outputs
[noutput
].slot_index
= 0;
2131 outputs
[noutput
].usage_mask
= 0x1;
2132 outputs
[noutput
].values
[0] =
2133 ac_get_arg(&ctx
->ac
, ctx
->args
->vs_prim_id
);
2134 for (unsigned j
= 1; j
< 4; j
++)
2135 outputs
[noutput
].values
[j
] = ctx
->ac
.f32_0
;
2139 radv_llvm_export_vs(ctx
, outputs
, noutput
, outinfo
, export_clip_dists
);
2145 handle_es_outputs_post(struct radv_shader_context
*ctx
,
2146 struct radv_es_output_info
*outinfo
)
2149 LLVMValueRef lds_base
= NULL
;
2151 if (ctx
->ac
.chip_class
>= GFX9
) {
2152 unsigned itemsize_dw
= outinfo
->esgs_itemsize
/ 4;
2153 LLVMValueRef vertex_idx
= ac_get_thread_id(&ctx
->ac
);
2154 LLVMValueRef wave_idx
=
2155 ac_unpack_param(&ctx
->ac
,
2156 ac_get_arg(&ctx
->ac
, ctx
->args
->merged_wave_info
), 24, 4);
2157 vertex_idx
= LLVMBuildOr(ctx
->ac
.builder
, vertex_idx
,
2158 LLVMBuildMul(ctx
->ac
.builder
, wave_idx
,
2159 LLVMConstInt(ctx
->ac
.i32
,
2160 ctx
->ac
.wave_size
, false), ""), "");
2161 lds_base
= LLVMBuildMul(ctx
->ac
.builder
, vertex_idx
,
2162 LLVMConstInt(ctx
->ac
.i32
, itemsize_dw
, 0), "");
2165 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
2166 LLVMValueRef dw_addr
= NULL
;
2167 LLVMValueRef
*out_ptr
= &ctx
->abi
.outputs
[i
* 4];
2168 unsigned output_usage_mask
;
2171 if (!(ctx
->output_mask
& (1ull << i
)))
2174 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
2176 ctx
->args
->shader_info
->vs
.output_usage_mask
[i
];
2178 assert(ctx
->stage
== MESA_SHADER_TESS_EVAL
);
2180 ctx
->args
->shader_info
->tes
.output_usage_mask
[i
];
2183 param_index
= shader_io_get_unique_index(i
);
2186 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
2187 LLVMConstInt(ctx
->ac
.i32
, param_index
* 4, false),
2191 for (j
= 0; j
< 4; j
++) {
2192 if (!(output_usage_mask
& (1 << j
)))
2195 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->ac
.builder
, out_ptr
[j
], "");
2196 out_val
= ac_to_integer(&ctx
->ac
, out_val
);
2197 out_val
= LLVMBuildZExtOrBitCast(ctx
->ac
.builder
, out_val
, ctx
->ac
.i32
, "");
2199 if (ctx
->ac
.chip_class
>= GFX9
) {
2200 LLVMValueRef dw_addr_offset
=
2201 LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
,
2202 LLVMConstInt(ctx
->ac
.i32
,
2205 ac_lds_store(&ctx
->ac
, dw_addr_offset
, out_val
);
2207 ac_build_buffer_store_dword(&ctx
->ac
,
2211 ac_get_arg(&ctx
->ac
, ctx
->args
->es2gs_offset
),
2212 (4 * param_index
+ j
) * 4,
2213 ac_glc
| ac_slc
| ac_swizzled
);
2220 handle_ls_outputs_post(struct radv_shader_context
*ctx
)
2222 LLVMValueRef vertex_id
= ctx
->rel_auto_id
;
2223 uint32_t num_tcs_inputs
= util_last_bit64(ctx
->args
->shader_info
->vs
.ls_outputs_written
);
2224 LLVMValueRef vertex_dw_stride
= LLVMConstInt(ctx
->ac
.i32
, num_tcs_inputs
* 4, false);
2225 LLVMValueRef base_dw_addr
= LLVMBuildMul(ctx
->ac
.builder
, vertex_id
,
2226 vertex_dw_stride
, "");
2228 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
2229 LLVMValueRef
*out_ptr
= &ctx
->abi
.outputs
[i
* 4];
2231 if (!(ctx
->output_mask
& (1ull << i
)))
2234 int param
= shader_io_get_unique_index(i
);
2235 LLVMValueRef dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, base_dw_addr
,
2236 LLVMConstInt(ctx
->ac
.i32
, param
* 4, false),
2238 for (unsigned j
= 0; j
< 4; j
++) {
2239 LLVMValueRef value
= LLVMBuildLoad(ctx
->ac
.builder
, out_ptr
[j
], "");
2240 value
= ac_to_integer(&ctx
->ac
, value
);
2241 value
= LLVMBuildZExtOrBitCast(ctx
->ac
.builder
, value
, ctx
->ac
.i32
, "");
2242 ac_lds_store(&ctx
->ac
, dw_addr
, value
);
2243 dw_addr
= LLVMBuildAdd(ctx
->ac
.builder
, dw_addr
, ctx
->ac
.i32_1
, "");
2248 static LLVMValueRef
get_wave_id_in_tg(struct radv_shader_context
*ctx
)
2250 return ac_unpack_param(&ctx
->ac
,
2251 ac_get_arg(&ctx
->ac
, ctx
->args
->merged_wave_info
), 24, 4);
2254 static LLVMValueRef
get_tgsize(struct radv_shader_context
*ctx
)
2256 return ac_unpack_param(&ctx
->ac
, ac_get_arg(&ctx
->ac
, ctx
->args
->merged_wave_info
), 28, 4);
2259 static LLVMValueRef
get_thread_id_in_tg(struct radv_shader_context
*ctx
)
2261 LLVMBuilderRef builder
= ctx
->ac
.builder
;
2263 tmp
= LLVMBuildMul(builder
, get_wave_id_in_tg(ctx
),
2264 LLVMConstInt(ctx
->ac
.i32
, ctx
->ac
.wave_size
, false), "");
2265 return LLVMBuildAdd(builder
, tmp
, ac_get_thread_id(&ctx
->ac
), "");
2268 static LLVMValueRef
ngg_get_vtx_cnt(struct radv_shader_context
*ctx
)
2270 return ac_build_bfe(&ctx
->ac
, ac_get_arg(&ctx
->ac
, ctx
->args
->gs_tg_info
),
2271 LLVMConstInt(ctx
->ac
.i32
, 12, false),
2272 LLVMConstInt(ctx
->ac
.i32
, 9, false),
2276 static LLVMValueRef
ngg_get_prim_cnt(struct radv_shader_context
*ctx
)
2278 return ac_build_bfe(&ctx
->ac
, ac_get_arg(&ctx
->ac
, ctx
->args
->gs_tg_info
),
2279 LLVMConstInt(ctx
->ac
.i32
, 22, false),
2280 LLVMConstInt(ctx
->ac
.i32
, 9, false),
2284 static LLVMValueRef
ngg_get_ordered_id(struct radv_shader_context
*ctx
)
2286 return ac_build_bfe(&ctx
->ac
, ac_get_arg(&ctx
->ac
, ctx
->args
->gs_tg_info
),
2288 LLVMConstInt(ctx
->ac
.i32
, 12, false),
2293 ngg_gs_get_vertex_storage(struct radv_shader_context
*ctx
)
2295 unsigned num_outputs
= util_bitcount64(ctx
->output_mask
);
2297 if (ctx
->args
->options
->key
.has_multiview_view_index
)
2300 LLVMTypeRef elements
[2] = {
2301 LLVMArrayType(ctx
->ac
.i32
, 4 * num_outputs
),
2302 LLVMArrayType(ctx
->ac
.i8
, 4),
2304 LLVMTypeRef type
= LLVMStructTypeInContext(ctx
->ac
.context
, elements
, 2, false);
2305 type
= LLVMPointerType(LLVMArrayType(type
, 0), AC_ADDR_SPACE_LDS
);
2306 return LLVMBuildBitCast(ctx
->ac
.builder
, ctx
->gs_ngg_emit
, type
, "");
2310 * Return a pointer to the LDS storage reserved for the N'th vertex, where N
2311 * is in emit order; that is:
2312 * - during the epilogue, N is the threadidx (relative to the entire threadgroup)
2313 * - during vertex emit, i.e. while the API GS shader invocation is running,
2314 * N = threadidx * gs_max_out_vertices + emitidx
2316 * Goals of the LDS memory layout:
2317 * 1. Eliminate bank conflicts on write for geometry shaders that have all emits
2318 * in uniform control flow
2319 * 2. Eliminate bank conflicts on read for export if, additionally, there is no
2321 * 3. Agnostic to the number of waves (since we don't know it before compiling)
2322 * 4. Allow coalescing of LDS instructions (ds_write_b128 etc.)
2323 * 5. Avoid wasting memory.
2325 * We use an AoS layout due to point 4 (this also helps point 3). In an AoS
2326 * layout, elimination of bank conflicts requires that each vertex occupy an
2327 * odd number of dwords. We use the additional dword to store the output stream
2328 * index as well as a flag to indicate whether this vertex ends a primitive
2329 * for rasterization.
2331 * Swizzling is required to satisfy points 1 and 2 simultaneously.
2333 * Vertices are stored in export order (gsthread * gs_max_out_vertices + emitidx).
2334 * Indices are swizzled in groups of 32, which ensures point 1 without
2335 * disturbing point 2.
2337 * \return an LDS pointer to type {[N x i32], [4 x i8]}
2340 ngg_gs_vertex_ptr(struct radv_shader_context
*ctx
, LLVMValueRef vertexidx
)
2342 LLVMBuilderRef builder
= ctx
->ac
.builder
;
2343 LLVMValueRef storage
= ngg_gs_get_vertex_storage(ctx
);
2345 /* gs_max_out_vertices = 2^(write_stride_2exp) * some odd number */
2346 unsigned write_stride_2exp
= ffs(ctx
->shader
->info
.gs
.vertices_out
) - 1;
2347 if (write_stride_2exp
) {
2349 LLVMBuildLShr(builder
, vertexidx
,
2350 LLVMConstInt(ctx
->ac
.i32
, 5, false), "");
2351 LLVMValueRef swizzle
=
2352 LLVMBuildAnd(builder
, row
,
2353 LLVMConstInt(ctx
->ac
.i32
, (1u << write_stride_2exp
) - 1,
2355 vertexidx
= LLVMBuildXor(builder
, vertexidx
, swizzle
, "");
2358 return ac_build_gep0(&ctx
->ac
, storage
, vertexidx
);
2362 ngg_gs_emit_vertex_ptr(struct radv_shader_context
*ctx
, LLVMValueRef gsthread
,
2363 LLVMValueRef emitidx
)
2365 LLVMBuilderRef builder
= ctx
->ac
.builder
;
2368 tmp
= LLVMConstInt(ctx
->ac
.i32
, ctx
->shader
->info
.gs
.vertices_out
, false);
2369 tmp
= LLVMBuildMul(builder
, tmp
, gsthread
, "");
2370 const LLVMValueRef vertexidx
= LLVMBuildAdd(builder
, tmp
, emitidx
, "");
2371 return ngg_gs_vertex_ptr(ctx
, vertexidx
);
2375 ngg_gs_get_emit_output_ptr(struct radv_shader_context
*ctx
, LLVMValueRef vertexptr
,
2378 LLVMValueRef gep_idx
[3] = {
2379 ctx
->ac
.i32_0
, /* implied C-style array */
2380 ctx
->ac
.i32_0
, /* first struct entry */
2381 LLVMConstInt(ctx
->ac
.i32
, out_idx
, false),
2383 return LLVMBuildGEP(ctx
->ac
.builder
, vertexptr
, gep_idx
, 3, "");
2387 ngg_gs_get_emit_primflag_ptr(struct radv_shader_context
*ctx
, LLVMValueRef vertexptr
,
2390 LLVMValueRef gep_idx
[3] = {
2391 ctx
->ac
.i32_0
, /* implied C-style array */
2392 ctx
->ac
.i32_1
, /* second struct entry */
2393 LLVMConstInt(ctx
->ac
.i32
, stream
, false),
2395 return LLVMBuildGEP(ctx
->ac
.builder
, vertexptr
, gep_idx
, 3, "");
2398 static struct radv_stream_output
*
2399 radv_get_stream_output_by_loc(struct radv_streamout_info
*so
, unsigned location
)
2401 for (unsigned i
= 0; i
< so
->num_outputs
; ++i
) {
2402 if (so
->outputs
[i
].location
== location
)
2403 return &so
->outputs
[i
];
2409 static void build_streamout_vertex(struct radv_shader_context
*ctx
,
2410 LLVMValueRef
*so_buffer
, LLVMValueRef
*wg_offset_dw
,
2411 unsigned stream
, LLVMValueRef offset_vtx
,
2412 LLVMValueRef vertexptr
)
2414 struct radv_streamout_info
*so
= &ctx
->args
->shader_info
->so
;
2415 LLVMBuilderRef builder
= ctx
->ac
.builder
;
2416 LLVMValueRef offset
[4] = {};
2419 for (unsigned buffer
= 0; buffer
< 4; ++buffer
) {
2420 if (!wg_offset_dw
[buffer
])
2423 tmp
= LLVMBuildMul(builder
, offset_vtx
,
2424 LLVMConstInt(ctx
->ac
.i32
, so
->strides
[buffer
], false), "");
2425 tmp
= LLVMBuildAdd(builder
, wg_offset_dw
[buffer
], tmp
, "");
2426 offset
[buffer
] = LLVMBuildShl(builder
, tmp
, LLVMConstInt(ctx
->ac
.i32
, 2, false), "");
2429 if (ctx
->stage
== MESA_SHADER_GEOMETRY
) {
2430 struct radv_shader_output_values outputs
[AC_LLVM_MAX_OUTPUTS
];
2431 unsigned noutput
= 0;
2432 unsigned out_idx
= 0;
2434 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
2435 unsigned output_usage_mask
=
2436 ctx
->args
->shader_info
->gs
.output_usage_mask
[i
];
2437 uint8_t output_stream
=
2438 output_stream
= ctx
->args
->shader_info
->gs
.output_streams
[i
];
2440 if (!(ctx
->output_mask
& (1ull << i
)) ||
2441 output_stream
!= stream
)
2444 outputs
[noutput
].slot_name
= i
;
2445 outputs
[noutput
].slot_index
= i
== VARYING_SLOT_CLIP_DIST1
;
2446 outputs
[noutput
].usage_mask
= output_usage_mask
;
2448 int length
= util_last_bit(output_usage_mask
);
2450 for (unsigned j
= 0; j
< length
; j
++, out_idx
++) {
2451 if (!(output_usage_mask
& (1 << j
)))
2454 tmp
= ac_build_gep0(&ctx
->ac
, vertexptr
,
2455 LLVMConstInt(ctx
->ac
.i32
, out_idx
, false));
2456 outputs
[noutput
].values
[j
] = LLVMBuildLoad(builder
, tmp
, "");
2459 for (unsigned j
= length
; j
< 4; j
++)
2460 outputs
[noutput
].values
[j
] = LLVMGetUndef(ctx
->ac
.f32
);
2465 for (unsigned i
= 0; i
< noutput
; i
++) {
2466 struct radv_stream_output
*output
=
2467 radv_get_stream_output_by_loc(so
, outputs
[i
].slot_name
);
2470 output
->stream
!= stream
)
2473 struct radv_shader_output_values out
= {};
2475 for (unsigned j
= 0; j
< 4; j
++) {
2476 out
.values
[j
] = outputs
[i
].values
[j
];
2479 radv_emit_stream_output(ctx
, so_buffer
, offset
, output
, &out
);
2482 for (unsigned i
= 0; i
< so
->num_outputs
; ++i
) {
2483 struct radv_stream_output
*output
=
2484 &ctx
->args
->shader_info
->so
.outputs
[i
];
2486 if (stream
!= output
->stream
)
2489 struct radv_shader_output_values out
= {};
2491 for (unsigned comp
= 0; comp
< 4; comp
++) {
2492 if (!(output
->component_mask
& (1 << comp
)))
2495 tmp
= ac_build_gep0(&ctx
->ac
, vertexptr
,
2496 LLVMConstInt(ctx
->ac
.i32
, 4 * i
+ comp
, false));
2497 out
.values
[comp
] = LLVMBuildLoad(builder
, tmp
, "");
2500 radv_emit_stream_output(ctx
, so_buffer
, offset
, output
, &out
);
2505 struct ngg_streamout
{
2506 LLVMValueRef num_vertices
;
2508 /* per-thread data */
2509 LLVMValueRef prim_enable
[4]; /* i1 per stream */
2510 LLVMValueRef vertices
[3]; /* [N x i32] addrspace(LDS)* */
2513 LLVMValueRef emit
[4]; /* per-stream emitted primitives (only valid for used streams) */
2517 * Build streamout logic.
2519 * Implies a barrier.
2521 * Writes number of emitted primitives to gs_ngg_scratch[4:7].
2523 * Clobbers gs_ngg_scratch[8:].
2525 static void build_streamout(struct radv_shader_context
*ctx
,
2526 struct ngg_streamout
*nggso
)
2528 struct radv_streamout_info
*so
= &ctx
->args
->shader_info
->so
;
2529 LLVMBuilderRef builder
= ctx
->ac
.builder
;
2530 LLVMValueRef buf_ptr
= ac_get_arg(&ctx
->ac
, ctx
->args
->streamout_buffers
);
2531 LLVMValueRef tid
= get_thread_id_in_tg(ctx
);
2532 LLVMValueRef cond
, tmp
, tmp2
;
2533 LLVMValueRef i32_2
= LLVMConstInt(ctx
->ac
.i32
, 2, false);
2534 LLVMValueRef i32_4
= LLVMConstInt(ctx
->ac
.i32
, 4, false);
2535 LLVMValueRef i32_8
= LLVMConstInt(ctx
->ac
.i32
, 8, false);
2536 LLVMValueRef so_buffer
[4] = {};
2537 unsigned max_num_vertices
= 1 + (nggso
->vertices
[1] ? 1 : 0) +
2538 (nggso
->vertices
[2] ? 1 : 0);
2539 LLVMValueRef prim_stride_dw
[4] = {};
2540 LLVMValueRef prim_stride_dw_vgpr
= LLVMGetUndef(ctx
->ac
.i32
);
2541 int stream_for_buffer
[4] = { -1, -1, -1, -1 };
2542 unsigned bufmask_for_stream
[4] = {};
2543 bool isgs
= ctx
->stage
== MESA_SHADER_GEOMETRY
;
2544 unsigned scratch_emit_base
= isgs
? 4 : 0;
2545 LLVMValueRef scratch_emit_basev
= isgs
? i32_4
: ctx
->ac
.i32_0
;
2546 unsigned scratch_offset_base
= isgs
? 8 : 4;
2547 LLVMValueRef scratch_offset_basev
= isgs
? i32_8
: i32_4
;
2549 ac_llvm_add_target_dep_function_attr(ctx
->main_function
,
2550 "amdgpu-gds-size", 256);
2552 /* Determine the mapping of streamout buffers to vertex streams. */
2553 for (unsigned i
= 0; i
< so
->num_outputs
; ++i
) {
2554 unsigned buf
= so
->outputs
[i
].buffer
;
2555 unsigned stream
= so
->outputs
[i
].stream
;
2556 assert(stream_for_buffer
[buf
] < 0 || stream_for_buffer
[buf
] == stream
);
2557 stream_for_buffer
[buf
] = stream
;
2558 bufmask_for_stream
[stream
] |= 1 << buf
;
2561 for (unsigned buffer
= 0; buffer
< 4; ++buffer
) {
2562 if (stream_for_buffer
[buffer
] == -1)
2565 assert(so
->strides
[buffer
]);
2567 LLVMValueRef stride_for_buffer
=
2568 LLVMConstInt(ctx
->ac
.i32
, so
->strides
[buffer
], false);
2569 prim_stride_dw
[buffer
] =
2570 LLVMBuildMul(builder
, stride_for_buffer
,
2571 nggso
->num_vertices
, "");
2572 prim_stride_dw_vgpr
= ac_build_writelane(
2573 &ctx
->ac
, prim_stride_dw_vgpr
, prim_stride_dw
[buffer
],
2574 LLVMConstInt(ctx
->ac
.i32
, buffer
, false));
2576 LLVMValueRef offset
= LLVMConstInt(ctx
->ac
.i32
, buffer
, false);
2577 so_buffer
[buffer
] = ac_build_load_to_sgpr(&ctx
->ac
, buf_ptr
,
2581 cond
= LLVMBuildICmp(builder
, LLVMIntEQ
, get_wave_id_in_tg(ctx
), ctx
->ac
.i32_0
, "");
2582 ac_build_ifcc(&ctx
->ac
, cond
, 5200);
2584 LLVMTypeRef gdsptr
= LLVMPointerType(ctx
->ac
.i32
, AC_ADDR_SPACE_GDS
);
2585 LLVMValueRef gdsbase
= LLVMBuildIntToPtr(builder
, ctx
->ac
.i32_0
, gdsptr
, "");
2587 /* Advance the streamout offsets in GDS. */
2588 LLVMValueRef offsets_vgpr
= ac_build_alloca_undef(&ctx
->ac
, ctx
->ac
.i32
, "");
2589 LLVMValueRef generated_by_stream_vgpr
= ac_build_alloca_undef(&ctx
->ac
, ctx
->ac
.i32
, "");
2591 cond
= LLVMBuildICmp(builder
, LLVMIntULT
, ac_get_thread_id(&ctx
->ac
), i32_4
, "");
2592 ac_build_ifcc(&ctx
->ac
, cond
, 5210);
2594 /* Fetch the number of generated primitives and store
2595 * it in GDS for later use.
2598 tmp
= ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, tid
);
2599 tmp
= LLVMBuildLoad(builder
, tmp
, "");
2601 tmp
= ac_build_writelane(&ctx
->ac
, ctx
->ac
.i32_0
,
2602 ngg_get_prim_cnt(ctx
), ctx
->ac
.i32_0
);
2604 LLVMBuildStore(builder
, tmp
, generated_by_stream_vgpr
);
2606 unsigned swizzle
[4];
2607 int unused_stream
= -1;
2608 for (unsigned stream
= 0; stream
< 4; ++stream
) {
2609 if (!ctx
->args
->shader_info
->gs
.num_stream_output_components
[stream
]) {
2610 unused_stream
= stream
;
2614 for (unsigned buffer
= 0; buffer
< 4; ++buffer
) {
2615 if (stream_for_buffer
[buffer
] >= 0) {
2616 swizzle
[buffer
] = stream_for_buffer
[buffer
];
2618 assert(unused_stream
>= 0);
2619 swizzle
[buffer
] = unused_stream
;
2623 tmp
= ac_build_quad_swizzle(&ctx
->ac
, tmp
,
2624 swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
2625 tmp
= LLVMBuildMul(builder
, tmp
, prim_stride_dw_vgpr
, "");
2627 LLVMValueRef args
[] = {
2628 LLVMBuildIntToPtr(builder
, ngg_get_ordered_id(ctx
), gdsptr
, ""),
2630 ctx
->ac
.i32_0
, // ordering
2631 ctx
->ac
.i32_0
, // scope
2632 ctx
->ac
.i1false
, // isVolatile
2633 LLVMConstInt(ctx
->ac
.i32
, 4 << 24, false), // OA index
2634 ctx
->ac
.i1true
, // wave release
2635 ctx
->ac
.i1true
, // wave done
2638 tmp
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.ds.ordered.add",
2639 ctx
->ac
.i32
, args
, ARRAY_SIZE(args
), 0);
2641 /* Keep offsets in a VGPR for quick retrieval via readlane by
2642 * the first wave for bounds checking, and also store in LDS
2643 * for retrieval by all waves later. */
2644 LLVMBuildStore(builder
, tmp
, offsets_vgpr
);
2646 tmp2
= LLVMBuildAdd(builder
, ac_get_thread_id(&ctx
->ac
),
2647 scratch_offset_basev
, "");
2648 tmp2
= ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, tmp2
);
2649 LLVMBuildStore(builder
, tmp
, tmp2
);
2651 ac_build_endif(&ctx
->ac
, 5210);
2653 /* Determine the max emit per buffer. This is done via the SALU, in part
2654 * because LLVM can't generate divide-by-multiply if we try to do this
2655 * via VALU with one lane per buffer.
2657 LLVMValueRef max_emit
[4] = {};
2658 for (unsigned buffer
= 0; buffer
< 4; ++buffer
) {
2659 if (stream_for_buffer
[buffer
] == -1)
2662 /* Compute the streamout buffer size in DWORD. */
2663 LLVMValueRef bufsize_dw
=
2664 LLVMBuildLShr(builder
,
2665 LLVMBuildExtractElement(builder
, so_buffer
[buffer
], i32_2
, ""),
2668 /* Load the streamout buffer offset from GDS. */
2669 tmp
= LLVMBuildLoad(builder
, offsets_vgpr
, "");
2670 LLVMValueRef offset_dw
=
2671 ac_build_readlane(&ctx
->ac
, tmp
,
2672 LLVMConstInt(ctx
->ac
.i32
, buffer
, false));
2674 /* Compute the remaining size to emit. */
2675 LLVMValueRef remaining_dw
=
2676 LLVMBuildSub(builder
, bufsize_dw
, offset_dw
, "");
2677 tmp
= LLVMBuildUDiv(builder
, remaining_dw
,
2678 prim_stride_dw
[buffer
], "");
2680 cond
= LLVMBuildICmp(builder
, LLVMIntULT
,
2681 bufsize_dw
, offset_dw
, "");
2682 max_emit
[buffer
] = LLVMBuildSelect(builder
, cond
,
2683 ctx
->ac
.i32_0
, tmp
, "");
2686 /* Determine the number of emitted primitives per stream and fixup the
2687 * GDS counter if necessary.
2689 * This is complicated by the fact that a single stream can emit to
2690 * multiple buffers (but luckily not vice versa).
2692 LLVMValueRef emit_vgpr
= ctx
->ac
.i32_0
;
2694 for (unsigned stream
= 0; stream
< 4; ++stream
) {
2695 if (!ctx
->args
->shader_info
->gs
.num_stream_output_components
[stream
])
2698 /* Load the number of generated primitives from GDS and
2699 * determine that number for the given stream.
2701 tmp
= LLVMBuildLoad(builder
, generated_by_stream_vgpr
, "");
2702 LLVMValueRef generated
=
2703 ac_build_readlane(&ctx
->ac
, tmp
,
2704 LLVMConstInt(ctx
->ac
.i32
, stream
, false));
2707 /* Compute the number of emitted primitives. */
2708 LLVMValueRef emit
= generated
;
2709 for (unsigned buffer
= 0; buffer
< 4; ++buffer
) {
2710 if (stream_for_buffer
[buffer
] == stream
)
2711 emit
= ac_build_umin(&ctx
->ac
, emit
, max_emit
[buffer
]);
2714 /* Store the number of emitted primitives for that
2717 emit_vgpr
= ac_build_writelane(&ctx
->ac
, emit_vgpr
, emit
,
2718 LLVMConstInt(ctx
->ac
.i32
, stream
, false));
2720 /* Fixup the offset using a plain GDS atomic if we overflowed. */
2721 cond
= LLVMBuildICmp(builder
, LLVMIntULT
, emit
, generated
, "");
2722 ac_build_ifcc(&ctx
->ac
, cond
, 5221); /* scalar branch */
2723 tmp
= LLVMBuildLShr(builder
,
2724 LLVMConstInt(ctx
->ac
.i32
, bufmask_for_stream
[stream
], false),
2725 ac_get_thread_id(&ctx
->ac
), "");
2726 tmp
= LLVMBuildTrunc(builder
, tmp
, ctx
->ac
.i1
, "");
2727 ac_build_ifcc(&ctx
->ac
, tmp
, 5222);
2729 tmp
= LLVMBuildSub(builder
, generated
, emit
, "");
2730 tmp
= LLVMBuildMul(builder
, tmp
, prim_stride_dw_vgpr
, "");
2731 tmp2
= LLVMBuildGEP(builder
, gdsbase
, &tid
, 1, "");
2732 LLVMBuildAtomicRMW(builder
, LLVMAtomicRMWBinOpSub
, tmp2
, tmp
,
2733 LLVMAtomicOrderingMonotonic
, false);
2735 ac_build_endif(&ctx
->ac
, 5222);
2736 ac_build_endif(&ctx
->ac
, 5221);
2739 /* Store the number of emitted primitives to LDS for later use. */
2740 cond
= LLVMBuildICmp(builder
, LLVMIntULT
, ac_get_thread_id(&ctx
->ac
), i32_4
, "");
2741 ac_build_ifcc(&ctx
->ac
, cond
, 5225);
2743 tmp
= LLVMBuildAdd(builder
, ac_get_thread_id(&ctx
->ac
),
2744 scratch_emit_basev
, "");
2745 tmp
= ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, tmp
);
2746 LLVMBuildStore(builder
, emit_vgpr
, tmp
);
2748 ac_build_endif(&ctx
->ac
, 5225);
2750 ac_build_endif(&ctx
->ac
, 5200);
2752 /* Determine the workgroup-relative per-thread / primitive offset into
2753 * the streamout buffers */
2754 struct ac_wg_scan primemit_scan
[4] = {};
2757 for (unsigned stream
= 0; stream
< 4; ++stream
) {
2758 if (!ctx
->args
->shader_info
->gs
.num_stream_output_components
[stream
])
2761 primemit_scan
[stream
].enable_exclusive
= true;
2762 primemit_scan
[stream
].op
= nir_op_iadd
;
2763 primemit_scan
[stream
].src
= nggso
->prim_enable
[stream
];
2764 primemit_scan
[stream
].scratch
=
2765 ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
,
2766 LLVMConstInt(ctx
->ac
.i32
, 12 + 8 * stream
, false));
2767 primemit_scan
[stream
].waveidx
= get_wave_id_in_tg(ctx
);
2768 primemit_scan
[stream
].numwaves
= get_tgsize(ctx
);
2769 primemit_scan
[stream
].maxwaves
= 8;
2770 ac_build_wg_scan_top(&ctx
->ac
, &primemit_scan
[stream
]);
2774 ac_build_s_barrier(&ctx
->ac
);
2776 /* Fetch the per-buffer offsets and per-stream emit counts in all waves. */
2777 LLVMValueRef wgoffset_dw
[4] = {};
2780 LLVMValueRef scratch_vgpr
;
2782 tmp
= ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, ac_get_thread_id(&ctx
->ac
));
2783 scratch_vgpr
= LLVMBuildLoad(builder
, tmp
, "");
2785 for (unsigned buffer
= 0; buffer
< 4; ++buffer
) {
2786 if (stream_for_buffer
[buffer
] >= 0) {
2787 wgoffset_dw
[buffer
] = ac_build_readlane(
2788 &ctx
->ac
, scratch_vgpr
,
2789 LLVMConstInt(ctx
->ac
.i32
, scratch_offset_base
+ buffer
, false));
2793 for (unsigned stream
= 0; stream
< 4; ++stream
) {
2794 if (ctx
->args
->shader_info
->gs
.num_stream_output_components
[stream
]) {
2795 nggso
->emit
[stream
] = ac_build_readlane(
2796 &ctx
->ac
, scratch_vgpr
,
2797 LLVMConstInt(ctx
->ac
.i32
, scratch_emit_base
+ stream
, false));
2802 /* Write out primitive data */
2803 for (unsigned stream
= 0; stream
< 4; ++stream
) {
2804 if (!ctx
->args
->shader_info
->gs
.num_stream_output_components
[stream
])
2808 ac_build_wg_scan_bottom(&ctx
->ac
, &primemit_scan
[stream
]);
2810 primemit_scan
[stream
].result_exclusive
= tid
;
2813 cond
= LLVMBuildICmp(builder
, LLVMIntULT
,
2814 primemit_scan
[stream
].result_exclusive
,
2815 nggso
->emit
[stream
], "");
2816 cond
= LLVMBuildAnd(builder
, cond
, nggso
->prim_enable
[stream
], "");
2817 ac_build_ifcc(&ctx
->ac
, cond
, 5240);
2819 LLVMValueRef offset_vtx
=
2820 LLVMBuildMul(builder
, primemit_scan
[stream
].result_exclusive
,
2821 nggso
->num_vertices
, "");
2823 for (unsigned i
= 0; i
< max_num_vertices
; ++i
) {
2824 cond
= LLVMBuildICmp(builder
, LLVMIntULT
,
2825 LLVMConstInt(ctx
->ac
.i32
, i
, false),
2826 nggso
->num_vertices
, "");
2827 ac_build_ifcc(&ctx
->ac
, cond
, 5241);
2828 build_streamout_vertex(ctx
, so_buffer
, wgoffset_dw
,
2829 stream
, offset_vtx
, nggso
->vertices
[i
]);
2830 ac_build_endif(&ctx
->ac
, 5241);
2831 offset_vtx
= LLVMBuildAdd(builder
, offset_vtx
, ctx
->ac
.i32_1
, "");
2834 ac_build_endif(&ctx
->ac
, 5240);
2838 static unsigned ngg_nogs_vertex_size(struct radv_shader_context
*ctx
)
2840 unsigned lds_vertex_size
= 0;
2842 if (ctx
->args
->shader_info
->so
.num_outputs
)
2843 lds_vertex_size
= 4 * ctx
->args
->shader_info
->so
.num_outputs
+ 1;
2845 return lds_vertex_size
;
2849 * Returns an `[N x i32] addrspace(LDS)*` pointing at contiguous LDS storage
2850 * for the vertex outputs.
2852 static LLVMValueRef
ngg_nogs_vertex_ptr(struct radv_shader_context
*ctx
,
2855 /* The extra dword is used to avoid LDS bank conflicts. */
2856 unsigned vertex_size
= ngg_nogs_vertex_size(ctx
);
2857 LLVMTypeRef ai32
= LLVMArrayType(ctx
->ac
.i32
, vertex_size
);
2858 LLVMTypeRef pai32
= LLVMPointerType(ai32
, AC_ADDR_SPACE_LDS
);
2859 LLVMValueRef tmp
= LLVMBuildBitCast(ctx
->ac
.builder
, ctx
->esgs_ring
, pai32
, "");
2860 return LLVMBuildGEP(ctx
->ac
.builder
, tmp
, &vtxid
, 1, "");
2864 handle_ngg_outputs_post_1(struct radv_shader_context
*ctx
)
2866 struct radv_streamout_info
*so
= &ctx
->args
->shader_info
->so
;
2867 LLVMBuilderRef builder
= ctx
->ac
.builder
;
2868 LLVMValueRef vertex_ptr
= NULL
;
2869 LLVMValueRef tmp
, tmp2
;
2871 assert((ctx
->stage
== MESA_SHADER_VERTEX
||
2872 ctx
->stage
== MESA_SHADER_TESS_EVAL
) && !ctx
->args
->is_gs_copy_shader
);
2874 if (!ctx
->args
->shader_info
->so
.num_outputs
)
2877 vertex_ptr
= ngg_nogs_vertex_ptr(ctx
, get_thread_id_in_tg(ctx
));
2879 for (unsigned i
= 0; i
< so
->num_outputs
; ++i
) {
2880 struct radv_stream_output
*output
=
2881 &ctx
->args
->shader_info
->so
.outputs
[i
];
2883 unsigned loc
= output
->location
;
2885 for (unsigned comp
= 0; comp
< 4; comp
++) {
2886 if (!(output
->component_mask
& (1 << comp
)))
2889 tmp
= ac_build_gep0(&ctx
->ac
, vertex_ptr
,
2890 LLVMConstInt(ctx
->ac
.i32
, 4 * i
+ comp
, false));
2891 tmp2
= LLVMBuildLoad(builder
,
2892 ctx
->abi
.outputs
[4 * loc
+ comp
], "");
2893 tmp2
= ac_to_integer(&ctx
->ac
, tmp2
);
2894 LLVMBuildStore(builder
, tmp2
, tmp
);
2900 handle_ngg_outputs_post_2(struct radv_shader_context
*ctx
)
2902 LLVMBuilderRef builder
= ctx
->ac
.builder
;
2905 assert((ctx
->stage
== MESA_SHADER_VERTEX
||
2906 ctx
->stage
== MESA_SHADER_TESS_EVAL
) && !ctx
->args
->is_gs_copy_shader
);
2908 LLVMValueRef prims_in_wave
= ac_unpack_param(&ctx
->ac
,
2909 ac_get_arg(&ctx
->ac
, ctx
->args
->merged_wave_info
), 8, 8);
2910 LLVMValueRef vtx_in_wave
= ac_unpack_param(&ctx
->ac
,
2911 ac_get_arg(&ctx
->ac
, ctx
->args
->merged_wave_info
), 0, 8);
2912 LLVMValueRef is_gs_thread
= LLVMBuildICmp(builder
, LLVMIntULT
,
2913 ac_get_thread_id(&ctx
->ac
), prims_in_wave
, "");
2914 LLVMValueRef is_es_thread
= LLVMBuildICmp(builder
, LLVMIntULT
,
2915 ac_get_thread_id(&ctx
->ac
), vtx_in_wave
, "");
2916 LLVMValueRef vtxindex
[] = {
2917 ac_unpack_param(&ctx
->ac
, ac_get_arg(&ctx
->ac
, ctx
->args
->gs_vtx_offset
[0]), 0, 16),
2918 ac_unpack_param(&ctx
->ac
, ac_get_arg(&ctx
->ac
, ctx
->args
->gs_vtx_offset
[0]), 16, 16),
2919 ac_unpack_param(&ctx
->ac
, ac_get_arg(&ctx
->ac
, ctx
->args
->gs_vtx_offset
[2]), 0, 16),
2922 /* Determine the number of vertices per primitive. */
2923 unsigned num_vertices
;
2924 LLVMValueRef num_vertices_val
;
2926 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
2927 LLVMValueRef outprim_val
=
2928 LLVMConstInt(ctx
->ac
.i32
,
2929 ctx
->args
->options
->key
.vs
.outprim
, false);
2930 num_vertices_val
= LLVMBuildAdd(builder
, outprim_val
,
2932 num_vertices
= 3; /* TODO: optimize for points & lines */
2934 assert(ctx
->stage
== MESA_SHADER_TESS_EVAL
);
2936 if (ctx
->shader
->info
.tess
.point_mode
)
2938 else if (ctx
->shader
->info
.tess
.primitive_mode
== GL_ISOLINES
)
2943 num_vertices_val
= LLVMConstInt(ctx
->ac
.i32
, num_vertices
, false);
2947 if (ctx
->args
->shader_info
->so
.num_outputs
) {
2948 struct ngg_streamout nggso
= {};
2950 nggso
.num_vertices
= num_vertices_val
;
2951 nggso
.prim_enable
[0] = is_gs_thread
;
2953 for (unsigned i
= 0; i
< num_vertices
; ++i
)
2954 nggso
.vertices
[i
] = ngg_nogs_vertex_ptr(ctx
, vtxindex
[i
]);
2956 build_streamout(ctx
, &nggso
);
2959 /* Copy Primitive IDs from GS threads to the LDS address corresponding
2960 * to the ES thread of the provoking vertex.
2962 if (ctx
->stage
== MESA_SHADER_VERTEX
&&
2963 ctx
->args
->options
->key
.vs_common_out
.export_prim_id
) {
2964 if (ctx
->args
->shader_info
->so
.num_outputs
)
2965 ac_build_s_barrier(&ctx
->ac
);
2967 ac_build_ifcc(&ctx
->ac
, is_gs_thread
, 5400);
2968 /* Extract the PROVOKING_VTX_INDEX field. */
2969 LLVMValueRef provoking_vtx_in_prim
=
2970 LLVMConstInt(ctx
->ac
.i32
, 0, false);
2972 /* provoking_vtx_index = vtxindex[provoking_vtx_in_prim]; */
2973 LLVMValueRef indices
= ac_build_gather_values(&ctx
->ac
, vtxindex
, 3);
2974 LLVMValueRef provoking_vtx_index
=
2975 LLVMBuildExtractElement(builder
, indices
, provoking_vtx_in_prim
, "");
2977 LLVMBuildStore(builder
, ac_get_arg(&ctx
->ac
, ctx
->args
->ac
.gs_prim_id
),
2978 ac_build_gep0(&ctx
->ac
, ctx
->esgs_ring
, provoking_vtx_index
));
2979 ac_build_endif(&ctx
->ac
, 5400);
2982 /* TODO: primitive culling */
2984 ac_build_sendmsg_gs_alloc_req(&ctx
->ac
, get_wave_id_in_tg(ctx
),
2985 ngg_get_vtx_cnt(ctx
), ngg_get_prim_cnt(ctx
));
2987 /* TODO: streamout queries */
2988 /* Export primitive data to the index buffer.
2990 * For the first version, we will always build up all three indices
2991 * independent of the primitive type. The additional garbage data
2994 * TODO: culling depends on the primitive type, so can have some
2997 ac_build_ifcc(&ctx
->ac
, is_gs_thread
, 6001);
2999 struct ac_ngg_prim prim
= {};
3001 if (ctx
->args
->options
->key
.vs_common_out
.as_ngg_passthrough
) {
3002 prim
.passthrough
= ac_get_arg(&ctx
->ac
, ctx
->args
->gs_vtx_offset
[0]);
3004 prim
.num_vertices
= num_vertices
;
3005 prim
.isnull
= ctx
->ac
.i1false
;
3006 memcpy(prim
.index
, vtxindex
, sizeof(vtxindex
[0]) * 3);
3008 for (unsigned i
= 0; i
< num_vertices
; ++i
) {
3009 tmp
= LLVMBuildLShr(builder
,
3010 ac_get_arg(&ctx
->ac
, ctx
->args
->ac
.gs_invocation_id
),
3011 LLVMConstInt(ctx
->ac
.i32
, 8 + i
, false), "");
3012 prim
.edgeflag
[i
] = LLVMBuildTrunc(builder
, tmp
, ctx
->ac
.i1
, "");
3016 ac_build_export_prim(&ctx
->ac
, &prim
);
3018 ac_build_endif(&ctx
->ac
, 6001);
3020 /* Export per-vertex data (positions and parameters). */
3021 ac_build_ifcc(&ctx
->ac
, is_es_thread
, 6002);
3023 struct radv_vs_output_info
*outinfo
=
3024 ctx
->stage
== MESA_SHADER_TESS_EVAL
?
3025 &ctx
->args
->shader_info
->tes
.outinfo
: &ctx
->args
->shader_info
->vs
.outinfo
;
3027 /* Exporting the primitive ID is handled below. */
3028 /* TODO: use the new VS export path */
3029 handle_vs_outputs_post(ctx
, false,
3030 ctx
->args
->options
->key
.vs_common_out
.export_clip_dists
,
3033 if (ctx
->args
->options
->key
.vs_common_out
.export_prim_id
) {
3034 unsigned param_count
= outinfo
->param_exports
;
3035 LLVMValueRef values
[4];
3037 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
3038 /* Wait for GS stores to finish. */
3039 ac_build_s_barrier(&ctx
->ac
);
3041 tmp
= ac_build_gep0(&ctx
->ac
, ctx
->esgs_ring
,
3042 get_thread_id_in_tg(ctx
));
3043 values
[0] = LLVMBuildLoad(builder
, tmp
, "");
3045 assert(ctx
->stage
== MESA_SHADER_TESS_EVAL
);
3046 values
[0] = ac_get_arg(&ctx
->ac
, ctx
->args
->ac
.tes_patch_id
);
3049 values
[0] = ac_to_float(&ctx
->ac
, values
[0]);
3050 for (unsigned j
= 1; j
< 4; j
++)
3051 values
[j
] = ctx
->ac
.f32_0
;
3053 radv_export_param(ctx
, param_count
, values
, 0x1);
3055 outinfo
->vs_output_param_offset
[VARYING_SLOT_PRIMITIVE_ID
] = param_count
++;
3056 outinfo
->param_exports
= param_count
;
3059 ac_build_endif(&ctx
->ac
, 6002);
3062 static void gfx10_ngg_gs_emit_prologue(struct radv_shader_context
*ctx
)
3064 /* Zero out the part of LDS scratch that is used to accumulate the
3065 * per-stream generated primitive count.
3067 LLVMBuilderRef builder
= ctx
->ac
.builder
;
3068 LLVMValueRef scratchptr
= ctx
->gs_ngg_scratch
;
3069 LLVMValueRef tid
= get_thread_id_in_tg(ctx
);
3070 LLVMBasicBlockRef merge_block
;
3073 LLVMValueRef fn
= LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
->ac
.builder
));
3074 LLVMBasicBlockRef then_block
= LLVMAppendBasicBlockInContext(ctx
->ac
.context
, fn
, "");
3075 merge_block
= LLVMAppendBasicBlockInContext(ctx
->ac
.context
, fn
, "");
3077 cond
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
3078 LLVMBuildCondBr(ctx
->ac
.builder
, cond
, then_block
, merge_block
);
3079 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, then_block
);
3081 LLVMValueRef ptr
= ac_build_gep0(&ctx
->ac
, scratchptr
, tid
);
3082 LLVMBuildStore(builder
, ctx
->ac
.i32_0
, ptr
);
3084 LLVMBuildBr(ctx
->ac
.builder
, merge_block
);
3085 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, merge_block
);
3087 ac_build_s_barrier(&ctx
->ac
);
3090 static void gfx10_ngg_gs_emit_epilogue_1(struct radv_shader_context
*ctx
)
3092 LLVMBuilderRef builder
= ctx
->ac
.builder
;
3093 LLVMValueRef i8_0
= LLVMConstInt(ctx
->ac
.i8
, 0, false);
3096 /* Zero out remaining (non-emitted) primitive flags.
3098 * Note: Alternatively, we could pass the relevant gs_next_vertex to
3099 * the emit threads via LDS. This is likely worse in the expected
3100 * typical case where each GS thread emits the full set of
3103 for (unsigned stream
= 0; stream
< 4; ++stream
) {
3104 unsigned num_components
;
3107 ctx
->args
->shader_info
->gs
.num_stream_output_components
[stream
];
3108 if (!num_components
)
3111 const LLVMValueRef gsthread
= get_thread_id_in_tg(ctx
);
3113 ac_build_bgnloop(&ctx
->ac
, 5100);
3115 const LLVMValueRef vertexidx
=
3116 LLVMBuildLoad(builder
, ctx
->gs_next_vertex
[stream
], "");
3117 tmp
= LLVMBuildICmp(builder
, LLVMIntUGE
, vertexidx
,
3118 LLVMConstInt(ctx
->ac
.i32
, ctx
->shader
->info
.gs
.vertices_out
, false), "");
3119 ac_build_ifcc(&ctx
->ac
, tmp
, 5101);
3120 ac_build_break(&ctx
->ac
);
3121 ac_build_endif(&ctx
->ac
, 5101);
3123 tmp
= LLVMBuildAdd(builder
, vertexidx
, ctx
->ac
.i32_1
, "");
3124 LLVMBuildStore(builder
, tmp
, ctx
->gs_next_vertex
[stream
]);
3126 tmp
= ngg_gs_emit_vertex_ptr(ctx
, gsthread
, vertexidx
);
3127 LLVMBuildStore(builder
, i8_0
,
3128 ngg_gs_get_emit_primflag_ptr(ctx
, tmp
, stream
));
3130 ac_build_endloop(&ctx
->ac
, 5100);
3133 /* Accumulate generated primitives counts across the entire threadgroup. */
3134 for (unsigned stream
= 0; stream
< 4; ++stream
) {
3135 unsigned num_components
;
3138 ctx
->args
->shader_info
->gs
.num_stream_output_components
[stream
];
3139 if (!num_components
)
3142 LLVMValueRef numprims
=
3143 LLVMBuildLoad(builder
, ctx
->gs_generated_prims
[stream
], "");
3144 numprims
= ac_build_reduce(&ctx
->ac
, numprims
, nir_op_iadd
, ctx
->ac
.wave_size
);
3146 tmp
= LLVMBuildICmp(builder
, LLVMIntEQ
, ac_get_thread_id(&ctx
->ac
), ctx
->ac
.i32_0
, "");
3147 ac_build_ifcc(&ctx
->ac
, tmp
, 5105);
3149 LLVMBuildAtomicRMW(builder
, LLVMAtomicRMWBinOpAdd
,
3150 ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
,
3151 LLVMConstInt(ctx
->ac
.i32
, stream
, false)),
3152 numprims
, LLVMAtomicOrderingMonotonic
, false);
3154 ac_build_endif(&ctx
->ac
, 5105);
3158 static void gfx10_ngg_gs_emit_epilogue_2(struct radv_shader_context
*ctx
)
3160 const unsigned verts_per_prim
= si_conv_gl_prim_to_vertices(ctx
->shader
->info
.gs
.output_primitive
);
3161 LLVMBuilderRef builder
= ctx
->ac
.builder
;
3162 LLVMValueRef tmp
, tmp2
;
3164 ac_build_s_barrier(&ctx
->ac
);
3166 const LLVMValueRef tid
= get_thread_id_in_tg(ctx
);
3167 LLVMValueRef num_emit_threads
= ngg_get_prim_cnt(ctx
);
3170 if (ctx
->args
->shader_info
->so
.num_outputs
) {
3171 struct ngg_streamout nggso
= {};
3173 nggso
.num_vertices
= LLVMConstInt(ctx
->ac
.i32
, verts_per_prim
, false);
3175 LLVMValueRef vertexptr
= ngg_gs_vertex_ptr(ctx
, tid
);
3176 for (unsigned stream
= 0; stream
< 4; ++stream
) {
3177 if (!ctx
->args
->shader_info
->gs
.num_stream_output_components
[stream
])
3180 tmp
= LLVMBuildLoad(builder
,
3181 ngg_gs_get_emit_primflag_ptr(ctx
, vertexptr
, stream
), "");
3182 tmp
= LLVMBuildTrunc(builder
, tmp
, ctx
->ac
.i1
, "");
3183 tmp2
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, num_emit_threads
, "");
3184 nggso
.prim_enable
[stream
] = LLVMBuildAnd(builder
, tmp
, tmp2
, "");
3187 for (unsigned i
= 0; i
< verts_per_prim
; ++i
) {
3188 tmp
= LLVMBuildSub(builder
, tid
,
3189 LLVMConstInt(ctx
->ac
.i32
, verts_per_prim
- i
- 1, false), "");
3190 tmp
= ngg_gs_vertex_ptr(ctx
, tmp
);
3191 nggso
.vertices
[i
] = ac_build_gep0(&ctx
->ac
, tmp
, ctx
->ac
.i32_0
);
3194 build_streamout(ctx
, &nggso
);
3199 /* Determine vertex liveness. */
3200 LLVMValueRef vertliveptr
= ac_build_alloca(&ctx
->ac
, ctx
->ac
.i1
, "vertexlive");
3202 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, num_emit_threads
, "");
3203 ac_build_ifcc(&ctx
->ac
, tmp
, 5120);
3205 for (unsigned i
= 0; i
< verts_per_prim
; ++i
) {
3206 const LLVMValueRef primidx
=
3207 LLVMBuildAdd(builder
, tid
,
3208 LLVMConstInt(ctx
->ac
.i32
, i
, false), "");
3211 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, primidx
, num_emit_threads
, "");
3212 ac_build_ifcc(&ctx
->ac
, tmp
, 5121 + i
);
3215 /* Load primitive liveness */
3216 tmp
= ngg_gs_vertex_ptr(ctx
, primidx
);
3217 tmp
= LLVMBuildLoad(builder
,
3218 ngg_gs_get_emit_primflag_ptr(ctx
, tmp
, 0), "");
3219 const LLVMValueRef primlive
=
3220 LLVMBuildTrunc(builder
, tmp
, ctx
->ac
.i1
, "");
3222 tmp
= LLVMBuildLoad(builder
, vertliveptr
, "");
3223 tmp
= LLVMBuildOr(builder
, tmp
, primlive
, ""),
3224 LLVMBuildStore(builder
, tmp
, vertliveptr
);
3227 ac_build_endif(&ctx
->ac
, 5121 + i
);
3230 ac_build_endif(&ctx
->ac
, 5120);
3232 /* Inclusive scan addition across the current wave. */
3233 LLVMValueRef vertlive
= LLVMBuildLoad(builder
, vertliveptr
, "");
3234 struct ac_wg_scan vertlive_scan
= {};
3235 vertlive_scan
.op
= nir_op_iadd
;
3236 vertlive_scan
.enable_reduce
= true;
3237 vertlive_scan
.enable_exclusive
= true;
3238 vertlive_scan
.src
= vertlive
;
3239 vertlive_scan
.scratch
= ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, ctx
->ac
.i32_0
);
3240 vertlive_scan
.waveidx
= get_wave_id_in_tg(ctx
);
3241 vertlive_scan
.numwaves
= get_tgsize(ctx
);
3242 vertlive_scan
.maxwaves
= 8;
3244 ac_build_wg_scan(&ctx
->ac
, &vertlive_scan
);
3246 /* Skip all exports (including index exports) when possible. At least on
3247 * early gfx10 revisions this is also to avoid hangs.
3249 LLVMValueRef have_exports
=
3250 LLVMBuildICmp(builder
, LLVMIntNE
, vertlive_scan
.result_reduce
, ctx
->ac
.i32_0
, "");
3252 LLVMBuildSelect(builder
, have_exports
, num_emit_threads
, ctx
->ac
.i32_0
, "");
3254 /* Allocate export space. Send this message as early as possible, to
3255 * hide the latency of the SQ <-> SPI roundtrip.
3257 * Note: We could consider compacting primitives for export as well.
3258 * PA processes 1 non-null prim / clock, but it fetches 4 DW of
3259 * prim data per clock and skips null primitives at no additional
3260 * cost. So compacting primitives can only be beneficial when
3261 * there are 4 or more contiguous null primitives in the export
3262 * (in the common case of single-dword prim exports).
3264 ac_build_sendmsg_gs_alloc_req(&ctx
->ac
, get_wave_id_in_tg(ctx
),
3265 vertlive_scan
.result_reduce
, num_emit_threads
);
3267 /* Setup the reverse vertex compaction permutation. We re-use stream 1
3268 * of the primitive liveness flags, relying on the fact that each
3269 * threadgroup can have at most 256 threads. */
3270 ac_build_ifcc(&ctx
->ac
, vertlive
, 5130);
3272 tmp
= ngg_gs_vertex_ptr(ctx
, vertlive_scan
.result_exclusive
);
3273 tmp2
= LLVMBuildTrunc(builder
, tid
, ctx
->ac
.i8
, "");
3274 LLVMBuildStore(builder
, tmp2
,
3275 ngg_gs_get_emit_primflag_ptr(ctx
, tmp
, 1));
3277 ac_build_endif(&ctx
->ac
, 5130);
3279 ac_build_s_barrier(&ctx
->ac
);
3281 /* Export primitive data */
3282 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, num_emit_threads
, "");
3283 ac_build_ifcc(&ctx
->ac
, tmp
, 5140);
3286 struct ac_ngg_prim prim
= {};
3287 prim
.num_vertices
= verts_per_prim
;
3289 tmp
= ngg_gs_vertex_ptr(ctx
, tid
);
3290 flags
= LLVMBuildLoad(builder
,
3291 ngg_gs_get_emit_primflag_ptr(ctx
, tmp
, 0), "");
3292 prim
.isnull
= LLVMBuildNot(builder
, LLVMBuildTrunc(builder
, flags
, ctx
->ac
.i1
, ""), "");
3294 for (unsigned i
= 0; i
< verts_per_prim
; ++i
) {
3295 prim
.index
[i
] = LLVMBuildSub(builder
, vertlive_scan
.result_exclusive
,
3296 LLVMConstInt(ctx
->ac
.i32
, verts_per_prim
- i
- 1, false), "");
3297 prim
.edgeflag
[i
] = ctx
->ac
.i1false
;
3300 /* Geometry shaders output triangle strips, but NGG expects
3301 * triangles. We need to change the vertex order for odd
3302 * triangles to get correct front/back facing by swapping 2
3303 * vertex indices, but we also have to keep the provoking
3304 * vertex in the same place.
3306 if (verts_per_prim
== 3) {
3307 LLVMValueRef is_odd
= LLVMBuildLShr(builder
, flags
, ctx
->ac
.i8_1
, "");
3308 is_odd
= LLVMBuildTrunc(builder
, is_odd
, ctx
->ac
.i1
, "");
3310 struct ac_ngg_prim in
= prim
;
3311 prim
.index
[0] = in
.index
[0];
3312 prim
.index
[1] = LLVMBuildSelect(builder
, is_odd
,
3313 in
.index
[2], in
.index
[1], "");
3314 prim
.index
[2] = LLVMBuildSelect(builder
, is_odd
,
3315 in
.index
[1], in
.index
[2], "");
3318 ac_build_export_prim(&ctx
->ac
, &prim
);
3320 ac_build_endif(&ctx
->ac
, 5140);
3322 /* Export position and parameter data */
3323 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, vertlive_scan
.result_reduce
, "");
3324 ac_build_ifcc(&ctx
->ac
, tmp
, 5145);
3326 struct radv_vs_output_info
*outinfo
= &ctx
->args
->shader_info
->vs
.outinfo
;
3327 bool export_view_index
= ctx
->args
->options
->key
.has_multiview_view_index
;
3328 struct radv_shader_output_values
*outputs
;
3329 unsigned noutput
= 0;
3331 /* Allocate a temporary array for the output values. */
3332 unsigned num_outputs
= util_bitcount64(ctx
->output_mask
) + export_view_index
;
3333 outputs
= calloc(num_outputs
, sizeof(outputs
[0]));
3335 memset(outinfo
->vs_output_param_offset
, AC_EXP_PARAM_UNDEFINED
,
3336 sizeof(outinfo
->vs_output_param_offset
));
3337 outinfo
->pos_exports
= 0;
3339 tmp
= ngg_gs_vertex_ptr(ctx
, tid
);
3340 tmp
= LLVMBuildLoad(builder
,
3341 ngg_gs_get_emit_primflag_ptr(ctx
, tmp
, 1), "");
3342 tmp
= LLVMBuildZExt(builder
, tmp
, ctx
->ac
.i32
, "");
3343 const LLVMValueRef vertexptr
= ngg_gs_vertex_ptr(ctx
, tmp
);
3345 unsigned out_idx
= 0;
3346 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
3347 unsigned output_usage_mask
=
3348 ctx
->args
->shader_info
->gs
.output_usage_mask
[i
];
3349 int length
= util_last_bit(output_usage_mask
);
3351 if (!(ctx
->output_mask
& (1ull << i
)))
3354 outputs
[noutput
].slot_name
= i
;
3355 outputs
[noutput
].slot_index
= i
== VARYING_SLOT_CLIP_DIST1
;
3356 outputs
[noutput
].usage_mask
= output_usage_mask
;
3358 for (unsigned j
= 0; j
< length
; j
++, out_idx
++) {
3359 if (!(output_usage_mask
& (1 << j
)))
3362 tmp
= ngg_gs_get_emit_output_ptr(ctx
, vertexptr
, out_idx
);
3363 tmp
= LLVMBuildLoad(builder
, tmp
, "");
3365 LLVMTypeRef type
= LLVMGetAllocatedType(ctx
->abi
.outputs
[ac_llvm_reg_index_soa(i
, j
)]);
3366 if (ac_get_type_size(type
) == 2) {
3367 tmp
= ac_to_integer(&ctx
->ac
, tmp
);
3368 tmp
= LLVMBuildTrunc(ctx
->ac
.builder
, tmp
, ctx
->ac
.i16
, "");
3371 outputs
[noutput
].values
[j
] = ac_to_float(&ctx
->ac
, tmp
);
3374 for (unsigned j
= length
; j
< 4; j
++)
3375 outputs
[noutput
].values
[j
] = LLVMGetUndef(ctx
->ac
.f32
);
3380 /* Export ViewIndex. */
3381 if (export_view_index
) {
3382 outputs
[noutput
].slot_name
= VARYING_SLOT_LAYER
;
3383 outputs
[noutput
].slot_index
= 0;
3384 outputs
[noutput
].usage_mask
= 0x1;
3385 outputs
[noutput
].values
[0] =
3386 ac_to_float(&ctx
->ac
, ac_get_arg(&ctx
->ac
, ctx
->args
->ac
.view_index
));
3387 for (unsigned j
= 1; j
< 4; j
++)
3388 outputs
[noutput
].values
[j
] = ctx
->ac
.f32_0
;
3392 radv_llvm_export_vs(ctx
, outputs
, noutput
, outinfo
,
3393 ctx
->args
->options
->key
.vs_common_out
.export_clip_dists
);
3396 ac_build_endif(&ctx
->ac
, 5145);
3399 static void gfx10_ngg_gs_emit_vertex(struct radv_shader_context
*ctx
,
3401 LLVMValueRef
*addrs
)
3403 LLVMBuilderRef builder
= ctx
->ac
.builder
;
3405 const LLVMValueRef vertexidx
=
3406 LLVMBuildLoad(builder
, ctx
->gs_next_vertex
[stream
], "");
3408 /* If this thread has already emitted the declared maximum number of
3409 * vertices, skip the write: excessive vertex emissions are not
3410 * supposed to have any effect.
3412 const LLVMValueRef can_emit
=
3413 LLVMBuildICmp(builder
, LLVMIntULT
, vertexidx
,
3414 LLVMConstInt(ctx
->ac
.i32
, ctx
->shader
->info
.gs
.vertices_out
, false), "");
3415 ac_build_ifcc(&ctx
->ac
, can_emit
, 9001);
3417 tmp
= LLVMBuildAdd(builder
, vertexidx
, ctx
->ac
.i32_1
, "");
3418 tmp
= LLVMBuildSelect(builder
, can_emit
, tmp
, vertexidx
, "");
3419 LLVMBuildStore(builder
, tmp
, ctx
->gs_next_vertex
[stream
]);
3421 const LLVMValueRef vertexptr
=
3422 ngg_gs_emit_vertex_ptr(ctx
, get_thread_id_in_tg(ctx
), vertexidx
);
3423 unsigned out_idx
= 0;
3424 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
3425 unsigned output_usage_mask
=
3426 ctx
->args
->shader_info
->gs
.output_usage_mask
[i
];
3427 uint8_t output_stream
=
3428 ctx
->args
->shader_info
->gs
.output_streams
[i
];
3429 LLVMValueRef
*out_ptr
= &addrs
[i
* 4];
3430 int length
= util_last_bit(output_usage_mask
);
3432 if (!(ctx
->output_mask
& (1ull << i
)) ||
3433 output_stream
!= stream
)
3436 for (unsigned j
= 0; j
< length
; j
++, out_idx
++) {
3437 if (!(output_usage_mask
& (1 << j
)))
3440 LLVMValueRef out_val
= LLVMBuildLoad(ctx
->ac
.builder
,
3442 out_val
= ac_to_integer(&ctx
->ac
, out_val
);
3443 out_val
= LLVMBuildZExtOrBitCast(ctx
->ac
.builder
, out_val
, ctx
->ac
.i32
, "");
3445 LLVMBuildStore(builder
, out_val
,
3446 ngg_gs_get_emit_output_ptr(ctx
, vertexptr
, out_idx
));
3449 assert(out_idx
* 4 <= ctx
->args
->shader_info
->gs
.gsvs_vertex_size
);
3451 /* Determine and store whether this vertex completed a primitive. */
3452 const LLVMValueRef curverts
= LLVMBuildLoad(builder
, ctx
->gs_curprim_verts
[stream
], "");
3454 tmp
= LLVMConstInt(ctx
->ac
.i32
, si_conv_gl_prim_to_vertices(ctx
->shader
->info
.gs
.output_primitive
) - 1, false);
3455 const LLVMValueRef iscompleteprim
=
3456 LLVMBuildICmp(builder
, LLVMIntUGE
, curverts
, tmp
, "");
3458 /* Since the geometry shader emits triangle strips, we need to
3459 * track which primitive is odd and swap vertex indices to get
3460 * the correct vertex order.
3462 LLVMValueRef is_odd
= ctx
->ac
.i1false
;
3464 si_conv_gl_prim_to_vertices(ctx
->shader
->info
.gs
.output_primitive
) == 3) {
3465 tmp
= LLVMBuildAnd(builder
, curverts
, ctx
->ac
.i32_1
, "");
3466 is_odd
= LLVMBuildICmp(builder
, LLVMIntEQ
, tmp
, ctx
->ac
.i32_1
, "");
3469 tmp
= LLVMBuildAdd(builder
, curverts
, ctx
->ac
.i32_1
, "");
3470 LLVMBuildStore(builder
, tmp
, ctx
->gs_curprim_verts
[stream
]);
3472 /* The per-vertex primitive flag encoding:
3473 * bit 0: whether this vertex finishes a primitive
3474 * bit 1: whether the primitive is odd (if we are emitting triangle strips)
3476 tmp
= LLVMBuildZExt(builder
, iscompleteprim
, ctx
->ac
.i8
, "");
3477 tmp
= LLVMBuildOr(builder
, tmp
,
3478 LLVMBuildShl(builder
,
3479 LLVMBuildZExt(builder
, is_odd
, ctx
->ac
.i8
, ""),
3480 ctx
->ac
.i8_1
, ""), "");
3481 LLVMBuildStore(builder
, tmp
,
3482 ngg_gs_get_emit_primflag_ptr(ctx
, vertexptr
, stream
));
3484 tmp
= LLVMBuildLoad(builder
, ctx
->gs_generated_prims
[stream
], "");
3485 tmp
= LLVMBuildAdd(builder
, tmp
, LLVMBuildZExt(builder
, iscompleteprim
, ctx
->ac
.i32
, ""), "");
3486 LLVMBuildStore(builder
, tmp
, ctx
->gs_generated_prims
[stream
]);
3488 ac_build_endif(&ctx
->ac
, 9001);
3492 write_tess_factors(struct radv_shader_context
*ctx
)
3494 unsigned stride
, outer_comps
, inner_comps
;
3495 LLVMValueRef tcs_rel_ids
= ac_get_arg(&ctx
->ac
, ctx
->args
->ac
.tcs_rel_ids
);
3496 LLVMValueRef invocation_id
= ac_unpack_param(&ctx
->ac
, tcs_rel_ids
, 8, 5);
3497 LLVMValueRef rel_patch_id
= ac_unpack_param(&ctx
->ac
, tcs_rel_ids
, 0, 8);
3498 unsigned tess_inner_index
= 0, tess_outer_index
;
3499 LLVMValueRef lds_base
, lds_inner
= NULL
, lds_outer
, byteoffset
, buffer
;
3500 LLVMValueRef out
[6], vec0
, vec1
, tf_base
, inner
[4], outer
[4];
3502 ac_emit_barrier(&ctx
->ac
, ctx
->stage
);
3504 switch (ctx
->args
->options
->key
.tcs
.primitive_mode
) {
3524 ac_build_ifcc(&ctx
->ac
,
3525 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
,
3526 invocation_id
, ctx
->ac
.i32_0
, ""), 6503);
3528 lds_base
= get_tcs_out_current_patch_data_offset(ctx
);
3531 tess_inner_index
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_INNER
);
3532 lds_inner
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
3533 LLVMConstInt(ctx
->ac
.i32
, tess_inner_index
* 4, false), "");
3536 tess_outer_index
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_OUTER
);
3537 lds_outer
= LLVMBuildAdd(ctx
->ac
.builder
, lds_base
,
3538 LLVMConstInt(ctx
->ac
.i32
, tess_outer_index
* 4, false), "");
3540 for (i
= 0; i
< 4; i
++) {
3541 inner
[i
] = LLVMGetUndef(ctx
->ac
.i32
);
3542 outer
[i
] = LLVMGetUndef(ctx
->ac
.i32
);
3546 if (ctx
->args
->options
->key
.tcs
.primitive_mode
== GL_ISOLINES
) {
3547 outer
[0] = out
[1] = ac_lds_load(&ctx
->ac
, lds_outer
);
3548 lds_outer
= LLVMBuildAdd(ctx
->ac
.builder
, lds_outer
,
3550 outer
[1] = out
[0] = ac_lds_load(&ctx
->ac
, lds_outer
);
3552 for (i
= 0; i
< outer_comps
; i
++) {
3554 ac_lds_load(&ctx
->ac
, lds_outer
);
3555 lds_outer
= LLVMBuildAdd(ctx
->ac
.builder
, lds_outer
,
3558 for (i
= 0; i
< inner_comps
; i
++) {
3559 inner
[i
] = out
[outer_comps
+i
] =
3560 ac_lds_load(&ctx
->ac
, lds_inner
);
3561 lds_inner
= LLVMBuildAdd(ctx
->ac
.builder
, lds_inner
,
3566 /* Convert the outputs to vectors for stores. */
3567 vec0
= ac_build_gather_values(&ctx
->ac
, out
, MIN2(stride
, 4));
3571 vec1
= ac_build_gather_values(&ctx
->ac
, out
+ 4, stride
- 4);
3574 buffer
= ctx
->hs_ring_tess_factor
;
3575 tf_base
= ac_get_arg(&ctx
->ac
, ctx
->args
->tess_factor_offset
);
3576 byteoffset
= LLVMBuildMul(ctx
->ac
.builder
, rel_patch_id
,
3577 LLVMConstInt(ctx
->ac
.i32
, 4 * stride
, false), "");
3578 unsigned tf_offset
= 0;
3580 if (ctx
->ac
.chip_class
<= GFX8
) {
3581 ac_build_ifcc(&ctx
->ac
,
3582 LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
,
3583 rel_patch_id
, ctx
->ac
.i32_0
, ""), 6504);
3585 /* Store the dynamic HS control word. */
3586 ac_build_buffer_store_dword(&ctx
->ac
, buffer
,
3587 LLVMConstInt(ctx
->ac
.i32
, 0x80000000, false),
3588 1, ctx
->ac
.i32_0
, tf_base
,
3592 ac_build_endif(&ctx
->ac
, 6504);
3595 /* Store the tessellation factors. */
3596 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec0
,
3597 MIN2(stride
, 4), byteoffset
, tf_base
,
3600 ac_build_buffer_store_dword(&ctx
->ac
, buffer
, vec1
,
3601 stride
- 4, byteoffset
, tf_base
,
3602 16 + tf_offset
, ac_glc
);
3604 //store to offchip for TES to read - only if TES reads them
3605 if (ctx
->args
->options
->key
.tcs
.tes_reads_tess_factors
) {
3606 LLVMValueRef inner_vec
, outer_vec
, tf_outer_offset
;
3607 LLVMValueRef tf_inner_offset
;
3608 unsigned param_outer
, param_inner
;
3610 param_outer
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_OUTER
);
3611 tf_outer_offset
= get_tcs_tes_buffer_address(ctx
, NULL
,
3612 LLVMConstInt(ctx
->ac
.i32
, param_outer
, 0));
3614 outer_vec
= ac_build_gather_values(&ctx
->ac
, outer
,
3615 util_next_power_of_two(outer_comps
));
3617 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, outer_vec
,
3618 outer_comps
, tf_outer_offset
,
3619 ac_get_arg(&ctx
->ac
, ctx
->args
->oc_lds
),
3622 param_inner
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_INNER
);
3623 tf_inner_offset
= get_tcs_tes_buffer_address(ctx
, NULL
,
3624 LLVMConstInt(ctx
->ac
.i32
, param_inner
, 0));
3626 inner_vec
= inner_comps
== 1 ? inner
[0] :
3627 ac_build_gather_values(&ctx
->ac
, inner
, inner_comps
);
3628 ac_build_buffer_store_dword(&ctx
->ac
, ctx
->hs_ring_tess_offchip
, inner_vec
,
3629 inner_comps
, tf_inner_offset
,
3630 ac_get_arg(&ctx
->ac
, ctx
->args
->oc_lds
),
3635 ac_build_endif(&ctx
->ac
, 6503);
3639 handle_tcs_outputs_post(struct radv_shader_context
*ctx
)
3641 write_tess_factors(ctx
);
3645 si_export_mrt_color(struct radv_shader_context
*ctx
,
3646 LLVMValueRef
*color
, unsigned index
,
3647 struct ac_export_args
*args
)
3650 si_llvm_init_export_args(ctx
, color
, 0xf,
3651 V_008DFC_SQ_EXP_MRT
+ index
, args
);
3652 if (!args
->enabled_channels
)
3653 return false; /* unnecessary NULL export */
3659 radv_export_mrt_z(struct radv_shader_context
*ctx
,
3660 LLVMValueRef depth
, LLVMValueRef stencil
,
3661 LLVMValueRef samplemask
)
3663 struct ac_export_args args
;
3665 ac_export_mrt_z(&ctx
->ac
, depth
, stencil
, samplemask
, &args
);
3667 ac_build_export(&ctx
->ac
, &args
);
3671 handle_fs_outputs_post(struct radv_shader_context
*ctx
)
3674 LLVMValueRef depth
= NULL
, stencil
= NULL
, samplemask
= NULL
;
3675 struct ac_export_args color_args
[8];
3677 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
3678 LLVMValueRef values
[4];
3680 if (!(ctx
->output_mask
& (1ull << i
)))
3683 if (i
< FRAG_RESULT_DATA0
)
3686 for (unsigned j
= 0; j
< 4; j
++)
3687 values
[j
] = ac_to_float(&ctx
->ac
,
3688 radv_load_output(ctx
, i
, j
));
3690 bool ret
= si_export_mrt_color(ctx
, values
,
3691 i
- FRAG_RESULT_DATA0
,
3692 &color_args
[index
]);
3697 /* Process depth, stencil, samplemask. */
3698 if (ctx
->args
->shader_info
->ps
.writes_z
) {
3699 depth
= ac_to_float(&ctx
->ac
,
3700 radv_load_output(ctx
, FRAG_RESULT_DEPTH
, 0));
3702 if (ctx
->args
->shader_info
->ps
.writes_stencil
) {
3703 stencil
= ac_to_float(&ctx
->ac
,
3704 radv_load_output(ctx
, FRAG_RESULT_STENCIL
, 0));
3706 if (ctx
->args
->shader_info
->ps
.writes_sample_mask
) {
3707 samplemask
= ac_to_float(&ctx
->ac
,
3708 radv_load_output(ctx
, FRAG_RESULT_SAMPLE_MASK
, 0));
3711 /* Set the DONE bit on last non-null color export only if Z isn't
3715 !ctx
->args
->shader_info
->ps
.writes_z
&&
3716 !ctx
->args
->shader_info
->ps
.writes_stencil
&&
3717 !ctx
->args
->shader_info
->ps
.writes_sample_mask
) {
3718 unsigned last
= index
- 1;
3720 color_args
[last
].valid_mask
= 1; /* whether the EXEC mask is valid */
3721 color_args
[last
].done
= 1; /* DONE bit */
3724 /* Export PS outputs. */
3725 for (unsigned i
= 0; i
< index
; i
++)
3726 ac_build_export(&ctx
->ac
, &color_args
[i
]);
3728 if (depth
|| stencil
|| samplemask
)
3729 radv_export_mrt_z(ctx
, depth
, stencil
, samplemask
);
3731 ac_build_export_null(&ctx
->ac
);
3735 emit_gs_epilogue(struct radv_shader_context
*ctx
)
3737 if (ctx
->args
->options
->key
.vs_common_out
.as_ngg
) {
3738 gfx10_ngg_gs_emit_epilogue_1(ctx
);
3742 if (ctx
->ac
.chip_class
>= GFX10
)
3743 LLVMBuildFence(ctx
->ac
.builder
, LLVMAtomicOrderingRelease
, false, "");
3745 ac_build_sendmsg(&ctx
->ac
, AC_SENDMSG_GS_OP_NOP
| AC_SENDMSG_GS_DONE
, ctx
->gs_wave_id
);
3749 handle_shader_outputs_post(struct ac_shader_abi
*abi
, unsigned max_outputs
,
3750 LLVMValueRef
*addrs
)
3752 struct radv_shader_context
*ctx
= radv_shader_context_from_abi(abi
);
3754 switch (ctx
->stage
) {
3755 case MESA_SHADER_VERTEX
:
3756 if (ctx
->args
->options
->key
.vs_common_out
.as_ls
)
3757 handle_ls_outputs_post(ctx
);
3758 else if (ctx
->args
->options
->key
.vs_common_out
.as_es
)
3759 handle_es_outputs_post(ctx
, &ctx
->args
->shader_info
->vs
.es_info
);
3760 else if (ctx
->args
->options
->key
.vs_common_out
.as_ngg
)
3761 handle_ngg_outputs_post_1(ctx
);
3763 handle_vs_outputs_post(ctx
, ctx
->args
->options
->key
.vs_common_out
.export_prim_id
,
3764 ctx
->args
->options
->key
.vs_common_out
.export_clip_dists
,
3765 &ctx
->args
->shader_info
->vs
.outinfo
);
3767 case MESA_SHADER_FRAGMENT
:
3768 handle_fs_outputs_post(ctx
);
3770 case MESA_SHADER_GEOMETRY
:
3771 emit_gs_epilogue(ctx
);
3773 case MESA_SHADER_TESS_CTRL
:
3774 handle_tcs_outputs_post(ctx
);
3776 case MESA_SHADER_TESS_EVAL
:
3777 if (ctx
->args
->options
->key
.vs_common_out
.as_es
)
3778 handle_es_outputs_post(ctx
, &ctx
->args
->shader_info
->tes
.es_info
);
3779 else if (ctx
->args
->options
->key
.vs_common_out
.as_ngg
)
3780 handle_ngg_outputs_post_1(ctx
);
3782 handle_vs_outputs_post(ctx
, ctx
->args
->options
->key
.vs_common_out
.export_prim_id
,
3783 ctx
->args
->options
->key
.vs_common_out
.export_clip_dists
,
3784 &ctx
->args
->shader_info
->tes
.outinfo
);
3791 static void ac_llvm_finalize_module(struct radv_shader_context
*ctx
,
3792 LLVMPassManagerRef passmgr
,
3793 const struct radv_nir_compiler_options
*options
)
3795 LLVMRunPassManager(passmgr
, ctx
->ac
.module
);
3796 LLVMDisposeBuilder(ctx
->ac
.builder
);
3798 ac_llvm_context_dispose(&ctx
->ac
);
3802 ac_nir_eliminate_const_vs_outputs(struct radv_shader_context
*ctx
)
3804 struct radv_vs_output_info
*outinfo
;
3806 switch (ctx
->stage
) {
3807 case MESA_SHADER_FRAGMENT
:
3808 case MESA_SHADER_COMPUTE
:
3809 case MESA_SHADER_TESS_CTRL
:
3810 case MESA_SHADER_GEOMETRY
:
3812 case MESA_SHADER_VERTEX
:
3813 if (ctx
->args
->options
->key
.vs_common_out
.as_ls
||
3814 ctx
->args
->options
->key
.vs_common_out
.as_es
)
3816 outinfo
= &ctx
->args
->shader_info
->vs
.outinfo
;
3818 case MESA_SHADER_TESS_EVAL
:
3819 if (ctx
->args
->options
->key
.vs_common_out
.as_es
)
3821 outinfo
= &ctx
->args
->shader_info
->tes
.outinfo
;
3824 unreachable("Unhandled shader type");
3827 ac_optimize_vs_outputs(&ctx
->ac
,
3829 outinfo
->vs_output_param_offset
,
3831 &outinfo
->param_exports
);
3835 ac_setup_rings(struct radv_shader_context
*ctx
)
3837 if (ctx
->args
->options
->chip_class
<= GFX8
&&
3838 (ctx
->stage
== MESA_SHADER_GEOMETRY
||
3839 ctx
->args
->options
->key
.vs_common_out
.as_es
|| ctx
->args
->options
->key
.vs_common_out
.as_es
)) {
3840 unsigned ring
= ctx
->stage
== MESA_SHADER_GEOMETRY
? RING_ESGS_GS
3842 LLVMValueRef offset
= LLVMConstInt(ctx
->ac
.i32
, ring
, false);
3844 ctx
->esgs_ring
= ac_build_load_to_sgpr(&ctx
->ac
,
3849 if (ctx
->args
->is_gs_copy_shader
) {
3851 ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
,
3852 LLVMConstInt(ctx
->ac
.i32
,
3853 RING_GSVS_VS
, false));
3856 if (ctx
->stage
== MESA_SHADER_GEOMETRY
) {
3857 /* The conceptual layout of the GSVS ring is
3858 * v0c0 .. vLv0 v0c1 .. vLc1 ..
3859 * but the real memory layout is swizzled across
3861 * t0v0c0 .. t15v0c0 t0v1c0 .. t15v1c0 ... t15vLcL
3863 * Override the buffer descriptor accordingly.
3865 LLVMTypeRef v2i64
= LLVMVectorType(ctx
->ac
.i64
, 2);
3866 uint64_t stream_offset
= 0;
3867 unsigned num_records
= ctx
->ac
.wave_size
;
3868 LLVMValueRef base_ring
;
3871 ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
,
3872 LLVMConstInt(ctx
->ac
.i32
,
3873 RING_GSVS_GS
, false));
3875 for (unsigned stream
= 0; stream
< 4; stream
++) {
3876 unsigned num_components
, stride
;
3877 LLVMValueRef ring
, tmp
;
3880 ctx
->args
->shader_info
->gs
.num_stream_output_components
[stream
];
3882 if (!num_components
)
3885 stride
= 4 * num_components
* ctx
->shader
->info
.gs
.vertices_out
;
3887 /* Limit on the stride field for <= GFX7. */
3888 assert(stride
< (1 << 14));
3890 ring
= LLVMBuildBitCast(ctx
->ac
.builder
,
3891 base_ring
, v2i64
, "");
3892 tmp
= LLVMBuildExtractElement(ctx
->ac
.builder
,
3893 ring
, ctx
->ac
.i32_0
, "");
3894 tmp
= LLVMBuildAdd(ctx
->ac
.builder
, tmp
,
3895 LLVMConstInt(ctx
->ac
.i64
,
3896 stream_offset
, 0), "");
3897 ring
= LLVMBuildInsertElement(ctx
->ac
.builder
,
3898 ring
, tmp
, ctx
->ac
.i32_0
, "");
3900 stream_offset
+= stride
* ctx
->ac
.wave_size
;
3902 ring
= LLVMBuildBitCast(ctx
->ac
.builder
, ring
,
3905 tmp
= LLVMBuildExtractElement(ctx
->ac
.builder
, ring
,
3907 tmp
= LLVMBuildOr(ctx
->ac
.builder
, tmp
,
3908 LLVMConstInt(ctx
->ac
.i32
,
3909 S_008F04_STRIDE(stride
), false), "");
3910 ring
= LLVMBuildInsertElement(ctx
->ac
.builder
, ring
, tmp
,
3913 ring
= LLVMBuildInsertElement(ctx
->ac
.builder
, ring
,
3914 LLVMConstInt(ctx
->ac
.i32
,
3915 num_records
, false),
3916 LLVMConstInt(ctx
->ac
.i32
, 2, false), "");
3918 ctx
->gsvs_ring
[stream
] = ring
;
3922 if (ctx
->stage
== MESA_SHADER_TESS_CTRL
||
3923 ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
3924 ctx
->hs_ring_tess_offchip
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_HS_TESS_OFFCHIP
, false));
3925 ctx
->hs_ring_tess_factor
= ac_build_load_to_sgpr(&ctx
->ac
, ctx
->ring_offsets
, LLVMConstInt(ctx
->ac
.i32
, RING_HS_TESS_FACTOR
, false));
3930 radv_nir_get_max_workgroup_size(enum chip_class chip_class
,
3931 gl_shader_stage stage
,
3932 const struct nir_shader
*nir
)
3934 const unsigned backup_sizes
[] = {chip_class
>= GFX9
? 128 : 64, 1, 1};
3936 for (unsigned i
= 0; i
< 3; i
++)
3937 sizes
[i
] = nir
? nir
->info
.cs
.local_size
[i
] : backup_sizes
[i
];
3938 return radv_get_max_workgroup_size(chip_class
, stage
, sizes
);
3941 /* Fixup the HW not emitting the TCS regs if there are no HS threads. */
3942 static void ac_nir_fixup_ls_hs_input_vgprs(struct radv_shader_context
*ctx
)
3944 LLVMValueRef count
=
3945 ac_unpack_param(&ctx
->ac
, ac_get_arg(&ctx
->ac
, ctx
->args
->merged_wave_info
), 8, 8);
3946 LLVMValueRef hs_empty
= LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntEQ
, count
,
3948 ctx
->abi
.instance_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
,
3949 ac_get_arg(&ctx
->ac
, ctx
->args
->rel_auto_id
),
3950 ctx
->abi
.instance_id
, "");
3951 ctx
->rel_auto_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
,
3952 ac_get_arg(&ctx
->ac
, ctx
->args
->ac
.tcs_rel_ids
),
3955 ctx
->abi
.vertex_id
= LLVMBuildSelect(ctx
->ac
.builder
, hs_empty
,
3956 ac_get_arg(&ctx
->ac
, ctx
->args
->ac
.tcs_patch_id
),
3957 ctx
->abi
.vertex_id
, "");
3960 static void prepare_gs_input_vgprs(struct radv_shader_context
*ctx
, bool merged
)
3963 for(int i
= 5; i
>= 0; --i
) {
3964 ctx
->gs_vtx_offset
[i
] =
3965 ac_unpack_param(&ctx
->ac
,
3966 ac_get_arg(&ctx
->ac
, ctx
->args
->gs_vtx_offset
[i
& ~1]),
3970 ctx
->gs_wave_id
= ac_unpack_param(&ctx
->ac
,
3971 ac_get_arg(&ctx
->ac
, ctx
->args
->merged_wave_info
),
3974 for (int i
= 0; i
< 6; i
++)
3975 ctx
->gs_vtx_offset
[i
] = ac_get_arg(&ctx
->ac
, ctx
->args
->gs_vtx_offset
[i
]);
3976 ctx
->gs_wave_id
= ac_get_arg(&ctx
->ac
, ctx
->args
->gs_wave_id
);
3980 /* Ensure that the esgs ring is declared.
3982 * We declare it with 64KB alignment as a hint that the
3983 * pointer value will always be 0.
3985 static void declare_esgs_ring(struct radv_shader_context
*ctx
)
3990 assert(!LLVMGetNamedGlobal(ctx
->ac
.module
, "esgs_ring"));
3992 ctx
->esgs_ring
= LLVMAddGlobalInAddressSpace(
3993 ctx
->ac
.module
, LLVMArrayType(ctx
->ac
.i32
, 0),
3996 LLVMSetLinkage(ctx
->esgs_ring
, LLVMExternalLinkage
);
3997 LLVMSetAlignment(ctx
->esgs_ring
, 64 * 1024);
4001 LLVMModuleRef
ac_translate_nir_to_llvm(struct ac_llvm_compiler
*ac_llvm
,
4002 struct nir_shader
*const *shaders
,
4004 const struct radv_shader_args
*args
)
4006 struct radv_shader_context ctx
= {0};
4009 enum ac_float_mode float_mode
= AC_FLOAT_MODE_DEFAULT
;
4011 if (args
->shader_info
->float_controls_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP32
) {
4012 float_mode
= AC_FLOAT_MODE_DENORM_FLUSH_TO_ZERO
;
4015 ac_llvm_context_init(&ctx
.ac
, ac_llvm
, args
->options
->chip_class
,
4016 args
->options
->family
, float_mode
,
4017 args
->shader_info
->wave_size
, 64);
4018 ctx
.context
= ctx
.ac
.context
;
4020 ctx
.max_workgroup_size
= 0;
4021 for (int i
= 0; i
< shader_count
; ++i
) {
4022 ctx
.max_workgroup_size
= MAX2(ctx
.max_workgroup_size
,
4023 radv_nir_get_max_workgroup_size(args
->options
->chip_class
,
4024 shaders
[i
]->info
.stage
,
4028 if (ctx
.ac
.chip_class
>= GFX10
) {
4029 if (is_pre_gs_stage(shaders
[0]->info
.stage
) &&
4030 args
->options
->key
.vs_common_out
.as_ngg
) {
4031 ctx
.max_workgroup_size
= 128;
4035 create_function(&ctx
, shaders
[shader_count
- 1]->info
.stage
, shader_count
>= 2);
4037 ctx
.abi
.inputs
= &ctx
.inputs
[0];
4038 ctx
.abi
.emit_outputs
= handle_shader_outputs_post
;
4039 ctx
.abi
.emit_vertex
= visit_emit_vertex
;
4040 ctx
.abi
.load_ubo
= radv_load_ubo
;
4041 ctx
.abi
.load_ssbo
= radv_load_ssbo
;
4042 ctx
.abi
.load_sampler_desc
= radv_get_sampler_desc
;
4043 ctx
.abi
.load_resource
= radv_load_resource
;
4044 ctx
.abi
.clamp_shadow_reference
= false;
4045 ctx
.abi
.robust_buffer_access
= args
->options
->robust_buffer_access
;
4047 bool is_ngg
= is_pre_gs_stage(shaders
[0]->info
.stage
) && args
->options
->key
.vs_common_out
.as_ngg
;
4048 if (shader_count
>= 2 || is_ngg
)
4049 ac_init_exec_full_mask(&ctx
.ac
);
4051 if (args
->ac
.vertex_id
.used
)
4052 ctx
.abi
.vertex_id
= ac_get_arg(&ctx
.ac
, args
->ac
.vertex_id
);
4053 if (args
->rel_auto_id
.used
)
4054 ctx
.rel_auto_id
= ac_get_arg(&ctx
.ac
, args
->rel_auto_id
);
4055 if (args
->ac
.instance_id
.used
)
4056 ctx
.abi
.instance_id
= ac_get_arg(&ctx
.ac
, args
->ac
.instance_id
);
4058 if (args
->options
->has_ls_vgpr_init_bug
&&
4059 shaders
[shader_count
- 1]->info
.stage
== MESA_SHADER_TESS_CTRL
)
4060 ac_nir_fixup_ls_hs_input_vgprs(&ctx
);
4063 /* Declare scratch space base for streamout and vertex
4064 * compaction. Whether space is actually allocated is
4065 * determined during linking / PM4 creation.
4067 * Add an extra dword per vertex to ensure an odd stride, which
4068 * avoids bank conflicts for SoA accesses.
4070 if (!args
->options
->key
.vs_common_out
.as_ngg_passthrough
)
4071 declare_esgs_ring(&ctx
);
4073 /* This is really only needed when streamout and / or vertex
4074 * compaction is enabled.
4076 if (args
->shader_info
->so
.num_outputs
) {
4077 LLVMTypeRef asi32
= LLVMArrayType(ctx
.ac
.i32
, 8);
4078 ctx
.gs_ngg_scratch
= LLVMAddGlobalInAddressSpace(ctx
.ac
.module
,
4079 asi32
, "ngg_scratch", AC_ADDR_SPACE_LDS
);
4080 LLVMSetInitializer(ctx
.gs_ngg_scratch
, LLVMGetUndef(asi32
));
4081 LLVMSetAlignment(ctx
.gs_ngg_scratch
, 4);
4085 for(int i
= 0; i
< shader_count
; ++i
) {
4086 ctx
.stage
= shaders
[i
]->info
.stage
;
4087 ctx
.shader
= shaders
[i
];
4088 ctx
.output_mask
= 0;
4090 if (shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
) {
4091 for (int i
= 0; i
< 4; i
++) {
4092 ctx
.gs_next_vertex
[i
] =
4093 ac_build_alloca(&ctx
.ac
, ctx
.ac
.i32
, "");
4095 if (args
->options
->key
.vs_common_out
.as_ngg
) {
4096 for (unsigned i
= 0; i
< 4; ++i
) {
4097 ctx
.gs_curprim_verts
[i
] =
4098 ac_build_alloca(&ctx
.ac
, ctx
.ac
.i32
, "");
4099 ctx
.gs_generated_prims
[i
] =
4100 ac_build_alloca(&ctx
.ac
, ctx
.ac
.i32
, "");
4103 unsigned scratch_size
= 8;
4104 if (args
->shader_info
->so
.num_outputs
)
4107 LLVMTypeRef ai32
= LLVMArrayType(ctx
.ac
.i32
, scratch_size
);
4108 ctx
.gs_ngg_scratch
=
4109 LLVMAddGlobalInAddressSpace(ctx
.ac
.module
,
4110 ai32
, "ngg_scratch", AC_ADDR_SPACE_LDS
);
4111 LLVMSetInitializer(ctx
.gs_ngg_scratch
, LLVMGetUndef(ai32
));
4112 LLVMSetAlignment(ctx
.gs_ngg_scratch
, 4);
4114 ctx
.gs_ngg_emit
= LLVMAddGlobalInAddressSpace(ctx
.ac
.module
,
4115 LLVMArrayType(ctx
.ac
.i32
, 0), "ngg_emit", AC_ADDR_SPACE_LDS
);
4116 LLVMSetLinkage(ctx
.gs_ngg_emit
, LLVMExternalLinkage
);
4117 LLVMSetAlignment(ctx
.gs_ngg_emit
, 4);
4120 ctx
.abi
.load_inputs
= load_gs_input
;
4121 ctx
.abi
.emit_primitive
= visit_end_primitive
;
4122 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_TESS_CTRL
) {
4123 ctx
.abi
.load_tess_varyings
= load_tcs_varyings
;
4124 ctx
.abi
.load_patch_vertices_in
= load_patch_vertices_in
;
4125 ctx
.abi
.store_tcs_outputs
= store_tcs_output
;
4126 if (shader_count
== 1)
4127 ctx
.tcs_num_inputs
= args
->options
->key
.tcs
.num_inputs
;
4129 ctx
.tcs_num_inputs
= util_last_bit64(args
->shader_info
->vs
.ls_outputs_written
);
4130 ctx
.tcs_num_patches
= get_tcs_num_patches(&ctx
);
4131 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_TESS_EVAL
) {
4132 ctx
.abi
.load_tess_varyings
= load_tes_input
;
4133 ctx
.abi
.load_tess_coord
= load_tess_coord
;
4134 ctx
.abi
.load_patch_vertices_in
= load_patch_vertices_in
;
4135 ctx
.tcs_num_patches
= args
->options
->key
.tes
.num_patches
;
4136 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_VERTEX
) {
4137 ctx
.abi
.load_base_vertex
= radv_load_base_vertex
;
4138 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_FRAGMENT
) {
4139 ctx
.abi
.load_sample_position
= load_sample_position
;
4140 ctx
.abi
.load_sample_mask_in
= load_sample_mask_in
;
4141 ctx
.abi
.emit_kill
= radv_emit_kill
;
4144 if (shaders
[i
]->info
.stage
== MESA_SHADER_VERTEX
&&
4145 args
->options
->key
.vs_common_out
.as_ngg
&&
4146 args
->options
->key
.vs_common_out
.export_prim_id
) {
4147 declare_esgs_ring(&ctx
);
4150 bool nested_barrier
= false;
4153 if (shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
&&
4154 args
->options
->key
.vs_common_out
.as_ngg
) {
4155 gfx10_ngg_gs_emit_prologue(&ctx
);
4156 nested_barrier
= false;
4158 nested_barrier
= true;
4162 if (nested_barrier
) {
4163 /* Execute a barrier before the second shader in
4166 * Execute the barrier inside the conditional block,
4167 * so that empty waves can jump directly to s_endpgm,
4168 * which will also signal the barrier.
4170 * This is possible in gfx9, because an empty wave
4171 * for the second shader does not participate in
4172 * the epilogue. With NGG, empty waves may still
4173 * be required to export data (e.g. GS output vertices),
4174 * so we cannot let them exit early.
4176 * If the shader is TCS and the TCS epilog is present
4177 * and contains a barrier, it will wait there and then
4180 ac_emit_barrier(&ctx
.ac
, ctx
.stage
);
4183 nir_foreach_variable(variable
, &shaders
[i
]->outputs
)
4184 scan_shader_output_decl(&ctx
, variable
, shaders
[i
], shaders
[i
]->info
.stage
);
4186 ac_setup_rings(&ctx
);
4188 LLVMBasicBlockRef merge_block
;
4189 if (shader_count
>= 2 || is_ngg
) {
4190 LLVMValueRef fn
= LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
.ac
.builder
));
4191 LLVMBasicBlockRef then_block
= LLVMAppendBasicBlockInContext(ctx
.ac
.context
, fn
, "");
4192 merge_block
= LLVMAppendBasicBlockInContext(ctx
.ac
.context
, fn
, "");
4194 LLVMValueRef count
=
4195 ac_unpack_param(&ctx
.ac
,
4196 ac_get_arg(&ctx
.ac
, args
->merged_wave_info
),
4198 LLVMValueRef thread_id
= ac_get_thread_id(&ctx
.ac
);
4199 LLVMValueRef cond
= LLVMBuildICmp(ctx
.ac
.builder
, LLVMIntULT
,
4200 thread_id
, count
, "");
4201 LLVMBuildCondBr(ctx
.ac
.builder
, cond
, then_block
, merge_block
);
4203 LLVMPositionBuilderAtEnd(ctx
.ac
.builder
, then_block
);
4206 if (shaders
[i
]->info
.stage
== MESA_SHADER_FRAGMENT
)
4207 prepare_interp_optimize(&ctx
, shaders
[i
]);
4208 else if(shaders
[i
]->info
.stage
== MESA_SHADER_VERTEX
)
4209 handle_vs_inputs(&ctx
, shaders
[i
]);
4210 else if(shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
)
4211 prepare_gs_input_vgprs(&ctx
, shader_count
>= 2);
4213 ac_nir_translate(&ctx
.ac
, &ctx
.abi
, &args
->ac
, shaders
[i
]);
4215 if (shader_count
>= 2 || is_ngg
) {
4216 LLVMBuildBr(ctx
.ac
.builder
, merge_block
);
4217 LLVMPositionBuilderAtEnd(ctx
.ac
.builder
, merge_block
);
4220 /* This needs to be outside the if wrapping the shader body, as sometimes
4221 * the HW generates waves with 0 es/vs threads. */
4222 if (is_pre_gs_stage(shaders
[i
]->info
.stage
) &&
4223 args
->options
->key
.vs_common_out
.as_ngg
&&
4224 i
== shader_count
- 1) {
4225 handle_ngg_outputs_post_2(&ctx
);
4226 } else if (shaders
[i
]->info
.stage
== MESA_SHADER_GEOMETRY
&&
4227 args
->options
->key
.vs_common_out
.as_ngg
) {
4228 gfx10_ngg_gs_emit_epilogue_2(&ctx
);
4231 if (shaders
[i
]->info
.stage
== MESA_SHADER_TESS_CTRL
) {
4232 args
->shader_info
->tcs
.num_patches
= ctx
.tcs_num_patches
;
4233 args
->shader_info
->tcs
.lds_size
= calculate_tess_lds_size(&ctx
);
4237 LLVMBuildRetVoid(ctx
.ac
.builder
);
4239 if (args
->options
->dump_preoptir
) {
4240 fprintf(stderr
, "%s LLVM IR:\n\n",
4241 radv_get_shader_name(args
->shader_info
,
4242 shaders
[shader_count
- 1]->info
.stage
));
4243 ac_dump_module(ctx
.ac
.module
);
4244 fprintf(stderr
, "\n");
4247 ac_llvm_finalize_module(&ctx
, ac_llvm
->passmgr
, args
->options
);
4249 if (shader_count
== 1)
4250 ac_nir_eliminate_const_vs_outputs(&ctx
);
4252 if (args
->options
->dump_shader
) {
4253 args
->shader_info
->private_mem_vgprs
=
4254 ac_count_scratch_private_memory(ctx
.main_function
);
4257 return ctx
.ac
.module
;
4260 static void ac_diagnostic_handler(LLVMDiagnosticInfoRef di
, void *context
)
4262 unsigned *retval
= (unsigned *)context
;
4263 LLVMDiagnosticSeverity severity
= LLVMGetDiagInfoSeverity(di
);
4264 char *description
= LLVMGetDiagInfoDescription(di
);
4266 if (severity
== LLVMDSError
) {
4268 fprintf(stderr
, "LLVM triggered Diagnostic Handler: %s\n",
4272 LLVMDisposeMessage(description
);
4275 static unsigned radv_llvm_compile(LLVMModuleRef M
,
4276 char **pelf_buffer
, size_t *pelf_size
,
4277 struct ac_llvm_compiler
*ac_llvm
)
4279 unsigned retval
= 0;
4280 LLVMContextRef llvm_ctx
;
4282 /* Setup Diagnostic Handler*/
4283 llvm_ctx
= LLVMGetModuleContext(M
);
4285 LLVMContextSetDiagnosticHandler(llvm_ctx
, ac_diagnostic_handler
,
4289 if (!radv_compile_to_elf(ac_llvm
, M
, pelf_buffer
, pelf_size
))
4294 static void ac_compile_llvm_module(struct ac_llvm_compiler
*ac_llvm
,
4295 LLVMModuleRef llvm_module
,
4296 struct radv_shader_binary
**rbinary
,
4297 gl_shader_stage stage
,
4299 const struct radv_nir_compiler_options
*options
)
4301 char *elf_buffer
= NULL
;
4302 size_t elf_size
= 0;
4303 char *llvm_ir_string
= NULL
;
4305 if (options
->dump_shader
) {
4306 fprintf(stderr
, "%s LLVM IR:\n\n", name
);
4307 ac_dump_module(llvm_module
);
4308 fprintf(stderr
, "\n");
4311 if (options
->record_ir
) {
4312 char *llvm_ir
= LLVMPrintModuleToString(llvm_module
);
4313 llvm_ir_string
= strdup(llvm_ir
);
4314 LLVMDisposeMessage(llvm_ir
);
4317 int v
= radv_llvm_compile(llvm_module
, &elf_buffer
, &elf_size
, ac_llvm
);
4319 fprintf(stderr
, "compile failed\n");
4322 LLVMContextRef ctx
= LLVMGetModuleContext(llvm_module
);
4323 LLVMDisposeModule(llvm_module
);
4324 LLVMContextDispose(ctx
);
4326 size_t llvm_ir_size
= llvm_ir_string
? strlen(llvm_ir_string
) : 0;
4327 size_t alloc_size
= sizeof(struct radv_shader_binary_rtld
) + elf_size
+ llvm_ir_size
+ 1;
4328 struct radv_shader_binary_rtld
*rbin
= calloc(1, alloc_size
);
4329 memcpy(rbin
->data
, elf_buffer
, elf_size
);
4331 memcpy(rbin
->data
+ elf_size
, llvm_ir_string
, llvm_ir_size
+ 1);
4333 rbin
->base
.type
= RADV_BINARY_TYPE_RTLD
;
4334 rbin
->base
.stage
= stage
;
4335 rbin
->base
.total_size
= alloc_size
;
4336 rbin
->elf_size
= elf_size
;
4337 rbin
->llvm_ir_size
= llvm_ir_size
;
4338 *rbinary
= &rbin
->base
;
4340 free(llvm_ir_string
);
4345 radv_compile_nir_shader(struct ac_llvm_compiler
*ac_llvm
,
4346 struct radv_shader_binary
**rbinary
,
4347 const struct radv_shader_args
*args
,
4348 struct nir_shader
*const *nir
,
4352 LLVMModuleRef llvm_module
;
4354 llvm_module
= ac_translate_nir_to_llvm(ac_llvm
, nir
, nir_count
, args
);
4356 ac_compile_llvm_module(ac_llvm
, llvm_module
, rbinary
,
4357 nir
[nir_count
- 1]->info
.stage
,
4358 radv_get_shader_name(args
->shader_info
,
4359 nir
[nir_count
- 1]->info
.stage
),
4362 /* Determine the ES type (VS or TES) for the GS on GFX9. */
4363 if (args
->options
->chip_class
>= GFX9
) {
4364 if (nir_count
== 2 &&
4365 nir
[1]->info
.stage
== MESA_SHADER_GEOMETRY
) {
4366 args
->shader_info
->gs
.es_type
= nir
[0]->info
.stage
;
4372 ac_gs_copy_shader_emit(struct radv_shader_context
*ctx
)
4374 LLVMValueRef vtx_offset
=
4375 LLVMBuildMul(ctx
->ac
.builder
, ac_get_arg(&ctx
->ac
, ctx
->args
->ac
.vertex_id
),
4376 LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
4377 LLVMValueRef stream_id
;
4379 /* Fetch the vertex stream ID. */
4380 if (!ctx
->args
->options
->use_ngg_streamout
&&
4381 ctx
->args
->shader_info
->so
.num_outputs
) {
4383 ac_unpack_param(&ctx
->ac
,
4384 ac_get_arg(&ctx
->ac
,
4385 ctx
->args
->streamout_config
),
4388 stream_id
= ctx
->ac
.i32_0
;
4391 LLVMBasicBlockRef end_bb
;
4392 LLVMValueRef switch_inst
;
4394 end_bb
= LLVMAppendBasicBlockInContext(ctx
->ac
.context
,
4395 ctx
->main_function
, "end");
4396 switch_inst
= LLVMBuildSwitch(ctx
->ac
.builder
, stream_id
, end_bb
, 4);
4398 for (unsigned stream
= 0; stream
< 4; stream
++) {
4399 unsigned num_components
=
4400 ctx
->args
->shader_info
->gs
.num_stream_output_components
[stream
];
4401 LLVMBasicBlockRef bb
;
4404 if (stream
> 0 && !num_components
)
4407 if (stream
> 0 && !ctx
->args
->shader_info
->so
.num_outputs
)
4410 bb
= LLVMInsertBasicBlockInContext(ctx
->ac
.context
, end_bb
, "out");
4411 LLVMAddCase(switch_inst
, LLVMConstInt(ctx
->ac
.i32
, stream
, 0), bb
);
4412 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, bb
);
4415 for (unsigned i
= 0; i
< AC_LLVM_MAX_OUTPUTS
; ++i
) {
4416 unsigned output_usage_mask
=
4417 ctx
->args
->shader_info
->gs
.output_usage_mask
[i
];
4418 unsigned output_stream
=
4419 ctx
->args
->shader_info
->gs
.output_streams
[i
];
4420 int length
= util_last_bit(output_usage_mask
);
4422 if (!(ctx
->output_mask
& (1ull << i
)) ||
4423 output_stream
!= stream
)
4426 for (unsigned j
= 0; j
< length
; j
++) {
4427 LLVMValueRef value
, soffset
;
4429 if (!(output_usage_mask
& (1 << j
)))
4432 soffset
= LLVMConstInt(ctx
->ac
.i32
,
4434 ctx
->shader
->info
.gs
.vertices_out
* 16 * 4, false);
4438 value
= ac_build_buffer_load(&ctx
->ac
,
4441 vtx_offset
, soffset
,
4442 0, ac_glc
| ac_slc
, true, false);
4444 LLVMTypeRef type
= LLVMGetAllocatedType(ctx
->abi
.outputs
[ac_llvm_reg_index_soa(i
, j
)]);
4445 if (ac_get_type_size(type
) == 2) {
4446 value
= LLVMBuildBitCast(ctx
->ac
.builder
, value
, ctx
->ac
.i32
, "");
4447 value
= LLVMBuildTrunc(ctx
->ac
.builder
, value
, ctx
->ac
.i16
, "");
4450 LLVMBuildStore(ctx
->ac
.builder
,
4451 ac_to_float(&ctx
->ac
, value
), ctx
->abi
.outputs
[ac_llvm_reg_index_soa(i
, j
)]);
4455 if (!ctx
->args
->options
->use_ngg_streamout
&&
4456 ctx
->args
->shader_info
->so
.num_outputs
)
4457 radv_emit_streamout(ctx
, stream
);
4460 handle_vs_outputs_post(ctx
, false, true,
4461 &ctx
->args
->shader_info
->vs
.outinfo
);
4464 LLVMBuildBr(ctx
->ac
.builder
, end_bb
);
4467 LLVMPositionBuilderAtEnd(ctx
->ac
.builder
, end_bb
);
4471 radv_compile_gs_copy_shader(struct ac_llvm_compiler
*ac_llvm
,
4472 struct nir_shader
*geom_shader
,
4473 struct radv_shader_binary
**rbinary
,
4474 const struct radv_shader_args
*args
)
4476 struct radv_shader_context ctx
= {0};
4479 assert(args
->is_gs_copy_shader
);
4481 ac_llvm_context_init(&ctx
.ac
, ac_llvm
, args
->options
->chip_class
,
4482 args
->options
->family
, AC_FLOAT_MODE_DEFAULT
, 64, 64);
4483 ctx
.context
= ctx
.ac
.context
;
4485 ctx
.stage
= MESA_SHADER_VERTEX
;
4486 ctx
.shader
= geom_shader
;
4488 create_function(&ctx
, MESA_SHADER_VERTEX
, false);
4490 ac_setup_rings(&ctx
);
4492 nir_foreach_variable(variable
, &geom_shader
->outputs
) {
4493 scan_shader_output_decl(&ctx
, variable
, geom_shader
, MESA_SHADER_VERTEX
);
4494 ac_handle_shader_output_decl(&ctx
.ac
, &ctx
.abi
, geom_shader
,
4495 variable
, MESA_SHADER_VERTEX
);
4498 ac_gs_copy_shader_emit(&ctx
);
4500 LLVMBuildRetVoid(ctx
.ac
.builder
);
4502 ac_llvm_finalize_module(&ctx
, ac_llvm
->passmgr
, args
->options
);
4504 ac_compile_llvm_module(ac_llvm
, ctx
.ac
.module
, rbinary
,
4505 MESA_SHADER_VERTEX
, "GS Copy Shader", args
->options
);
4506 (*rbinary
)->is_gs_copy_shader
= true;