1 /*
2 * Copyright 2016 Advanced Micro Devices, Inc.
3 * All Rights Reserved.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * on the rights to use, copy, modify, merge, publish, distribute, sub
9 * license, and/or sell copies of the Software, and to permit persons to whom
10 * the Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22 * USE OR OTHER DEALINGS IN THE SOFTWARE.
23 */
24
25 #ifndef SI_SHADER_PRIVATE_H
26 #define SI_SHADER_PRIVATE_H
27
28 #include "ac_shader_abi.h"
29 #include "si_shader.h"
30
31 struct pipe_debug_callback;
32
33 #define RADEON_LLVM_MAX_INPUTS 32 * 4
34
35 /* Ideally pass the sample mask input to the PS epilog as v14, which
36 * is its usual location, so that the shader doesn't have to add v_mov.
37 */
38 #define PS_EPILOG_SAMPLEMASK_MIN_LOC 14
39
40 struct si_shader_output_values {
41 LLVMValueRef values[4];
42 unsigned semantic_name;
43 unsigned semantic_index;
44 ubyte vertex_stream[4];
45 };
46
47 struct si_shader_context {
48 struct ac_llvm_context ac;
49 struct si_shader *shader;
50 struct si_screen *screen;
51
52 unsigned type; /* PIPE_SHADER_* specifies the type of shader. */
53
54 /* For clamping the non-constant index in resource indexing: */
55 unsigned num_const_buffers;
56 unsigned num_shader_buffers;
57 unsigned num_images;
58 unsigned num_samplers;
59
60 struct ac_shader_args args;
61 struct ac_shader_abi abi;
62
63 LLVMValueRef inputs[RADEON_LLVM_MAX_INPUTS];
64
65 LLVMBasicBlockRef merged_wrap_if_entry_block;
66 int merged_wrap_if_label;
67
68 LLVMValueRef main_fn;
69 LLVMTypeRef return_type;
70
71 struct ac_arg const_and_shader_buffers;
72 struct ac_arg samplers_and_images;
73
74 /* For merged shaders, the per-stage descriptors for the stage other
75 * than the one we're processing, used to pass them through from the
76 * first stage to the second.
77 */
78 struct ac_arg other_const_and_shader_buffers;
79 struct ac_arg other_samplers_and_images;
80
81 struct ac_arg rw_buffers;
82 struct ac_arg bindless_samplers_and_images;
83 /* Common inputs for merged shaders. */
84 struct ac_arg merged_wave_info;
85 struct ac_arg merged_scratch_offset;
86 struct ac_arg small_prim_cull_info;
87 /* API VS */
88 struct ac_arg vertex_buffers;
89 struct ac_arg vb_descriptors[5];
90 struct ac_arg rel_auto_id;
91 struct ac_arg vs_prim_id;
92 struct ac_arg vertex_index0;
93 /* VS states and layout of LS outputs / TCS inputs at the end
94 * [0] = clamp vertex color
95 * [1] = indexed
96 * [2:3] = NGG: output primitive type
97 * [4:5] = NGG: provoking vertex index
98 * [6] = NGG: streamout queries enabled
99 * [7:10] = NGG: small prim filter precision = num_samples / quant_mode,
100 * but in reality it's: 1/2^n, from 1/16 to 1/4096 = 1/2^4 to 1/2^12
101 * Only the first 4 bits of the exponent are stored.
102 * Set it like this: (fui(num_samples / quant_mode) >> 23)
103 * Expand to FP32 like this: ((0x70 | value) << 23);
104 * With 0x70 = 112, we get 2^(112 + value - 127) = 2^(value - 15)
105 * = 1/2^(15 - value) in FP32
106 * [11:23] = stride between patches in DW = num_inputs * num_vertices * 4
107 * max = 32*32*4 + 32*4
108 * [24:31] = stride between vertices in DW = num_inputs * 4
109 * max = 32*4
110 */
111 struct ac_arg vs_state_bits;
112 struct ac_arg vs_blit_inputs;
113 struct ac_arg ngg_old_thread_id; /* generated by the NGG cull shader */
114 /* HW VS */
115 struct ac_arg streamout_config;
116 struct ac_arg streamout_write_index;
117 struct ac_arg streamout_offset[4];
118
119 /* API TCS & TES */
120 /* Layout of TCS outputs in the offchip buffer
121 * # 6 bits
122 * [0:5] = the number of patches per threadgroup, max = NUM_PATCHES (40)
123 * # 6 bits
124 * [6:11] = the number of output vertices per patch, max = 32
125 * # 20 bits
126 * [12:31] = the offset of per patch attributes in the buffer in bytes.
127 * max = NUM_PATCHES*32*32*16
128 */
129 struct ac_arg tcs_offchip_layout;
130
131 /* API TCS */
132 /* Offsets where TCS outputs and TCS patch outputs live in LDS:
133 * [0:15] = TCS output patch0 offset / 16, max = NUM_PATCHES * 32 * 32
134 * [16:31] = TCS output patch0 offset for per-patch / 16
135 * max = (NUM_PATCHES + 1) * 32*32
136 */
137 struct ac_arg tcs_out_lds_offsets;
138 /* Layout of TCS outputs / TES inputs:
139 * [0:12] = stride between output patches in DW, num_outputs * num_vertices * 4
140 * max = 32*32*4 + 32*4
141 * [13:18] = gl_PatchVerticesIn, max = 32
142 * [19:31] = high 13 bits of the 32-bit address of tessellation ring buffers
143 */
144 struct ac_arg tcs_out_lds_layout;
145 struct ac_arg tcs_offchip_offset;
146 struct ac_arg tcs_factor_offset;
147
148 /* API TES */
149 struct ac_arg tes_offchip_addr;
150 struct ac_arg tes_u;
151 struct ac_arg tes_v;
152 struct ac_arg tes_rel_patch_id;
153 /* HW ES */
154 struct ac_arg es2gs_offset;
155 /* HW GS */
156 /* On gfx10:
157 * - bits 0..11: ordered_wave_id
158 * - bits 12..20: number of vertices in group
159 * - bits 22..30: number of primitives in group
160 */
161 struct ac_arg gs_tg_info;
162 /* API GS */
163 struct ac_arg gs2vs_offset;
164 struct ac_arg gs_wave_id; /* GFX6 */
165 struct ac_arg gs_vtx_offset[6]; /* in dwords (GFX6) */
166 struct ac_arg gs_vtx01_offset; /* in dwords (GFX9) */
167 struct ac_arg gs_vtx23_offset; /* in dwords (GFX9) */
168 struct ac_arg gs_vtx45_offset; /* in dwords (GFX9) */
169 /* PS */
170 struct ac_arg pos_fixed_pt;
171 /* CS */
172 struct ac_arg block_size;
173 struct ac_arg cs_user_data;
174 struct ac_arg cs_shaderbuf[3];
175 struct ac_arg cs_image[3];
176
177 struct ac_llvm_compiler *compiler;
178
179 /* Preloaded descriptors. */
180 LLVMValueRef esgs_ring;
181 LLVMValueRef gsvs_ring[4];
182 LLVMValueRef tess_offchip_ring;
183
184 LLVMValueRef invoc0_tess_factors[6]; /* outer[4], inner[2] */
185 LLVMValueRef gs_next_vertex[4];
186 LLVMValueRef gs_curprim_verts[4];
187 LLVMValueRef gs_generated_prims[4];
188 LLVMValueRef gs_ngg_emit;
189 LLVMValueRef gs_ngg_scratch;
190 LLVMValueRef postponed_kill;
191 LLVMValueRef return_value;
192 };
193
si_shader_context_from_abi(struct ac_shader_abi * abi)194 static inline struct si_shader_context *si_shader_context_from_abi(struct ac_shader_abi *abi)
195 {
196 struct si_shader_context *ctx = NULL;
197 return container_of(abi, ctx, abi);
198 }
199
200 bool si_is_multi_part_shader(struct si_shader *shader);
201 bool si_is_merged_shader(struct si_shader *shader);
202 void si_add_arg_checked(struct ac_shader_args *args, enum ac_arg_regfile file, unsigned registers,
203 enum ac_arg_type type, struct ac_arg *arg, unsigned idx);
204 unsigned si_get_max_workgroup_size(const struct si_shader *shader);
205 bool si_need_ps_prolog(const union si_shader_part_key *key);
206 void si_get_ps_prolog_key(struct si_shader *shader, union si_shader_part_key *key,
207 bool separate_prolog);
208 void si_get_ps_epilog_key(struct si_shader *shader, union si_shader_part_key *key);
209 void si_fix_resource_usage(struct si_screen *sscreen, struct si_shader *shader);
210 void si_create_function(struct si_shader_context *ctx, bool ngg_cull_shader);
211
212 bool gfx10_ngg_export_prim_early(struct si_shader *shader);
213 void gfx10_ngg_build_sendmsg_gs_alloc_req(struct si_shader_context *ctx);
214 void gfx10_ngg_build_export_prim(struct si_shader_context *ctx, LLVMValueRef user_edgeflags[3],
215 LLVMValueRef prim_passthrough);
216 void gfx10_emit_ngg_culling_epilogue(struct ac_shader_abi *abi, unsigned max_outputs,
217 LLVMValueRef *addrs);
218 void gfx10_emit_ngg_epilogue(struct ac_shader_abi *abi, unsigned max_outputs, LLVMValueRef *addrs);
219 void gfx10_ngg_gs_emit_vertex(struct si_shader_context *ctx, unsigned stream, LLVMValueRef *addrs);
220 void gfx10_ngg_gs_emit_prologue(struct si_shader_context *ctx);
221 void gfx10_ngg_gs_emit_epilogue(struct si_shader_context *ctx);
222 unsigned gfx10_ngg_get_scratch_dw_size(struct si_shader *shader);
223 bool gfx10_ngg_calculate_subgroup_info(struct si_shader *shader);
224
225 /* si_shader_llvm.c */
226 bool si_compile_llvm(struct si_screen *sscreen, struct si_shader_binary *binary,
227 struct ac_shader_config *conf, struct ac_llvm_compiler *compiler,
228 struct ac_llvm_context *ac, struct pipe_debug_callback *debug,
229 enum pipe_shader_type shader_type, const char *name, bool less_optimized);
230 void si_llvm_context_init(struct si_shader_context *ctx, struct si_screen *sscreen,
231 struct ac_llvm_compiler *compiler, unsigned wave_size);
232 void si_llvm_create_func(struct si_shader_context *ctx, const char *name, LLVMTypeRef *return_types,
233 unsigned num_return_elems, unsigned max_workgroup_size);
234 void si_llvm_optimize_module(struct si_shader_context *ctx);
235 void si_llvm_dispose(struct si_shader_context *ctx);
236 LLVMValueRef si_buffer_load_const(struct si_shader_context *ctx, LLVMValueRef resource,
237 LLVMValueRef offset);
238 void si_llvm_build_ret(struct si_shader_context *ctx, LLVMValueRef ret);
239 LLVMValueRef si_insert_input_ret(struct si_shader_context *ctx, LLVMValueRef ret,
240 struct ac_arg param, unsigned return_index);
241 LLVMValueRef si_insert_input_ret_float(struct si_shader_context *ctx, LLVMValueRef ret,
242 struct ac_arg param, unsigned return_index);
243 LLVMValueRef si_insert_input_ptr(struct si_shader_context *ctx, LLVMValueRef ret,
244 struct ac_arg param, unsigned return_index);
245 LLVMValueRef si_prolog_get_rw_buffers(struct si_shader_context *ctx);
246 LLVMValueRef si_build_gather_64bit(struct si_shader_context *ctx, LLVMTypeRef type,
247 LLVMValueRef val1, LLVMValueRef val2);
248 void si_llvm_emit_barrier(struct si_shader_context *ctx);
249 void si_llvm_declare_esgs_ring(struct si_shader_context *ctx);
250 void si_init_exec_from_input(struct si_shader_context *ctx, struct ac_arg param,
251 unsigned bitoffset);
252 LLVMValueRef si_unpack_param(struct si_shader_context *ctx, struct ac_arg param, unsigned rshift,
253 unsigned bitwidth);
254 LLVMValueRef si_get_primitive_id(struct si_shader_context *ctx, unsigned swizzle);
255 LLVMValueRef si_llvm_get_block_size(struct ac_shader_abi *abi);
256 void si_llvm_declare_compute_memory(struct si_shader_context *ctx);
257 bool si_nir_build_llvm(struct si_shader_context *ctx, struct nir_shader *nir);
258 void si_build_wrapper_function(struct si_shader_context *ctx, LLVMValueRef *parts,
259 unsigned num_parts, unsigned main_part,
260 unsigned next_shader_first_part);
261
262 /* si_shader_llvm_gs.c */
263 LLVMValueRef si_is_es_thread(struct si_shader_context *ctx);
264 LLVMValueRef si_is_gs_thread(struct si_shader_context *ctx);
265 void si_llvm_emit_es_epilogue(struct ac_shader_abi *abi, unsigned max_outputs, LLVMValueRef *addrs);
266 void si_preload_esgs_ring(struct si_shader_context *ctx);
267 void si_preload_gs_rings(struct si_shader_context *ctx);
268 void si_llvm_build_gs_prolog(struct si_shader_context *ctx, union si_shader_part_key *key);
269 void si_llvm_init_gs_callbacks(struct si_shader_context *ctx);
270
271 /* si_shader_llvm_tess.c */
272 void si_llvm_preload_tes_rings(struct si_shader_context *ctx);
273 void si_llvm_emit_ls_epilogue(struct ac_shader_abi *abi, unsigned max_outputs, LLVMValueRef *addrs);
274 void si_llvm_build_tcs_epilog(struct si_shader_context *ctx, union si_shader_part_key *key);
275 void si_llvm_init_tcs_callbacks(struct si_shader_context *ctx);
276 void si_llvm_init_tes_callbacks(struct si_shader_context *ctx, bool ngg_cull_shader);
277
278 /* si_shader_llvm_ps.c */
279 LLVMValueRef si_get_sample_id(struct si_shader_context *ctx);
280 void si_llvm_build_ps_prolog(struct si_shader_context *ctx, union si_shader_part_key *key);
281 void si_llvm_build_ps_epilog(struct si_shader_context *ctx, union si_shader_part_key *key);
282 void si_llvm_build_monolithic_ps(struct si_shader_context *ctx, struct si_shader *shader);
283 void si_llvm_init_ps_callbacks(struct si_shader_context *ctx);
284
285 /* si_shader_llvm_resources.c */
286 void si_llvm_init_resource_callbacks(struct si_shader_context *ctx);
287
288 /* si_shader_llvm_vs.c */
289 void si_llvm_load_vs_inputs(struct si_shader_context *ctx, struct nir_shader *nir);
290 void si_llvm_streamout_store_output(struct si_shader_context *ctx, LLVMValueRef const *so_buffers,
291 LLVMValueRef const *so_write_offsets,
292 struct pipe_stream_output *stream_out,
293 struct si_shader_output_values *shader_out);
294 void si_llvm_emit_streamout(struct si_shader_context *ctx, struct si_shader_output_values *outputs,
295 unsigned noutput, unsigned stream);
296 void si_llvm_build_vs_exports(struct si_shader_context *ctx,
297 struct si_shader_output_values *outputs, unsigned noutput);
298 void si_llvm_emit_vs_epilogue(struct ac_shader_abi *abi, unsigned max_outputs, LLVMValueRef *addrs);
299 void si_llvm_build_vs_prolog(struct si_shader_context *ctx, union si_shader_part_key *key);
300 void si_llvm_init_vs_callbacks(struct si_shader_context *ctx, bool ngg_cull_shader);
301
302 #endif
303