Christian König | ce40e47 | 2012-08-02 12:14:59 +0200 | [diff] [blame] | 1 | /* |
| 2 | * Copyright 2012 Advanced Micro Devices, Inc. |
| 3 | * |
| 4 | * Permission is hereby granted, free of charge, to any person obtaining a |
| 5 | * copy of this software and associated documentation files (the "Software"), |
| 6 | * to deal in the Software without restriction, including without limitation |
| 7 | * on the rights to use, copy, modify, merge, publish, distribute, sub |
| 8 | * license, and/or sell copies of the Software, and to permit persons to whom |
| 9 | * the Software is furnished to do so, subject to the following conditions: |
| 10 | * |
| 11 | * The above copyright notice and this permission notice (including the next |
| 12 | * paragraph) shall be included in all copies or substantial portions of the |
| 13 | * Software. |
| 14 | * |
| 15 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 16 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 17 | * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL |
| 18 | * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM, |
| 19 | * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR |
| 20 | * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE |
| 21 | * USE OR OTHER DEALINGS IN THE SOFTWARE. |
| 22 | * |
| 23 | * Authors: |
| 24 | * Tom Stellard <thomas.stellard@amd.com> |
| 25 | * Michel Dänzer <michel.daenzer@amd.com> |
| 26 | * Christian König <christian.koenig@amd.com> |
| 27 | */ |
| 28 | |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 29 | #include "gallivm/lp_bld_const.h" |
Michel Dänzer | c2bae6b | 2012-08-02 17:19:22 +0200 | [diff] [blame] | 30 | #include "gallivm/lp_bld_gather.h" |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 31 | #include "gallivm/lp_bld_intr.h" |
Michel Dänzer | 7708a86 | 2012-11-02 15:57:30 +0100 | [diff] [blame] | 32 | #include "gallivm/lp_bld_logic.h" |
Christian König | 5e616cf | 2013-03-07 11:58:56 +0100 | [diff] [blame] | 33 | #include "gallivm/lp_bld_arit.h" |
Marek Olšák | 8d03d92 | 2013-09-01 23:59:06 +0200 | [diff] [blame] | 34 | #include "gallivm/lp_bld_flow.h" |
Tom Stellard | 2397a72 | 2014-12-10 09:13:59 -0500 | [diff] [blame] | 35 | #include "radeon/r600_cs.h" |
Emil Velikov | a131263 | 2014-08-16 17:58:25 +0100 | [diff] [blame] | 36 | #include "radeon/radeon_llvm.h" |
Tom Stellard | 1f4e48d | 2014-09-25 18:11:24 -0700 | [diff] [blame] | 37 | #include "radeon/radeon_elf_util.h" |
Emil Velikov | a131263 | 2014-08-16 17:58:25 +0100 | [diff] [blame] | 38 | #include "radeon/radeon_llvm_emit.h" |
Christian König | 0f6cf2b | 2013-03-15 15:53:25 +0100 | [diff] [blame] | 39 | #include "util/u_memory.h" |
Marek Olšák | 6c5af1d | 2015-01-31 21:43:37 +0100 | [diff] [blame] | 40 | #include "util/u_pstipple.h" |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 41 | #include "tgsi/tgsi_parse.h" |
Michel Dänzer | beaa5eb | 2013-05-24 13:23:26 +0100 | [diff] [blame] | 42 | #include "tgsi/tgsi_util.h" |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 43 | #include "tgsi/tgsi_dump.h" |
| 44 | |
Andreas Hartmetz | 786af2f | 2014-01-04 18:44:33 +0100 | [diff] [blame] | 45 | #include "si_pipe.h" |
| 46 | #include "si_shader.h" |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 47 | #include "sid.h" |
| 48 | |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 49 | #include <errno.h> |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 50 | |
Tom Stellard | 2397a72 | 2014-12-10 09:13:59 -0500 | [diff] [blame] | 51 | static const char *scratch_rsrc_dword0_symbol = |
| 52 | "SCRATCH_RSRC_DWORD0"; |
| 53 | |
| 54 | static const char *scratch_rsrc_dword1_symbol = |
| 55 | "SCRATCH_RSRC_DWORD1"; |
| 56 | |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 57 | struct si_shader_output_values |
| 58 | { |
| 59 | LLVMValueRef values[4]; |
| 60 | unsigned name; |
Michel Dänzer | 67e385b | 2014-01-08 17:48:21 +0900 | [diff] [blame] | 61 | unsigned sid; |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 62 | }; |
| 63 | |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 64 | struct si_shader_context |
| 65 | { |
| 66 | struct radeon_llvm_context radeon_bld; |
Marek Olšák | 8c37c16 | 2014-09-16 18:40:07 +0200 | [diff] [blame] | 67 | struct si_shader *shader; |
Marek Olšák | 368b0a7 | 2014-12-28 21:51:35 +0100 | [diff] [blame] | 68 | struct si_screen *screen; |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 69 | unsigned type; /* TGSI_PROCESSOR_* specifies the type of shader. */ |
Marek Olšák | 8d03d92 | 2013-09-01 23:59:06 +0200 | [diff] [blame] | 70 | int param_streamout_config; |
| 71 | int param_streamout_write_index; |
| 72 | int param_streamout_offset[4]; |
| 73 | int param_vertex_id; |
| 74 | int param_instance_id; |
Christian König | 206f059 | 2013-03-20 14:37:21 +0100 | [diff] [blame] | 75 | LLVMValueRef const_md; |
Marek Olšák | ee2a818 | 2014-07-07 23:27:19 +0200 | [diff] [blame] | 76 | LLVMValueRef const_resource[SI_NUM_CONST_BUFFERS]; |
Michel Dänzer | a06ee5a | 2013-06-19 18:14:01 +0200 | [diff] [blame] | 77 | LLVMValueRef ddxy_lds; |
Marek Olšák | ee2a818 | 2014-07-07 23:27:19 +0200 | [diff] [blame] | 78 | LLVMValueRef *constants[SI_NUM_CONST_BUFFERS]; |
Marek Olšák | 55a9b77 | 2014-10-05 12:38:54 +0200 | [diff] [blame] | 79 | LLVMValueRef resources[SI_NUM_SAMPLER_VIEWS]; |
| 80 | LLVMValueRef samplers[SI_NUM_SAMPLER_STATES]; |
Marek Olšák | 8d03d92 | 2013-09-01 23:59:06 +0200 | [diff] [blame] | 81 | LLVMValueRef so_buffers[4]; |
Marek Olšák | fc3b335 | 2014-10-05 13:33:40 +0200 | [diff] [blame] | 82 | LLVMValueRef esgs_ring; |
| 83 | LLVMValueRef gsvs_ring; |
Michel Dänzer | f07a96d | 2014-01-08 18:45:10 +0900 | [diff] [blame] | 84 | LLVMValueRef gs_next_vertex; |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 85 | }; |
| 86 | |
| 87 | static struct si_shader_context * si_shader_context( |
| 88 | struct lp_build_tgsi_context * bld_base) |
| 89 | { |
| 90 | return (struct si_shader_context *)bld_base; |
| 91 | } |
| 92 | |
| 93 | |
| 94 | #define PERSPECTIVE_BASE 0 |
| 95 | #define LINEAR_BASE 9 |
| 96 | |
| 97 | #define SAMPLE_OFFSET 0 |
| 98 | #define CENTER_OFFSET 2 |
| 99 | #define CENTROID_OFSET 4 |
| 100 | |
| 101 | #define USE_SGPR_MAX_SUFFIX_LEN 5 |
Tom Stellard | 467f516 | 2012-05-16 15:15:35 -0400 | [diff] [blame] | 102 | #define CONST_ADDR_SPACE 2 |
Michel Dänzer | a06ee5a | 2013-06-19 18:14:01 +0200 | [diff] [blame] | 103 | #define LOCAL_ADDR_SPACE 3 |
Tom Stellard | 89ece08 | 2012-05-29 11:36:29 -0400 | [diff] [blame] | 104 | #define USER_SGPR_ADDR_SPACE 8 |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 105 | |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 106 | |
| 107 | #define SENDMSG_GS 2 |
| 108 | #define SENDMSG_GS_DONE 3 |
| 109 | |
| 110 | #define SENDMSG_GS_OP_NOP (0 << 4) |
| 111 | #define SENDMSG_GS_OP_CUT (1 << 4) |
| 112 | #define SENDMSG_GS_OP_EMIT (2 << 4) |
| 113 | #define SENDMSG_GS_OP_EMIT_CUT (3 << 4) |
| 114 | |
Marek Olšák | e29353f | 2014-09-17 22:17:02 +0200 | [diff] [blame] | 115 | /** |
| 116 | * Returns a unique index for a semantic name and index. The index must be |
| 117 | * less than 64, so that a 64-bit bitmask of used inputs or outputs can be |
| 118 | * calculated. |
| 119 | */ |
Marek Olšák | 0a2d6f0 | 2014-09-30 16:55:36 +0200 | [diff] [blame] | 120 | unsigned si_shader_io_get_unique_index(unsigned semantic_name, unsigned index) |
Marek Olšák | e29353f | 2014-09-17 22:17:02 +0200 | [diff] [blame] | 121 | { |
| 122 | switch (semantic_name) { |
| 123 | case TGSI_SEMANTIC_POSITION: |
| 124 | return 0; |
| 125 | case TGSI_SEMANTIC_PSIZE: |
| 126 | return 1; |
| 127 | case TGSI_SEMANTIC_CLIPDIST: |
| 128 | assert(index <= 1); |
| 129 | return 2 + index; |
| 130 | case TGSI_SEMANTIC_CLIPVERTEX: |
| 131 | return 4; |
| 132 | case TGSI_SEMANTIC_COLOR: |
| 133 | assert(index <= 1); |
| 134 | return 5 + index; |
| 135 | case TGSI_SEMANTIC_BCOLOR: |
| 136 | assert(index <= 1); |
| 137 | return 7 + index; |
| 138 | case TGSI_SEMANTIC_FOG: |
| 139 | return 9; |
| 140 | case TGSI_SEMANTIC_EDGEFLAG: |
| 141 | return 10; |
| 142 | case TGSI_SEMANTIC_GENERIC: |
| 143 | assert(index <= 63-11); |
| 144 | return 11 + index; |
| 145 | default: |
| 146 | assert(0); |
| 147 | return 63; |
| 148 | } |
| 149 | } |
| 150 | |
| 151 | /** |
| 152 | * Given a semantic name and index of a parameter and a mask of used parameters |
| 153 | * (inputs or outputs), return the index of the parameter in the list of all |
| 154 | * used parameters. |
| 155 | * |
| 156 | * For example, assume this list of parameters: |
| 157 | * POSITION, PSIZE, GENERIC0, GENERIC2 |
| 158 | * which has the mask: |
| 159 | * 11000000000101 |
| 160 | * Then: |
| 161 | * querying POSITION returns 0, |
| 162 | * querying PSIZE returns 1, |
| 163 | * querying GENERIC0 returns 2, |
| 164 | * querying GENERIC2 returns 3. |
| 165 | * |
| 166 | * Which can be used as an offset to a parameter buffer in units of vec4s. |
| 167 | */ |
| 168 | static int get_param_index(unsigned semantic_name, unsigned index, |
| 169 | uint64_t mask) |
| 170 | { |
Marek Olšák | 0a2d6f0 | 2014-09-30 16:55:36 +0200 | [diff] [blame] | 171 | unsigned unique_index = si_shader_io_get_unique_index(semantic_name, index); |
Marek Olšák | e29353f | 2014-09-17 22:17:02 +0200 | [diff] [blame] | 172 | int i, param_index = 0; |
| 173 | |
| 174 | /* If not present... */ |
| 175 | if (!((1llu << unique_index) & mask)) |
| 176 | return -1; |
| 177 | |
| 178 | for (i = 0; mask; i++) { |
| 179 | uint64_t bit = 1llu << i; |
| 180 | |
| 181 | if (bit & mask) { |
| 182 | if (i == unique_index) |
| 183 | return param_index; |
| 184 | |
| 185 | mask &= ~bit; |
| 186 | param_index++; |
| 187 | } |
| 188 | } |
| 189 | |
| 190 | assert(!"unreachable"); |
| 191 | return -1; |
| 192 | } |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 193 | |
Tom Stellard | 467f516 | 2012-05-16 15:15:35 -0400 | [diff] [blame] | 194 | /** |
Marek Olšák | d787608 | 2014-09-26 23:06:32 +0200 | [diff] [blame] | 195 | * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad. |
| 196 | * It's equivalent to doing a load from &base_ptr[index]. |
Tom Stellard | 467f516 | 2012-05-16 15:15:35 -0400 | [diff] [blame] | 197 | * |
Marek Olšák | d787608 | 2014-09-26 23:06:32 +0200 | [diff] [blame] | 198 | * \param base_ptr Where the array starts. |
| 199 | * \param index The element index into the array. |
Tom Stellard | 467f516 | 2012-05-16 15:15:35 -0400 | [diff] [blame] | 200 | */ |
Marek Olšák | d787608 | 2014-09-26 23:06:32 +0200 | [diff] [blame] | 201 | static LLVMValueRef build_indexed_load(struct si_shader_context *si_shader_ctx, |
| 202 | LLVMValueRef base_ptr, LLVMValueRef index) |
Tom Stellard | 467f516 | 2012-05-16 15:15:35 -0400 | [diff] [blame] | 203 | { |
Marek Olšák | d787608 | 2014-09-26 23:06:32 +0200 | [diff] [blame] | 204 | struct lp_build_tgsi_context *bld_base = &si_shader_ctx->radeon_bld.soa.bld_base; |
| 205 | struct gallivm_state *gallivm = bld_base->base.gallivm; |
| 206 | LLVMValueRef indices[2], pointer; |
Tom Stellard | 467f516 | 2012-05-16 15:15:35 -0400 | [diff] [blame] | 207 | |
Marek Olšák | d787608 | 2014-09-26 23:06:32 +0200 | [diff] [blame] | 208 | indices[0] = bld_base->uint_bld.zero; |
| 209 | indices[1] = index; |
Christian König | 206f059 | 2013-03-20 14:37:21 +0100 | [diff] [blame] | 210 | |
Marek Olšák | d787608 | 2014-09-26 23:06:32 +0200 | [diff] [blame] | 211 | pointer = LLVMBuildGEP(gallivm->builder, base_ptr, indices, 2, ""); |
| 212 | return LLVMBuildLoad(gallivm->builder, pointer, ""); |
| 213 | } |
| 214 | |
| 215 | /** |
| 216 | * Do a load from &base_ptr[index], but also add a flag that it's loading |
| 217 | * a constant. |
| 218 | */ |
| 219 | static LLVMValueRef build_indexed_load_const( |
| 220 | struct si_shader_context * si_shader_ctx, |
| 221 | LLVMValueRef base_ptr, LLVMValueRef index) |
| 222 | { |
| 223 | LLVMValueRef result = build_indexed_load(si_shader_ctx, base_ptr, index); |
Christian König | 206f059 | 2013-03-20 14:37:21 +0100 | [diff] [blame] | 224 | LLVMSetMetadata(result, 1, si_shader_ctx->const_md); |
| 225 | return result; |
Tom Stellard | 467f516 | 2012-05-16 15:15:35 -0400 | [diff] [blame] | 226 | } |
| 227 | |
Marek Olšák | f317ce5 | 2013-09-05 15:39:57 +0200 | [diff] [blame] | 228 | static LLVMValueRef get_instance_index_for_fetch( |
Christian König | a0dca44 | 2013-03-22 15:59:22 +0100 | [diff] [blame] | 229 | struct radeon_llvm_context * radeon_bld, |
| 230 | unsigned divisor) |
| 231 | { |
Marek Olšák | 8d03d92 | 2013-09-01 23:59:06 +0200 | [diff] [blame] | 232 | struct si_shader_context *si_shader_ctx = |
| 233 | si_shader_context(&radeon_bld->soa.bld_base); |
Christian König | a0dca44 | 2013-03-22 15:59:22 +0100 | [diff] [blame] | 234 | struct gallivm_state * gallivm = radeon_bld->soa.bld_base.base.gallivm; |
| 235 | |
Marek Olšák | 8d03d92 | 2013-09-01 23:59:06 +0200 | [diff] [blame] | 236 | LLVMValueRef result = LLVMGetParam(radeon_bld->main_fn, |
| 237 | si_shader_ctx->param_instance_id); |
Christian König | a0dca44 | 2013-03-22 15:59:22 +0100 | [diff] [blame] | 238 | |
Marek Olšák | 50908a8 | 2015-01-31 19:00:44 +0100 | [diff] [blame] | 239 | /* The division must be done before START_INSTANCE is added. */ |
Christian König | a0dca44 | 2013-03-22 15:59:22 +0100 | [diff] [blame] | 240 | if (divisor > 1) |
| 241 | result = LLVMBuildUDiv(gallivm->builder, result, |
| 242 | lp_build_const_int32(gallivm, divisor), ""); |
| 243 | |
Marek Olšák | 50908a8 | 2015-01-31 19:00:44 +0100 | [diff] [blame] | 244 | return LLVMBuildAdd(gallivm->builder, result, LLVMGetParam( |
| 245 | radeon_bld->main_fn, SI_PARAM_START_INSTANCE), ""); |
Christian König | a0dca44 | 2013-03-22 15:59:22 +0100 | [diff] [blame] | 246 | } |
| 247 | |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 248 | static void declare_input_vs( |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 249 | struct radeon_llvm_context *radeon_bld, |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 250 | unsigned input_index, |
| 251 | const struct tgsi_full_declaration *decl) |
| 252 | { |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 253 | struct lp_build_context *base = &radeon_bld->soa.bld_base.base; |
| 254 | struct gallivm_state *gallivm = base->gallivm; |
| 255 | struct si_shader_context *si_shader_ctx = |
| 256 | si_shader_context(&radeon_bld->soa.bld_base); |
Christian König | a0dca44 | 2013-03-22 15:59:22 +0100 | [diff] [blame] | 257 | unsigned divisor = si_shader_ctx->shader->key.vs.instance_divisors[input_index]; |
| 258 | |
| 259 | unsigned chan; |
| 260 | |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 261 | LLVMValueRef t_list_ptr; |
| 262 | LLVMValueRef t_offset; |
Tom Stellard | 467f516 | 2012-05-16 15:15:35 -0400 | [diff] [blame] | 263 | LLVMValueRef t_list; |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 264 | LLVMValueRef attribute_offset; |
Christian König | a0dca44 | 2013-03-22 15:59:22 +0100 | [diff] [blame] | 265 | LLVMValueRef buffer_index; |
Tom Stellard | 467f516 | 2012-05-16 15:15:35 -0400 | [diff] [blame] | 266 | LLVMValueRef args[3]; |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 267 | LLVMTypeRef vec4_type; |
| 268 | LLVMValueRef input; |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 269 | |
Tom Stellard | 467f516 | 2012-05-16 15:15:35 -0400 | [diff] [blame] | 270 | /* Load the T list */ |
Christian König | 55fe5cc | 2013-03-04 16:30:06 +0100 | [diff] [blame] | 271 | t_list_ptr = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_VERTEX_BUFFER); |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 272 | |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 273 | t_offset = lp_build_const_int32(gallivm, input_index); |
Tom Stellard | 467f516 | 2012-05-16 15:15:35 -0400 | [diff] [blame] | 274 | |
Marek Olšák | d787608 | 2014-09-26 23:06:32 +0200 | [diff] [blame] | 275 | t_list = build_indexed_load_const(si_shader_ctx, t_list_ptr, t_offset); |
Tom Stellard | 467f516 | 2012-05-16 15:15:35 -0400 | [diff] [blame] | 276 | |
| 277 | /* Build the attribute offset */ |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 278 | attribute_offset = lp_build_const_int32(gallivm, 0); |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 279 | |
Christian König | a0dca44 | 2013-03-22 15:59:22 +0100 | [diff] [blame] | 280 | if (divisor) { |
| 281 | /* Build index from instance ID, start instance and divisor */ |
Marek Olšák | 8c37c16 | 2014-09-16 18:40:07 +0200 | [diff] [blame] | 282 | si_shader_ctx->shader->uses_instanceid = true; |
Marek Olšák | f317ce5 | 2013-09-05 15:39:57 +0200 | [diff] [blame] | 283 | buffer_index = get_instance_index_for_fetch(&si_shader_ctx->radeon_bld, divisor); |
Christian König | a0dca44 | 2013-03-22 15:59:22 +0100 | [diff] [blame] | 284 | } else { |
Marek Olšák | 09056b3 | 2014-04-23 16:15:36 +0200 | [diff] [blame] | 285 | /* Load the buffer index for vertices. */ |
| 286 | LLVMValueRef vertex_id = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, |
| 287 | si_shader_ctx->param_vertex_id); |
| 288 | LLVMValueRef base_vertex = LLVMGetParam(radeon_bld->main_fn, |
| 289 | SI_PARAM_BASE_VERTEX); |
| 290 | buffer_index = LLVMBuildAdd(gallivm->builder, base_vertex, vertex_id, ""); |
Christian König | a0dca44 | 2013-03-22 15:59:22 +0100 | [diff] [blame] | 291 | } |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 292 | |
| 293 | vec4_type = LLVMVectorType(base->elem_type, 4); |
Tom Stellard | 467f516 | 2012-05-16 15:15:35 -0400 | [diff] [blame] | 294 | args[0] = t_list; |
| 295 | args[1] = attribute_offset; |
Christian König | a0dca44 | 2013-03-22 15:59:22 +0100 | [diff] [blame] | 296 | args[2] = buffer_index; |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 297 | input = build_intrinsic(gallivm->builder, |
Christian König | 44e3224 | 2013-03-20 12:10:35 +0100 | [diff] [blame] | 298 | "llvm.SI.vs.load.input", vec4_type, args, 3, |
| 299 | LLVMReadNoneAttribute | LLVMNoUnwindAttribute); |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 300 | |
| 301 | /* Break up the vec4 into individual components */ |
| 302 | for (chan = 0; chan < 4; chan++) { |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 303 | LLVMValueRef llvm_chan = lp_build_const_int32(gallivm, chan); |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 304 | /* XXX: Use a helper function for this. There is one in |
| 305 | * tgsi_llvm.c. */ |
| 306 | si_shader_ctx->radeon_bld.inputs[radeon_llvm_reg_index_soa(input_index, chan)] = |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 307 | LLVMBuildExtractElement(gallivm->builder, |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 308 | input, llvm_chan, ""); |
| 309 | } |
| 310 | } |
| 311 | |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 312 | static LLVMValueRef fetch_input_gs( |
| 313 | struct lp_build_tgsi_context *bld_base, |
| 314 | const struct tgsi_full_src_register *reg, |
| 315 | enum tgsi_opcode_type type, |
| 316 | unsigned swizzle) |
| 317 | { |
| 318 | struct lp_build_context *base = &bld_base->base; |
| 319 | struct si_shader_context *si_shader_ctx = si_shader_context(bld_base); |
Marek Olšák | 8c37c16 | 2014-09-16 18:40:07 +0200 | [diff] [blame] | 320 | struct si_shader *shader = si_shader_ctx->shader; |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 321 | struct lp_build_context *uint = &si_shader_ctx->radeon_bld.soa.bld_base.uint_bld; |
| 322 | struct gallivm_state *gallivm = base->gallivm; |
| 323 | LLVMTypeRef i32 = LLVMInt32TypeInContext(gallivm->context); |
| 324 | LLVMValueRef vtx_offset; |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 325 | LLVMValueRef args[9]; |
| 326 | unsigned vtx_offset_param; |
Marek Olšák | e23fec1 | 2014-10-04 17:40:39 +0200 | [diff] [blame] | 327 | struct tgsi_shader_info *info = &shader->selector->info; |
| 328 | unsigned semantic_name = info->input_semantic_name[reg->Register.Index]; |
| 329 | unsigned semantic_index = info->input_semantic_index[reg->Register.Index]; |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 330 | |
Marek Olšák | e23fec1 | 2014-10-04 17:40:39 +0200 | [diff] [blame] | 331 | if (swizzle != ~0 && semantic_name == TGSI_SEMANTIC_PRIMID) { |
Michel Dänzer | d8b3d80 | 2014-01-09 12:55:26 +0900 | [diff] [blame] | 332 | if (swizzle == 0) |
| 333 | return LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, |
| 334 | SI_PARAM_PRIMITIVE_ID); |
| 335 | else |
| 336 | return uint->zero; |
| 337 | } |
| 338 | |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 339 | if (!reg->Register.Dimension) |
| 340 | return NULL; |
| 341 | |
| 342 | if (swizzle == ~0) { |
| 343 | LLVMValueRef values[TGSI_NUM_CHANNELS]; |
| 344 | unsigned chan; |
| 345 | for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) { |
| 346 | values[chan] = fetch_input_gs(bld_base, reg, type, chan); |
| 347 | } |
| 348 | return lp_build_gather_values(bld_base->base.gallivm, values, |
| 349 | TGSI_NUM_CHANNELS); |
| 350 | } |
| 351 | |
| 352 | /* Get the vertex offset parameter */ |
| 353 | vtx_offset_param = reg->Dimension.Index; |
| 354 | if (vtx_offset_param < 2) { |
| 355 | vtx_offset_param += SI_PARAM_VTX0_OFFSET; |
| 356 | } else { |
| 357 | assert(vtx_offset_param < 6); |
| 358 | vtx_offset_param += SI_PARAM_VTX2_OFFSET - 2; |
| 359 | } |
| 360 | vtx_offset = lp_build_mul_imm(uint, |
| 361 | LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, |
| 362 | vtx_offset_param), |
| 363 | 4); |
| 364 | |
Marek Olšák | fc3b335 | 2014-10-05 13:33:40 +0200 | [diff] [blame] | 365 | args[0] = si_shader_ctx->esgs_ring; |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 366 | args[1] = vtx_offset; |
| 367 | args[2] = lp_build_const_int32(gallivm, |
Marek Olšák | e23fec1 | 2014-10-04 17:40:39 +0200 | [diff] [blame] | 368 | (get_param_index(semantic_name, semantic_index, |
Marek Olšák | 0a2d6f0 | 2014-09-30 16:55:36 +0200 | [diff] [blame] | 369 | shader->selector->gs_used_inputs) * 4 + |
Michel Dänzer | 7c7d738 | 2014-01-09 15:33:34 +0900 | [diff] [blame] | 370 | swizzle) * 256); |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 371 | args[3] = uint->zero; |
| 372 | args[4] = uint->one; /* OFFEN */ |
| 373 | args[5] = uint->zero; /* IDXEN */ |
| 374 | args[6] = uint->one; /* GLC */ |
| 375 | args[7] = uint->zero; /* SLC */ |
| 376 | args[8] = uint->zero; /* TFE */ |
| 377 | |
| 378 | return LLVMBuildBitCast(gallivm->builder, |
| 379 | build_intrinsic(gallivm->builder, |
| 380 | "llvm.SI.buffer.load.dword.i32.i32", |
| 381 | i32, args, 9, |
| 382 | LLVMReadOnlyAttribute | LLVMNoUnwindAttribute), |
| 383 | tgsi2llvmtype(bld_base, type), ""); |
| 384 | } |
| 385 | |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 386 | static void declare_input_fs( |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 387 | struct radeon_llvm_context *radeon_bld, |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 388 | unsigned input_index, |
| 389 | const struct tgsi_full_declaration *decl) |
| 390 | { |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 391 | struct lp_build_context *base = &radeon_bld->soa.bld_base.base; |
| 392 | struct si_shader_context *si_shader_ctx = |
| 393 | si_shader_context(&radeon_bld->soa.bld_base); |
Marek Olšák | 8c37c16 | 2014-09-16 18:40:07 +0200 | [diff] [blame] | 394 | struct si_shader *shader = si_shader_ctx->shader; |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 395 | struct lp_build_context *uint = &radeon_bld->soa.bld_base.uint_bld; |
| 396 | struct gallivm_state *gallivm = base->gallivm; |
Tom Stellard | 0fb1e68 | 2012-09-06 16:18:11 -0400 | [diff] [blame] | 397 | LLVMTypeRef input_type = LLVMFloatTypeInContext(gallivm->context); |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 398 | LLVMValueRef main_fn = radeon_bld->main_fn; |
Christian König | 0666ffd | 2013-03-05 15:07:39 +0100 | [diff] [blame] | 399 | |
| 400 | LLVMValueRef interp_param; |
| 401 | const char * intr_name; |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 402 | |
| 403 | /* This value is: |
| 404 | * [15:0] NewPrimMask (Bit mask for each quad. It is set it the |
| 405 | * quad begins a new primitive. Bit 0 always needs |
| 406 | * to be unset) |
| 407 | * [32:16] ParamOffset |
| 408 | * |
| 409 | */ |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 410 | LLVMValueRef params = LLVMGetParam(main_fn, SI_PARAM_PRIM_MASK); |
Michel Dänzer | c3db19e | 2012-09-27 20:01:33 +0200 | [diff] [blame] | 411 | LLVMValueRef attr_number; |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 412 | |
Christian König | 0666ffd | 2013-03-05 15:07:39 +0100 | [diff] [blame] | 413 | unsigned chan; |
| 414 | |
Tom Stellard | 0fb1e68 | 2012-09-06 16:18:11 -0400 | [diff] [blame] | 415 | if (decl->Semantic.Name == TGSI_SEMANTIC_POSITION) { |
| 416 | for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) { |
Tom Stellard | 0fb1e68 | 2012-09-06 16:18:11 -0400 | [diff] [blame] | 417 | unsigned soa_index = |
| 418 | radeon_llvm_reg_index_soa(input_index, chan); |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 419 | radeon_bld->inputs[soa_index] = |
Christian König | 0666ffd | 2013-03-05 15:07:39 +0100 | [diff] [blame] | 420 | LLVMGetParam(main_fn, SI_PARAM_POS_X_FLOAT + chan); |
Michel Dänzer | 954bc4a | 2013-02-13 15:57:23 +0100 | [diff] [blame] | 421 | |
| 422 | if (chan == 3) |
| 423 | /* RCP for fragcoord.w */ |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 424 | radeon_bld->inputs[soa_index] = |
Michel Dänzer | 954bc4a | 2013-02-13 15:57:23 +0100 | [diff] [blame] | 425 | LLVMBuildFDiv(gallivm->builder, |
| 426 | lp_build_const_float(gallivm, 1.0f), |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 427 | radeon_bld->inputs[soa_index], |
Michel Dänzer | 954bc4a | 2013-02-13 15:57:23 +0100 | [diff] [blame] | 428 | ""); |
Tom Stellard | 0fb1e68 | 2012-09-06 16:18:11 -0400 | [diff] [blame] | 429 | } |
| 430 | return; |
| 431 | } |
| 432 | |
Michel Dänzer | 97078b1 | 2012-09-25 12:41:31 +0200 | [diff] [blame] | 433 | if (decl->Semantic.Name == TGSI_SEMANTIC_FACE) { |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 434 | radeon_bld->inputs[radeon_llvm_reg_index_soa(input_index, 0)] = |
Marek Olšák | e827bb6 | 2014-10-16 16:20:26 +0200 | [diff] [blame] | 435 | LLVMGetParam(main_fn, SI_PARAM_FRONT_FACE); |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 436 | radeon_bld->inputs[radeon_llvm_reg_index_soa(input_index, 1)] = |
| 437 | radeon_bld->inputs[radeon_llvm_reg_index_soa(input_index, 2)] = |
Michel Dänzer | 97078b1 | 2012-09-25 12:41:31 +0200 | [diff] [blame] | 438 | lp_build_const_float(gallivm, 0.0f); |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 439 | radeon_bld->inputs[radeon_llvm_reg_index_soa(input_index, 3)] = |
Michel Dänzer | 97078b1 | 2012-09-25 12:41:31 +0200 | [diff] [blame] | 440 | lp_build_const_float(gallivm, 1.0f); |
| 441 | |
| 442 | return; |
| 443 | } |
| 444 | |
Marek Olšák | 8b057dd | 2014-10-04 22:15:07 +0200 | [diff] [blame] | 445 | shader->ps_input_param_offset[input_index] = shader->nparam++; |
Michel Dänzer | c3db19e | 2012-09-27 20:01:33 +0200 | [diff] [blame] | 446 | attr_number = lp_build_const_int32(gallivm, |
Marek Olšák | 8b057dd | 2014-10-04 22:15:07 +0200 | [diff] [blame] | 447 | shader->ps_input_param_offset[input_index]); |
Michel Dänzer | c3db19e | 2012-09-27 20:01:33 +0200 | [diff] [blame] | 448 | |
Francisco Jerez | 1279923 | 2012-04-30 18:27:52 +0200 | [diff] [blame] | 449 | switch (decl->Interp.Interpolate) { |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 450 | case TGSI_INTERPOLATE_CONSTANT: |
Christian König | 0666ffd | 2013-03-05 15:07:39 +0100 | [diff] [blame] | 451 | interp_param = 0; |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 452 | break; |
| 453 | case TGSI_INTERPOLATE_LINEAR: |
Marek Olšák | 10e386f | 2014-09-30 17:09:13 +0200 | [diff] [blame] | 454 | if (decl->Interp.Location == TGSI_INTERPOLATE_LOC_SAMPLE) |
Marek Olšák | 99df120 | 2014-05-06 19:10:52 +0200 | [diff] [blame] | 455 | interp_param = LLVMGetParam(main_fn, SI_PARAM_LINEAR_SAMPLE); |
Ilia Mirkin | 4c97ed4 | 2014-07-01 20:54:01 -0400 | [diff] [blame] | 456 | else if (decl->Interp.Location == TGSI_INTERPOLATE_LOC_CENTROID) |
Christian König | 0666ffd | 2013-03-05 15:07:39 +0100 | [diff] [blame] | 457 | interp_param = LLVMGetParam(main_fn, SI_PARAM_LINEAR_CENTROID); |
Michel Dänzer | 1deb2be | 2012-05-14 16:26:19 +0200 | [diff] [blame] | 458 | else |
Christian König | 0666ffd | 2013-03-05 15:07:39 +0100 | [diff] [blame] | 459 | interp_param = LLVMGetParam(main_fn, SI_PARAM_LINEAR_CENTER); |
Michel Dänzer | 1deb2be | 2012-05-14 16:26:19 +0200 | [diff] [blame] | 460 | break; |
Marek Olšák | 99df120 | 2014-05-06 19:10:52 +0200 | [diff] [blame] | 461 | case TGSI_INTERPOLATE_COLOR: |
Michel Dänzer | 1deb2be | 2012-05-14 16:26:19 +0200 | [diff] [blame] | 462 | case TGSI_INTERPOLATE_PERSPECTIVE: |
Marek Olšák | 10e386f | 2014-09-30 17:09:13 +0200 | [diff] [blame] | 463 | if (decl->Interp.Location == TGSI_INTERPOLATE_LOC_SAMPLE) |
Marek Olšák | 99df120 | 2014-05-06 19:10:52 +0200 | [diff] [blame] | 464 | interp_param = LLVMGetParam(main_fn, SI_PARAM_PERSP_SAMPLE); |
Ilia Mirkin | 4c97ed4 | 2014-07-01 20:54:01 -0400 | [diff] [blame] | 465 | else if (decl->Interp.Location == TGSI_INTERPOLATE_LOC_CENTROID) |
Christian König | 0666ffd | 2013-03-05 15:07:39 +0100 | [diff] [blame] | 466 | interp_param = LLVMGetParam(main_fn, SI_PARAM_PERSP_CENTROID); |
Michel Dänzer | 1deb2be | 2012-05-14 16:26:19 +0200 | [diff] [blame] | 467 | else |
Christian König | 0666ffd | 2013-03-05 15:07:39 +0100 | [diff] [blame] | 468 | interp_param = LLVMGetParam(main_fn, SI_PARAM_PERSP_CENTER); |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 469 | break; |
| 470 | default: |
| 471 | fprintf(stderr, "Warning: Unhandled interpolation mode.\n"); |
| 472 | return; |
| 473 | } |
| 474 | |
Marek Olšák | 13de947 | 2015-01-04 20:09:51 +0100 | [diff] [blame] | 475 | /* fs.constant returns the param from the middle vertex, so it's not |
| 476 | * really useful for flat shading. It's meant to be used for custom |
| 477 | * interpolation (but the intrinsic can't fetch from the other two |
| 478 | * vertices). |
| 479 | * |
| 480 | * Luckily, it doesn't matter, because we rely on the FLAT_SHADE state |
| 481 | * to do the right thing. The only reason we use fs.constant is that |
| 482 | * fs.interp cannot be used on integers, because they can be equal |
| 483 | * to NaN. |
| 484 | */ |
Christian König | 0666ffd | 2013-03-05 15:07:39 +0100 | [diff] [blame] | 485 | intr_name = interp_param ? "llvm.SI.fs.interp" : "llvm.SI.fs.constant"; |
| 486 | |
Michel Dänzer | 691f08d | 2012-09-06 18:03:38 +0200 | [diff] [blame] | 487 | if (decl->Semantic.Name == TGSI_SEMANTIC_COLOR && |
Christian König | a0dca44 | 2013-03-22 15:59:22 +0100 | [diff] [blame] | 488 | si_shader_ctx->shader->key.ps.color_two_side) { |
Christian König | 0666ffd | 2013-03-05 15:07:39 +0100 | [diff] [blame] | 489 | LLVMValueRef args[4]; |
Michel Dänzer | 691f08d | 2012-09-06 18:03:38 +0200 | [diff] [blame] | 490 | LLVMValueRef face, is_face_positive; |
| 491 | LLVMValueRef back_attr_number = |
| 492 | lp_build_const_int32(gallivm, |
Marek Olšák | 8b057dd | 2014-10-04 22:15:07 +0200 | [diff] [blame] | 493 | shader->ps_input_param_offset[input_index] + 1); |
Michel Dänzer | 691f08d | 2012-09-06 18:03:38 +0200 | [diff] [blame] | 494 | |
Christian König | 0666ffd | 2013-03-05 15:07:39 +0100 | [diff] [blame] | 495 | face = LLVMGetParam(main_fn, SI_PARAM_FRONT_FACE); |
| 496 | |
Michel Dänzer | 691f08d | 2012-09-06 18:03:38 +0200 | [diff] [blame] | 497 | is_face_positive = LLVMBuildFCmp(gallivm->builder, |
Marek Olšák | d1d2af2 | 2015-01-04 20:45:35 +0100 | [diff] [blame] | 498 | LLVMRealOGT, face, |
Michel Dänzer | 691f08d | 2012-09-06 18:03:38 +0200 | [diff] [blame] | 499 | lp_build_const_float(gallivm, 0.0f), |
| 500 | ""); |
| 501 | |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 502 | args[2] = params; |
Christian König | 0666ffd | 2013-03-05 15:07:39 +0100 | [diff] [blame] | 503 | args[3] = interp_param; |
Michel Dänzer | 691f08d | 2012-09-06 18:03:38 +0200 | [diff] [blame] | 504 | for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) { |
| 505 | LLVMValueRef llvm_chan = lp_build_const_int32(gallivm, chan); |
| 506 | unsigned soa_index = radeon_llvm_reg_index_soa(input_index, chan); |
| 507 | LLVMValueRef front, back; |
| 508 | |
| 509 | args[0] = llvm_chan; |
| 510 | args[1] = attr_number; |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 511 | front = build_intrinsic(gallivm->builder, intr_name, |
Christian König | 0666ffd | 2013-03-05 15:07:39 +0100 | [diff] [blame] | 512 | input_type, args, args[3] ? 4 : 3, |
Christian König | 44e3224 | 2013-03-20 12:10:35 +0100 | [diff] [blame] | 513 | LLVMReadNoneAttribute | LLVMNoUnwindAttribute); |
Michel Dänzer | 691f08d | 2012-09-06 18:03:38 +0200 | [diff] [blame] | 514 | |
| 515 | args[1] = back_attr_number; |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 516 | back = build_intrinsic(gallivm->builder, intr_name, |
Christian König | 0666ffd | 2013-03-05 15:07:39 +0100 | [diff] [blame] | 517 | input_type, args, args[3] ? 4 : 3, |
Christian König | 44e3224 | 2013-03-20 12:10:35 +0100 | [diff] [blame] | 518 | LLVMReadNoneAttribute | LLVMNoUnwindAttribute); |
Michel Dänzer | 691f08d | 2012-09-06 18:03:38 +0200 | [diff] [blame] | 519 | |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 520 | radeon_bld->inputs[soa_index] = |
Michel Dänzer | 691f08d | 2012-09-06 18:03:38 +0200 | [diff] [blame] | 521 | LLVMBuildSelect(gallivm->builder, |
| 522 | is_face_positive, |
| 523 | front, |
| 524 | back, |
| 525 | ""); |
| 526 | } |
| 527 | |
Michel Dänzer | 7c7d738 | 2014-01-09 15:33:34 +0900 | [diff] [blame] | 528 | shader->nparam++; |
Michel Dänzer | 237cb07 | 2013-08-21 18:00:35 +0200 | [diff] [blame] | 529 | } else if (decl->Semantic.Name == TGSI_SEMANTIC_FOG) { |
| 530 | LLVMValueRef args[4]; |
| 531 | |
| 532 | args[0] = uint->zero; |
| 533 | args[1] = attr_number; |
| 534 | args[2] = params; |
| 535 | args[3] = interp_param; |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 536 | radeon_bld->inputs[radeon_llvm_reg_index_soa(input_index, 0)] = |
| 537 | build_intrinsic(gallivm->builder, intr_name, |
| 538 | input_type, args, args[3] ? 4 : 3, |
| 539 | LLVMReadNoneAttribute | LLVMNoUnwindAttribute); |
| 540 | radeon_bld->inputs[radeon_llvm_reg_index_soa(input_index, 1)] = |
| 541 | radeon_bld->inputs[radeon_llvm_reg_index_soa(input_index, 2)] = |
Michel Dänzer | 237cb07 | 2013-08-21 18:00:35 +0200 | [diff] [blame] | 542 | lp_build_const_float(gallivm, 0.0f); |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 543 | radeon_bld->inputs[radeon_llvm_reg_index_soa(input_index, 3)] = |
Michel Dänzer | 237cb07 | 2013-08-21 18:00:35 +0200 | [diff] [blame] | 544 | lp_build_const_float(gallivm, 1.0f); |
Michel Dänzer | 691f08d | 2012-09-06 18:03:38 +0200 | [diff] [blame] | 545 | } else { |
| 546 | for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) { |
Christian König | 0666ffd | 2013-03-05 15:07:39 +0100 | [diff] [blame] | 547 | LLVMValueRef args[4]; |
Michel Dänzer | 691f08d | 2012-09-06 18:03:38 +0200 | [diff] [blame] | 548 | LLVMValueRef llvm_chan = lp_build_const_int32(gallivm, chan); |
| 549 | unsigned soa_index = radeon_llvm_reg_index_soa(input_index, chan); |
| 550 | args[0] = llvm_chan; |
| 551 | args[1] = attr_number; |
| 552 | args[2] = params; |
Christian König | 0666ffd | 2013-03-05 15:07:39 +0100 | [diff] [blame] | 553 | args[3] = interp_param; |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 554 | radeon_bld->inputs[soa_index] = |
| 555 | build_intrinsic(gallivm->builder, intr_name, |
Christian König | 0666ffd | 2013-03-05 15:07:39 +0100 | [diff] [blame] | 556 | input_type, args, args[3] ? 4 : 3, |
Christian König | 44e3224 | 2013-03-20 12:10:35 +0100 | [diff] [blame] | 557 | LLVMReadNoneAttribute | LLVMNoUnwindAttribute); |
Michel Dänzer | 691f08d | 2012-09-06 18:03:38 +0200 | [diff] [blame] | 558 | } |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 559 | } |
| 560 | } |
| 561 | |
Marek Olšák | 5b06fc3 | 2014-05-06 18:12:40 +0200 | [diff] [blame] | 562 | static LLVMValueRef get_sample_id(struct radeon_llvm_context *radeon_bld) |
| 563 | { |
| 564 | struct gallivm_state *gallivm = &radeon_bld->gallivm; |
| 565 | LLVMValueRef value = LLVMGetParam(radeon_bld->main_fn, |
| 566 | SI_PARAM_ANCILLARY); |
| 567 | value = LLVMBuildLShr(gallivm->builder, value, |
| 568 | lp_build_const_int32(gallivm, 8), ""); |
| 569 | value = LLVMBuildAnd(gallivm->builder, value, |
| 570 | lp_build_const_int32(gallivm, 0xf), ""); |
| 571 | return value; |
| 572 | } |
| 573 | |
Marek Olšák | d787608 | 2014-09-26 23:06:32 +0200 | [diff] [blame] | 574 | /** |
| 575 | * Load a dword from a constant buffer. |
| 576 | */ |
| 577 | static LLVMValueRef buffer_load_const(LLVMBuilderRef builder, LLVMValueRef resource, |
| 578 | LLVMValueRef offset, LLVMTypeRef return_type) |
Marek Olšák | 250aa93 | 2014-05-06 14:10:47 +0200 | [diff] [blame] | 579 | { |
| 580 | LLVMValueRef args[2] = {resource, offset}; |
| 581 | |
| 582 | return build_intrinsic(builder, "llvm.SI.load.const", return_type, args, 2, |
| 583 | LLVMReadNoneAttribute | LLVMNoUnwindAttribute); |
| 584 | } |
| 585 | |
Christian König | e4ed587 | 2013-03-21 18:02:52 +0100 | [diff] [blame] | 586 | static void declare_system_value( |
| 587 | struct radeon_llvm_context * radeon_bld, |
| 588 | unsigned index, |
| 589 | const struct tgsi_full_declaration *decl) |
| 590 | { |
Marek Olšák | 8d03d92 | 2013-09-01 23:59:06 +0200 | [diff] [blame] | 591 | struct si_shader_context *si_shader_ctx = |
| 592 | si_shader_context(&radeon_bld->soa.bld_base); |
Marek Olšák | 99d9d7c | 2014-05-06 18:20:58 +0200 | [diff] [blame] | 593 | struct lp_build_context *uint_bld = &radeon_bld->soa.bld_base.uint_bld; |
| 594 | struct gallivm_state *gallivm = &radeon_bld->gallivm; |
Christian König | e4ed587 | 2013-03-21 18:02:52 +0100 | [diff] [blame] | 595 | LLVMValueRef value = 0; |
| 596 | |
| 597 | switch (decl->Semantic.Name) { |
| 598 | case TGSI_SEMANTIC_INSTANCEID: |
Marek Olšák | f317ce5 | 2013-09-05 15:39:57 +0200 | [diff] [blame] | 599 | value = LLVMGetParam(radeon_bld->main_fn, |
| 600 | si_shader_ctx->param_instance_id); |
Christian König | e4ed587 | 2013-03-21 18:02:52 +0100 | [diff] [blame] | 601 | break; |
| 602 | |
| 603 | case TGSI_SEMANTIC_VERTEXID: |
Marek Olšák | d7c6f39 | 2015-01-04 14:41:49 +0100 | [diff] [blame] | 604 | value = LLVMBuildAdd(gallivm->builder, |
| 605 | LLVMGetParam(radeon_bld->main_fn, |
| 606 | si_shader_ctx->param_vertex_id), |
| 607 | LLVMGetParam(radeon_bld->main_fn, |
| 608 | SI_PARAM_BASE_VERTEX), ""); |
Christian König | e4ed587 | 2013-03-21 18:02:52 +0100 | [diff] [blame] | 609 | break; |
| 610 | |
Marek Olšák | e3d4bdd | 2015-01-04 14:51:01 +0100 | [diff] [blame] | 611 | case TGSI_SEMANTIC_VERTEXID_NOBASE: |
| 612 | value = LLVMGetParam(radeon_bld->main_fn, |
| 613 | si_shader_ctx->param_vertex_id); |
| 614 | break; |
| 615 | |
| 616 | case TGSI_SEMANTIC_BASEVERTEX: |
| 617 | value = LLVMGetParam(radeon_bld->main_fn, |
| 618 | SI_PARAM_BASE_VERTEX); |
| 619 | break; |
| 620 | |
Marek Olšák | 5b06fc3 | 2014-05-06 18:12:40 +0200 | [diff] [blame] | 621 | case TGSI_SEMANTIC_SAMPLEID: |
| 622 | value = get_sample_id(radeon_bld); |
| 623 | break; |
| 624 | |
Marek Olšák | 99d9d7c | 2014-05-06 18:20:58 +0200 | [diff] [blame] | 625 | case TGSI_SEMANTIC_SAMPLEPOS: |
| 626 | { |
| 627 | LLVMBuilderRef builder = gallivm->builder; |
| 628 | LLVMValueRef desc = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_CONST); |
Marek Olšák | ee2a818 | 2014-07-07 23:27:19 +0200 | [diff] [blame] | 629 | LLVMValueRef buf_index = lp_build_const_int32(gallivm, SI_DRIVER_STATE_CONST_BUF); |
Marek Olšák | d787608 | 2014-09-26 23:06:32 +0200 | [diff] [blame] | 630 | LLVMValueRef resource = build_indexed_load_const(si_shader_ctx, desc, buf_index); |
Marek Olšák | 99d9d7c | 2014-05-06 18:20:58 +0200 | [diff] [blame] | 631 | |
| 632 | /* offset = sample_id * 8 (8 = 2 floats containing samplepos.xy) */ |
| 633 | LLVMValueRef offset0 = lp_build_mul_imm(uint_bld, get_sample_id(radeon_bld), 8); |
| 634 | LLVMValueRef offset1 = LLVMBuildAdd(builder, offset0, lp_build_const_int32(gallivm, 4), ""); |
| 635 | |
| 636 | LLVMValueRef pos[4] = { |
Marek Olšák | d787608 | 2014-09-26 23:06:32 +0200 | [diff] [blame] | 637 | buffer_load_const(builder, resource, offset0, radeon_bld->soa.bld_base.base.elem_type), |
| 638 | buffer_load_const(builder, resource, offset1, radeon_bld->soa.bld_base.base.elem_type), |
Marek Olšák | 99d9d7c | 2014-05-06 18:20:58 +0200 | [diff] [blame] | 639 | lp_build_const_float(gallivm, 0), |
| 640 | lp_build_const_float(gallivm, 0) |
| 641 | }; |
| 642 | value = lp_build_gather_values(gallivm, pos, 4); |
| 643 | break; |
| 644 | } |
| 645 | |
Christian König | e4ed587 | 2013-03-21 18:02:52 +0100 | [diff] [blame] | 646 | default: |
| 647 | assert(!"unknown system value"); |
| 648 | return; |
| 649 | } |
| 650 | |
| 651 | radeon_bld->system_values[index] = value; |
| 652 | } |
| 653 | |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 654 | static LLVMValueRef fetch_constant( |
| 655 | struct lp_build_tgsi_context * bld_base, |
| 656 | const struct tgsi_full_src_register *reg, |
| 657 | enum tgsi_opcode_type type, |
| 658 | unsigned swizzle) |
| 659 | { |
Christian König | 55fe5cc | 2013-03-04 16:30:06 +0100 | [diff] [blame] | 660 | struct si_shader_context *si_shader_ctx = si_shader_context(bld_base); |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 661 | struct lp_build_context * base = &bld_base->base; |
Christian König | 0f6cf2b | 2013-03-15 15:53:25 +0100 | [diff] [blame] | 662 | const struct tgsi_ind_register *ireg = ®->Indirect; |
Marek Olšák | 2fd4200 | 2013-10-25 11:45:47 +0200 | [diff] [blame] | 663 | unsigned buf, idx; |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 664 | |
Christian König | 0f6cf2b | 2013-03-15 15:53:25 +0100 | [diff] [blame] | 665 | LLVMValueRef addr; |
Christian König | f5298b0 | 2013-02-28 14:50:07 +0100 | [diff] [blame] | 666 | LLVMValueRef result; |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 667 | |
Christian König | 8514f5a | 2013-02-04 17:46:42 +0100 | [diff] [blame] | 668 | if (swizzle == LP_CHAN_ALL) { |
| 669 | unsigned chan; |
| 670 | LLVMValueRef values[4]; |
| 671 | for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) |
| 672 | values[chan] = fetch_constant(bld_base, reg, type, chan); |
| 673 | |
| 674 | return lp_build_gather_values(bld_base->base.gallivm, values, 4); |
| 675 | } |
| 676 | |
Marek Olšák | 2fd4200 | 2013-10-25 11:45:47 +0200 | [diff] [blame] | 677 | buf = reg->Register.Dimension ? reg->Dimension.Index : 0; |
Christian König | 0f6cf2b | 2013-03-15 15:53:25 +0100 | [diff] [blame] | 678 | idx = reg->Register.Index * 4 + swizzle; |
Christian König | f5298b0 | 2013-02-28 14:50:07 +0100 | [diff] [blame] | 679 | |
Marek Olšák | 2fd4200 | 2013-10-25 11:45:47 +0200 | [diff] [blame] | 680 | if (!reg->Register.Indirect) |
| 681 | return bitcast(bld_base, type, si_shader_ctx->constants[buf][idx]); |
| 682 | |
Christian König | 0f6cf2b | 2013-03-15 15:53:25 +0100 | [diff] [blame] | 683 | addr = si_shader_ctx->radeon_bld.soa.addr[ireg->Index][ireg->Swizzle]; |
| 684 | addr = LLVMBuildLoad(base->gallivm->builder, addr, "load addr reg"); |
| 685 | addr = lp_build_mul_imm(&bld_base->uint_bld, addr, 16); |
Marek Olšák | 250aa93 | 2014-05-06 14:10:47 +0200 | [diff] [blame] | 686 | addr = lp_build_add(&bld_base->uint_bld, addr, |
| 687 | lp_build_const_int32(base->gallivm, idx * 4)); |
Christian König | e7723b5 | 2012-08-24 12:55:34 +0200 | [diff] [blame] | 688 | |
Marek Olšák | d787608 | 2014-09-26 23:06:32 +0200 | [diff] [blame] | 689 | result = buffer_load_const(base->gallivm->builder, si_shader_ctx->const_resource[buf], |
Marek Olšák | 250aa93 | 2014-05-06 14:10:47 +0200 | [diff] [blame] | 690 | addr, base->elem_type); |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 691 | |
Christian König | f5298b0 | 2013-02-28 14:50:07 +0100 | [diff] [blame] | 692 | return bitcast(bld_base, type, result); |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 693 | } |
| 694 | |
Michel Dänzer | 26c7139 | 2012-08-24 12:03:11 +0200 | [diff] [blame] | 695 | /* Initialize arguments for the shader export intrinsic */ |
| 696 | static void si_llvm_init_export_args(struct lp_build_tgsi_context *bld_base, |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 697 | LLVMValueRef *values, |
Michel Dänzer | 26c7139 | 2012-08-24 12:03:11 +0200 | [diff] [blame] | 698 | unsigned target, |
| 699 | LLVMValueRef *args) |
| 700 | { |
| 701 | struct si_shader_context *si_shader_ctx = si_shader_context(bld_base); |
| 702 | struct lp_build_context *uint = |
| 703 | &si_shader_ctx->radeon_bld.soa.bld_base.uint_bld; |
| 704 | struct lp_build_context *base = &bld_base->base; |
| 705 | unsigned compressed = 0; |
| 706 | unsigned chan; |
| 707 | |
Michel Dänzer | f402acd | 2012-08-22 18:15:36 +0200 | [diff] [blame] | 708 | if (si_shader_ctx->type == TGSI_PROCESSOR_FRAGMENT) { |
| 709 | int cbuf = target - V_008DFC_SQ_EXP_MRT; |
| 710 | |
| 711 | if (cbuf >= 0 && cbuf < 8) { |
Christian König | a0dca44 | 2013-03-22 15:59:22 +0100 | [diff] [blame] | 712 | compressed = (si_shader_ctx->shader->key.ps.export_16bpc >> cbuf) & 0x1; |
Michel Dänzer | 1ace200 | 2012-12-21 15:39:26 +0100 | [diff] [blame] | 713 | |
| 714 | if (compressed) |
| 715 | si_shader_ctx->shader->spi_shader_col_format |= |
| 716 | V_028714_SPI_SHADER_FP16_ABGR << (4 * cbuf); |
| 717 | else |
| 718 | si_shader_ctx->shader->spi_shader_col_format |= |
| 719 | V_028714_SPI_SHADER_32_ABGR << (4 * cbuf); |
Michel Dänzer | e369f40 | 2013-04-30 16:34:10 +0200 | [diff] [blame] | 720 | |
| 721 | si_shader_ctx->shader->cb_shader_mask |= 0xf << (4 * cbuf); |
Michel Dänzer | f402acd | 2012-08-22 18:15:36 +0200 | [diff] [blame] | 722 | } |
| 723 | } |
| 724 | |
| 725 | if (compressed) { |
| 726 | /* Pixel shader needs to pack output values before export */ |
| 727 | for (chan = 0; chan < 2; chan++ ) { |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 728 | args[0] = values[2 * chan]; |
| 729 | args[1] = values[2 * chan + 1]; |
Michel Dänzer | f402acd | 2012-08-22 18:15:36 +0200 | [diff] [blame] | 730 | args[chan + 5] = |
| 731 | build_intrinsic(base->gallivm->builder, |
| 732 | "llvm.SI.packf16", |
| 733 | LLVMInt32TypeInContext(base->gallivm->context), |
| 734 | args, 2, |
Christian König | e4188ee | 2013-02-27 22:39:26 +0100 | [diff] [blame] | 735 | LLVMReadNoneAttribute | LLVMNoUnwindAttribute); |
Michel Dänzer | 8b6aec6 | 2012-11-27 19:53:58 +0100 | [diff] [blame] | 736 | args[chan + 7] = args[chan + 5] = |
| 737 | LLVMBuildBitCast(base->gallivm->builder, |
| 738 | args[chan + 5], |
| 739 | LLVMFloatTypeInContext(base->gallivm->context), |
| 740 | ""); |
Michel Dänzer | f402acd | 2012-08-22 18:15:36 +0200 | [diff] [blame] | 741 | } |
| 742 | |
| 743 | /* Set COMPR flag */ |
| 744 | args[4] = uint->one; |
| 745 | } else { |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 746 | for (chan = 0; chan < 4; chan++ ) |
Michel Dänzer | f402acd | 2012-08-22 18:15:36 +0200 | [diff] [blame] | 747 | /* +5 because the first output value will be |
| 748 | * the 6th argument to the intrinsic. */ |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 749 | args[chan + 5] = values[chan]; |
Michel Dänzer | f402acd | 2012-08-22 18:15:36 +0200 | [diff] [blame] | 750 | |
| 751 | /* Clear COMPR flag */ |
| 752 | args[4] = uint->zero; |
Michel Dänzer | 26c7139 | 2012-08-24 12:03:11 +0200 | [diff] [blame] | 753 | } |
| 754 | |
| 755 | /* XXX: This controls which components of the output |
| 756 | * registers actually get exported. (e.g bit 0 means export |
| 757 | * X component, bit 1 means export Y component, etc.) I'm |
| 758 | * hard coding this to 0xf for now. In the future, we might |
| 759 | * want to do something else. */ |
| 760 | args[0] = lp_build_const_int32(base->gallivm, 0xf); |
| 761 | |
| 762 | /* Specify whether the EXEC mask represents the valid mask */ |
| 763 | args[1] = uint->zero; |
| 764 | |
| 765 | /* Specify whether this is the last export */ |
| 766 | args[2] = uint->zero; |
| 767 | |
| 768 | /* Specify the target we are exporting */ |
| 769 | args[3] = lp_build_const_int32(base->gallivm, target); |
| 770 | |
Michel Dänzer | 26c7139 | 2012-08-24 12:03:11 +0200 | [diff] [blame] | 771 | /* XXX: We probably need to keep track of the output |
| 772 | * values, so we know what we are passing to the next |
| 773 | * stage. */ |
| 774 | } |
| 775 | |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 776 | /* Load from output pointers and initialize arguments for the shader export intrinsic */ |
| 777 | static void si_llvm_init_export_args_load(struct lp_build_tgsi_context *bld_base, |
| 778 | LLVMValueRef *out_ptr, |
| 779 | unsigned target, |
| 780 | LLVMValueRef *args) |
| 781 | { |
| 782 | struct gallivm_state *gallivm = bld_base->base.gallivm; |
| 783 | LLVMValueRef values[4]; |
| 784 | int i; |
| 785 | |
| 786 | for (i = 0; i < 4; i++) |
| 787 | values[i] = LLVMBuildLoad(gallivm->builder, out_ptr[i], ""); |
| 788 | |
| 789 | si_llvm_init_export_args(bld_base, values, target, args); |
| 790 | } |
| 791 | |
Michel Dänzer | 7708a86 | 2012-11-02 15:57:30 +0100 | [diff] [blame] | 792 | static void si_alpha_test(struct lp_build_tgsi_context *bld_base, |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 793 | LLVMValueRef *out_ptr) |
Michel Dänzer | 7708a86 | 2012-11-02 15:57:30 +0100 | [diff] [blame] | 794 | { |
| 795 | struct si_shader_context *si_shader_ctx = si_shader_context(bld_base); |
| 796 | struct gallivm_state *gallivm = bld_base->base.gallivm; |
| 797 | |
Christian König | a0dca44 | 2013-03-22 15:59:22 +0100 | [diff] [blame] | 798 | if (si_shader_ctx->shader->key.ps.alpha_func != PIPE_FUNC_NEVER) { |
Vadim Girlin | 453ea2d | 2013-10-13 19:53:54 +0400 | [diff] [blame] | 799 | LLVMValueRef alpha_ref = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, |
| 800 | SI_PARAM_ALPHA_REF); |
| 801 | |
Michel Dänzer | 7708a86 | 2012-11-02 15:57:30 +0100 | [diff] [blame] | 802 | LLVMValueRef alpha_pass = |
| 803 | lp_build_cmp(&bld_base->base, |
Christian König | a0dca44 | 2013-03-22 15:59:22 +0100 | [diff] [blame] | 804 | si_shader_ctx->shader->key.ps.alpha_func, |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 805 | LLVMBuildLoad(gallivm->builder, out_ptr[3], ""), |
Vadim Girlin | 453ea2d | 2013-10-13 19:53:54 +0400 | [diff] [blame] | 806 | alpha_ref); |
Michel Dänzer | 7708a86 | 2012-11-02 15:57:30 +0100 | [diff] [blame] | 807 | LLVMValueRef arg = |
| 808 | lp_build_select(&bld_base->base, |
| 809 | alpha_pass, |
| 810 | lp_build_const_float(gallivm, 1.0f), |
| 811 | lp_build_const_float(gallivm, -1.0f)); |
| 812 | |
| 813 | build_intrinsic(gallivm->builder, |
| 814 | "llvm.AMDGPU.kill", |
| 815 | LLVMVoidTypeInContext(gallivm->context), |
| 816 | &arg, 1, 0); |
| 817 | } else { |
| 818 | build_intrinsic(gallivm->builder, |
| 819 | "llvm.AMDGPU.kilp", |
| 820 | LLVMVoidTypeInContext(gallivm->context), |
| 821 | NULL, 0, 0); |
| 822 | } |
Marek Olšák | adc5797 | 2014-09-19 17:07:07 +0200 | [diff] [blame] | 823 | |
| 824 | si_shader_ctx->shader->db_shader_control |= S_02880C_KILL_ENABLE(1); |
Michel Dänzer | 7708a86 | 2012-11-02 15:57:30 +0100 | [diff] [blame] | 825 | } |
| 826 | |
Michel Dänzer | e3befbc | 2013-05-15 18:09:50 +0200 | [diff] [blame] | 827 | static void si_llvm_emit_clipvertex(struct lp_build_tgsi_context * bld_base, |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 828 | LLVMValueRef (*pos)[9], LLVMValueRef *out_elts) |
Michel Dänzer | e3befbc | 2013-05-15 18:09:50 +0200 | [diff] [blame] | 829 | { |
| 830 | struct si_shader_context *si_shader_ctx = si_shader_context(bld_base); |
| 831 | struct lp_build_context *base = &bld_base->base; |
| 832 | struct lp_build_context *uint = &si_shader_ctx->radeon_bld.soa.bld_base.uint_bld; |
Michel Dänzer | e3befbc | 2013-05-15 18:09:50 +0200 | [diff] [blame] | 833 | unsigned reg_index; |
| 834 | unsigned chan; |
| 835 | unsigned const_chan; |
Michel Dänzer | e3befbc | 2013-05-15 18:09:50 +0200 | [diff] [blame] | 836 | LLVMValueRef base_elt; |
| 837 | LLVMValueRef ptr = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_CONST); |
Marek Olšák | ee2a818 | 2014-07-07 23:27:19 +0200 | [diff] [blame] | 838 | LLVMValueRef constbuf_index = lp_build_const_int32(base->gallivm, SI_DRIVER_STATE_CONST_BUF); |
Marek Olšák | d787608 | 2014-09-26 23:06:32 +0200 | [diff] [blame] | 839 | LLVMValueRef const_resource = build_indexed_load_const(si_shader_ctx, ptr, constbuf_index); |
Michel Dänzer | e3befbc | 2013-05-15 18:09:50 +0200 | [diff] [blame] | 840 | |
Michel Dänzer | e3befbc | 2013-05-15 18:09:50 +0200 | [diff] [blame] | 841 | for (reg_index = 0; reg_index < 2; reg_index ++) { |
Michel Dänzer | b00269a | 2013-08-07 18:14:16 +0200 | [diff] [blame] | 842 | LLVMValueRef *args = pos[2 + reg_index]; |
| 843 | |
Michel Dänzer | e3befbc | 2013-05-15 18:09:50 +0200 | [diff] [blame] | 844 | args[5] = |
| 845 | args[6] = |
| 846 | args[7] = |
| 847 | args[8] = lp_build_const_float(base->gallivm, 0.0f); |
| 848 | |
| 849 | /* Compute dot products of position and user clip plane vectors */ |
| 850 | for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) { |
| 851 | for (const_chan = 0; const_chan < TGSI_NUM_CHANNELS; const_chan++) { |
Michel Dänzer | e3befbc | 2013-05-15 18:09:50 +0200 | [diff] [blame] | 852 | args[1] = lp_build_const_int32(base->gallivm, |
| 853 | ((reg_index * 4 + chan) * 4 + |
| 854 | const_chan) * 4); |
Marek Olšák | d787608 | 2014-09-26 23:06:32 +0200 | [diff] [blame] | 855 | base_elt = buffer_load_const(base->gallivm->builder, const_resource, |
Marek Olšák | 250aa93 | 2014-05-06 14:10:47 +0200 | [diff] [blame] | 856 | args[1], base->elem_type); |
Michel Dänzer | e3befbc | 2013-05-15 18:09:50 +0200 | [diff] [blame] | 857 | args[5 + chan] = |
| 858 | lp_build_add(base, args[5 + chan], |
| 859 | lp_build_mul(base, base_elt, |
| 860 | out_elts[const_chan])); |
| 861 | } |
| 862 | } |
| 863 | |
| 864 | args[0] = lp_build_const_int32(base->gallivm, 0xf); |
| 865 | args[1] = uint->zero; |
| 866 | args[2] = uint->zero; |
| 867 | args[3] = lp_build_const_int32(base->gallivm, |
| 868 | V_008DFC_SQ_EXP_POS + 2 + reg_index); |
| 869 | args[4] = uint->zero; |
Michel Dänzer | e3befbc | 2013-05-15 18:09:50 +0200 | [diff] [blame] | 870 | } |
| 871 | } |
| 872 | |
Marek Olšák | 8d03d92 | 2013-09-01 23:59:06 +0200 | [diff] [blame] | 873 | static void si_dump_streamout(struct pipe_stream_output_info *so) |
| 874 | { |
| 875 | unsigned i; |
| 876 | |
| 877 | if (so->num_outputs) |
| 878 | fprintf(stderr, "STREAMOUT\n"); |
| 879 | |
| 880 | for (i = 0; i < so->num_outputs; i++) { |
| 881 | unsigned mask = ((1 << so->output[i].num_components) - 1) << |
| 882 | so->output[i].start_component; |
| 883 | fprintf(stderr, " %i: BUF%i[%i..%i] <- OUT[%i].%s%s%s%s\n", |
| 884 | i, so->output[i].output_buffer, |
| 885 | so->output[i].dst_offset, so->output[i].dst_offset + so->output[i].num_components - 1, |
| 886 | so->output[i].register_index, |
| 887 | mask & 1 ? "x" : "", |
| 888 | mask & 2 ? "y" : "", |
| 889 | mask & 4 ? "z" : "", |
| 890 | mask & 8 ? "w" : ""); |
| 891 | } |
| 892 | } |
| 893 | |
| 894 | /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4. |
| 895 | * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2), |
| 896 | * or v4i32 (num_channels=3,4). */ |
| 897 | static void build_tbuffer_store(struct si_shader_context *shader, |
| 898 | LLVMValueRef rsrc, |
| 899 | LLVMValueRef vdata, |
| 900 | unsigned num_channels, |
| 901 | LLVMValueRef vaddr, |
| 902 | LLVMValueRef soffset, |
| 903 | unsigned inst_offset, |
| 904 | unsigned dfmt, |
| 905 | unsigned nfmt, |
| 906 | unsigned offen, |
| 907 | unsigned idxen, |
| 908 | unsigned glc, |
| 909 | unsigned slc, |
| 910 | unsigned tfe) |
| 911 | { |
| 912 | struct gallivm_state *gallivm = &shader->radeon_bld.gallivm; |
| 913 | LLVMTypeRef i32 = LLVMInt32TypeInContext(gallivm->context); |
| 914 | LLVMValueRef args[] = { |
| 915 | rsrc, |
| 916 | vdata, |
| 917 | LLVMConstInt(i32, num_channels, 0), |
| 918 | vaddr, |
| 919 | soffset, |
| 920 | LLVMConstInt(i32, inst_offset, 0), |
| 921 | LLVMConstInt(i32, dfmt, 0), |
| 922 | LLVMConstInt(i32, nfmt, 0), |
| 923 | LLVMConstInt(i32, offen, 0), |
| 924 | LLVMConstInt(i32, idxen, 0), |
| 925 | LLVMConstInt(i32, glc, 0), |
| 926 | LLVMConstInt(i32, slc, 0), |
| 927 | LLVMConstInt(i32, tfe, 0) |
| 928 | }; |
| 929 | |
Michel Dänzer | db9d6af | 2014-01-24 16:46:27 +0900 | [diff] [blame] | 930 | /* The instruction offset field has 12 bits */ |
| 931 | assert(offen || inst_offset < (1 << 12)); |
| 932 | |
Marek Olšák | 8d03d92 | 2013-09-01 23:59:06 +0200 | [diff] [blame] | 933 | /* The intrinsic is overloaded, we need to add a type suffix for overloading to work. */ |
| 934 | unsigned func = CLAMP(num_channels, 1, 3) - 1; |
| 935 | const char *types[] = {"i32", "v2i32", "v4i32"}; |
| 936 | char name[256]; |
| 937 | snprintf(name, sizeof(name), "llvm.SI.tbuffer.store.%s", types[func]); |
| 938 | |
| 939 | lp_build_intrinsic(gallivm->builder, name, |
| 940 | LLVMVoidTypeInContext(gallivm->context), |
| 941 | args, Elements(args)); |
| 942 | } |
| 943 | |
| 944 | static void build_streamout_store(struct si_shader_context *shader, |
| 945 | LLVMValueRef rsrc, |
| 946 | LLVMValueRef vdata, |
| 947 | unsigned num_channels, |
| 948 | LLVMValueRef vaddr, |
| 949 | LLVMValueRef soffset, |
| 950 | unsigned inst_offset) |
| 951 | { |
| 952 | static unsigned dfmt[] = { |
| 953 | V_008F0C_BUF_DATA_FORMAT_32, |
| 954 | V_008F0C_BUF_DATA_FORMAT_32_32, |
| 955 | V_008F0C_BUF_DATA_FORMAT_32_32_32, |
| 956 | V_008F0C_BUF_DATA_FORMAT_32_32_32_32 |
| 957 | }; |
| 958 | assert(num_channels >= 1 && num_channels <= 4); |
| 959 | |
| 960 | build_tbuffer_store(shader, rsrc, vdata, num_channels, vaddr, soffset, |
| 961 | inst_offset, dfmt[num_channels-1], |
| 962 | V_008F0C_BUF_NUM_FORMAT_UINT, 1, 0, 1, 1, 0); |
| 963 | } |
| 964 | |
| 965 | /* On SI, the vertex shader is responsible for writing streamout data |
| 966 | * to buffers. */ |
Michel Dänzer | 67e385b | 2014-01-08 17:48:21 +0900 | [diff] [blame] | 967 | static void si_llvm_emit_streamout(struct si_shader_context *shader, |
| 968 | struct si_shader_output_values *outputs, |
| 969 | unsigned noutput) |
Marek Olšák | 8d03d92 | 2013-09-01 23:59:06 +0200 | [diff] [blame] | 970 | { |
| 971 | struct pipe_stream_output_info *so = &shader->shader->selector->so; |
| 972 | struct gallivm_state *gallivm = &shader->radeon_bld.gallivm; |
| 973 | LLVMBuilderRef builder = gallivm->builder; |
| 974 | int i, j; |
| 975 | struct lp_build_if_state if_ctx; |
| 976 | |
| 977 | LLVMTypeRef i32 = LLVMInt32TypeInContext(gallivm->context); |
| 978 | |
| 979 | LLVMValueRef so_param = |
| 980 | LLVMGetParam(shader->radeon_bld.main_fn, |
| 981 | shader->param_streamout_config); |
| 982 | |
| 983 | /* Get bits [22:16], i.e. (so_param >> 16) & 127; */ |
| 984 | LLVMValueRef so_vtx_count = |
| 985 | LLVMBuildAnd(builder, |
| 986 | LLVMBuildLShr(builder, so_param, |
| 987 | LLVMConstInt(i32, 16, 0), ""), |
| 988 | LLVMConstInt(i32, 127, 0), ""); |
| 989 | |
| 990 | LLVMValueRef tid = build_intrinsic(builder, "llvm.SI.tid", i32, |
| 991 | NULL, 0, LLVMReadNoneAttribute); |
| 992 | |
| 993 | /* can_emit = tid < so_vtx_count; */ |
| 994 | LLVMValueRef can_emit = |
| 995 | LLVMBuildICmp(builder, LLVMIntULT, tid, so_vtx_count, ""); |
| 996 | |
| 997 | /* Emit the streamout code conditionally. This actually avoids |
| 998 | * out-of-bounds buffer access. The hw tells us via the SGPR |
| 999 | * (so_vtx_count) which threads are allowed to emit streamout data. */ |
| 1000 | lp_build_if(&if_ctx, gallivm, can_emit); |
| 1001 | { |
| 1002 | /* The buffer offset is computed as follows: |
| 1003 | * ByteOffset = streamout_offset[buffer_id]*4 + |
| 1004 | * (streamout_write_index + thread_id)*stride[buffer_id] + |
| 1005 | * attrib_offset |
| 1006 | */ |
| 1007 | |
| 1008 | LLVMValueRef so_write_index = |
| 1009 | LLVMGetParam(shader->radeon_bld.main_fn, |
| 1010 | shader->param_streamout_write_index); |
| 1011 | |
| 1012 | /* Compute (streamout_write_index + thread_id). */ |
| 1013 | so_write_index = LLVMBuildAdd(builder, so_write_index, tid, ""); |
| 1014 | |
| 1015 | /* Compute the write offset for each enabled buffer. */ |
| 1016 | LLVMValueRef so_write_offset[4] = {}; |
| 1017 | for (i = 0; i < 4; i++) { |
| 1018 | if (!so->stride[i]) |
| 1019 | continue; |
| 1020 | |
| 1021 | LLVMValueRef so_offset = LLVMGetParam(shader->radeon_bld.main_fn, |
| 1022 | shader->param_streamout_offset[i]); |
| 1023 | so_offset = LLVMBuildMul(builder, so_offset, LLVMConstInt(i32, 4, 0), ""); |
| 1024 | |
| 1025 | so_write_offset[i] = LLVMBuildMul(builder, so_write_index, |
| 1026 | LLVMConstInt(i32, so->stride[i]*4, 0), ""); |
| 1027 | so_write_offset[i] = LLVMBuildAdd(builder, so_write_offset[i], so_offset, ""); |
| 1028 | } |
| 1029 | |
Marek Olšák | 8d03d92 | 2013-09-01 23:59:06 +0200 | [diff] [blame] | 1030 | /* Write streamout data. */ |
| 1031 | for (i = 0; i < so->num_outputs; i++) { |
| 1032 | unsigned buf_idx = so->output[i].output_buffer; |
| 1033 | unsigned reg = so->output[i].register_index; |
| 1034 | unsigned start = so->output[i].start_component; |
| 1035 | unsigned num_comps = so->output[i].num_components; |
| 1036 | LLVMValueRef out[4]; |
| 1037 | |
| 1038 | assert(num_comps && num_comps <= 4); |
| 1039 | if (!num_comps || num_comps > 4) |
| 1040 | continue; |
| 1041 | |
Marek Olšák | 558f777 | 2014-10-04 22:37:23 +0200 | [diff] [blame] | 1042 | if (reg >= noutput) |
| 1043 | continue; |
| 1044 | |
Marek Olšák | 8d03d92 | 2013-09-01 23:59:06 +0200 | [diff] [blame] | 1045 | /* Load the output as int. */ |
| 1046 | for (j = 0; j < num_comps; j++) { |
Marek Olšák | 558f777 | 2014-10-04 22:37:23 +0200 | [diff] [blame] | 1047 | out[j] = LLVMBuildBitCast(builder, |
| 1048 | outputs[reg].values[start+j], |
| 1049 | i32, ""); |
Marek Olšák | 8d03d92 | 2013-09-01 23:59:06 +0200 | [diff] [blame] | 1050 | } |
| 1051 | |
| 1052 | /* Pack the output. */ |
| 1053 | LLVMValueRef vdata = NULL; |
| 1054 | |
| 1055 | switch (num_comps) { |
| 1056 | case 1: /* as i32 */ |
| 1057 | vdata = out[0]; |
| 1058 | break; |
| 1059 | case 2: /* as v2i32 */ |
| 1060 | case 3: /* as v4i32 (aligned to 4) */ |
| 1061 | case 4: /* as v4i32 */ |
| 1062 | vdata = LLVMGetUndef(LLVMVectorType(i32, util_next_power_of_two(num_comps))); |
| 1063 | for (j = 0; j < num_comps; j++) { |
| 1064 | vdata = LLVMBuildInsertElement(builder, vdata, out[j], |
| 1065 | LLVMConstInt(i32, j, 0), ""); |
| 1066 | } |
| 1067 | break; |
| 1068 | } |
| 1069 | |
| 1070 | build_streamout_store(shader, shader->so_buffers[buf_idx], |
| 1071 | vdata, num_comps, |
| 1072 | so_write_offset[buf_idx], |
| 1073 | LLVMConstInt(i32, 0, 0), |
| 1074 | so->output[i].dst_offset*4); |
| 1075 | } |
| 1076 | } |
| 1077 | lp_build_endif(&if_ctx); |
| 1078 | } |
| 1079 | |
Michel Dänzer | 7435d9f | 2013-12-04 13:37:07 +0900 | [diff] [blame] | 1080 | |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1081 | /* Generate export instructions for hardware VS shader stage */ |
| 1082 | static void si_llvm_export_vs(struct lp_build_tgsi_context *bld_base, |
| 1083 | struct si_shader_output_values *outputs, |
| 1084 | unsigned noutput) |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 1085 | { |
| 1086 | struct si_shader_context * si_shader_ctx = si_shader_context(bld_base); |
Marek Olšák | 8c37c16 | 2014-09-16 18:40:07 +0200 | [diff] [blame] | 1087 | struct si_shader * shader = si_shader_ctx->shader; |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 1088 | struct lp_build_context * base = &bld_base->base; |
| 1089 | struct lp_build_context * uint = |
| 1090 | &si_shader_ctx->radeon_bld.soa.bld_base.uint_bld; |
Michel Dänzer | 1a616c1 | 2012-11-13 17:35:09 +0100 | [diff] [blame] | 1091 | LLVMValueRef args[9]; |
Michel Dänzer | b00269a | 2013-08-07 18:14:16 +0200 | [diff] [blame] | 1092 | LLVMValueRef pos_args[4][9] = { { 0 } }; |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1093 | LLVMValueRef psize_value = NULL, edgeflag_value = NULL, layer_value = NULL; |
Marek Olšák | a9592cd | 2014-10-04 19:09:09 +0200 | [diff] [blame] | 1094 | unsigned semantic_name, semantic_index; |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1095 | unsigned target; |
Christian König | 3508815 | 2012-08-01 22:35:24 +0200 | [diff] [blame] | 1096 | unsigned param_count = 0; |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 1097 | unsigned pos_idx; |
Michel Dänzer | b00269a | 2013-08-07 18:14:16 +0200 | [diff] [blame] | 1098 | int i; |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 1099 | |
Michel Dänzer | 67e385b | 2014-01-08 17:48:21 +0900 | [diff] [blame] | 1100 | if (outputs && si_shader_ctx->shader->selector->so.num_outputs) { |
| 1101 | si_llvm_emit_streamout(si_shader_ctx, outputs, noutput); |
Marek Olšák | 8d03d92 | 2013-09-01 23:59:06 +0200 | [diff] [blame] | 1102 | } |
| 1103 | |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1104 | for (i = 0; i < noutput; i++) { |
| 1105 | semantic_name = outputs[i].name; |
Michel Dänzer | 67e385b | 2014-01-08 17:48:21 +0900 | [diff] [blame] | 1106 | semantic_index = outputs[i].sid; |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 1107 | |
Michel Dänzer | 0afeea5 | 2013-05-02 14:53:17 +0200 | [diff] [blame] | 1108 | handle_semantic: |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1109 | /* Select the correct target */ |
| 1110 | switch(semantic_name) { |
| 1111 | case TGSI_SEMANTIC_PSIZE: |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1112 | psize_value = outputs[i].values[0]; |
| 1113 | continue; |
| 1114 | case TGSI_SEMANTIC_EDGEFLAG: |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1115 | edgeflag_value = outputs[i].values[0]; |
| 1116 | continue; |
| 1117 | case TGSI_SEMANTIC_LAYER: |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1118 | layer_value = outputs[i].values[0]; |
| 1119 | continue; |
| 1120 | case TGSI_SEMANTIC_POSITION: |
| 1121 | target = V_008DFC_SQ_EXP_POS; |
| 1122 | break; |
| 1123 | case TGSI_SEMANTIC_COLOR: |
| 1124 | case TGSI_SEMANTIC_BCOLOR: |
| 1125 | target = V_008DFC_SQ_EXP_PARAM + param_count; |
Marek Olšák | 8b057dd | 2014-10-04 22:15:07 +0200 | [diff] [blame] | 1126 | shader->vs_output_param_offset[i] = param_count; |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1127 | param_count++; |
| 1128 | break; |
| 1129 | case TGSI_SEMANTIC_CLIPDIST: |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1130 | target = V_008DFC_SQ_EXP_POS + 2 + semantic_index; |
| 1131 | break; |
| 1132 | case TGSI_SEMANTIC_CLIPVERTEX: |
| 1133 | si_llvm_emit_clipvertex(bld_base, pos_args, outputs[i].values); |
| 1134 | continue; |
Michel Dänzer | d8b3d80 | 2014-01-09 12:55:26 +0900 | [diff] [blame] | 1135 | case TGSI_SEMANTIC_PRIMID: |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1136 | case TGSI_SEMANTIC_FOG: |
| 1137 | case TGSI_SEMANTIC_GENERIC: |
| 1138 | target = V_008DFC_SQ_EXP_PARAM + param_count; |
Marek Olšák | 8b057dd | 2014-10-04 22:15:07 +0200 | [diff] [blame] | 1139 | shader->vs_output_param_offset[i] = param_count; |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1140 | param_count++; |
| 1141 | break; |
| 1142 | default: |
| 1143 | target = 0; |
| 1144 | fprintf(stderr, |
| 1145 | "Warning: SI unhandled vs output type:%d\n", |
| 1146 | semantic_name); |
| 1147 | } |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 1148 | |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1149 | si_llvm_init_export_args(bld_base, outputs[i].values, target, args); |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 1150 | |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1151 | if (target >= V_008DFC_SQ_EXP_POS && |
| 1152 | target <= (V_008DFC_SQ_EXP_POS + 3)) { |
| 1153 | memcpy(pos_args[target - V_008DFC_SQ_EXP_POS], |
| 1154 | args, sizeof(args)); |
| 1155 | } else { |
| 1156 | lp_build_intrinsic(base->gallivm->builder, |
| 1157 | "llvm.SI.export", |
| 1158 | LLVMVoidTypeInContext(base->gallivm->context), |
| 1159 | args, 9); |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 1160 | } |
| 1161 | |
| 1162 | if (semantic_name == TGSI_SEMANTIC_CLIPDIST) { |
| 1163 | semantic_name = TGSI_SEMANTIC_GENERIC; |
| 1164 | goto handle_semantic; |
| 1165 | } |
| 1166 | } |
| 1167 | |
| 1168 | /* We need to add the position output manually if it's missing. */ |
| 1169 | if (!pos_args[0][0]) { |
| 1170 | pos_args[0][0] = lp_build_const_int32(base->gallivm, 0xf); /* writemask */ |
| 1171 | pos_args[0][1] = uint->zero; /* EXEC mask */ |
| 1172 | pos_args[0][2] = uint->zero; /* last export? */ |
| 1173 | pos_args[0][3] = lp_build_const_int32(base->gallivm, V_008DFC_SQ_EXP_POS); |
| 1174 | pos_args[0][4] = uint->zero; /* COMPR flag */ |
| 1175 | pos_args[0][5] = base->zero; /* X */ |
| 1176 | pos_args[0][6] = base->zero; /* Y */ |
| 1177 | pos_args[0][7] = base->zero; /* Z */ |
| 1178 | pos_args[0][8] = base->one; /* W */ |
| 1179 | } |
| 1180 | |
| 1181 | /* Write the misc vector (point size, edgeflag, layer, viewport). */ |
Marek Olšák | 1615347 | 2014-12-07 19:30:08 +0100 | [diff] [blame] | 1182 | if (shader->selector->info.writes_psize || |
| 1183 | shader->selector->info.writes_edgeflag || |
| 1184 | shader->selector->info.writes_layer) { |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 1185 | pos_args[1][0] = lp_build_const_int32(base->gallivm, /* writemask */ |
Marek Olšák | 1615347 | 2014-12-07 19:30:08 +0100 | [diff] [blame] | 1186 | shader->selector->info.writes_psize | |
| 1187 | (shader->selector->info.writes_edgeflag << 1) | |
| 1188 | (shader->selector->info.writes_layer << 2)); |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 1189 | pos_args[1][1] = uint->zero; /* EXEC mask */ |
| 1190 | pos_args[1][2] = uint->zero; /* last export? */ |
| 1191 | pos_args[1][3] = lp_build_const_int32(base->gallivm, V_008DFC_SQ_EXP_POS + 1); |
| 1192 | pos_args[1][4] = uint->zero; /* COMPR flag */ |
| 1193 | pos_args[1][5] = base->zero; /* X */ |
| 1194 | pos_args[1][6] = base->zero; /* Y */ |
| 1195 | pos_args[1][7] = base->zero; /* Z */ |
| 1196 | pos_args[1][8] = base->zero; /* W */ |
| 1197 | |
Marek Olšák | 1615347 | 2014-12-07 19:30:08 +0100 | [diff] [blame] | 1198 | if (shader->selector->info.writes_psize) |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1199 | pos_args[1][5] = psize_value; |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 1200 | |
Marek Olšák | 1615347 | 2014-12-07 19:30:08 +0100 | [diff] [blame] | 1201 | if (shader->selector->info.writes_edgeflag) { |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 1202 | /* The output is a float, but the hw expects an integer |
| 1203 | * with the first bit containing the edge flag. */ |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1204 | edgeflag_value = LLVMBuildFPToUI(base->gallivm->builder, |
| 1205 | edgeflag_value, |
| 1206 | bld_base->uint_bld.elem_type, ""); |
| 1207 | edgeflag_value = lp_build_min(&bld_base->int_bld, |
| 1208 | edgeflag_value, |
| 1209 | bld_base->int_bld.one); |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 1210 | |
| 1211 | /* The LLVM intrinsic expects a float. */ |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1212 | pos_args[1][6] = LLVMBuildBitCast(base->gallivm->builder, |
| 1213 | edgeflag_value, |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 1214 | base->elem_type, ""); |
| 1215 | } |
| 1216 | |
Marek Olšák | 1615347 | 2014-12-07 19:30:08 +0100 | [diff] [blame] | 1217 | if (shader->selector->info.writes_layer) |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1218 | pos_args[1][7] = layer_value; |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 1219 | } |
| 1220 | |
| 1221 | for (i = 0; i < 4; i++) |
| 1222 | if (pos_args[i][0]) |
| 1223 | shader->nr_pos_exports++; |
| 1224 | |
| 1225 | pos_idx = 0; |
| 1226 | for (i = 0; i < 4; i++) { |
| 1227 | if (!pos_args[i][0]) |
| 1228 | continue; |
| 1229 | |
| 1230 | /* Specify the target we are exporting */ |
| 1231 | pos_args[i][3] = lp_build_const_int32(base->gallivm, V_008DFC_SQ_EXP_POS + pos_idx++); |
| 1232 | |
| 1233 | if (pos_idx == shader->nr_pos_exports) |
| 1234 | /* Specify that this is the last export */ |
| 1235 | pos_args[i][2] = uint->one; |
| 1236 | |
| 1237 | lp_build_intrinsic(base->gallivm->builder, |
| 1238 | "llvm.SI.export", |
| 1239 | LLVMVoidTypeInContext(base->gallivm->context), |
| 1240 | pos_args[i], 9); |
| 1241 | } |
| 1242 | } |
| 1243 | |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1244 | static void si_llvm_emit_es_epilogue(struct lp_build_tgsi_context * bld_base) |
| 1245 | { |
| 1246 | struct si_shader_context *si_shader_ctx = si_shader_context(bld_base); |
| 1247 | struct gallivm_state *gallivm = bld_base->base.gallivm; |
Marek Olšák | 8c37c16 | 2014-09-16 18:40:07 +0200 | [diff] [blame] | 1248 | struct si_shader *es = si_shader_ctx->shader; |
Marek Olšák | 216cf86 | 2014-10-04 17:04:05 +0200 | [diff] [blame] | 1249 | struct tgsi_shader_info *info = &es->selector->info; |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1250 | LLVMTypeRef i32 = LLVMInt32TypeInContext(gallivm->context); |
Michel Dänzer | db9d6af | 2014-01-24 16:46:27 +0900 | [diff] [blame] | 1251 | LLVMValueRef soffset = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, |
| 1252 | SI_PARAM_ES2GS_OFFSET); |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1253 | unsigned chan; |
| 1254 | int i; |
| 1255 | |
Marek Olšák | 216cf86 | 2014-10-04 17:04:05 +0200 | [diff] [blame] | 1256 | for (i = 0; i < info->num_outputs; i++) { |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1257 | LLVMValueRef *out_ptr = |
Marek Olšák | 216cf86 | 2014-10-04 17:04:05 +0200 | [diff] [blame] | 1258 | si_shader_ctx->radeon_bld.soa.outputs[i]; |
| 1259 | int param_index = get_param_index(info->output_semantic_name[i], |
| 1260 | info->output_semantic_index[i], |
Marek Olšák | e29353f | 2014-09-17 22:17:02 +0200 | [diff] [blame] | 1261 | es->key.vs.gs_used_inputs); |
Michel Dänzer | e884c56 | 2014-01-15 15:24:14 +0900 | [diff] [blame] | 1262 | |
Marek Olšák | e29353f | 2014-09-17 22:17:02 +0200 | [diff] [blame] | 1263 | if (param_index < 0) |
Michel Dänzer | e884c56 | 2014-01-15 15:24:14 +0900 | [diff] [blame] | 1264 | continue; |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1265 | |
| 1266 | for (chan = 0; chan < 4; chan++) { |
| 1267 | LLVMValueRef out_val = LLVMBuildLoad(gallivm->builder, out_ptr[chan], ""); |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1268 | out_val = LLVMBuildBitCast(gallivm->builder, out_val, i32, ""); |
| 1269 | |
Marek Olšák | fc3b335 | 2014-10-05 13:33:40 +0200 | [diff] [blame] | 1270 | build_tbuffer_store(si_shader_ctx, |
| 1271 | si_shader_ctx->esgs_ring, |
| 1272 | out_val, 1, |
Michel Dänzer | db9d6af | 2014-01-24 16:46:27 +0900 | [diff] [blame] | 1273 | LLVMGetUndef(i32), soffset, |
Marek Olšák | e29353f | 2014-09-17 22:17:02 +0200 | [diff] [blame] | 1274 | (4 * param_index + chan) * 4, |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1275 | V_008F0C_BUF_DATA_FORMAT_32, |
| 1276 | V_008F0C_BUF_NUM_FORMAT_UINT, |
Michel Dänzer | db9d6af | 2014-01-24 16:46:27 +0900 | [diff] [blame] | 1277 | 0, 0, 1, 1, 0); |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1278 | } |
| 1279 | } |
| 1280 | } |
| 1281 | |
| 1282 | static void si_llvm_emit_gs_epilogue(struct lp_build_tgsi_context *bld_base) |
| 1283 | { |
| 1284 | struct si_shader_context *si_shader_ctx = si_shader_context(bld_base); |
| 1285 | struct gallivm_state *gallivm = bld_base->base.gallivm; |
| 1286 | LLVMValueRef args[2]; |
| 1287 | |
| 1288 | args[0] = lp_build_const_int32(gallivm, SENDMSG_GS_OP_NOP | SENDMSG_GS_DONE); |
| 1289 | args[1] = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_GS_WAVE_ID); |
| 1290 | build_intrinsic(gallivm->builder, "llvm.SI.sendmsg", |
| 1291 | LLVMVoidTypeInContext(gallivm->context), args, 2, |
| 1292 | LLVMNoUnwindAttribute); |
| 1293 | } |
| 1294 | |
| 1295 | static void si_llvm_emit_vs_epilogue(struct lp_build_tgsi_context * bld_base) |
| 1296 | { |
| 1297 | struct si_shader_context *si_shader_ctx = si_shader_context(bld_base); |
| 1298 | struct gallivm_state *gallivm = bld_base->base.gallivm; |
Marek Olšák | ec0d168 | 2014-10-04 22:33:36 +0200 | [diff] [blame] | 1299 | struct tgsi_shader_info *info = &si_shader_ctx->shader->selector->info; |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1300 | struct si_shader_output_values *outputs = NULL; |
Marek Olšák | ec0d168 | 2014-10-04 22:33:36 +0200 | [diff] [blame] | 1301 | int i,j; |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1302 | |
Marek Olšák | ec0d168 | 2014-10-04 22:33:36 +0200 | [diff] [blame] | 1303 | outputs = MALLOC(info->num_outputs * sizeof(outputs[0])); |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1304 | |
Marek Olšák | ec0d168 | 2014-10-04 22:33:36 +0200 | [diff] [blame] | 1305 | for (i = 0; i < info->num_outputs; i++) { |
| 1306 | outputs[i].name = info->output_semantic_name[i]; |
| 1307 | outputs[i].sid = info->output_semantic_index[i]; |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1308 | |
Marek Olšák | ec0d168 | 2014-10-04 22:33:36 +0200 | [diff] [blame] | 1309 | for (j = 0; j < 4; j++) |
| 1310 | outputs[i].values[j] = |
| 1311 | LLVMBuildLoad(gallivm->builder, |
| 1312 | si_shader_ctx->radeon_bld.soa.outputs[i][j], |
| 1313 | ""); |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1314 | } |
| 1315 | |
Marek Olšák | ec0d168 | 2014-10-04 22:33:36 +0200 | [diff] [blame] | 1316 | si_llvm_export_vs(bld_base, outputs, info->num_outputs); |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1317 | FREE(outputs); |
| 1318 | } |
| 1319 | |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 1320 | static void si_llvm_emit_fs_epilogue(struct lp_build_tgsi_context * bld_base) |
| 1321 | { |
| 1322 | struct si_shader_context * si_shader_ctx = si_shader_context(bld_base); |
Marek Olšák | 8c37c16 | 2014-09-16 18:40:07 +0200 | [diff] [blame] | 1323 | struct si_shader * shader = si_shader_ctx->shader; |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 1324 | struct lp_build_context * base = &bld_base->base; |
Marek Olšák | 75e97e2 | 2014-10-04 23:13:50 +0200 | [diff] [blame] | 1325 | struct lp_build_context * uint = &bld_base->uint_bld; |
| 1326 | struct tgsi_shader_info *info = &shader->selector->info; |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 1327 | LLVMValueRef args[9]; |
| 1328 | LLVMValueRef last_args[9] = { 0 }; |
Marek Olšák | d0e8b65 | 2014-05-06 20:04:31 +0200 | [diff] [blame] | 1329 | int depth_index = -1, stencil_index = -1, samplemask_index = -1; |
Marek Olšák | 75e97e2 | 2014-10-04 23:13:50 +0200 | [diff] [blame] | 1330 | int i; |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 1331 | |
Marek Olšák | 75e97e2 | 2014-10-04 23:13:50 +0200 | [diff] [blame] | 1332 | for (i = 0; i < info->num_outputs; i++) { |
| 1333 | unsigned semantic_name = info->output_semantic_name[i]; |
| 1334 | unsigned semantic_index = info->output_semantic_index[i]; |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 1335 | unsigned target; |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 1336 | |
Marek Olšák | 75e97e2 | 2014-10-04 23:13:50 +0200 | [diff] [blame] | 1337 | /* Select the correct target */ |
| 1338 | switch (semantic_name) { |
| 1339 | case TGSI_SEMANTIC_POSITION: |
| 1340 | depth_index = i; |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 1341 | continue; |
Marek Olšák | 75e97e2 | 2014-10-04 23:13:50 +0200 | [diff] [blame] | 1342 | case TGSI_SEMANTIC_STENCIL: |
| 1343 | stencil_index = i; |
| 1344 | continue; |
| 1345 | case TGSI_SEMANTIC_SAMPLEMASK: |
| 1346 | samplemask_index = i; |
| 1347 | continue; |
| 1348 | case TGSI_SEMANTIC_COLOR: |
| 1349 | target = V_008DFC_SQ_EXP_MRT + semantic_index; |
| 1350 | if (si_shader_ctx->shader->key.ps.alpha_to_one) |
| 1351 | LLVMBuildStore(bld_base->base.gallivm->builder, |
| 1352 | bld_base->base.one, |
| 1353 | si_shader_ctx->radeon_bld.soa.outputs[i][3]); |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 1354 | |
Marek Olšák | 75e97e2 | 2014-10-04 23:13:50 +0200 | [diff] [blame] | 1355 | if (semantic_index == 0 && |
| 1356 | si_shader_ctx->shader->key.ps.alpha_func != PIPE_FUNC_ALWAYS) |
| 1357 | si_alpha_test(bld_base, |
| 1358 | si_shader_ctx->radeon_bld.soa.outputs[i]); |
| 1359 | break; |
| 1360 | default: |
| 1361 | target = 0; |
| 1362 | fprintf(stderr, |
| 1363 | "Warning: SI unhandled fs output type:%d\n", |
| 1364 | semantic_name); |
| 1365 | } |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 1366 | |
Marek Olšák | 75e97e2 | 2014-10-04 23:13:50 +0200 | [diff] [blame] | 1367 | si_llvm_init_export_args_load(bld_base, |
| 1368 | si_shader_ctx->radeon_bld.soa.outputs[i], |
| 1369 | target, args); |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 1370 | |
Marek Olšák | 75e97e2 | 2014-10-04 23:13:50 +0200 | [diff] [blame] | 1371 | if (semantic_name == TGSI_SEMANTIC_COLOR) { |
| 1372 | /* If there is an export instruction waiting to be emitted, do so now. */ |
| 1373 | if (last_args[0]) { |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 1374 | lp_build_intrinsic(base->gallivm->builder, |
| 1375 | "llvm.SI.export", |
| 1376 | LLVMVoidTypeInContext(base->gallivm->context), |
Marek Olšák | 75e97e2 | 2014-10-04 23:13:50 +0200 | [diff] [blame] | 1377 | last_args, 9); |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 1378 | } |
Marek Olšák | 75e97e2 | 2014-10-04 23:13:50 +0200 | [diff] [blame] | 1379 | |
| 1380 | /* This instruction will be emitted at the end of the shader. */ |
| 1381 | memcpy(last_args, args, sizeof(args)); |
| 1382 | |
| 1383 | /* Handle FS_COLOR0_WRITES_ALL_CBUFS. */ |
| 1384 | if (shader->selector->info.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS] && |
| 1385 | semantic_index == 0 && |
| 1386 | si_shader_ctx->shader->key.ps.last_cbuf > 0) { |
| 1387 | for (int c = 1; c <= si_shader_ctx->shader->key.ps.last_cbuf; c++) { |
| 1388 | si_llvm_init_export_args_load(bld_base, |
| 1389 | si_shader_ctx->radeon_bld.soa.outputs[i], |
| 1390 | V_008DFC_SQ_EXP_MRT + c, args); |
| 1391 | lp_build_intrinsic(base->gallivm->builder, |
| 1392 | "llvm.SI.export", |
| 1393 | LLVMVoidTypeInContext(base->gallivm->context), |
| 1394 | args, 9); |
| 1395 | } |
| 1396 | } |
| 1397 | } else { |
| 1398 | lp_build_intrinsic(base->gallivm->builder, |
| 1399 | "llvm.SI.export", |
| 1400 | LLVMVoidTypeInContext(base->gallivm->context), |
| 1401 | args, 9); |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 1402 | } |
| 1403 | } |
| 1404 | |
Marek Olšák | d0e8b65 | 2014-05-06 20:04:31 +0200 | [diff] [blame] | 1405 | if (depth_index >= 0 || stencil_index >= 0 || samplemask_index >= 0) { |
Michel Dänzer | 1a616c1 | 2012-11-13 17:35:09 +0100 | [diff] [blame] | 1406 | LLVMValueRef out_ptr; |
| 1407 | unsigned mask = 0; |
| 1408 | |
| 1409 | /* Specify the target we are exporting */ |
| 1410 | args[3] = lp_build_const_int32(base->gallivm, V_008DFC_SQ_EXP_MRTZ); |
| 1411 | |
Marek Olšák | 9baaa5d | 2014-05-06 19:55:48 +0200 | [diff] [blame] | 1412 | args[5] = base->zero; /* R, depth */ |
| 1413 | args[6] = base->zero; /* G, stencil test value[0:7], stencil op value[8:15] */ |
| 1414 | args[7] = base->zero; /* B, sample mask */ |
| 1415 | args[8] = base->zero; /* A, alpha to mask */ |
| 1416 | |
Michel Dänzer | 1a616c1 | 2012-11-13 17:35:09 +0100 | [diff] [blame] | 1417 | if (depth_index >= 0) { |
| 1418 | out_ptr = si_shader_ctx->radeon_bld.soa.outputs[depth_index][2]; |
| 1419 | args[5] = LLVMBuildLoad(base->gallivm->builder, out_ptr, ""); |
| 1420 | mask |= 0x1; |
Marek Olšák | d9e102b | 2014-05-06 19:59:53 +0200 | [diff] [blame] | 1421 | si_shader_ctx->shader->db_shader_control |= S_02880C_Z_EXPORT_ENABLE(1); |
Michel Dänzer | 1a616c1 | 2012-11-13 17:35:09 +0100 | [diff] [blame] | 1422 | } |
| 1423 | |
| 1424 | if (stencil_index >= 0) { |
| 1425 | out_ptr = si_shader_ctx->radeon_bld.soa.outputs[stencil_index][1]; |
Michel Dänzer | 1a616c1 | 2012-11-13 17:35:09 +0100 | [diff] [blame] | 1426 | args[6] = LLVMBuildLoad(base->gallivm->builder, out_ptr, ""); |
Marek Olšák | 368b0a7 | 2014-12-28 21:51:35 +0100 | [diff] [blame] | 1427 | mask |= 0x2; |
Marek Olšák | d9e102b | 2014-05-06 19:59:53 +0200 | [diff] [blame] | 1428 | si_shader_ctx->shader->db_shader_control |= |
| 1429 | S_02880C_STENCIL_TEST_VAL_EXPORT_ENABLE(1); |
Michel Dänzer | 1a616c1 | 2012-11-13 17:35:09 +0100 | [diff] [blame] | 1430 | } |
| 1431 | |
Marek Olšák | d0e8b65 | 2014-05-06 20:04:31 +0200 | [diff] [blame] | 1432 | if (samplemask_index >= 0) { |
| 1433 | out_ptr = si_shader_ctx->radeon_bld.soa.outputs[samplemask_index][0]; |
| 1434 | args[7] = LLVMBuildLoad(base->gallivm->builder, out_ptr, ""); |
Marek Olšák | 368b0a7 | 2014-12-28 21:51:35 +0100 | [diff] [blame] | 1435 | mask |= 0x4; |
Marek Olšák | d0e8b65 | 2014-05-06 20:04:31 +0200 | [diff] [blame] | 1436 | si_shader_ctx->shader->db_shader_control |= S_02880C_MASK_EXPORT_ENABLE(1); |
| 1437 | } |
| 1438 | |
Marek Olšák | 368b0a7 | 2014-12-28 21:51:35 +0100 | [diff] [blame] | 1439 | /* SI (except OLAND) has a bug that it only looks |
| 1440 | * at the X writemask component. */ |
| 1441 | if (si_shader_ctx->screen->b.chip_class == SI && |
| 1442 | si_shader_ctx->screen->b.family != CHIP_OLAND) |
| 1443 | mask |= 0x1; |
| 1444 | |
Marek Olšák | d0e8b65 | 2014-05-06 20:04:31 +0200 | [diff] [blame] | 1445 | if (samplemask_index >= 0) |
| 1446 | si_shader_ctx->shader->spi_shader_z_format = V_028710_SPI_SHADER_32_ABGR; |
| 1447 | else if (stencil_index >= 0) |
Marek Olšák | d9e102b | 2014-05-06 19:59:53 +0200 | [diff] [blame] | 1448 | si_shader_ctx->shader->spi_shader_z_format = V_028710_SPI_SHADER_32_GR; |
| 1449 | else |
| 1450 | si_shader_ctx->shader->spi_shader_z_format = V_028710_SPI_SHADER_32_R; |
| 1451 | |
Michel Dänzer | 1a616c1 | 2012-11-13 17:35:09 +0100 | [diff] [blame] | 1452 | /* Specify which components to enable */ |
| 1453 | args[0] = lp_build_const_int32(base->gallivm, mask); |
| 1454 | |
| 1455 | args[1] = |
| 1456 | args[2] = |
| 1457 | args[4] = uint->zero; |
| 1458 | |
| 1459 | if (last_args[0]) |
| 1460 | lp_build_intrinsic(base->gallivm->builder, |
| 1461 | "llvm.SI.export", |
| 1462 | LLVMVoidTypeInContext(base->gallivm->context), |
| 1463 | args, 9); |
| 1464 | else |
| 1465 | memcpy(last_args, args, sizeof(args)); |
| 1466 | } |
| 1467 | |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 1468 | if (!last_args[0]) { |
| 1469 | /* Specify which components to enable */ |
| 1470 | last_args[0] = lp_build_const_int32(base->gallivm, 0x0); |
Christian König | f18fd25 | 2012-07-25 21:58:46 +0200 | [diff] [blame] | 1471 | |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 1472 | /* Specify the target we are exporting */ |
| 1473 | last_args[3] = lp_build_const_int32(base->gallivm, V_008DFC_SQ_EXP_MRT); |
Marek Olšák | 48784f3 | 2013-10-23 16:10:38 +0200 | [diff] [blame] | 1474 | |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 1475 | /* Set COMPR flag to zero to export data as 32-bit */ |
| 1476 | last_args[4] = uint->zero; |
Marek Olšák | 053606d | 2013-11-19 22:07:30 +0100 | [diff] [blame] | 1477 | |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 1478 | /* dummy bits */ |
| 1479 | last_args[5]= uint->zero; |
| 1480 | last_args[6]= uint->zero; |
| 1481 | last_args[7]= uint->zero; |
| 1482 | last_args[8]= uint->zero; |
Michel Dänzer | c840270 | 2013-02-12 18:37:22 +0100 | [diff] [blame] | 1483 | } |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 1484 | |
| 1485 | /* Specify whether the EXEC mask represents the valid mask */ |
| 1486 | last_args[1] = uint->one; |
| 1487 | |
| 1488 | /* Specify that this is the last export */ |
| 1489 | last_args[2] = lp_build_const_int32(base->gallivm, 1); |
| 1490 | |
| 1491 | lp_build_intrinsic(base->gallivm->builder, |
| 1492 | "llvm.SI.export", |
| 1493 | LLVMVoidTypeInContext(base->gallivm->context), |
| 1494 | last_args, 9); |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 1495 | } |
| 1496 | |
Marek Olšák | 4855acd | 2013-08-06 15:08:54 +0200 | [diff] [blame] | 1497 | static void build_tex_intrinsic(const struct lp_build_tgsi_action * action, |
| 1498 | struct lp_build_tgsi_context * bld_base, |
| 1499 | struct lp_build_emit_data * emit_data); |
| 1500 | |
Marek Olšák | c7b5a5c | 2014-06-06 03:00:18 +0200 | [diff] [blame] | 1501 | static bool tgsi_is_shadow_sampler(unsigned target) |
| 1502 | { |
| 1503 | return target == TGSI_TEXTURE_SHADOW1D || |
| 1504 | target == TGSI_TEXTURE_SHADOW1D_ARRAY || |
| 1505 | target == TGSI_TEXTURE_SHADOW2D || |
| 1506 | target == TGSI_TEXTURE_SHADOW2D_ARRAY || |
| 1507 | target == TGSI_TEXTURE_SHADOWCUBE || |
| 1508 | target == TGSI_TEXTURE_SHADOWCUBE_ARRAY || |
| 1509 | target == TGSI_TEXTURE_SHADOWRECT; |
| 1510 | } |
| 1511 | |
Marek Olšák | 877bb52 | 2014-06-25 03:12:46 +0200 | [diff] [blame] | 1512 | static const struct lp_build_tgsi_action tex_action; |
| 1513 | |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 1514 | static void tex_fetch_args( |
| 1515 | struct lp_build_tgsi_context * bld_base, |
| 1516 | struct lp_build_emit_data * emit_data) |
| 1517 | { |
Christian König | 55fe5cc | 2013-03-04 16:30:06 +0100 | [diff] [blame] | 1518 | struct si_shader_context *si_shader_ctx = si_shader_context(bld_base); |
Michel Dänzer | e5fb734 | 2013-01-24 18:54:51 +0100 | [diff] [blame] | 1519 | struct gallivm_state *gallivm = bld_base->base.gallivm; |
Michel Dänzer | c2bae6b | 2012-08-02 17:19:22 +0200 | [diff] [blame] | 1520 | const struct tgsi_full_instruction * inst = emit_data->inst; |
Michel Dänzer | 120efee | 2013-01-25 12:10:11 +0100 | [diff] [blame] | 1521 | unsigned opcode = inst->Instruction.Opcode; |
| 1522 | unsigned target = inst->Texture.Texture; |
Michel Dänzer | 120efee | 2013-01-25 12:10:11 +0100 | [diff] [blame] | 1523 | LLVMValueRef coords[4]; |
| 1524 | LLVMValueRef address[16]; |
Michel Dänzer | beaa5eb | 2013-05-24 13:23:26 +0100 | [diff] [blame] | 1525 | int ref_pos; |
| 1526 | unsigned num_coords = tgsi_util_get_texture_coord_dim(target, &ref_pos); |
Michel Dänzer | 120efee | 2013-01-25 12:10:11 +0100 | [diff] [blame] | 1527 | unsigned count = 0; |
Michel Dänzer | e5fb734 | 2013-01-24 18:54:51 +0100 | [diff] [blame] | 1528 | unsigned chan; |
Marek Olšák | dbeedbb | 2013-10-31 15:08:49 +0100 | [diff] [blame] | 1529 | unsigned sampler_src = emit_data->inst->Instruction.NumSrcRegs - 1; |
| 1530 | unsigned sampler_index = emit_data->inst->Src[sampler_src].Register.Index; |
Marek Olšák | 877bb52 | 2014-06-25 03:12:46 +0200 | [diff] [blame] | 1531 | bool has_offset = HAVE_LLVM >= 0x0305 ? inst->Texture.NumOffsets > 0 : false; |
Marek Olšák | dbeedbb | 2013-10-31 15:08:49 +0100 | [diff] [blame] | 1532 | |
| 1533 | if (target == TGSI_TEXTURE_BUFFER) { |
| 1534 | LLVMTypeRef i128 = LLVMIntTypeInContext(gallivm->context, 128); |
| 1535 | LLVMTypeRef v2i128 = LLVMVectorType(i128, 2); |
| 1536 | LLVMTypeRef i8 = LLVMInt8TypeInContext(gallivm->context); |
| 1537 | LLVMTypeRef v16i8 = LLVMVectorType(i8, 16); |
| 1538 | |
Marek Olšák | 4f3f043 | 2014-07-07 22:49:15 +0200 | [diff] [blame] | 1539 | /* Bitcast and truncate v8i32 to v16i8. */ |
Marek Olšák | dbeedbb | 2013-10-31 15:08:49 +0100 | [diff] [blame] | 1540 | LLVMValueRef res = si_shader_ctx->resources[sampler_index]; |
| 1541 | res = LLVMBuildBitCast(gallivm->builder, res, v2i128, ""); |
Marek Olšák | b142dd2 | 2015-02-01 16:58:08 +0100 | [diff] [blame] | 1542 | res = LLVMBuildExtractElement(gallivm->builder, res, bld_base->uint_bld.one, ""); |
Marek Olšák | dbeedbb | 2013-10-31 15:08:49 +0100 | [diff] [blame] | 1543 | res = LLVMBuildBitCast(gallivm->builder, res, v16i8, ""); |
| 1544 | |
| 1545 | emit_data->dst_type = LLVMVectorType(bld_base->base.elem_type, 4); |
| 1546 | emit_data->args[0] = res; |
| 1547 | emit_data->args[1] = bld_base->uint_bld.zero; |
| 1548 | emit_data->args[2] = lp_build_emit_fetch(bld_base, emit_data->inst, 0, 0); |
| 1549 | emit_data->arg_count = 3; |
| 1550 | return; |
| 1551 | } |
Tom Stellard | 467f516 | 2012-05-16 15:15:35 -0400 | [diff] [blame] | 1552 | |
Michel Dänzer | 120efee | 2013-01-25 12:10:11 +0100 | [diff] [blame] | 1553 | /* Fetch and project texture coordinates */ |
| 1554 | coords[3] = lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_W); |
Michel Dänzer | e5fb734 | 2013-01-24 18:54:51 +0100 | [diff] [blame] | 1555 | for (chan = 0; chan < 3; chan++ ) { |
| 1556 | coords[chan] = lp_build_emit_fetch(bld_base, |
| 1557 | emit_data->inst, 0, |
| 1558 | chan); |
Michel Dänzer | 120efee | 2013-01-25 12:10:11 +0100 | [diff] [blame] | 1559 | if (opcode == TGSI_OPCODE_TXP) |
Michel Dänzer | c2bae6b | 2012-08-02 17:19:22 +0200 | [diff] [blame] | 1560 | coords[chan] = lp_build_emit_llvm_binary(bld_base, |
| 1561 | TGSI_OPCODE_DIV, |
Michel Dänzer | e5fb734 | 2013-01-24 18:54:51 +0100 | [diff] [blame] | 1562 | coords[chan], |
| 1563 | coords[3]); |
| 1564 | } |
| 1565 | |
Michel Dänzer | 120efee | 2013-01-25 12:10:11 +0100 | [diff] [blame] | 1566 | if (opcode == TGSI_OPCODE_TXP) |
| 1567 | coords[3] = bld_base->base.one; |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 1568 | |
Marek Olšák | c7b5a5c | 2014-06-06 03:00:18 +0200 | [diff] [blame] | 1569 | /* Pack offsets. */ |
Marek Olšák | 877bb52 | 2014-06-25 03:12:46 +0200 | [diff] [blame] | 1570 | if (has_offset && opcode != TGSI_OPCODE_TXF) { |
Marek Olšák | c7b5a5c | 2014-06-06 03:00:18 +0200 | [diff] [blame] | 1571 | /* The offsets are six-bit signed integers packed like this: |
| 1572 | * X=[5:0], Y=[13:8], and Z=[21:16]. |
| 1573 | */ |
| 1574 | LLVMValueRef offset[3], pack; |
| 1575 | |
| 1576 | assert(inst->Texture.NumOffsets == 1); |
| 1577 | |
| 1578 | for (chan = 0; chan < 3; chan++) { |
| 1579 | offset[chan] = lp_build_emit_fetch_texoffset(bld_base, |
| 1580 | emit_data->inst, 0, chan); |
| 1581 | offset[chan] = LLVMBuildAnd(gallivm->builder, offset[chan], |
| 1582 | lp_build_const_int32(gallivm, 0x3f), ""); |
| 1583 | if (chan) |
| 1584 | offset[chan] = LLVMBuildShl(gallivm->builder, offset[chan], |
| 1585 | lp_build_const_int32(gallivm, chan*8), ""); |
| 1586 | } |
| 1587 | |
| 1588 | pack = LLVMBuildOr(gallivm->builder, offset[0], offset[1], ""); |
| 1589 | pack = LLVMBuildOr(gallivm->builder, pack, offset[2], ""); |
| 1590 | address[count++] = pack; |
| 1591 | } |
| 1592 | |
Michel Dänzer | 120efee | 2013-01-25 12:10:11 +0100 | [diff] [blame] | 1593 | /* Pack LOD bias value */ |
| 1594 | if (opcode == TGSI_OPCODE_TXB) |
| 1595 | address[count++] = coords[3]; |
Marek Olšák | 2484daa | 2014-04-22 21:23:29 +0200 | [diff] [blame] | 1596 | if (opcode == TGSI_OPCODE_TXB2) |
| 1597 | address[count++] = lp_build_emit_fetch(bld_base, inst, 1, 0); |
Vadim Girlin | 8cf552b | 2012-12-18 17:39:19 +0400 | [diff] [blame] | 1598 | |
Michel Dänzer | 120efee | 2013-01-25 12:10:11 +0100 | [diff] [blame] | 1599 | /* Pack depth comparison value */ |
Marek Olšák | 57f3da9 | 2014-06-16 16:53:42 +0200 | [diff] [blame] | 1600 | if (tgsi_is_shadow_sampler(target) && opcode != TGSI_OPCODE_LODQ) { |
Marek Olšák | 6818e11 | 2014-06-06 03:04:21 +0200 | [diff] [blame] | 1601 | if (target == TGSI_TEXTURE_SHADOWCUBE_ARRAY) { |
| 1602 | address[count++] = lp_build_emit_fetch(bld_base, inst, 1, 0); |
| 1603 | } else { |
| 1604 | assert(ref_pos >= 0); |
| 1605 | address[count++] = coords[ref_pos]; |
| 1606 | } |
Michel Dänzer | e0f2ffc | 2012-12-03 12:46:30 +0100 | [diff] [blame] | 1607 | } |
| 1608 | |
Marek Olšák | b279f014 | 2014-07-04 03:19:58 +0200 | [diff] [blame] | 1609 | if (target == TGSI_TEXTURE_CUBE || |
| 1610 | target == TGSI_TEXTURE_CUBE_ARRAY || |
| 1611 | target == TGSI_TEXTURE_SHADOWCUBE || |
| 1612 | target == TGSI_TEXTURE_SHADOWCUBE_ARRAY) |
| 1613 | radeon_llvm_emit_prepare_cube_coords(bld_base, emit_data, coords); |
| 1614 | |
Michel Dänzer | a6b83c0 | 2013-02-21 16:10:55 +0100 | [diff] [blame] | 1615 | /* Pack user derivatives */ |
| 1616 | if (opcode == TGSI_OPCODE_TXD) { |
Marek Olšák | 04aa2bd | 2014-07-02 03:55:47 +0200 | [diff] [blame] | 1617 | int num_deriv_channels, param; |
| 1618 | |
| 1619 | switch (target) { |
| 1620 | case TGSI_TEXTURE_3D: |
| 1621 | num_deriv_channels = 3; |
| 1622 | break; |
| 1623 | case TGSI_TEXTURE_2D: |
| 1624 | case TGSI_TEXTURE_SHADOW2D: |
| 1625 | case TGSI_TEXTURE_RECT: |
| 1626 | case TGSI_TEXTURE_SHADOWRECT: |
| 1627 | case TGSI_TEXTURE_2D_ARRAY: |
| 1628 | case TGSI_TEXTURE_SHADOW2D_ARRAY: |
| 1629 | case TGSI_TEXTURE_CUBE: |
| 1630 | case TGSI_TEXTURE_SHADOWCUBE: |
| 1631 | case TGSI_TEXTURE_CUBE_ARRAY: |
| 1632 | case TGSI_TEXTURE_SHADOWCUBE_ARRAY: |
| 1633 | num_deriv_channels = 2; |
| 1634 | break; |
| 1635 | case TGSI_TEXTURE_1D: |
| 1636 | case TGSI_TEXTURE_SHADOW1D: |
| 1637 | case TGSI_TEXTURE_1D_ARRAY: |
| 1638 | case TGSI_TEXTURE_SHADOW1D_ARRAY: |
| 1639 | num_deriv_channels = 1; |
| 1640 | break; |
| 1641 | default: |
| 1642 | assert(0); /* no other targets are valid here */ |
Michel Dänzer | a6b83c0 | 2013-02-21 16:10:55 +0100 | [diff] [blame] | 1643 | } |
Marek Olšák | 04aa2bd | 2014-07-02 03:55:47 +0200 | [diff] [blame] | 1644 | |
| 1645 | for (param = 1; param <= 2; param++) |
| 1646 | for (chan = 0; chan < num_deriv_channels; chan++) |
| 1647 | address[count++] = lp_build_emit_fetch(bld_base, inst, param, chan); |
Michel Dänzer | a6b83c0 | 2013-02-21 16:10:55 +0100 | [diff] [blame] | 1648 | } |
| 1649 | |
Michel Dänzer | 120efee | 2013-01-25 12:10:11 +0100 | [diff] [blame] | 1650 | /* Pack texture coordinates */ |
| 1651 | address[count++] = coords[0]; |
Michel Dänzer | beaa5eb | 2013-05-24 13:23:26 +0100 | [diff] [blame] | 1652 | if (num_coords > 1) |
Michel Dänzer | 120efee | 2013-01-25 12:10:11 +0100 | [diff] [blame] | 1653 | address[count++] = coords[1]; |
Michel Dänzer | beaa5eb | 2013-05-24 13:23:26 +0100 | [diff] [blame] | 1654 | if (num_coords > 2) |
Michel Dänzer | 120efee | 2013-01-25 12:10:11 +0100 | [diff] [blame] | 1655 | address[count++] = coords[2]; |
Michel Dänzer | e5fb734 | 2013-01-24 18:54:51 +0100 | [diff] [blame] | 1656 | |
Marek Olšák | d2bd634 | 2013-09-18 15:40:21 +0200 | [diff] [blame] | 1657 | /* Pack LOD or sample index */ |
Michel Dänzer | 3623111 | 2013-05-02 09:44:45 +0200 | [diff] [blame] | 1658 | if (opcode == TGSI_OPCODE_TXL || opcode == TGSI_OPCODE_TXF) |
Michel Dänzer | 120efee | 2013-01-25 12:10:11 +0100 | [diff] [blame] | 1659 | address[count++] = coords[3]; |
Marek Olšák | 6818e11 | 2014-06-06 03:04:21 +0200 | [diff] [blame] | 1660 | else if (opcode == TGSI_OPCODE_TXL2) |
Marek Olšák | 2484daa | 2014-04-22 21:23:29 +0200 | [diff] [blame] | 1661 | address[count++] = lp_build_emit_fetch(bld_base, inst, 1, 0); |
Michel Dänzer | 120efee | 2013-01-25 12:10:11 +0100 | [diff] [blame] | 1662 | |
| 1663 | if (count > 16) { |
| 1664 | assert(!"Cannot handle more than 16 texture address parameters"); |
| 1665 | count = 16; |
| 1666 | } |
| 1667 | |
| 1668 | for (chan = 0; chan < count; chan++ ) { |
| 1669 | address[chan] = LLVMBuildBitCast(gallivm->builder, |
| 1670 | address[chan], |
| 1671 | LLVMInt32TypeInContext(gallivm->context), |
| 1672 | ""); |
| 1673 | } |
| 1674 | |
Marek Olšák | 4855acd | 2013-08-06 15:08:54 +0200 | [diff] [blame] | 1675 | /* Adjust the sample index according to FMASK. |
| 1676 | * |
| 1677 | * For uncompressed MSAA surfaces, FMASK should return 0x76543210, |
| 1678 | * which is the identity mapping. Each nibble says which physical sample |
| 1679 | * should be fetched to get that sample. |
| 1680 | * |
| 1681 | * For example, 0x11111100 means there are only 2 samples stored and |
| 1682 | * the second sample covers 3/4 of the pixel. When reading samples 0 |
| 1683 | * and 1, return physical sample 0 (determined by the first two 0s |
| 1684 | * in FMASK), otherwise return physical sample 1. |
| 1685 | * |
| 1686 | * The sample index should be adjusted as follows: |
| 1687 | * sample_index = (fmask >> (sample_index * 4)) & 0xF; |
| 1688 | */ |
| 1689 | if (target == TGSI_TEXTURE_2D_MSAA || |
| 1690 | target == TGSI_TEXTURE_2D_ARRAY_MSAA) { |
| 1691 | struct lp_build_context *uint_bld = &bld_base->uint_bld; |
| 1692 | struct lp_build_emit_data txf_emit_data = *emit_data; |
Marek Olšák | d2bd634 | 2013-09-18 15:40:21 +0200 | [diff] [blame] | 1693 | LLVMValueRef txf_address[4]; |
Marek Olšák | 4855acd | 2013-08-06 15:08:54 +0200 | [diff] [blame] | 1694 | unsigned txf_count = count; |
Marek Olšák | 877bb52 | 2014-06-25 03:12:46 +0200 | [diff] [blame] | 1695 | struct tgsi_full_instruction inst = {}; |
Marek Olšák | 4855acd | 2013-08-06 15:08:54 +0200 | [diff] [blame] | 1696 | |
Marek Olšák | d2bd634 | 2013-09-18 15:40:21 +0200 | [diff] [blame] | 1697 | memcpy(txf_address, address, sizeof(txf_address)); |
| 1698 | |
| 1699 | if (target == TGSI_TEXTURE_2D_MSAA) { |
| 1700 | txf_address[2] = bld_base->uint_bld.zero; |
| 1701 | } |
| 1702 | txf_address[3] = bld_base->uint_bld.zero; |
Marek Olšák | 4855acd | 2013-08-06 15:08:54 +0200 | [diff] [blame] | 1703 | |
| 1704 | /* Pad to a power-of-two size. */ |
| 1705 | while (txf_count < util_next_power_of_two(txf_count)) |
| 1706 | txf_address[txf_count++] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context)); |
| 1707 | |
| 1708 | /* Read FMASK using TXF. */ |
Marek Olšák | 877bb52 | 2014-06-25 03:12:46 +0200 | [diff] [blame] | 1709 | inst.Instruction.Opcode = TGSI_OPCODE_TXF; |
| 1710 | inst.Texture.Texture = target == TGSI_TEXTURE_2D_MSAA ? TGSI_TEXTURE_2D : TGSI_TEXTURE_2D_ARRAY; |
| 1711 | txf_emit_data.inst = &inst; |
Marek Olšák | 4855acd | 2013-08-06 15:08:54 +0200 | [diff] [blame] | 1712 | txf_emit_data.chan = 0; |
| 1713 | txf_emit_data.dst_type = LLVMVectorType( |
Marek Olšák | 6818e11 | 2014-06-06 03:04:21 +0200 | [diff] [blame] | 1714 | LLVMInt32TypeInContext(gallivm->context), 4); |
Marek Olšák | 4855acd | 2013-08-06 15:08:54 +0200 | [diff] [blame] | 1715 | txf_emit_data.args[0] = lp_build_gather_values(gallivm, txf_address, txf_count); |
Marek Olšák | ee2a818 | 2014-07-07 23:27:19 +0200 | [diff] [blame] | 1716 | txf_emit_data.args[1] = si_shader_ctx->resources[SI_FMASK_TEX_OFFSET + sampler_index]; |
Marek Olšák | 877bb52 | 2014-06-25 03:12:46 +0200 | [diff] [blame] | 1717 | txf_emit_data.args[2] = lp_build_const_int32(gallivm, inst.Texture.Texture); |
Marek Olšák | 4855acd | 2013-08-06 15:08:54 +0200 | [diff] [blame] | 1718 | txf_emit_data.arg_count = 3; |
| 1719 | |
Marek Olšák | 877bb52 | 2014-06-25 03:12:46 +0200 | [diff] [blame] | 1720 | build_tex_intrinsic(&tex_action, bld_base, &txf_emit_data); |
Marek Olšák | 4855acd | 2013-08-06 15:08:54 +0200 | [diff] [blame] | 1721 | |
| 1722 | /* Initialize some constants. */ |
Marek Olšák | 4855acd | 2013-08-06 15:08:54 +0200 | [diff] [blame] | 1723 | LLVMValueRef four = LLVMConstInt(uint_bld->elem_type, 4, 0); |
| 1724 | LLVMValueRef F = LLVMConstInt(uint_bld->elem_type, 0xF, 0); |
| 1725 | |
| 1726 | /* Apply the formula. */ |
| 1727 | LLVMValueRef fmask = |
| 1728 | LLVMBuildExtractElement(gallivm->builder, |
| 1729 | txf_emit_data.output[0], |
| 1730 | uint_bld->zero, ""); |
| 1731 | |
Marek Olšák | d2bd634 | 2013-09-18 15:40:21 +0200 | [diff] [blame] | 1732 | unsigned sample_chan = target == TGSI_TEXTURE_2D_MSAA ? 2 : 3; |
Marek Olšák | 4855acd | 2013-08-06 15:08:54 +0200 | [diff] [blame] | 1733 | |
| 1734 | LLVMValueRef sample_index4 = |
Marek Olšák | d2bd634 | 2013-09-18 15:40:21 +0200 | [diff] [blame] | 1735 | LLVMBuildMul(gallivm->builder, address[sample_chan], four, ""); |
Marek Olšák | 4855acd | 2013-08-06 15:08:54 +0200 | [diff] [blame] | 1736 | |
| 1737 | LLVMValueRef shifted_fmask = |
| 1738 | LLVMBuildLShr(gallivm->builder, fmask, sample_index4, ""); |
| 1739 | |
| 1740 | LLVMValueRef final_sample = |
| 1741 | LLVMBuildAnd(gallivm->builder, shifted_fmask, F, ""); |
| 1742 | |
| 1743 | /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK |
| 1744 | * resource descriptor is 0 (invalid), |
| 1745 | */ |
| 1746 | LLVMValueRef fmask_desc = |
| 1747 | LLVMBuildBitCast(gallivm->builder, |
Marek Olšák | ee2a818 | 2014-07-07 23:27:19 +0200 | [diff] [blame] | 1748 | si_shader_ctx->resources[SI_FMASK_TEX_OFFSET + sampler_index], |
Marek Olšák | 4855acd | 2013-08-06 15:08:54 +0200 | [diff] [blame] | 1749 | LLVMVectorType(uint_bld->elem_type, 8), ""); |
| 1750 | |
| 1751 | LLVMValueRef fmask_word1 = |
| 1752 | LLVMBuildExtractElement(gallivm->builder, fmask_desc, |
| 1753 | uint_bld->one, ""); |
| 1754 | |
| 1755 | LLVMValueRef word1_is_nonzero = |
| 1756 | LLVMBuildICmp(gallivm->builder, LLVMIntNE, |
| 1757 | fmask_word1, uint_bld->zero, ""); |
| 1758 | |
Marek Olšák | d2bd634 | 2013-09-18 15:40:21 +0200 | [diff] [blame] | 1759 | /* Replace the MSAA sample index. */ |
| 1760 | address[sample_chan] = |
Marek Olšák | 4855acd | 2013-08-06 15:08:54 +0200 | [diff] [blame] | 1761 | LLVMBuildSelect(gallivm->builder, word1_is_nonzero, |
Marek Olšák | d2bd634 | 2013-09-18 15:40:21 +0200 | [diff] [blame] | 1762 | final_sample, address[sample_chan], ""); |
Marek Olšák | 4855acd | 2013-08-06 15:08:54 +0200 | [diff] [blame] | 1763 | } |
Michel Dänzer | a6b83c0 | 2013-02-21 16:10:55 +0100 | [diff] [blame] | 1764 | |
Michel Dänzer | 3623111 | 2013-05-02 09:44:45 +0200 | [diff] [blame] | 1765 | /* Resource */ |
Marek Olšák | 4855acd | 2013-08-06 15:08:54 +0200 | [diff] [blame] | 1766 | emit_data->args[1] = si_shader_ctx->resources[sampler_index]; |
Michel Dänzer | 3623111 | 2013-05-02 09:44:45 +0200 | [diff] [blame] | 1767 | |
| 1768 | if (opcode == TGSI_OPCODE_TXF) { |
| 1769 | /* add tex offsets */ |
| 1770 | if (inst->Texture.NumOffsets) { |
| 1771 | struct lp_build_context *uint_bld = &bld_base->uint_bld; |
| 1772 | struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base); |
| 1773 | const struct tgsi_texture_offset * off = inst->TexOffsets; |
| 1774 | |
| 1775 | assert(inst->Texture.NumOffsets == 1); |
| 1776 | |
Marek Olšák | defedc0 | 2013-09-18 15:36:38 +0200 | [diff] [blame] | 1777 | switch (target) { |
| 1778 | case TGSI_TEXTURE_3D: |
| 1779 | address[2] = lp_build_add(uint_bld, address[2], |
| 1780 | bld->immediates[off->Index][off->SwizzleZ]); |
| 1781 | /* fall through */ |
| 1782 | case TGSI_TEXTURE_2D: |
| 1783 | case TGSI_TEXTURE_SHADOW2D: |
| 1784 | case TGSI_TEXTURE_RECT: |
| 1785 | case TGSI_TEXTURE_SHADOWRECT: |
| 1786 | case TGSI_TEXTURE_2D_ARRAY: |
| 1787 | case TGSI_TEXTURE_SHADOW2D_ARRAY: |
Michel Dänzer | 3623111 | 2013-05-02 09:44:45 +0200 | [diff] [blame] | 1788 | address[1] = |
| 1789 | lp_build_add(uint_bld, address[1], |
Marek Olšák | defedc0 | 2013-09-18 15:36:38 +0200 | [diff] [blame] | 1790 | bld->immediates[off->Index][off->SwizzleY]); |
| 1791 | /* fall through */ |
| 1792 | case TGSI_TEXTURE_1D: |
| 1793 | case TGSI_TEXTURE_SHADOW1D: |
| 1794 | case TGSI_TEXTURE_1D_ARRAY: |
| 1795 | case TGSI_TEXTURE_SHADOW1D_ARRAY: |
| 1796 | address[0] = |
| 1797 | lp_build_add(uint_bld, address[0], |
| 1798 | bld->immediates[off->Index][off->SwizzleX]); |
| 1799 | break; |
| 1800 | /* texture offsets do not apply to other texture targets */ |
| 1801 | } |
Michel Dänzer | 3623111 | 2013-05-02 09:44:45 +0200 | [diff] [blame] | 1802 | } |
| 1803 | |
Marek Olšák | c7b5a5c | 2014-06-06 03:00:18 +0200 | [diff] [blame] | 1804 | emit_data->args[2] = lp_build_const_int32(gallivm, target); |
| 1805 | emit_data->arg_count = 3; |
| 1806 | |
Michel Dänzer | 3623111 | 2013-05-02 09:44:45 +0200 | [diff] [blame] | 1807 | emit_data->dst_type = LLVMVectorType( |
Marek Olšák | 6818e11 | 2014-06-06 03:04:21 +0200 | [diff] [blame] | 1808 | LLVMInt32TypeInContext(gallivm->context), |
Michel Dänzer | 3623111 | 2013-05-02 09:44:45 +0200 | [diff] [blame] | 1809 | 4); |
Marek Olšák | 57f3da9 | 2014-06-16 16:53:42 +0200 | [diff] [blame] | 1810 | } else if (opcode == TGSI_OPCODE_TG4 || |
Marek Olšák | 877bb52 | 2014-06-25 03:12:46 +0200 | [diff] [blame] | 1811 | opcode == TGSI_OPCODE_LODQ || |
| 1812 | has_offset) { |
Marek Olšák | c7b5a5c | 2014-06-06 03:00:18 +0200 | [diff] [blame] | 1813 | unsigned is_array = target == TGSI_TEXTURE_1D_ARRAY || |
| 1814 | target == TGSI_TEXTURE_SHADOW1D_ARRAY || |
| 1815 | target == TGSI_TEXTURE_2D_ARRAY || |
| 1816 | target == TGSI_TEXTURE_SHADOW2D_ARRAY || |
| 1817 | target == TGSI_TEXTURE_CUBE_ARRAY || |
| 1818 | target == TGSI_TEXTURE_SHADOWCUBE_ARRAY; |
| 1819 | unsigned is_rect = target == TGSI_TEXTURE_RECT; |
Marek Olšák | 57f3da9 | 2014-06-16 16:53:42 +0200 | [diff] [blame] | 1820 | unsigned dmask = 0xf; |
Michel Dänzer | 3623111 | 2013-05-02 09:44:45 +0200 | [diff] [blame] | 1821 | |
Marek Olšák | 57f3da9 | 2014-06-16 16:53:42 +0200 | [diff] [blame] | 1822 | if (opcode == TGSI_OPCODE_TG4) { |
| 1823 | unsigned gather_comp = 0; |
Marek Olšák | c7b5a5c | 2014-06-06 03:00:18 +0200 | [diff] [blame] | 1824 | |
Marek Olšák | 57f3da9 | 2014-06-16 16:53:42 +0200 | [diff] [blame] | 1825 | /* DMASK was repurposed for GATHER4. 4 components are always |
| 1826 | * returned and DMASK works like a swizzle - it selects |
| 1827 | * the component to fetch. The only valid DMASK values are |
| 1828 | * 1=red, 2=green, 4=blue, 8=alpha. (e.g. 1 returns |
| 1829 | * (red,red,red,red) etc.) The ISA document doesn't mention |
| 1830 | * this. |
| 1831 | */ |
Marek Olšák | c7b5a5c | 2014-06-06 03:00:18 +0200 | [diff] [blame] | 1832 | |
Marek Olšák | 57f3da9 | 2014-06-16 16:53:42 +0200 | [diff] [blame] | 1833 | /* Get the component index from src1.x for Gather4. */ |
| 1834 | if (!tgsi_is_shadow_sampler(target)) { |
| 1835 | LLVMValueRef (*imms)[4] = lp_soa_context(bld_base)->immediates; |
| 1836 | LLVMValueRef comp_imm; |
| 1837 | struct tgsi_src_register src1 = inst->Src[1].Register; |
Marek Olšák | c7b5a5c | 2014-06-06 03:00:18 +0200 | [diff] [blame] | 1838 | |
Marek Olšák | 57f3da9 | 2014-06-16 16:53:42 +0200 | [diff] [blame] | 1839 | assert(src1.File == TGSI_FILE_IMMEDIATE); |
| 1840 | |
| 1841 | comp_imm = imms[src1.Index][src1.SwizzleX]; |
| 1842 | gather_comp = LLVMConstIntGetZExtValue(comp_imm); |
| 1843 | gather_comp = CLAMP(gather_comp, 0, 3); |
| 1844 | } |
| 1845 | |
| 1846 | dmask = 1 << gather_comp; |
Marek Olšák | c7b5a5c | 2014-06-06 03:00:18 +0200 | [diff] [blame] | 1847 | } |
| 1848 | |
Marek Olšák | 4855acd | 2013-08-06 15:08:54 +0200 | [diff] [blame] | 1849 | emit_data->args[2] = si_shader_ctx->samplers[sampler_index]; |
Marek Olšák | 57f3da9 | 2014-06-16 16:53:42 +0200 | [diff] [blame] | 1850 | emit_data->args[3] = lp_build_const_int32(gallivm, dmask); |
Marek Olšák | c7b5a5c | 2014-06-06 03:00:18 +0200 | [diff] [blame] | 1851 | emit_data->args[4] = lp_build_const_int32(gallivm, is_rect); /* unorm */ |
| 1852 | emit_data->args[5] = lp_build_const_int32(gallivm, 0); /* r128 */ |
| 1853 | emit_data->args[6] = lp_build_const_int32(gallivm, is_array); /* da */ |
| 1854 | emit_data->args[7] = lp_build_const_int32(gallivm, 0); /* glc */ |
| 1855 | emit_data->args[8] = lp_build_const_int32(gallivm, 0); /* slc */ |
| 1856 | emit_data->args[9] = lp_build_const_int32(gallivm, 0); /* tfe */ |
| 1857 | emit_data->args[10] = lp_build_const_int32(gallivm, 0); /* lwe */ |
| 1858 | |
| 1859 | emit_data->arg_count = 11; |
Michel Dänzer | 3623111 | 2013-05-02 09:44:45 +0200 | [diff] [blame] | 1860 | |
| 1861 | emit_data->dst_type = LLVMVectorType( |
Marek Olšák | 6818e11 | 2014-06-06 03:04:21 +0200 | [diff] [blame] | 1862 | LLVMFloatTypeInContext(gallivm->context), |
Michel Dänzer | 3623111 | 2013-05-02 09:44:45 +0200 | [diff] [blame] | 1863 | 4); |
Marek Olšák | c7b5a5c | 2014-06-06 03:00:18 +0200 | [diff] [blame] | 1864 | } else { |
| 1865 | emit_data->args[2] = si_shader_ctx->samplers[sampler_index]; |
| 1866 | emit_data->args[3] = lp_build_const_int32(gallivm, target); |
Michel Dänzer | 3623111 | 2013-05-02 09:44:45 +0200 | [diff] [blame] | 1867 | emit_data->arg_count = 4; |
Marek Olšák | c7b5a5c | 2014-06-06 03:00:18 +0200 | [diff] [blame] | 1868 | |
| 1869 | emit_data->dst_type = LLVMVectorType( |
| 1870 | LLVMFloatTypeInContext(gallivm->context), |
| 1871 | 4); |
Michel Dänzer | 3623111 | 2013-05-02 09:44:45 +0200 | [diff] [blame] | 1872 | } |
| 1873 | |
Marek Olšák | 2484daa | 2014-04-22 21:23:29 +0200 | [diff] [blame] | 1874 | /* The fetch opcode has been converted to a 2D array fetch. |
| 1875 | * This simplifies the LLVM backend. */ |
| 1876 | if (target == TGSI_TEXTURE_CUBE_ARRAY) |
| 1877 | target = TGSI_TEXTURE_2D_ARRAY; |
| 1878 | else if (target == TGSI_TEXTURE_SHADOWCUBE_ARRAY) |
| 1879 | target = TGSI_TEXTURE_SHADOW2D_ARRAY; |
| 1880 | |
Michel Dänzer | 120efee | 2013-01-25 12:10:11 +0100 | [diff] [blame] | 1881 | /* Pad to power of two vector */ |
| 1882 | while (count < util_next_power_of_two(count)) |
| 1883 | address[count++] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context)); |
| 1884 | |
Christian König | ccf3e8f | 2013-03-26 15:09:27 +0100 | [diff] [blame] | 1885 | emit_data->args[0] = lp_build_gather_values(gallivm, address, count); |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 1886 | } |
| 1887 | |
Michel Dänzer | 07eddc4 | 2013-02-06 15:43:10 +0100 | [diff] [blame] | 1888 | static void build_tex_intrinsic(const struct lp_build_tgsi_action * action, |
| 1889 | struct lp_build_tgsi_context * bld_base, |
| 1890 | struct lp_build_emit_data * emit_data) |
| 1891 | { |
| 1892 | struct lp_build_context * base = &bld_base->base; |
Marek Olšák | 877bb52 | 2014-06-25 03:12:46 +0200 | [diff] [blame] | 1893 | unsigned opcode = emit_data->inst->Instruction.Opcode; |
| 1894 | unsigned target = emit_data->inst->Texture.Texture; |
Kai Wasserbäch | bbb77fc | 2013-10-27 19:36:07 +0100 | [diff] [blame] | 1895 | char intr_name[127]; |
Marek Olšák | 877bb52 | 2014-06-25 03:12:46 +0200 | [diff] [blame] | 1896 | bool has_offset = HAVE_LLVM >= 0x0305 ? |
| 1897 | emit_data->inst->Texture.NumOffsets > 0 : false; |
Michel Dänzer | 07eddc4 | 2013-02-06 15:43:10 +0100 | [diff] [blame] | 1898 | |
Marek Olšák | 877bb52 | 2014-06-25 03:12:46 +0200 | [diff] [blame] | 1899 | if (target == TGSI_TEXTURE_BUFFER) { |
Marek Olšák | dbeedbb | 2013-10-31 15:08:49 +0100 | [diff] [blame] | 1900 | emit_data->output[emit_data->chan] = build_intrinsic( |
| 1901 | base->gallivm->builder, |
| 1902 | "llvm.SI.vs.load.input", emit_data->dst_type, |
| 1903 | emit_data->args, emit_data->arg_count, |
| 1904 | LLVMReadNoneAttribute | LLVMNoUnwindAttribute); |
| 1905 | return; |
| 1906 | } |
| 1907 | |
Marek Olšák | 877bb52 | 2014-06-25 03:12:46 +0200 | [diff] [blame] | 1908 | if (opcode == TGSI_OPCODE_TG4 || |
| 1909 | opcode == TGSI_OPCODE_LODQ || |
| 1910 | (opcode != TGSI_OPCODE_TXF && has_offset)) { |
| 1911 | bool is_shadow = tgsi_is_shadow_sampler(target); |
| 1912 | const char *name = "llvm.SI.image.sample"; |
| 1913 | const char *infix = ""; |
Michel Dänzer | 07eddc4 | 2013-02-06 15:43:10 +0100 | [diff] [blame] | 1914 | |
Marek Olšák | 877bb52 | 2014-06-25 03:12:46 +0200 | [diff] [blame] | 1915 | switch (opcode) { |
| 1916 | case TGSI_OPCODE_TEX: |
| 1917 | case TGSI_OPCODE_TEX2: |
| 1918 | case TGSI_OPCODE_TXP: |
| 1919 | break; |
| 1920 | case TGSI_OPCODE_TXB: |
| 1921 | case TGSI_OPCODE_TXB2: |
| 1922 | infix = ".b"; |
| 1923 | break; |
| 1924 | case TGSI_OPCODE_TXL: |
| 1925 | case TGSI_OPCODE_TXL2: |
| 1926 | infix = ".l"; |
| 1927 | break; |
| 1928 | case TGSI_OPCODE_TXD: |
| 1929 | infix = ".d"; |
| 1930 | break; |
| 1931 | case TGSI_OPCODE_TG4: |
| 1932 | name = "llvm.SI.gather4"; |
| 1933 | break; |
| 1934 | case TGSI_OPCODE_LODQ: |
| 1935 | name = "llvm.SI.getlod"; |
| 1936 | is_shadow = false; |
| 1937 | has_offset = false; |
| 1938 | break; |
| 1939 | default: |
| 1940 | assert(0); |
| 1941 | return; |
| 1942 | } |
Michel Dänzer | 07eddc4 | 2013-02-06 15:43:10 +0100 | [diff] [blame] | 1943 | |
Marek Olšák | 877bb52 | 2014-06-25 03:12:46 +0200 | [diff] [blame] | 1944 | /* Add the type and suffixes .c, .o if needed. */ |
| 1945 | sprintf(intr_name, "%s%s%s%s.v%ui32", name, |
| 1946 | is_shadow ? ".c" : "", infix, has_offset ? ".o" : "", |
| 1947 | LLVMGetVectorSize(LLVMTypeOf(emit_data->args[0]))); |
Marek Olšák | c7b5a5c | 2014-06-06 03:00:18 +0200 | [diff] [blame] | 1948 | |
Marek Olšák | 877bb52 | 2014-06-25 03:12:46 +0200 | [diff] [blame] | 1949 | emit_data->output[emit_data->chan] = build_intrinsic( |
| 1950 | base->gallivm->builder, intr_name, emit_data->dst_type, |
| 1951 | emit_data->args, emit_data->arg_count, |
| 1952 | LLVMReadNoneAttribute | LLVMNoUnwindAttribute); |
| 1953 | } else { |
Marek Olšák | d859bdb | 2014-07-14 19:40:14 +0200 | [diff] [blame] | 1954 | LLVMTypeRef i8, v16i8, v32i8; |
Marek Olšák | 877bb52 | 2014-06-25 03:12:46 +0200 | [diff] [blame] | 1955 | const char *name; |
Marek Olšák | c7b5a5c | 2014-06-06 03:00:18 +0200 | [diff] [blame] | 1956 | |
Marek Olšák | 877bb52 | 2014-06-25 03:12:46 +0200 | [diff] [blame] | 1957 | switch (opcode) { |
| 1958 | case TGSI_OPCODE_TEX: |
| 1959 | case TGSI_OPCODE_TEX2: |
| 1960 | case TGSI_OPCODE_TXP: |
| 1961 | name = "llvm.SI.sample"; |
| 1962 | break; |
| 1963 | case TGSI_OPCODE_TXB: |
| 1964 | case TGSI_OPCODE_TXB2: |
| 1965 | name = "llvm.SI.sampleb"; |
| 1966 | break; |
| 1967 | case TGSI_OPCODE_TXD: |
| 1968 | name = "llvm.SI.sampled"; |
| 1969 | break; |
| 1970 | case TGSI_OPCODE_TXF: |
| 1971 | name = "llvm.SI.imageload"; |
| 1972 | break; |
| 1973 | case TGSI_OPCODE_TXL: |
| 1974 | case TGSI_OPCODE_TXL2: |
| 1975 | name = "llvm.SI.samplel"; |
| 1976 | break; |
| 1977 | default: |
| 1978 | assert(0); |
| 1979 | return; |
| 1980 | } |
| 1981 | |
Marek Olšák | d859bdb | 2014-07-14 19:40:14 +0200 | [diff] [blame] | 1982 | i8 = LLVMInt8TypeInContext(base->gallivm->context); |
| 1983 | v16i8 = LLVMVectorType(i8, 16); |
| 1984 | v32i8 = LLVMVectorType(i8, 32); |
| 1985 | |
| 1986 | emit_data->args[1] = LLVMBuildBitCast(base->gallivm->builder, |
| 1987 | emit_data->args[1], v32i8, ""); |
| 1988 | if (opcode != TGSI_OPCODE_TXF) { |
| 1989 | emit_data->args[2] = LLVMBuildBitCast(base->gallivm->builder, |
| 1990 | emit_data->args[2], v16i8, ""); |
| 1991 | } |
| 1992 | |
Marek Olšák | 877bb52 | 2014-06-25 03:12:46 +0200 | [diff] [blame] | 1993 | sprintf(intr_name, "%s.v%ui32", name, |
| 1994 | LLVMGetVectorSize(LLVMTypeOf(emit_data->args[0]))); |
| 1995 | |
| 1996 | emit_data->output[emit_data->chan] = build_intrinsic( |
| 1997 | base->gallivm->builder, intr_name, emit_data->dst_type, |
| 1998 | emit_data->args, emit_data->arg_count, |
| 1999 | LLVMReadNoneAttribute | LLVMNoUnwindAttribute); |
| 2000 | } |
Marek Olšák | c7b5a5c | 2014-06-06 03:00:18 +0200 | [diff] [blame] | 2001 | } |
| 2002 | |
Michel Dänzer | 0495adb | 2013-05-06 12:45:14 +0200 | [diff] [blame] | 2003 | static void txq_fetch_args( |
| 2004 | struct lp_build_tgsi_context * bld_base, |
| 2005 | struct lp_build_emit_data * emit_data) |
| 2006 | { |
| 2007 | struct si_shader_context *si_shader_ctx = si_shader_context(bld_base); |
| 2008 | const struct tgsi_full_instruction *inst = emit_data->inst; |
Marek Olšák | dbeedbb | 2013-10-31 15:08:49 +0100 | [diff] [blame] | 2009 | struct gallivm_state *gallivm = bld_base->base.gallivm; |
Marek Olšák | 2484daa | 2014-04-22 21:23:29 +0200 | [diff] [blame] | 2010 | unsigned target = inst->Texture.Texture; |
Marek Olšák | dbeedbb | 2013-10-31 15:08:49 +0100 | [diff] [blame] | 2011 | |
Marek Olšák | 2484daa | 2014-04-22 21:23:29 +0200 | [diff] [blame] | 2012 | if (target == TGSI_TEXTURE_BUFFER) { |
Marek Olšák | dbeedbb | 2013-10-31 15:08:49 +0100 | [diff] [blame] | 2013 | LLVMTypeRef i32 = LLVMInt32TypeInContext(gallivm->context); |
| 2014 | LLVMTypeRef v8i32 = LLVMVectorType(i32, 8); |
| 2015 | |
| 2016 | /* Read the size from the buffer descriptor directly. */ |
| 2017 | LLVMValueRef size = si_shader_ctx->resources[inst->Src[1].Register.Index]; |
| 2018 | size = LLVMBuildBitCast(gallivm->builder, size, v8i32, ""); |
| 2019 | size = LLVMBuildExtractElement(gallivm->builder, size, |
Marek Olšák | b142dd2 | 2015-02-01 16:58:08 +0100 | [diff] [blame] | 2020 | lp_build_const_int32(gallivm, 6), ""); |
Marek Olšák | dbeedbb | 2013-10-31 15:08:49 +0100 | [diff] [blame] | 2021 | emit_data->args[0] = size; |
| 2022 | return; |
| 2023 | } |
Michel Dänzer | 0495adb | 2013-05-06 12:45:14 +0200 | [diff] [blame] | 2024 | |
| 2025 | /* Mip level */ |
| 2026 | emit_data->args[0] = lp_build_emit_fetch(bld_base, inst, 0, TGSI_CHAN_X); |
| 2027 | |
| 2028 | /* Resource */ |
| 2029 | emit_data->args[1] = si_shader_ctx->resources[inst->Src[1].Register.Index]; |
| 2030 | |
Marek Olšák | 2484daa | 2014-04-22 21:23:29 +0200 | [diff] [blame] | 2031 | /* Texture target */ |
| 2032 | if (target == TGSI_TEXTURE_CUBE_ARRAY || |
| 2033 | target == TGSI_TEXTURE_SHADOWCUBE_ARRAY) |
| 2034 | target = TGSI_TEXTURE_2D_ARRAY; |
| 2035 | |
Michel Dänzer | 0495adb | 2013-05-06 12:45:14 +0200 | [diff] [blame] | 2036 | emit_data->args[2] = lp_build_const_int32(bld_base->base.gallivm, |
Marek Olšák | 6818e11 | 2014-06-06 03:04:21 +0200 | [diff] [blame] | 2037 | target); |
Michel Dänzer | 0495adb | 2013-05-06 12:45:14 +0200 | [diff] [blame] | 2038 | |
| 2039 | emit_data->arg_count = 3; |
| 2040 | |
| 2041 | emit_data->dst_type = LLVMVectorType( |
| 2042 | LLVMInt32TypeInContext(bld_base->base.gallivm->context), |
| 2043 | 4); |
| 2044 | } |
| 2045 | |
Marek Olšák | dbeedbb | 2013-10-31 15:08:49 +0100 | [diff] [blame] | 2046 | static void build_txq_intrinsic(const struct lp_build_tgsi_action * action, |
| 2047 | struct lp_build_tgsi_context * bld_base, |
| 2048 | struct lp_build_emit_data * emit_data) |
| 2049 | { |
Marek Olšák | 2484daa | 2014-04-22 21:23:29 +0200 | [diff] [blame] | 2050 | unsigned target = emit_data->inst->Texture.Texture; |
| 2051 | |
| 2052 | if (target == TGSI_TEXTURE_BUFFER) { |
Marek Olšák | dbeedbb | 2013-10-31 15:08:49 +0100 | [diff] [blame] | 2053 | /* Just return the buffer size. */ |
| 2054 | emit_data->output[emit_data->chan] = emit_data->args[0]; |
| 2055 | return; |
| 2056 | } |
| 2057 | |
| 2058 | build_tgsi_intrinsic_nomem(action, bld_base, emit_data); |
Marek Olšák | 2484daa | 2014-04-22 21:23:29 +0200 | [diff] [blame] | 2059 | |
| 2060 | /* Divide the number of layers by 6 to get the number of cubes. */ |
| 2061 | if (target == TGSI_TEXTURE_CUBE_ARRAY || |
| 2062 | target == TGSI_TEXTURE_SHADOWCUBE_ARRAY) { |
| 2063 | LLVMBuilderRef builder = bld_base->base.gallivm->builder; |
| 2064 | LLVMValueRef two = lp_build_const_int32(bld_base->base.gallivm, 2); |
| 2065 | LLVMValueRef six = lp_build_const_int32(bld_base->base.gallivm, 6); |
| 2066 | |
| 2067 | LLVMValueRef v4 = emit_data->output[emit_data->chan]; |
| 2068 | LLVMValueRef z = LLVMBuildExtractElement(builder, v4, two, ""); |
| 2069 | z = LLVMBuildSDiv(builder, z, six, ""); |
| 2070 | |
| 2071 | emit_data->output[emit_data->chan] = |
| 2072 | LLVMBuildInsertElement(builder, v4, z, two, ""); |
| 2073 | } |
Marek Olšák | dbeedbb | 2013-10-31 15:08:49 +0100 | [diff] [blame] | 2074 | } |
| 2075 | |
Michel Dänzer | a06ee5a | 2013-06-19 18:14:01 +0200 | [diff] [blame] | 2076 | static void si_llvm_emit_ddxy( |
| 2077 | const struct lp_build_tgsi_action * action, |
| 2078 | struct lp_build_tgsi_context * bld_base, |
| 2079 | struct lp_build_emit_data * emit_data) |
| 2080 | { |
| 2081 | struct si_shader_context *si_shader_ctx = si_shader_context(bld_base); |
| 2082 | struct gallivm_state *gallivm = bld_base->base.gallivm; |
| 2083 | struct lp_build_context * base = &bld_base->base; |
| 2084 | const struct tgsi_full_instruction *inst = emit_data->inst; |
| 2085 | unsigned opcode = inst->Instruction.Opcode; |
| 2086 | LLVMValueRef indices[2]; |
| 2087 | LLVMValueRef store_ptr, load_ptr0, load_ptr1; |
| 2088 | LLVMValueRef tl, trbl, result[4]; |
| 2089 | LLVMTypeRef i32; |
| 2090 | unsigned swizzle[4]; |
| 2091 | unsigned c; |
| 2092 | |
| 2093 | i32 = LLVMInt32TypeInContext(gallivm->context); |
| 2094 | |
| 2095 | indices[0] = bld_base->uint_bld.zero; |
| 2096 | indices[1] = build_intrinsic(gallivm->builder, "llvm.SI.tid", i32, |
| 2097 | NULL, 0, LLVMReadNoneAttribute); |
| 2098 | store_ptr = LLVMBuildGEP(gallivm->builder, si_shader_ctx->ddxy_lds, |
| 2099 | indices, 2, ""); |
| 2100 | |
| 2101 | indices[1] = LLVMBuildAnd(gallivm->builder, indices[1], |
| 2102 | lp_build_const_int32(gallivm, 0xfffffffc), ""); |
| 2103 | load_ptr0 = LLVMBuildGEP(gallivm->builder, si_shader_ctx->ddxy_lds, |
| 2104 | indices, 2, ""); |
| 2105 | |
| 2106 | indices[1] = LLVMBuildAdd(gallivm->builder, indices[1], |
| 2107 | lp_build_const_int32(gallivm, |
| 2108 | opcode == TGSI_OPCODE_DDX ? 1 : 2), |
| 2109 | ""); |
| 2110 | load_ptr1 = LLVMBuildGEP(gallivm->builder, si_shader_ctx->ddxy_lds, |
| 2111 | indices, 2, ""); |
| 2112 | |
| 2113 | for (c = 0; c < 4; ++c) { |
| 2114 | unsigned i; |
| 2115 | |
| 2116 | swizzle[c] = tgsi_util_get_full_src_register_swizzle(&inst->Src[0], c); |
| 2117 | for (i = 0; i < c; ++i) { |
| 2118 | if (swizzle[i] == swizzle[c]) { |
| 2119 | result[c] = result[i]; |
| 2120 | break; |
| 2121 | } |
| 2122 | } |
| 2123 | if (i != c) |
| 2124 | continue; |
| 2125 | |
| 2126 | LLVMBuildStore(gallivm->builder, |
| 2127 | LLVMBuildBitCast(gallivm->builder, |
| 2128 | lp_build_emit_fetch(bld_base, inst, 0, c), |
| 2129 | i32, ""), |
| 2130 | store_ptr); |
| 2131 | |
| 2132 | tl = LLVMBuildLoad(gallivm->builder, load_ptr0, ""); |
| 2133 | tl = LLVMBuildBitCast(gallivm->builder, tl, base->elem_type, ""); |
| 2134 | |
| 2135 | trbl = LLVMBuildLoad(gallivm->builder, load_ptr1, ""); |
| 2136 | trbl = LLVMBuildBitCast(gallivm->builder, trbl, base->elem_type, ""); |
| 2137 | |
| 2138 | result[c] = LLVMBuildFSub(gallivm->builder, trbl, tl, ""); |
| 2139 | } |
| 2140 | |
| 2141 | emit_data->output[0] = lp_build_gather_values(gallivm, result, 4); |
| 2142 | } |
| 2143 | |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2144 | /* Emit one vertex from the geometry shader */ |
| 2145 | static void si_llvm_emit_vertex( |
| 2146 | const struct lp_build_tgsi_action *action, |
| 2147 | struct lp_build_tgsi_context *bld_base, |
| 2148 | struct lp_build_emit_data *emit_data) |
| 2149 | { |
| 2150 | struct si_shader_context *si_shader_ctx = si_shader_context(bld_base); |
Michel Dänzer | f07a96d | 2014-01-08 18:45:10 +0900 | [diff] [blame] | 2151 | struct lp_build_context *uint = &bld_base->uint_bld; |
Marek Olšák | 8c37c16 | 2014-09-16 18:40:07 +0200 | [diff] [blame] | 2152 | struct si_shader *shader = si_shader_ctx->shader; |
Marek Olšák | 02134cf | 2014-10-04 22:07:50 +0200 | [diff] [blame] | 2153 | struct tgsi_shader_info *info = &shader->selector->info; |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2154 | struct gallivm_state *gallivm = bld_base->base.gallivm; |
| 2155 | LLVMTypeRef i32 = LLVMInt32TypeInContext(gallivm->context); |
Michel Dänzer | db9d6af | 2014-01-24 16:46:27 +0900 | [diff] [blame] | 2156 | LLVMValueRef soffset = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, |
| 2157 | SI_PARAM_GS2VS_OFFSET); |
Michel Dänzer | f07a96d | 2014-01-08 18:45:10 +0900 | [diff] [blame] | 2158 | LLVMValueRef gs_next_vertex; |
Michel Daenzer | 59936a4 | 2014-02-13 15:37:11 +0900 | [diff] [blame] | 2159 | LLVMValueRef can_emit, kill; |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2160 | LLVMValueRef args[2]; |
| 2161 | unsigned chan; |
| 2162 | int i; |
| 2163 | |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2164 | /* Write vertex attribute values to GSVS ring */ |
Michel Dänzer | f07a96d | 2014-01-08 18:45:10 +0900 | [diff] [blame] | 2165 | gs_next_vertex = LLVMBuildLoad(gallivm->builder, si_shader_ctx->gs_next_vertex, ""); |
Michel Daenzer | 59936a4 | 2014-02-13 15:37:11 +0900 | [diff] [blame] | 2166 | |
| 2167 | /* If this thread has already emitted the declared maximum number of |
| 2168 | * vertices, kill it: excessive vertex emissions are not supposed to |
| 2169 | * have any effect, and GS threads have no externally observable |
| 2170 | * effects other than emitting vertices. |
| 2171 | */ |
| 2172 | can_emit = LLVMBuildICmp(gallivm->builder, LLVMIntULE, gs_next_vertex, |
| 2173 | lp_build_const_int32(gallivm, |
Marek Olšák | 0a2d6f0 | 2014-09-30 16:55:36 +0200 | [diff] [blame] | 2174 | shader->selector->gs_max_out_vertices), ""); |
Michel Daenzer | 59936a4 | 2014-02-13 15:37:11 +0900 | [diff] [blame] | 2175 | kill = lp_build_select(&bld_base->base, can_emit, |
| 2176 | lp_build_const_float(gallivm, 1.0f), |
| 2177 | lp_build_const_float(gallivm, -1.0f)); |
| 2178 | build_intrinsic(gallivm->builder, "llvm.AMDGPU.kill", |
| 2179 | LLVMVoidTypeInContext(gallivm->context), &kill, 1, 0); |
| 2180 | |
Marek Olšák | 02134cf | 2014-10-04 22:07:50 +0200 | [diff] [blame] | 2181 | for (i = 0; i < info->num_outputs; i++) { |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2182 | LLVMValueRef *out_ptr = |
Marek Olšák | 02134cf | 2014-10-04 22:07:50 +0200 | [diff] [blame] | 2183 | si_shader_ctx->radeon_bld.soa.outputs[i]; |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2184 | |
| 2185 | for (chan = 0; chan < 4; chan++) { |
| 2186 | LLVMValueRef out_val = LLVMBuildLoad(gallivm->builder, out_ptr[chan], ""); |
Michel Dänzer | f07a96d | 2014-01-08 18:45:10 +0900 | [diff] [blame] | 2187 | LLVMValueRef voffset = |
| 2188 | lp_build_const_int32(gallivm, (i * 4 + chan) * |
Marek Olšák | 0a2d6f0 | 2014-09-30 16:55:36 +0200 | [diff] [blame] | 2189 | shader->selector->gs_max_out_vertices); |
Michel Dänzer | f07a96d | 2014-01-08 18:45:10 +0900 | [diff] [blame] | 2190 | |
| 2191 | voffset = lp_build_add(uint, voffset, gs_next_vertex); |
| 2192 | voffset = lp_build_mul_imm(uint, voffset, 4); |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2193 | |
| 2194 | out_val = LLVMBuildBitCast(gallivm->builder, out_val, i32, ""); |
| 2195 | |
Marek Olšák | fc3b335 | 2014-10-05 13:33:40 +0200 | [diff] [blame] | 2196 | build_tbuffer_store(si_shader_ctx, |
| 2197 | si_shader_ctx->gsvs_ring, |
| 2198 | out_val, 1, |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2199 | voffset, soffset, 0, |
| 2200 | V_008F0C_BUF_DATA_FORMAT_32, |
| 2201 | V_008F0C_BUF_NUM_FORMAT_UINT, |
| 2202 | 1, 0, 1, 1, 0); |
| 2203 | } |
| 2204 | } |
Michel Dänzer | f07a96d | 2014-01-08 18:45:10 +0900 | [diff] [blame] | 2205 | gs_next_vertex = lp_build_add(uint, gs_next_vertex, |
| 2206 | lp_build_const_int32(gallivm, 1)); |
| 2207 | LLVMBuildStore(gallivm->builder, gs_next_vertex, si_shader_ctx->gs_next_vertex); |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2208 | |
| 2209 | /* Signal vertex emission */ |
| 2210 | args[0] = lp_build_const_int32(gallivm, SENDMSG_GS_OP_EMIT | SENDMSG_GS); |
| 2211 | args[1] = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_GS_WAVE_ID); |
| 2212 | build_intrinsic(gallivm->builder, "llvm.SI.sendmsg", |
| 2213 | LLVMVoidTypeInContext(gallivm->context), args, 2, |
| 2214 | LLVMNoUnwindAttribute); |
| 2215 | } |
| 2216 | |
| 2217 | /* Cut one primitive from the geometry shader */ |
| 2218 | static void si_llvm_emit_primitive( |
| 2219 | const struct lp_build_tgsi_action *action, |
| 2220 | struct lp_build_tgsi_context *bld_base, |
| 2221 | struct lp_build_emit_data *emit_data) |
| 2222 | { |
| 2223 | struct si_shader_context *si_shader_ctx = si_shader_context(bld_base); |
| 2224 | struct gallivm_state *gallivm = bld_base->base.gallivm; |
| 2225 | LLVMValueRef args[2]; |
| 2226 | |
| 2227 | /* Signal primitive cut */ |
| 2228 | args[0] = lp_build_const_int32(gallivm, SENDMSG_GS_OP_CUT | SENDMSG_GS); |
| 2229 | args[1] = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_GS_WAVE_ID); |
| 2230 | build_intrinsic(gallivm->builder, "llvm.SI.sendmsg", |
| 2231 | LLVMVoidTypeInContext(gallivm->context), args, 2, |
| 2232 | LLVMNoUnwindAttribute); |
| 2233 | } |
| 2234 | |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 2235 | static const struct lp_build_tgsi_action tex_action = { |
| 2236 | .fetch_args = tex_fetch_args, |
Michel Dänzer | 07eddc4 | 2013-02-06 15:43:10 +0100 | [diff] [blame] | 2237 | .emit = build_tex_intrinsic, |
Michel Dänzer | 56ae9be | 2012-11-06 17:41:50 +0100 | [diff] [blame] | 2238 | }; |
| 2239 | |
Michel Dänzer | 0495adb | 2013-05-06 12:45:14 +0200 | [diff] [blame] | 2240 | static const struct lp_build_tgsi_action txq_action = { |
| 2241 | .fetch_args = txq_fetch_args, |
Marek Olšák | dbeedbb | 2013-10-31 15:08:49 +0100 | [diff] [blame] | 2242 | .emit = build_txq_intrinsic, |
Michel Dänzer | 0495adb | 2013-05-06 12:45:14 +0200 | [diff] [blame] | 2243 | .intr_name = "llvm.SI.resinfo" |
| 2244 | }; |
| 2245 | |
Christian König | 206f059 | 2013-03-20 14:37:21 +0100 | [diff] [blame] | 2246 | static void create_meta_data(struct si_shader_context *si_shader_ctx) |
| 2247 | { |
| 2248 | struct gallivm_state *gallivm = si_shader_ctx->radeon_bld.soa.bld_base.base.gallivm; |
| 2249 | LLVMValueRef args[3]; |
| 2250 | |
| 2251 | args[0] = LLVMMDStringInContext(gallivm->context, "const", 5); |
| 2252 | args[1] = 0; |
| 2253 | args[2] = lp_build_const_int32(gallivm, 1); |
| 2254 | |
| 2255 | si_shader_ctx->const_md = LLVMMDNodeInContext(gallivm->context, args, 3); |
| 2256 | } |
| 2257 | |
Marek Olšák | 4f3f043 | 2014-07-07 22:49:15 +0200 | [diff] [blame] | 2258 | static LLVMTypeRef const_array(LLVMTypeRef elem_type, int num_elements) |
| 2259 | { |
| 2260 | return LLVMPointerType(LLVMArrayType(elem_type, num_elements), |
| 2261 | CONST_ADDR_SPACE); |
| 2262 | } |
| 2263 | |
Christian König | 55fe5cc | 2013-03-04 16:30:06 +0100 | [diff] [blame] | 2264 | static void create_function(struct si_shader_context *si_shader_ctx) |
| 2265 | { |
Michel Dänzer | a06ee5a | 2013-06-19 18:14:01 +0200 | [diff] [blame] | 2266 | struct lp_build_tgsi_context *bld_base = &si_shader_ctx->radeon_bld.soa.bld_base; |
| 2267 | struct gallivm_state *gallivm = bld_base->base.gallivm; |
Marek Olšák | 8c37c16 | 2014-09-16 18:40:07 +0200 | [diff] [blame] | 2268 | struct si_shader *shader = si_shader_ctx->shader; |
Marek Olšák | 4f3f043 | 2014-07-07 22:49:15 +0200 | [diff] [blame] | 2269 | LLVMTypeRef params[SI_NUM_PARAMS], f32, i8, i32, v2i32, v3i32, v16i8, v4i32, v8i32; |
Marek Olšák | aeb05f0 | 2014-09-30 18:13:06 +0200 | [diff] [blame] | 2270 | unsigned i, last_array_pointer, last_sgpr, num_params; |
Christian König | 55fe5cc | 2013-03-04 16:30:06 +0100 | [diff] [blame] | 2271 | |
Christian König | 55fe5cc | 2013-03-04 16:30:06 +0100 | [diff] [blame] | 2272 | i8 = LLVMInt8TypeInContext(gallivm->context); |
Christian König | c497321 | 2013-03-05 12:14:02 +0100 | [diff] [blame] | 2273 | i32 = LLVMInt32TypeInContext(gallivm->context); |
Christian König | 0666ffd | 2013-03-05 15:07:39 +0100 | [diff] [blame] | 2274 | f32 = LLVMFloatTypeInContext(gallivm->context); |
| 2275 | v2i32 = LLVMVectorType(i32, 2); |
| 2276 | v3i32 = LLVMVectorType(i32, 3); |
Marek Olšák | 4f3f043 | 2014-07-07 22:49:15 +0200 | [diff] [blame] | 2277 | v4i32 = LLVMVectorType(i32, 4); |
| 2278 | v8i32 = LLVMVectorType(i32, 8); |
| 2279 | v16i8 = LLVMVectorType(i8, 16); |
Christian König | c497321 | 2013-03-05 12:14:02 +0100 | [diff] [blame] | 2280 | |
Marek Olšák | ee2a818 | 2014-07-07 23:27:19 +0200 | [diff] [blame] | 2281 | params[SI_PARAM_RW_BUFFERS] = const_array(v16i8, SI_NUM_RW_BUFFERS); |
Marek Olšák | 1f6c0b5 | 2014-09-23 19:42:28 +0200 | [diff] [blame] | 2282 | params[SI_PARAM_CONST] = const_array(v16i8, SI_NUM_CONST_BUFFERS); |
Marek Olšák | ee2a818 | 2014-07-07 23:27:19 +0200 | [diff] [blame] | 2283 | params[SI_PARAM_SAMPLER] = const_array(v4i32, SI_NUM_SAMPLER_STATES); |
| 2284 | params[SI_PARAM_RESOURCE] = const_array(v8i32, SI_NUM_SAMPLER_VIEWS); |
Marek Olšák | aeb05f0 | 2014-09-30 18:13:06 +0200 | [diff] [blame] | 2285 | last_array_pointer = SI_PARAM_RESOURCE; |
Christian König | 55fe5cc | 2013-03-04 16:30:06 +0100 | [diff] [blame] | 2286 | |
Marek Olšák | 13a1a8b | 2013-08-18 01:57:40 +0200 | [diff] [blame] | 2287 | switch (si_shader_ctx->type) { |
| 2288 | case TGSI_PROCESSOR_VERTEX: |
Marek Olšák | ee2a818 | 2014-07-07 23:27:19 +0200 | [diff] [blame] | 2289 | params[SI_PARAM_VERTEX_BUFFER] = const_array(v16i8, SI_NUM_VERTEX_BUFFERS); |
Marek Olšák | aeb05f0 | 2014-09-30 18:13:06 +0200 | [diff] [blame] | 2290 | last_array_pointer = SI_PARAM_VERTEX_BUFFER; |
Marek Olšák | 09056b3 | 2014-04-23 16:15:36 +0200 | [diff] [blame] | 2291 | params[SI_PARAM_BASE_VERTEX] = i32; |
Christian König | cf9b31f | 2013-03-21 18:30:23 +0100 | [diff] [blame] | 2292 | params[SI_PARAM_START_INSTANCE] = i32; |
Marek Olšák | 8d03d92 | 2013-09-01 23:59:06 +0200 | [diff] [blame] | 2293 | num_params = SI_PARAM_START_INSTANCE+1; |
Marek Olšák | 1f6c0b5 | 2014-09-23 19:42:28 +0200 | [diff] [blame] | 2294 | |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2295 | if (shader->key.vs.as_es) { |
| 2296 | params[SI_PARAM_ES2GS_OFFSET] = i32; |
| 2297 | num_params++; |
| 2298 | } else { |
Marek Olšák | 1f6c0b5 | 2014-09-23 19:42:28 +0200 | [diff] [blame] | 2299 | if (shader->is_gs_copy_shader) { |
| 2300 | last_array_pointer = SI_PARAM_CONST; |
| 2301 | num_params = SI_PARAM_CONST+1; |
| 2302 | } |
| 2303 | |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2304 | /* The locations of the other parameters are assigned dynamically. */ |
Marek Olšák | 8d03d92 | 2013-09-01 23:59:06 +0200 | [diff] [blame] | 2305 | |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2306 | /* Streamout SGPRs. */ |
| 2307 | if (shader->selector->so.num_outputs) { |
| 2308 | params[si_shader_ctx->param_streamout_config = num_params++] = i32; |
| 2309 | params[si_shader_ctx->param_streamout_write_index = num_params++] = i32; |
| 2310 | } |
| 2311 | /* A streamout buffer offset is loaded if the stride is non-zero. */ |
| 2312 | for (i = 0; i < 4; i++) { |
| 2313 | if (!shader->selector->so.stride[i]) |
| 2314 | continue; |
Marek Olšák | 8d03d92 | 2013-09-01 23:59:06 +0200 | [diff] [blame] | 2315 | |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2316 | params[si_shader_ctx->param_streamout_offset[i] = num_params++] = i32; |
| 2317 | } |
Marek Olšák | 8d03d92 | 2013-09-01 23:59:06 +0200 | [diff] [blame] | 2318 | } |
| 2319 | |
| 2320 | last_sgpr = num_params-1; |
| 2321 | |
| 2322 | /* VGPRs */ |
| 2323 | params[si_shader_ctx->param_vertex_id = num_params++] = i32; |
| 2324 | params[num_params++] = i32; /* unused*/ |
| 2325 | params[num_params++] = i32; /* unused */ |
| 2326 | params[si_shader_ctx->param_instance_id = num_params++] = i32; |
Marek Olšák | 13a1a8b | 2013-08-18 01:57:40 +0200 | [diff] [blame] | 2327 | break; |
Christian König | 0666ffd | 2013-03-05 15:07:39 +0100 | [diff] [blame] | 2328 | |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2329 | case TGSI_PROCESSOR_GEOMETRY: |
| 2330 | params[SI_PARAM_GS2VS_OFFSET] = i32; |
| 2331 | params[SI_PARAM_GS_WAVE_ID] = i32; |
| 2332 | last_sgpr = SI_PARAM_GS_WAVE_ID; |
| 2333 | |
| 2334 | /* VGPRs */ |
| 2335 | params[SI_PARAM_VTX0_OFFSET] = i32; |
| 2336 | params[SI_PARAM_VTX1_OFFSET] = i32; |
| 2337 | params[SI_PARAM_PRIMITIVE_ID] = i32; |
| 2338 | params[SI_PARAM_VTX2_OFFSET] = i32; |
| 2339 | params[SI_PARAM_VTX3_OFFSET] = i32; |
| 2340 | params[SI_PARAM_VTX4_OFFSET] = i32; |
| 2341 | params[SI_PARAM_VTX5_OFFSET] = i32; |
| 2342 | params[SI_PARAM_GS_INSTANCE_ID] = i32; |
| 2343 | num_params = SI_PARAM_GS_INSTANCE_ID+1; |
| 2344 | break; |
| 2345 | |
Marek Olšák | 13a1a8b | 2013-08-18 01:57:40 +0200 | [diff] [blame] | 2346 | case TGSI_PROCESSOR_FRAGMENT: |
Vadim Girlin | 453ea2d | 2013-10-13 19:53:54 +0400 | [diff] [blame] | 2347 | params[SI_PARAM_ALPHA_REF] = f32; |
Christian König | 0666ffd | 2013-03-05 15:07:39 +0100 | [diff] [blame] | 2348 | params[SI_PARAM_PRIM_MASK] = i32; |
Marek Olšák | 13a1a8b | 2013-08-18 01:57:40 +0200 | [diff] [blame] | 2349 | last_sgpr = SI_PARAM_PRIM_MASK; |
Christian König | 0666ffd | 2013-03-05 15:07:39 +0100 | [diff] [blame] | 2350 | params[SI_PARAM_PERSP_SAMPLE] = v2i32; |
| 2351 | params[SI_PARAM_PERSP_CENTER] = v2i32; |
| 2352 | params[SI_PARAM_PERSP_CENTROID] = v2i32; |
| 2353 | params[SI_PARAM_PERSP_PULL_MODEL] = v3i32; |
| 2354 | params[SI_PARAM_LINEAR_SAMPLE] = v2i32; |
| 2355 | params[SI_PARAM_LINEAR_CENTER] = v2i32; |
| 2356 | params[SI_PARAM_LINEAR_CENTROID] = v2i32; |
| 2357 | params[SI_PARAM_LINE_STIPPLE_TEX] = f32; |
| 2358 | params[SI_PARAM_POS_X_FLOAT] = f32; |
| 2359 | params[SI_PARAM_POS_Y_FLOAT] = f32; |
| 2360 | params[SI_PARAM_POS_Z_FLOAT] = f32; |
| 2361 | params[SI_PARAM_POS_W_FLOAT] = f32; |
| 2362 | params[SI_PARAM_FRONT_FACE] = f32; |
Marek Olšák | 5b06fc3 | 2014-05-06 18:12:40 +0200 | [diff] [blame] | 2363 | params[SI_PARAM_ANCILLARY] = i32; |
Christian König | 0666ffd | 2013-03-05 15:07:39 +0100 | [diff] [blame] | 2364 | params[SI_PARAM_SAMPLE_COVERAGE] = f32; |
| 2365 | params[SI_PARAM_POS_FIXED_PT] = f32; |
Marek Olšák | 13a1a8b | 2013-08-18 01:57:40 +0200 | [diff] [blame] | 2366 | num_params = SI_PARAM_POS_FIXED_PT+1; |
| 2367 | break; |
| 2368 | |
| 2369 | default: |
| 2370 | assert(0 && "unimplemented shader"); |
| 2371 | return; |
Christian König | c497321 | 2013-03-05 12:14:02 +0100 | [diff] [blame] | 2372 | } |
Christian König | 55fe5cc | 2013-03-04 16:30:06 +0100 | [diff] [blame] | 2373 | |
Marek Olšák | 13a1a8b | 2013-08-18 01:57:40 +0200 | [diff] [blame] | 2374 | assert(num_params <= Elements(params)); |
| 2375 | radeon_llvm_create_func(&si_shader_ctx->radeon_bld, params, num_params); |
Christian König | 55fe5cc | 2013-03-04 16:30:06 +0100 | [diff] [blame] | 2376 | radeon_llvm_shader_type(si_shader_ctx->radeon_bld.main_fn, si_shader_ctx->type); |
Christian König | cf9b31f | 2013-03-21 18:30:23 +0100 | [diff] [blame] | 2377 | |
Marek Olšák | 1829f9c | 2015-01-04 17:08:57 +0100 | [diff] [blame] | 2378 | if (shader->dx10_clamp_mode) |
| 2379 | LLVMAddTargetDependentFunctionAttr(si_shader_ctx->radeon_bld.main_fn, |
| 2380 | "enable-no-nans-fp-math", "true"); |
| 2381 | |
Marek Olšák | 13a1a8b | 2013-08-18 01:57:40 +0200 | [diff] [blame] | 2382 | for (i = 0; i <= last_sgpr; ++i) { |
| 2383 | LLVMValueRef P = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, i); |
Marek Olšák | aeb05f0 | 2014-09-30 18:13:06 +0200 | [diff] [blame] | 2384 | |
Vincent Lejeune | 6e51c2a | 2013-10-05 16:04:48 +0200 | [diff] [blame] | 2385 | /* We tell llvm that array inputs are passed by value to allow Sinking pass |
| 2386 | * to move load. Inputs are constant so this is fine. */ |
Marek Olšák | aeb05f0 | 2014-09-30 18:13:06 +0200 | [diff] [blame] | 2387 | if (i <= last_array_pointer) |
Vincent Lejeune | 6e51c2a | 2013-10-05 16:04:48 +0200 | [diff] [blame] | 2388 | LLVMAddAttribute(P, LLVMByValAttribute); |
Marek Olšák | aeb05f0 | 2014-09-30 18:13:06 +0200 | [diff] [blame] | 2389 | else |
| 2390 | LLVMAddAttribute(P, LLVMInRegAttribute); |
Christian König | cf9b31f | 2013-03-21 18:30:23 +0100 | [diff] [blame] | 2391 | } |
Michel Dänzer | a06ee5a | 2013-06-19 18:14:01 +0200 | [diff] [blame] | 2392 | |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2393 | if (bld_base->info && |
| 2394 | (bld_base->info->opcode_count[TGSI_OPCODE_DDX] > 0 || |
| 2395 | bld_base->info->opcode_count[TGSI_OPCODE_DDY] > 0)) |
Michel Dänzer | a06ee5a | 2013-06-19 18:14:01 +0200 | [diff] [blame] | 2396 | si_shader_ctx->ddxy_lds = |
| 2397 | LLVMAddGlobalInAddressSpace(gallivm->module, |
| 2398 | LLVMArrayType(i32, 64), |
| 2399 | "ddxy_lds", |
| 2400 | LOCAL_ADDR_SPACE); |
Christian König | 55fe5cc | 2013-03-04 16:30:06 +0100 | [diff] [blame] | 2401 | } |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 2402 | |
Christian König | 0f6cf2b | 2013-03-15 15:53:25 +0100 | [diff] [blame] | 2403 | static void preload_constants(struct si_shader_context *si_shader_ctx) |
| 2404 | { |
| 2405 | struct lp_build_tgsi_context * bld_base = &si_shader_ctx->radeon_bld.soa.bld_base; |
| 2406 | struct gallivm_state * gallivm = bld_base->base.gallivm; |
| 2407 | const struct tgsi_shader_info * info = bld_base->info; |
Marek Olšák | 2fd4200 | 2013-10-25 11:45:47 +0200 | [diff] [blame] | 2408 | unsigned buf; |
| 2409 | LLVMValueRef ptr = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_CONST); |
Christian König | 0f6cf2b | 2013-03-15 15:53:25 +0100 | [diff] [blame] | 2410 | |
Marek Olšák | ee2a818 | 2014-07-07 23:27:19 +0200 | [diff] [blame] | 2411 | for (buf = 0; buf < SI_NUM_CONST_BUFFERS; buf++) { |
Marek Olšák | 2fd4200 | 2013-10-25 11:45:47 +0200 | [diff] [blame] | 2412 | unsigned i, num_const = info->const_file_max[buf] + 1; |
Christian König | 0f6cf2b | 2013-03-15 15:53:25 +0100 | [diff] [blame] | 2413 | |
Marek Olšák | 2fd4200 | 2013-10-25 11:45:47 +0200 | [diff] [blame] | 2414 | if (num_const == 0) |
| 2415 | continue; |
Christian König | 0f6cf2b | 2013-03-15 15:53:25 +0100 | [diff] [blame] | 2416 | |
Marek Olšák | 2fd4200 | 2013-10-25 11:45:47 +0200 | [diff] [blame] | 2417 | /* Allocate space for the constant values */ |
| 2418 | si_shader_ctx->constants[buf] = CALLOC(num_const * 4, sizeof(LLVMValueRef)); |
Christian König | 0f6cf2b | 2013-03-15 15:53:25 +0100 | [diff] [blame] | 2419 | |
Marek Olšák | 2fd4200 | 2013-10-25 11:45:47 +0200 | [diff] [blame] | 2420 | /* Load the resource descriptor */ |
| 2421 | si_shader_ctx->const_resource[buf] = |
Marek Olšák | d787608 | 2014-09-26 23:06:32 +0200 | [diff] [blame] | 2422 | build_indexed_load_const(si_shader_ctx, ptr, lp_build_const_int32(gallivm, buf)); |
Christian König | 0f6cf2b | 2013-03-15 15:53:25 +0100 | [diff] [blame] | 2423 | |
Marek Olšák | 2fd4200 | 2013-10-25 11:45:47 +0200 | [diff] [blame] | 2424 | /* Load the constants, we rely on the code sinking to do the rest */ |
| 2425 | for (i = 0; i < num_const * 4; ++i) { |
Marek Olšák | 2fd4200 | 2013-10-25 11:45:47 +0200 | [diff] [blame] | 2426 | si_shader_ctx->constants[buf][i] = |
Marek Olšák | d787608 | 2014-09-26 23:06:32 +0200 | [diff] [blame] | 2427 | buffer_load_const(gallivm->builder, |
Marek Olšák | 250aa93 | 2014-05-06 14:10:47 +0200 | [diff] [blame] | 2428 | si_shader_ctx->const_resource[buf], |
| 2429 | lp_build_const_int32(gallivm, i * 4), |
| 2430 | bld_base->base.elem_type); |
Marek Olšák | 2fd4200 | 2013-10-25 11:45:47 +0200 | [diff] [blame] | 2431 | } |
Christian König | 0f6cf2b | 2013-03-15 15:53:25 +0100 | [diff] [blame] | 2432 | } |
| 2433 | } |
| 2434 | |
Christian König | 1c10018 | 2013-03-17 16:02:42 +0100 | [diff] [blame] | 2435 | static void preload_samplers(struct si_shader_context *si_shader_ctx) |
| 2436 | { |
| 2437 | struct lp_build_tgsi_context * bld_base = &si_shader_ctx->radeon_bld.soa.bld_base; |
| 2438 | struct gallivm_state * gallivm = bld_base->base.gallivm; |
| 2439 | const struct tgsi_shader_info * info = bld_base->info; |
| 2440 | |
| 2441 | unsigned i, num_samplers = info->file_max[TGSI_FILE_SAMPLER] + 1; |
| 2442 | |
| 2443 | LLVMValueRef res_ptr, samp_ptr; |
| 2444 | LLVMValueRef offset; |
| 2445 | |
| 2446 | if (num_samplers == 0) |
| 2447 | return; |
| 2448 | |
Christian König | 1c10018 | 2013-03-17 16:02:42 +0100 | [diff] [blame] | 2449 | res_ptr = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_RESOURCE); |
| 2450 | samp_ptr = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_SAMPLER); |
| 2451 | |
| 2452 | /* Load the resources and samplers, we rely on the code sinking to do the rest */ |
| 2453 | for (i = 0; i < num_samplers; ++i) { |
Christian König | 1c10018 | 2013-03-17 16:02:42 +0100 | [diff] [blame] | 2454 | /* Resource */ |
| 2455 | offset = lp_build_const_int32(gallivm, i); |
Marek Olšák | d787608 | 2014-09-26 23:06:32 +0200 | [diff] [blame] | 2456 | si_shader_ctx->resources[i] = build_indexed_load_const(si_shader_ctx, res_ptr, offset); |
Christian König | 1c10018 | 2013-03-17 16:02:42 +0100 | [diff] [blame] | 2457 | |
| 2458 | /* Sampler */ |
| 2459 | offset = lp_build_const_int32(gallivm, i); |
Marek Olšák | d787608 | 2014-09-26 23:06:32 +0200 | [diff] [blame] | 2460 | si_shader_ctx->samplers[i] = build_indexed_load_const(si_shader_ctx, samp_ptr, offset); |
Marek Olšák | 4855acd | 2013-08-06 15:08:54 +0200 | [diff] [blame] | 2461 | |
| 2462 | /* FMASK resource */ |
| 2463 | if (info->is_msaa_sampler[i]) { |
Marek Olšák | ee2a818 | 2014-07-07 23:27:19 +0200 | [diff] [blame] | 2464 | offset = lp_build_const_int32(gallivm, SI_FMASK_TEX_OFFSET + i); |
| 2465 | si_shader_ctx->resources[SI_FMASK_TEX_OFFSET + i] = |
Marek Olšák | d787608 | 2014-09-26 23:06:32 +0200 | [diff] [blame] | 2466 | build_indexed_load_const(si_shader_ctx, res_ptr, offset); |
Marek Olšák | 4855acd | 2013-08-06 15:08:54 +0200 | [diff] [blame] | 2467 | } |
Christian König | 1c10018 | 2013-03-17 16:02:42 +0100 | [diff] [blame] | 2468 | } |
| 2469 | } |
| 2470 | |
Marek Olšák | 8d03d92 | 2013-09-01 23:59:06 +0200 | [diff] [blame] | 2471 | static void preload_streamout_buffers(struct si_shader_context *si_shader_ctx) |
| 2472 | { |
| 2473 | struct lp_build_tgsi_context * bld_base = &si_shader_ctx->radeon_bld.soa.bld_base; |
| 2474 | struct gallivm_state * gallivm = bld_base->base.gallivm; |
| 2475 | unsigned i; |
| 2476 | |
Michel Dänzer | 67e385b | 2014-01-08 17:48:21 +0900 | [diff] [blame] | 2477 | if (si_shader_ctx->type != TGSI_PROCESSOR_VERTEX || |
| 2478 | si_shader_ctx->shader->key.vs.as_es || |
| 2479 | !si_shader_ctx->shader->selector->so.num_outputs) |
Marek Olšák | 8d03d92 | 2013-09-01 23:59:06 +0200 | [diff] [blame] | 2480 | return; |
| 2481 | |
| 2482 | LLVMValueRef buf_ptr = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, |
Michel Dänzer | f8e1601 | 2014-01-28 15:39:30 +0900 | [diff] [blame] | 2483 | SI_PARAM_RW_BUFFERS); |
Marek Olšák | 8d03d92 | 2013-09-01 23:59:06 +0200 | [diff] [blame] | 2484 | |
| 2485 | /* Load the resources, we rely on the code sinking to do the rest */ |
| 2486 | for (i = 0; i < 4; ++i) { |
| 2487 | if (si_shader_ctx->shader->selector->so.stride[i]) { |
Michel Dänzer | f8e1601 | 2014-01-28 15:39:30 +0900 | [diff] [blame] | 2488 | LLVMValueRef offset = lp_build_const_int32(gallivm, |
Marek Olšák | ee2a818 | 2014-07-07 23:27:19 +0200 | [diff] [blame] | 2489 | SI_SO_BUF_OFFSET + i); |
Marek Olšák | 8d03d92 | 2013-09-01 23:59:06 +0200 | [diff] [blame] | 2490 | |
Marek Olšák | d787608 | 2014-09-26 23:06:32 +0200 | [diff] [blame] | 2491 | si_shader_ctx->so_buffers[i] = build_indexed_load_const(si_shader_ctx, buf_ptr, offset); |
Marek Olšák | 8d03d92 | 2013-09-01 23:59:06 +0200 | [diff] [blame] | 2492 | } |
| 2493 | } |
| 2494 | } |
| 2495 | |
Marek Olšák | fc3b335 | 2014-10-05 13:33:40 +0200 | [diff] [blame] | 2496 | /** |
| 2497 | * Load ESGS and GSVS ring buffer resource descriptors and save the variables |
| 2498 | * for later use. |
| 2499 | */ |
| 2500 | static void preload_ring_buffers(struct si_shader_context *si_shader_ctx) |
| 2501 | { |
| 2502 | struct gallivm_state *gallivm = |
| 2503 | si_shader_ctx->radeon_bld.soa.bld_base.base.gallivm; |
| 2504 | |
| 2505 | LLVMValueRef buf_ptr = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, |
| 2506 | SI_PARAM_RW_BUFFERS); |
| 2507 | |
| 2508 | if ((si_shader_ctx->type == TGSI_PROCESSOR_VERTEX && |
| 2509 | si_shader_ctx->shader->key.vs.as_es) || |
| 2510 | si_shader_ctx->type == TGSI_PROCESSOR_GEOMETRY) { |
| 2511 | LLVMValueRef offset = lp_build_const_int32(gallivm, SI_RING_ESGS); |
| 2512 | |
| 2513 | si_shader_ctx->esgs_ring = |
| 2514 | build_indexed_load_const(si_shader_ctx, buf_ptr, offset); |
| 2515 | } |
| 2516 | |
| 2517 | if (si_shader_ctx->type == TGSI_PROCESSOR_GEOMETRY || |
| 2518 | si_shader_ctx->shader->is_gs_copy_shader) { |
| 2519 | LLVMValueRef offset = lp_build_const_int32(gallivm, SI_RING_GSVS); |
| 2520 | |
| 2521 | si_shader_ctx->gsvs_ring = |
| 2522 | build_indexed_load_const(si_shader_ctx, buf_ptr, offset); |
| 2523 | } |
| 2524 | } |
| 2525 | |
Tom Stellard | 2397a72 | 2014-12-10 09:13:59 -0500 | [diff] [blame] | 2526 | void si_shader_binary_read_config(const struct si_screen *sscreen, |
Tom Stellard | 1f4e48d | 2014-09-25 18:11:24 -0700 | [diff] [blame] | 2527 | struct si_shader *shader, |
| 2528 | unsigned symbol_offset) |
Tom Stellard | 302f53d | 2012-10-25 13:50:10 -0400 | [diff] [blame] | 2529 | { |
Tom Stellard | 302f53d | 2012-10-25 13:50:10 -0400 | [diff] [blame] | 2530 | unsigned i; |
Tom Stellard | 1f4e48d | 2014-09-25 18:11:24 -0700 | [diff] [blame] | 2531 | const unsigned char *config = |
Tom Stellard | 2397a72 | 2014-12-10 09:13:59 -0500 | [diff] [blame] | 2532 | radeon_shader_binary_config_start(&shader->binary, |
| 2533 | symbol_offset); |
Tom Stellard | 302f53d | 2012-10-25 13:50:10 -0400 | [diff] [blame] | 2534 | |
Tom Stellard | d50343d | 2013-04-04 16:21:06 -0400 | [diff] [blame] | 2535 | /* XXX: We may be able to emit some of these values directly rather than |
| 2536 | * extracting fields to be emitted later. |
| 2537 | */ |
Tom Stellard | 1f4e48d | 2014-09-25 18:11:24 -0700 | [diff] [blame] | 2538 | |
Tom Stellard | 2397a72 | 2014-12-10 09:13:59 -0500 | [diff] [blame] | 2539 | for (i = 0; i < shader->binary.config_size_per_symbol; i+= 8) { |
Tom Stellard | 1f4e48d | 2014-09-25 18:11:24 -0700 | [diff] [blame] | 2540 | unsigned reg = util_le32_to_cpu(*(uint32_t*)(config + i)); |
| 2541 | unsigned value = util_le32_to_cpu(*(uint32_t*)(config + i + 4)); |
Tom Stellard | d50343d | 2013-04-04 16:21:06 -0400 | [diff] [blame] | 2542 | switch (reg) { |
| 2543 | case R_00B028_SPI_SHADER_PGM_RSRC1_PS: |
| 2544 | case R_00B128_SPI_SHADER_PGM_RSRC1_VS: |
| 2545 | case R_00B228_SPI_SHADER_PGM_RSRC1_GS: |
| 2546 | case R_00B848_COMPUTE_PGM_RSRC1: |
Tom Stellard | 1f4e48d | 2014-09-25 18:11:24 -0700 | [diff] [blame] | 2547 | shader->num_sgprs = MAX2(shader->num_sgprs, (G_00B028_SGPRS(value) + 1) * 8); |
| 2548 | shader->num_vgprs = MAX2(shader->num_vgprs, (G_00B028_VGPRS(value) + 1) * 4); |
Tom Stellard | bbfa1c3 | 2015-03-06 14:53:00 +0000 | [diff] [blame^] | 2549 | shader->float_mode = G_00B028_FLOAT_MODE(value); |
Tom Stellard | d50343d | 2013-04-04 16:21:06 -0400 | [diff] [blame] | 2550 | break; |
Michel Dänzer | a06ee5a | 2013-06-19 18:14:01 +0200 | [diff] [blame] | 2551 | case R_00B02C_SPI_SHADER_PGM_RSRC2_PS: |
Tom Stellard | 1f4e48d | 2014-09-25 18:11:24 -0700 | [diff] [blame] | 2552 | shader->lds_size = MAX2(shader->lds_size, G_00B02C_EXTRA_LDS_SIZE(value)); |
Michel Dänzer | a06ee5a | 2013-06-19 18:14:01 +0200 | [diff] [blame] | 2553 | break; |
| 2554 | case R_00B84C_COMPUTE_PGM_RSRC2: |
Tom Stellard | 1f4e48d | 2014-09-25 18:11:24 -0700 | [diff] [blame] | 2555 | shader->lds_size = MAX2(shader->lds_size, G_00B84C_LDS_SIZE(value)); |
Michel Dänzer | a06ee5a | 2013-06-19 18:14:01 +0200 | [diff] [blame] | 2556 | break; |
Tom Stellard | d50343d | 2013-04-04 16:21:06 -0400 | [diff] [blame] | 2557 | case R_0286CC_SPI_PS_INPUT_ENA: |
| 2558 | shader->spi_ps_input_ena = value; |
| 2559 | break; |
Tom Stellard | 2397a72 | 2014-12-10 09:13:59 -0500 | [diff] [blame] | 2560 | case R_0286E8_SPI_TMPRING_SIZE: |
Tom Stellard | b0f7803 | 2014-07-18 14:45:18 -0400 | [diff] [blame] | 2561 | case R_00B860_COMPUTE_TMPRING_SIZE: |
| 2562 | /* WAVESIZE is in units of 256 dwords. */ |
| 2563 | shader->scratch_bytes_per_wave = |
| 2564 | G_00B860_WAVESIZE(value) * 256 * 4 * 1; |
| 2565 | break; |
Tom Stellard | d50343d | 2013-04-04 16:21:06 -0400 | [diff] [blame] | 2566 | default: |
| 2567 | fprintf(stderr, "Warning: Compiler emitted unknown " |
| 2568 | "config register: 0x%x\n", reg); |
| 2569 | break; |
| 2570 | } |
| 2571 | } |
Tom Stellard | 1f4e48d | 2014-09-25 18:11:24 -0700 | [diff] [blame] | 2572 | } |
| 2573 | |
Tom Stellard | 2397a72 | 2014-12-10 09:13:59 -0500 | [diff] [blame] | 2574 | void si_shader_apply_scratch_relocs(struct si_context *sctx, |
| 2575 | struct si_shader *shader, |
| 2576 | uint64_t scratch_va) |
| 2577 | { |
| 2578 | unsigned i; |
| 2579 | uint32_t scratch_rsrc_dword0 = scratch_va & 0xffffffff; |
| 2580 | uint32_t scratch_rsrc_dword1 = |
| 2581 | S_008F04_BASE_ADDRESS_HI(scratch_va >> 32) |
| 2582 | | S_008F04_STRIDE(shader->scratch_bytes_per_wave / 64); |
| 2583 | |
| 2584 | for (i = 0 ; i < shader->binary.reloc_count; i++) { |
| 2585 | const struct radeon_shader_reloc *reloc = |
| 2586 | &shader->binary.relocs[i]; |
| 2587 | if (!strcmp(scratch_rsrc_dword0_symbol, reloc->name)) { |
| 2588 | util_memcpy_cpu_to_le32(shader->binary.code + reloc->offset, |
| 2589 | &scratch_rsrc_dword0, 4); |
| 2590 | } else if (!strcmp(scratch_rsrc_dword1_symbol, reloc->name)) { |
| 2591 | util_memcpy_cpu_to_le32(shader->binary.code + reloc->offset, |
| 2592 | &scratch_rsrc_dword1, 4); |
| 2593 | } |
| 2594 | } |
| 2595 | } |
| 2596 | |
Tom Stellard | 1f4e48d | 2014-09-25 18:11:24 -0700 | [diff] [blame] | 2597 | int si_shader_binary_read(struct si_screen *sscreen, |
| 2598 | struct si_shader *shader, |
| 2599 | const struct radeon_shader_binary *binary) |
| 2600 | { |
| 2601 | |
| 2602 | unsigned i; |
| 2603 | unsigned code_size; |
| 2604 | unsigned char *ptr; |
| 2605 | bool dump = r600_can_dump_shader(&sscreen->b, |
| 2606 | shader->selector ? shader->selector->tokens : NULL); |
| 2607 | |
| 2608 | if (dump && !binary->disassembled) { |
| 2609 | fprintf(stderr, "SI CODE:\n"); |
| 2610 | for (i = 0; i < binary->code_size; i+=4 ) { |
| 2611 | fprintf(stderr, "@0x%x: %02x%02x%02x%02x\n", i, binary->code[i + 3], |
| 2612 | binary->code[i + 2], binary->code[i + 1], |
| 2613 | binary->code[i]); |
| 2614 | } |
| 2615 | } |
| 2616 | |
Tom Stellard | 2397a72 | 2014-12-10 09:13:59 -0500 | [diff] [blame] | 2617 | si_shader_binary_read_config(sscreen, shader, 0); |
Tom Stellard | 302f53d | 2012-10-25 13:50:10 -0400 | [diff] [blame] | 2618 | |
| 2619 | /* copy new shader */ |
Tom Stellard | 1f4e48d | 2014-09-25 18:11:24 -0700 | [diff] [blame] | 2620 | code_size = binary->code_size + binary->rodata_size; |
Marek Olšák | a81c3e0 | 2013-08-14 01:04:39 +0200 | [diff] [blame] | 2621 | r600_resource_reference(&shader->bo, NULL); |
Marek Olšák | 1abb1a9 | 2014-09-17 22:44:22 +0200 | [diff] [blame] | 2622 | shader->bo = si_resource_create_custom(&sscreen->b.b, PIPE_USAGE_IMMUTABLE, |
Tom Stellard | 9ba3105 | 2014-07-14 16:49:08 -0400 | [diff] [blame] | 2623 | code_size); |
Tom Stellard | 302f53d | 2012-10-25 13:50:10 -0400 | [diff] [blame] | 2624 | if (shader->bo == NULL) { |
| 2625 | return -ENOMEM; |
| 2626 | } |
| 2627 | |
Tom Stellard | 1f4e48d | 2014-09-25 18:11:24 -0700 | [diff] [blame] | 2628 | |
| 2629 | ptr = sscreen->b.ws->buffer_map(shader->bo->cs_buf, NULL, PIPE_TRANSFER_READ_WRITE); |
| 2630 | util_memcpy_cpu_to_le32(ptr, binary->code, binary->code_size); |
| 2631 | if (binary->rodata_size > 0) { |
| 2632 | ptr += binary->code_size; |
| 2633 | util_memcpy_cpu_to_le32(ptr, binary->rodata, binary->rodata_size); |
Tom Stellard | 302f53d | 2012-10-25 13:50:10 -0400 | [diff] [blame] | 2634 | } |
Tom Stellard | 6f0c1f2 | 2014-07-18 15:10:52 -0400 | [diff] [blame] | 2635 | |
Marek Olšák | 1abb1a9 | 2014-09-17 22:44:22 +0200 | [diff] [blame] | 2636 | sscreen->b.ws->buffer_unmap(shader->bo->cs_buf); |
Tom Stellard | 302f53d | 2012-10-25 13:50:10 -0400 | [diff] [blame] | 2637 | |
Tom Stellard | 1f4e48d | 2014-09-25 18:11:24 -0700 | [diff] [blame] | 2638 | return 0; |
| 2639 | } |
Tom Stellard | 302f53d | 2012-10-25 13:50:10 -0400 | [diff] [blame] | 2640 | |
Tom Stellard | 1f4e48d | 2014-09-25 18:11:24 -0700 | [diff] [blame] | 2641 | int si_compile_llvm(struct si_screen *sscreen, struct si_shader *shader, |
| 2642 | LLVMModuleRef mod) |
| 2643 | { |
| 2644 | int r = 0; |
Tom Stellard | 1f4e48d | 2014-09-25 18:11:24 -0700 | [diff] [blame] | 2645 | bool dump = r600_can_dump_shader(&sscreen->b, |
| 2646 | shader->selector ? shader->selector->tokens : NULL); |
Tom Stellard | 2397a72 | 2014-12-10 09:13:59 -0500 | [diff] [blame] | 2647 | r = radeon_llvm_compile(mod, &shader->binary, |
Tom Stellard | 761e36b | 2014-10-15 12:24:30 -0400 | [diff] [blame] | 2648 | r600_get_llvm_processor_name(sscreen->b.family), dump, sscreen->tm); |
Tom Stellard | 1f4e48d | 2014-09-25 18:11:24 -0700 | [diff] [blame] | 2649 | |
| 2650 | if (r) { |
| 2651 | return r; |
| 2652 | } |
Tom Stellard | 2397a72 | 2014-12-10 09:13:59 -0500 | [diff] [blame] | 2653 | r = si_shader_binary_read(sscreen, shader, &shader->binary); |
| 2654 | |
| 2655 | FREE(shader->binary.config); |
| 2656 | FREE(shader->binary.rodata); |
| 2657 | FREE(shader->binary.global_symbol_offsets); |
| 2658 | if (shader->scratch_bytes_per_wave == 0) { |
| 2659 | FREE(shader->binary.code); |
| 2660 | FREE(shader->binary.relocs); |
| 2661 | memset(&shader->binary, 0, sizeof(shader->binary)); |
| 2662 | } |
Darren Powell | bc86690 | 2014-03-31 18:00:28 -0400 | [diff] [blame] | 2663 | return r; |
Tom Stellard | 302f53d | 2012-10-25 13:50:10 -0400 | [diff] [blame] | 2664 | } |
| 2665 | |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2666 | /* Generate code for the hardware VS shader stage to go with a geometry shader */ |
Marek Olšák | 1abb1a9 | 2014-09-17 22:44:22 +0200 | [diff] [blame] | 2667 | static int si_generate_gs_copy_shader(struct si_screen *sscreen, |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2668 | struct si_shader_context *si_shader_ctx, |
Marek Olšák | 68d36c0 | 2014-09-30 18:15:17 +0200 | [diff] [blame] | 2669 | struct si_shader *gs, bool dump) |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2670 | { |
| 2671 | struct gallivm_state *gallivm = &si_shader_ctx->radeon_bld.gallivm; |
| 2672 | struct lp_build_tgsi_context *bld_base = &si_shader_ctx->radeon_bld.soa.bld_base; |
| 2673 | struct lp_build_context *base = &bld_base->base; |
| 2674 | struct lp_build_context *uint = &bld_base->uint_bld; |
Marek Olšák | 8c37c16 | 2014-09-16 18:40:07 +0200 | [diff] [blame] | 2675 | struct si_shader *shader = si_shader_ctx->shader; |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2676 | struct si_shader_output_values *outputs; |
Marek Olšák | 02134cf | 2014-10-04 22:07:50 +0200 | [diff] [blame] | 2677 | struct tgsi_shader_info *gsinfo = &gs->selector->info; |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2678 | LLVMValueRef args[9]; |
| 2679 | int i, r; |
| 2680 | |
Marek Olšák | 02134cf | 2014-10-04 22:07:50 +0200 | [diff] [blame] | 2681 | outputs = MALLOC(gsinfo->num_outputs * sizeof(outputs[0])); |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2682 | |
| 2683 | si_shader_ctx->type = TGSI_PROCESSOR_VERTEX; |
Marek Olšák | 1f6c0b5 | 2014-09-23 19:42:28 +0200 | [diff] [blame] | 2684 | shader->is_gs_copy_shader = true; |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2685 | |
| 2686 | radeon_llvm_context_init(&si_shader_ctx->radeon_bld); |
| 2687 | |
| 2688 | create_meta_data(si_shader_ctx); |
| 2689 | create_function(si_shader_ctx); |
| 2690 | preload_streamout_buffers(si_shader_ctx); |
Marek Olšák | fc3b335 | 2014-10-05 13:33:40 +0200 | [diff] [blame] | 2691 | preload_ring_buffers(si_shader_ctx); |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2692 | |
Marek Olšák | fc3b335 | 2014-10-05 13:33:40 +0200 | [diff] [blame] | 2693 | args[0] = si_shader_ctx->gsvs_ring; |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2694 | args[1] = lp_build_mul_imm(uint, |
| 2695 | LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, |
| 2696 | si_shader_ctx->param_vertex_id), |
| 2697 | 4); |
| 2698 | args[3] = uint->zero; |
| 2699 | args[4] = uint->one; /* OFFEN */ |
| 2700 | args[5] = uint->zero; /* IDXEN */ |
| 2701 | args[6] = uint->one; /* GLC */ |
| 2702 | args[7] = uint->one; /* SLC */ |
| 2703 | args[8] = uint->zero; /* TFE */ |
| 2704 | |
| 2705 | /* Fetch vertex data from GSVS ring */ |
Marek Olšák | 02134cf | 2014-10-04 22:07:50 +0200 | [diff] [blame] | 2706 | for (i = 0; i < gsinfo->num_outputs; ++i) { |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2707 | unsigned chan; |
| 2708 | |
Marek Olšák | 02134cf | 2014-10-04 22:07:50 +0200 | [diff] [blame] | 2709 | outputs[i].name = gsinfo->output_semantic_name[i]; |
Marek Olšák | 02134cf | 2014-10-04 22:07:50 +0200 | [diff] [blame] | 2710 | outputs[i].sid = gsinfo->output_semantic_index[i]; |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2711 | |
| 2712 | for (chan = 0; chan < 4; chan++) { |
| 2713 | args[2] = lp_build_const_int32(gallivm, |
| 2714 | (i * 4 + chan) * |
Marek Olšák | 0a2d6f0 | 2014-09-30 16:55:36 +0200 | [diff] [blame] | 2715 | gs->selector->gs_max_out_vertices * 16 * 4); |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2716 | |
| 2717 | outputs[i].values[chan] = |
| 2718 | LLVMBuildBitCast(gallivm->builder, |
| 2719 | build_intrinsic(gallivm->builder, |
| 2720 | "llvm.SI.buffer.load.dword.i32.i32", |
| 2721 | LLVMInt32TypeInContext(gallivm->context), |
| 2722 | args, 9, |
| 2723 | LLVMReadOnlyAttribute | LLVMNoUnwindAttribute), |
| 2724 | base->elem_type, ""); |
| 2725 | } |
| 2726 | } |
| 2727 | |
Marek Olšák | 02134cf | 2014-10-04 22:07:50 +0200 | [diff] [blame] | 2728 | si_llvm_export_vs(bld_base, outputs, gsinfo->num_outputs); |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2729 | |
| 2730 | radeon_llvm_finalize_module(&si_shader_ctx->radeon_bld); |
| 2731 | |
| 2732 | if (dump) |
| 2733 | fprintf(stderr, "Copy Vertex Shader for Geometry Shader:\n\n"); |
| 2734 | |
Marek Olšák | 1abb1a9 | 2014-09-17 22:44:22 +0200 | [diff] [blame] | 2735 | r = si_compile_llvm(sscreen, si_shader_ctx->shader, |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2736 | bld_base->base.gallivm->module); |
| 2737 | |
| 2738 | radeon_llvm_dispose(&si_shader_ctx->radeon_bld); |
| 2739 | |
| 2740 | FREE(outputs); |
| 2741 | return r; |
| 2742 | } |
| 2743 | |
Marek Olšák | ffd701e | 2015-02-22 14:58:30 +0100 | [diff] [blame] | 2744 | static void si_dump_key(unsigned shader, union si_shader_key *key) |
| 2745 | { |
| 2746 | int i; |
| 2747 | |
| 2748 | fprintf(stderr, "SHADER KEY\n"); |
| 2749 | |
| 2750 | switch (shader) { |
| 2751 | case PIPE_SHADER_VERTEX: |
| 2752 | fprintf(stderr, " instance_divisors = {"); |
| 2753 | for (i = 0; i < Elements(key->vs.instance_divisors); i++) |
| 2754 | fprintf(stderr, !i ? "%u" : ", %u", |
| 2755 | key->vs.instance_divisors[i]); |
| 2756 | fprintf(stderr, "}\n"); |
| 2757 | |
| 2758 | if (key->vs.as_es) |
| 2759 | fprintf(stderr, " gs_used_inputs = 0x%"PRIx64"\n", |
| 2760 | key->vs.gs_used_inputs); |
| 2761 | fprintf(stderr, " as_es = %u\n", key->vs.as_es); |
| 2762 | break; |
| 2763 | |
| 2764 | case PIPE_SHADER_GEOMETRY: |
| 2765 | break; |
| 2766 | |
| 2767 | case PIPE_SHADER_FRAGMENT: |
| 2768 | fprintf(stderr, " export_16bpc = 0x%X\n", key->ps.export_16bpc); |
| 2769 | fprintf(stderr, " last_cbuf = %u\n", key->ps.last_cbuf); |
| 2770 | fprintf(stderr, " color_two_side = %u\n", key->ps.color_two_side); |
| 2771 | fprintf(stderr, " alpha_func = %u\n", key->ps.alpha_func); |
| 2772 | fprintf(stderr, " alpha_to_one = %u\n", key->ps.alpha_to_one); |
| 2773 | fprintf(stderr, " poly_stipple = %u\n", key->ps.poly_stipple); |
| 2774 | break; |
| 2775 | |
| 2776 | default: |
| 2777 | assert(0); |
| 2778 | } |
| 2779 | } |
| 2780 | |
Marek Olšák | 1abb1a9 | 2014-09-17 22:44:22 +0200 | [diff] [blame] | 2781 | int si_shader_create(struct si_screen *sscreen, struct si_shader *shader) |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 2782 | { |
Marek Olšák | 2774abd | 2014-09-16 18:45:33 +0200 | [diff] [blame] | 2783 | struct si_shader_selector *sel = shader->selector; |
Marek Olšák | 6c5af1d | 2015-01-31 21:43:37 +0100 | [diff] [blame] | 2784 | struct tgsi_token *tokens = sel->tokens; |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 2785 | struct si_shader_context si_shader_ctx; |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 2786 | struct lp_build_tgsi_context * bld_base; |
Marek Olšák | 6c5af1d | 2015-01-31 21:43:37 +0100 | [diff] [blame] | 2787 | struct tgsi_shader_info stipple_shader_info; |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 2788 | LLVMModuleRef mod; |
Tom Stellard | 302f53d | 2012-10-25 13:50:10 -0400 | [diff] [blame] | 2789 | int r = 0; |
Marek Olšák | 6c5af1d | 2015-01-31 21:43:37 +0100 | [diff] [blame] | 2790 | bool poly_stipple = sel->type == PIPE_SHADER_FRAGMENT && |
| 2791 | shader->key.ps.poly_stipple; |
Marek Olšák | 1abb1a9 | 2014-09-17 22:44:22 +0200 | [diff] [blame] | 2792 | bool dump = r600_can_dump_shader(&sscreen->b, sel->tokens); |
Michel Dänzer | e1df0d4 | 2014-01-15 12:31:07 +0900 | [diff] [blame] | 2793 | |
Marek Olšák | 6c5af1d | 2015-01-31 21:43:37 +0100 | [diff] [blame] | 2794 | if (poly_stipple) { |
| 2795 | tokens = util_pstipple_create_fragment_shader(tokens, NULL, |
| 2796 | SI_POLY_STIPPLE_SAMPLER); |
| 2797 | tgsi_scan_shader(tokens, &stipple_shader_info); |
| 2798 | } |
| 2799 | |
Michel Dänzer | e1df0d4 | 2014-01-15 12:31:07 +0900 | [diff] [blame] | 2800 | /* Dump TGSI code before doing TGSI->LLVM conversion in case the |
| 2801 | * conversion fails. */ |
| 2802 | if (dump) { |
Marek Olšák | ffd701e | 2015-02-22 14:58:30 +0100 | [diff] [blame] | 2803 | si_dump_key(sel->type, &shader->key); |
Marek Olšák | 6c5af1d | 2015-01-31 21:43:37 +0100 | [diff] [blame] | 2804 | tgsi_dump(tokens, 0); |
Michel Dänzer | e1df0d4 | 2014-01-15 12:31:07 +0900 | [diff] [blame] | 2805 | si_dump_streamout(&sel->so); |
| 2806 | } |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 2807 | |
Marek Olšák | 8c37c16 | 2014-09-16 18:40:07 +0200 | [diff] [blame] | 2808 | assert(shader->nparam == 0); |
Michel Dänzer | 82e38ac | 2012-09-27 16:39:26 +0200 | [diff] [blame] | 2809 | |
Michel Dänzer | cfebaf9 | 2012-08-31 19:04:08 +0200 | [diff] [blame] | 2810 | memset(&si_shader_ctx, 0, sizeof(si_shader_ctx)); |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 2811 | radeon_llvm_context_init(&si_shader_ctx.radeon_bld); |
| 2812 | bld_base = &si_shader_ctx.radeon_bld.soa.bld_base; |
| 2813 | |
Marek Olšák | 1829f9c | 2015-01-04 17:08:57 +0100 | [diff] [blame] | 2814 | if (sel->type != PIPE_SHADER_COMPUTE) |
| 2815 | shader->dx10_clamp_mode = true; |
| 2816 | |
Marek Olšák | 5233568 | 2014-09-30 15:12:09 +0200 | [diff] [blame] | 2817 | if (sel->info.uses_kill) |
Marek Olšák | adc5797 | 2014-09-19 17:07:07 +0200 | [diff] [blame] | 2818 | shader->db_shader_control |= S_02880C_KILL_ENABLE(1); |
| 2819 | |
Marek Olšák | 5233568 | 2014-09-30 15:12:09 +0200 | [diff] [blame] | 2820 | shader->uses_instanceid = sel->info.uses_instanceid; |
Marek Olšák | 6c5af1d | 2015-01-31 21:43:37 +0100 | [diff] [blame] | 2821 | bld_base->info = poly_stipple ? &stipple_shader_info : &sel->info; |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 2822 | bld_base->emit_fetch_funcs[TGSI_FILE_CONSTANT] = fetch_constant; |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 2823 | |
| 2824 | bld_base->op_actions[TGSI_OPCODE_TEX] = tex_action; |
Marek Olšák | 2484daa | 2014-04-22 21:23:29 +0200 | [diff] [blame] | 2825 | bld_base->op_actions[TGSI_OPCODE_TEX2] = tex_action; |
Marek Olšák | 877bb52 | 2014-06-25 03:12:46 +0200 | [diff] [blame] | 2826 | bld_base->op_actions[TGSI_OPCODE_TXB] = tex_action; |
| 2827 | bld_base->op_actions[TGSI_OPCODE_TXB2] = tex_action; |
Marek Olšák | 877bb52 | 2014-06-25 03:12:46 +0200 | [diff] [blame] | 2828 | bld_base->op_actions[TGSI_OPCODE_TXD] = tex_action; |
Marek Olšák | 877bb52 | 2014-06-25 03:12:46 +0200 | [diff] [blame] | 2829 | bld_base->op_actions[TGSI_OPCODE_TXF] = tex_action; |
| 2830 | bld_base->op_actions[TGSI_OPCODE_TXL] = tex_action; |
| 2831 | bld_base->op_actions[TGSI_OPCODE_TXL2] = tex_action; |
Michel Dänzer | c2bae6b | 2012-08-02 17:19:22 +0200 | [diff] [blame] | 2832 | bld_base->op_actions[TGSI_OPCODE_TXP] = tex_action; |
Michel Dänzer | 0495adb | 2013-05-06 12:45:14 +0200 | [diff] [blame] | 2833 | bld_base->op_actions[TGSI_OPCODE_TXQ] = txq_action; |
Marek Olšák | 877bb52 | 2014-06-25 03:12:46 +0200 | [diff] [blame] | 2834 | bld_base->op_actions[TGSI_OPCODE_TG4] = tex_action; |
| 2835 | bld_base->op_actions[TGSI_OPCODE_LODQ] = tex_action; |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 2836 | |
Michel Dänzer | a06ee5a | 2013-06-19 18:14:01 +0200 | [diff] [blame] | 2837 | bld_base->op_actions[TGSI_OPCODE_DDX].emit = si_llvm_emit_ddxy; |
| 2838 | bld_base->op_actions[TGSI_OPCODE_DDY].emit = si_llvm_emit_ddxy; |
Michel Dänzer | a06ee5a | 2013-06-19 18:14:01 +0200 | [diff] [blame] | 2839 | |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2840 | bld_base->op_actions[TGSI_OPCODE_EMIT].emit = si_llvm_emit_vertex; |
| 2841 | bld_base->op_actions[TGSI_OPCODE_ENDPRIM].emit = si_llvm_emit_primitive; |
| 2842 | |
Marek Olšák | 645b471 | 2014-11-20 22:16:09 +0100 | [diff] [blame] | 2843 | if (HAVE_LLVM >= 0x0306) { |
| 2844 | bld_base->op_actions[TGSI_OPCODE_MAX].emit = build_tgsi_intrinsic_nomem; |
| 2845 | bld_base->op_actions[TGSI_OPCODE_MAX].intr_name = "llvm.maxnum.f32"; |
| 2846 | bld_base->op_actions[TGSI_OPCODE_MIN].emit = build_tgsi_intrinsic_nomem; |
| 2847 | bld_base->op_actions[TGSI_OPCODE_MIN].intr_name = "llvm.minnum.f32"; |
| 2848 | } |
| 2849 | |
Christian König | e4ed587 | 2013-03-21 18:02:52 +0100 | [diff] [blame] | 2850 | si_shader_ctx.radeon_bld.load_system_value = declare_system_value; |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 2851 | si_shader_ctx.shader = shader; |
Marek Olšák | 6c5af1d | 2015-01-31 21:43:37 +0100 | [diff] [blame] | 2852 | si_shader_ctx.type = tgsi_get_processor_type(tokens); |
Marek Olšák | 368b0a7 | 2014-12-28 21:51:35 +0100 | [diff] [blame] | 2853 | si_shader_ctx.screen = sscreen; |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 2854 | |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 2855 | switch (si_shader_ctx.type) { |
| 2856 | case TGSI_PROCESSOR_VERTEX: |
| 2857 | si_shader_ctx.radeon_bld.load_input = declare_input_vs; |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2858 | if (shader->key.vs.as_es) { |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2859 | bld_base->emit_epilogue = si_llvm_emit_es_epilogue; |
| 2860 | } else { |
| 2861 | bld_base->emit_epilogue = si_llvm_emit_vs_epilogue; |
| 2862 | } |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 2863 | break; |
Marek Olšák | 8908fae | 2014-09-30 15:48:22 +0200 | [diff] [blame] | 2864 | case TGSI_PROCESSOR_GEOMETRY: |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2865 | bld_base->emit_fetch_funcs[TGSI_FILE_INPUT] = fetch_input_gs; |
| 2866 | bld_base->emit_epilogue = si_llvm_emit_gs_epilogue; |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2867 | break; |
Marek Olšák | 8908fae | 2014-09-30 15:48:22 +0200 | [diff] [blame] | 2868 | case TGSI_PROCESSOR_FRAGMENT: |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 2869 | si_shader_ctx.radeon_bld.load_input = declare_input_fs; |
| 2870 | bld_base->emit_epilogue = si_llvm_emit_fs_epilogue; |
Marek Olšák | 6a2b383 | 2014-06-14 17:58:30 +0200 | [diff] [blame] | 2871 | |
Marek Olšák | 0c4bc1e | 2014-10-02 16:36:51 +0200 | [diff] [blame] | 2872 | switch (sel->info.properties[TGSI_PROPERTY_FS_DEPTH_LAYOUT]) { |
Marek Olšák | 8908fae | 2014-09-30 15:48:22 +0200 | [diff] [blame] | 2873 | case TGSI_FS_DEPTH_LAYOUT_GREATER: |
| 2874 | shader->db_shader_control |= |
| 2875 | S_02880C_CONSERVATIVE_Z_EXPORT(V_02880C_EXPORT_GREATER_THAN_Z); |
| 2876 | break; |
| 2877 | case TGSI_FS_DEPTH_LAYOUT_LESS: |
| 2878 | shader->db_shader_control |= |
| 2879 | S_02880C_CONSERVATIVE_Z_EXPORT(V_02880C_EXPORT_LESS_THAN_Z); |
| 2880 | break; |
Marek Olšák | 6a2b383 | 2014-06-14 17:58:30 +0200 | [diff] [blame] | 2881 | } |
Michel Dänzer | 51f89a0 | 2013-12-09 15:33:53 +0900 | [diff] [blame] | 2882 | break; |
| 2883 | default: |
| 2884 | assert(!"Unsupported shader type"); |
| 2885 | return -1; |
| 2886 | } |
| 2887 | |
Christian König | 206f059 | 2013-03-20 14:37:21 +0100 | [diff] [blame] | 2888 | create_meta_data(&si_shader_ctx); |
Christian König | 55fe5cc | 2013-03-04 16:30:06 +0100 | [diff] [blame] | 2889 | create_function(&si_shader_ctx); |
Christian König | 0f6cf2b | 2013-03-15 15:53:25 +0100 | [diff] [blame] | 2890 | preload_constants(&si_shader_ctx); |
Christian König | 1c10018 | 2013-03-17 16:02:42 +0100 | [diff] [blame] | 2891 | preload_samplers(&si_shader_ctx); |
Marek Olšák | 8d03d92 | 2013-09-01 23:59:06 +0200 | [diff] [blame] | 2892 | preload_streamout_buffers(&si_shader_ctx); |
Marek Olšák | fc3b335 | 2014-10-05 13:33:40 +0200 | [diff] [blame] | 2893 | preload_ring_buffers(&si_shader_ctx); |
Christian König | b8f4ca3 | 2013-03-04 15:35:30 +0100 | [diff] [blame] | 2894 | |
Michel Dänzer | f07a96d | 2014-01-08 18:45:10 +0900 | [diff] [blame] | 2895 | if (si_shader_ctx.type == TGSI_PROCESSOR_GEOMETRY) { |
| 2896 | si_shader_ctx.gs_next_vertex = |
| 2897 | lp_build_alloca(bld_base->base.gallivm, |
| 2898 | bld_base->uint_bld.elem_type, ""); |
| 2899 | } |
| 2900 | |
Marek Olšák | 6c5af1d | 2015-01-31 21:43:37 +0100 | [diff] [blame] | 2901 | if (!lp_build_tgsi_llvm(bld_base, tokens)) { |
Michel Dänzer | 82cd9c0 | 2012-08-08 15:35:42 +0200 | [diff] [blame] | 2902 | fprintf(stderr, "Failed to translate shader from TGSI to LLVM\n"); |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2903 | goto out; |
Michel Dänzer | 82cd9c0 | 2012-08-08 15:35:42 +0200 | [diff] [blame] | 2904 | } |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 2905 | |
| 2906 | radeon_llvm_finalize_module(&si_shader_ctx.radeon_bld); |
| 2907 | |
| 2908 | mod = bld_base->base.gallivm->module; |
Marek Olšák | 1abb1a9 | 2014-09-17 22:44:22 +0200 | [diff] [blame] | 2909 | r = si_compile_llvm(sscreen, shader, mod); |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2910 | if (r) { |
| 2911 | fprintf(stderr, "LLVM failed to compile shader\n"); |
| 2912 | goto out; |
| 2913 | } |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 2914 | |
Michel Dänzer | 4b64fa2 | 2012-08-15 18:22:46 +0200 | [diff] [blame] | 2915 | radeon_llvm_dispose(&si_shader_ctx.radeon_bld); |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2916 | |
| 2917 | if (si_shader_ctx.type == TGSI_PROCESSOR_GEOMETRY) { |
Marek Olšák | 8c37c16 | 2014-09-16 18:40:07 +0200 | [diff] [blame] | 2918 | shader->gs_copy_shader = CALLOC_STRUCT(si_shader); |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2919 | shader->gs_copy_shader->selector = shader->selector; |
Michel Dänzer | 7b19c39 | 2014-01-09 18:18:26 +0900 | [diff] [blame] | 2920 | shader->gs_copy_shader->key = shader->key; |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2921 | si_shader_ctx.shader = shader->gs_copy_shader; |
Marek Olšák | 68d36c0 | 2014-09-30 18:15:17 +0200 | [diff] [blame] | 2922 | if ((r = si_generate_gs_copy_shader(sscreen, &si_shader_ctx, |
| 2923 | shader, dump))) { |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2924 | free(shader->gs_copy_shader); |
| 2925 | shader->gs_copy_shader = NULL; |
| 2926 | goto out; |
| 2927 | } |
| 2928 | } |
| 2929 | |
Michel Dänzer | 404b29d | 2013-11-21 16:45:28 +0900 | [diff] [blame] | 2930 | out: |
Marek Olšák | ee2a818 | 2014-07-07 23:27:19 +0200 | [diff] [blame] | 2931 | for (int i = 0; i < SI_NUM_CONST_BUFFERS; i++) |
Marek Olšák | 2fd4200 | 2013-10-25 11:45:47 +0200 | [diff] [blame] | 2932 | FREE(si_shader_ctx.constants[i]); |
Marek Olšák | 6c5af1d | 2015-01-31 21:43:37 +0100 | [diff] [blame] | 2933 | if (poly_stipple) |
| 2934 | tgsi_free_tokens(tokens); |
Tom Stellard | 302f53d | 2012-10-25 13:50:10 -0400 | [diff] [blame] | 2935 | return r; |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 2936 | } |
| 2937 | |
Marek Olšák | 2774abd | 2014-09-16 18:45:33 +0200 | [diff] [blame] | 2938 | void si_shader_destroy(struct pipe_context *ctx, struct si_shader *shader) |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 2939 | { |
Marek Olšák | dc05a9e | 2014-09-18 23:48:04 +0200 | [diff] [blame] | 2940 | if (shader->gs_copy_shader) |
| 2941 | si_shader_destroy(ctx, shader->gs_copy_shader); |
| 2942 | |
Tom Stellard | 5dcd97f | 2015-01-27 16:35:21 +0000 | [diff] [blame] | 2943 | if (shader->scratch_bo) |
| 2944 | r600_resource_reference(&shader->scratch_bo, NULL); |
| 2945 | |
Marek Olšák | a81c3e0 | 2013-08-14 01:04:39 +0200 | [diff] [blame] | 2946 | r600_resource_reference(&shader->bo, NULL); |
Tom Stellard | 2397a72 | 2014-12-10 09:13:59 -0500 | [diff] [blame] | 2947 | |
| 2948 | FREE(shader->binary.code); |
| 2949 | FREE(shader->binary.relocs); |
Tom Stellard | a75c616 | 2012-01-06 17:38:37 -0500 | [diff] [blame] | 2950 | } |