1 /* 2 Copyright (C) Intel Corp. 2006. All Rights Reserved. 3 Intel funded Tungsten Graphics to 4 develop this 3D driver. 5 6 Permission is hereby granted, free of charge, to any person obtaining 7 a copy of this software and associated documentation files (the 8 "Software"), to deal in the Software without restriction, including 9 without limitation the rights to use, copy, modify, merge, publish, 10 distribute, sublicense, and/or sell copies of the Software, and to 11 permit persons to whom the Software is furnished to do so, subject to 12 the following conditions: 13 14 The above copyright notice and this permission notice (including the 15 next paragraph) shall be included in all copies or substantial 16 portions of the Software. 17 18 THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 19 EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 20 MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. 21 IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE 22 LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION 23 OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION 24 WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. 25 26 **********************************************************************/ 27 /* 28 * Authors: 29 * Keith Whitwell <keithw (at) vmware.com> 30 */ 31 32 /** @file brw_curbe.c 33 * 34 * Push constant handling for gen4/5. 35 * 36 * Push constants are constant values (such as GLSL uniforms) that are 37 * pre-loaded into a shader stage's register space at thread spawn time. On 38 * gen4 and gen5, we create a blob in memory containing all the push constants 39 * for all the stages in order. At CMD_CONST_BUFFER time that blob is loaded 40 * into URB space as a constant URB entry (CURBE) so that it can be accessed 41 * quickly at thread setup time. Each individual fixed function unit's state 42 * (brw_vs_state.c for example) tells the hardware which subset of the CURBE 43 * it wants in its register space, and we calculate those areas here under the 44 * BRW_NEW_PUSH_CONSTANT_ALLOCATION state flag. The brw_urb.c allocation will control 45 * how many CURBEs can be loaded into the hardware at once before a pipeline 46 * stall occurs at CMD_CONST_BUFFER time. 47 * 48 * On gen6+, constant handling becomes a much simpler set of per-unit state. 49 * See gen6_upload_vec4_push_constants() in gen6_vs_state.c for that code. 50 */ 51 52 53 #include "compiler/nir/nir.h" 54 #include "main/context.h" 55 #include "main/macros.h" 56 #include "main/enums.h" 57 #include "program/prog_parameter.h" 58 #include "program/prog_print.h" 59 #include "program/prog_statevars.h" 60 #include "util/bitscan.h" 61 #include "intel_batchbuffer.h" 62 #include "intel_buffer_objects.h" 63 #include "brw_context.h" 64 #include "brw_defines.h" 65 #include "brw_state.h" 66 #include "brw_util.h" 67 68 69 /** 70 * Partition the CURBE between the various users of constant values. 71 * 72 * If the users all fit within the previous allocatation, we avoid changing 73 * the layout because that means reuploading all unit state and uploading new 74 * constant buffers. 75 */ 76 static void calculate_curbe_offsets( struct brw_context *brw ) 77 { 78 struct gl_context *ctx = &brw->ctx; 79 /* BRW_NEW_FS_PROG_DATA */ 80 const GLuint nr_fp_regs = (brw->wm.base.prog_data->nr_params + 15) / 16; 81 82 /* BRW_NEW_VS_PROG_DATA */ 83 const GLuint nr_vp_regs = (brw->vs.base.prog_data->nr_params + 15) / 16; 84 GLuint nr_clip_regs = 0; 85 GLuint total_regs; 86 87 /* _NEW_TRANSFORM */ 88 if (ctx->Transform.ClipPlanesEnabled) { 89 GLuint nr_planes = 6 + _mesa_bitcount(ctx->Transform.ClipPlanesEnabled); 90 nr_clip_regs = (nr_planes * 4 + 15) / 16; 91 } 92 93 94 total_regs = nr_fp_regs + nr_vp_regs + nr_clip_regs; 95 96 /* The CURBE allocation size is limited to 32 512-bit units (128 EU 97 * registers, or 1024 floats). See CS_URB_STATE in the gen4 or gen5 98 * (volume 1, part 1) PRMs. 99 * 100 * Note that in brw_fs.cpp we're only loading up to 16 EU registers of 101 * values as push constants before spilling to pull constants, and in 102 * brw_vec4.cpp we're loading up to 32 registers of push constants. An EU 103 * register is 1/2 of one of these URB entry units, so that leaves us 16 EU 104 * regs for clip. 105 */ 106 assert(total_regs <= 32); 107 108 /* Lazy resize: 109 */ 110 if (nr_fp_regs > brw->curbe.wm_size || 111 nr_vp_regs > brw->curbe.vs_size || 112 nr_clip_regs != brw->curbe.clip_size || 113 (total_regs < brw->curbe.total_size / 4 && 114 brw->curbe.total_size > 16)) { 115 116 GLuint reg = 0; 117 118 /* Calculate a new layout: 119 */ 120 reg = 0; 121 brw->curbe.wm_start = reg; 122 brw->curbe.wm_size = nr_fp_regs; reg += nr_fp_regs; 123 brw->curbe.clip_start = reg; 124 brw->curbe.clip_size = nr_clip_regs; reg += nr_clip_regs; 125 brw->curbe.vs_start = reg; 126 brw->curbe.vs_size = nr_vp_regs; reg += nr_vp_regs; 127 brw->curbe.total_size = reg; 128 129 if (0) 130 fprintf(stderr, "curbe wm %d+%d clip %d+%d vs %d+%d\n", 131 brw->curbe.wm_start, 132 brw->curbe.wm_size, 133 brw->curbe.clip_start, 134 brw->curbe.clip_size, 135 brw->curbe.vs_start, 136 brw->curbe.vs_size ); 137 138 brw->ctx.NewDriverState |= BRW_NEW_PUSH_CONSTANT_ALLOCATION; 139 } 140 } 141 142 143 const struct brw_tracked_state brw_curbe_offsets = { 144 .dirty = { 145 .mesa = _NEW_TRANSFORM, 146 .brw = BRW_NEW_CONTEXT | 147 BRW_NEW_BLORP | 148 BRW_NEW_FS_PROG_DATA | 149 BRW_NEW_VS_PROG_DATA, 150 }, 151 .emit = calculate_curbe_offsets 152 }; 153 154 155 156 157 /** Uploads the CS_URB_STATE packet. 158 * 159 * Just like brw_vs_state.c and brw_wm_state.c define a URB entry size and 160 * number of entries for their stages, constant buffers do so using this state 161 * packet. Having multiple CURBEs in the URB at the same time allows the 162 * hardware to avoid a pipeline stall between primitives using different 163 * constant buffer contents. 164 */ 165 void brw_upload_cs_urb_state(struct brw_context *brw) 166 { 167 BEGIN_BATCH(2); 168 OUT_BATCH(CMD_CS_URB_STATE << 16 | (2-2)); 169 170 /* BRW_NEW_URB_FENCE */ 171 if (brw->urb.csize == 0) { 172 OUT_BATCH(0); 173 } else { 174 /* BRW_NEW_URB_FENCE */ 175 assert(brw->urb.nr_cs_entries); 176 OUT_BATCH((brw->urb.csize - 1) << 4 | brw->urb.nr_cs_entries); 177 } 178 ADVANCE_BATCH(); 179 } 180 181 static const GLfloat fixed_plane[6][4] = { 182 { 0, 0, -1, 1 }, 183 { 0, 0, 1, 1 }, 184 { 0, -1, 0, 1 }, 185 { 0, 1, 0, 1 }, 186 {-1, 0, 0, 1 }, 187 { 1, 0, 0, 1 } 188 }; 189 190 /** 191 * Gathers together all the uniform values into a block of memory to be 192 * uploaded into the CURBE, then emits the state packet telling the hardware 193 * the new location. 194 */ 195 static void 196 brw_upload_constant_buffer(struct brw_context *brw) 197 { 198 const struct gen_device_info *devinfo = &brw->screen->devinfo; 199 struct gl_context *ctx = &brw->ctx; 200 /* BRW_NEW_PUSH_CONSTANT_ALLOCATION */ 201 const GLuint sz = brw->curbe.total_size; 202 const GLuint bufsz = sz * 16 * sizeof(GLfloat); 203 gl_constant_value *buf; 204 GLuint i; 205 gl_clip_plane *clip_planes; 206 207 /* BRW_NEW_FRAGMENT_PROGRAM */ 208 struct gl_program *fp = brw->programs[MESA_SHADER_FRAGMENT]; 209 210 /* BRW_NEW_VERTEX_PROGRAM */ 211 struct gl_program *vp = brw->programs[MESA_SHADER_VERTEX]; 212 213 if (sz == 0) { 214 goto emit; 215 } 216 217 buf = intel_upload_space(brw, bufsz, 64, 218 &brw->curbe.curbe_bo, &brw->curbe.curbe_offset); 219 220 STATIC_ASSERT(sizeof(gl_constant_value) == sizeof(float)); 221 222 /* fragment shader constants */ 223 if (brw->curbe.wm_size) { 224 _mesa_load_state_parameters(ctx, fp->Parameters); 225 226 /* BRW_NEW_PUSH_CONSTANT_ALLOCATION */ 227 GLuint offset = brw->curbe.wm_start * 16; 228 229 /* BRW_NEW_FS_PROG_DATA | _NEW_PROGRAM_CONSTANTS: copy uniform values */ 230 brw_populate_constant_data(brw, fp, &brw->wm.base, &buf[offset], 231 brw->wm.base.prog_data->param, 232 brw->wm.base.prog_data->nr_params); 233 } 234 235 /* clipper constants */ 236 if (brw->curbe.clip_size) { 237 GLuint offset = brw->curbe.clip_start * 16; 238 GLbitfield mask; 239 240 /* If any planes are going this way, send them all this way: 241 */ 242 for (i = 0; i < 6; i++) { 243 buf[offset + i * 4 + 0].f = fixed_plane[i][0]; 244 buf[offset + i * 4 + 1].f = fixed_plane[i][1]; 245 buf[offset + i * 4 + 2].f = fixed_plane[i][2]; 246 buf[offset + i * 4 + 3].f = fixed_plane[i][3]; 247 } 248 249 /* Clip planes: _NEW_TRANSFORM plus _NEW_PROJECTION to get to 250 * clip-space: 251 */ 252 clip_planes = brw_select_clip_planes(ctx); 253 mask = ctx->Transform.ClipPlanesEnabled; 254 while (mask) { 255 const int j = u_bit_scan(&mask); 256 buf[offset + i * 4 + 0].f = clip_planes[j][0]; 257 buf[offset + i * 4 + 1].f = clip_planes[j][1]; 258 buf[offset + i * 4 + 2].f = clip_planes[j][2]; 259 buf[offset + i * 4 + 3].f = clip_planes[j][3]; 260 i++; 261 } 262 } 263 264 /* vertex shader constants */ 265 if (brw->curbe.vs_size) { 266 _mesa_load_state_parameters(ctx, vp->Parameters); 267 268 GLuint offset = brw->curbe.vs_start * 16; 269 270 /* BRW_NEW_VS_PROG_DATA | _NEW_PROGRAM_CONSTANTS: copy uniform values */ 271 brw_populate_constant_data(brw, vp, &brw->vs.base, &buf[offset], 272 brw->vs.base.prog_data->param, 273 brw->vs.base.prog_data->nr_params); 274 } 275 276 if (0) { 277 for (i = 0; i < sz*16; i+=4) 278 fprintf(stderr, "curbe %d.%d: %f %f %f %f\n", i/8, i&4, 279 buf[i+0].f, buf[i+1].f, buf[i+2].f, buf[i+3].f); 280 } 281 282 /* Because this provokes an action (ie copy the constants into the 283 * URB), it shouldn't be shortcircuited if identical to the 284 * previous time - because eg. the urb destination may have 285 * changed, or the urb contents different to last time. 286 * 287 * Note that the data referred to is actually copied internally, 288 * not just used in place according to passed pointer. 289 * 290 * It appears that the CS unit takes care of using each available 291 * URB entry (Const URB Entry == CURBE) in turn, and issuing 292 * flushes as necessary when doublebuffering of CURBEs isn't 293 * possible. 294 */ 295 296 emit: 297 /* BRW_NEW_URB_FENCE: From the gen4 PRM, volume 1, section 3.9.8 298 * (CONSTANT_BUFFER (CURBE Load)): 299 * 300 * "Modifying the CS URB allocation via URB_FENCE invalidates any 301 * previous CURBE entries. Therefore software must subsequently 302 * [re]issue a CONSTANT_BUFFER command before CURBE data can be used 303 * in the pipeline." 304 */ 305 BEGIN_BATCH(2); 306 if (brw->curbe.total_size == 0) { 307 OUT_BATCH((CMD_CONST_BUFFER << 16) | (2 - 2)); 308 OUT_BATCH(0); 309 } else { 310 OUT_BATCH((CMD_CONST_BUFFER << 16) | (1 << 8) | (2 - 2)); 311 OUT_RELOC(brw->curbe.curbe_bo, 0, 312 (brw->curbe.total_size - 1) + brw->curbe.curbe_offset); 313 } 314 ADVANCE_BATCH(); 315 316 /* Work around a Broadwater/Crestline depth interpolator bug. The 317 * following sequence will cause GPU hangs: 318 * 319 * 1. Change state so that all depth related fields in CC_STATE are 320 * disabled, and in WM_STATE, only "PS Use Source Depth" is enabled. 321 * 2. Emit a CONSTANT_BUFFER packet. 322 * 3. Draw via 3DPRIMITIVE. 323 * 324 * The recommended workaround is to emit a non-pipelined state change after 325 * emitting CONSTANT_BUFFER, in order to drain the windowizer pipeline. 326 * 327 * We arbitrarily choose 3DSTATE_GLOBAL_DEPTH_CLAMP_OFFSET (as it's small), 328 * and always emit it when "PS Use Source Depth" is set. We could be more 329 * precise, but the additional complexity is probably not worth it. 330 * 331 * BRW_NEW_FRAGMENT_PROGRAM 332 */ 333 if (devinfo->gen == 4 && !devinfo->is_g4x && 334 (fp->info.inputs_read & (1 << VARYING_SLOT_POS))) { 335 BEGIN_BATCH(2); 336 OUT_BATCH(_3DSTATE_GLOBAL_DEPTH_OFFSET_CLAMP << 16 | (2 - 2)); 337 OUT_BATCH(0); 338 ADVANCE_BATCH(); 339 } 340 } 341 342 const struct brw_tracked_state brw_constant_buffer = { 343 .dirty = { 344 .mesa = _NEW_PROGRAM_CONSTANTS, 345 .brw = BRW_NEW_BATCH | 346 BRW_NEW_BLORP | 347 BRW_NEW_PUSH_CONSTANT_ALLOCATION | 348 BRW_NEW_FRAGMENT_PROGRAM | 349 BRW_NEW_FS_PROG_DATA | 350 BRW_NEW_PSP | /* Implicit - hardware requires this, not used above */ 351 BRW_NEW_URB_FENCE | 352 BRW_NEW_VS_PROG_DATA, 353 }, 354 .emit = brw_upload_constant_buffer, 355 }; 356