brw_curbe.c revision 118a47623a11a374df371d52ed0294224e6a62dc
1/* 2 Copyright (C) Intel Corp. 2006. All Rights Reserved. 3 Intel funded Tungsten Graphics (http://www.tungstengraphics.com) 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 <keith@tungstengraphics.com> 30 */ 31 32 33 34#include "main/glheader.h" 35#include "main/context.h" 36#include "main/macros.h" 37#include "main/enums.h" 38#include "shader/prog_parameter.h" 39#include "shader/prog_print.h" 40#include "shader/prog_statevars.h" 41#include "intel_batchbuffer.h" 42#include "intel_regions.h" 43#include "brw_context.h" 44#include "brw_defines.h" 45#include "brw_state.h" 46#include "brw_util.h" 47 48 49/** 50 * Partition the CURBE between the various users of constant values: 51 * Note that vertex and fragment shaders can now fetch constants out 52 * of constant buffers. We no longer allocatea block of the GRF for 53 * constants. That greatly reduces the demand for space in the CURBE. 54 * Some of the comments within are dated... 55 */ 56static void calculate_curbe_offsets( struct brw_context *brw ) 57{ 58 GLcontext *ctx = &brw->intel.ctx; 59 /* CACHE_NEW_WM_PROG */ 60 const GLuint nr_fp_regs = (brw->wm.prog_data->nr_params + 15) / 16; 61 62 /* BRW_NEW_VERTEX_PROGRAM */ 63 const GLuint nr_vp_regs = (brw->vs.prog_data->nr_params + 15) / 16; 64 GLuint nr_clip_regs = 0; 65 GLuint total_regs; 66 67 /* _NEW_TRANSFORM */ 68 if (ctx->Transform.ClipPlanesEnabled) { 69 GLuint nr_planes = 6 + brw_count_bits(ctx->Transform.ClipPlanesEnabled); 70 nr_clip_regs = (nr_planes * 4 + 15) / 16; 71 } 72 73 74 total_regs = nr_fp_regs + nr_vp_regs + nr_clip_regs; 75 76 /* This can happen - what to do? Probably rather than falling 77 * back, the best thing to do is emit programs which code the 78 * constants as immediate values. Could do this either as a static 79 * cap on WM and VS, or adaptively. 80 * 81 * Unfortunately, this is currently dependent on the results of the 82 * program generation process (in the case of wm), so this would 83 * introduce the need to re-generate programs in the event of a 84 * curbe allocation failure. 85 */ 86 /* Max size is 32 - just large enough to 87 * hold the 128 parameters allowed by 88 * the fragment and vertex program 89 * api's. It's not clear what happens 90 * when both VP and FP want to use 128 91 * parameters, though. 92 */ 93 assert(total_regs <= 32); 94 95 /* Lazy resize: 96 */ 97 if (nr_fp_regs > brw->curbe.wm_size || 98 nr_vp_regs > brw->curbe.vs_size || 99 nr_clip_regs != brw->curbe.clip_size || 100 (total_regs < brw->curbe.total_size / 4 && 101 brw->curbe.total_size > 16)) { 102 103 GLuint reg = 0; 104 105 /* Calculate a new layout: 106 */ 107 reg = 0; 108 brw->curbe.wm_start = reg; 109 brw->curbe.wm_size = nr_fp_regs; reg += nr_fp_regs; 110 brw->curbe.clip_start = reg; 111 brw->curbe.clip_size = nr_clip_regs; reg += nr_clip_regs; 112 brw->curbe.vs_start = reg; 113 brw->curbe.vs_size = nr_vp_regs; reg += nr_vp_regs; 114 brw->curbe.total_size = reg; 115 116 if (0) 117 printf("curbe wm %d+%d clip %d+%d vs %d+%d\n", 118 brw->curbe.wm_start, 119 brw->curbe.wm_size, 120 brw->curbe.clip_start, 121 brw->curbe.clip_size, 122 brw->curbe.vs_start, 123 brw->curbe.vs_size ); 124 125 brw->state.dirty.brw |= BRW_NEW_CURBE_OFFSETS; 126 } 127} 128 129 130const struct brw_tracked_state brw_curbe_offsets = { 131 .dirty = { 132 .mesa = _NEW_TRANSFORM, 133 .brw = BRW_NEW_VERTEX_PROGRAM | BRW_NEW_CONTEXT, 134 .cache = CACHE_NEW_WM_PROG 135 }, 136 .prepare = calculate_curbe_offsets 137}; 138 139 140 141 142/* Define the number of curbes within CS's urb allocation. Multiple 143 * urb entries -> multiple curbes. These will be used by 144 * fixed-function hardware in a double-buffering scheme to avoid a 145 * pipeline stall each time the contents of the curbe is changed. 146 */ 147void brw_upload_cs_urb_state(struct brw_context *brw) 148{ 149 struct brw_cs_urb_state cs_urb; 150 memset(&cs_urb, 0, sizeof(cs_urb)); 151 152 /* It appears that this is the state packet for the CS unit, ie. the 153 * urb entries detailed here are housed in the CS range from the 154 * URB_FENCE command. 155 */ 156 cs_urb.header.opcode = CMD_CS_URB_STATE; 157 cs_urb.header.length = sizeof(cs_urb)/4 - 2; 158 159 /* BRW_NEW_URB_FENCE */ 160 cs_urb.bits0.nr_urb_entries = brw->urb.nr_cs_entries; 161 cs_urb.bits0.urb_entry_size = brw->urb.csize - 1; 162 163 assert(brw->urb.nr_cs_entries); 164 BRW_CACHED_BATCH_STRUCT(brw, &cs_urb); 165} 166 167static GLfloat fixed_plane[6][4] = { 168 { 0, 0, -1, 1 }, 169 { 0, 0, 1, 1 }, 170 { 0, -1, 0, 1 }, 171 { 0, 1, 0, 1 }, 172 {-1, 0, 0, 1 }, 173 { 1, 0, 0, 1 } 174}; 175 176/* Upload a new set of constants. Too much variability to go into the 177 * cache mechanism, but maybe would benefit from a comparison against 178 * the current uploaded set of constants. 179 */ 180static void prepare_constant_buffer(struct brw_context *brw) 181{ 182 GLcontext *ctx = &brw->intel.ctx; 183 const struct brw_vertex_program *vp = 184 brw_vertex_program_const(brw->vertex_program); 185 const struct brw_fragment_program *fp = 186 brw_fragment_program_const(brw->fragment_program); 187 const GLuint sz = brw->curbe.total_size; 188 const GLuint bufsz = sz * 16 * sizeof(GLfloat); 189 GLfloat *buf; 190 GLuint i; 191 192 if (sz == 0) { 193 brw->curbe.last_bufsz = 0; 194 return; 195 } 196 197 buf = brw->curbe.next_buf; 198 199 /* fragment shader constants */ 200 if (brw->curbe.wm_size) { 201 GLuint offset = brw->curbe.wm_start * 16; 202 203 /* copy float constants */ 204 for (i = 0; i < brw->wm.prog_data->nr_params; i++) 205 buf[offset + i] = *brw->wm.prog_data->param[i]; 206 } 207 208 209 /* The clipplanes are actually delivered to both CLIP and VS units. 210 * VS uses them to calculate the outcode bitmasks. 211 */ 212 if (brw->curbe.clip_size) { 213 GLuint offset = brw->curbe.clip_start * 16; 214 GLuint j; 215 216 /* If any planes are going this way, send them all this way: 217 */ 218 for (i = 0; i < 6; i++) { 219 buf[offset + i * 4 + 0] = fixed_plane[i][0]; 220 buf[offset + i * 4 + 1] = fixed_plane[i][1]; 221 buf[offset + i * 4 + 2] = fixed_plane[i][2]; 222 buf[offset + i * 4 + 3] = fixed_plane[i][3]; 223 } 224 225 /* Clip planes: _NEW_TRANSFORM plus _NEW_PROJECTION to get to 226 * clip-space: 227 */ 228 assert(MAX_CLIP_PLANES == 6); 229 for (j = 0; j < MAX_CLIP_PLANES; j++) { 230 if (ctx->Transform.ClipPlanesEnabled & (1<<j)) { 231 buf[offset + i * 4 + 0] = ctx->Transform._ClipUserPlane[j][0]; 232 buf[offset + i * 4 + 1] = ctx->Transform._ClipUserPlane[j][1]; 233 buf[offset + i * 4 + 2] = ctx->Transform._ClipUserPlane[j][2]; 234 buf[offset + i * 4 + 3] = ctx->Transform._ClipUserPlane[j][3]; 235 i++; 236 } 237 } 238 } 239 240 /* vertex shader constants */ 241 if (brw->curbe.vs_size) { 242 GLuint offset = brw->curbe.vs_start * 16; 243 GLuint nr = brw->vs.prog_data->nr_params / 4; 244 245 if (vp->use_const_buffer) { 246 /* Load the subset of push constants that will get used when 247 * we also have a pull constant buffer. 248 */ 249 for (i = 0; i < vp->program.Base.Parameters->NumParameters; i++) { 250 if (brw->vs.constant_map[i] != -1) { 251 assert(brw->vs.constant_map[i] <= nr); 252 memcpy(buf + offset + brw->vs.constant_map[i] * 4, 253 vp->program.Base.Parameters->ParameterValues[i], 254 4 * sizeof(float)); 255 } 256 } 257 } else { 258 for (i = 0; i < nr; i++) { 259 memcpy(buf + offset + i * 4, 260 vp->program.Base.Parameters->ParameterValues[i], 261 4 * sizeof(float)); 262 } 263 } 264 } 265 266 if (0) { 267 for (i = 0; i < sz*16; i+=4) 268 printf("curbe %d.%d: %f %f %f %f\n", i/8, i&4, 269 buf[i+0], buf[i+1], buf[i+2], buf[i+3]); 270 271 printf("last_buf %p buf %p sz %d/%d cmp %d\n", 272 brw->curbe.last_buf, buf, 273 bufsz, brw->curbe.last_bufsz, 274 brw->curbe.last_buf ? memcmp(buf, brw->curbe.last_buf, bufsz) : -1); 275 } 276 277 if (brw->curbe.curbe_bo != NULL && 278 bufsz == brw->curbe.last_bufsz && 279 memcmp(buf, brw->curbe.last_buf, bufsz) == 0) { 280 /* constants have not changed */ 281 } else { 282 /* Update the record of what our last set of constants was. We 283 * don't just flip the pointers because we don't fill in the 284 * data in the padding between the entries. 285 */ 286 memcpy(brw->curbe.last_buf, buf, bufsz); 287 brw->curbe.last_bufsz = bufsz; 288 289 if (brw->curbe.curbe_bo != NULL && 290 brw->curbe.curbe_next_offset + bufsz > brw->curbe.curbe_bo->size) 291 { 292 drm_intel_gem_bo_unmap_gtt(brw->curbe.curbe_bo); 293 drm_intel_bo_unreference(brw->curbe.curbe_bo); 294 brw->curbe.curbe_bo = NULL; 295 } 296 297 if (brw->curbe.curbe_bo == NULL) { 298 /* Allocate a single page for CURBE entries for this batchbuffer. 299 * They're generally around 64b. 300 */ 301 brw->curbe.curbe_bo = drm_intel_bo_alloc(brw->intel.bufmgr, "CURBE", 302 4096, 1 << 6); 303 brw->curbe.curbe_next_offset = 0; 304 drm_intel_gem_bo_map_gtt(brw->curbe.curbe_bo); 305 assert(bufsz < 4096); 306 } 307 308 brw->curbe.curbe_offset = brw->curbe.curbe_next_offset; 309 brw->curbe.curbe_next_offset += bufsz; 310 brw->curbe.curbe_next_offset = ALIGN(brw->curbe.curbe_next_offset, 64); 311 312 /* Copy data to the buffer: 313 */ 314 memcpy(brw->curbe.curbe_bo->virtual + brw->curbe.curbe_offset, 315 buf, 316 bufsz); 317 } 318 319 brw_add_validated_bo(brw, brw->curbe.curbe_bo); 320 321 /* Because this provokes an action (ie copy the constants into the 322 * URB), it shouldn't be shortcircuited if identical to the 323 * previous time - because eg. the urb destination may have 324 * changed, or the urb contents different to last time. 325 * 326 * Note that the data referred to is actually copied internally, 327 * not just used in place according to passed pointer. 328 * 329 * It appears that the CS unit takes care of using each available 330 * URB entry (Const URB Entry == CURBE) in turn, and issuing 331 * flushes as necessary when doublebuffering of CURBEs isn't 332 * possible. 333 */ 334} 335 336static void emit_constant_buffer(struct brw_context *brw) 337{ 338 struct intel_context *intel = &brw->intel; 339 GLuint sz = brw->curbe.total_size; 340 341 BEGIN_BATCH(2); 342 if (sz == 0) { 343 OUT_BATCH((CMD_CONST_BUFFER << 16) | (2 - 2)); 344 OUT_BATCH(0); 345 } else { 346 OUT_BATCH((CMD_CONST_BUFFER << 16) | (1 << 8) | (2 - 2)); 347 OUT_RELOC(brw->curbe.curbe_bo, 348 I915_GEM_DOMAIN_INSTRUCTION, 0, 349 (sz - 1) + brw->curbe.curbe_offset); 350 } 351 ADVANCE_BATCH(); 352} 353 354/* This tracked state is unique in that the state it monitors varies 355 * dynamically depending on the parameters tracked by the fragment and 356 * vertex programs. This is the template used as a starting point, 357 * each context will maintain a copy of this internally and update as 358 * required. 359 */ 360const struct brw_tracked_state brw_constant_buffer = { 361 .dirty = { 362 .mesa = _NEW_PROGRAM_CONSTANTS, 363 .brw = (BRW_NEW_FRAGMENT_PROGRAM | 364 BRW_NEW_VERTEX_PROGRAM | 365 BRW_NEW_URB_FENCE | /* Implicit - hardware requires this, not used above */ 366 BRW_NEW_PSP | /* Implicit - hardware requires this, not used above */ 367 BRW_NEW_CURBE_OFFSETS | 368 BRW_NEW_BATCH), 369 .cache = (CACHE_NEW_WM_PROG) 370 }, 371 .prepare = prepare_constant_buffer, 372 .emit = emit_constant_buffer, 373}; 374 375