2 * Copyright © 2012 Intel Corporation
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 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
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
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 NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
25 #include "intel_batchbuffer.h"
26 #include "intel_fbo.h"
28 #include "brw_blorp.h"
29 #include "brw_compiler.h"
31 #include "brw_state.h"
33 #define FILE_DEBUG_FLAG DEBUG_BLORP
36 brw_blorp_surface_info_init(struct brw_context *brw,
37 struct brw_blorp_surface_info *info,
38 struct intel_mipmap_tree *mt,
39 unsigned int level, unsigned int layer,
40 mesa_format format, bool is_render_target)
42 /* Layer is a physical layer, so if this is a 2D multisample array texture
43 * using INTEL_MSAA_LAYOUT_UMS or INTEL_MSAA_LAYOUT_CMS, then it had better
44 * be a multiple of num_samples.
46 unsigned layer_multiplier = 1;
47 if (mt->msaa_layout == INTEL_MSAA_LAYOUT_UMS ||
48 mt->msaa_layout == INTEL_MSAA_LAYOUT_CMS) {
49 assert(mt->num_samples <= 1 || layer % mt->num_samples == 0);
50 layer_multiplier = MAX2(mt->num_samples, 1);
53 intel_miptree_check_level_layer(mt, level, layer);
57 intel_miptree_get_isl_surf(brw, mt, &info->surf);
60 intel_miptree_get_aux_isl_surf(brw, mt, &info->aux_surf,
63 info->aux_usage = ISL_AUX_USAGE_NONE;
66 info->view = (struct isl_view) {
67 .usage = is_render_target ? ISL_SURF_USAGE_RENDER_TARGET_BIT :
68 ISL_SURF_USAGE_TEXTURE_BIT,
69 .format = ISL_FORMAT_UNSUPPORTED, /* Set later */
72 .base_array_layer = layer / layer_multiplier,
75 ISL_CHANNEL_SELECT_RED,
76 ISL_CHANNEL_SELECT_GREEN,
77 ISL_CHANNEL_SELECT_BLUE,
78 ISL_CHANNEL_SELECT_ALPHA,
84 info->width = minify(mt->physical_width0, level - mt->first_level);
85 info->height = minify(mt->physical_height0, level - mt->first_level);
87 if (format == MESA_FORMAT_NONE)
91 case MESA_FORMAT_S_UINT8:
92 assert(info->surf.tiling == ISL_TILING_W);
93 /* Prior to Broadwell, we can't render to R8_UINT */
94 info->view.format = brw->gen >= 8 ? BRW_SURFACEFORMAT_R8_UINT :
95 BRW_SURFACEFORMAT_R8_UNORM;
97 case MESA_FORMAT_Z24_UNORM_X8_UINT:
98 /* It would make sense to use BRW_SURFACEFORMAT_R24_UNORM_X8_TYPELESS
99 * here, but unfortunately it isn't supported as a render target, which
100 * would prevent us from blitting to 24-bit depth.
102 * The miptree consists of 32 bits per pixel, arranged as 24-bit depth
103 * values interleaved with 8 "don't care" bits. Since depth values don't
104 * require any blending, it doesn't matter how we interpret the bit
105 * pattern as long as we copy the right amount of data, so just map it
108 info->view.format = BRW_SURFACEFORMAT_B8G8R8A8_UNORM;
110 case MESA_FORMAT_Z_FLOAT32:
111 info->view.format = BRW_SURFACEFORMAT_R32_FLOAT;
113 case MESA_FORMAT_Z_UNORM16:
114 info->view.format = BRW_SURFACEFORMAT_R16_UNORM;
117 if (is_render_target) {
118 assert(brw->format_supported_as_render_target[format]);
119 info->view.format = brw->render_target_format[format];
121 info->view.format = brw_format_for_mesa_format(format);
127 uint32_t x_offset, y_offset;
128 intel_miptree_get_image_offset(mt, level, layer, &x_offset, &y_offset);
130 uint8_t bs = isl_format_get_layout(info->view.format)->bpb / 8;
131 isl_tiling_get_intratile_offset_el(&brw->isl_dev, info->surf.tiling, bs,
132 info->surf.row_pitch, x_offset, y_offset,
134 &info->tile_x_sa, &info->tile_y_sa);
139 brw_blorp_params_init(struct brw_blorp_params *params)
141 memset(params, 0, sizeof(*params));
142 params->hiz_op = GEN6_HIZ_OP_NONE;
143 params->fast_clear_op = 0;
144 params->num_draw_buffers = 1;
145 params->num_layers = 1;
149 brw_blorp_init_wm_prog_key(struct brw_wm_prog_key *wm_key)
151 memset(wm_key, 0, sizeof(*wm_key));
152 wm_key->nr_color_regions = 1;
153 for (int i = 0; i < MAX_SAMPLERS; i++)
154 wm_key->tex.swizzles[i] = SWIZZLE_XYZW;
158 nir_uniform_type_size(const struct glsl_type *type)
160 /* Only very basic types are allowed */
161 assert(glsl_type_is_vector_or_scalar(type));
162 assert(glsl_get_bit_size(type) == 32);
164 return glsl_get_vector_elements(type) * 4;
168 brw_blorp_compile_nir_shader(struct brw_context *brw, struct nir_shader *nir,
169 const struct brw_wm_prog_key *wm_key,
171 struct brw_blorp_prog_data *prog_data,
172 unsigned *program_size)
174 const struct brw_compiler *compiler = brw->intelScreen->compiler;
176 void *mem_ctx = ralloc_context(NULL);
178 /* Calling brw_preprocess_nir and friends is destructive and, if cloning is
179 * enabled, may end up completely replacing the nir_shader. Therefore, we
180 * own it and might as well put it in our context for easy cleanup.
182 ralloc_steal(mem_ctx, nir);
184 compiler->glsl_compiler_options[MESA_SHADER_FRAGMENT].NirOptions;
186 struct brw_wm_prog_data wm_prog_data;
187 memset(&wm_prog_data, 0, sizeof(wm_prog_data));
189 wm_prog_data.base.nr_params = 0;
190 wm_prog_data.base.param = NULL;
192 /* BLORP always just uses the first two binding table entries */
193 wm_prog_data.binding_table.render_target_start = 0;
194 wm_prog_data.base.binding_table.texture_start = 1;
196 nir = brw_preprocess_nir(compiler, nir);
197 nir_remove_dead_variables(nir, nir_var_shader_in);
198 nir_shader_gather_info(nir, nir_shader_get_entrypoint(nir)->impl);
200 /* Uniforms are required to be lowered before going into compile_fs. For
201 * BLORP, we'll assume that whoever builds the shader sets the location
202 * they want so we just need to lower them and figure out how many we have
205 nir->num_uniforms = 0;
206 nir_foreach_variable(var, &nir->uniforms) {
207 var->data.driver_location = var->data.location;
208 unsigned end = var->data.location + nir_uniform_type_size(var->type);
209 nir->num_uniforms = MAX2(nir->num_uniforms, end);
211 nir_lower_io(nir, nir_var_uniform, nir_uniform_type_size);
213 const unsigned *program =
214 brw_compile_fs(compiler, brw, mem_ctx, wm_key, &wm_prog_data, nir,
215 NULL, -1, -1, false, use_repclear, program_size, NULL);
217 /* Copy the relavent bits of wm_prog_data over into the blorp prog data */
218 prog_data->dispatch_8 = wm_prog_data.dispatch_8;
219 prog_data->dispatch_16 = wm_prog_data.dispatch_16;
220 prog_data->first_curbe_grf_0 = wm_prog_data.base.dispatch_grf_start_reg;
221 prog_data->first_curbe_grf_2 = wm_prog_data.dispatch_grf_start_reg_2;
222 prog_data->ksp_offset_2 = wm_prog_data.prog_offset_2;
223 prog_data->persample_msaa_dispatch = wm_prog_data.persample_dispatch;
224 prog_data->flat_inputs = wm_prog_data.flat_inputs;
225 prog_data->num_varying_inputs = wm_prog_data.num_varying_inputs;
226 prog_data->inputs_read = nir->info.inputs_read;
228 assert(wm_prog_data.base.nr_params == 0);
233 struct surface_state_info {
235 unsigned ss_align; /* Required alignment of RENDER_SURFACE_STATE in bytes */
237 unsigned aux_reloc_dw;
242 static const struct surface_state_info surface_state_infos[] = {
244 [7] = {8, 32, 1, 6, GEN7_MOCS_L3, GEN7_MOCS_L3},
245 [8] = {13, 64, 8, 10, BDW_MOCS_WB, BDW_MOCS_PTE},
246 [9] = {16, 64, 8, 10, SKL_MOCS_WB, SKL_MOCS_PTE},
250 brw_blorp_emit_surface_state(struct brw_context *brw,
251 const struct brw_blorp_surface_info *surface,
252 uint32_t read_domains, uint32_t write_domain,
253 bool is_render_target)
255 const struct surface_state_info ss_info = surface_state_infos[brw->gen];
257 struct isl_surf surf = surface->surf;
259 /* Stomp surface dimensions and tiling (if needed) with info from blorp */
260 surf.dim = ISL_SURF_DIM_2D;
261 surf.dim_layout = ISL_DIM_LAYOUT_GEN4_2D;
262 surf.logical_level0_px.width = surface->width;
263 surf.logical_level0_px.height = surface->height;
264 surf.logical_level0_px.depth = 1;
265 surf.logical_level0_px.array_len = 1;
268 /* Alignment doesn't matter since we have 1 miplevel and 1 array slice so
269 * just pick something that works for everybody.
271 surf.image_alignment_el = isl_extent3d(4, 4, 1);
273 if (brw->gen == 6 && surf.samples > 1) {
274 /* Since gen6 uses INTEL_MSAA_LAYOUT_IMS, width and height are measured
275 * in samples. But SURFACE_STATE wants them in pixels, so we need to
276 * divide them each by 2.
278 surf.logical_level0_px.width /= 2;
279 surf.logical_level0_px.height /= 2;
282 if (brw->gen == 6 && surf.image_alignment_el.height > 4) {
283 /* This can happen on stencil buffers on Sandy Bridge due to the
284 * single-LOD work-around. It's fairly harmless as long as we don't
285 * pass a bogus value into isl_surf_fill_state().
287 surf.image_alignment_el = isl_extent3d(4, 2, 1);
290 union isl_color_value clear_color = { .u32 = { 0, 0, 0, 0 } };
292 const struct isl_surf *aux_surf = NULL;
293 uint64_t aux_offset = 0;
294 if (surface->mt->mcs_mt) {
295 aux_surf = &surface->aux_surf;
296 assert(surface->mt->mcs_mt->offset == 0);
297 aux_offset = surface->mt->mcs_mt->bo->offset64;
299 /* We only really need a clear color if we also have an auxiliary
300 * surface. Without one, it does nothing.
302 clear_color = intel_miptree_get_isl_clear_color(brw, surface->mt);
305 struct isl_view view = {
306 .format = surface->view.format,
309 .base_array_layer = 0,
312 ISL_CHANNEL_SELECT_RED,
313 ISL_CHANNEL_SELECT_GREEN,
314 ISL_CHANNEL_SELECT_BLUE,
315 ISL_CHANNEL_SELECT_ALPHA,
317 .usage = is_render_target ? ISL_SURF_USAGE_RENDER_TARGET_BIT :
318 ISL_SURF_USAGE_TEXTURE_BIT,
321 uint32_t surf_offset;
322 uint32_t *dw = brw_state_batch(brw, AUB_TRACE_SURFACE_STATE,
323 ss_info.num_dwords * 4, ss_info.ss_align,
326 const uint32_t mocs = is_render_target ? ss_info.rb_mocs : ss_info.tex_mocs;
328 isl_surf_fill_state(&brw->isl_dev, dw, .surf = &surf, .view = &view,
329 .address = surface->mt->bo->offset64 + surface->bo_offset,
330 .aux_surf = aux_surf, .aux_usage = surface->aux_usage,
331 .aux_address = aux_offset,
332 .mocs = mocs, .clear_color = clear_color,
333 .x_offset_sa = surface->tile_x_sa,
334 .y_offset_sa = surface->tile_y_sa);
336 /* Emit relocation to surface contents */
337 drm_intel_bo_emit_reloc(brw->batch.bo,
338 surf_offset + ss_info.reloc_dw * 4,
340 dw[ss_info.reloc_dw] - surface->mt->bo->offset64,
341 read_domains, write_domain);
344 /* On gen7 and prior, the bottom 12 bits of the MCS base address are
345 * used to store other information. This should be ok, however, because
346 * surface buffer addresses are always 4K page alinged.
348 assert((aux_offset & 0xfff) == 0);
349 drm_intel_bo_emit_reloc(brw->batch.bo,
350 surf_offset + ss_info.aux_reloc_dw * 4,
351 surface->mt->mcs_mt->bo,
352 dw[ss_info.aux_reloc_dw] & 0xfff,
353 read_domains, write_domain);
360 * Perform a HiZ or depth resolve operation.
362 * For an overview of HiZ ops, see the following sections of the Sandy Bridge
363 * PRM, Volume 1, Part 2:
364 * - 7.5.3.1 Depth Buffer Clear
365 * - 7.5.3.2 Depth Buffer Resolve
366 * - 7.5.3.3 Hierarchical Depth Buffer Resolve
369 intel_hiz_exec(struct brw_context *brw, struct intel_mipmap_tree *mt,
370 unsigned int level, unsigned int layer, enum gen6_hiz_op op)
372 const char *opname = NULL;
375 case GEN6_HIZ_OP_DEPTH_RESOLVE:
376 opname = "depth resolve";
378 case GEN6_HIZ_OP_HIZ_RESOLVE:
379 opname = "hiz ambiguate";
381 case GEN6_HIZ_OP_DEPTH_CLEAR:
382 opname = "depth clear";
384 case GEN6_HIZ_OP_NONE:
389 DBG("%s %s to mt %p level %d layer %d\n",
390 __func__, opname, mt, level, layer);
393 gen8_hiz_exec(brw, mt, level, layer, op);
395 gen6_blorp_hiz_exec(brw, mt, level, layer, op);
400 brw_blorp_exec(struct brw_context *brw, const struct brw_blorp_params *params)
402 struct gl_context *ctx = &brw->ctx;
403 const uint32_t estimated_max_batch_usage = brw->gen >= 8 ? 1800 : 1500;
404 bool check_aperture_failed_once = false;
406 /* Flush the sampler and render caches. We definitely need to flush the
407 * sampler cache so that we get updated contents from the render cache for
408 * the glBlitFramebuffer() source. Also, we are sometimes warned in the
409 * docs to flush the cache between reinterpretations of the same surface
410 * data with different formats, which blorp does for stencil and depth
413 brw_emit_mi_flush(brw);
415 brw_select_pipeline(brw, BRW_RENDER_PIPELINE);
418 intel_batchbuffer_require_space(brw, estimated_max_batch_usage, RENDER_RING);
419 intel_batchbuffer_save_state(brw);
420 drm_intel_bo *saved_bo = brw->batch.bo;
421 uint32_t saved_used = USED_BATCH(brw->batch);
422 uint32_t saved_state_batch_offset = brw->batch.state_batch_offset;
426 gen6_blorp_exec(brw, params);
429 gen7_blorp_exec(brw, params);
433 gen8_blorp_exec(brw, params);
436 /* BLORP is not supported before Gen6. */
437 unreachable("not reached");
440 /* Make sure we didn't wrap the batch unintentionally, and make sure we
441 * reserved enough space that a wrap will never happen.
443 assert(brw->batch.bo == saved_bo);
444 assert((USED_BATCH(brw->batch) - saved_used) * 4 +
445 (saved_state_batch_offset - brw->batch.state_batch_offset) <
446 estimated_max_batch_usage);
447 /* Shut up compiler warnings on release build */
450 (void)saved_state_batch_offset;
452 /* Check if the blorp op we just did would make our batch likely to fail to
453 * map all the BOs into the GPU at batch exec time later. If so, flush the
454 * batch and try again with nothing else in the batch.
456 if (dri_bufmgr_check_aperture_space(&brw->batch.bo, 1)) {
457 if (!check_aperture_failed_once) {
458 check_aperture_failed_once = true;
459 intel_batchbuffer_reset_to_saved(brw);
460 intel_batchbuffer_flush(brw);
463 int ret = intel_batchbuffer_flush(brw);
464 WARN_ONCE(ret == -ENOSPC,
465 "i965: blorp emit exceeded available aperture space\n");
469 if (unlikely(brw->always_flush_batch))
470 intel_batchbuffer_flush(brw);
472 /* We've smashed all state compared to what the normal 3D pipeline
473 * rendering tracks for GL.
475 brw->ctx.NewDriverState |= BRW_NEW_BLORP;
476 brw->no_depth_or_stencil = false;
479 /* Flush the sampler cache so any texturing from the destination is
482 brw_emit_mi_flush(brw);
486 gen6_blorp_hiz_exec(struct brw_context *brw, struct intel_mipmap_tree *mt,
487 unsigned int level, unsigned int layer, enum gen6_hiz_op op)
489 struct brw_blorp_params params;
490 brw_blorp_params_init(¶ms);
494 brw_blorp_surface_info_init(brw, ¶ms.depth, mt, level, layer,
497 /* Align the rectangle primitive to 8x4 pixels.
499 * During fast depth clears, the emitted rectangle primitive must be
500 * aligned to 8x4 pixels. From the Ivybridge PRM, Vol 2 Part 1 Section
501 * 11.5.3.1 Depth Buffer Clear (and the matching section in the Sandybridge
503 * If Number of Multisamples is NUMSAMPLES_1, the rectangle must be
504 * aligned to an 8x4 pixel block relative to the upper left corner
505 * of the depth buffer [...]
507 * For hiz resolves, the rectangle must also be 8x4 aligned. Item
508 * WaHizAmbiguate8x4Aligned from the Haswell workarounds page and the
509 * Ivybridge simulator require the alignment.
511 * To be safe, let's just align the rect for all hiz operations and all
512 * hardware generations.
514 * However, for some miptree slices of a Z24 texture, emitting an 8x4
515 * aligned rectangle that covers the slice may clobber adjacent slices if
516 * we strictly adhered to the texture alignments specified in the PRM. The
517 * Ivybridge PRM, Section "Alignment Unit Size", states that
518 * SURFACE_STATE.Surface_Horizontal_Alignment should be 4 for Z24 surfaces,
519 * not 8. But commit 1f112cc increased the alignment from 4 to 8, which
520 * prevents the clobbering.
522 params.dst.surf.samples = MAX2(mt->num_samples, 1);
523 if (params.depth.surf.samples > 1) {
524 params.depth.width = ALIGN(mt->logical_width0, 8);
525 params.depth.height = ALIGN(mt->logical_height0, 4);
527 params.depth.width = ALIGN(params.depth.width, 8);
528 params.depth.height = ALIGN(params.depth.height, 4);
531 params.x1 = params.depth.width;
532 params.y1 = params.depth.height;
534 assert(intel_miptree_level_has_hiz(mt, level));
536 switch (mt->format) {
537 case MESA_FORMAT_Z_UNORM16:
538 params.depth_format = BRW_DEPTHFORMAT_D16_UNORM;
540 case MESA_FORMAT_Z_FLOAT32:
541 params.depth_format = BRW_DEPTHFORMAT_D32_FLOAT;
543 case MESA_FORMAT_Z24_UNORM_X8_UINT:
544 params.depth_format = BRW_DEPTHFORMAT_D24_UNORM_X8_UINT;
547 unreachable("not reached");
550 brw_blorp_exec(brw, ¶ms);