xref: /aosp_15_r20/external/mesa3d/src/intel/blorp/blorp.c (revision 6104692788411f58d303aa86923a9ff6ecaded22)
1 /*
2  * Copyright © 2012 Intel Corporation
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  * 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:
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 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
21  * IN THE SOFTWARE.
22  */
23 
24 #include "blorp_priv.h"
25 #include "compiler/intel_nir.h"
26 #include "dev/intel_device_info.h"
27 #include "util/u_math.h"
28 
29 enum intel_measure_snapshot_type
blorp_op_to_intel_measure_snapshot(enum blorp_op op)30 blorp_op_to_intel_measure_snapshot(enum blorp_op op)
31 {
32    enum intel_measure_snapshot_type vals[] = {
33 #define MAP(name) [BLORP_OP_##name] = INTEL_SNAPSHOT_##name
34       MAP(BLIT),
35       MAP(COPY),
36       MAP(CCS_AMBIGUATE),
37       MAP(CCS_COLOR_CLEAR),
38       MAP(CCS_PARTIAL_RESOLVE),
39       MAP(CCS_RESOLVE),
40       MAP(HIZ_AMBIGUATE),
41       MAP(HIZ_CLEAR),
42       MAP(HIZ_RESOLVE),
43       MAP(MCS_AMBIGUATE),
44       MAP(MCS_COLOR_CLEAR),
45       MAP(MCS_PARTIAL_RESOLVE),
46       MAP(SLOW_COLOR_CLEAR),
47       MAP(SLOW_DEPTH_CLEAR),
48 #undef MAP
49    };
50    assert(op < ARRAY_SIZE(vals));
51 
52    return vals[op];
53 }
54 
blorp_op_to_name(enum blorp_op op)55 const char *blorp_op_to_name(enum blorp_op op)
56 {
57    const char *names[] = {
58 #define MAP(name) [BLORP_OP_##name] = #name
59       MAP(BLIT),
60       MAP(COPY),
61       MAP(CCS_AMBIGUATE),
62       MAP(CCS_COLOR_CLEAR),
63       MAP(CCS_PARTIAL_RESOLVE),
64       MAP(CCS_RESOLVE),
65       MAP(HIZ_AMBIGUATE),
66       MAP(HIZ_CLEAR),
67       MAP(HIZ_RESOLVE),
68       MAP(MCS_AMBIGUATE),
69       MAP(MCS_COLOR_CLEAR),
70       MAP(MCS_PARTIAL_RESOLVE),
71       MAP(SLOW_COLOR_CLEAR),
72       MAP(SLOW_DEPTH_CLEAR),
73 #undef MAP
74    };
75    assert(op < ARRAY_SIZE(names));
76 
77    return names[op];
78 }
79 
80 const char *
blorp_shader_type_to_name(enum blorp_shader_type type)81 blorp_shader_type_to_name(enum blorp_shader_type type)
82 {
83    static const char *shader_name[] = {
84       [BLORP_SHADER_TYPE_COPY]                = "BLORP-copy",
85       [BLORP_SHADER_TYPE_BLIT]                = "BLORP-blit",
86       [BLORP_SHADER_TYPE_CLEAR]               = "BLORP-clear",
87       [BLORP_SHADER_TYPE_MCS_PARTIAL_RESOLVE] = "BLORP-mcs-partial-resolve",
88       [BLORP_SHADER_TYPE_LAYER_OFFSET_VS]     = "BLORP-layer-offset-vs",
89       [BLORP_SHADER_TYPE_GFX4_SF]             = "BLORP-gfx4-sf",
90    };
91    assert(type < ARRAY_SIZE(shader_name));
92 
93    return shader_name[type];
94 }
95 
96 const char *
blorp_shader_pipeline_to_name(enum blorp_shader_pipeline pipe)97 blorp_shader_pipeline_to_name(enum blorp_shader_pipeline pipe)
98 {
99    static const char *pipeline_name[] = {
100       [BLORP_SHADER_PIPELINE_RENDER]  = "render",
101       [BLORP_SHADER_PIPELINE_COMPUTE] = "compute",
102    };
103    assert(pipe < ARRAY_SIZE(pipeline_name));
104 
105    return pipeline_name[pipe];
106 }
107 
108 void
blorp_init(struct blorp_context * blorp,void * driver_ctx,struct isl_device * isl_dev,const struct blorp_config * config)109 blorp_init(struct blorp_context *blorp, void *driver_ctx,
110            struct isl_device *isl_dev, const struct blorp_config *config)
111 {
112    memset(blorp, 0, sizeof(*blorp));
113 
114    blorp->driver_ctx = driver_ctx;
115    blorp->isl_dev = isl_dev;
116    if (config)
117       blorp->config = *config;
118 
119    blorp->compiler = rzalloc(NULL, struct blorp_compiler);
120 }
121 
122 void
blorp_finish(struct blorp_context * blorp)123 blorp_finish(struct blorp_context *blorp)
124 {
125    ralloc_free(blorp->compiler);
126    blorp->driver_ctx = NULL;
127 }
128 
129 void
blorp_batch_init(struct blorp_context * blorp,struct blorp_batch * batch,void * driver_batch,enum blorp_batch_flags flags)130 blorp_batch_init(struct blorp_context *blorp,
131                  struct blorp_batch *batch, void *driver_batch,
132                  enum blorp_batch_flags flags)
133 {
134    batch->blorp = blorp;
135    batch->driver_batch = driver_batch;
136    batch->flags = flags;
137 }
138 
139 void
blorp_batch_finish(struct blorp_batch * batch)140 blorp_batch_finish(struct blorp_batch *batch)
141 {
142    batch->blorp = NULL;
143 }
144 
145 void
blorp_surface_info_init(struct blorp_batch * batch,struct blorp_surface_info * info,const struct blorp_surf * surf,unsigned int level,float layer,enum isl_format format,bool is_dest)146 blorp_surface_info_init(struct blorp_batch *batch,
147                             struct blorp_surface_info *info,
148                             const struct blorp_surf *surf,
149                             unsigned int level, float layer,
150                             enum isl_format format, bool is_dest)
151 {
152    struct blorp_context *blorp = batch->blorp;
153    memset(info, 0, sizeof(*info));
154    assert(level < surf->surf->levels);
155    assert(layer < MAX2(surf->surf->logical_level0_px.depth >> level,
156                        surf->surf->logical_level0_px.array_len));
157 
158    info->enabled = true;
159 
160    if (format == ISL_FORMAT_UNSUPPORTED)
161       format = surf->surf->format;
162 
163    info->surf = *surf->surf;
164    info->addr = surf->addr;
165 
166    info->aux_usage = surf->aux_usage;
167    if (info->aux_usage != ISL_AUX_USAGE_NONE) {
168       info->aux_surf = *surf->aux_surf;
169       info->aux_addr = surf->aux_addr;
170    }
171 
172    info->clear_color = surf->clear_color;
173    info->clear_color_addr = surf->clear_color_addr;
174 
175    isl_surf_usage_flags_t view_usage;
176    if (is_dest) {
177       if (batch->flags & BLORP_BATCH_USE_COMPUTE)
178          view_usage = ISL_SURF_USAGE_STORAGE_BIT;
179       else
180          view_usage = ISL_SURF_USAGE_RENDER_TARGET_BIT;
181    } else {
182       view_usage = ISL_SURF_USAGE_TEXTURE_BIT;
183    }
184 
185    info->view = (struct isl_view) {
186       .usage = view_usage,
187       .format = format,
188       .base_level = level,
189       .levels = 1,
190       .swizzle = ISL_SWIZZLE_IDENTITY,
191    };
192 
193    info->view.array_len =
194       MAX2(u_minify(info->surf.logical_level0_px.depth, level),
195            info->surf.logical_level0_px.array_len);
196 
197    if (!is_dest &&
198        (info->surf.dim == ISL_SURF_DIM_3D ||
199         info->surf.msaa_layout == ISL_MSAA_LAYOUT_ARRAY)) {
200       /* 3-D textures don't support base_array layer and neither do 2-D
201        * multisampled textures on IVB so we need to pass it through the
202        * sampler in those cases.  These are also two cases where we are
203        * guaranteed that we won't be doing any funny surface hacks.
204        */
205       info->view.base_array_layer = 0;
206       info->z_offset = layer;
207    } else {
208       info->view.base_array_layer = layer;
209 
210       assert(info->view.array_len >= info->view.base_array_layer);
211       info->view.array_len -= info->view.base_array_layer;
212       info->z_offset = 0;
213    }
214 
215    /* Sandy Bridge and earlier have a limit of a maximum of 512 layers for
216     * layered rendering.
217     */
218    if (is_dest && blorp->isl_dev->info->ver <= 6)
219       info->view.array_len = MIN2(info->view.array_len, 512);
220 
221    if (surf->tile_x_sa || surf->tile_y_sa) {
222       /* This is only allowed on simple 2D surfaces without MSAA */
223       assert(info->surf.dim == ISL_SURF_DIM_2D);
224       assert(info->surf.samples == 1);
225       assert(info->surf.levels == 1);
226       assert(info->surf.logical_level0_px.array_len == 1);
227       assert(info->aux_usage == ISL_AUX_USAGE_NONE);
228 
229       info->tile_x_sa = surf->tile_x_sa;
230       info->tile_y_sa = surf->tile_y_sa;
231 
232       /* Instead of using the X/Y Offset fields in RENDER_SURFACE_STATE, we
233        * place the image at the tile boundary and offset our sampling or
234        * rendering.  For this reason, we need to grow the image by the offset
235        * to ensure that the hardware doesn't think we've gone past the edge.
236        */
237       info->surf.logical_level0_px.w += surf->tile_x_sa;
238       info->surf.logical_level0_px.h += surf->tile_y_sa;
239       info->surf.phys_level0_sa.w += surf->tile_x_sa;
240       info->surf.phys_level0_sa.h += surf->tile_y_sa;
241    }
242 }
243 
244 
245 void
blorp_params_init(struct blorp_params * params)246 blorp_params_init(struct blorp_params *params)
247 {
248    memset(params, 0, sizeof(*params));
249    params->num_samples = 1;
250    params->num_draw_buffers = 1;
251    params->num_layers = 1;
252 }
253 
254 void
blorp_hiz_op(struct blorp_batch * batch,struct blorp_surf * surf,uint32_t level,uint32_t start_layer,uint32_t num_layers,enum isl_aux_op op)255 blorp_hiz_op(struct blorp_batch *batch, struct blorp_surf *surf,
256              uint32_t level, uint32_t start_layer, uint32_t num_layers,
257              enum isl_aux_op op)
258 {
259    const struct intel_device_info *devinfo = batch->blorp->isl_dev->info;
260 
261    struct blorp_params params;
262    blorp_params_init(&params);
263 
264    params.hiz_op = op;
265    params.full_surface_hiz_op = true;
266    switch (op) {
267    case ISL_AUX_OP_FULL_RESOLVE:
268       params.op = BLORP_OP_HIZ_RESOLVE;
269       break;
270    case ISL_AUX_OP_AMBIGUATE:
271       params.op = BLORP_OP_HIZ_AMBIGUATE;
272       break;
273    case ISL_AUX_OP_FAST_CLEAR:
274       params.op = BLORP_OP_HIZ_CLEAR;
275       break;
276    case ISL_AUX_OP_PARTIAL_RESOLVE:
277    case ISL_AUX_OP_NONE:
278       unreachable("Invalid HiZ op");
279    }
280 
281    for (uint32_t a = 0; a < num_layers; a++) {
282       const uint32_t layer = start_layer + a;
283 
284       blorp_surface_info_init(batch, &params.depth, surf, level,
285                                   layer, surf->surf->format, true);
286 
287       /* Align the rectangle primitive to 8x4 pixels.
288        *
289        * During fast depth clears, the emitted rectangle primitive  must be
290        * aligned to 8x4 pixels.  From the Ivybridge PRM, Vol 2 Part 1 Section
291        * 11.5.3.1 Depth Buffer Clear (and the matching section in the
292        * Sandybridge PRM):
293        *
294        *     If Number of Multisamples is NUMSAMPLES_1, the rectangle must be
295        *     aligned to an 8x4 pixel block relative to the upper left corner
296        *     of the depth buffer [...]
297        *
298        * For hiz resolves, the rectangle must also be 8x4 aligned. Item
299        * WaHizAmbiguate8x4Aligned from the Haswell workarounds page and the
300        * Ivybridge simulator require the alignment.
301        *
302        * To be safe, let's just align the rect for all hiz operations and all
303        * hardware generations.
304        *
305        * However, for some miptree slices of a Z24 texture, emitting an 8x4
306        * aligned rectangle that covers the slice may clobber adjacent slices
307        * if we strictly adhered to the texture alignments specified in the
308        * PRM.  The Ivybridge PRM, Section "Alignment Unit Size", states that
309        * SURFACE_STATE.Surface_Horizontal_Alignment should be 4 for Z24
310        * surfaces, not 8. But commit 1f112cc increased the alignment from 4 to
311        * 8, which prevents the clobbering.
312        */
313       params.x1 = u_minify(params.depth.surf.logical_level0_px.width,
314                            params.depth.view.base_level);
315       params.y1 = u_minify(params.depth.surf.logical_level0_px.height,
316                            params.depth.view.base_level);
317       params.x1 = ALIGN(params.x1, 8);
318       params.y1 = ALIGN(params.y1, 4);
319 
320       if (params.depth.view.base_level == 0) {
321          /* TODO: What about MSAA? */
322          params.depth.surf.logical_level0_px.width = params.x1;
323          params.depth.surf.logical_level0_px.height = params.y1;
324       } else if (devinfo->ver >= 8 && devinfo->ver <= 9 &&
325                  op == ISL_AUX_OP_AMBIGUATE) {
326          /* On some platforms, it's not enough to just adjust the clear
327           * rectangle when the LOD is greater than 0.
328           *
329           * From the BDW and SKL PRMs, Vol 7, "Optimized Hierarchical Depth
330           * Buffer Resolve":
331           *
332           *    The following is required when performing a hierarchical depth
333           *    buffer resolve:
334           *
335           *    - A rectangle primitive covering the full render target must be
336           *      programmed on Xmin, Ymin, Xmax, and Ymax in the
337           *      3DSTATE_WM_HZ_OP command.
338           *
339           *    - The rectangle primitive size must be aligned to 8x4 pixels.
340           *
341           * And from the Clear Rectangle programming note in 3DSTATE_WM_HZ_OP
342           * (Vol 2a):
343           *
344           *    Hence the max values must be less than or equal to: ( Surface
345           *    Width » LOD ) and ( Surface Height » LOD ) for X Max and Y Max
346           *    respectively.
347           *
348           * This means that the extent of the LOD must be naturally
349           * 8x4-aligned after minification of the base LOD. Since the base LOD
350           * dimensions affect the placement of smaller LODs, it's not trivial
351           * (nor possible, at times) to satisfy the requirement by adjusting
352           * the base LOD extent. Just assert that the caller is accessing an
353           * LOD that satisfies this requirement.
354           */
355          assert(u_minify(params.depth.surf.logical_level0_px.width,
356                          params.depth.view.base_level) == params.x1);
357          assert(u_minify(params.depth.surf.logical_level0_px.height,
358                          params.depth.view.base_level) == params.y1);
359       }
360 
361       params.dst.surf.samples = params.depth.surf.samples;
362       params.dst.surf.logical_level0_px = params.depth.surf.logical_level0_px;
363       params.depth_format =
364          isl_format_get_depth_format(surf->surf->format, false);
365       params.num_samples = params.depth.surf.samples;
366 
367       batch->blorp->exec(batch, &params);
368    }
369 }
370