xref: /aosp_15_r20/external/mesa3d/src/broadcom/common/v3d_util.c (revision 6104692788411f58d303aa86923a9ff6ecaded22)
1 /*
2  * Copyright © 2021 Raspberry Pi Ltd
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 "v3d_util.h"
25 #include "util/macros.h"
26 
27 /* Choose a number of workgroups per supergroup that maximizes
28  * lane occupancy. We can pack up to 16 workgroups into a supergroup.
29  */
30 uint32_t
v3d_csd_choose_workgroups_per_supergroup(struct v3d_device_info * devinfo,bool has_subgroups,bool has_tsy_barrier,uint32_t threads,uint32_t num_wgs,uint32_t wg_size)31 v3d_csd_choose_workgroups_per_supergroup(struct v3d_device_info *devinfo,
32                                          bool has_subgroups,
33                                          bool has_tsy_barrier,
34                                          uint32_t threads,
35                                          uint32_t num_wgs,
36                                          uint32_t wg_size)
37 {
38    /* FIXME: subgroups may restrict supergroup packing. For now, we disable it
39     * completely if the shader uses subgroups.
40     */
41    if (has_subgroups)
42            return 1;
43 
44    /* Compute maximum number of batches in a supergroup for this workgroup size.
45     * Each batch is 16 elements, and we can have up to 16 work groups in a
46     * supergroup:
47     *
48     * max_batches_per_sg = (wg_size * max_wgs_per_sg) / elements_per_batch
49     * since max_wgs_per_sg = 16 and elements_per_batch = 16, we get:
50     * max_batches_per_sg = wg_size
51     */
52    uint32_t max_batches_per_sg = wg_size;
53 
54    /* QPU threads will stall at TSY barriers until the entire supergroup
55     * reaches the barrier. Limit the supergroup size to half the QPU threads
56     * available, so we can have at least 2 supergroups executing in parallel
57     * and we don't stall all our QPU threads when a supergroup hits a barrier.
58     */
59    if (has_tsy_barrier) {
60       uint32_t max_qpu_threads = devinfo->qpu_count * threads;
61       max_batches_per_sg = MIN2(max_batches_per_sg, max_qpu_threads / 2);
62    }
63    uint32_t max_wgs_per_sg = max_batches_per_sg * 16 / wg_size;
64 
65    uint32_t best_wgs_per_sg = 1;
66    uint32_t best_unused_lanes = 16;
67    for (uint32_t wgs_per_sg = 1; wgs_per_sg <= max_wgs_per_sg; wgs_per_sg++) {
68       /* Don't try to pack more workgroups per supergroup than the total amount
69        * of workgroups dispatched.
70        */
71       if (wgs_per_sg > num_wgs)
72          return best_wgs_per_sg;
73 
74       /* Compute wasted lines for this configuration and keep track of the
75        * config with less waste.
76        */
77       uint32_t unused_lanes = (16 - ((wgs_per_sg * wg_size) % 16)) & 0x0f;
78       if (unused_lanes == 0)
79          return wgs_per_sg;
80 
81       if (unused_lanes < best_unused_lanes) {
82          best_wgs_per_sg = wgs_per_sg;
83          best_unused_lanes = unused_lanes;
84       }
85    }
86 
87    return best_wgs_per_sg;
88 }
89 
90 #define V3D71_TLB_COLOR_SIZE     (16 * 1024)
91 #define V3D71_TLB_DETPH_SIZE     (16 * 1024)
92 #define V3D71_TLB_AUX_DETPH_SIZE  (8 * 1024)
93 
94 static bool
tile_size_valid(uint32_t pixel_count,uint32_t color_bpp,uint32_t depth_bpp)95 tile_size_valid(uint32_t pixel_count, uint32_t color_bpp, uint32_t depth_bpp)
96 {
97    /* First, we check if we can fit this tile size allocating the depth
98     * TLB memory to color.
99     */
100    if (pixel_count * depth_bpp <= V3D71_TLB_AUX_DETPH_SIZE &&
101        pixel_count * color_bpp <= V3D71_TLB_COLOR_SIZE + V3D71_TLB_DETPH_SIZE) {
102       return true;
103    }
104 
105    /* Otherwise the tile must fit in the main TLB buffers */
106    return pixel_count * depth_bpp <= V3D71_TLB_DETPH_SIZE &&
107           pixel_count * color_bpp <= V3D71_TLB_COLOR_SIZE;
108 }
109 
110 void
v3d_choose_tile_size(const struct v3d_device_info * devinfo,uint32_t color_attachment_count,uint32_t max_internal_bpp,uint32_t total_color_bpp,bool msaa,bool double_buffer,uint32_t * width,uint32_t * height)111 v3d_choose_tile_size(const struct v3d_device_info *devinfo,
112                      uint32_t color_attachment_count,
113                      /* V3D 4.x max internal bpp of all RTs */
114                      uint32_t max_internal_bpp,
115                      /* V3D 7.x accumulated bpp for all RTs (in bytes) */
116                      uint32_t total_color_bpp,
117                      bool msaa,
118                      bool double_buffer,
119                      uint32_t *width,
120                      uint32_t *height)
121 {
122    static const uint8_t tile_sizes[] = {
123       64, 64,
124       64, 32,
125       32, 32,
126       32, 16,
127       16, 16,
128       16,  8,
129        8,  8
130    };
131 
132    uint32_t idx = 0;
133    if (devinfo->ver >= 71) {
134       /* In V3D 7.x, we use the actual bpp used by color attachments to compute
135        * the tile size instead of the maximum bpp. This may allow us to choose a
136        * larger tile size than we would in 4.x in scenarios with multiple RTs
137        * with different bpps.
138        *
139        * Also, the TLB has an auxiliary buffer of 8KB that will be automatically
140        * used for depth instead of the main 16KB depth TLB buffer when the depth
141        * tile fits in the auxiliary buffer, allowing the hardware to allocate
142        * the 16KB from the main depth TLB to the color TLB. If we can do that,
143        * then we are effectively doubling the memory we have for color and we
144        * can also select a larger tile size. This is necessary to support
145        * the most expensive configuration: 8x128bpp RTs + MSAA.
146        *
147        * FIXME: the docs state that depth TLB memory can be used for color
148        * if depth testing is not used by setting the 'depth disable' bit in the
149        * rendering configuration. However, this comes with a requirement that
150        * occlussion queries must not be active. We need to clarify if this means
151        * active at the point at which we emit a tile rendering configuration
152        * item, meaning that the we have a query spanning a full render pass
153        * (this is something we can tell before we emit the rendering
154        * configuration item) or active in the subpass for which we are enabling
155        * the bit (which we can't tell until later, when we record commands for
156        * the subpass). If it is the latter, then we cannot use this feature.
157        *
158        * FIXME: pending handling double_buffer.
159        */
160       const uint32_t color_bpp = total_color_bpp * (msaa ? 4 : 1);
161       const uint32_t depth_bpp = 4 * (msaa ? 4 : 1);
162       do {
163          const uint32_t tile_w = tile_sizes[idx * 2];
164          const uint32_t tile_h = tile_sizes[idx * 2 + 1];
165          if (tile_size_valid(tile_w * tile_h, color_bpp, depth_bpp))
166             break;
167          idx++;
168       } while (idx < ARRAY_SIZE(tile_sizes) / 2);
169 
170       /* FIXME: pending handling double_buffer */
171       assert(!double_buffer);
172    } else {
173       /* On V3D 4.x tile size is selected based on the number of RTs, the
174        * maximum bpp across all of them and whether 4x MSAA is used.
175        */
176       if (color_attachment_count > 4)
177          idx += 3;
178       else if (color_attachment_count > 2)
179          idx += 2;
180       else if (color_attachment_count > 1)
181          idx += 1;
182 
183       /* MSAA and double-buffer are mutually exclusive */
184       assert(!msaa || !double_buffer);
185       if (msaa)
186          idx += 2;
187       else if (double_buffer)
188          idx += 1;
189 
190       idx += max_internal_bpp;
191    }
192 
193    assert(idx < ARRAY_SIZE(tile_sizes) / 2);
194 
195    *width = tile_sizes[idx * 2];
196    *height = tile_sizes[idx * 2 + 1];
197 }
198 
199 /* Translates a pipe swizzle to the swizzle values used in the
200  * TEXTURE_SHADER_STATE packet.
201  */
202 uint32_t
v3d_translate_pipe_swizzle(enum pipe_swizzle swizzle)203 v3d_translate_pipe_swizzle(enum pipe_swizzle swizzle)
204 {
205    switch (swizzle) {
206    case PIPE_SWIZZLE_0:
207       return 0;
208    case PIPE_SWIZZLE_1:
209       return 1;
210    case PIPE_SWIZZLE_X:
211    case PIPE_SWIZZLE_Y:
212    case PIPE_SWIZZLE_Z:
213    case PIPE_SWIZZLE_W:
214       return 2 + swizzle;
215    default:
216       unreachable("unknown swizzle");
217    }
218 }
219 
220 /* Translates a pipe primitive type to a hw value we can use in the various
221  * draw packets.
222  */
223 uint32_t
v3d_hw_prim_type(enum mesa_prim prim_type)224 v3d_hw_prim_type(enum mesa_prim prim_type)
225 {
226    switch (prim_type) {
227    case MESA_PRIM_POINTS:
228    case MESA_PRIM_LINES:
229    case MESA_PRIM_LINE_LOOP:
230    case MESA_PRIM_LINE_STRIP:
231    case MESA_PRIM_TRIANGLES:
232    case MESA_PRIM_TRIANGLE_STRIP:
233    case MESA_PRIM_TRIANGLE_FAN:
234       return prim_type;
235 
236    case MESA_PRIM_LINES_ADJACENCY:
237    case MESA_PRIM_LINE_STRIP_ADJACENCY:
238    case MESA_PRIM_TRIANGLES_ADJACENCY:
239    case MESA_PRIM_TRIANGLE_STRIP_ADJACENCY:
240       return 8 + (prim_type - MESA_PRIM_LINES_ADJACENCY);
241 
242    default:
243       unreachable("Unsupported primitive type");
244    }
245 }
246 
247 uint32_t
v3d_internal_bpp_words(uint32_t internal_bpp)248 v3d_internal_bpp_words(uint32_t internal_bpp)
249 {
250         switch (internal_bpp) {
251         case 0 /* V3D_INTERNAL_BPP_32 */:
252                 return 1;
253         case 1 /* V3D_INTERNAL_BPP_64 */:
254                 return 2;
255         case 2 /* V3D_INTERNAL_BPP_128 */:
256                 return 4;
257         default:
258                 unreachable("Unsupported internal BPP");
259         }
260 }
261 
262 uint32_t
v3d_compute_rt_row_row_stride_128_bits(uint32_t tile_width,uint32_t bpp)263 v3d_compute_rt_row_row_stride_128_bits(uint32_t tile_width,
264                                        uint32_t bpp)
265 {
266         /* stride in multiples of 128 bits, and covers 2 rows. This is the
267          * reason we divide by 2 instead of 4, as we divide number of 32-bit
268          * words per row by 2.
269          */
270 
271         return (tile_width * bpp) / 2;
272 }
273