xref: /aosp_15_r20/external/mesa3d/src/intel/compiler/brw_rt.h (revision 6104692788411f58d303aa86923a9ff6ecaded22)
1 /*
2  * Copyright © 2020 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 #ifndef BRW_RT_H
25 #define BRW_RT_H
26 
27 #include <stdint.h>
28 
29 #include "compiler/shader_enums.h"
30 #include "util/macros.h"
31 
32 #ifdef __cplusplus
33 extern "C" {
34 #endif
35 
36 /** Vulkan defines shaderGroupHandleSize = 32 */
37 #define BRW_RT_SBT_HANDLE_SIZE 32
38 
39 /** RT_DISPATCH_GLOBALS size (see gen_rt.xml) */
40 #define BRW_RT_DISPATCH_GLOBALS_SIZE 80
41 
42 /** Offset after the RT dispatch globals at which "push" constants live */
43 #define BRW_RT_PUSH_CONST_OFFSET 128
44 
45 /** Stride of the resume SBT */
46 #define BRW_BTD_RESUME_SBT_STRIDE 8
47 
48 /* Vulkan always uses exactly two levels of BVH: world and object.  At the API
49  * level, these are referred to as top and bottom.
50  */
51 enum brw_rt_bvh_level {
52    BRW_RT_BVH_LEVEL_WORLD = 0,
53    BRW_RT_BVH_LEVEL_OBJECT = 1,
54 };
55 #define BRW_RT_MAX_BVH_LEVELS 2
56 
57 enum brw_rt_bvh_node_type {
58    BRW_RT_BVH_NODE_TYPE_INTERNAL = 0,
59    BRW_RT_BVH_NODE_TYPE_INSTANCE = 1,
60    BRW_RT_BVH_NODE_TYPE_PROCEDURAL = 3,
61    BRW_RT_BVH_NODE_TYPE_QUAD = 4,
62 };
63 
64 /** HitKind values returned for triangle geometry
65  *
66  * This enum must match the SPIR-V enum.
67  */
68 enum brw_rt_hit_kind {
69    BRW_RT_HIT_KIND_FRONT_FACE = 0xfe,
70    BRW_RT_HIT_KIND_BACK_FACE = 0xff,
71 };
72 
73 /** Ray flags
74  *
75  * This enum must match the SPIR-V RayFlags enum.
76  */
77 enum brw_rt_ray_flags {
78    BRW_RT_RAY_FLAG_FORCE_OPAQUE                    = 0x01,
79    BRW_RT_RAY_FLAG_FORCE_NON_OPAQUE                = 0x02,
80    BRW_RT_RAY_FLAG_TERMINATE_ON_FIRST_HIT          = 0x04,
81    BRW_RT_RAY_FLAG_SKIP_CLOSEST_HIT_SHADER         = 0x08,
82    BRW_RT_RAY_FLAG_CULL_BACK_FACING_TRIANGLES      = 0x10,
83    BRW_RT_RAY_FLAG_CULL_FRONT_FACING_TRIANGLES     = 0x20,
84    BRW_RT_RAY_FLAG_CULL_OPAQUE                     = 0x40,
85    BRW_RT_RAY_FLAG_CULL_NON_OPAQUE                 = 0x80,
86    BRW_RT_RAY_FLAG_SKIP_TRIANGLES                  = 0x100,
87    BRW_RT_RAY_FLAG_SKIP_AABBS                      = 0x200,
88 };
89 
90 struct brw_rt_scratch_layout {
91    /** Number of stack IDs per DSS */
92    uint32_t stack_ids_per_dss;
93 
94    /** Start offset (in bytes) of the hardware MemRay stack */
95    uint32_t ray_stack_start;
96 
97    /** Stride (in bytes) of the hardware MemRay stack */
98    uint32_t ray_stack_stride;
99 
100    /** Start offset (in bytes) of the SW stacks */
101    uint64_t sw_stack_start;
102 
103    /** Size (in bytes) of the SW stack for a single shader invocation */
104    uint32_t sw_stack_size;
105 
106    /** Total size (in bytes) of the RT scratch memory area */
107    uint64_t total_size;
108 };
109 
110 /** Parameters passed to the raygen trampoline shader
111  *
112  * This struct is carefully construected to be 32B and must be passed to the
113  * raygen trampoline shader as as inline constant data.
114  */
115 struct brw_rt_raygen_trampoline_params {
116    /** The GPU address of the RT_DISPATCH_GLOBALS */
117    uint64_t rt_disp_globals_addr;
118 
119    /** The GPU address of the BINDLESS_SHADER_RECORD for the raygen shader */
120    uint64_t raygen_bsr_addr;
121 
122    /** 1 if this is an indirect dispatch, 0 otherwise */
123    uint8_t is_indirect;
124 
125    /** The integer log2 of the local group size
126     *
127     * Ray-tracing shaders don't have a concept of local vs. global workgroup
128     * size.  They only have a single 3D launch size.  The raygen trampoline
129     * shader is always dispatched with a local workgroup size equal to the
130     * SIMD width but the shape of the local workgroup is determined at
131     * dispatch time based on the shape of the launch and passed to the
132     * trampoline via this field.  (There's no sense having a Z dimension on
133     * the local workgroup if the launch is 2D.)
134     *
135     * We use the integer log2 of the size because there's no point in
136     * non-power-of-two sizes and  shifts are cheaper than division.
137     */
138    uint8_t local_group_size_log2[3];
139 
140    uint32_t pad[3];
141 };
142 
143 /** Size of the "hot zone" in bytes
144  *
145  * The hot zone is a SW-defined data structure which is a single uvec4
146  * containing two bits of information:
147  *
148  *  - hotzone.x: Stack offset (in bytes)
149  *
150  *    This is the offset (in bytes) into the per-thread scratch space at which
151  *    the current shader's stack starts.  This is incremented by the calling
152  *    shader prior to any shader call type instructions and gets decremented
153  *    by the resume shader as part of completing the return operation.
154  *
155  *
156  *  - hotzone.yzw: The launch ID associated with the current thread
157  *
158  *    Inside a bindless shader, the only information we have is the DSS ID
159  *    from the hardware EU and a per-DSS stack ID.  In particular, the three-
160  *    dimensional launch ID is lost the moment we leave the raygen trampoline.
161  */
162 #define BRW_RT_SIZEOF_HOTZONE 16
163 
164 /* From the BSpec "Address Computation for Memory Based Data Structures:
165  * Ray and TraversalStack (Async Ray Tracing)":
166  *
167  *    sizeof(Ray) = 64B, sizeof(HitInfo) = 32B, sizeof(TravStack) = 32B.
168  */
169 #define BRW_RT_SIZEOF_RAY 64
170 #define BRW_RT_SIZEOF_HIT_INFO 32
171 #define BRW_RT_SIZEOF_TRAV_STACK 32
172 
173 /* From the BSpec:
174  *
175  *    syncStackSize = (maxBVHLevels % 2 == 1) ?
176  *       (sizeof(HitInfo) * 2 +
177  *          (sizeof(Ray) + sizeof(TravStack)) * maxBVHLevels + 32B) :
178  *       (sizeof(HitInfo) * 2 +
179  *          (sizeof(Ray) + sizeof(TravStack)) * maxBVHLevels);
180  *
181  * The select is just to align to 64B.
182  */
183 #define BRW_RT_SIZEOF_RAY_QUERY \
184    (BRW_RT_SIZEOF_HIT_INFO * 2 + \
185     (BRW_RT_SIZEOF_RAY + BRW_RT_SIZEOF_TRAV_STACK) * BRW_RT_MAX_BVH_LEVELS + \
186     (BRW_RT_MAX_BVH_LEVELS % 2 ? 32 : 0))
187 
188 #define BRW_RT_SIZEOF_SHADOW_RAY_QUERY  \
189    (BRW_RT_SIZEOF_HIT_INFO * 2 + \
190     (BRW_RT_SIZEOF_RAY + BRW_RT_SIZEOF_TRAV_STACK) * BRW_RT_MAX_BVH_LEVELS)
191 
192 #define BRW_RT_SIZEOF_HW_STACK \
193    (BRW_RT_SIZEOF_HIT_INFO * 2 + \
194     BRW_RT_SIZEOF_RAY * BRW_RT_MAX_BVH_LEVELS + \
195     BRW_RT_SIZEOF_TRAV_STACK * BRW_RT_MAX_BVH_LEVELS)
196 
197 /* This is a mesa-defined region for hit attribute data */
198 #define BRW_RT_SIZEOF_HIT_ATTRIB_DATA 64
199 #define BRW_RT_OFFSETOF_HIT_ATTRIB_DATA BRW_RT_SIZEOF_HW_STACK
200 
201 #define BRW_RT_ASYNC_STACK_STRIDE \
202    ALIGN_POT(BRW_RT_OFFSETOF_HIT_ATTRIB_DATA + \
203              BRW_RT_SIZEOF_HIT_ATTRIB_DATA, 64)
204 
205 static inline void
brw_rt_compute_scratch_layout(struct brw_rt_scratch_layout * layout,const struct intel_device_info * devinfo,uint32_t stack_ids_per_dss,uint32_t sw_stack_size)206 brw_rt_compute_scratch_layout(struct brw_rt_scratch_layout *layout,
207                               const struct intel_device_info *devinfo,
208                               uint32_t stack_ids_per_dss,
209                               uint32_t sw_stack_size)
210 {
211    layout->stack_ids_per_dss = stack_ids_per_dss;
212 
213    const uint32_t dss_count = intel_device_info_dual_subslice_id_bound(devinfo);
214    const uint32_t num_stack_ids = dss_count * stack_ids_per_dss;
215 
216    uint64_t size = 0;
217 
218    /* The first thing in our scratch area is an array of "hot zones" which
219     * store the stack offset as well as the launch IDs for each active
220     * invocation.
221     */
222    size += BRW_RT_SIZEOF_HOTZONE * num_stack_ids;
223 
224    /* Next, we place the HW ray stacks */
225    assert(size % 64 == 0); /* Cache-line aligned */
226    assert(size < UINT32_MAX);
227    layout->ray_stack_start = size;
228    layout->ray_stack_stride = BRW_RT_ASYNC_STACK_STRIDE;
229    size += num_stack_ids * layout->ray_stack_stride;
230 
231    /* Finally, we place the SW stacks for the individual ray-tracing shader
232     * invocations.  We align these to 64B to ensure that we don't have any
233     * shared cache lines which could hurt performance.
234     */
235    assert(size % 64 == 0);
236    layout->sw_stack_start = size;
237    layout->sw_stack_size = ALIGN(sw_stack_size, 64);
238 
239    /* Currently it's always the case that sw_stack_size is a power of
240     * two, but power-of-two SW stack sizes are prone to causing
241     * collisions in the hashing function used by the L3 to map memory
242     * addresses to banks, which can cause stack accesses from most
243     * DSSes to bottleneck on a single L3 bank.  Fix it by padding the
244     * SW stack by a single cacheline if it was a power of two.
245     */
246    if (layout->sw_stack_size > 64 &&
247        util_is_power_of_two_nonzero(layout->sw_stack_size))
248       layout->sw_stack_size += 64;
249 
250    size += num_stack_ids * layout->sw_stack_size;
251 
252    layout->total_size = size;
253 }
254 
255 static inline uint32_t
brw_rt_ray_queries_hw_stacks_size(const struct intel_device_info * devinfo)256 brw_rt_ray_queries_hw_stacks_size(const struct intel_device_info *devinfo)
257 {
258    /* Maximum slice/subslice/EU ID can be computed from the max_scratch_ids
259     * which includes all the threads.
260     */
261    uint32_t max_eu_id = devinfo->max_scratch_ids[MESA_SHADER_COMPUTE];
262    uint32_t max_simd_size = 16; /* Cannot run in SIMD32 with ray queries */
263    return max_eu_id * max_simd_size * BRW_RT_SIZEOF_RAY_QUERY;
264 }
265 
266 static inline uint32_t
brw_rt_ray_queries_shadow_stack_size(const struct intel_device_info * devinfo)267 brw_rt_ray_queries_shadow_stack_size(const struct intel_device_info *devinfo)
268 {
269    /* Maximum slice/subslice/EU ID can be computed from the max_scratch_ids
270     * which includes all the threads.
271     */
272    uint32_t max_eu_id = devinfo->max_scratch_ids[MESA_SHADER_COMPUTE];
273    uint32_t max_simd_size = 16; /* Cannot run in SIMD32 with ray queries */
274    return max_eu_id * max_simd_size * BRW_RT_SIZEOF_SHADOW_RAY_QUERY;
275 }
276 
277 static inline uint32_t
brw_rt_ray_queries_shadow_stacks_size(const struct intel_device_info * devinfo,uint32_t ray_queries)278 brw_rt_ray_queries_shadow_stacks_size(const struct intel_device_info *devinfo,
279                                       uint32_t ray_queries)
280 {
281    /* Don't bother a shadow stack if we only have a single query. We can
282     * directly write in the HW buffer.
283     */
284    return (ray_queries > 1 ? ray_queries : 0) * brw_rt_ray_queries_shadow_stack_size(devinfo) +
285           ray_queries * 4; /* Ctrl + Level data */
286 }
287 
288 #ifdef __cplusplus
289 }
290 #endif
291 
292 #endif /* BRW_RT_H */
293