1 /*
2 * Copyright © 2018 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 "brw_nir.h"
25 #include "compiler/nir/nir_builder.h"
26 #include "util/u_math.h"
27 #include "util/bitscan.h"
28
29 static nir_ssa_def *
dup_mem_intrinsic(nir_builder * b,nir_intrinsic_instr * intrin,nir_ssa_def * store_src,int offset,unsigned num_components,unsigned bit_size,unsigned align)30 dup_mem_intrinsic(nir_builder *b, nir_intrinsic_instr *intrin,
31 nir_ssa_def *store_src, int offset,
32 unsigned num_components, unsigned bit_size,
33 unsigned align)
34 {
35 const nir_intrinsic_info *info = &nir_intrinsic_infos[intrin->intrinsic];
36
37 nir_intrinsic_instr *dup =
38 nir_intrinsic_instr_create(b->shader, intrin->intrinsic);
39
40 nir_src *intrin_offset_src = nir_get_io_offset_src(intrin);
41 for (unsigned i = 0; i < info->num_srcs; i++) {
42 assert(intrin->src[i].is_ssa);
43 if (i == 0 && store_src) {
44 assert(!info->has_dest);
45 assert(&intrin->src[i] != intrin_offset_src);
46 dup->src[i] = nir_src_for_ssa(store_src);
47 } else if (&intrin->src[i] == intrin_offset_src) {
48 dup->src[i] = nir_src_for_ssa(nir_iadd_imm(b, intrin->src[i].ssa,
49 offset));
50 } else {
51 dup->src[i] = nir_src_for_ssa(intrin->src[i].ssa);
52 }
53 }
54
55 dup->num_components = num_components;
56 if (intrin->intrinsic == nir_intrinsic_load_scratch ||
57 intrin->intrinsic == nir_intrinsic_store_scratch)
58 assert(num_components == 1);
59
60 for (unsigned i = 0; i < info->num_indices; i++)
61 dup->const_index[i] = intrin->const_index[i];
62
63 if (nir_intrinsic_has_access(intrin))
64 nir_intrinsic_set_access(dup, nir_intrinsic_access(intrin));
65
66 nir_intrinsic_set_align(dup, align, 0);
67
68 if (info->has_dest) {
69 assert(intrin->dest.is_ssa);
70 nir_ssa_dest_init(&dup->instr, &dup->dest,
71 num_components, bit_size, NULL);
72 } else {
73 nir_intrinsic_set_write_mask(dup, (1 << num_components) - 1);
74 }
75
76 nir_builder_instr_insert(b, &dup->instr);
77
78 return info->has_dest ? &dup->dest.ssa : NULL;
79 }
80
81 static bool
lower_mem_load_bit_size(nir_builder * b,nir_intrinsic_instr * intrin,const struct intel_device_info * devinfo)82 lower_mem_load_bit_size(nir_builder *b, nir_intrinsic_instr *intrin,
83 const struct intel_device_info *devinfo)
84 {
85 const bool needs_scalar =
86 intrin->intrinsic == nir_intrinsic_load_scratch;
87
88 assert(intrin->dest.is_ssa);
89 const unsigned bit_size = intrin->dest.ssa.bit_size;
90 const unsigned num_components = intrin->dest.ssa.num_components;
91 const unsigned bytes_read = num_components * (bit_size / 8);
92 const unsigned align = nir_intrinsic_align(intrin);
93
94 if (bit_size == 32 && align >= 32 && intrin->num_components <= 4 &&
95 (!needs_scalar || intrin->num_components == 1))
96 return false;
97
98 nir_ssa_def *result;
99 nir_src *offset_src = nir_get_io_offset_src(intrin);
100 if (bit_size < 32 && !needs_scalar && nir_src_is_const(*offset_src)) {
101 /* The offset is constant so we can use a 32-bit load and just shift it
102 * around as needed.
103 */
104 const int load_offset = nir_src_as_uint(*offset_src) % 4;
105 assert(load_offset % (bit_size / 8) == 0);
106 const unsigned load_comps32 = DIV_ROUND_UP(bytes_read + load_offset, 4);
107 /* A 16-bit vec4 is a 32-bit vec2. We add an extra component in case
108 * we offset into a component with load_offset.
109 */
110 assert(load_comps32 <= 3);
111
112 nir_ssa_def *load = dup_mem_intrinsic(b, intrin, NULL, -load_offset,
113 load_comps32, 32, 4);
114 result = nir_extract_bits(b, &load, 1, load_offset * 8,
115 num_components, bit_size);
116 } else {
117 /* Otherwise, we have to break it into smaller loads. We could end up
118 * with as many as 32 loads if we're loading a u64vec16 from scratch.
119 */
120 nir_ssa_def *loads[32];
121 unsigned num_loads = 0;
122 int load_offset = 0;
123 while (load_offset < bytes_read) {
124 const unsigned bytes_left = bytes_read - load_offset;
125 unsigned load_bit_size, load_comps;
126 if (align < 4) {
127 load_comps = 1;
128 /* Choose a byte, word, or dword */
129 load_bit_size = util_next_power_of_two(MIN2(bytes_left, 4)) * 8;
130 } else {
131 assert(load_offset % 4 == 0);
132 load_bit_size = 32;
133 load_comps = needs_scalar ? 1 :
134 DIV_ROUND_UP(MIN2(bytes_left, 16), 4);
135 }
136
137 loads[num_loads++] = dup_mem_intrinsic(b, intrin, NULL, load_offset,
138 load_comps, load_bit_size,
139 align);
140
141 load_offset += load_comps * (load_bit_size / 8);
142 }
143 assert(num_loads <= ARRAY_SIZE(loads));
144 result = nir_extract_bits(b, loads, num_loads, 0,
145 num_components, bit_size);
146 }
147
148 nir_ssa_def_rewrite_uses(&intrin->dest.ssa,
149 result);
150 nir_instr_remove(&intrin->instr);
151
152 return true;
153 }
154
155 static bool
lower_mem_store_bit_size(nir_builder * b,nir_intrinsic_instr * intrin,const struct intel_device_info * devinfo)156 lower_mem_store_bit_size(nir_builder *b, nir_intrinsic_instr *intrin,
157 const struct intel_device_info *devinfo)
158 {
159 const bool needs_scalar =
160 intrin->intrinsic == nir_intrinsic_store_scratch;
161
162 assert(intrin->src[0].is_ssa);
163 nir_ssa_def *value = intrin->src[0].ssa;
164
165 assert(intrin->num_components == value->num_components);
166 const unsigned bit_size = value->bit_size;
167 const unsigned num_components = intrin->num_components;
168 const unsigned bytes_written = num_components * (bit_size / 8);
169 const unsigned align_mul = nir_intrinsic_align_mul(intrin);
170 const unsigned align_offset = nir_intrinsic_align_offset(intrin);
171 const unsigned align = nir_intrinsic_align(intrin);
172
173 nir_component_mask_t writemask = nir_intrinsic_write_mask(intrin);
174 assert(writemask < (1 << num_components));
175
176 if ((value->bit_size <= 32 && num_components == 1) ||
177 (value->bit_size == 32 && num_components <= 4 && align >= 32 &&
178 writemask == (1 << num_components) - 1 &&
179 !needs_scalar))
180 return false;
181
182 nir_src *offset_src = nir_get_io_offset_src(intrin);
183 const bool offset_is_const = nir_src_is_const(*offset_src);
184 const unsigned const_offset =
185 offset_is_const ? nir_src_as_uint(*offset_src) : 0;
186
187 const unsigned byte_size = bit_size / 8;
188 assert(byte_size <= sizeof(uint64_t));
189
190 BITSET_DECLARE(mask, NIR_MAX_VEC_COMPONENTS * sizeof(uint64_t));
191 BITSET_ZERO(mask);
192
193 for (unsigned i = 0; i < num_components; i++) {
194 if (writemask & (1u << i))
195 BITSET_SET_RANGE_INSIDE_WORD(mask, i * byte_size, ((i + 1) * byte_size) - 1);
196 }
197
198 while (BITSET_FFS(mask) != 0) {
199 const int start = BITSET_FFS(mask) - 1;
200
201 int end;
202 for (end = start + 1; end < bytes_written; end++) {
203 if (!(BITSET_TEST(mask, end)))
204 break;
205 }
206 /* The size of the current contiguous chunk in bytes */
207 const unsigned chunk_bytes = end - start;
208
209 const bool is_dword_aligned =
210 (align_mul >= 4 && (align_offset + start) % 4 == 0) ||
211 (offset_is_const && (start + const_offset) % 4 == 0);
212
213 unsigned store_comps, store_bit_size, store_align;
214 if (chunk_bytes >= 4 && is_dword_aligned) {
215 store_align = MAX2(align, 4);
216 store_bit_size = 32;
217 store_comps = needs_scalar ? 1 : MIN2(chunk_bytes, 16) / 4;
218 } else {
219 store_align = align;
220 store_comps = 1;
221 store_bit_size = MIN2(chunk_bytes, 4) * 8;
222 /* The bit size must be a power of two */
223 if (store_bit_size == 24)
224 store_bit_size = 16;
225 }
226 const unsigned store_bytes = store_comps * (store_bit_size / 8);
227
228 nir_ssa_def *packed = nir_extract_bits(b, &value, 1, start * 8,
229 store_comps, store_bit_size);
230
231 dup_mem_intrinsic(b, intrin, packed, start,
232 store_comps, store_bit_size, store_align);
233
234 BITSET_CLEAR_RANGE(mask, start, (start + store_bytes - 1));
235 }
236
237 nir_instr_remove(&intrin->instr);
238
239 return true;
240 }
241
242 static bool
lower_mem_access_bit_sizes_instr(nir_builder * b,nir_instr * instr,void * cb_data)243 lower_mem_access_bit_sizes_instr(nir_builder *b,
244 nir_instr *instr,
245 void *cb_data)
246 {
247 const struct intel_device_info *devinfo = cb_data;
248
249 if (instr->type != nir_instr_type_intrinsic)
250 return false;
251
252 b->cursor = nir_after_instr(instr);
253
254 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
255 switch (intrin->intrinsic) {
256 case nir_intrinsic_load_global:
257 case nir_intrinsic_load_global_constant:
258 case nir_intrinsic_load_ssbo:
259 case nir_intrinsic_load_shared:
260 case nir_intrinsic_load_scratch:
261 return lower_mem_load_bit_size(b, intrin, devinfo);
262
263 case nir_intrinsic_store_global:
264 case nir_intrinsic_store_ssbo:
265 case nir_intrinsic_store_shared:
266 case nir_intrinsic_store_scratch:
267 return lower_mem_store_bit_size(b, intrin, devinfo);
268
269 default:
270 return false;
271 }
272 }
273
274 /**
275 * This pass loads arbitrary SSBO and shared memory load/store operations to
276 * intrinsics which are natively handleable by GEN hardware. In particular,
277 * we have two general types of memory load/store messages:
278 *
279 * - Untyped surface read/write: These can load/store between one and four
280 * dword components to/from a dword-aligned offset.
281 *
282 * - Byte scattered read/write: These can load/store a single byte, word, or
283 * dword scalar to/from an unaligned byte offset.
284 *
285 * Neither type of message can do a write-masked store. This pass converts
286 * all nir load/store intrinsics into a series of either 8 or 32-bit
287 * load/store intrinsics with a number of components that we can directly
288 * handle in hardware and with a trivial write-mask.
289 *
290 * For scratch access, additional consideration has to be made due to the way
291 * that we swizzle the memory addresses to achieve decent cache locality. In
292 * particular, even though untyped surface read/write messages exist and work,
293 * we can't use them to load multiple components in a single SEND. For more
294 * detail on the scratch swizzle, see fs_visitor::swizzle_nir_scratch_addr.
295 */
296 bool
brw_nir_lower_mem_access_bit_sizes(nir_shader * shader,const struct intel_device_info * devinfo)297 brw_nir_lower_mem_access_bit_sizes(nir_shader *shader,
298 const struct intel_device_info *devinfo)
299 {
300 return nir_shader_instructions_pass(shader, lower_mem_access_bit_sizes_instr,
301 nir_metadata_block_index |
302 nir_metadata_dominance,
303 (void *)devinfo);
304 }
305