1 /*
2 * Copyright (C) 2021 Collabora, 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 FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21 * SOFTWARE.
22 */
23
24 #include "bi_builder.h"
25 #include "compiler.h"
26
27 #define XXH_INLINE_ALL
28 #include "util/xxhash.h"
29
30 /* Fuse pairs of TEXS_2D instructions into a single dual texture TEXC, if both
31 * sample at the same coordinate with the default LOD mode for the shader stage
32 * (computed LOD in fragment shaders, else zero LOD) and immediate
33 * texture/sampler indices 0...3.
34 *
35 * Fusing across basic block boundaries is not expected to be useful, as it
36 * increases register pressure and causes redundant memory traffic. As such, we
37 * use a local optimization pass.
38 *
39 * To pair ops efficiently, we maintain a set (backed by a hash table) using
40 * only the coordinate sources for hashing and equality. Hence the pass runs in
41 * O(n) worst case expected time for n insturctions in a block. We reject
42 * invalid texture instructions quickly to reduce the constant factor.
43 *
44 * Dual texture instructions have skip flags, like normal texture instructions.
45 * Adding a skip flag to an instruction that doesn't have it is illegal, but
46 * removing a skip flag from one that has it is legal. Accordingly, set the
47 * fused TEXC's skip to the logical AND of the unfused TEXS flags. We run the
48 * optimization pass to run after bi_analyze_helper_requirements.
49 */
50
51 static inline bool
bi_can_fuse_dual_tex(bi_instr * I,bool fuse_zero_lod)52 bi_can_fuse_dual_tex(bi_instr *I, bool fuse_zero_lod)
53 {
54 return (I->op == BI_OPCODE_TEXS_2D_F32 || I->op == BI_OPCODE_TEXS_2D_F16) &&
55 (I->texture_index < 4 && I->sampler_index < 4) &&
56 (I->lod_mode == fuse_zero_lod);
57 }
58
59 static enum bifrost_texture_format
bi_format_for_texs_2d(enum bi_opcode op)60 bi_format_for_texs_2d(enum bi_opcode op)
61 {
62 switch (op) {
63 case BI_OPCODE_TEXS_2D_F32:
64 return BIFROST_TEXTURE_FORMAT_F32;
65 case BI_OPCODE_TEXS_2D_F16:
66 return BIFROST_TEXTURE_FORMAT_F16;
67 default:
68 unreachable("Invalid TEXS_2D instruction");
69 }
70 }
71
72 static void
bi_fuse_dual(bi_context * ctx,bi_instr * I1,bi_instr * I2)73 bi_fuse_dual(bi_context *ctx, bi_instr *I1, bi_instr *I2)
74 {
75 /* Construct a texture operation descriptor for the dual texture */
76 struct bifrost_dual_texture_operation desc = {
77 .mode = BIFROST_TEXTURE_OPERATION_DUAL,
78
79 .primary_texture_index = I1->texture_index,
80 .primary_sampler_index = I1->sampler_index,
81 .primary_format = bi_format_for_texs_2d(I1->op),
82 .primary_mask = 0xF,
83
84 .secondary_texture_index = I2->texture_index,
85 .secondary_sampler_index = I2->sampler_index,
86 .secondary_format = bi_format_for_texs_2d(I2->op),
87 .secondary_mask = 0xF,
88 };
89
90 /* LOD mode is implied in a shader stage */
91 assert(I1->lod_mode == I2->lod_mode);
92
93 /* Insert before the earlier instruction in case its result is consumed
94 * before the later instruction
95 */
96 bi_builder b = bi_init_builder(ctx, bi_before_instr(I1));
97
98 bi_instr *I = bi_texc_dual_to(
99 &b, I1->dest[0], I2->dest[0], bi_null(), /* staging */
100 I1->src[0], I1->src[1], /* coordinates */
101 bi_imm_u32(bi_dual_tex_as_u32(desc)), I1->lod_mode,
102 bi_count_write_registers(I1, 0), bi_count_write_registers(I2, 0));
103
104 I->skip = I1->skip && I2->skip;
105
106 bi_remove_instruction(I1);
107 bi_remove_instruction(I2);
108 }
109
110 #define HASH(hash, data) XXH32(&(data), sizeof(data), hash)
111
112 static uint32_t
coord_hash(const void * key)113 coord_hash(const void *key)
114 {
115 const bi_instr *I = key;
116
117 return XXH32(&I->src[0], sizeof(I->src[0]) + sizeof(I->src[1]), 0);
118 }
119
120 static bool
coord_equal(const void * key1,const void * key2)121 coord_equal(const void *key1, const void *key2)
122 {
123 const bi_instr *I = key1;
124 const bi_instr *J = key2;
125
126 return memcmp(&I->src[0], &J->src[0],
127 sizeof(I->src[0]) + sizeof(I->src[1])) == 0;
128 }
129
130 static void
bi_opt_fuse_dual_texture_block(bi_context * ctx,bi_block * block)131 bi_opt_fuse_dual_texture_block(bi_context *ctx, bi_block *block)
132 {
133 struct set *set = _mesa_set_create(ctx, coord_hash, coord_equal);
134 bool fuse_zero_lod = (ctx->stage != MESA_SHADER_FRAGMENT);
135 bool found = false;
136
137 bi_foreach_instr_in_block_safe(block, I) {
138 if (!bi_can_fuse_dual_tex(I, fuse_zero_lod))
139 continue;
140
141 struct set_entry *ent = _mesa_set_search_or_add(set, I, &found);
142
143 if (found) {
144 bi_fuse_dual(ctx, (bi_instr *)ent->key, I);
145 _mesa_set_remove(set, ent);
146 }
147 }
148 }
149
150 void
bi_opt_fuse_dual_texture(bi_context * ctx)151 bi_opt_fuse_dual_texture(bi_context *ctx)
152 {
153 bi_foreach_block(ctx, block) {
154 bi_opt_fuse_dual_texture_block(ctx, block);
155 }
156 }
157