1bf215546Sopenharmony_ci/*
2bf215546Sopenharmony_ci * Copyright © 2022 Imagination Technologies Ltd.
3bf215546Sopenharmony_ci *
4bf215546Sopenharmony_ci * Permission is hereby granted, free of charge, to any person obtaining a copy
5bf215546Sopenharmony_ci * of this software and associated documentation files (the "Software"), to deal
6bf215546Sopenharmony_ci * in the Software without restriction, including without limitation the rights
7bf215546Sopenharmony_ci * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
8bf215546Sopenharmony_ci * copies of the Software, and to permit persons to whom the Software is
9bf215546Sopenharmony_ci * furnished to do so, subject to the following conditions:
10bf215546Sopenharmony_ci *
11bf215546Sopenharmony_ci * The above copyright notice and this permission notice (including the next
12bf215546Sopenharmony_ci * paragraph) shall be included in all copies or substantial portions of the
13bf215546Sopenharmony_ci * Software.
14bf215546Sopenharmony_ci *
15bf215546Sopenharmony_ci * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16bf215546Sopenharmony_ci * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17bf215546Sopenharmony_ci * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18bf215546Sopenharmony_ci * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19bf215546Sopenharmony_ci * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20bf215546Sopenharmony_ci * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21bf215546Sopenharmony_ci * SOFTWARE.
22bf215546Sopenharmony_ci */
23bf215546Sopenharmony_ci
24bf215546Sopenharmony_ci#include "compiler/spirv/nir_spirv.h"
25bf215546Sopenharmony_ci#include "nir/nir.h"
26bf215546Sopenharmony_ci#include "nir/nir_schedule.h"
27bf215546Sopenharmony_ci#include "rogue_nir.h"
28bf215546Sopenharmony_ci#include "rogue_operand.h"
29bf215546Sopenharmony_ci
30bf215546Sopenharmony_ci/**
31bf215546Sopenharmony_ci * \file rogue_nir.c
32bf215546Sopenharmony_ci *
33bf215546Sopenharmony_ci * \brief Contains NIR-specific functions.
34bf215546Sopenharmony_ci */
35bf215546Sopenharmony_ci
36bf215546Sopenharmony_ci/**
37bf215546Sopenharmony_ci * \brief SPIR-V to NIR compilation options.
38bf215546Sopenharmony_ci */
39bf215546Sopenharmony_cistatic const struct spirv_to_nir_options spirv_options = {
40bf215546Sopenharmony_ci   .environment = NIR_SPIRV_VULKAN,
41bf215546Sopenharmony_ci
42bf215546Sopenharmony_ci   /* Buffer address: (descriptor_set, binding), offset. */
43bf215546Sopenharmony_ci   .ubo_addr_format = nir_address_format_vec2_index_32bit_offset,
44bf215546Sopenharmony_ci};
45bf215546Sopenharmony_ci
46bf215546Sopenharmony_cistatic const nir_shader_compiler_options nir_options = {
47bf215546Sopenharmony_ci   .lower_fsat = true,
48bf215546Sopenharmony_ci   .fuse_ffma32 = true,
49bf215546Sopenharmony_ci};
50bf215546Sopenharmony_ci
51bf215546Sopenharmony_ciconst struct spirv_to_nir_options *
52bf215546Sopenharmony_cirogue_get_spirv_options(const struct rogue_compiler *compiler)
53bf215546Sopenharmony_ci{
54bf215546Sopenharmony_ci   return &spirv_options;
55bf215546Sopenharmony_ci}
56bf215546Sopenharmony_ci
57bf215546Sopenharmony_ciconst nir_shader_compiler_options *
58bf215546Sopenharmony_cirogue_get_compiler_options(const struct rogue_compiler *compiler)
59bf215546Sopenharmony_ci{
60bf215546Sopenharmony_ci   return &nir_options;
61bf215546Sopenharmony_ci}
62bf215546Sopenharmony_ci
63bf215546Sopenharmony_cistatic int rogue_glsl_type_size(const struct glsl_type *type, bool bindless)
64bf215546Sopenharmony_ci{
65bf215546Sopenharmony_ci   return glsl_count_attribute_slots(type, false);
66bf215546Sopenharmony_ci}
67bf215546Sopenharmony_ci
68bf215546Sopenharmony_ci/**
69bf215546Sopenharmony_ci * \brief Applies optimizations and passes required to lower the NIR shader into
70bf215546Sopenharmony_ci * a form suitable for lowering to Rogue IR.
71bf215546Sopenharmony_ci *
72bf215546Sopenharmony_ci * \param[in] ctx Shared multi-stage build context.
73bf215546Sopenharmony_ci * \param[in] shader Rogue shader.
74bf215546Sopenharmony_ci * \param[in] stage Shader stage.
75bf215546Sopenharmony_ci * \return true if successful, otherwise false.
76bf215546Sopenharmony_ci */
77bf215546Sopenharmony_cibool rogue_nir_passes(struct rogue_build_ctx *ctx,
78bf215546Sopenharmony_ci                      nir_shader *nir,
79bf215546Sopenharmony_ci                      gl_shader_stage stage)
80bf215546Sopenharmony_ci{
81bf215546Sopenharmony_ci   bool progress;
82bf215546Sopenharmony_ci
83bf215546Sopenharmony_ci   nir_validate_shader(nir, "after spirv_to_nir");
84bf215546Sopenharmony_ci
85bf215546Sopenharmony_ci   /* Splitting. */
86bf215546Sopenharmony_ci   NIR_PASS_V(nir, nir_split_var_copies);
87bf215546Sopenharmony_ci   NIR_PASS_V(nir, nir_split_per_member_structs);
88bf215546Sopenharmony_ci
89bf215546Sopenharmony_ci   /* Ensure fs outputs are in the [0.0f...1.0f] range. */
90bf215546Sopenharmony_ci   NIR_PASS_V(nir, nir_lower_clamp_color_outputs);
91bf215546Sopenharmony_ci
92bf215546Sopenharmony_ci   /* Replace references to I/O variables with intrinsics. */
93bf215546Sopenharmony_ci   NIR_PASS_V(nir,
94bf215546Sopenharmony_ci              nir_lower_io,
95bf215546Sopenharmony_ci              nir_var_shader_in | nir_var_shader_out,
96bf215546Sopenharmony_ci              rogue_glsl_type_size,
97bf215546Sopenharmony_ci              (nir_lower_io_options)0);
98bf215546Sopenharmony_ci
99bf215546Sopenharmony_ci   /* Load inputs to scalars (single registers later). */
100bf215546Sopenharmony_ci   NIR_PASS_V(nir, nir_lower_io_to_scalar, nir_var_shader_in);
101bf215546Sopenharmony_ci
102bf215546Sopenharmony_ci   /* Optimize GL access qualifiers. */
103bf215546Sopenharmony_ci   const nir_opt_access_options opt_access_options = {
104bf215546Sopenharmony_ci      .is_vulkan = true,
105bf215546Sopenharmony_ci      .infer_non_readable = true,
106bf215546Sopenharmony_ci   };
107bf215546Sopenharmony_ci   NIR_PASS_V(nir, nir_opt_access, &opt_access_options);
108bf215546Sopenharmony_ci
109bf215546Sopenharmony_ci   /* Apply PFO code to the fragment shader output. */
110bf215546Sopenharmony_ci   if (nir->info.stage == MESA_SHADER_FRAGMENT)
111bf215546Sopenharmony_ci      NIR_PASS_V(nir, rogue_nir_pfo);
112bf215546Sopenharmony_ci
113bf215546Sopenharmony_ci   /* Load outputs to scalars (single registers later). */
114bf215546Sopenharmony_ci   NIR_PASS_V(nir, nir_lower_io_to_scalar, nir_var_shader_out);
115bf215546Sopenharmony_ci
116bf215546Sopenharmony_ci   /* Lower ALU operations to scalars. */
117bf215546Sopenharmony_ci   NIR_PASS_V(nir, nir_lower_alu_to_scalar, NULL, NULL);
118bf215546Sopenharmony_ci
119bf215546Sopenharmony_ci   /* Algebraic opts. */
120bf215546Sopenharmony_ci   do {
121bf215546Sopenharmony_ci      progress = false;
122bf215546Sopenharmony_ci
123bf215546Sopenharmony_ci      NIR_PASS(progress, nir, nir_copy_prop);
124bf215546Sopenharmony_ci      NIR_PASS(progress, nir, nir_opt_cse);
125bf215546Sopenharmony_ci      NIR_PASS(progress, nir, nir_opt_algebraic);
126bf215546Sopenharmony_ci      NIR_PASS(progress, nir, nir_opt_constant_folding);
127bf215546Sopenharmony_ci      NIR_PASS(progress, nir, nir_opt_dce);
128bf215546Sopenharmony_ci      NIR_PASS_V(nir, nir_opt_gcm, false);
129bf215546Sopenharmony_ci   } while (progress);
130bf215546Sopenharmony_ci
131bf215546Sopenharmony_ci   /* Additional I/O lowering. */
132bf215546Sopenharmony_ci   NIR_PASS_V(nir,
133bf215546Sopenharmony_ci              nir_lower_explicit_io,
134bf215546Sopenharmony_ci              nir_var_mem_ubo,
135bf215546Sopenharmony_ci              spirv_options.ubo_addr_format);
136bf215546Sopenharmony_ci   NIR_PASS_V(nir, rogue_nir_lower_io, NULL);
137bf215546Sopenharmony_ci
138bf215546Sopenharmony_ci   /* Late algebraic opts. */
139bf215546Sopenharmony_ci   do {
140bf215546Sopenharmony_ci      progress = false;
141bf215546Sopenharmony_ci
142bf215546Sopenharmony_ci      NIR_PASS(progress, nir, nir_opt_algebraic_late);
143bf215546Sopenharmony_ci      NIR_PASS_V(nir, nir_opt_constant_folding);
144bf215546Sopenharmony_ci      NIR_PASS_V(nir, nir_copy_prop);
145bf215546Sopenharmony_ci      NIR_PASS_V(nir, nir_opt_dce);
146bf215546Sopenharmony_ci      NIR_PASS_V(nir, nir_opt_cse);
147bf215546Sopenharmony_ci   } while (progress);
148bf215546Sopenharmony_ci
149bf215546Sopenharmony_ci   /* Replace SSA constant references with a register that loads the value. */
150bf215546Sopenharmony_ci   NIR_PASS_V(nir, rogue_nir_constreg);
151bf215546Sopenharmony_ci   /* Remove unused constant registers. */
152bf215546Sopenharmony_ci   NIR_PASS_V(nir, nir_opt_dce);
153bf215546Sopenharmony_ci
154bf215546Sopenharmony_ci   /* Move loads to just before they're needed. */
155bf215546Sopenharmony_ci   NIR_PASS_V(nir, nir_opt_move, nir_move_load_ubo | nir_move_load_input);
156bf215546Sopenharmony_ci
157bf215546Sopenharmony_ci   /* Convert vecNs to movs so we can sequentially allocate them later. */
158bf215546Sopenharmony_ci   NIR_PASS_V(nir, nir_lower_vec_to_movs, NULL, NULL);
159bf215546Sopenharmony_ci
160bf215546Sopenharmony_ci   /* Out of SSA pass. */
161bf215546Sopenharmony_ci   NIR_PASS_V(nir, nir_convert_from_ssa, false);
162bf215546Sopenharmony_ci
163bf215546Sopenharmony_ci   /* TODO: Re-enable scheduling after register pressure tweaks. */
164bf215546Sopenharmony_ci#if 0
165bf215546Sopenharmony_ci	/* Instruction scheduling. */
166bf215546Sopenharmony_ci	struct nir_schedule_options schedule_options = {
167bf215546Sopenharmony_ci		.threshold = ROGUE_MAX_REG_TEMP / 2,
168bf215546Sopenharmony_ci	};
169bf215546Sopenharmony_ci	NIR_PASS_V(nir, nir_schedule, &schedule_options);
170bf215546Sopenharmony_ci#endif
171bf215546Sopenharmony_ci
172bf215546Sopenharmony_ci   /* Assign I/O locations. */
173bf215546Sopenharmony_ci   nir_assign_io_var_locations(nir,
174bf215546Sopenharmony_ci                               nir_var_shader_in,
175bf215546Sopenharmony_ci                               &nir->num_inputs,
176bf215546Sopenharmony_ci                               nir->info.stage);
177bf215546Sopenharmony_ci   nir_assign_io_var_locations(nir,
178bf215546Sopenharmony_ci                               nir_var_shader_out,
179bf215546Sopenharmony_ci                               &nir->num_outputs,
180bf215546Sopenharmony_ci                               nir->info.stage);
181bf215546Sopenharmony_ci
182bf215546Sopenharmony_ci   /* Gather info into nir shader struct. */
183bf215546Sopenharmony_ci   nir_shader_gather_info(nir, nir_shader_get_entrypoint(nir));
184bf215546Sopenharmony_ci
185bf215546Sopenharmony_ci   /* Clean-up after passes. */
186bf215546Sopenharmony_ci   nir_sweep(nir);
187bf215546Sopenharmony_ci
188bf215546Sopenharmony_ci   nir_validate_shader(nir, "after passes");
189bf215546Sopenharmony_ci
190bf215546Sopenharmony_ci   return true;
191bf215546Sopenharmony_ci}
192