1/*
2 * Copyright (C) 2019 Collabora, Ltd.
3 * Copyright (C) 2019-2020 Collabora, Ltd.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 *
24 * Authors (Collabora):
25 *   Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
26 */
27
28#include "compiler.h"
29
30/* Derivatives in Midgard are implemented on the texture pipe, rather than the
31 * ALU pipe as suggested by NIR. The rationale is that normal texture
32 * instructions require (implicit) derivatives to be calculated anyway, so it
33 * makes sense to reuse the derivative logic. Thus, in addition to the usual
34 * texturing ops that calculate derivatives, there are two explicit texture ops
35 * dFdx/dFdy that perform differencing across helper invocations in either
36 * horizontal or vertical directions.
37 *
38 * One major caveat is that derivatives can only be calculated on up to a vec2
39 * at a time. This restriction presumably is to save some silicon, as 99% of
40 * derivatives will be vec2 (autocalculating mip levels of 2D texture
41 * coordinates). Admittedly I'm not sure why 3D textures can have their levels
42 * calculated automatically, umm... Pressing on.
43 *
44 * This caveat is handled in two steps. During the first pass (code
45 * generation), we generate texture ops 1:1 to the incoming NIR derivatives.
46 * This works for float/vec2 but not for vec3/vec4. A later lowering pass will
47 * scan for vec3/vec4 derivatives and lower (split) to multiple instructions.
48 * This pass is separated as we'll have to rewrite th e destination into a
49 * register (rather than SSA) and we'd rather do this after we have the whole
50 * IR in front of us to do it at once.
51 */
52
53static unsigned
54mir_derivative_mode(nir_op op)
55{
56        switch (op) {
57        case nir_op_fddx:
58        case nir_op_fddx_fine:
59        case nir_op_fddx_coarse:
60                return TEXTURE_DFDX;
61
62        case nir_op_fddy:
63        case nir_op_fddy_fine:
64        case nir_op_fddy_coarse:
65                return TEXTURE_DFDY;
66
67        default:
68                unreachable("Invalid derivative op");
69        }
70}
71
72/* Returns true if a texturing op computes derivatives either explicitly or
73 * implicitly */
74
75bool
76mir_op_computes_derivatives(gl_shader_stage stage, unsigned op)
77{
78        /* Only fragment shaders may compute derivatives, but the sense of
79         * "normal" changes in vertex shaders on certain GPUs */
80
81        if (op == midgard_tex_op_normal && stage != MESA_SHADER_FRAGMENT)
82                return false;
83
84        switch (op) {
85        case midgard_tex_op_normal:
86        case midgard_tex_op_derivative:
87                assert(stage == MESA_SHADER_FRAGMENT);
88                return true;
89        default:
90                return false;
91        }
92}
93
94void
95midgard_emit_derivatives(compiler_context *ctx, nir_alu_instr *instr)
96{
97        /* Create texture instructions */
98
99        unsigned nr_components = nir_dest_num_components(instr->dest.dest);
100
101        midgard_instruction ins = {
102                .type = TAG_TEXTURE_4,
103                .mask = mask_of(nr_components),
104                .dest = nir_dest_index(&instr->dest.dest),
105                .dest_type = nir_type_float32,
106                .src = { ~0, nir_src_index(ctx, &instr->src[0].src), ~0, ~0 },
107                .swizzle = SWIZZLE_IDENTITY_4,
108                .src_types = { nir_type_float32, nir_type_float32 },
109                .op = midgard_tex_op_derivative,
110                .texture = {
111                        .mode = mir_derivative_mode(instr->op),
112                        .format = 2,
113                        .in_reg_full = 1,
114                        .out_full = 1,
115                        .sampler_type = MALI_SAMPLER_FLOAT,
116                }
117        };
118
119        if (!instr->dest.dest.is_ssa)
120                ins.mask &= instr->dest.write_mask;
121
122        emit_mir_instruction(ctx, ins);
123}
124
125void
126midgard_lower_derivatives(compiler_context *ctx, midgard_block *block)
127{
128        mir_foreach_instr_in_block_safe(block, ins) {
129                if (ins->type != TAG_TEXTURE_4) continue;
130                if (ins->op != midgard_tex_op_derivative) continue;
131
132                /* Check if we need to split */
133
134                bool upper = ins->mask & 0b1100;
135                bool lower = ins->mask & 0b0011;
136
137                if (!(upper && lower)) continue;
138
139                /* Duplicate for dedicated upper instruction */
140
141                midgard_instruction dup;
142                memcpy(&dup, ins, sizeof(dup));
143
144                /* Fixup masks. Make original just lower and dupe just upper */
145
146                ins->mask &= 0b0011;
147                dup.mask &= 0b1100;
148
149                /* Fixup swizzles */
150                dup.swizzle[0][0] = dup.swizzle[0][1] = dup.swizzle[0][2] = COMPONENT_X;
151                dup.swizzle[0][3] = COMPONENT_Y;
152
153                dup.swizzle[1][0] = COMPONENT_Z;
154                dup.swizzle[1][1] = dup.swizzle[1][2] = dup.swizzle[1][3] = COMPONENT_W;
155
156                /* Insert the new instruction */
157                mir_insert_instruction_before(ctx, mir_next_op(ins), dup);
158
159                /* We'll need both instructions to write to the same index, so
160                 * rewrite to use a register */
161
162                unsigned new = make_compiler_temp_reg(ctx);
163                mir_rewrite_index(ctx, ins->dest, new);
164        }
165}
166