2
* Copyright © Microsoft Corporation
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:
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
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
25
#include "nir_builder.h"
26
#include "nir_format_convert.h"
28
#include "pipe/p_state.h"
29
#include "util/format/u_format.h"
31
#include "d3d12_compiler.h"
32
#include "d3d12_nir_passes.h"
35
convert_value(nir_builder *b, nir_ssa_def *value,
36
const struct util_format_description *from_desc,
37
const struct util_format_description *to_desc)
39
if (from_desc->format == to_desc->format)
42
assert(value->num_components == 4);
43
/* No support for 16 or 64 bit data in the shader for image loads/stores */
44
assert(value->bit_size == 32);
45
/* Overall format size needs to be the same */
46
assert(from_desc->block.bits == to_desc->block.bits);
47
assert(from_desc->nr_channels <= 4 && to_desc->nr_channels <= 4);
49
const unsigned rgba1010102_bits[] = { 10, 10, 10, 2 };
51
/* First, construct a "tightly packed" vector of the input values. For unorm/snorm, convert
52
* from the float we're given into the original bits (only happens while storing). For packed
53
* formats that don't fall on a nice bit size, convert/pack them into 32bit values. Otherwise,
54
* just produce a vecNx4 where N is the expected bit size.
56
nir_ssa_def *src_as_vec;
57
if (from_desc->format == PIPE_FORMAT_R10G10B10A2_UINT ||
58
from_desc->format == PIPE_FORMAT_R10G10B10A2_UNORM) {
59
if (from_desc->format == PIPE_FORMAT_R10G10B10A2_UNORM)
60
value = nir_format_float_to_unorm(b, value, rgba1010102_bits);
61
nir_ssa_def *channels[4];
62
for (unsigned i = 0; i < 4; ++i)
63
channels[i] = nir_channel(b, value, i);
65
src_as_vec = channels[0];
66
src_as_vec = nir_mask_shift_or(b, src_as_vec, channels[1], (1 << 10) - 1, 10);
67
src_as_vec = nir_mask_shift_or(b, src_as_vec, channels[2], (1 << 10) - 1, 20);
68
src_as_vec = nir_mask_shift_or(b, src_as_vec, channels[3], (1 << 2) - 1, 30);
69
} else if (from_desc->format == PIPE_FORMAT_R11G11B10_FLOAT) {
70
src_as_vec = nir_format_pack_11f11f10f(b, value);
71
} else if (from_desc->is_unorm) {
72
if (from_desc->channel[0].size == 8)
73
src_as_vec = nir_pack_unorm_4x8(b, value);
75
nir_ssa_def *packed_channels[2];
76
packed_channels[0] = nir_pack_unorm_2x16(b, nir_channels(b, value, 0x3));
77
packed_channels[1] = nir_pack_unorm_2x16(b, nir_channels(b, value, 0x3 << 2));
78
src_as_vec = nir_vec(b, packed_channels, 2);
80
} else if (from_desc->is_snorm) {
81
if (from_desc->channel[0].size == 8)
82
src_as_vec = nir_pack_snorm_4x8(b, value);
84
nir_ssa_def *packed_channels[2];
85
packed_channels[0] = nir_pack_snorm_2x16(b, nir_channels(b, value, 0x3));
86
packed_channels[1] = nir_pack_snorm_2x16(b, nir_channels(b, value, 0x3 << 2));
87
src_as_vec = nir_vec(b, packed_channels, 2);
89
} else if (util_format_is_float(from_desc->format)) {
90
src_as_vec = nir_f2fN(b, value, from_desc->channel[0].size);
91
} else if (util_format_is_pure_sint(from_desc->format)) {
92
src_as_vec = nir_i2iN(b, value, from_desc->channel[0].size);
94
src_as_vec = nir_u2uN(b, value, from_desc->channel[0].size);
97
/* Now that we have the tightly packed bits, we can use nir_extract_bits to get it into a
98
* vector of differently-sized components. For producing packed formats, get a 32-bit
99
* value and manually extract the bits. For unorm/snorm, get one or two 32-bit values,
100
* and extract it using helpers. Otherwise, get a format-sized dest vector and use a
101
* cast to expand it back to 32-bit.
103
* Pay extra attention for changing semantics for alpha as 1.
105
if (to_desc->format == PIPE_FORMAT_R10G10B10A2_UINT ||
106
to_desc->format == PIPE_FORMAT_R10G10B10A2_UNORM) {
107
nir_ssa_def *u32 = nir_extract_bits(b, &src_as_vec, 1, 0, 1, 32);
108
nir_ssa_def *channels[4] = {
109
nir_iand(b, u32, nir_imm_int(b, (1 << 10) - 1)),
110
nir_iand(b, nir_ushr(b, u32, nir_imm_int(b, 10)), nir_imm_int(b, (1 << 10) - 1)),
111
nir_iand(b, nir_ushr(b, u32, nir_imm_int(b, 20)), nir_imm_int(b, (1 << 10) - 1)),
112
nir_ushr(b, u32, nir_imm_int(b, 30))
114
nir_ssa_def *vec = nir_vec(b, channels, 4);
115
if (to_desc->format == PIPE_FORMAT_R10G10B10A2_UNORM)
116
vec = nir_format_unorm_to_float(b, vec, rgba1010102_bits);
118
} else if (to_desc->format == PIPE_FORMAT_R11G11B10_FLOAT) {
119
nir_ssa_def *u32 = nir_extract_bits(b, &src_as_vec, 1, 0, 1, 32);
120
nir_ssa_def *vec3 = nir_format_unpack_11f11f10f(b, u32);
121
return nir_vec4(b, nir_channel(b, vec3, 0),
122
nir_channel(b, vec3, 1),
123
nir_channel(b, vec3, 2),
124
nir_imm_float(b, 1.0f));
125
} else if (to_desc->is_unorm || to_desc->is_snorm) {
126
nir_ssa_def *dest_packed = nir_extract_bits(b, &src_as_vec, 1, 0,
127
DIV_ROUND_UP(to_desc->nr_channels * to_desc->channel[0].size, 32), 32);
128
if (to_desc->is_unorm) {
129
if (to_desc->channel[0].size == 8) {
130
nir_ssa_def *unpacked = nir_unpack_unorm_4x8(b, nir_channel(b, dest_packed, 0));
131
if (to_desc->nr_channels < 4)
132
unpacked = nir_vector_insert_imm(b, unpacked, nir_imm_float(b, 1.0f), 3);
135
nir_ssa_def *vec2s[2] = {
136
nir_unpack_unorm_2x16(b, nir_channel(b, dest_packed, 0)),
137
to_desc->nr_channels > 2 ?
138
nir_unpack_unorm_2x16(b, nir_channel(b, dest_packed, 1)) :
139
nir_vec2(b, nir_imm_float(b, 0.0f), nir_imm_float(b, 1.0f))
141
if (to_desc->nr_channels == 1)
142
vec2s[0] = nir_vector_insert_imm(b, vec2s[0], nir_imm_float(b, 0.0f), 1);
143
return nir_vec4(b, nir_channel(b, vec2s[0], 0),
144
nir_channel(b, vec2s[0], 1),
145
nir_channel(b, vec2s[1], 0),
146
nir_channel(b, vec2s[1], 1));
148
if (to_desc->channel[0].size == 8) {
149
nir_ssa_def *unpacked = nir_unpack_snorm_4x8(b, nir_channel(b, dest_packed, 0));
150
if (to_desc->nr_channels < 4)
151
unpacked = nir_vector_insert_imm(b, unpacked, nir_imm_float(b, 1.0f), 3);
154
nir_ssa_def *vec2s[2] = {
155
nir_unpack_snorm_2x16(b, nir_channel(b, dest_packed, 0)),
156
to_desc->nr_channels > 2 ?
157
nir_unpack_snorm_2x16(b, nir_channel(b, dest_packed, 1)) :
158
nir_vec2(b, nir_imm_float(b, 0.0f), nir_imm_float(b, 1.0f))
160
if (to_desc->nr_channels == 1)
161
vec2s[0] = nir_vector_insert_imm(b, vec2s[0], nir_imm_float(b, 0.0f), 1);
162
return nir_vec4(b, nir_channel(b, vec2s[0], 0),
163
nir_channel(b, vec2s[0], 1),
164
nir_channel(b, vec2s[1], 0),
165
nir_channel(b, vec2s[1], 1));
168
nir_ssa_def *dest_packed = nir_extract_bits(b, &src_as_vec, 1, 0,
169
to_desc->nr_channels, to_desc->channel[0].size);
170
nir_ssa_def *final_channels[4];
171
for (unsigned i = 0; i < 4; ++i) {
172
if (i >= dest_packed->num_components)
173
final_channels[i] = util_format_is_float(to_desc->format) ?
174
nir_imm_floatN_t(b, i == 3 ? 1.0f : 0.0f, to_desc->channel[0].size) :
175
nir_imm_intN_t(b, i == 3 ? 1 : 0, to_desc->channel[0].size);
177
final_channels[i] = nir_channel(b, dest_packed, i);
179
nir_ssa_def *final_vec = nir_vec(b, final_channels, 4);
180
if (util_format_is_float(to_desc->format))
181
return nir_f2f32(b, final_vec);
182
else if (util_format_is_pure_sint(to_desc->format))
183
return nir_i2i32(b, final_vec);
185
return nir_u2u32(b, final_vec);
190
lower_image_cast_instr(nir_builder *b, nir_instr *instr, void *_data)
192
if (instr->type != nir_instr_type_intrinsic)
195
nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
196
if (intr->intrinsic != nir_intrinsic_image_deref_load &&
197
intr->intrinsic != nir_intrinsic_image_deref_store)
200
const struct d3d12_image_format_conversion_info *info = _data;
201
nir_variable *image = nir_deref_instr_get_variable(nir_src_as_deref(intr->src[0]));
204
enum pipe_format emulation_format = info[image->data.driver_location].emulated_format;
205
if (emulation_format == PIPE_FORMAT_NONE)
208
enum pipe_format real_format = info[image->data.driver_location].view_format;
209
assert(real_format != emulation_format);
212
const struct util_format_description *from_desc, *to_desc;
213
if (intr->intrinsic == nir_intrinsic_image_deref_load) {
214
b->cursor = nir_after_instr(instr);
215
value = &intr->dest.ssa;
216
from_desc = util_format_description(emulation_format);
217
to_desc = util_format_description(real_format);
219
b->cursor = nir_before_instr(instr);
220
value = intr->src[3].ssa;
221
from_desc = util_format_description(real_format);
222
to_desc = util_format_description(emulation_format);
225
nir_ssa_def *new_value = convert_value(b, value, from_desc, to_desc);
227
nir_alu_type alu_type = util_format_is_pure_uint(emulation_format) ?
228
nir_type_uint : (util_format_is_pure_sint(emulation_format) ?
229
nir_type_int : nir_type_float);
231
if (intr->intrinsic == nir_intrinsic_image_deref_load) {
232
nir_ssa_def_rewrite_uses_after(value, new_value, new_value->parent_instr);
233
nir_intrinsic_set_dest_type(intr, alu_type);
235
nir_instr_rewrite_src_ssa(instr, &intr->src[3], new_value);
236
nir_intrinsic_set_src_type(intr, alu_type);
238
nir_intrinsic_set_format(intr, emulation_format);
242
/* Given a shader that does image loads/stores expecting to load from the format embedded in the intrinsic,
243
* if the corresponding entry in formats is not PIPE_FORMAT_NONE, replace the image format and convert
244
* the data being loaded/stored to/from the app's expected format.
247
d3d12_lower_image_casts(nir_shader *s, struct d3d12_image_format_conversion_info *info)
249
bool progress = nir_shader_instructions_pass(s, lower_image_cast_instr,
250
nir_metadata_block_index | nir_metadata_dominance, info);
253
nir_foreach_image_variable(var, s) {
254
if (info[var->data.driver_location].emulated_format != PIPE_FORMAT_NONE) {
255
var->data.image.format = info[var->data.driver_location].emulated_format;