1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
|
// SPDX-FileCopyrightText: Copyright 2021 yuzu Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later
#include "shader_recompiler/backend/spirv/emit_spirv_instructions.h"
#include "shader_recompiler/backend/spirv/spirv_emit_context.h"
namespace Shader::Backend::SPIRV {
namespace {
Id SubgroupScope(EmitContext& ctx) {
return ctx.Const(static_cast<u32>(spv::Scope::Subgroup));
}
Id GetThreadId(EmitContext& ctx) {
return ctx.OpLoad(ctx.U32[1], ctx.subgroup_local_invocation_id);
}
Id WarpExtract(EmitContext& ctx, Id value) {
const Id thread_id{GetThreadId(ctx)};
const Id local_index{ctx.OpShiftRightArithmetic(ctx.U32[1], thread_id, ctx.Const(5U))};
return ctx.OpVectorExtractDynamic(ctx.U32[1], value, local_index);
}
Id LoadMask(EmitContext& ctx, Id mask) {
const Id value{ctx.OpLoad(ctx.U32[4], mask)};
if (!ctx.profile.warp_size_potentially_larger_than_guest) {
return ctx.OpCompositeExtract(ctx.U32[1], value, 0U);
}
return WarpExtract(ctx, value);
}
void SetInBoundsFlag(IR::Inst* inst, Id result) {
IR::Inst* const in_bounds{inst->GetAssociatedPseudoOperation(IR::Opcode::GetInBoundsFromOp)};
if (!in_bounds) {
return;
}
in_bounds->SetDefinition(result);
in_bounds->Invalidate();
}
Id ComputeMinThreadId(EmitContext& ctx, Id thread_id, Id segmentation_mask) {
return ctx.OpBitwiseAnd(ctx.U32[1], thread_id, segmentation_mask);
}
Id ComputeMaxThreadId(EmitContext& ctx, Id min_thread_id, Id clamp, Id not_seg_mask) {
return ctx.OpBitwiseOr(ctx.U32[1], min_thread_id,
ctx.OpBitwiseAnd(ctx.U32[1], clamp, not_seg_mask));
}
Id GetMaxThreadId(EmitContext& ctx, Id thread_id, Id clamp, Id segmentation_mask) {
const Id not_seg_mask{ctx.OpNot(ctx.U32[1], segmentation_mask)};
const Id min_thread_id{ComputeMinThreadId(ctx, thread_id, segmentation_mask)};
return ComputeMaxThreadId(ctx, min_thread_id, clamp, not_seg_mask);
}
Id SelectValue(EmitContext& ctx, Id in_range, Id value, Id src_thread_id) {
return ctx.OpSelect(
ctx.U32[1], in_range,
ctx.OpGroupNonUniformShuffle(ctx.U32[1], SubgroupScope(ctx), value, src_thread_id), value);
}
Id GetUpperClamp(EmitContext& ctx, Id invocation_id, Id clamp) {
const Id thirty_two{ctx.Const(32u)};
const Id is_upper_partition{ctx.OpSGreaterThanEqual(ctx.U1, invocation_id, thirty_two)};
const Id upper_clamp{ctx.OpIAdd(ctx.U32[1], thirty_two, clamp)};
return ctx.OpSelect(ctx.U32[1], is_upper_partition, upper_clamp, clamp);
}
} // Anonymous namespace
Id EmitLaneId(EmitContext& ctx) {
const Id id{GetThreadId(ctx)};
if (!ctx.profile.warp_size_potentially_larger_than_guest) {
return id;
}
return ctx.OpBitwiseAnd(ctx.U32[1], id, ctx.Const(31U));
}
Id EmitVoteAll(EmitContext& ctx, Id pred) {
if (!ctx.profile.warp_size_potentially_larger_than_guest) {
return ctx.OpGroupNonUniformAll(ctx.U1, SubgroupScope(ctx), pred);
}
const Id mask_ballot{
ctx.OpGroupNonUniformBallot(ctx.U32[4], SubgroupScope(ctx), ctx.true_value)};
const Id active_mask{WarpExtract(ctx, mask_ballot)};
const Id ballot{
WarpExtract(ctx, ctx.OpGroupNonUniformBallot(ctx.U32[4], SubgroupScope(ctx), pred))};
const Id lhs{ctx.OpBitwiseAnd(ctx.U32[1], ballot, active_mask)};
return ctx.OpIEqual(ctx.U1, lhs, active_mask);
}
Id EmitVoteAny(EmitContext& ctx, Id pred) {
if (!ctx.profile.warp_size_potentially_larger_than_guest) {
return ctx.OpGroupNonUniformAny(ctx.U1, SubgroupScope(ctx), pred);
}
const Id mask_ballot{
ctx.OpGroupNonUniformBallot(ctx.U32[4], SubgroupScope(ctx), ctx.true_value)};
const Id active_mask{WarpExtract(ctx, mask_ballot)};
const Id ballot{
WarpExtract(ctx, ctx.OpGroupNonUniformBallot(ctx.U32[4], SubgroupScope(ctx), pred))};
const Id lhs{ctx.OpBitwiseAnd(ctx.U32[1], ballot, active_mask)};
return ctx.OpINotEqual(ctx.U1, lhs, ctx.u32_zero_value);
}
Id EmitVoteEqual(EmitContext& ctx, Id pred) {
if (!ctx.profile.warp_size_potentially_larger_than_guest) {
return ctx.OpGroupNonUniformAllEqual(ctx.U1, SubgroupScope(ctx), pred);
}
const Id mask_ballot{
ctx.OpGroupNonUniformBallot(ctx.U32[4], SubgroupScope(ctx), ctx.true_value)};
const Id active_mask{WarpExtract(ctx, mask_ballot)};
const Id ballot{
WarpExtract(ctx, ctx.OpGroupNonUniformBallot(ctx.U32[4], SubgroupScope(ctx), pred))};
const Id lhs{ctx.OpBitwiseXor(ctx.U32[1], ballot, active_mask)};
return ctx.OpLogicalOr(ctx.U1, ctx.OpIEqual(ctx.U1, lhs, ctx.u32_zero_value),
ctx.OpIEqual(ctx.U1, lhs, active_mask));
}
Id EmitSubgroupBallot(EmitContext& ctx, Id pred) {
const Id ballot{ctx.OpGroupNonUniformBallot(ctx.U32[4], SubgroupScope(ctx), pred)};
if (!ctx.profile.warp_size_potentially_larger_than_guest) {
return ctx.OpCompositeExtract(ctx.U32[1], ballot, 0U);
}
return WarpExtract(ctx, ballot);
}
Id EmitSubgroupEqMask(EmitContext& ctx) {
return LoadMask(ctx, ctx.subgroup_mask_eq);
}
Id EmitSubgroupLtMask(EmitContext& ctx) {
return LoadMask(ctx, ctx.subgroup_mask_lt);
}
Id EmitSubgroupLeMask(EmitContext& ctx) {
return LoadMask(ctx, ctx.subgroup_mask_le);
}
Id EmitSubgroupGtMask(EmitContext& ctx) {
return LoadMask(ctx, ctx.subgroup_mask_gt);
}
Id EmitSubgroupGeMask(EmitContext& ctx) {
return LoadMask(ctx, ctx.subgroup_mask_ge);
}
Id EmitShuffleIndex(EmitContext& ctx, IR::Inst* inst, Id value, Id index, Id clamp,
Id segmentation_mask) {
const Id not_seg_mask{ctx.OpNot(ctx.U32[1], segmentation_mask)};
const Id thread_id{GetThreadId(ctx)};
if (ctx.profile.warp_size_potentially_larger_than_guest) {
const Id thirty_two{ctx.Const(32u)};
const Id is_upper_partition{ctx.OpSGreaterThanEqual(ctx.U1, thread_id, thirty_two)};
const Id upper_index{ctx.OpIAdd(ctx.U32[1], thirty_two, index)};
const Id upper_clamp{ctx.OpIAdd(ctx.U32[1], thirty_two, clamp)};
index = ctx.OpSelect(ctx.U32[1], is_upper_partition, upper_index, index);
clamp = ctx.OpSelect(ctx.U32[1], is_upper_partition, upper_clamp, clamp);
}
const Id min_thread_id{ComputeMinThreadId(ctx, thread_id, segmentation_mask)};
const Id max_thread_id{ComputeMaxThreadId(ctx, min_thread_id, clamp, not_seg_mask)};
const Id lhs{ctx.OpBitwiseAnd(ctx.U32[1], index, not_seg_mask)};
const Id src_thread_id{ctx.OpBitwiseOr(ctx.U32[1], lhs, min_thread_id)};
const Id in_range{ctx.OpSLessThanEqual(ctx.U1, src_thread_id, max_thread_id)};
SetInBoundsFlag(inst, in_range);
return SelectValue(ctx, in_range, value, src_thread_id);
}
Id EmitShuffleUp(EmitContext& ctx, IR::Inst* inst, Id value, Id index, Id clamp,
Id segmentation_mask) {
const Id thread_id{GetThreadId(ctx)};
if (ctx.profile.warp_size_potentially_larger_than_guest) {
clamp = GetUpperClamp(ctx, thread_id, clamp);
}
const Id max_thread_id{GetMaxThreadId(ctx, thread_id, clamp, segmentation_mask)};
const Id src_thread_id{ctx.OpISub(ctx.U32[1], thread_id, index)};
const Id in_range{ctx.OpSGreaterThanEqual(ctx.U1, src_thread_id, max_thread_id)};
SetInBoundsFlag(inst, in_range);
return SelectValue(ctx, in_range, value, src_thread_id);
}
Id EmitShuffleDown(EmitContext& ctx, IR::Inst* inst, Id value, Id index, Id clamp,
Id segmentation_mask) {
const Id thread_id{GetThreadId(ctx)};
if (ctx.profile.warp_size_potentially_larger_than_guest) {
clamp = GetUpperClamp(ctx, thread_id, clamp);
}
const Id max_thread_id{GetMaxThreadId(ctx, thread_id, clamp, segmentation_mask)};
const Id src_thread_id{ctx.OpIAdd(ctx.U32[1], thread_id, index)};
const Id in_range{ctx.OpSLessThanEqual(ctx.U1, src_thread_id, max_thread_id)};
SetInBoundsFlag(inst, in_range);
return SelectValue(ctx, in_range, value, src_thread_id);
}
Id EmitShuffleButterfly(EmitContext& ctx, IR::Inst* inst, Id value, Id index, Id clamp,
Id segmentation_mask) {
const Id thread_id{GetThreadId(ctx)};
if (ctx.profile.warp_size_potentially_larger_than_guest) {
clamp = GetUpperClamp(ctx, thread_id, clamp);
}
const Id max_thread_id{GetMaxThreadId(ctx, thread_id, clamp, segmentation_mask)};
const Id src_thread_id{ctx.OpBitwiseXor(ctx.U32[1], thread_id, index)};
const Id in_range{ctx.OpSLessThanEqual(ctx.U1, src_thread_id, max_thread_id)};
SetInBoundsFlag(inst, in_range);
return SelectValue(ctx, in_range, value, src_thread_id);
}
Id EmitFSwizzleAdd(EmitContext& ctx, Id op_a, Id op_b, Id swizzle) {
const Id three{ctx.Const(3U)};
Id mask{ctx.OpLoad(ctx.U32[1], ctx.subgroup_local_invocation_id)};
mask = ctx.OpBitwiseAnd(ctx.U32[1], mask, three);
mask = ctx.OpShiftLeftLogical(ctx.U32[1], mask, ctx.Const(1U));
mask = ctx.OpShiftRightLogical(ctx.U32[1], swizzle, mask);
mask = ctx.OpBitwiseAnd(ctx.U32[1], mask, three);
const Id modifier_a{ctx.OpVectorExtractDynamic(ctx.F32[1], ctx.fswzadd_lut_a, mask)};
const Id modifier_b{ctx.OpVectorExtractDynamic(ctx.F32[1], ctx.fswzadd_lut_b, mask)};
const Id result_a{ctx.OpFMul(ctx.F32[1], op_a, modifier_a)};
const Id result_b{ctx.OpFMul(ctx.F32[1], op_b, modifier_b)};
return ctx.OpFAdd(ctx.F32[1], result_a, result_b);
}
Id EmitDPdxFine(EmitContext& ctx, Id op_a) {
return ctx.OpDPdxFine(ctx.F32[1], op_a);
}
Id EmitDPdyFine(EmitContext& ctx, Id op_a) {
return ctx.OpDPdyFine(ctx.F32[1], op_a);
}
Id EmitDPdxCoarse(EmitContext& ctx, Id op_a) {
return ctx.OpDPdxCoarse(ctx.F32[1], op_a);
}
Id EmitDPdyCoarse(EmitContext& ctx, Id op_a) {
return ctx.OpDPdyCoarse(ctx.F32[1], op_a);
}
} // namespace Shader::Backend::SPIRV
|
