Flutter Engine
The Flutter Engine
compute_pass_mtl.mm
Go to the documentation of this file.
1// Copyright 2013 The Flutter Authors. All rights reserved.
2// Use of this source code is governed by a BSD-style license that can be
3// found in the LICENSE file.
4
6
7#include <Metal/Metal.h>
8#include <memory>
9
10#include "flutter/fml/backtrace.h"
11#include "flutter/fml/closure.h"
12#include "flutter/fml/logging.h"
13#include "flutter/fml/trace_event.h"
14#include "fml/status.h"
26
27namespace impeller {
28
29ComputePassMTL::ComputePassMTL(std::shared_ptr<const Context> context,
30 id<MTLCommandBuffer> buffer)
31 : ComputePass(std::move(context)), buffer_(buffer) {
32 if (!buffer_) {
33 return;
34 }
35 encoder_ = [buffer_ computeCommandEncoderWithDispatchType:
36 MTLDispatchType::MTLDispatchTypeConcurrent];
37 if (!encoder_) {
38 return;
39 }
40 pass_bindings_cache_.SetEncoder(encoder_);
41 is_valid_ = true;
42}
43
44ComputePassMTL::~ComputePassMTL() = default;
45
46bool ComputePassMTL::IsValid() const {
47 return is_valid_;
48}
49
50void ComputePassMTL::OnSetLabel(const std::string& label) {
51#ifdef IMPELLER_DEBUG
52 if (label.empty()) {
53 return;
54 }
55 [encoder_ setLabel:@(label.c_str())];
56#endif // IMPELLER_DEBUG
57}
58
59void ComputePassMTL::SetCommandLabel(std::string_view label) {
60 has_label_ = true;
61 [encoder_ pushDebugGroup:@(label.data())];
62}
63
64// |ComputePass|
65void ComputePassMTL::SetPipeline(
66 const std::shared_ptr<Pipeline<ComputePipelineDescriptor>>& pipeline) {
67 pass_bindings_cache_.SetComputePipelineState(
68 ComputePipelineMTL::Cast(*pipeline).GetMTLComputePipelineState());
69}
70
71// |ComputePass|
72void ComputePassMTL::AddBufferMemoryBarrier() {
73 [encoder_ memoryBarrierWithScope:MTLBarrierScopeBuffers];
74}
75
76// |ComputePass|
77void ComputePassMTL::AddTextureMemoryBarrier() {
78 [encoder_ memoryBarrierWithScope:MTLBarrierScopeTextures];
79}
80
81// |ComputePass|
82bool ComputePassMTL::BindResource(ShaderStage stage,
84 const ShaderUniformSlot& slot,
85 const ShaderMetadata& metadata,
86 BufferView view) {
87 if (!view.buffer) {
88 return false;
89 }
90
91 const std::shared_ptr<const DeviceBuffer>& device_buffer = view.buffer;
92 if (!device_buffer) {
93 return false;
94 }
95
96 id<MTLBuffer> buffer = DeviceBufferMTL::Cast(*device_buffer).GetMTLBuffer();
97 // The Metal call is a void return and we don't want to make it on nil.
98 if (!buffer) {
99 return false;
100 }
101
102 pass_bindings_cache_.SetBuffer(slot.ext_res_0, view.range.offset, buffer);
103 return true;
104}
105
106// |ComputePass|
107bool ComputePassMTL::BindResource(
108 ShaderStage stage,
110 const SampledImageSlot& slot,
111 const ShaderMetadata& metadata,
112 std::shared_ptr<const Texture> texture,
113 const std::unique_ptr<const Sampler>& sampler) {
114 if (!sampler || !texture->IsValid()) {
115 return false;
116 }
117
118 pass_bindings_cache_.SetTexture(slot.texture_index,
119 TextureMTL::Cast(*texture).GetMTLTexture());
120 pass_bindings_cache_.SetSampler(
121 slot.texture_index, SamplerMTL::Cast(*sampler).GetMTLSamplerState());
122 return true;
123}
124
125fml::Status ComputePassMTL::Compute(const ISize& grid_size) {
126 if (grid_size.IsEmpty()) {
128 "Invalid grid size for compute command.");
129 }
130 // TODO(dnfield): use feature detection to support non-uniform threadgroup
131 // sizes.
132 // https://github.com/flutter/flutter/issues/110619
133 auto width = grid_size.width;
134 auto height = grid_size.height;
135
136 auto max_total_threads_per_threadgroup = static_cast<int64_t>(
137 pass_bindings_cache_.GetPipeline().maxTotalThreadsPerThreadgroup);
138
139 // Special case for linear processing.
140 if (height == 1) {
141 int64_t thread_groups = std::max(
142 static_cast<int64_t>(
143 std::ceil(width * 1.0 / max_total_threads_per_threadgroup * 1.0)),
144 1LL);
145 [encoder_
146 dispatchThreadgroups:MTLSizeMake(thread_groups, 1, 1)
147 threadsPerThreadgroup:MTLSizeMake(max_total_threads_per_threadgroup, 1,
148 1)];
149 } else {
150 while (width * height > max_total_threads_per_threadgroup) {
151 width = std::max(1LL, width / 2);
152 height = std::max(1LL, height / 2);
153 }
154
155 auto size = MTLSizeMake(width, height, 1);
156 [encoder_ dispatchThreadgroups:size threadsPerThreadgroup:size];
157 }
158
159#ifdef IMPELLER_DEBUG
160 if (has_label_) {
161 [encoder_ popDebugGroup];
162 }
163 has_label_ = false;
164#endif // IMPELLER_DEBUG
165 return fml::Status();
166}
167
168bool ComputePassMTL::EncodeCommands() const {
169 [encoder_ endEncoding];
170 return true;
171}
172
173} // namespace impeller
GLenum type
static float max(float r, float g, float b)
Definition: hsl.cpp:49
FlTexture * texture
DEF_SWITCHES_START aot vmservice shared library Name of the *so containing AOT compiled Dart assets for launching the service isolate vm snapshot The VM snapshot data that will be memory mapped as read only SnapshotAssetPath must be present isolate snapshot The isolate snapshot data that will be memory mapped as read only SnapshotAssetPath must be present cache dir Path to the cache directory This is different from the persistent_cache_path in embedder which is used for Skia shader cache icu native lib Path to the library file that exports the ICU data vm service The hostname IP address on which the Dart VM Service should be served If not defaults to or::depending on whether ipv6 is specified vm service A custom Dart VM Service port The default is to pick a randomly available open port disable vm Disable the Dart VM Service The Dart VM Service is never available in release mode disable vm service Disable mDNS Dart VM Service publication Bind to the IPv6 localhost address for the Dart VM Service Ignored if vm service host is set endless trace buffer
Definition: switches.h:126
it will be possible to load the file into Perfetto s trace viewer disable asset Prevents usage of any non test fonts unless they were explicitly Loaded via prefetched default font Indicates whether the embedding started a prefetch of the default font manager before creating the engine run In non interactive keep the shell running after the Dart script has completed enable serial On low power devices with low core running concurrent GC tasks on threads can cause them to contend with the UI thread which could potentially lead to jank This option turns off all concurrent GC activities domain network JSON encoded network policy per domain This overrides the DisallowInsecureConnections switch Embedder can specify whether to allow or disallow insecure connections at a domain level old gen heap size
Definition: switches.h:259
ISize64 ISize
Definition: size.h:140
Task::Status Status
Definition: TaskList.cpp:15
SIN Vec< N, float > ceil(const Vec< N, float > &x)
Definition: SkVx.h:702
Definition: ref_ptr.h:256
int32_t height
int32_t width