@@ -134,13 +134,16 @@ bool BlitPassVK::OnCopyTextureToTextureCommand(
134134 return true ;
135135 }
136136
137+ // The copy must complete before a later render or compute pass samples the
138+ // destination.
137139 BarrierVK barrier;
138140 barrier.cmd_buffer = cmd_buffer;
139141 barrier.new_layout = vk::ImageLayout::eShaderReadOnlyOptimal;
140- barrier.src_access = {} ;
141- barrier.src_stage = vk::PipelineStageFlagBits::eTopOfPipe ;
142+ barrier.src_access = vk::AccessFlagBits::eTransferWrite ;
143+ barrier.src_stage = vk::PipelineStageFlagBits::eTransfer ;
142144 barrier.dst_access = vk::AccessFlagBits::eShaderRead;
143- barrier.dst_stage = vk::PipelineStageFlagBits::eFragmentShader;
145+ barrier.dst_stage = vk::PipelineStageFlagBits::eFragmentShader |
146+ vk::PipelineStageFlagBits::eComputeShader;
144147
145148 return dst.SetLayout (barrier);
146149}
@@ -170,9 +173,9 @@ bool BlitPassVK::OnCopyTextureToBufferCommand(
170173 barrier.src_stage = vk::PipelineStageFlagBits::eFragmentShader |
171174 vk::PipelineStageFlagBits::eTransfer |
172175 vk::PipelineStageFlagBits::eColorAttachmentOutput;
173- barrier. dst_access = vk::AccessFlagBits::eShaderRead;
174- barrier.dst_stage = vk::PipelineStageFlagBits::eVertexShader |
175- vk::PipelineStageFlagBits::eFragmentShader ;
176+ // The copy below reads the image.
177+ barrier.dst_access = vk::AccessFlagBits::eTransferRead;
178+ barrier. dst_stage = vk::PipelineStageFlagBits::eTransfer ;
176179
177180 const auto & dst = DeviceBufferVK::Cast (*destination);
178181
@@ -198,16 +201,22 @@ bool BlitPassVK::OnCopyTextureToBufferCommand(
198201 image_copy //
199202 );
200203
201- // If the buffer is used for readback, then apply a transfer -> host memory
202- // barrier .
203- if (destination-> GetDeviceBufferDescriptor (). readback ) {
204+ // The copy must complete before a later compute pass reads or writes the
205+ // buffer. If the buffer is used for readback, the host reads it too .
206+ {
204207 vk::MemoryBarrier barrier;
205208 barrier.srcAccessMask = vk::AccessFlagBits::eTransferWrite;
206- barrier.dstAccessMask = vk::AccessFlagBits::eHostRead;
209+ barrier.dstAccessMask =
210+ vk::AccessFlagBits::eShaderRead | vk::AccessFlagBits::eShaderWrite;
211+ vk::PipelineStageFlags dst_stage =
212+ vk::PipelineStageFlagBits::eComputeShader;
213+ if (destination->GetDeviceBufferDescriptor ().readback ) {
214+ barrier.dstAccessMask |= vk::AccessFlagBits::eHostRead;
215+ dst_stage |= vk::PipelineStageFlagBits::eHost;
216+ }
207217
208- cmd_buffer.pipelineBarrier (vk::PipelineStageFlagBits::eTransfer,
209- vk::PipelineStageFlagBits::eHost, {}, 1 ,
210- &barrier, 0 , {}, 0 , {});
218+ cmd_buffer.pipelineBarrier (vk::PipelineStageFlagBits::eTransfer, dst_stage,
219+ {}, 1 , &barrier, 0 , {}, 0 , {});
211220 }
212221
213222 return true ;
@@ -222,7 +231,8 @@ bool BlitPassVK::ConvertTextureToShaderRead(
222231 barrier.src_access = vk::AccessFlagBits::eTransferWrite;
223232 barrier.src_stage = vk::PipelineStageFlagBits::eTransfer;
224233 barrier.dst_access = vk::AccessFlagBits::eShaderRead;
225- barrier.dst_stage = vk::PipelineStageFlagBits::eFragmentShader;
234+ barrier.dst_stage = vk::PipelineStageFlagBits::eFragmentShader |
235+ vk::PipelineStageFlagBits::eComputeShader;
226236
227237 barrier.new_layout = vk::ImageLayout::eShaderReadOnlyOptimal;
228238
@@ -300,7 +310,8 @@ bool BlitPassVK::OnCopyBufferToTextureCommand(
300310 barrier.src_access = vk::AccessFlagBits::eTransferWrite;
301311 barrier.src_stage = vk::PipelineStageFlagBits::eTransfer;
302312 barrier.dst_access = vk::AccessFlagBits::eShaderRead;
303- barrier.dst_stage = vk::PipelineStageFlagBits::eFragmentShader;
313+ barrier.dst_stage = vk::PipelineStageFlagBits::eFragmentShader |
314+ vk::PipelineStageFlagBits::eComputeShader;
304315
305316 barrier.new_layout = vk::ImageLayout::eShaderReadOnlyOptimal;
306317
@@ -382,15 +393,17 @@ bool BlitPassVK::ResizeTexture(const std::shared_ptr<Texture>& source,
382393
383394 );
384395
385- // Convert back to shader read
396+ // Convert back to shader read once the blit completes, for a later render or
397+ // compute pass.
386398
387399 BarrierVK barrier;
388400 barrier.cmd_buffer = cmd_buffer;
389401 barrier.new_layout = vk::ImageLayout::eShaderReadOnlyOptimal;
390- barrier.src_access = {} ;
391- barrier.src_stage = vk::PipelineStageFlagBits::eTopOfPipe ;
402+ barrier.src_access = vk::AccessFlagBits::eTransferWrite ;
403+ barrier.src_stage = vk::PipelineStageFlagBits::eTransfer ;
392404 barrier.dst_access = vk::AccessFlagBits::eShaderRead;
393- barrier.dst_stage = vk::PipelineStageFlagBits::eFragmentShader;
405+ barrier.dst_stage = vk::PipelineStageFlagBits::eFragmentShader |
406+ vk::PipelineStageFlagBits::eComputeShader;
394407
395408 return dst.SetLayout (barrier);
396409}
@@ -520,8 +533,9 @@ bool BlitPassVK::OnGenerateMipmapCommand(std::shared_ptr<Texture> texture,
520533 barrier.dstAccessMask = vk::AccessFlagBits::eShaderRead;
521534
522535 cmd.pipelineBarrier (vk::PipelineStageFlagBits::eTransfer,
523- vk::PipelineStageFlagBits::eFragmentShader, {}, {}, {},
524- {barrier});
536+ vk::PipelineStageFlagBits::eFragmentShader |
537+ vk::PipelineStageFlagBits::eComputeShader,
538+ {}, {}, {}, {barrier});
525539
526540 // We modified the layouts of this image from underneath it. Tell it its new
527541 // state so it doesn't try to perform redundant transitions under the hood.
0 commit comments