diff options
Diffstat (limited to '')
| -rw-r--r-- | src/video_core/CMakeLists.txt | 4 | ||||
| -rw-r--r-- | src/video_core/renderer_vulkan/vk_compute_pass.cpp | 339 | ||||
| -rw-r--r-- | src/video_core/renderer_vulkan/vk_compute_pass.h | 77 | ||||
| -rw-r--r-- | src/video_core/renderer_vulkan/vk_shader_util.cpp | 34 | ||||
| -rw-r--r-- | src/video_core/renderer_vulkan/vk_shader_util.h | 17 |
5 files changed, 471 insertions, 0 deletions
diff --git a/src/video_core/CMakeLists.txt b/src/video_core/CMakeLists.txt index caf03c2ae..142852082 100644 --- a/src/video_core/CMakeLists.txt +++ b/src/video_core/CMakeLists.txt | |||
| @@ -155,6 +155,8 @@ if (ENABLE_VULKAN) | |||
| 155 | renderer_vulkan/maxwell_to_vk.h | 155 | renderer_vulkan/maxwell_to_vk.h |
| 156 | renderer_vulkan/vk_buffer_cache.cpp | 156 | renderer_vulkan/vk_buffer_cache.cpp |
| 157 | renderer_vulkan/vk_buffer_cache.h | 157 | renderer_vulkan/vk_buffer_cache.h |
| 158 | renderer_vulkan/vk_compute_pass.cpp | ||
| 159 | renderer_vulkan/vk_compute_pass.h | ||
| 158 | renderer_vulkan/vk_compute_pipeline.cpp | 160 | renderer_vulkan/vk_compute_pipeline.cpp |
| 159 | renderer_vulkan/vk_compute_pipeline.h | 161 | renderer_vulkan/vk_compute_pipeline.h |
| 160 | renderer_vulkan/vk_descriptor_pool.cpp | 162 | renderer_vulkan/vk_descriptor_pool.cpp |
| @@ -180,6 +182,8 @@ if (ENABLE_VULKAN) | |||
| 180 | renderer_vulkan/vk_scheduler.h | 182 | renderer_vulkan/vk_scheduler.h |
| 181 | renderer_vulkan/vk_shader_decompiler.cpp | 183 | renderer_vulkan/vk_shader_decompiler.cpp |
| 182 | renderer_vulkan/vk_shader_decompiler.h | 184 | renderer_vulkan/vk_shader_decompiler.h |
| 185 | renderer_vulkan/vk_shader_util.cpp | ||
| 186 | renderer_vulkan/vk_shader_util.h | ||
| 183 | renderer_vulkan/vk_staging_buffer_pool.cpp | 187 | renderer_vulkan/vk_staging_buffer_pool.cpp |
| 184 | renderer_vulkan/vk_staging_buffer_pool.h | 188 | renderer_vulkan/vk_staging_buffer_pool.h |
| 185 | renderer_vulkan/vk_stream_buffer.cpp | 189 | renderer_vulkan/vk_stream_buffer.cpp |
diff --git a/src/video_core/renderer_vulkan/vk_compute_pass.cpp b/src/video_core/renderer_vulkan/vk_compute_pass.cpp new file mode 100644 index 000000000..7bdda3d79 --- /dev/null +++ b/src/video_core/renderer_vulkan/vk_compute_pass.cpp | |||
| @@ -0,0 +1,339 @@ | |||
| 1 | // Copyright 2019 yuzu Emulator Project | ||
| 2 | // Licensed under GPLv2 or any later version | ||
| 3 | // Refer to the license.txt file included. | ||
| 4 | |||
| 5 | #include <cstring> | ||
| 6 | #include <memory> | ||
| 7 | #include <optional> | ||
| 8 | #include <utility> | ||
| 9 | #include <vector> | ||
| 10 | #include "common/alignment.h" | ||
| 11 | #include "common/assert.h" | ||
| 12 | #include "common/common_types.h" | ||
| 13 | #include "video_core/renderer_vulkan/declarations.h" | ||
| 14 | #include "video_core/renderer_vulkan/vk_compute_pass.h" | ||
| 15 | #include "video_core/renderer_vulkan/vk_descriptor_pool.h" | ||
| 16 | #include "video_core/renderer_vulkan/vk_device.h" | ||
| 17 | #include "video_core/renderer_vulkan/vk_scheduler.h" | ||
| 18 | #include "video_core/renderer_vulkan/vk_staging_buffer_pool.h" | ||
| 19 | #include "video_core/renderer_vulkan/vk_update_descriptor.h" | ||
| 20 | |||
| 21 | namespace Vulkan { | ||
| 22 | |||
| 23 | namespace { | ||
| 24 | |||
| 25 | // Quad array SPIR-V module. Generated from the "shaders/" directory, read the instructions there. | ||
| 26 | constexpr u8 quad_array[] = { | ||
| 27 | 0x03, 0x02, 0x23, 0x07, 0x00, 0x00, 0x01, 0x00, 0x07, 0x00, 0x08, 0x00, 0x54, 0x00, 0x00, 0x00, | ||
| 28 | 0x00, 0x00, 0x00, 0x00, 0x11, 0x00, 0x02, 0x00, 0x01, 0x00, 0x00, 0x00, 0x0b, 0x00, 0x06, 0x00, | ||
| 29 | 0x01, 0x00, 0x00, 0x00, 0x47, 0x4c, 0x53, 0x4c, 0x2e, 0x73, 0x74, 0x64, 0x2e, 0x34, 0x35, 0x30, | ||
| 30 | 0x00, 0x00, 0x00, 0x00, 0x0e, 0x00, 0x03, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, | ||
| 31 | 0x0f, 0x00, 0x06, 0x00, 0x05, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x6d, 0x61, 0x69, 0x6e, | ||
| 32 | 0x00, 0x00, 0x00, 0x00, 0x0b, 0x00, 0x00, 0x00, 0x10, 0x00, 0x06, 0x00, 0x04, 0x00, 0x00, 0x00, | ||
| 33 | 0x11, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, | ||
| 34 | 0x47, 0x00, 0x04, 0x00, 0x0b, 0x00, 0x00, 0x00, 0x0b, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, | ||
| 35 | 0x47, 0x00, 0x04, 0x00, 0x13, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, | ||
| 36 | 0x48, 0x00, 0x05, 0x00, 0x14, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x23, 0x00, 0x00, 0x00, | ||
| 37 | 0x00, 0x00, 0x00, 0x00, 0x47, 0x00, 0x03, 0x00, 0x14, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, | ||
| 38 | 0x47, 0x00, 0x04, 0x00, 0x16, 0x00, 0x00, 0x00, 0x22, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, | ||
| 39 | 0x47, 0x00, 0x04, 0x00, 0x16, 0x00, 0x00, 0x00, 0x21, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, | ||
| 40 | 0x48, 0x00, 0x05, 0x00, 0x29, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x23, 0x00, 0x00, 0x00, | ||
| 41 | 0x00, 0x00, 0x00, 0x00, 0x47, 0x00, 0x03, 0x00, 0x29, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, | ||
| 42 | 0x47, 0x00, 0x04, 0x00, 0x4a, 0x00, 0x00, 0x00, 0x0b, 0x00, 0x00, 0x00, 0x19, 0x00, 0x00, 0x00, | ||
| 43 | 0x13, 0x00, 0x02, 0x00, 0x02, 0x00, 0x00, 0x00, 0x21, 0x00, 0x03, 0x00, 0x03, 0x00, 0x00, 0x00, | ||
| 44 | 0x02, 0x00, 0x00, 0x00, 0x15, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, | ||
| 45 | 0x00, 0x00, 0x00, 0x00, 0x20, 0x00, 0x04, 0x00, 0x07, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, | ||
| 46 | 0x06, 0x00, 0x00, 0x00, 0x17, 0x00, 0x04, 0x00, 0x09, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, | ||
| 47 | 0x03, 0x00, 0x00, 0x00, 0x20, 0x00, 0x04, 0x00, 0x0a, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, | ||
| 48 | 0x09, 0x00, 0x00, 0x00, 0x3b, 0x00, 0x04, 0x00, 0x0a, 0x00, 0x00, 0x00, 0x0b, 0x00, 0x00, 0x00, | ||
| 49 | 0x01, 0x00, 0x00, 0x00, 0x2b, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x00, 0x00, | ||
| 50 | 0x00, 0x00, 0x00, 0x00, 0x20, 0x00, 0x04, 0x00, 0x0d, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, | ||
| 51 | 0x06, 0x00, 0x00, 0x00, 0x2b, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x11, 0x00, 0x00, 0x00, | ||
| 52 | 0x06, 0x00, 0x00, 0x00, 0x1d, 0x00, 0x03, 0x00, 0x13, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, | ||
| 53 | 0x1e, 0x00, 0x03, 0x00, 0x14, 0x00, 0x00, 0x00, 0x13, 0x00, 0x00, 0x00, 0x20, 0x00, 0x04, 0x00, | ||
| 54 | 0x15, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x3b, 0x00, 0x04, 0x00, | ||
| 55 | 0x15, 0x00, 0x00, 0x00, 0x16, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x15, 0x00, 0x04, 0x00, | ||
| 56 | 0x18, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x14, 0x00, 0x02, 0x00, | ||
| 57 | 0x1b, 0x00, 0x00, 0x00, 0x1e, 0x00, 0x03, 0x00, 0x29, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, | ||
| 58 | 0x20, 0x00, 0x04, 0x00, 0x2a, 0x00, 0x00, 0x00, 0x09, 0x00, 0x00, 0x00, 0x29, 0x00, 0x00, 0x00, | ||
| 59 | 0x3b, 0x00, 0x04, 0x00, 0x2a, 0x00, 0x00, 0x00, 0x2b, 0x00, 0x00, 0x00, 0x09, 0x00, 0x00, 0x00, | ||
| 60 | 0x2b, 0x00, 0x04, 0x00, 0x18, 0x00, 0x00, 0x00, 0x2c, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, | ||
| 61 | 0x20, 0x00, 0x04, 0x00, 0x2d, 0x00, 0x00, 0x00, 0x09, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, | ||
| 62 | 0x2b, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x31, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, | ||
| 63 | 0x1c, 0x00, 0x04, 0x00, 0x34, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x11, 0x00, 0x00, 0x00, | ||
| 64 | 0x2b, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x35, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, | ||
| 65 | 0x2b, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x36, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, | ||
| 66 | 0x2b, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x37, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, | ||
| 67 | 0x2c, 0x00, 0x09, 0x00, 0x34, 0x00, 0x00, 0x00, 0x38, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x00, 0x00, | ||
| 68 | 0x35, 0x00, 0x00, 0x00, 0x36, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x00, 0x00, 0x36, 0x00, 0x00, 0x00, | ||
| 69 | 0x37, 0x00, 0x00, 0x00, 0x20, 0x00, 0x04, 0x00, 0x3a, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, | ||
| 70 | 0x34, 0x00, 0x00, 0x00, 0x20, 0x00, 0x04, 0x00, 0x44, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, | ||
| 71 | 0x06, 0x00, 0x00, 0x00, 0x2b, 0x00, 0x04, 0x00, 0x18, 0x00, 0x00, 0x00, 0x47, 0x00, 0x00, 0x00, | ||
| 72 | 0x01, 0x00, 0x00, 0x00, 0x2b, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x49, 0x00, 0x00, 0x00, | ||
| 73 | 0x00, 0x04, 0x00, 0x00, 0x2c, 0x00, 0x06, 0x00, 0x09, 0x00, 0x00, 0x00, 0x4a, 0x00, 0x00, 0x00, | ||
| 74 | 0x49, 0x00, 0x00, 0x00, 0x35, 0x00, 0x00, 0x00, 0x35, 0x00, 0x00, 0x00, 0x36, 0x00, 0x05, 0x00, | ||
| 75 | 0x02, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, | ||
| 76 | 0xf8, 0x00, 0x02, 0x00, 0x05, 0x00, 0x00, 0x00, 0x3b, 0x00, 0x04, 0x00, 0x3a, 0x00, 0x00, 0x00, | ||
| 77 | 0x3b, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0xf9, 0x00, 0x02, 0x00, 0x4c, 0x00, 0x00, 0x00, | ||
| 78 | 0xf8, 0x00, 0x02, 0x00, 0x4c, 0x00, 0x00, 0x00, 0xf6, 0x00, 0x04, 0x00, 0x4b, 0x00, 0x00, 0x00, | ||
| 79 | 0x4e, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xf9, 0x00, 0x02, 0x00, 0x4d, 0x00, 0x00, 0x00, | ||
| 80 | 0xf8, 0x00, 0x02, 0x00, 0x4d, 0x00, 0x00, 0x00, 0x41, 0x00, 0x05, 0x00, 0x0d, 0x00, 0x00, 0x00, | ||
| 81 | 0x0e, 0x00, 0x00, 0x00, 0x0b, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x00, 0x00, 0x3d, 0x00, 0x04, 0x00, | ||
| 82 | 0x06, 0x00, 0x00, 0x00, 0x0f, 0x00, 0x00, 0x00, 0x0e, 0x00, 0x00, 0x00, 0x84, 0x00, 0x05, 0x00, | ||
| 83 | 0x06, 0x00, 0x00, 0x00, 0x12, 0x00, 0x00, 0x00, 0x0f, 0x00, 0x00, 0x00, 0x11, 0x00, 0x00, 0x00, | ||
| 84 | 0x44, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, 0x17, 0x00, 0x00, 0x00, 0x16, 0x00, 0x00, 0x00, | ||
| 85 | 0x00, 0x00, 0x00, 0x00, 0x7c, 0x00, 0x04, 0x00, 0x18, 0x00, 0x00, 0x00, 0x19, 0x00, 0x00, 0x00, | ||
| 86 | 0x17, 0x00, 0x00, 0x00, 0x7c, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x1a, 0x00, 0x00, 0x00, | ||
| 87 | 0x19, 0x00, 0x00, 0x00, 0xae, 0x00, 0x05, 0x00, 0x1b, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, | ||
| 88 | 0x12, 0x00, 0x00, 0x00, 0x1a, 0x00, 0x00, 0x00, 0xf7, 0x00, 0x03, 0x00, 0x1e, 0x00, 0x00, 0x00, | ||
| 89 | 0x00, 0x00, 0x00, 0x00, 0xfa, 0x00, 0x04, 0x00, 0x1c, 0x00, 0x00, 0x00, 0x1d, 0x00, 0x00, 0x00, | ||
| 90 | 0x1e, 0x00, 0x00, 0x00, 0xf8, 0x00, 0x02, 0x00, 0x1d, 0x00, 0x00, 0x00, 0xf9, 0x00, 0x02, 0x00, | ||
| 91 | 0x4b, 0x00, 0x00, 0x00, 0xf8, 0x00, 0x02, 0x00, 0x1e, 0x00, 0x00, 0x00, 0xf9, 0x00, 0x02, 0x00, | ||
| 92 | 0x21, 0x00, 0x00, 0x00, 0xf8, 0x00, 0x02, 0x00, 0x21, 0x00, 0x00, 0x00, 0xf5, 0x00, 0x07, 0x00, | ||
| 93 | 0x06, 0x00, 0x00, 0x00, 0x53, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x00, 0x00, 0x1e, 0x00, 0x00, 0x00, | ||
| 94 | 0x48, 0x00, 0x00, 0x00, 0x22, 0x00, 0x00, 0x00, 0xb0, 0x00, 0x05, 0x00, 0x1b, 0x00, 0x00, 0x00, | ||
| 95 | 0x27, 0x00, 0x00, 0x00, 0x53, 0x00, 0x00, 0x00, 0x11, 0x00, 0x00, 0x00, 0xf6, 0x00, 0x04, 0x00, | ||
| 96 | 0x23, 0x00, 0x00, 0x00, 0x22, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xfa, 0x00, 0x04, 0x00, | ||
| 97 | 0x27, 0x00, 0x00, 0x00, 0x22, 0x00, 0x00, 0x00, 0x23, 0x00, 0x00, 0x00, 0xf8, 0x00, 0x02, 0x00, | ||
| 98 | 0x22, 0x00, 0x00, 0x00, 0x41, 0x00, 0x05, 0x00, 0x2d, 0x00, 0x00, 0x00, 0x2e, 0x00, 0x00, 0x00, | ||
| 99 | 0x2b, 0x00, 0x00, 0x00, 0x2c, 0x00, 0x00, 0x00, 0x3d, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, | ||
| 100 | 0x2f, 0x00, 0x00, 0x00, 0x2e, 0x00, 0x00, 0x00, 0x84, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, | ||
| 101 | 0x32, 0x00, 0x00, 0x00, 0x0f, 0x00, 0x00, 0x00, 0x31, 0x00, 0x00, 0x00, 0x80, 0x00, 0x05, 0x00, | ||
| 102 | 0x06, 0x00, 0x00, 0x00, 0x33, 0x00, 0x00, 0x00, 0x2f, 0x00, 0x00, 0x00, 0x32, 0x00, 0x00, 0x00, | ||
| 103 | 0x3e, 0x00, 0x03, 0x00, 0x3b, 0x00, 0x00, 0x00, 0x38, 0x00, 0x00, 0x00, 0x41, 0x00, 0x05, 0x00, | ||
| 104 | 0x07, 0x00, 0x00, 0x00, 0x3c, 0x00, 0x00, 0x00, 0x3b, 0x00, 0x00, 0x00, 0x53, 0x00, 0x00, 0x00, | ||
| 105 | 0x3d, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x3d, 0x00, 0x00, 0x00, 0x3c, 0x00, 0x00, 0x00, | ||
| 106 | 0x80, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, 0x3e, 0x00, 0x00, 0x00, 0x33, 0x00, 0x00, 0x00, | ||
| 107 | 0x3d, 0x00, 0x00, 0x00, 0x80, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, 0x42, 0x00, 0x00, 0x00, | ||
| 108 | 0x12, 0x00, 0x00, 0x00, 0x53, 0x00, 0x00, 0x00, 0x41, 0x00, 0x06, 0x00, 0x44, 0x00, 0x00, 0x00, | ||
| 109 | 0x45, 0x00, 0x00, 0x00, 0x16, 0x00, 0x00, 0x00, 0x2c, 0x00, 0x00, 0x00, 0x42, 0x00, 0x00, 0x00, | ||
| 110 | 0x3e, 0x00, 0x03, 0x00, 0x45, 0x00, 0x00, 0x00, 0x3e, 0x00, 0x00, 0x00, 0x80, 0x00, 0x05, 0x00, | ||
| 111 | 0x06, 0x00, 0x00, 0x00, 0x48, 0x00, 0x00, 0x00, 0x53, 0x00, 0x00, 0x00, 0x47, 0x00, 0x00, 0x00, | ||
| 112 | 0xf9, 0x00, 0x02, 0x00, 0x21, 0x00, 0x00, 0x00, 0xf8, 0x00, 0x02, 0x00, 0x23, 0x00, 0x00, 0x00, | ||
| 113 | 0xf9, 0x00, 0x02, 0x00, 0x4b, 0x00, 0x00, 0x00, 0xf8, 0x00, 0x02, 0x00, 0x4e, 0x00, 0x00, 0x00, | ||
| 114 | 0xf9, 0x00, 0x02, 0x00, 0x4c, 0x00, 0x00, 0x00, 0xf8, 0x00, 0x02, 0x00, 0x4b, 0x00, 0x00, 0x00, | ||
| 115 | 0xfd, 0x00, 0x01, 0x00, 0x38, 0x00, 0x01, 0x00}; | ||
| 116 | |||
| 117 | // Uint8 SPIR-V module. Generated from the "shaders/" directory. | ||
| 118 | constexpr u8 uint8_pass[] = { | ||
| 119 | 0x03, 0x02, 0x23, 0x07, 0x00, 0x00, 0x01, 0x00, 0x07, 0x00, 0x08, 0x00, 0x2f, 0x00, 0x00, 0x00, | ||
| 120 | 0x00, 0x00, 0x00, 0x00, 0x11, 0x00, 0x02, 0x00, 0x01, 0x00, 0x00, 0x00, 0x11, 0x00, 0x02, 0x00, | ||
| 121 | 0x51, 0x11, 0x00, 0x00, 0x11, 0x00, 0x02, 0x00, 0x61, 0x11, 0x00, 0x00, 0x0a, 0x00, 0x07, 0x00, | ||
| 122 | 0x53, 0x50, 0x56, 0x5f, 0x4b, 0x48, 0x52, 0x5f, 0x31, 0x36, 0x62, 0x69, 0x74, 0x5f, 0x73, 0x74, | ||
| 123 | 0x6f, 0x72, 0x61, 0x67, 0x65, 0x00, 0x00, 0x00, 0x0a, 0x00, 0x07, 0x00, 0x53, 0x50, 0x56, 0x5f, | ||
| 124 | 0x4b, 0x48, 0x52, 0x5f, 0x38, 0x62, 0x69, 0x74, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, | ||
| 125 | 0x00, 0x00, 0x00, 0x00, 0x0b, 0x00, 0x06, 0x00, 0x01, 0x00, 0x00, 0x00, 0x47, 0x4c, 0x53, 0x4c, | ||
| 126 | 0x2e, 0x73, 0x74, 0x64, 0x2e, 0x34, 0x35, 0x30, 0x00, 0x00, 0x00, 0x00, 0x0e, 0x00, 0x03, 0x00, | ||
| 127 | 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x0f, 0x00, 0x06, 0x00, 0x05, 0x00, 0x00, 0x00, | ||
| 128 | 0x04, 0x00, 0x00, 0x00, 0x6d, 0x61, 0x69, 0x6e, 0x00, 0x00, 0x00, 0x00, 0x0b, 0x00, 0x00, 0x00, | ||
| 129 | 0x10, 0x00, 0x06, 0x00, 0x04, 0x00, 0x00, 0x00, 0x11, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, | ||
| 130 | 0x01, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x47, 0x00, 0x04, 0x00, 0x0b, 0x00, 0x00, 0x00, | ||
| 131 | 0x0b, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, 0x47, 0x00, 0x04, 0x00, 0x12, 0x00, 0x00, 0x00, | ||
| 132 | 0x06, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x48, 0x00, 0x04, 0x00, 0x13, 0x00, 0x00, 0x00, | ||
| 133 | 0x00, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, 0x48, 0x00, 0x05, 0x00, 0x13, 0x00, 0x00, 0x00, | ||
| 134 | 0x00, 0x00, 0x00, 0x00, 0x23, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x47, 0x00, 0x03, 0x00, | ||
| 135 | 0x13, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x47, 0x00, 0x04, 0x00, 0x15, 0x00, 0x00, 0x00, | ||
| 136 | 0x22, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x47, 0x00, 0x04, 0x00, 0x15, 0x00, 0x00, 0x00, | ||
| 137 | 0x21, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x47, 0x00, 0x04, 0x00, 0x1f, 0x00, 0x00, 0x00, | ||
| 138 | 0x06, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x48, 0x00, 0x04, 0x00, 0x20, 0x00, 0x00, 0x00, | ||
| 139 | 0x00, 0x00, 0x00, 0x00, 0x19, 0x00, 0x00, 0x00, 0x48, 0x00, 0x05, 0x00, 0x20, 0x00, 0x00, 0x00, | ||
| 140 | 0x00, 0x00, 0x00, 0x00, 0x23, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x47, 0x00, 0x03, 0x00, | ||
| 141 | 0x20, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x47, 0x00, 0x04, 0x00, 0x22, 0x00, 0x00, 0x00, | ||
| 142 | 0x22, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x47, 0x00, 0x04, 0x00, 0x22, 0x00, 0x00, 0x00, | ||
| 143 | 0x21, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x47, 0x00, 0x04, 0x00, 0x2e, 0x00, 0x00, 0x00, | ||
| 144 | 0x0b, 0x00, 0x00, 0x00, 0x19, 0x00, 0x00, 0x00, 0x13, 0x00, 0x02, 0x00, 0x02, 0x00, 0x00, 0x00, | ||
| 145 | 0x21, 0x00, 0x03, 0x00, 0x03, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x15, 0x00, 0x04, 0x00, | ||
| 146 | 0x06, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x20, 0x00, 0x04, 0x00, | ||
| 147 | 0x07, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x17, 0x00, 0x04, 0x00, | ||
| 148 | 0x09, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x20, 0x00, 0x04, 0x00, | ||
| 149 | 0x0a, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x09, 0x00, 0x00, 0x00, 0x3b, 0x00, 0x04, 0x00, | ||
| 150 | 0x0a, 0x00, 0x00, 0x00, 0x0b, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x2b, 0x00, 0x04, 0x00, | ||
| 151 | 0x06, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x20, 0x00, 0x04, 0x00, | ||
| 152 | 0x0d, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x15, 0x00, 0x04, 0x00, | ||
| 153 | 0x11, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x1d, 0x00, 0x03, 0x00, | ||
| 154 | 0x12, 0x00, 0x00, 0x00, 0x11, 0x00, 0x00, 0x00, 0x1e, 0x00, 0x03, 0x00, 0x13, 0x00, 0x00, 0x00, | ||
| 155 | 0x12, 0x00, 0x00, 0x00, 0x20, 0x00, 0x04, 0x00, 0x14, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, | ||
| 156 | 0x13, 0x00, 0x00, 0x00, 0x3b, 0x00, 0x04, 0x00, 0x14, 0x00, 0x00, 0x00, 0x15, 0x00, 0x00, 0x00, | ||
| 157 | 0x02, 0x00, 0x00, 0x00, 0x15, 0x00, 0x04, 0x00, 0x17, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, | ||
| 158 | 0x01, 0x00, 0x00, 0x00, 0x14, 0x00, 0x02, 0x00, 0x1a, 0x00, 0x00, 0x00, 0x15, 0x00, 0x04, 0x00, | ||
| 159 | 0x1e, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x1d, 0x00, 0x03, 0x00, | ||
| 160 | 0x1f, 0x00, 0x00, 0x00, 0x1e, 0x00, 0x00, 0x00, 0x1e, 0x00, 0x03, 0x00, 0x20, 0x00, 0x00, 0x00, | ||
| 161 | 0x1f, 0x00, 0x00, 0x00, 0x20, 0x00, 0x04, 0x00, 0x21, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, | ||
| 162 | 0x20, 0x00, 0x00, 0x00, 0x3b, 0x00, 0x04, 0x00, 0x21, 0x00, 0x00, 0x00, 0x22, 0x00, 0x00, 0x00, | ||
| 163 | 0x02, 0x00, 0x00, 0x00, 0x2b, 0x00, 0x04, 0x00, 0x17, 0x00, 0x00, 0x00, 0x23, 0x00, 0x00, 0x00, | ||
| 164 | 0x00, 0x00, 0x00, 0x00, 0x20, 0x00, 0x04, 0x00, 0x26, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, | ||
| 165 | 0x11, 0x00, 0x00, 0x00, 0x20, 0x00, 0x04, 0x00, 0x2a, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, | ||
| 166 | 0x1e, 0x00, 0x00, 0x00, 0x2b, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x2c, 0x00, 0x00, 0x00, | ||
| 167 | 0x00, 0x04, 0x00, 0x00, 0x2b, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x2d, 0x00, 0x00, 0x00, | ||
| 168 | 0x01, 0x00, 0x00, 0x00, 0x2c, 0x00, 0x06, 0x00, 0x09, 0x00, 0x00, 0x00, 0x2e, 0x00, 0x00, 0x00, | ||
| 169 | 0x2c, 0x00, 0x00, 0x00, 0x2d, 0x00, 0x00, 0x00, 0x2d, 0x00, 0x00, 0x00, 0x36, 0x00, 0x05, 0x00, | ||
| 170 | 0x02, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, | ||
| 171 | 0xf8, 0x00, 0x02, 0x00, 0x05, 0x00, 0x00, 0x00, 0x3b, 0x00, 0x04, 0x00, 0x07, 0x00, 0x00, 0x00, | ||
| 172 | 0x08, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x41, 0x00, 0x05, 0x00, 0x0d, 0x00, 0x00, 0x00, | ||
| 173 | 0x0e, 0x00, 0x00, 0x00, 0x0b, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x00, 0x00, 0x3d, 0x00, 0x04, 0x00, | ||
| 174 | 0x06, 0x00, 0x00, 0x00, 0x0f, 0x00, 0x00, 0x00, 0x0e, 0x00, 0x00, 0x00, 0x3e, 0x00, 0x03, 0x00, | ||
| 175 | 0x08, 0x00, 0x00, 0x00, 0x0f, 0x00, 0x00, 0x00, 0x3d, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, | ||
| 176 | 0x10, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x44, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, | ||
| 177 | 0x16, 0x00, 0x00, 0x00, 0x15, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x7c, 0x00, 0x04, 0x00, | ||
| 178 | 0x17, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, 0x16, 0x00, 0x00, 0x00, 0x7c, 0x00, 0x04, 0x00, | ||
| 179 | 0x06, 0x00, 0x00, 0x00, 0x19, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, 0xb0, 0x00, 0x05, 0x00, | ||
| 180 | 0x1a, 0x00, 0x00, 0x00, 0x1b, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x19, 0x00, 0x00, 0x00, | ||
| 181 | 0xf7, 0x00, 0x03, 0x00, 0x1d, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xfa, 0x00, 0x04, 0x00, | ||
| 182 | 0x1b, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, 0x1d, 0x00, 0x00, 0x00, 0xf8, 0x00, 0x02, 0x00, | ||
| 183 | 0x1c, 0x00, 0x00, 0x00, 0x3d, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x24, 0x00, 0x00, 0x00, | ||
| 184 | 0x08, 0x00, 0x00, 0x00, 0x3d, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x25, 0x00, 0x00, 0x00, | ||
| 185 | 0x08, 0x00, 0x00, 0x00, 0x41, 0x00, 0x06, 0x00, 0x26, 0x00, 0x00, 0x00, 0x27, 0x00, 0x00, 0x00, | ||
| 186 | 0x15, 0x00, 0x00, 0x00, 0x23, 0x00, 0x00, 0x00, 0x25, 0x00, 0x00, 0x00, 0x3d, 0x00, 0x04, 0x00, | ||
| 187 | 0x11, 0x00, 0x00, 0x00, 0x28, 0x00, 0x00, 0x00, 0x27, 0x00, 0x00, 0x00, 0x71, 0x00, 0x04, 0x00, | ||
| 188 | 0x1e, 0x00, 0x00, 0x00, 0x29, 0x00, 0x00, 0x00, 0x28, 0x00, 0x00, 0x00, 0x41, 0x00, 0x06, 0x00, | ||
| 189 | 0x2a, 0x00, 0x00, 0x00, 0x2b, 0x00, 0x00, 0x00, 0x22, 0x00, 0x00, 0x00, 0x23, 0x00, 0x00, 0x00, | ||
| 190 | 0x24, 0x00, 0x00, 0x00, 0x3e, 0x00, 0x03, 0x00, 0x2b, 0x00, 0x00, 0x00, 0x29, 0x00, 0x00, 0x00, | ||
| 191 | 0xf9, 0x00, 0x02, 0x00, 0x1d, 0x00, 0x00, 0x00, 0xf8, 0x00, 0x02, 0x00, 0x1d, 0x00, 0x00, 0x00, | ||
| 192 | 0xfd, 0x00, 0x01, 0x00, 0x38, 0x00, 0x01, 0x00}; | ||
| 193 | |||
| 194 | } // Anonymous namespace | ||
| 195 | |||
| 196 | VKComputePass::VKComputePass(const VKDevice& device, VKDescriptorPool& descriptor_pool, | ||
| 197 | const std::vector<vk::DescriptorSetLayoutBinding>& bindings, | ||
| 198 | const std::vector<vk::DescriptorUpdateTemplateEntry>& templates, | ||
| 199 | const std::vector<vk::PushConstantRange> push_constants, | ||
| 200 | std::size_t code_size, const u8* code) { | ||
| 201 | const auto dev = device.GetLogical(); | ||
| 202 | const auto& dld = device.GetDispatchLoader(); | ||
| 203 | |||
| 204 | const vk::DescriptorSetLayoutCreateInfo descriptor_layout_ci( | ||
| 205 | {}, static_cast<u32>(bindings.size()), bindings.data()); | ||
| 206 | descriptor_set_layout = dev.createDescriptorSetLayoutUnique(descriptor_layout_ci, nullptr, dld); | ||
| 207 | |||
| 208 | const vk::PipelineLayoutCreateInfo pipeline_layout_ci({}, 1, &*descriptor_set_layout, | ||
| 209 | static_cast<u32>(push_constants.size()), | ||
| 210 | push_constants.data()); | ||
| 211 | layout = dev.createPipelineLayoutUnique(pipeline_layout_ci, nullptr, dld); | ||
| 212 | |||
| 213 | if (!templates.empty()) { | ||
| 214 | const vk::DescriptorUpdateTemplateCreateInfo template_ci( | ||
| 215 | {}, static_cast<u32>(templates.size()), templates.data(), | ||
| 216 | vk::DescriptorUpdateTemplateType::eDescriptorSet, *descriptor_set_layout, | ||
| 217 | vk::PipelineBindPoint::eGraphics, *layout, 0); | ||
| 218 | descriptor_template = dev.createDescriptorUpdateTemplateUnique(template_ci, nullptr, dld); | ||
| 219 | |||
| 220 | descriptor_allocator.emplace(descriptor_pool, *descriptor_set_layout); | ||
| 221 | } | ||
| 222 | |||
| 223 | auto code_copy = std::make_unique<u32[]>(code_size / sizeof(u32) + 1); | ||
| 224 | std::memcpy(code_copy.get(), code, code_size); | ||
| 225 | const vk::ShaderModuleCreateInfo module_ci({}, code_size, code_copy.get()); | ||
| 226 | module = dev.createShaderModuleUnique(module_ci, nullptr, dld); | ||
| 227 | |||
| 228 | const vk::PipelineShaderStageCreateInfo stage_ci({}, vk::ShaderStageFlagBits::eCompute, *module, | ||
| 229 | "main", nullptr); | ||
| 230 | |||
| 231 | const vk::ComputePipelineCreateInfo pipeline_ci({}, stage_ci, *layout, nullptr, 0); | ||
| 232 | pipeline = dev.createComputePipelineUnique(nullptr, pipeline_ci, nullptr, dld); | ||
| 233 | } | ||
| 234 | |||
| 235 | VKComputePass::~VKComputePass() = default; | ||
| 236 | |||
| 237 | vk::DescriptorSet VKComputePass::CommitDescriptorSet( | ||
| 238 | VKUpdateDescriptorQueue& update_descriptor_queue, VKFence& fence) { | ||
| 239 | if (!descriptor_template) { | ||
| 240 | return {}; | ||
| 241 | } | ||
| 242 | const auto set = descriptor_allocator->Commit(fence); | ||
| 243 | update_descriptor_queue.Send(*descriptor_template, set); | ||
| 244 | return set; | ||
| 245 | } | ||
| 246 | |||
| 247 | QuadArrayPass::QuadArrayPass(const VKDevice& device, VKScheduler& scheduler, | ||
| 248 | VKDescriptorPool& descriptor_pool, | ||
| 249 | VKStagingBufferPool& staging_buffer_pool, | ||
| 250 | VKUpdateDescriptorQueue& update_descriptor_queue) | ||
| 251 | : VKComputePass(device, descriptor_pool, | ||
| 252 | {vk::DescriptorSetLayoutBinding(0, vk::DescriptorType::eStorageBuffer, 1, | ||
| 253 | vk::ShaderStageFlagBits::eCompute, nullptr)}, | ||
| 254 | {vk::DescriptorUpdateTemplateEntry(0, 0, 1, vk::DescriptorType::eStorageBuffer, | ||
| 255 | 0, sizeof(DescriptorUpdateEntry))}, | ||
| 256 | {vk::PushConstantRange(vk::ShaderStageFlagBits::eCompute, 0, sizeof(u32))}, | ||
| 257 | std::size(quad_array), quad_array), | ||
| 258 | scheduler{scheduler}, staging_buffer_pool{staging_buffer_pool}, | ||
| 259 | update_descriptor_queue{update_descriptor_queue} {} | ||
| 260 | |||
| 261 | QuadArrayPass::~QuadArrayPass() = default; | ||
| 262 | |||
| 263 | std::pair<const vk::Buffer&, vk::DeviceSize> QuadArrayPass::Assemble(u32 num_vertices, u32 first) { | ||
| 264 | const u32 num_triangle_vertices = num_vertices * 6 / 4; | ||
| 265 | const std::size_t staging_size = num_triangle_vertices * sizeof(u32); | ||
| 266 | auto& buffer = staging_buffer_pool.GetUnusedBuffer(staging_size, false); | ||
| 267 | |||
| 268 | update_descriptor_queue.Acquire(); | ||
| 269 | update_descriptor_queue.AddBuffer(&*buffer.handle, 0, staging_size); | ||
| 270 | const auto set = CommitDescriptorSet(update_descriptor_queue, scheduler.GetFence()); | ||
| 271 | |||
| 272 | scheduler.RequestOutsideRenderPassOperationContext(); | ||
| 273 | |||
| 274 | ASSERT(num_vertices % 4 == 0); | ||
| 275 | const u32 num_quads = num_vertices / 4; | ||
| 276 | scheduler.Record([layout = *layout, pipeline = *pipeline, buffer = *buffer.handle, num_quads, | ||
| 277 | first, set](auto cmdbuf, auto& dld) { | ||
| 278 | constexpr u32 dispatch_size = 1024; | ||
| 279 | cmdbuf.bindPipeline(vk::PipelineBindPoint::eCompute, pipeline, dld); | ||
| 280 | cmdbuf.bindDescriptorSets(vk::PipelineBindPoint::eCompute, layout, 0, {set}, {}, dld); | ||
| 281 | cmdbuf.pushConstants(layout, vk::ShaderStageFlagBits::eCompute, 0, sizeof(first), &first, | ||
| 282 | dld); | ||
| 283 | cmdbuf.dispatch(Common::AlignUp(num_quads, dispatch_size) / dispatch_size, 1, 1, dld); | ||
| 284 | |||
| 285 | const vk::BufferMemoryBarrier barrier( | ||
| 286 | vk::AccessFlagBits::eShaderWrite, vk::AccessFlagBits::eVertexAttributeRead, | ||
| 287 | VK_QUEUE_FAMILY_IGNORED, VK_QUEUE_FAMILY_IGNORED, buffer, 0, | ||
| 288 | static_cast<vk::DeviceSize>(num_quads) * 6 * sizeof(u32)); | ||
| 289 | cmdbuf.pipelineBarrier(vk::PipelineStageFlagBits::eComputeShader, | ||
| 290 | vk::PipelineStageFlagBits::eVertexInput, {}, {}, {barrier}, {}, dld); | ||
| 291 | }); | ||
| 292 | return {*buffer.handle, 0}; | ||
| 293 | } | ||
| 294 | |||
| 295 | Uint8Pass::Uint8Pass(const VKDevice& device, VKScheduler& scheduler, | ||
| 296 | VKDescriptorPool& descriptor_pool, VKStagingBufferPool& staging_buffer_pool, | ||
| 297 | VKUpdateDescriptorQueue& update_descriptor_queue) | ||
| 298 | : VKComputePass(device, descriptor_pool, | ||
| 299 | {vk::DescriptorSetLayoutBinding(0, vk::DescriptorType::eStorageBuffer, 1, | ||
| 300 | vk::ShaderStageFlagBits::eCompute, nullptr), | ||
| 301 | vk::DescriptorSetLayoutBinding(1, vk::DescriptorType::eStorageBuffer, 1, | ||
| 302 | vk::ShaderStageFlagBits::eCompute, nullptr)}, | ||
| 303 | {vk::DescriptorUpdateTemplateEntry(0, 0, 2, vk::DescriptorType::eStorageBuffer, | ||
| 304 | 0, sizeof(DescriptorUpdateEntry))}, | ||
| 305 | {}, std::size(uint8_pass), uint8_pass), | ||
| 306 | scheduler{scheduler}, staging_buffer_pool{staging_buffer_pool}, | ||
| 307 | update_descriptor_queue{update_descriptor_queue} {} | ||
| 308 | |||
| 309 | Uint8Pass::~Uint8Pass() = default; | ||
| 310 | |||
| 311 | std::pair<const vk::Buffer*, u64> Uint8Pass::Assemble(u32 num_vertices, vk::Buffer src_buffer, | ||
| 312 | u64 src_offset) { | ||
| 313 | const auto staging_size = static_cast<u32>(num_vertices * sizeof(u16)); | ||
| 314 | auto& buffer = staging_buffer_pool.GetUnusedBuffer(staging_size, false); | ||
| 315 | |||
| 316 | update_descriptor_queue.Acquire(); | ||
| 317 | update_descriptor_queue.AddBuffer(&src_buffer, src_offset, num_vertices); | ||
| 318 | update_descriptor_queue.AddBuffer(&*buffer.handle, 0, staging_size); | ||
| 319 | const auto set = CommitDescriptorSet(update_descriptor_queue, scheduler.GetFence()); | ||
| 320 | |||
| 321 | scheduler.RequestOutsideRenderPassOperationContext(); | ||
| 322 | scheduler.Record([layout = *layout, pipeline = *pipeline, buffer = *buffer.handle, set, | ||
| 323 | num_vertices](auto cmdbuf, auto& dld) { | ||
| 324 | constexpr u32 dispatch_size = 1024; | ||
| 325 | cmdbuf.bindPipeline(vk::PipelineBindPoint::eCompute, pipeline, dld); | ||
| 326 | cmdbuf.bindDescriptorSets(vk::PipelineBindPoint::eCompute, layout, 0, {set}, {}, dld); | ||
| 327 | cmdbuf.dispatch(Common::AlignUp(num_vertices, dispatch_size) / dispatch_size, 1, 1, dld); | ||
| 328 | |||
| 329 | const vk::BufferMemoryBarrier barrier( | ||
| 330 | vk::AccessFlagBits::eShaderWrite, vk::AccessFlagBits::eVertexAttributeRead, | ||
| 331 | VK_QUEUE_FAMILY_IGNORED, VK_QUEUE_FAMILY_IGNORED, buffer, 0, | ||
| 332 | static_cast<vk::DeviceSize>(num_vertices) * sizeof(u16)); | ||
| 333 | cmdbuf.pipelineBarrier(vk::PipelineStageFlagBits::eComputeShader, | ||
| 334 | vk::PipelineStageFlagBits::eVertexInput, {}, {}, {barrier}, {}, dld); | ||
| 335 | }); | ||
| 336 | return {&*buffer.handle, 0}; | ||
| 337 | } | ||
| 338 | |||
| 339 | } // namespace Vulkan | ||
diff --git a/src/video_core/renderer_vulkan/vk_compute_pass.h b/src/video_core/renderer_vulkan/vk_compute_pass.h new file mode 100644 index 000000000..7057eb837 --- /dev/null +++ b/src/video_core/renderer_vulkan/vk_compute_pass.h | |||
| @@ -0,0 +1,77 @@ | |||
| 1 | // Copyright 2019 yuzu Emulator Project | ||
| 2 | // Licensed under GPLv2 or any later version | ||
| 3 | // Refer to the license.txt file included. | ||
| 4 | |||
| 5 | #pragma once | ||
| 6 | |||
| 7 | #include <optional> | ||
| 8 | #include <utility> | ||
| 9 | #include <vector> | ||
| 10 | #include "common/common_types.h" | ||
| 11 | #include "video_core/renderer_vulkan/declarations.h" | ||
| 12 | #include "video_core/renderer_vulkan/vk_descriptor_pool.h" | ||
| 13 | |||
| 14 | namespace Vulkan { | ||
| 15 | |||
| 16 | class VKDevice; | ||
| 17 | class VKFence; | ||
| 18 | class VKScheduler; | ||
| 19 | class VKStagingBufferPool; | ||
| 20 | class VKUpdateDescriptorQueue; | ||
| 21 | |||
| 22 | class VKComputePass { | ||
| 23 | public: | ||
| 24 | explicit VKComputePass(const VKDevice& device, VKDescriptorPool& descriptor_pool, | ||
| 25 | const std::vector<vk::DescriptorSetLayoutBinding>& bindings, | ||
| 26 | const std::vector<vk::DescriptorUpdateTemplateEntry>& templates, | ||
| 27 | const std::vector<vk::PushConstantRange> push_constants, | ||
| 28 | std::size_t code_size, const u8* code); | ||
| 29 | ~VKComputePass(); | ||
| 30 | |||
| 31 | protected: | ||
| 32 | vk::DescriptorSet CommitDescriptorSet(VKUpdateDescriptorQueue& update_descriptor_queue, | ||
| 33 | VKFence& fence); | ||
| 34 | |||
| 35 | UniqueDescriptorUpdateTemplate descriptor_template; | ||
| 36 | UniquePipelineLayout layout; | ||
| 37 | UniquePipeline pipeline; | ||
| 38 | |||
| 39 | private: | ||
| 40 | UniqueDescriptorSetLayout descriptor_set_layout; | ||
| 41 | std::optional<DescriptorAllocator> descriptor_allocator; | ||
| 42 | UniqueShaderModule module; | ||
| 43 | }; | ||
| 44 | |||
| 45 | class QuadArrayPass final : public VKComputePass { | ||
| 46 | public: | ||
| 47 | explicit QuadArrayPass(const VKDevice& device, VKScheduler& scheduler, | ||
| 48 | VKDescriptorPool& descriptor_pool, | ||
| 49 | VKStagingBufferPool& staging_buffer_pool, | ||
| 50 | VKUpdateDescriptorQueue& update_descriptor_queue); | ||
| 51 | ~QuadArrayPass(); | ||
| 52 | |||
| 53 | std::pair<const vk::Buffer&, vk::DeviceSize> Assemble(u32 num_vertices, u32 first); | ||
| 54 | |||
| 55 | private: | ||
| 56 | VKScheduler& scheduler; | ||
| 57 | VKStagingBufferPool& staging_buffer_pool; | ||
| 58 | VKUpdateDescriptorQueue& update_descriptor_queue; | ||
| 59 | }; | ||
| 60 | |||
| 61 | class Uint8Pass final : public VKComputePass { | ||
| 62 | public: | ||
| 63 | explicit Uint8Pass(const VKDevice& device, VKScheduler& scheduler, | ||
| 64 | VKDescriptorPool& descriptor_pool, VKStagingBufferPool& staging_buffer_pool, | ||
| 65 | VKUpdateDescriptorQueue& update_descriptor_queue); | ||
| 66 | ~Uint8Pass(); | ||
| 67 | |||
| 68 | std::pair<const vk::Buffer*, u64> Assemble(u32 num_vertices, vk::Buffer src_buffer, | ||
| 69 | u64 src_offset); | ||
| 70 | |||
| 71 | private: | ||
| 72 | VKScheduler& scheduler; | ||
| 73 | VKStagingBufferPool& staging_buffer_pool; | ||
| 74 | VKUpdateDescriptorQueue& update_descriptor_queue; | ||
| 75 | }; | ||
| 76 | |||
| 77 | } // namespace Vulkan | ||
diff --git a/src/video_core/renderer_vulkan/vk_shader_util.cpp b/src/video_core/renderer_vulkan/vk_shader_util.cpp new file mode 100644 index 000000000..b97c4cb3d --- /dev/null +++ b/src/video_core/renderer_vulkan/vk_shader_util.cpp | |||
| @@ -0,0 +1,34 @@ | |||
| 1 | // Copyright 2018 yuzu Emulator Project | ||
| 2 | // Licensed under GPLv2 or any later version | ||
| 3 | // Refer to the license.txt file included. | ||
| 4 | |||
| 5 | #include <cstring> | ||
| 6 | #include <memory> | ||
| 7 | #include <vector> | ||
| 8 | #include "common/alignment.h" | ||
| 9 | #include "common/assert.h" | ||
| 10 | #include "common/common_types.h" | ||
| 11 | #include "video_core/renderer_vulkan/declarations.h" | ||
| 12 | #include "video_core/renderer_vulkan/vk_device.h" | ||
| 13 | #include "video_core/renderer_vulkan/vk_shader_util.h" | ||
| 14 | |||
| 15 | namespace Vulkan { | ||
| 16 | |||
| 17 | UniqueShaderModule BuildShader(const VKDevice& device, std::size_t code_size, const u8* code_data) { | ||
| 18 | // Avoid undefined behavior by copying to a staging allocation | ||
| 19 | ASSERT(code_size % sizeof(u32) == 0); | ||
| 20 | const auto data = std::make_unique<u32[]>(code_size / sizeof(u32)); | ||
| 21 | std::memcpy(data.get(), code_data, code_size); | ||
| 22 | |||
| 23 | const auto dev = device.GetLogical(); | ||
| 24 | const auto& dld = device.GetDispatchLoader(); | ||
| 25 | const vk::ShaderModuleCreateInfo shader_ci({}, code_size, data.get()); | ||
| 26 | vk::ShaderModule shader_module; | ||
| 27 | if (dev.createShaderModule(&shader_ci, nullptr, &shader_module, dld) != vk::Result::eSuccess) { | ||
| 28 | UNREACHABLE_MSG("Shader module failed to build!"); | ||
| 29 | } | ||
| 30 | |||
| 31 | return UniqueShaderModule(shader_module, vk::ObjectDestroy(dev, nullptr, dld)); | ||
| 32 | } | ||
| 33 | |||
| 34 | } // namespace Vulkan | ||
diff --git a/src/video_core/renderer_vulkan/vk_shader_util.h b/src/video_core/renderer_vulkan/vk_shader_util.h new file mode 100644 index 000000000..c06d65970 --- /dev/null +++ b/src/video_core/renderer_vulkan/vk_shader_util.h | |||
| @@ -0,0 +1,17 @@ | |||
| 1 | // Copyright 2018 yuzu Emulator Project | ||
| 2 | // Licensed under GPLv2 or any later version | ||
| 3 | // Refer to the license.txt file included. | ||
| 4 | |||
| 5 | #pragma once | ||
| 6 | |||
| 7 | #include <vector> | ||
| 8 | #include "common/common_types.h" | ||
| 9 | #include "video_core/renderer_vulkan/declarations.h" | ||
| 10 | |||
| 11 | namespace Vulkan { | ||
| 12 | |||
| 13 | class VKDevice; | ||
| 14 | |||
| 15 | UniqueShaderModule BuildShader(const VKDevice& device, std::size_t code_size, const u8* code_data); | ||
| 16 | |||
| 17 | } // namespace Vulkan | ||