vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Add multi-submit for command buffers

Browse source
This commit is contained in:
0cc4m 2023-09-30 10:32:11 +02:00
parent 42bfa889a6
commit da09a02b81
1 changed files with 7 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

9
ggml-vulkan.cpp
View file

@ -50,8 +50,6 @@
#define VK_DEVICE_DESCRIPTOR_POOL_MODE_MULTI 1 #define VK_DEVICE_DESCRIPTOR_POOL_MODE_MULTI 1
#define VK_DEVICE_DESCRIPTOR_POOL_MODE_SINGLE 2 #define VK_DEVICE_DESCRIPTOR_POOL_MODE_SINGLE 2
#define VK_SUBMIT_BATCH 3
#define VK_NUM_TYPES 16 #define VK_NUM_TYPES 16
#ifndef K_QUANTS_PER_ITERATION #ifndef K_QUANTS_PER_ITERATION
@ -166,6 +164,7 @@ struct vk_staging_memcpy {
}; };
struct ggml_vk_tensor_extra_gpu { struct ggml_vk_tensor_extra_gpu {
uint32_t batch_size;
std::vector<uint32_t> buffer_idx; std::vector<uint32_t> buffer_idx;
std::vector<vk_staging_memcpy> memcpys; std::vector<vk_staging_memcpy> memcpys;
@ -1243,10 +1242,14 @@ void ggml_vk_host_free(void* ptr) {
vk_pinned_memory.erase(vk_pinned_memory.begin() + index); vk_pinned_memory.erase(vk_pinned_memory.begin() + index);
} }
static vk_submission ggml_vk_begin_submission(vk_queue& q) { static vk_submission ggml_vk_begin_submission(vk_queue& q, bool one_time = true) {
vk_submission s; vk_submission s;
s.buffer = ggml_vk_create_cmd_buffer(q); s.buffer = ggml_vk_create_cmd_buffer(q);
if (one_time) {
s.buffer.begin({ vk::CommandBufferUsageFlagBits::eOneTimeSubmit }); s.buffer.begin({ vk::CommandBufferUsageFlagBits::eOneTimeSubmit });
} else {
s.buffer.begin({ vk::CommandBufferUsageFlags{} });
}
return s; return s;
} }