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

Reduce usage of waitIdle

Browse source
This commit is contained in:
0cc4m 2023-07-22 18:25:07 +02:00
parent 3452095089
commit f2d4ca34bf
1 changed files with 16 additions and 6 deletions
Download patch file Download diff file Expand all files Collapse all files

22
ggml-vulkan.cpp
View file

@ -380,7 +380,7 @@ static void ggml_vk_queue_cleanup(vk_queue& q) {
#ifdef VK_DEBUG
std::cerr << "ggml_vk_queue_cleanup()" << std::endl;
#endif
q.queue.waitIdle();
// Requires semaphores and command buffers to be done
q.semaphore_idx = 0;
@ -702,7 +702,7 @@ void ggml_vk_init(void) {
static vk_pipeline* ggml_vk_get_to_fp32(ggml_type type) {
#ifdef VK_DEBUG
std::cerr << "ggml_vk_get_to_fp32_vk()" << std::endl;
std::cerr << "ggml_vk_get_to_fp32()" << std::endl;
#endif
switch (type) {
case GGML_TYPE_Q4_0:
@ -723,6 +723,9 @@ static vk_pipeline* ggml_vk_get_to_fp32(ggml_type type) {
}
static vk_pipeline* ggml_vk_get_dequantize_mul_mat_vec(ggml_type type) {
#ifdef VK_DEBUG
std::cerr << "ggml_vk_get_dequantize_mul_mat_vec()" << std::endl;
#endif
switch (type) {
case GGML_TYPE_Q4_0:
return &vk_pipeline_dequant_mul_mat_vec_q4_0;
@ -828,6 +831,9 @@ static void ggml_vk_pool_free(vk_buffer& buffer) {
}
void ggml_vk_free_data(const struct ggml_tensor* tensor) {
#ifdef VK_DEBUG
std::cerr << "ggml_vk_free_data(" << tensor << ")" << std::endl;
#endif
if (tensor->backend != GGML_BACKEND_GPU) {
return;
}
@ -1452,7 +1458,8 @@ static void ggml_vk_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
ggml_vk_submit(vk_transfer_queues[0], transfer_0_seqs, VK_NULL_HANDLE);
// cleanup waits for the queue to be done
vk_transfer_queues[0].queue.waitIdle();
ggml_vk_queue_cleanup(vk_transfer_queues[0]);
ggml_vk_queue_cleanup(vk_transfer_queues[1]);
ggml_vk_queue_cleanup(vk_compute_queue);
@ -1605,7 +1612,8 @@ static void ggml_vk_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
ggml_vk_submit(vk_transfer_queues[0], transfer_0_seqs, VK_NULL_HANDLE);
// cleanup waits for the queue to be done
vk_transfer_queues[0].queue.waitIdle();
ggml_vk_queue_cleanup(vk_transfer_queues[0]);
ggml_vk_queue_cleanup(vk_transfer_queues[1]);
ggml_vk_queue_cleanup(vk_compute_queue);
@ -1781,7 +1789,8 @@ static void ggml_vk_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
ggml_vk_submit(vk_transfer_queues[0], transfer_0_seqs, VK_NULL_HANDLE);
// cleanup waits for the queue to be done
vk_transfer_queues[0].queue.waitIdle();
ggml_vk_queue_cleanup(vk_transfer_queues[0]);
ggml_vk_queue_cleanup(vk_transfer_queues[1]);
ggml_vk_queue_cleanup(vk_compute_queue);
@ -2000,7 +2009,8 @@ static void ggml_vk_mul_f32(const ggml_tensor * src0, const ggml_tensor * src1,
}
}
// cleanup waits for the queue to be done
vk_transfer_queues[1].queue.waitIdle();
ggml_vk_queue_cleanup(vk_transfer_queues[0]);
ggml_vk_queue_cleanup(vk_transfer_queues[1]);
ggml_vk_queue_cleanup(vk_compute_queue);