vrr/prima.cpp
2
0
Fork 0
mirror of https://github.com/Lizonghang/prima.cpp.git synced 2025-09-06 13:39:02 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

add EPS in device_compute_delay

Browse source
This commit is contained in:
Lizonghang 2024-12-29 22:31:45 +04:00
parent fa210d2034
commit d9beb030ee
2 changed files with 44 additions and 44 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

48
common/profiler.cpp
View file

@ -1239,29 +1239,29 @@ static float device_compute_delay(struct device_info & dev_info, int n_layers, c
#ifdef GGML_USE_CUDA
struct gpu_props gpu = dev_info.gpu_props;
gpu_latency_per_layer += (double)n_flops.layer_f32_f32 / (double)gpu.cuda_flops_f32_f32 / 1e9;
gpu_latency_per_layer += (double)n_flops.layer_f16_f32 / (double)gpu.cuda_flops_f16_f32 / 1e9;
gpu_latency_per_layer += (double)n_flops.layer_q4k_f32 / (double)gpu.cuda_flops_q4k_f32 / 1e9;
gpu_latency_per_layer += (double)n_flops.layer_q5k_f32 / (double)gpu.cuda_flops_q5k_f32 / 1e9;
gpu_latency_per_layer += (double)n_flops.layer_q6k_f32 / (double)gpu.cuda_flops_q6k_f32 / 1e9;
gpu_latency_per_layer += (double)n_flops.layer_q80_f32 / (double)gpu.cuda_flops_q80_f32 / 1e9;
gpu_latency_per_layer += (double)n_flops.layer_f32_f32 / ((double)gpu.cuda_flops_f32_f32 + EPS) / 1e9;
gpu_latency_per_layer += (double)n_flops.layer_f16_f32 / ((double)gpu.cuda_flops_f16_f32 + EPS) / 1e9;
gpu_latency_per_layer += (double)n_flops.layer_q4k_f32 / ((double)gpu.cuda_flops_q4k_f32 + EPS) / 1e9;
gpu_latency_per_layer += (double)n_flops.layer_q5k_f32 / ((double)gpu.cuda_flops_q5k_f32 + EPS) / 1e9;
gpu_latency_per_layer += (double)n_flops.layer_q6k_f32 / ((double)gpu.cuda_flops_q6k_f32 + EPS) / 1e9;
gpu_latency_per_layer += (double)n_flops.layer_q80_f32 / ((double)gpu.cuda_flops_q80_f32 + EPS) / 1e9;
#elif GGML_USE_METAL
struct gpu_props gpu = dev_info.gpu_props;
gpu_latency_per_layer += (double)n_flops.layer_f32_f32 / (double)gpu.metal_flops_f32_f32 / 1e9;
gpu_latency_per_layer += (double)n_flops.layer_f16_f32 / (double)gpu.metal_flops_f16_f32 / 1e9;
gpu_latency_per_layer += (double)n_flops.layer_q4k_f32 / (double)gpu.metal_flops_q4k_f32 / 1e9;
gpu_latency_per_layer += (double)n_flops.layer_q5k_f32 / (double)gpu.metal_flops_q5k_f32 / 1e9;
gpu_latency_per_layer += (double)n_flops.layer_q6k_f32 / (double)gpu.metal_flops_q6k_f32 / 1e9;
gpu_latency_per_layer += (double)n_flops.layer_q80_f32 / (double)gpu.metal_flops_q80_f32 / 1e9;
gpu_latency_per_layer += (double)n_flops.layer_f32_f32 / ((double)gpu.metal_flops_f32_f32 + EPS) / 1e9;
gpu_latency_per_layer += (double)n_flops.layer_f16_f32 / ((double)gpu.metal_flops_f16_f32 + EPS) / 1e9;
gpu_latency_per_layer += (double)n_flops.layer_q4k_f32 / ((double)gpu.metal_flops_q4k_f32 + EPS) / 1e9;
gpu_latency_per_layer += (double)n_flops.layer_q5k_f32 / ((double)gpu.metal_flops_q5k_f32 + EPS) / 1e9;
gpu_latency_per_layer += (double)n_flops.layer_q6k_f32 / ((double)gpu.metal_flops_q6k_f32 + EPS) / 1e9;
gpu_latency_per_layer += (double)n_flops.layer_q80_f32 / ((double)gpu.metal_flops_q80_f32 + EPS) / 1e9;
#endif
cpu_latency_per_layer += (double)n_flops.layer_f32_f32 / (double)cpu.flops_f32_f32 / 1e9;
cpu_latency_per_layer += (double)n_flops.layer_f16_f32 / (double)cpu.flops_f16_f32 / 1e9;
cpu_latency_per_layer += (double)n_flops.layer_q4k_f32 / (double)cpu.flops_q4k_f32 / 1e9;
cpu_latency_per_layer += (double)n_flops.layer_q5k_f32 / (double)cpu.flops_q5k_f32 / 1e9;
cpu_latency_per_layer += (double)n_flops.layer_q6k_f32 / (double)cpu.flops_q6k_f32 / 1e9;
cpu_latency_per_layer += (double)n_flops.layer_q80_f32 / (double)cpu.flops_q80_f32 / 1e9;
cpu_latency_per_layer += (double)n_flops.layer_f32_f32 / ((double)cpu.flops_f32_f32 + EPS) / 1e9;
cpu_latency_per_layer += (double)n_flops.layer_f16_f32 / ((double)cpu.flops_f16_f32 + EPS) / 1e9;
cpu_latency_per_layer += (double)n_flops.layer_q4k_f32 / ((double)cpu.flops_q4k_f32 + EPS) / 1e9;
cpu_latency_per_layer += (double)n_flops.layer_q5k_f32 / ((double)cpu.flops_q5k_f32 + EPS) / 1e9;
cpu_latency_per_layer += (double)n_flops.layer_q6k_f32 / ((double)cpu.flops_q6k_f32 + EPS) / 1e9;
cpu_latency_per_layer += (double)n_flops.layer_q80_f32 / ((double)cpu.flops_q80_f32 + EPS) / 1e9;
double total_latency = 0.0f;
@ -1274,12 +1274,12 @@ static float device_compute_delay(struct device_info & dev_info, int n_layers, c
total_latency += cpu_latency_per_layer * n_layers;
#endif
total_latency += (double)n_flops.output_f32_f32 / (double)cpu.flops_f32_f32 / 1e9;
total_latency += (double)n_flops.output_f16_f32 / (double)cpu.flops_f16_f32 / 1e9;
total_latency += (double)n_flops.output_q4k_f32 / (double)cpu.flops_q4k_f32 / 1e9;
total_latency += (double)n_flops.output_q5k_f32 / (double)cpu.flops_q5k_f32 / 1e9;
total_latency += (double)n_flops.output_q6k_f32 / (double)cpu.flops_q6k_f32 / 1e9;
total_latency += (double)n_flops.output_q80_f32 / (double)cpu.flops_q80_f32 / 1e9;
total_latency += (double)n_flops.output_f32_f32 / ((double)cpu.flops_f32_f32 + EPS) / 1e9;
total_latency += (double)n_flops.output_f16_f32 / ((double)cpu.flops_f16_f32 + EPS) / 1e9;
total_latency += (double)n_flops.output_q4k_f32 / ((double)cpu.flops_q4k_f32 + EPS) / 1e9;
total_latency += (double)n_flops.output_q5k_f32 / ((double)cpu.flops_q5k_f32 + EPS) / 1e9;
total_latency += (double)n_flops.output_q6k_f32 / ((double)cpu.flops_q6k_f32 + EPS) / 1e9;
total_latency += (double)n_flops.output_q80_f32 / ((double)cpu.flops_q80_f32 + EPS) / 1e9;
total_latency *= 1000; // convert to ms

36
common/profiler.h
View file

@ -26,12 +26,12 @@ struct cpu_props {
name(""),
description(""),
cores(0),
flops_f32_f32(EPS),
flops_f16_f32(EPS),
flops_q4k_f32(EPS),
flops_q5k_f32(EPS),
flops_q6k_f32(EPS),
flops_q80_f32(EPS) {}
flops_f32_f32(0.0f),
flops_f16_f32(0.0f),
flops_q4k_f32(0.0f),
flops_q5k_f32(0.0f),
flops_q6k_f32(0.0f),
flops_q80_f32(0.0f) {}
};
struct memory_info {
@ -100,20 +100,20 @@ struct gpu_props {
memory_free (0.0f),
memory_total (0.0f),
metal_read_vram_bw (0.0f),
metal_flops_f32_f32(EPS),
metal_flops_f16_f32(EPS),
metal_flops_q4k_f32(EPS),
metal_flops_q5k_f32(EPS),
metal_flops_q6k_f32(EPS),
metal_flops_q80_f32(EPS),
metal_flops_f32_f32(0.0f),
metal_flops_f16_f32(0.0f),
metal_flops_q4k_f32(0.0f),
metal_flops_q5k_f32(0.0f),
metal_flops_q6k_f32(0.0f),
metal_flops_q80_f32(0.0f),
metal_mem_cpy_delay(0.0f),
cuda_read_vram_bw (0.0f),
cuda_flops_f32_f32 (EPS),
cuda_flops_f16_f32 (EPS),
cuda_flops_q4k_f32 (EPS),
cuda_flops_q5k_f32 (EPS),
cuda_flops_q6k_f32 (EPS),
cuda_flops_q80_f32 (EPS),
cuda_flops_f32_f32 (0.0f),
cuda_flops_f16_f32 (0.0f),
cuda_flops_q4k_f32 (0.0f),
cuda_flops_q5k_f32 (0.0f),
cuda_flops_q6k_f32 (0.0f),
cuda_flops_q80_f32 (0.0f),
cuda_mem_cpy_delay (0.0f) {}
};