Performance decreated between tag b1500 and b2581 on Windows ARM64 PC

[Billzhong2022]

Hi LLAMA team,

I use llama tag b2581 on Windows ARM64 PC, the performance is more lower than previous tag b1500. Please refer to below detailed information. What is the reason? Please help on this issue.

Thanks a lot!

[Detailed information]

Command: main.exe -m llama-2-7b-chat.ggufv3.q4_0.bin --color --ctx_size 2048 -n -1 -ins -b 256 --top_k 10000 --temp 0.2 --repeat_penalty 1.1 -t 10

Prompt: I have 3 years of experience as a software developer. Now I got bored with coding and want to transition to another career. My education qualifications are B. Tech in computer science, and I am well-versed in understanding the business side of software as well. Suggest a list of career options that are easy for me to transition.

system_info: n_threads = 10 / 12 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | FMA = 0 | NEON = 1 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 0 | SSSE3 = 0 | VSX = 0 |

Tag b1500 results: llama_print_timings: load time = 723.53 ms llama_print_timings: sample time = 925.29 ms / 624 runs ( 1.48 ms per token, 674.38 tokens per second) llama_print_timings: prompt eval time = 2583.12 ms / 91 tokens ( 28.39 ms per token, 35.23 tokens per second) llama_print_timings: eval time = 31693.17 ms / 625 runs ( 50.71 ms per token, 19.72 tokens per second) llama_print_timings: total time = 51797.58 ms

Tag b2581 results: llama_print_timings: load time = 963.25 ms llama_print_timings: sample time = 416.14 ms / 586 runs ( 0.71 ms per token, 1408.17 tokens per second) llama_print_timings: prompt eval time = 11847.94 ms / 94 tokens ( 126.04 ms per token, 7.93 tokens per second) llama_print_timings: eval time = 68542.50 ms / 585 runs ( 117.17 ms per token, 8.53 tokens per second) llama_print_timings: total time = 82696.57 ms / 679 tokens

[woachk]

Which platform is that on? Snapdragon 8cx Gen 3 I presume?

[Billzhong2022]

Hi LLAMA team,

I use platform https://www.qualcomm.com/products/mobile/snapdragon/pcs-and-tablets/snapdragon-x-elite.

Thanks!

[JohannesGaessler]

If you want any chance of getting this fixed, do a git bisect to identify the exact commit that caused performance regression and notify the corresponding dev.

[Billzhong2022]

Hi LLAMA team,

I tried to do "git bisect" to find root reason for it, but there're huge patches added between tag b1500 and b2581.

Can you please help check and analyze this issue?

Thanks!

[JohannesGaessler]

I tried to do "git bisect" to find root reason for it, but there're huge patches added between tag b1500 and b2581.

Download the model as the original weights and convert it from that at each git bisect iteration.

Can you please help check and analyze this issue?

As I said, if you cannot identify the commit that is causing the performance regression this has basically no chance of getting fixed. This issue seems to be hardware-specific so without the corresponding hardware it is impossible to find the bad commit. If there was a performance regression for the hardware of any of the devs they would have already reported it.

[woachk]

I get significantly lower perf than expected on 8cx Gen 3 too, and can also get access to X Elite hardware.

[Billzhong2022]

Hi LLAMA team,

Is it related with FP16 or NEON feature?

Thanks!

[Billzhong2022]

Hi LLAMA team,

Do you find more useful information for this issue?

Thanks!

[JohannesGaessler]

As I've said two times already: without a git bisect from one of the affected people nothing is going to happen.

[Billzhong2022]

Hi LLAMA team,

I've found the commit "780e24a" between tag b1951 & b1952 caused performance decreated on Windows ARM64 PC.

Please help on this issue continually.

Thanks!

[Commit]

commit 780e24a22eb595b705cbe8284771e9ceff1c4dd2 Author: Reinforce-II fate@eastal.com Date: Mon Jan 22 21:15:08 2024 +0800

ggml : parallelize FP32 conversion when using BLAS (#5045)

* make GGML_TASK_INIT phase can be run in multithread

* multithreaded dequantize in mul_mat when using blas library

* minor fixes

* update outdated comment
* fix coding style

* simplify code

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

==============================================================================

Please refer to below test results.

[Test results]

Command: main.exe -m llama-2-7b-chat.ggufv3.q4_0.bin --color --ctx_size 2048 -n -1 -ins -b 256 --top_k 10000 --temp 0.2 --repeat_penalty 1.1 -t 10

Prompt: I have 3 years of experience as a software developer. Now I got bored with coding and want to transition to another career. My education qualifications are B. Tech in computer science, and I am well-versed in understanding the business side of software as well. Suggest a list of career options that are easy for me to transition.

system_info: n_threads = 10 / 12 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | FMA = 0 | NEON = 1 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 0 | SSSE3 = 0 | VSX = 0 |

Tag b1951 results: llama_print_timings: load time = 839.94 ms llama_print_timings: sample time = 895.91 ms / 568 runs ( 1.58 ms per token, 633.99 tokens per second) llama_print_timings: prompt eval time = 2770.04 ms / 93 tokens ( 29.79 ms per token, 33.57 tokens per second) llama_print_timings: eval time = 28958.15 ms / 568 runs ( 50.98 ms per token, 19.61 tokens per second) llama_print_timings: total time = 43121.95 ms / 661 tokens

Tag b1952 results: llama_print_timings: load time = 1381.51 ms llama_print_timings: sample time = 1033.81 ms / 552 runs ( 1.87 ms per token, 533.95 tokens per second) llama_print_timings: prompt eval time = 9733.86 ms / 93 tokens ( 104.67 ms per token, 9.55 tokens per second) llama_print_timings: eval time = 69562.43 ms / 552 runs ( 126.02 ms per token, 7.94 tokens per second) llama_print_timings: total time = 83113.79 ms / 645 tokens

Tag b2581 default results: llama_print_timings: load time = 963.25 ms llama_print_timings: sample time = 416.14 ms / 586 runs ( 0.71 ms per token, 1408.17 tokens per second) llama_print_timings: prompt eval time = 11847.94 ms / 94 tokens ( 126.04 ms per token, 7.93 tokens per second) llama_print_timings: eval time = 68542.50 ms / 585 runs ( 117.17 ms per token, 8.53 tokens per second) llama_print_timings: total time = 82696.57 ms / 679 tokens

Tag b2581 after removing commit 780e24a2 results: llama_print_timings: load time = 1055.60 ms llama_print_timings: sample time = 436.21 ms / 621 runs ( 0.70 ms per token, 1423.64 tokens per second) llama_print_timings: prompt eval time = 18649.48 ms / 94 tokens ( 198.40 ms per token, 5.04 tokens per second) llama_print_timings: eval time = 32446.21 ms / 620 runs ( 52.33 ms per token, 19.11 tokens per second) llama_print_timings: total time = 53483.55 ms / 714 tokens

[JohannesGaessler]

I am seeing virtually no difference with 6.6.19-1-MANJARO and a Ryzen 5950X.

[woachk]

@Billzhong2022 what if you limit the cores used through -C to the two "big" complexes?

[ReinForce-II]

@Billzhong2022 could you provide those testing results

1. running on a single cluster, e.g. taskset -c 0-3 <...>, or something equivalent on windows
2. use -t 4, and set environemnt variable like OPENBLAS_NUM_THREADS=4 if you are using blas backend

[Billzhong2022]

Hi LLAMA team,

Do you have patch to fix this issue?

Thanks!

[Billzhong2022]

Hi LLAMA team,

Any update? Do you use MACRO "GGML_USE_OPENBLAS" for all your modified codes in commit https://github.com/ggerganov/llama.cpp/commit/780e24a22eb595b705cbe8284771e9ceff1c4dd2?

@ReinForce-II use -t 4: The performance is not good.

Thanks!

[ReinForce-II]

Hi LLAMA team,

Any update? Do you use MACRO "GGML_USE_OPENBLAS" for all your modified codes in commit 780e24a?

@ReinForce-II use -t 4: The performance is not good.

Thanks!

How about 1.? You can run cmd /c start /b /affinity 0xf main.exe -t 4 ... in powershell, better if you can run cmd /c start /b /affinity 0x1e main.exe -t 4 ... additionally

thanks.

[Billzhong2022]

Hi @ReinForce-II ,

How about 1.? You can run cmd /c start /b /affinity 0xf main.exe -t 4 ... in powershell, Answer: The performance is very very bad.

better if you can run cmd /c start /b /affinity 0x1e main.exe -t 4 ... additionally Answer: The performance is bad.

Please help on this issue with high priority.

Thanks!

[ReinForce-II]

Hi @ReinForce-II ,

How about 1.? You can run cmd /c start /b /affinity 0xf main.exe -t 4 ... in powershell, Answer: The performance is very very bad.

better if you can run cmd /c start /b /affinity 0x1e main.exe -t 4 ... additionally Answer: The performance is bad.

Please help on this issue with high priority.

Thanks!

This issue might to be specific to qualcomm laptop platforms. I haven't been able to reproduce it on several arm-based hardwares such as ampere, rockchip, graviton. Sorry, but I currently have no access to either 8cx gen3 or x elite, so I have no idea about it now.

[woachk]

For reference: the X Elite has 3 4-core clusters.

The cores across the 3 clusters are identical but the "lower power" cluster has 1/2 the link to fabric as the two other ones, and as such can only use half as much memory bandwidth.

For applications that expect somewhat uniform throughput between the cores, that can cause a breakdown in expected versus realised performance.

[Billzhong2022]

Hi @ReinForce-II ,

But after reverting commit https://github.com/ggerganov/llama.cpp/commit/780e24a22eb595b705cbe8284771e9ceff1c4dd2, the performance is much better on platform https://www.qualcomm.com/products/mobile/snapdragon/pcs-and-tablets/snapdragon-x-elite. How to explain it?

Thanks!

[ReinForce-II]

Hi @ReinForce-II ,

But after reverting commit 780e24a, the performance is much better on platform https://www.qualcomm.com/products/mobile/snapdragon/pcs-and-tablets/snapdragon-x-elite. How to explain it?

Thanks!

It might have something to do with more synchronize operations in the commit

[Billzhong2022]

Hi @ReinForce-II ,

Ok. Is it fixed?

Thanks!

[Billzhong2022]

Hi @ReinForce-II ,

Please help debug and fix this issue.

Thank you very much!

[ReinForce-II]

Hi @ReinForce-II ,

Please help debug and fix this issue.

Thank you very much!

It would be great help if you can kindly provide some sampling results from snapdragon profiler otherwise, there's a higher likelihood of the process getting stuck here, until the x elite products become publicly availiable

[ReinForce-II]

Hi, @Billzhong2022

Please take a look at 4ae60ad8 The commit is not specified for snapdragon device, but it might also alleviate your problem.

hope for your feedback.

[Billzhong2022]

Hi @ReinForce-II ,

After applying patch 4ae60ad8, the llama2 performance is not good on Snapdragon X Elite device.

Please dig out this issue further.

Thanks!

Logs: llama_print_timings: load time = 3453.65 ms llama_print_timings: sample time = 372.97 ms / 658 runs ( 0.57 ms per token, 1764.23 tokens per second) llama_print_timings: prompt eval time = 15051.02 ms / 94 tokens ( 160.12 ms per token, 6.25 tokens per second) llama_print_timings: eval time = 115842.64 ms / 657 runs ( 176.32 ms per token, 5.67 tokens per second) llama_print_timings: total time = 143545.05 ms / 751 tokens

[Billzhong2022]

Hi @ReinForce-II ,

Can you please refine commit https://github.com/ggerganov/llama.cpp/commit/780e24a22eb595b705cbe8284771e9ceff1c4dd2 directly?

Thank you very much!

[quic-zhanweiw]

@ReinForce-II It seems the performance issue is due to the function 'ggml_compute_forward()' was called twice in function 'ggml_graph_compute_thread()'. It was just called once in previous code. May we avoid this? Thanks in advance! By the way, I haven't enable Blas feature.

[ReinForce-II]

@ReinForce-II It seems the performance issue is due to the function 'ggml_compute_forward()' was called twice in function 'ggml_graph_compute_thread()'. It was just called once in previous code. May we avoid this? Thanks in advance! By the way, I haven't enable Blas feature.

In previous code, ggml_compute_forward is called once with GGML_TASK_INIT state, by only one of the threads, the remaining threads will spin waiting it returns. Then called once with GGML_TASK_COMPUTE state, by all threads. After, both GGML_TASK_INIT and GGML_TASK_COMPUTE are called by all threads. In GGML_TASK_INIT state, if you are running w/o BLAS feature, one of the threads will do the work, the remainging threads returns immediately and spin waiting it. I think there is no difference in this part.

[quic-zhanweiw]

Thanks @ReinForce-II for your update! In the latest Llama.cpp code, after add the patch below, the performance issue disappeared. May you help check the reason?

diff --git a/ggml.c b/ggml.c
index b96a82a4..f71369f6 100644
--- a/ggml.c
+++ b/ggml.c
@@ -19494,13 +19494,13 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {

                 params.nth = n_tasks;

-                if (n_tasks == 1) {
-                    /* INIT */
-                    if (GGML_OP_HAS_INIT[node->op]) {
-                        params.type = GGML_TASK_TYPE_INIT;
-                        ggml_compute_forward(&params, node);
-                    }
+                /* INIT */
+                if (GGML_OP_HAS_INIT[node->op]) {
+                    params.type = GGML_TASK_TYPE_INIT;
+                    ggml_compute_forward(&params, node);
+                }

+                if (n_tasks == 1) {
                     // TODO: maybe push node_n to the atomic but if other threads see n_tasks is 1,
                     // they do something more efficient than spinning (?)
                     params.type = GGML_TASK_TYPE_COMPUTE;
@@ -19524,10 +19524,10 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
             task_phase = GGML_TASK_TYPE_INIT;
             atomic_store(&state->shared->n_active,  n_threads);
             atomic_store(&state->shared->node_n,    node_n);
-            atomic_store(&state->shared->node_task, task_phase);
+            // atomic_store(&state->shared->node_task, task_phase);
         } else {
             ggml_graph_compute_thread_sync_node(&node_n,     state, false);
-            ggml_graph_compute_thread_sync_task(&task_phase, state, false);
+            // ggml_graph_compute_thread_sync_task(&task_phase, state, false);
         }

         // check if we should stop
@@ -19538,13 +19538,17 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
         const int n_tasks = ggml_get_n_tasks(node, n_threads, state->shared->n_threads);

         struct ggml_compute_params params = {
-            /*.type  =*/ GGML_TASK_TYPE_INIT,
+            /*.type  =*/ GGML_TASK_TYPE_COMPUTE,
             /*.ith   =*/ state->ith,
             /*.nth   =*/ n_tasks,
             /*.wsize =*/ cplan->work_size,
             /*.wdata =*/ cplan->work_data,
         };

+        if (state->ith < n_tasks) {
+            ggml_compute_forward(&params, node);
+        }
+/*
         if (state->ith < n_tasks) {
             if (GGML_OP_HAS_INIT[node->op]) {
                 ggml_compute_forward(&params, node);
@@ -19579,6 +19583,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
         else {
             ggml_graph_compute_thread_sync_task(&task_phase, state, false);
         }
+*/
     }

     return 0;

[ReinForce-II]

@quic-zhanweiw Thanks looking into the problem.

I have checked the changes. It looks like eliminated extra sync stages, 4 -> 2 per node; runs init stage single threaded. But I'm unable to determine why huge performance regression could happen according to these changes.

Could you provide a flamegraph with symbols or something alternative?

Or consider examining the impact of applying part of your changes.

1.

From 48283bb99ffd21fcd1637b4116d81c801506fbcb Mon Sep 17 00:00:00 2001
From: Reinforce-II <fate@eastal.com>
Date: Mon, 20 May 2024 05:51:33 +0000
Subject: [PATCH] -

---
 ggml.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/ggml.c b/ggml.c
index 53da231e..949d4e33 100644
--- a/ggml.c
+++ b/ggml.c
@@ -19834,6 +19834,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
             ggml_compute_forward(&params, node, state);
         }

+        /*
         if (atomic_fetch_sub(&state->shared->n_active, 1) == 1) {
             task_phase = GGML_TASK_TYPE_FINALIZE;
             atomic_store(&state->shared->n_active,  n_threads);
@@ -19842,6 +19843,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
         else {
             ggml_graph_compute_thread_sync_task(&task_phase, state, false);
         }
+        */
     }

     return 0;
--
2.32.0.windows.2

2.

From a21f0f623d2ee1460e892803c2d83fa14f7acd20 Mon Sep 17 00:00:00 2001
From: Reinforce-II <fate@eastal.com>
Date: Mon, 20 May 2024 05:52:12 +0000
Subject: [PATCH] -

---
 ggml.c | 8 +-------
 1 file changed, 1 insertion(+), 7 deletions(-)

diff --git a/ggml.c b/ggml.c
index 53da231e..0642f57c 100644
--- a/ggml.c
+++ b/ggml.c
@@ -19820,13 +19820,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
             atomic_store(&state->shared->node_task, task_phase);
         }
         else {
-            // TODO: this sched_yield can have significant impact on the performance - either positive or negative
-            //       depending on the workload and the operating system.
-            //       since it is not clear what is the best approach, it should potentially become user-configurable
-            //       ref: https://github.com/ggerganov/ggml/issues/291
-            // UPD:  adding the do_yield flag seems to resolve the issue universally
-            const bool do_yield = node_n < 0 || cgraph->nodes[node_n]->op == GGML_OP_MUL_MAT;
-            ggml_graph_compute_thread_sync_task(&task_phase, state, do_yield);
+            ggml_graph_compute_thread_sync_task(&task_phase, state, false);
         }

         if (state->ith < n_tasks) {
--
2.32.0.windows.2

[zhanweiw]

@ReinForce-II Your patch can't fix this issue. According to the picture, most of the time are cost in 'atomic_load' function.

[ReinForce-II]

@zhanweiw Thank you, the picture makes great help! It looks like LSE atomics extension is not enabled or used, please check compiler option used in msvc.

link

The default is /arch:armv8.0 For example, /arch:armv8.1 allows the _Interlocked* intrinsic functions to use the appropriate atomic instruction that was introduced with the ARMv8.1 extension, FEAT_LSE, but compiler support requires Visual Studio 2022 version 17.2 or later.

[quic-zhanweiw]

Thanks @ReinForce-II very much!

The below patch can fix the performance issue on WoA devices. May I get your support to mainline this change? Thanks in advance!

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 9cc60039..a365a40b 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -999,6 +999,7 @@ if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR CMAKE_GENERATOR_PLATFORM_LWR STR
         add_compile_definitions(__aarch64__) # MSVC defines _M_ARM64 instead
         add_compile_definitions(__ARM_NEON)
         add_compile_definitions(__ARM_FEATURE_FMA)
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:armv8.1")

         set(CMAKE_REQUIRED_FLAGS_PREV ${CMAKE_REQUIRED_FLAGS})
         string(JOIN " " CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS} "/arch:armv8.2")

[woachk]

hmm, this would break support for ARMv8.0 devices (Snapdragon 835 and crew) on Windows 10

IsProcessorFeaturePresent(PF_ARM_V81_ATOMIC_INSTRUCTIONS_AVAILABLE) is how to do a feature check if support for those doesn't matter

[ReinForce-II]

However, IsProcessorFeaturePresent is runtime checking, could we consider add option to build multiple variants of executable, akin to the x64 approach ?

[quic-zhanweiw]

The command below works, may we add this to the 'README'? Add it for Windows on ARM compiling.

cmake .. -A ARM64 -D CMAKE_CXX_FLAGS="/arch:armv8.1"

[ReinForce-II]

That's a solution without adding further complexity. Appears to not provide a long-term solution though. What if we seek the opinions of the maintainers?

[Billzhong2022]

Hi @ReinForce-II ,

As commit "780e24a22eb595b705cbe8284771e9ceff1c4dd2" is ony for BLAS, what is reason you didn't use one BLAS MACRO for functions "ggml_graph_compute_thread_sync_node()" and "ggml_graph_compute_thread_sync_task()"? Function atomic_load() called from both of them.

Thanks!

[Code snippets] static void ggml_graph_compute_thread_sync_node(int node_n, struct ggml_compute_state state, const bool do_yield) { // wait for other threads to finish const int last_node_n = * node_n;

while (true) {
    if (do_yield) {
        sched_yield();
    }

    * node_n = atomic_load(&state->shared->node_n);
    if (* node_n != last_node_n) break;
}

}

static void ggml_graph_compute_thread_sync_task(int task_phase, struct ggml_compute_state state, const bool do_yield) { // wait for other threads to finish const int last_task_phase = * task_phase;

while (true) { if (do_yield) { sched_yield(); }

    * task_phase = atomic_load(&state->shared->node_task);
    if (* task_phase != last_task_phase) break;
}

}

[ReinForce-II]

@Billzhong2022 Using BLAS macro can lead to unnecessary discrepancies in code execution between w/ and w/o BLAS as it's not expected to have observable difference in performance there. It seems to make sense now.

[zhanweiw]

@ReinForce-II

After enabled OpenBlas, even we have enabled the 'armv8.1', while handling the user input prompt, it was stuck there, no response after several minutes.

The 'prompt eval time' is very poor, but the 'eval time' is good. Any idea on the reason?

llama_print_timings:        load time =    1828.46 ms
llama_print_timings:      sample time =      10.72 ms /   113 runs   (    0.09 ms per token, 10538.10 tokens per second)
llama_print_timings: prompt eval time =  134070.93 ms /    93 tokens ( 1441.62 ms per token,     0.69 tokens per second)
llama_print_timings:        eval time =    5422.03 ms /   112 runs   (   48.41 ms per token,    20.66 tokens per second)
llama_print_timings:       total time =  141149.61 ms /   205 tokens

[ReinForce-II]

@zhanweiw Please check your openblas is also built with LSE enabled, it's going to lock bus to implement synchorization operations otherwise.

eval time is not affected, since llama.cpp is not going to invoke blas library on mat mul with size < 32 (you have size 1 here)

[zhanweiw]

After enabled LSE for both OpenBLAS & Llama.cpp, the 'prompt eval time' is still poor.

Enabled LSE in both OpenBLAS & Llama.cpp:

llama_print_timings:        load time =    1879.53 ms
llama_print_timings:      sample time =      10.98 ms /   103 runs   (    0.11 ms per token,  9384.11 tokens per second)
llama_print_timings: prompt eval time =   21516.55 ms /    93 tokens (  231.36 ms per token,     4.32 tokens per second)
llama_print_timings:        eval time =    5072.23 ms /   102 runs   (   49.73 ms per token,    20.11 tokens per second)

Enabled LSE in Llama.cpp without OpenBLAS:

llama_print_timings:        load time =    1816.11 ms
llama_print_timings:      sample time =      26.30 ms /   259 runs   (    0.10 ms per token,  9849.78 tokens per second)
llama_print_timings: prompt eval time =    2875.31 ms /    93 tokens (   30.92 ms per token,    32.34 tokens per second)
llama_print_timings:        eval time =   13077.06 ms /   258 runs   (   50.69 ms per token,    19.73 tokens per second)

[ReinForce-II]

0.69 -> 4.32 looks like the problem caused by bus lock has gone.

Please configure the amount of threads allocated by openblas (you can use environment vars) + allocated by llama.cpp (-t arg) not greater than cores you have. BLAS libraries typically manage their own threads and spins, if you dont configure the amount of threads carefully, it can have high pressure on context switching.

p.s., Using blas implementation is not likely going to have good performance on snapdragon platforms eventually. It will do mat mul in single precision, perform better in case hardware sgemm accelerator present on your platform (or something equivalent such as huge number of cpu cores).

[zhanweiw]

Thanks @ReinForce-II I've already set thread to 10 for both OpenBlas & Llama.cpp in previous test:

C:\llm\llama.cpp>Set OPENBLAS_NUM_THREADS=10
C:\llm\llama.cpp>main.exe -m models\llama-2-7b-chat.q4_0.gguf --color --ctx_size 2048 -n -1 -ins -b 256 --top_k 10000 --temp 0.0 --repeat_penalty 1.0 -t 10

It seems no benefit by enabling OpenBLAS for input prompt processing.

[ReinForce-II]

@zhanweiw Sorry for not making it clear, I means set threads allocated by OpenBLAS + llama.cpp not greater than cores you have, e.g. 6+6, 8+4, rather than 10+10.

[zhanweiw]

Thanks!

Just tried 6+6, 4+6, 6+4, almost got the same result(12 cores devices):

C:\llm\llama.cpp>Set OPENBLAS_NUM_THREADS=6
C:\llm\llama.cpp>main.exe -m models\llama-2-7b-chat.q4_0.gguf --color --ctx_size 2048 -n -1 -ins -b 256 --top_k 10000 --temp 0.0 --repeat_penalty 1.0 -t 6

[ReinForce-II]

Well, the platform may not providing much fp32 arithmetic power. How about 8+4, 8 on blas library. We'd better use llama-bench.exe to get some more detailed results.

[zhanweiw]

By checking the source code of OpenBLAS, the environment variable 'OPENBLAS_NUM_THREADS' only works for Linux system: https://github.com/OpenMathLib/OpenBLAS/blob/700ea74a378cb5bf9073b4447a089a029131fb8b/driver/others/init.c#L820

[zhanweiw]

It's my mistake, the environment was read here. https://github.com/OpenMathLib/OpenBLAS/blob/develop/driver/others/openblas_env.c#L75