perf: make ggml_conv_2d faster

.dockerignore

@@ -0,0 +1,6 @@
+build*/
+test/
+
+.cache/
+*.swp
+models/

Dockerfile

@@ -0,0 +1,17 @@
+ARG UBUNTU_VERSION=22.04
+
+FROM ubuntu:$UBUNTU_VERSION as build
+
+RUN apt-get update && apt-get install -y build-essential git cmake
+
+WORKDIR /sd.cpp
+
+COPY . .
+
+RUN mkdir build && cd build && cmake .. && cmake --build . --config Release
+
+FROM ubuntu:$UBUNTU_VERSION as runtime
+
+COPY --from=build /sd.cpp/build/bin/sd /sd
+
+ENTRYPOINT [ "/sd" ]

README.md

@@ -23,6 +23,7 @@ Inference of [Stable Diffusion](https://github.com/CompVis/stable-diffusion) in
     - Linux
     - Mac OS
     - Windows
+    - Android (via Termux)
 
 ### TODO
 
@@ -34,6 +35,7 @@ Inference of [Stable Diffusion](https://github.com/CompVis/stable-diffusion) in
 - [ ] LoRA support
 - [ ] k-quants support
 - [ ] Cross-platform reproducibility (perhaps ensuring consistency with the original SD)
+- [ ] Adapting to more weight formats
 
 ## Usage
 
@@ -49,6 +51,7 @@ cd stable-diffusion.cpp
 ```
 cd stable-diffusion.cpp
 git pull origin master
+git submodule init
 git submodule update
 ```
 
@@ -84,6 +87,8 @@ You can specify the output model format using the --out_type parameter
 
 ### Build
 
+#### Build from scratch
+
 ```shell
 mkdir build
 cd build
@@ -91,7 +96,7 @@ cmake ..
 cmake --build . --config Release
 ```
 
-#### Using OpenBLAS
+##### Using OpenBLAS
 
 ```
 cmake .. -DGGML_OPENBLAS=ON
@@ -149,6 +154,22 @@ Using formats of different precisions will yield results of varying quality.
   <img src="./assets/img2img_output.png" width="256x">
 </p>
 
+### Docker
+
+#### Building using Docker
+
+```shell
+docker build -t sd .
+```
+
+#### Run
+
+```shell
+docker run -v /path/to/models:/models -v /path/to/output/:/output sd [args...]
+# For example
+# docker run -v ./models:/models -v ./build:/output sd -m /models/sd-v1-4-ggml-model-f16.bin -p "a lovely cat" -v -o /output/output.png
+```
+
 ## Memory/Disk Requirements
 
 | precision | f32  | f16  |q8_0  |q5_0  |q5_1  |q4_0  |q4_1  |

examples/main.cpp

@@ -1,4 +1,5 @@
 #include <stdio.h>
+#include <ctime>
 #include <fstream>
 #include <iostream>
 #include <random>
@@ -28,18 +29,17 @@
 #define TXT2IMG "txt2img"
 #define IMG2IMG "img2img"
 
-// get_num_physical_cores is copy from 
+// get_num_physical_cores is copy from
 // https://github.com/ggerganov/llama.cpp/blob/master/examples/common.cpp
 // LICENSE: https://github.com/ggerganov/llama.cpp/blob/master/LICENSE
 int32_t get_num_physical_cores() {
 #ifdef __linux__
     // enumerate the set of thread siblings, num entries is num cores
     std::unordered_set<std::string> siblings;
-    for (uint32_t cpu=0; cpu < UINT32_MAX; ++cpu) {
-        std::ifstream thread_siblings("/sys/devices/system/cpu"
-            + std::to_string(cpu) + "/topology/thread_siblings");
+    for (uint32_t cpu = 0; cpu < UINT32_MAX; ++cpu) {
+        std::ifstream thread_siblings("/sys/devices/system/cpu" + std::to_string(cpu) + "/topology/thread_siblings");
         if (!thread_siblings.is_open()) {
-            break; // no more cpus
+            break;  // no more cpus
         }
         std::string line;
         if (std::getline(thread_siblings, line)) {
@@ -61,7 +61,7 @@ int32_t get_num_physical_cores() {
         return num_physical_cores;
     }
 #elif defined(_WIN32)
-    //TODO: Implement
+    // TODO: Implement
 #endif
     unsigned int n_threads = std::thread::hardware_concurrency();
     return n_threads > 0 ? (n_threads <= 4 ? n_threads : n_threads / 2) : 4;
@@ -282,6 +282,11 @@ void parse_args(int argc, const char* argv[], Option* opt) {
         fprintf(stderr, "error: can only work with strength in [0.0, 1.0]\n");
         exit(1);
     }
+
+    if (opt->seed < 0) {
+        srand((int)time(NULL));
+        opt->seed = rand();
+    }
 }
 
 int main(int argc, const char* argv[]) {

stable-diffusion.cpp

@@ -3256,6 +3256,10 @@ class StableDiffusionGGML {
                         }
                     }
                 }
+
+#ifdef GGML_PERF
+                ggml_graph_print(&diffusion_graph);
+#endif
                 int64_t t1 = ggml_time_ms();
                 LOG_INFO("step %d sampling completed, taking %.2fs", i + 1, (t1 - t0) * 1.0f / 1000);
                 LOG_DEBUG("diffusion graph use %.2fMB runtime memory: static %.2fMB, dynamic %.2fMB",
@@ -3345,6 +3349,10 @@ class StableDiffusionGGML {
             int64_t t0 = ggml_time_ms();
             ggml_graph_compute_with_ctx(ctx, &vae_graph, n_threads);
             int64_t t1 = ggml_time_ms();
+
+#ifdef GGML_PERF
+            ggml_graph_print(&vae_graph);
+#endif
             LOG_DEBUG("computing vae graph completed, taking %.2fs", (t1 - t0) * 1.0f / 1000);
 
             result = ggml_dup_tensor(res_ctx, moments);
@@ -3470,6 +3478,10 @@ class StableDiffusionGGML {
             int64_t t0 = ggml_time_ms();
             ggml_graph_compute_with_ctx(ctx, &vae_graph, n_threads);
             int64_t t1 = ggml_time_ms();
+
+#ifdef GGML_PERF
+            ggml_graph_print(&vae_graph);
+#endif
             LOG_DEBUG("computing vae graph completed, taking %.2fs", (t1 - t0) * 1.0f / 1000);
 
             result_img = ggml_dup_tensor(res_ctx, img);