wip

Author: sohzm

CMakeLists.txt

@@ -27,6 +27,7 @@ option(SD_BUILD_EXAMPLES             "sd: build examples" ${SD_STANDALONE})
 option(SD_CUBLAS                     "sd: cuda backend" OFF)
 option(SD_HIPBLAS                    "sd: rocm backend" OFF)
 option(SD_METAL                      "sd: metal backend" OFF)
+option(SD_VULKAN                     "sd: vulkan backend" OFF)
 option(SD_FLASH_ATTN                 "sd: use flash attention for x4 less memory usage" OFF)
 option(SD_FAST_SOFTMAX               "sd: x1.5 faster softmax, indeterministic (sometimes, same seed don't generate same image), cuda only" OFF)
 option(SD_BUILD_SHARED_LIBS          "sd: build shared libs" OFF)
@@ -44,6 +45,12 @@ if(SD_METAL)
     add_definitions(-DSD_USE_METAL)
 endif()
 
+if (SD_VULKAN)
+    message("Use Vulkan as backend stable-diffusion")
+    set(GGML_VULKAN ON)
+    add_definitions(-DSD_USE_VULKAN)
+endif ()
+
 if (SD_HIPBLAS)
     message("Use HIPBLAS as backend stable-diffusion")
     set(GGML_HIPBLAS ON)

ggml

@@ -32,6 +32,10 @@
 #include "ggml-metal.h"
 #endif
 
+#ifdef SD_USE_VULKAN
+#include "ggml-vulkan.h"
+#endif
+
 #include "rng.hpp"
 #include "util.h"
 
@@ -588,7 +592,7 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_attention(struct ggml_context* ctx
                                                         struct ggml_tensor* k,
                                                         struct ggml_tensor* v,
                                                         bool mask = false) {
-#if defined(SD_USE_FLASH_ATTENTION) && !defined(SD_USE_CUBLAS) && !defined(SD_USE_METAL)
+#if defined(SD_USE_FLASH_ATTENTION) && !defined(SD_USE_CUBLAS) && !defined(SD_USE_METAL) && !defined(SD_USE_VULKAN)
     struct ggml_tensor* kqv = ggml_flash_attn(ctx, q, k, v, false);  // [N * n_head, n_token, d_head]
 #else
     float d_head = (float)q->ne[0];

ggml_extend.hpp

@@ -21,6 +21,10 @@
 #include "ggml-metal.h"
 #endif
 
+#ifdef SD_USE_VULKAN
+#include "ggml-vulkan.h"
+#endif
+
 #define ST_HEADER_SIZE_LEN 8
 
 uint64_t read_u64(uint8_t* buffer) {

model.cpp

@@ -154,13 +154,17 @@ class StableDiffusionGGML {
         ggml_backend_metal_log_set_callback(ggml_log_callback_default, nullptr);
         backend = ggml_backend_metal_init();
 #endif
+#ifdef SD_USE_VULKAN
+        LOG_DEBUG("Using Vulkan backend");
+        backend = ggml_backend_vk_init();
+#endif
 
         if (!backend) {
             LOG_DEBUG("Using CPU backend");
             backend = ggml_backend_cpu_init();
         }
 #ifdef SD_USE_FLASH_ATTENTION
-#if defined(SD_USE_CUBLAS) || defined(SD_USE_METAL)
+#if defined(SD_USE_CUBLAS) || defined(SD_USE_METAL) || defined(SD_USE_VULKAN)
         LOG_WARN("Flash Attention not supported with GPU Backend");
 #else
         LOG_INFO("Flash Attention enabled");

stable-diffusion.cpp

@@ -24,6 +24,10 @@ struct UpscalerGGML {
         ggml_backend_metal_log_set_callback(ggml_log_callback_default, nullptr);
         backend = ggml_backend_metal_init();
 #endif
+#ifdef SD_USE_VULKAN
+        LOG_DEBUG("Using Vulkan backend");
+        backend = ggml_backend_vk_init(0);
+#endif
 
         if (!backend) {
             LOG_DEBUG("Using CPU backend");