ggml : add Vulkan backend (ggerganov#2059)

* Vulkan loader code

* Fix matmul kernel, continue implementation

* Continue implementation

* Vulkan memory management

* Vulkan development

* Matmul call

* Add aligned malloc and free for VMA

* Continue implementation

* First matmul success

* GEMM Kernel optimization

* 1D Blocktiling

* 2D Blocktiling

* Write coalescing

* Continue vulkan implementation and optimization

* First FP16 attempt, disabled for now

* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel

* Enable device extensions properly, restore fp16 matmul op

* Fix mulmat_f16

* Output FP32 in fp16 matmul shader

* Fix f16_to_f32 kernel

* dequant_q4_0 kernel

* Add VMA library

* Avoid requesting dedicated memory, VMA can decide that by itself

* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly

* add cmake commands

* Add 2d write operation, profiling code

* Fix 2d write

* Fix queue selection for AMD RADV

* Fix trailing whitespace in vk_mem_alloc.h

* Add WIP warp tile mat mul shaders

* Disable glslc optimization

* Disable glslc optimization for CMake

* Optimize warptile matmul shader, replace blocktile with it

* Add split-k optimization for small matrix multiplication

Use semaphores for synchronization instead of fences or waitidle

Rework async write/read for synchronization

* Fix validation errors, improve compatibility with AMD GPUs

* Rework command buffer handling

* Variable matmul kernel using specialization constants

* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints

* Reuse semaphores

* Handle stage flags during command buffer submission properly

* Increase matmul test runs for consistent results

* Fix F32 matmul

* Add vectorized loading and zeropadding for matrix multiplication

* Use pinned memory for f16 preprocessing

* Don't force aligned matmul

* Don't free before queue done

* Replace VMA library with native Vulkan buffer management

* Basic offloading support with mul_f32 and dmmv for q4_0

* Run glslc commands in parallel

* Unroll loops in dmmv shader

* Reduce usage of waitIdle

* Reuse pinned allocation for f16 conversion

* Handle devices with only a single queue

* Fix trailing whitespace in CMakeLists.txt

* Allow parallel execution of kernels, parallelize third and fourth dimension calls

* Add fallback for devices only supporting one DescriptorSet per DescriptorPool

* Move to graph function similar to CUDA implementation

* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function

* Add F32 dmmv shaders

* Batch submissions

* Add .spv to gitignore

* Split off matrix vector multiplication for separate optimization

* Use single command buffer for matrix vector multiplication ops

* Reduce overhead of mul_f32 calls by using a single command buffer

* Add submission batching to mul_f32

* Fix tests

* Add missing barrier

* Add further missing barrier

* Add further ops

* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions

* Remove unnecessary cblas link

* Fix descriptor set pre-allocation assert

* Add runtime shader compilation, start transferring shaders to this approach

* Transfer remaining shaders to header and compile on runtime

* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16

* Add support for q4_1, q5_0, q5_1 and q8_0

* Remove unnecessary scalar layout extension

* Parse graph early to pre-record command buffers

* Add q6_k support

* Add multi-submit for command buffers

* Fix q6_k dequant shader for AMD

* Fix q6_k for GPUs without fp16 support

* Simplify q6_k fp16 fix

* Minor fixes

* Fix wg_denom of m-mulmat shaders

* Add Python-based Vulkan shader generator

* Replace shaderc dependency with precompiled shaders

Fix python script to generate shaders

* Clean up code

* Fix shader generator script Windows compatibility

Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>

* Close file before deletion

* Fix vulkan shader fp32 name

* Add q2_k and q3_k support

Add validation check to compare shader results to cpu results

* Add q4_k support

* Add q5_k support

* Bake SPIR-V bytecode into the library instead of loading shaders from file

* Switch to signal semaphores for flexibility

Prepare broadcasting support for mul mat

* Finish broadcasting mul mat support for GQA

* Clean up unused functions

Add repeat op

* Add further ops, not yet enabled. Improve semaphore code

* Reduce number of used semaphores by utilizing timelines more properly

* Remove queue information

* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations

* Add Vulkan to llama-bench

* Remove cblas dependency

* Fix matmul k-split bug

* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader

* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug

* Fix issues with float16 overflows in shaders

* Fix issues with older Vulkan headers on Ubuntu 22.04

* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers

* Implement further ops, rework op_f32 calls, fix bugs

* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code

* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders

* Merge upstream changes, fix conflicts, adapt soft_max op

* Fix Python and shader header format

* Free model gpu buffers on exit

* Use single queue per device to simplify code

* Add matmul shader support for running multiple calculations in parallel

* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible

* Fix missing event cast

* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity

* Fix warning about empty C function parameters

* Fix compiler warnings

* Properly implement Vulkan backend buffer handling

* Fix oversized host staging buffers

* Simplify barrier synchronization calls

* Fix gcc warnings

* Implement max_size for backend buffer types to limit the size of a single allocation

* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size

* refactor multi buf

* Disable unsupported ops to fix tests

* Check for maintenance4 support before using it

* Handle devices with only a single queue

* Fix single queue logic

* propagate buffer usage in multi buffers

* Implement rope_neox op

* Cleanup header and other files

* Simplify gpu_extras by removing events and putting staging memcpys into contexts

* Move queue into context

Add not-yet-enabled async backend ops

* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization

* Add get_max_size to SYCL backend.

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

* llama : fix trailing whitespace

---------

Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

CMakeLists.txt

@@ -99,6 +99,7 @@ set(LLAMA_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
 option(LLAMA_HIPBLAS                         "llama: use hipBLAS"                               OFF)
 option(LLAMA_HIP_UMA                         "llama: use HIP unified memory architecture"       OFF)
 option(LLAMA_CLBLAST                         "llama: use CLBlast"                               OFF)
+option(LLAMA_VULKAN                          "llama: use Vulkan"                                OFF)
 option(LLAMA_METAL                           "llama: use Metal"                                 ${LLAMA_METAL_DEFAULT})
 option(LLAMA_METAL_NDEBUG                    "llama: disable Metal debugging"                   OFF)
 option(LLAMA_METAL_SHADER_DEBUG              "llama: compile Metal with -fno-fast-math"         OFF)
@@ -416,6 +417,22 @@ if (LLAMA_CLBLAST)
     endif()
 endif()
 
+if (LLAMA_VULKAN)
+    find_package(Vulkan)
+    if (Vulkan_FOUND)
+        message(STATUS "Vulkan found")
+
+        add_library(ggml-vulkan STATIC ggml-vulkan.cpp ggml-vulkan.h)
+        target_link_libraries(ggml-vulkan PRIVATE Vulkan::Vulkan)
+
+        add_compile_definitions(GGML_USE_VULKAN)
+
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ggml-vulkan)
+    else()
+        message(WARNING "Vulkan not found")
+    endif()
+endif()
+
 if (LLAMA_HIPBLAS)
     list(APPEND CMAKE_PREFIX_PATH /opt/rocm)
 

Makefile

@@ -448,6 +448,19 @@ ggml-opencl.o: ggml-opencl.cpp ggml-opencl.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 endif # LLAMA_CLBLAST
 
+ifdef LLAMA_VULKAN
+	MK_CPPFLAGS  += -DGGML_USE_VULKAN
+	MK_LDFLAGS += -lvulkan
+	OBJS    += ggml-vulkan.o
+
+ifdef LLAMA_VULKAN_CHECK_RESULTS
+	MK_CPPFLAGS  += -DGGML_VULKAN_CHECK_RESULTS
+endif
+
+ggml-vulkan.o: ggml-vulkan.cpp ggml-vulkan.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+endif # LLAMA_VULKAN
+
 ifdef LLAMA_HIPBLAS
 
 	ifeq ($(wildcard /opt/rocm),)

examples/llama-bench/llama-bench.cpp

@@ -562,6 +562,7 @@ struct test {
     static const int build_number;
     static const bool cuda;
     static const bool opencl;
+    static const bool vulkan;
     static const bool metal;
     static const bool gpu_blas;
     static const bool blas;
@@ -643,6 +644,9 @@ struct test {
         if (opencl) {
             return "OpenCL";
         }
+        if (vulkan) {
+            return "Vulkan";
+        }
         if (metal) {
             return "Metal";
         }
@@ -658,7 +662,7 @@ struct test {
     static const std::vector<std::string> & get_fields() {
         static const std::vector<std::string> fields = {
             "build_commit", "build_number",
-            "cuda", "opencl", "metal", "gpu_blas", "blas",
+            "cuda", "opencl", "vulkan", "metal", "gpu_blas", "blas",
             "cpu_info", "gpu_info",
             "model_filename", "model_type", "model_size", "model_n_params",
             "n_batch", "n_threads", "type_k", "type_v",
@@ -682,7 +686,7 @@ struct test {
             field == "avg_ns" || field == "stddev_ns") {
             return INT;
         }
-        if (field == "cuda" || field == "opencl" || field == "metal" || field == "gpu_blas" || field == "blas" ||
+        if (field == "cuda" || field == "opencl"  || field == "vulkan"|| field == "metal" || field == "gpu_blas" || field == "blas" ||
             field == "f16_kv" || field == "no_kv_offload" || field == "mul_mat_q") {
             return BOOL;
         }
@@ -710,7 +714,7 @@ struct test {
         }
         std::vector<std::string> values = {
             build_commit, std::to_string(build_number),
-            std::to_string(cuda), std::to_string(opencl), std::to_string(metal), std::to_string(gpu_blas), std::to_string(blas),
+            std::to_string(cuda), std::to_string(opencl), std::to_string(vulkan), std::to_string(metal), std::to_string(gpu_blas), std::to_string(blas),
             cpu_info, gpu_info,
             model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params),
             std::to_string(n_batch), std::to_string(n_threads), ggml_type_name(type_k), ggml_type_name(type_v),
@@ -738,6 +742,7 @@ const std::string test::build_commit = LLAMA_COMMIT;
 const int         test::build_number = LLAMA_BUILD_NUMBER;
 const bool        test::cuda         = !!ggml_cpu_has_cublas();
 const bool        test::opencl       = !!ggml_cpu_has_clblast();
+const bool        test::vulkan       = !!ggml_cpu_has_vulkan();
 const bool        test::metal        = !!ggml_cpu_has_metal();
 const bool        test::gpu_blas     = !!ggml_cpu_has_gpublas();
 const bool        test::blas         = !!ggml_cpu_has_blas();

ggml-alloc.c

@@ -778,38 +778,26 @@ size_t ggml_allocr_alloc_graph(ggml_allocr_t alloc, struct ggml_cgraph * graph)
 }
 
 // utils
-ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft) {
-    GGML_ASSERT(ggml_get_no_alloc(ctx) == true);
-
-    size_t alignment = ggml_backend_buft_get_alignment(buft);
-
-    size_t nbytes = 0;
-    for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
-        if (t->data == NULL && t->view_src == NULL) {
-            nbytes += GGML_PAD(ggml_backend_buft_get_alloc_size(buft, t), alignment);
-        }
-    }
-
-    if (nbytes == 0) {
-        // all the tensors in the context are already allocated
-#ifndef NDEBUG
-        fprintf(stderr, "%s: all tensors in the context are already allocated\n", __func__);
-#endif
-        return NULL;
-    }
 
-    ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, nbytes);
+static bool alloc_tensor_range(struct ggml_context * ctx,
+        struct ggml_tensor * first, struct ggml_tensor * last,
+        ggml_backend_buffer_type_t buft, size_t size,
+        ggml_backend_buffer_t ** buffers, size_t * n_buffers) {
+    ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, size);
     if (buffer == NULL) {
-        // failed to allocate buffer
 #ifndef NDEBUG
-        fprintf(stderr, "%s: failed to allocate buffer\n", __func__);
+        fprintf(stderr, "%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(buft), size);
 #endif
-        return NULL;
+        for (size_t i = 0; i < *n_buffers; i++) {
+            ggml_backend_buffer_free(*buffers[i]);
+        }
+        free(buffers);
+        return false;
     }
 
     ggml_tallocr_t tallocr = ggml_tallocr_new_from_buffer(buffer);
 
-    for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+    for (struct ggml_tensor * t = first; t != last; t = ggml_get_next_tensor(ctx, t)) {
         if (t->data == NULL) {
             if (t->view_src == NULL) {
                 ggml_tallocr_alloc(tallocr, t);
@@ -826,6 +814,76 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
 
     ggml_tallocr_free(tallocr);
 
+    *buffers = realloc(*buffers, sizeof(ggml_backend_buffer_t) * (*n_buffers + 1));
+    (*buffers)[(*n_buffers)++] = buffer;
+
+    return true;
+}
+
+ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft) {
+    GGML_ASSERT(ggml_get_no_alloc(ctx) == true);
+
+    size_t alignment = ggml_backend_buft_get_alignment(buft);
+    size_t max_size = ggml_backend_buft_get_max_size(buft);
+
+    ggml_backend_buffer_t * buffers = NULL;
+    size_t n_buffers = 0;
+
+    size_t cur_buf_size = 0;
+    struct ggml_tensor * first = ggml_get_first_tensor(ctx);
+    for (struct ggml_tensor * t = first; t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+        size_t this_size = 0;
+        if (t->data == NULL && t->view_src == NULL) {
+            this_size = GGML_PAD(ggml_backend_buft_get_alloc_size(buft, t), alignment);
+        }
+
+        if (this_size > max_size) {
+            // tensor is too large to fit in a single buffer
+            fprintf(stderr, "%s: tensor %s is too large to fit in a %s buffer (tensor size: %zu, max buffer size: %zu)\n",
+                    __func__, t->name,
+                    ggml_backend_buft_name(buft),
+                    this_size, max_size);
+            for (size_t i = 0; i < n_buffers; i++) {
+                ggml_backend_buffer_free(buffers[i]);
+            }
+            free(buffers);
+            return NULL;
+        }
+
+        if ((cur_buf_size + this_size) > max_size) {
+            // allocate tensors in the current buffer
+            if (!alloc_tensor_range(ctx, first, t, buft, cur_buf_size, &buffers, &n_buffers)) {
+                return NULL;
+            }
+            first = t;
+            cur_buf_size = this_size;
+        } else {
+            cur_buf_size += this_size;
+        }
+    }
+
+    // allocate remaining tensors
+    if (cur_buf_size > 0) {
+        if (!alloc_tensor_range(ctx, first, NULL, buft, cur_buf_size, &buffers, &n_buffers)) {
+            return NULL;
+        }
+    }
+
+    if (n_buffers == 0) {
+        // all the tensors in the context are already allocated
+#ifndef NDEBUG
+        fprintf(stderr, "%s: all tensors in the context are already allocated\n", __func__);
+#endif
+        return NULL;
+    }
+
+    ggml_backend_buffer_t buffer;
+    if (n_buffers == 1) {
+        buffer = buffers[0];
+    } else {
+        buffer = ggml_backend_multi_buffer_alloc_buffer(buffers, n_buffers);
+    }
+    free(buffers);
     return buffer;
 }
 

ggml-backend-impl.h

@@ -19,6 +19,7 @@ extern "C" {
         const char *          (*GGML_CALL get_name)        (ggml_backend_buffer_type_t buft);
         ggml_backend_buffer_t (*GGML_CALL alloc_buffer)    (ggml_backend_buffer_type_t buft, size_t size);
         size_t                (*GGML_CALL get_alignment)   (ggml_backend_buffer_type_t buft); // tensor alignment
+        size_t                (*GGML_CALL get_max_size)    (ggml_backend_buffer_type_t buft); // allocation max size
         size_t                (*GGML_CALL get_alloc_size)  (ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
         bool                  (*GGML_CALL supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
         // check if tensor data is in host memory
@@ -63,6 +64,11 @@ extern "C" {
     // do not use directly, use ggml_backend_tensor_copy instead
     bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst);
 
+    // buffer that contains a collection of buffers
+    GGML_CALL ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers);
+    GGML_CALL bool                  ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer);
+    GGML_CALL void                  ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
+
     //
     // Backend
     //