diff --git a/ggml/include/ggml-kompute.h b/ggml/include/ggml-kompute.h index 171465456a5b1..7d76236142d09 100644 --- a/ggml/include/ggml-kompute.h +++ b/ggml/include/ggml-kompute.h @@ -11,6 +11,8 @@ extern "C" { #endif +#define GGML_KOMPUTE_MAX_DEVICES 16 + struct ggml_vk_device { int index; int type; // same as VkPhysicalDeviceType @@ -23,10 +25,10 @@ struct ggml_vk_device { }; struct ggml_vk_device * ggml_vk_available_devices(size_t memoryRequired, size_t * count); +int ggml_backend_kompute_get_device_count(void); +void ggml_backend_kompute_get_device_memory(int device, size_t * free, size_t * total); bool ggml_vk_get_device(struct ggml_vk_device * device, size_t memoryRequired, const char * name); bool ggml_vk_has_vulkan(void); -bool ggml_vk_has_device(void); -struct ggml_vk_device ggml_vk_current_device(void); // // backend API diff --git a/ggml/src/ggml-kompute.cpp b/ggml/src/ggml-kompute.cpp index 41ac63fa48e0f..dfecf0881023c 100644 --- a/ggml/src/ggml-kompute.cpp +++ b/ggml/src/ggml-kompute.cpp @@ -66,43 +66,31 @@ static std::string ggml_kompute_format_name(int device) { return "Kompute" + std::to_string(device); } -struct ggml_kompute_context { +struct ggml_backend_kompute_context { int device; std::string name; + + kp::Manager manager; std::shared_ptr pool; - ggml_kompute_context(int device) - : device(device), name(ggml_kompute_format_name(device)) {} + ggml_backend_buffer_type buft; + + ggml_backend_kompute_context(int device) + : device(device), name(ggml_kompute_format_name(device)) { buft.context = nullptr; } }; -// FIXME: It would be good to consolidate the kompute manager and the kompute context into one object -// and consolidate the init functions and simplify object lifetime management. As it currently stands, -// we *have* to have the kompute manager no matter what for device discovery, but the kompute context -// is only created when a device is set and vulkan is explicitly turned on. -static ggml_kompute_context *s_kompute_context = nullptr; -class kompute_manager { - kp::Manager *s_mgr = nullptr; - -public: - kp::Manager *operator()() { - if (s_mgr && !s_mgr->hasInstance()) { - destroy(); - } - if (!s_mgr) { - s_mgr = new kp::Manager; - } - return s_mgr; - } +struct ggml_backend_kompute_buffer_type_context { + int device; + int device_ref = 0; + uint64_t buffer_alignment; + uint64_t max_alloc; + std::string name; - void destroy() { - delete s_mgr; - s_mgr = nullptr; - } + ggml_backend_kompute_buffer_type_context(int device, uint64_t buffer_alignment, uint64_t max_alloc) + : device(device), buffer_alignment(buffer_alignment), max_alloc(max_alloc), name(ggml_kompute_format_name(device)) {} }; -static kompute_manager komputeManager; - struct ggml_vk_memory { void *data = nullptr; size_t size = 0; @@ -112,6 +100,65 @@ struct ggml_vk_memory { vk::Buffer *stagingBuffer = nullptr; }; +struct ggml_backend_kompute_buffer_context { + struct ggml_vk_memory memory; +}; + +class kompute_manager { +public: + kompute_manager(); + ~kompute_manager(); + + kp::Manager *get_kp_manager(void); + ggml_backend_t create_backend(int device); + void destroy_backend(ggml_backend_t backend); + ggml_backend_t get_backend(int device); + +private: + // Only for global queries, not for creating devices + kp::Manager *m_kp_manager; + + std::vector m_backends; +}; + + +static kompute_manager komputeManager; + + +static ggml_backend_t kompute_backend(int device) +{ + return komputeManager.get_backend(device); +} + +static ggml_backend_t kompute_backend(ggml_backend_buffer_type_t buffer_type) +{ + auto *buft_ctx = static_cast(buffer_type->context); + return kompute_backend(buft_ctx->device); +} + +static ggml_backend_t kompute_backend(ggml_backend_buffer_t buffer) +{ + return kompute_backend(buffer->buft); +} + +static ggml_backend_kompute_context *kompute_backend_context(int device) +{ + auto * backend = kompute_backend(device); + return backend ? static_cast(backend->context) : nullptr; +} + +static ggml_backend_kompute_context *kompute_backend_context(ggml_backend_buffer_t buffer) +{ + auto * backend = kompute_backend(buffer); + return backend ? static_cast(backend->context) : nullptr; +} + +static ggml_backend_kompute_context *kompute_backend_context(ggml_backend_buffer_type_t buffer_type) +{ + auto * backend = kompute_backend(buffer_type); + return backend ? static_cast(backend->context) : nullptr; +} + #ifdef __linux__ __attribute__((constructor)) static void enable_sam() { @@ -167,12 +214,12 @@ static const char * ggml_vk_getVendorName(uint32_t vendorID) { static std::vector ggml_vk_available_devices_internal(size_t memoryRequired) { std::vector results; - if (!komputeManager()->hasVulkan() || !komputeManager()->hasInstance()) + if (!komputeManager.get_kp_manager()->hasVulkan() || !komputeManager.get_kp_manager()->hasInstance()) return results; std::vector physical_devices; try { - physical_devices = komputeManager()->listDevices(); + physical_devices = komputeManager.get_kp_manager()->listDevices(); } catch (vk::SystemError & err) { std::cerr << __func__ << ": ignoring Vulkan exception: " << err.what() << "\n"; return results; @@ -287,6 +334,24 @@ ggml_vk_device * ggml_vk_available_devices(size_t memoryRequired, size_t * count return arr; } +int ggml_backend_kompute_get_device_count(void) { + auto devices = ggml_vk_available_devices_internal(0); + return devices.size(); +} + + +void ggml_backend_kompute_get_device_memory(int device, size_t * free, size_t * total) { + auto devices = ggml_vk_available_devices_internal(0); + + for (std::size_t i = 0; i < devices.size(); i++) { + if (devices[i].index == device) { + *total = devices[i].heapSize; + *free = devices[i].heapSize; + break; + } + } +} + static void ggml_vk_filterByVendor(std::vector& devices, const std::string& targetVendor) { devices.erase( std::remove_if(devices.begin(), devices.end(), @@ -330,25 +395,25 @@ bool ggml_vk_get_device(ggml_vk_device * device, size_t memoryRequired, const ch } bool ggml_vk_has_vulkan() { - return komputeManager()->hasVulkan(); + return komputeManager.get_kp_manager()->hasVulkan(); } -bool ggml_vk_has_device() { - return komputeManager()->hasDevice(); +static bool ggml_vk_has_device(struct ggml_backend_kompute_context *ctx) { + return ctx->manager.hasDevice(); } -ggml_vk_device ggml_vk_current_device() { - if (!komputeManager()->hasDevice()) +static ggml_vk_device ggml_vk_current_device(struct ggml_backend_kompute_context *ctx) { + if (!ctx->manager.hasDevice()) return ggml_vk_device(); auto devices = ggml_vk_available_devices_internal(0); - ggml_vk_filterByName(devices, komputeManager()->physicalDevice()->getProperties().deviceName.data()); + ggml_vk_filterByName(devices, ctx->manager.physicalDevice()->getProperties().deviceName.data()); GGML_ASSERT(!devices.empty()); return devices.front(); } static -void ggml_vk_allocate_descriptor_pool(struct ggml_kompute_context * ctx, size_t size) { +void ggml_vk_allocate_descriptor_pool(struct ggml_backend_kompute_context * ctx, size_t size) { std::vector descriptorPoolSizes = { vk::DescriptorPoolSize( vk::DescriptorType::eStorageBuffer, @@ -363,16 +428,16 @@ void ggml_vk_allocate_descriptor_pool(struct ggml_kompute_context * ctx, size_t descriptorPoolSizes.data()); ctx->pool = std::make_shared(); - vk::Result r = komputeManager()->device()->createDescriptorPool( + vk::Result r = ctx->manager.device()->createDescriptorPool( &descriptorPoolInfo, nullptr, ctx->pool.get()); if (r != vk::Result::eSuccess) std::cerr << "Error allocating descriptor pool" << vk::to_string(r); } static -void ggml_vk_free_descriptor_pool(struct ggml_kompute_context * ctx) { +void ggml_vk_free_descriptor_pool(struct ggml_backend_kompute_context * ctx) { if (ctx->pool) { - komputeManager()->device()->destroy( + ctx->manager.device()->destroy( *ctx->pool, (vk::Optional)nullptr); ctx->pool = nullptr; @@ -380,7 +445,7 @@ void ggml_vk_free_descriptor_pool(struct ggml_kompute_context * ctx) { } static -vk::Buffer *ggml_vk_allocate_buffer(size_t size) { +vk::Buffer *ggml_vk_allocate_buffer(struct ggml_backend_kompute_context * ctx, size_t size) { vk::BufferCreateInfo bufferCreateInfo; bufferCreateInfo.size = size; bufferCreateInfo.usage = vk::BufferUsageFlagBits::eStorageBuffer | @@ -389,18 +454,18 @@ vk::Buffer *ggml_vk_allocate_buffer(size_t size) { bufferCreateInfo.sharingMode = vk::SharingMode::eExclusive; vk::Buffer *vkBuffer = new vk::Buffer; - vk::Result r = komputeManager()->device()->createBuffer(&bufferCreateInfo, nullptr, vkBuffer); + vk::Result r = ctx->manager.device()->createBuffer(&bufferCreateInfo, nullptr, vkBuffer); if (r != vk::Result::eSuccess) std::cerr << "Error allocating buffer " << vk::to_string(r) << std::endl; return vkBuffer; } static -vk::DeviceMemory *ggml_vk_allocate(size_t size, vk::MemoryPropertyFlags flags, vk::MemoryRequirements requirements, bool *isHostVisible) { +vk::DeviceMemory *ggml_vk_allocate(struct ggml_backend_kompute_context * ctx, size_t size, vk::MemoryPropertyFlags flags, vk::MemoryRequirements requirements, bool *isHostVisible) { uint32_t memoryTypeIndex = -1; bool memoryTypeIndexFound = false; - vk::PhysicalDeviceMemoryProperties memoryProperties = komputeManager()->physicalDevice()->getMemoryProperties(); + vk::PhysicalDeviceMemoryProperties memoryProperties = ctx->manager.physicalDevice()->getMemoryProperties(); for (uint32_t i = 0; i < memoryProperties.memoryTypeCount; i++) { const vk::MemoryType &memoryType = memoryProperties.memoryTypes[i]; const vk::MemoryHeap &memoryHeap = memoryProperties.memoryHeaps[memoryType.heapIndex]; @@ -429,7 +494,7 @@ vk::DeviceMemory *ggml_vk_allocate(size_t size, vk::MemoryPropertyFlags flags, v allocInfo.allocationSize = size; allocInfo.memoryTypeIndex = memoryTypeIndex; vk::DeviceMemory *vkDeviceMemory = new vk::DeviceMemory; - vk::Result r = komputeManager()->device()->allocateMemory(&allocInfo, nullptr, vkDeviceMemory); + vk::Result r = ctx->manager.device()->allocateMemory(&allocInfo, nullptr, vkDeviceMemory); if (r != vk::Result::eSuccess) { std::cerr << "Error allocating memory " << vk::to_string(r) << std::endl; throw std::runtime_error("Error allocating vulkan memory."); @@ -449,31 +514,31 @@ static size_t ggml_vk_aligned_offset(ggml_backend_buffer_t buffer, size_t offset return (offset / minStorageBufferOffsetAlignment) * minStorageBufferOffsetAlignment; } -static ggml_vk_memory ggml_vk_allocate(size_t size) { +static ggml_vk_memory ggml_vk_allocate(struct ggml_backend_kompute_context * ctx, size_t size) { ggml_vk_memory memory; bool isHostVisible = false; { - memory.primaryBuffer = ggml_vk_allocate_buffer(size); - vk::MemoryRequirements memoryRequirements = komputeManager()->device()->getBufferMemoryRequirements(*memory.primaryBuffer); + memory.primaryBuffer = ggml_vk_allocate_buffer(ctx, size); + vk::MemoryRequirements memoryRequirements = ctx->manager.device()->getBufferMemoryRequirements(*memory.primaryBuffer); vk::MemoryPropertyFlags memoryPropertyFlags = vk::MemoryPropertyFlagBits::eDeviceLocal; - memory.primaryMemory = ggml_vk_allocate(size, memoryPropertyFlags, memoryRequirements, &isHostVisible); - komputeManager()->device()->bindBufferMemory(*memory.primaryBuffer, *memory.primaryMemory, 0); + memory.primaryMemory = ggml_vk_allocate(ctx, size, memoryPropertyFlags, memoryRequirements, &isHostVisible); + ctx->manager.device()->bindBufferMemory(*memory.primaryBuffer, *memory.primaryMemory, 0); if (isHostVisible) { - vk::Result r = komputeManager()->device()->mapMemory(*memory.primaryMemory, 0, size, vk::MemoryMapFlags(), &memory.data); + vk::Result r = ctx->manager.device()->mapMemory(*memory.primaryMemory, 0, size, vk::MemoryMapFlags(), &memory.data); if (r != vk::Result::eSuccess) std::cerr << "Error mapping memory" << vk::to_string(r); } } if (!isHostVisible) { - memory.stagingBuffer = ggml_vk_allocate_buffer(size); - vk::MemoryRequirements memoryRequirements = komputeManager()->device()->getBufferMemoryRequirements(*memory.stagingBuffer); + memory.stagingBuffer = ggml_vk_allocate_buffer(ctx, size); + vk::MemoryRequirements memoryRequirements = ctx->manager.device()->getBufferMemoryRequirements(*memory.stagingBuffer); vk::MemoryPropertyFlags memoryPropertyFlags = vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached; - memory.stagingMemory = ggml_vk_allocate(size, memoryPropertyFlags, memoryRequirements, &isHostVisible); - komputeManager()->device()->bindBufferMemory(*memory.stagingBuffer, *memory.stagingMemory, 0); - vk::Result r = komputeManager()->device()->mapMemory(*memory.stagingMemory, 0, size, vk::MemoryMapFlags(), &memory.data); + memory.stagingMemory = ggml_vk_allocate(ctx, size, memoryPropertyFlags, memoryRequirements, &isHostVisible); + ctx->manager.device()->bindBufferMemory(*memory.stagingBuffer, *memory.stagingMemory, 0); + vk::Result r = ctx->manager.device()->mapMemory(*memory.stagingMemory, 0, size, vk::MemoryMapFlags(), &memory.data); if (r != vk::Result::eSuccess) std::cerr << "Error mapping memory" << vk::to_string(r); } @@ -482,21 +547,21 @@ static ggml_vk_memory ggml_vk_allocate(size_t size) { return memory; } -static void ggml_vk_free_memory(ggml_vk_memory &memory) +static void ggml_vk_free_memory(struct ggml_backend_kompute_context * ctx, ggml_vk_memory &memory) { - komputeManager()->device()->destroy( + ctx->manager.device()->destroy( *memory.primaryBuffer, (vk::Optional)nullptr); if (memory.stagingBuffer) { - komputeManager()->device()->destroy( + ctx->manager.device()->destroy( *memory.stagingBuffer, (vk::Optional)nullptr); } - komputeManager()->device()->freeMemory( + ctx->manager.device()->freeMemory( *memory.primaryMemory, (vk::Optional)nullptr); if (memory.stagingMemory) { - komputeManager()->device()->freeMemory( + ctx->manager.device()->freeMemory( *memory.stagingMemory, (vk::Optional)nullptr); } @@ -522,7 +587,7 @@ ggml_vk_memory * ggml_vk_find_tensor(const struct ggml_tensor * t, uint64_t & of } static -const std::shared_ptr ggml_vk_get_tensor(const struct ggml_tensor * t, uint32_t * alignedOffset = nullptr) { +const std::shared_ptr ggml_vk_get_tensor(struct ggml_backend_kompute_context * ctx, const struct ggml_tensor * t, uint32_t * alignedOffset = nullptr) { uint64_t originalOffset = 0; auto * res = ggml_vk_find_tensor(t, originalOffset); if (!res) { @@ -540,7 +605,7 @@ const std::shared_ptr ggml_vk_get_tensor(const struct ggml_tensor * nbytes += *alignedOffset; } - return komputeManager()->tensor( + return ctx->manager.tensor( t->data, nelements, nbytes, kp::Tensor::TensorDataTypes::eFloat, @@ -572,6 +637,7 @@ uint32_t safe_divide(uint32_t a, uint32_t b) { } static void ggml_vk_add( + struct ggml_backend_kompute_context *ctx, kp::Sequence& seq, const std::shared_ptr& inA, const std::shared_ptr& inB, @@ -606,19 +672,21 @@ static void ggml_vk_add( }; std::shared_ptr s_algo = nullptr; - if (!komputeManager()->hasAlgorithm(__func__)) { - s_algo = komputeManager()->algorithm(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts}); + if (!ctx->manager.hasAlgorithm(__func__)) { + s_algo = ctx->manager.algorithm(__func__, ctx->pool.get(), {inA, inB, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts}); } else { - s_algo = komputeManager()->getAlgorithm(__func__); + s_algo = ctx->manager.getAlgorithm(__func__); s_algo->setTensors({inA, inB, out}); s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)}); s_algo->setPushConstants({pushConsts}); - s_algo->updateDescriptors(s_kompute_context->pool.get()); + s_algo->updateDescriptors(ctx->pool.get()); } seq.record(s_algo); } -static void ggml_vk_addrow(kp::Sequence& seq, +static void ggml_vk_addrow( + struct ggml_backend_kompute_context *ctx, + kp::Sequence& seq, const std::shared_ptr& inA, const std::shared_ptr& inB, const std::shared_ptr& out, @@ -637,19 +705,20 @@ static void ggml_vk_addrow(kp::Sequence& seq, }; std::shared_ptr s_algo = nullptr; - if (!komputeManager()->hasAlgorithm(__func__)) - s_algo = komputeManager()->algorithm(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {size}, {}, {pushConsts}); + if (!ctx->manager.hasAlgorithm(__func__)) + s_algo = ctx->manager.algorithm(__func__, ctx->pool.get(), {inA, inB, out}, spirv, {size}, {}, {pushConsts}); else { - s_algo = komputeManager()->getAlgorithm(__func__); + s_algo = ctx->manager.getAlgorithm(__func__); s_algo->setTensors({inA, inB, out}); s_algo->setWorkgroup({size}); s_algo->setPushConstants({pushConsts}); - s_algo->updateDescriptors(s_kompute_context->pool.get()); + s_algo->updateDescriptors(ctx->pool.get()); } seq.record(s_algo); } static void ggml_vk_mul( + struct ggml_backend_kompute_context *ctx, kp::Sequence& seq, const std::shared_ptr& inA, const std::shared_ptr& inB, @@ -684,19 +753,21 @@ static void ggml_vk_mul( }; std::shared_ptr s_algo = nullptr; - if (!komputeManager()->hasAlgorithm(__func__)) { - s_algo = komputeManager()->algorithm(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts}); + if (!ctx->manager.hasAlgorithm(__func__)) { + s_algo = ctx->manager.algorithm(__func__, ctx->pool.get(), {inA, inB, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts}); } else { - s_algo = komputeManager()->getAlgorithm(__func__); + s_algo = ctx->manager.getAlgorithm(__func__); s_algo->setTensors({inA, inB, out}); s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)}); s_algo->setPushConstants({pushConsts}); - s_algo->updateDescriptors(s_kompute_context->pool.get()); + s_algo->updateDescriptors(ctx->pool.get()); } seq.record(s_algo); } -static void ggml_vk_scale(kp::Sequence& seq, +static void ggml_vk_scale( + struct ggml_backend_kompute_context *ctx, + kp::Sequence& seq, const std::shared_ptr& in, const std::shared_ptr& out, uint32_t inOff, uint32_t outOff, @@ -725,19 +796,20 @@ static void ggml_vk_scale(kp::Sequence& seq, } std::shared_ptr s_algo = nullptr; - if (!komputeManager()->hasAlgorithm(name)) { - s_algo = komputeManager()->algorithm(name, s_kompute_context->pool.get(), {in, out}, *spirv, {size}, {}, {pushConsts}); + if (!ctx->manager.hasAlgorithm(name)) { + s_algo = ctx->manager.algorithm(name, ctx->pool.get(), {in, out}, *spirv, {size}, {}, {pushConsts}); } else { - s_algo = komputeManager()->getAlgorithm(name); + s_algo = ctx->manager.getAlgorithm(name); s_algo->setTensors({in, out}); s_algo->setWorkgroup({size}); s_algo->setPushConstants({pushConsts}); - s_algo->updateDescriptors(s_kompute_context->pool.get()); + s_algo->updateDescriptors(ctx->pool.get()); } seq.record(s_algo); } static void ggml_vk_xxlu( + struct ggml_backend_kompute_context *ctx, const std::vector& spirv, const char * suffix, kp::Sequence& seq, const std::shared_ptr& in, const std::shared_ptr& out, @@ -752,43 +824,44 @@ static void ggml_vk_xxlu( auto name = std::string(__func__) + "_" + suffix; std::shared_ptr s_algo = nullptr; - if (!komputeManager()->hasAlgorithm(name)) { - s_algo = komputeManager()->algorithm(name, s_kompute_context->pool.get(), {in, out}, spirv, {size}, {}, {pushConsts}); + if (!ctx->manager.hasAlgorithm(name)) { + s_algo = ctx->manager.algorithm(name, ctx->pool.get(), {in, out}, spirv, {size}, {}, {pushConsts}); } else { - s_algo = komputeManager()->getAlgorithm(name); + s_algo = ctx->manager.getAlgorithm(name); s_algo->setTensors({in, out}); s_algo->setWorkgroup({size}); s_algo->setPushConstants({pushConsts}); - s_algo->updateDescriptors(s_kompute_context->pool.get()); + s_algo->updateDescriptors(ctx->pool.get()); } seq.record(s_algo); } template -static void ggml_vk_silu(Args&&... args) { +static void ggml_vk_silu(struct ggml_backend_kompute_context *ctx, Args&&... args) { const static auto spirv = getSpirvShader(kp::shader_data::op_silu_comp_spv, kp::shader_data::op_silu_comp_spv_len); - ggml_vk_xxlu(spirv, "silu", std::forward(args)...); + ggml_vk_xxlu(ctx, spirv, "silu", std::forward(args)...); } template -static void ggml_vk_relu(Args&&... args) { +static void ggml_vk_relu(struct ggml_backend_kompute_context *ctx, Args&&... args) { const static auto spirv = getSpirvShader(kp::shader_data::op_relu_comp_spv, kp::shader_data::op_relu_comp_spv_len); - ggml_vk_xxlu(spirv, "relu", std::forward(args)...); + ggml_vk_xxlu(ctx, spirv, "relu", std::forward(args)...); } template -static void ggml_vk_gelu(Args&&... args) { +static void ggml_vk_gelu(struct ggml_backend_kompute_context *ctx, Args&&... args) { const static auto spirv = getSpirvShader(kp::shader_data::op_gelu_comp_spv, kp::shader_data::op_gelu_comp_spv_len); - ggml_vk_xxlu(spirv, "gelu", std::forward(args)...); + ggml_vk_xxlu(ctx, spirv, "gelu", std::forward(args)...); } static void ggml_vk_soft_max( + struct ggml_backend_kompute_context *ctx, kp::Sequence& seq, const std::shared_ptr& inA, const std::shared_ptr& inB, @@ -815,21 +888,22 @@ static void ggml_vk_soft_max( auto & inB_ = inB ? inB : inA; std::shared_ptr s_algo = nullptr; - if (!komputeManager()->hasAlgorithm(__func__)) { + if (!ctx->manager.hasAlgorithm(__func__)) { // FIXME: The softmax kernel needs to be fixed to use the subgroupsize which can vary by device const uint32_t local_x = 32; - s_algo = komputeManager()->algorithm(__func__, s_kompute_context->pool.get(), {inA, inB_, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {local_x}, {pushConsts}); + s_algo = ctx->manager.algorithm(__func__, ctx->pool.get(), {inA, inB_, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {local_x}, {pushConsts}); } else { - s_algo = komputeManager()->getAlgorithm(__func__); + s_algo = ctx->manager.getAlgorithm(__func__); s_algo->setTensors({inA, inB_, out}); s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)}); s_algo->setPushConstants({pushConsts}); - s_algo->updateDescriptors(s_kompute_context->pool.get()); + s_algo->updateDescriptors(ctx->pool.get()); } seq.record(s_algo); } static void ggml_vk_norm_( + struct ggml_backend_kompute_context *ctx, const std::vector& spirv, const char * suffix, kp::Sequence& seq, const std::shared_ptr& in, const std::shared_ptr& out, @@ -851,35 +925,37 @@ static void ggml_vk_norm_( auto name = std::string(__func__) + "_" + suffix; std::shared_ptr s_algo = nullptr; - if (!komputeManager()->hasAlgorithm(name)) { - s_algo = komputeManager()->algorithm(name, s_kompute_context->pool.get(), {in, out}, spirv, {(uint32_t)nrows}, {}, {pushConsts}); + if (!ctx->manager.hasAlgorithm(name)) { + s_algo = ctx->manager.algorithm(name, ctx->pool.get(), {in, out}, spirv, {(uint32_t)nrows}, {}, {pushConsts}); } else { - s_algo = komputeManager()->getAlgorithm(name); + s_algo = ctx->manager.getAlgorithm(name); s_algo->setTensors({in, out}); s_algo->setWorkgroup({(uint32_t)nrows}); s_algo->setPushConstants({pushConsts}); - s_algo->updateDescriptors(s_kompute_context->pool.get()); + s_algo->updateDescriptors(ctx->pool.get()); } seq.record(s_algo); } template -static void ggml_vk_norm(Args&&... args) { +static void ggml_vk_norm(struct ggml_backend_kompute_context *ctx, Args&&... args) { const static auto spirv = getSpirvShader(kp::shader_data::op_norm_comp_spv, kp::shader_data::op_norm_comp_spv_len); - ggml_vk_norm_(spirv, "norm", std::forward(args)...); + ggml_vk_norm_(ctx, spirv, "norm", std::forward(args)...); } template -static void ggml_vk_rms_norm(Args&&... args) { +static void ggml_vk_rms_norm(struct ggml_backend_kompute_context *ctx, Args&&... args) { const static auto spirv = getSpirvShader(kp::shader_data::op_rmsnorm_comp_spv, kp::shader_data::op_rmsnorm_comp_spv_len); - ggml_vk_norm_(spirv, "rms", std::forward(args)...); + ggml_vk_norm_(ctx, spirv, "rms", std::forward(args)...); } -static void ggml_vk_diag_mask_inf(kp::Sequence& seq, +static void ggml_vk_diag_mask_inf( + struct ggml_backend_kompute_context *ctx, + kp::Sequence& seq, const std::shared_ptr& in, const std::shared_ptr& out, uint32_t inOff, uint32_t outOff, @@ -899,19 +975,20 @@ static void ggml_vk_diag_mask_inf(kp::Sequence& seq, }; std::shared_ptr s_algo = nullptr; - if (!komputeManager()->hasAlgorithm(__func__)) - s_algo = komputeManager()->algorithm(__func__, s_kompute_context->pool.get(), {in, out}, spirv, {unsigned(ne00), unsigned(ne01), unsigned(ne02)}, {}, {pushConsts}); + if (!ctx->manager.hasAlgorithm(__func__)) + s_algo = ctx->manager.algorithm(__func__, ctx->pool.get(), {in, out}, spirv, {unsigned(ne00), unsigned(ne01), unsigned(ne02)}, {}, {pushConsts}); else { - s_algo = komputeManager()->getAlgorithm(__func__); + s_algo = ctx->manager.getAlgorithm(__func__); s_algo->setTensors({in, out}); s_algo->setWorkgroup({unsigned(ne00), unsigned(ne01), unsigned(ne02)}); s_algo->setPushConstants({pushConsts}); - s_algo->updateDescriptors(s_kompute_context->pool.get()); + s_algo->updateDescriptors(ctx->pool.get()); } seq.record(s_algo); } static void ggml_vk_mul_mat_f16( + struct ggml_backend_kompute_context *ctx, kp::Sequence& seq, const std::shared_ptr& inA, const std::shared_ptr& inB, @@ -948,20 +1025,22 @@ static void ggml_vk_mul_mat_f16( const unsigned ny = unsigned((ne11 + 4 - 1)/4); std::shared_ptr s_algo = nullptr; - if (!komputeManager()->hasAlgorithm(__func__)) { - const uint32_t local_x = ggml_vk_current_device().subgroupSize * 2; - s_algo = komputeManager()->algorithm(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned(ne01), ny, unsigned(ne12*ne13)}, {local_x}, {pushConsts}); + if (!ctx->manager.hasAlgorithm(__func__)) { + const uint32_t local_x = ggml_vk_current_device(ctx).subgroupSize * 2; + s_algo = ctx->manager.algorithm(__func__, ctx->pool.get(), {inA, inB, out}, spirv, {unsigned(ne01), ny, unsigned(ne12*ne13)}, {local_x}, {pushConsts}); } else { - s_algo = komputeManager()->getAlgorithm(__func__); + s_algo = ctx->manager.getAlgorithm(__func__); s_algo->setTensors({inA, inB, out}); s_algo->setWorkgroup({unsigned(ne01), ny, unsigned(ne12*ne13)}); s_algo->setPushConstants({pushConsts}); - s_algo->updateDescriptors(s_kompute_context->pool.get()); + s_algo->updateDescriptors(ctx->pool.get()); } seq.record(s_algo); } -static void ggml_vk_mul_mat_mat_f32(kp::Sequence& seq, +static void ggml_vk_mul_mat_mat_f32( + struct ggml_backend_kompute_context *ctx, + kp::Sequence& seq, const std::shared_ptr& inA, const std::shared_ptr& inB, const std::shared_ptr& out, @@ -987,10 +1066,10 @@ static void ggml_vk_mul_mat_mat_f32(kp::Sequence& seq, nb1, nb2 }; - const uint32_t local_x = ggml_vk_current_device().subgroupSize; + const uint32_t local_x = ggml_vk_current_device(ctx).subgroupSize; std::shared_ptr s_algo = nullptr; - if (!komputeManager()->hasAlgorithm(__func__)) { - s_algo = komputeManager()->algorithm(__func__, s_kompute_context->pool.get(), + if (!ctx->manager.hasAlgorithm(__func__)) { + s_algo = ctx->manager.algorithm(__func__, ctx->pool.get(), {inA, inB, out}, spirv, {unsigned(ne01), unsigned(ne11), @@ -999,19 +1078,20 @@ static void ggml_vk_mul_mat_mat_f32(kp::Sequence& seq, {local_x}, {pushConsts}); } else { - s_algo = komputeManager()->getAlgorithm(__func__); + s_algo = ctx->manager.getAlgorithm(__func__); s_algo->setTensors({inA, inB, out}); s_algo->setWorkgroup({unsigned(ne01), unsigned(ne11), unsigned(std::max(ne12, ne02)), }); s_algo->setPushConstants({pushConsts}); - s_algo->updateDescriptors(s_kompute_context->pool.get()); + s_algo->updateDescriptors(ctx->pool.get()); } seq.record(s_algo); } static void ggml_vk_mul_mat_impl( + struct ggml_backend_kompute_context *ctx, const std::vector& spirv, const char * suffix, uint32_t block_size, kp::Sequence& seq, const std::shared_ptr& inA, const std::shared_ptr& inB, @@ -1038,44 +1118,45 @@ static void ggml_vk_mul_mat_impl( auto name = std::string(__func__) + "_" + suffix; std::shared_ptr s_algo = nullptr; - if (!komputeManager()->hasAlgorithm(name)) { - const uint32_t local_x = ggml_vk_current_device().subgroupSize * 2; - s_algo = komputeManager()->algorithm(name, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned((ne01 + 7)/8), unsigned(ne11), unsigned(ne12*ne13)}, {local_x}, {pushConsts}); + if (!ctx->manager.hasAlgorithm(name)) { + const uint32_t local_x = ggml_vk_current_device(ctx).subgroupSize * 2; + s_algo = ctx->manager.algorithm(name, ctx->pool.get(), {inA, inB, out}, spirv, {unsigned((ne01 + 7)/8), unsigned(ne11), unsigned(ne12*ne13)}, {local_x}, {pushConsts}); } else { - s_algo = komputeManager()->getAlgorithm(name); + s_algo = ctx->manager.getAlgorithm(name); s_algo->setTensors({inA, inB, out}); s_algo->setWorkgroup({unsigned((ne01 + 7)/8), unsigned(ne11), unsigned(ne12*ne13)}); s_algo->setPushConstants({pushConsts}); - s_algo->updateDescriptors(s_kompute_context->pool.get()); + s_algo->updateDescriptors(ctx->pool.get()); } seq.record(s_algo); } template -static void ggml_vk_mul_mat_q4_0(Args&&... args) { +static void ggml_vk_mul_mat_q4_0(struct ggml_backend_kompute_context *ctx, Args&&... args) { const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q4_0_comp_spv, kp::shader_data::op_mul_mat_q4_0_comp_spv_len); - ggml_vk_mul_mat_impl(spirv, "q4_0", 1/*We access blocks unaligned*/, std::forward(args)...); + ggml_vk_mul_mat_impl(ctx, spirv, "q4_0", 1/*We access blocks unaligned*/, std::forward(args)...); } template -static void ggml_vk_mul_mat_q4_1(Args&&... args) { +static void ggml_vk_mul_mat_q4_1(struct ggml_backend_kompute_context *ctx, Args&&... args) { const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q4_1_comp_spv, kp::shader_data::op_mul_mat_q4_1_comp_spv_len); - ggml_vk_mul_mat_impl(spirv, "q4_1", 1/*We access blocks unaligned*/, std::forward(args)...); + ggml_vk_mul_mat_impl(ctx, spirv, "q4_1", 1/*We access blocks unaligned*/, std::forward(args)...); } template -static void ggml_vk_mul_mat_q8_0(Args&&... args) { +static void ggml_vk_mul_mat_q8_0(struct ggml_backend_kompute_context *ctx, Args&&... args) { const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q8_0_comp_spv, kp::shader_data::op_mul_mat_q8_0_comp_spv_len); - ggml_vk_mul_mat_impl(spirv, "q8_0", 1/*We access blocks unaligned*/, std::forward(args)...); + ggml_vk_mul_mat_impl(ctx, spirv, "q8_0", 1/*We access blocks unaligned*/, std::forward(args)...); } static void ggml_vk_mul_mat_q6_k( + struct ggml_backend_kompute_context *ctx, kp::Sequence& seq, const std::shared_ptr& inA, const std::shared_ptr& inB, @@ -1096,20 +1177,21 @@ static void ggml_vk_mul_mat_q6_k( }; std::shared_ptr s_algo = nullptr; - if (!komputeManager()->hasAlgorithm(__func__)) { - const uint32_t local_x = ggml_vk_current_device().subgroupSize * 2; - s_algo = komputeManager()->algorithm(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned((ne01 + 1)/2), unsigned(ne11), unsigned(ne12)}, {local_x}, {pushConsts}); + if (!ctx->manager.hasAlgorithm(__func__)) { + const uint32_t local_x = ggml_vk_current_device(ctx).subgroupSize * 2; + s_algo = ctx->manager.algorithm(__func__, ctx->pool.get(), {inA, inB, out}, spirv, {unsigned((ne01 + 1)/2), unsigned(ne11), unsigned(ne12)}, {local_x}, {pushConsts}); } else { - s_algo = komputeManager()->getAlgorithm(__func__); + s_algo = ctx->manager.getAlgorithm(__func__); s_algo->setTensors({inA, inB, out}); s_algo->setWorkgroup({unsigned((ne01 + 1)/2), unsigned(ne11), unsigned(ne12)}); s_algo->setPushConstants({pushConsts}); - s_algo->updateDescriptors(s_kompute_context->pool.get()); + s_algo->updateDescriptors(ctx->pool.get()); } seq.record(s_algo); } static void ggml_vk_get_rows( + struct ggml_backend_kompute_context *ctx, const std::vector& spirv, const char * suffix, unsigned element_size, unsigned qk, @@ -1135,58 +1217,59 @@ static void ggml_vk_get_rows( auto name = std::string(__func__) + "_" + suffix; std::shared_ptr s_algo = nullptr; - if (!komputeManager()->hasAlgorithm(name)) { - s_algo = komputeManager()->algorithm(name, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {size}, {}, {pushConsts}); + if (!ctx->manager.hasAlgorithm(name)) { + s_algo = ctx->manager.algorithm(name, ctx->pool.get(), {inA, inB, out}, spirv, {size}, {}, {pushConsts}); } else { - s_algo = komputeManager()->getAlgorithm(name); + s_algo = ctx->manager.getAlgorithm(name); s_algo->setTensors({inA, inB, out}); s_algo->setWorkgroup({size}); s_algo->setPushConstants({pushConsts}); - s_algo->updateDescriptors(s_kompute_context->pool.get()); + s_algo->updateDescriptors(ctx->pool.get()); } seq.record(s_algo); } template -static void ggml_vk_get_rows_f32(Args&&... args) { +static void ggml_vk_get_rows_f32(struct ggml_backend_kompute_context *ctx, Args&&... args) { const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_f32_comp_spv, kp::shader_data::op_getrows_f32_comp_spv_len); - ggml_vk_get_rows(spirv, "f32", sizeof(float), 0, std::forward(args)...); + ggml_vk_get_rows(ctx, spirv, "f32", sizeof(float), 0, std::forward(args)...); } template -static void ggml_vk_get_rows_f16(Args&&... args) { +static void ggml_vk_get_rows_f16(struct ggml_backend_kompute_context *ctx, Args&&... args) { const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_f16_comp_spv, kp::shader_data::op_getrows_f16_comp_spv_len); - ggml_vk_get_rows(spirv, "f16", sizeof(half), 0, std::forward(args)...); + ggml_vk_get_rows(ctx, spirv, "f16", sizeof(half), 0, std::forward(args)...); } template -static void ggml_vk_get_rows_q4_0(Args&&... args) { +static void ggml_vk_get_rows_q4_0(struct ggml_backend_kompute_context *ctx, Args&&... args) { const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_q4_0_comp_spv, kp::shader_data::op_getrows_q4_0_comp_spv_len); - ggml_vk_get_rows(spirv, "q4_0", 1/*We access blocks unaligned*/, QK4_0, std::forward(args)...); + ggml_vk_get_rows(ctx, spirv, "q4_0", 1/*We access blocks unaligned*/, QK4_0, std::forward(args)...); } template -static void ggml_vk_get_rows_q4_1(Args&&... args) { +static void ggml_vk_get_rows_q4_1(struct ggml_backend_kompute_context *ctx, Args&&... args) { const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_q4_1_comp_spv, kp::shader_data::op_getrows_q4_1_comp_spv_len); - ggml_vk_get_rows(spirv, "q4_1", 1/*We access blocks unaligned*/, QK4_1, std::forward(args)...); + ggml_vk_get_rows(ctx, spirv, "q4_1", 1/*We access blocks unaligned*/, QK4_1, std::forward(args)...); } template -static void ggml_vk_get_rows_q6_k(Args&&... args) { +static void ggml_vk_get_rows_q6_k(struct ggml_backend_kompute_context *ctx, Args&&... args) { const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_q6_k_comp_spv, kp::shader_data::op_getrows_q6_k_comp_spv_len); - ggml_vk_get_rows(spirv, "q6_k", 1/*We access blocks unaligned*/, QK_NL, std::forward(args)...); + ggml_vk_get_rows(ctx, spirv, "q6_k", 1/*We access blocks unaligned*/, QK_NL, std::forward(args)...); } static void ggml_vk_rope( + struct ggml_backend_kompute_context *ctx, kp::Sequence& seq, const std::shared_ptr& inA, const std::shared_ptr& inB, @@ -1237,23 +1320,24 @@ static void ggml_vk_rope( auto name = std::string(__func__) + (src0t == GGML_TYPE_F16 ? "_f16" : "_f32"); std::shared_ptr s_algo = nullptr; - if (!komputeManager()->hasAlgorithm(name)) { - s_algo = komputeManager()->algorithm( - name, s_kompute_context->pool.get(), {inA, inB, out}, + if (!ctx->manager.hasAlgorithm(name)) { + s_algo = ctx->manager.algorithm( + name, ctx->pool.get(), {inA, inB, out}, src0t == GGML_TYPE_F16 ? spirv_f16 : spirv_f32, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts} ); } else { - s_algo = komputeManager()->getAlgorithm(name); + s_algo = ctx->manager.getAlgorithm(name); s_algo->setTensors({inA, inB, out}); s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)}); s_algo->setPushConstants({pushConsts}); - s_algo->updateDescriptors(s_kompute_context->pool.get()); + s_algo->updateDescriptors(ctx->pool.get()); } seq.record(s_algo); } static void ggml_vk_cpy( + struct ggml_backend_kompute_context *ctx, const std::vector& spirv, uint32_t in_element_size, uint32_t out_element_size, kp::Sequence& seq, @@ -1283,44 +1367,44 @@ static void ggml_vk_cpy( + "_i_" + std::to_string(in_element_size) + "_o_" + std::to_string(out_element_size); std::shared_ptr s_algo = nullptr; - if (!komputeManager()->hasAlgorithm(name)) - s_algo = komputeManager()->algorithm(name, s_kompute_context->pool.get(), {in, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts}); + if (!ctx->manager.hasAlgorithm(name)) + s_algo = ctx->manager.algorithm(name, ctx->pool.get(), {in, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts}); else { - s_algo = komputeManager()->getAlgorithm(name); + s_algo = ctx->manager.getAlgorithm(name); s_algo->setTensors({in, out}); s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)}); s_algo->setPushConstants({pushConsts}); - s_algo->updateDescriptors(s_kompute_context->pool.get()); + s_algo->updateDescriptors(ctx->pool.get()); } seq.record(s_algo); } template -static void ggml_vk_cpy_f32_f16(Args&&... args) { +static void ggml_vk_cpy_f32_f16(struct ggml_backend_kompute_context *ctx, Args&&... args) { const static auto spirv = getSpirvShader(kp::shader_data::op_cpy_f32_f16_comp_spv, kp::shader_data::op_cpy_f32_f16_comp_spv_len); - ggml_vk_cpy(spirv, 4, 2, std::forward(args)...); + ggml_vk_cpy(ctx, spirv, 4, 2, std::forward(args)...); } template -static void ggml_vk_cpy_f32_f32(Args&&... args) { +static void ggml_vk_cpy_f32_f32(struct ggml_backend_kompute_context *ctx, Args&&... args) { const static auto spirv = getSpirvShader(kp::shader_data::op_cpy_f32_f32_comp_spv, kp::shader_data::op_cpy_f32_f32_comp_spv_len); - ggml_vk_cpy(spirv, 4, 4, std::forward(args)...); + ggml_vk_cpy(ctx, spirv, 4, 4, std::forward(args)...); } template -static void ggml_vk_cpy_f16_f16(Args&&... args) { +static void ggml_vk_cpy_f16_f16(struct ggml_backend_kompute_context *ctx, Args&&... args) { const static auto spirv = getSpirvShader(kp::shader_data::op_cpy_f16_f16_comp_spv, kp::shader_data::op_cpy_f16_f16_comp_spv_len); - ggml_vk_cpy(spirv, 2, 2, std::forward(args)...); + ggml_vk_cpy(ctx, spirv, 2, 2, std::forward(args)...); } template -static void ggml_vk_cpy_f16_f32(Args&&... args) { +static void ggml_vk_cpy_f16_f32(struct ggml_backend_kompute_context *ctx, Args&&... args) { const static auto spirv = getSpirvShader(kp::shader_data::op_cpy_f16_f32_comp_spv, kp::shader_data::op_cpy_f16_f32_comp_spv_len); - ggml_vk_cpy(spirv, 2, 4, std::forward(args)...); + ggml_vk_cpy(ctx, spirv, 2, 4, std::forward(args)...); } static bool ggml_vk_supports_op(const struct ggml_tensor * op) { @@ -1412,7 +1496,7 @@ static bool ggml_vk_supports_op(const struct ggml_tensor * op) { return false; } -static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml_cgraph * gf) { +static void ggml_vk_graph_compute(struct ggml_backend_kompute_context * ctx, struct ggml_cgraph * gf) { const int n_seq = 8; // FIXME: Figure out if we can somehow optimize the size of the pool... right now we're setting @@ -1422,7 +1506,7 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml std::vector> sequences(n_seq); for (auto& sequence : sequences) { - sequence = komputeManager()->sequence(); + sequence = ctx->manager.sequence(); } for (int seq_idx = 0; seq_idx < n_seq; ++seq_idx) { const int n_nodes_per_seq = (gf->n_nodes + n_seq - 1) / n_seq; @@ -1501,19 +1585,19 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml uint32_t off_src0 = 0; uint32_t off_src1 = 0; uint32_t off_dst = 0; - const std::shared_ptr& id_src0 = src0 ? ggml_vk_get_tensor(src0, &off_src0) : nullTensor; - const std::shared_ptr& id_src1 = src1 ? ggml_vk_get_tensor(src1, &off_src1) : nullTensor; - const std::shared_ptr& id_dst = dst ? ggml_vk_get_tensor(dst, &off_dst) : nullTensor; + const std::shared_ptr& id_src0 = src0 ? ggml_vk_get_tensor(ctx, src0, &off_src0) : nullTensor; + const std::shared_ptr& id_src1 = src1 ? ggml_vk_get_tensor(ctx, src1, &off_src1) : nullTensor; + const std::shared_ptr& id_dst = dst ? ggml_vk_get_tensor(ctx, dst, &off_dst) : nullTensor; switch (dst->op) { case GGML_OP_ADD: { if (ggml_nelements(src1) == ne10 && ggml_is_contiguous(src1) && ne00 % 4 == 0 && ne10 % 4 == 0) { // src1 is a row - ggml_vk_addrow(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ggml_nelements(dst)/4, ne00); + ggml_vk_addrow(ctx, seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ggml_nelements(dst)/4, ne00); } else { ggml_vk_add( - seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, + ctx, seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne10, ne11, ne12, ne13, @@ -1526,7 +1610,7 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml case GGML_OP_MUL: { ggml_vk_mul( - seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, + ctx, seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne10, ne11, ne12, ne13, @@ -1539,7 +1623,7 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml { float scale; memcpy(&scale, dst->op_params, sizeof(float)); - ggml_vk_scale(seq, id_src0, id_dst, off_src0, off_dst, ggml_nelements(dst), scale); + ggml_vk_scale(ctx, seq, id_src0, id_dst, off_src0, off_dst, ggml_nelements(dst), scale); } break; case GGML_OP_UNARY: { @@ -1548,16 +1632,16 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml switch (ggml_get_unary_op(gf->nodes[i])) { case GGML_UNARY_OP_SILU: { - ggml_vk_silu(seq, id_src0, id_dst, off_src0, off_dst, n/4); + ggml_vk_silu(ctx, seq, id_src0, id_dst, off_src0, off_dst, n/4); } break; case GGML_UNARY_OP_RELU: { - ggml_vk_relu(seq, id_src0, id_dst, off_src0, off_dst, n/4); + ggml_vk_relu(ctx, seq, id_src0, id_dst, off_src0, off_dst, n/4); } break; case GGML_UNARY_OP_GELU: { GGML_ASSERT(n % 8 == 0); - ggml_vk_gelu(seq, id_src0, id_dst, off_src0, off_dst, n/8); + ggml_vk_gelu(ctx, seq, id_src0, id_dst, off_src0, off_dst, n/8); } break; default: { @@ -1582,18 +1666,18 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml #pragma message("ref: https://github.com/ggerganov/llama.cpp/pull/7192") GGML_ASSERT(max_bias == 0.0f); - ggml_vk_soft_max(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, ne01, ne02, ne03, scale); + ggml_vk_soft_max(ctx, seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, ne01, ne02, ne03, scale); } break; case GGML_OP_DIAG_MASK_INF: { const int n_past = ((int32_t *)(dst->op_params))[0]; - ggml_vk_diag_mask_inf(seq, id_src0, id_dst, off_src0, off_dst, n_past, ne00, ne01, ne02); + ggml_vk_diag_mask_inf(ctx, seq, id_src0, id_dst, off_src0, off_dst, n_past, ne00, ne01, ne02); } break; case GGML_OP_NORM: { float eps; memcpy(&eps, dst->op_params, sizeof(float)); - ggml_vk_norm(seq, id_src0, id_dst, off_src0, off_dst, ne00, nb01, ggml_nrows(src0), eps); + ggml_vk_norm(ctx, seq, id_src0, id_dst, off_src0, off_dst, ne00, nb01, ggml_nrows(src0), eps); } break; case GGML_OP_RMS_NORM: { @@ -1601,7 +1685,7 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml float eps; memcpy(&eps, dst->op_params, sizeof(float)); - ggml_vk_rms_norm(seq, id_src0, id_dst, off_src0, off_dst, ne00, nb01, ggml_nrows(src0), eps); + ggml_vk_rms_norm(ctx, seq, id_src0, id_dst, off_src0, off_dst, ne00, nb01, ggml_nrows(src0), eps); } break; case GGML_OP_MUL_MAT: { @@ -1627,38 +1711,38 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml switch (src0t) { case GGML_TYPE_F32: ggml_vk_mul_mat_mat_f32( - seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, + ctx, seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, ne01, ne02, nb01, nb02, ne11, ne12, nb11, nb12, nb1, nb2 ); break; case GGML_TYPE_F16: ggml_vk_mul_mat_f16( - seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, + ctx, seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, ne01, ne02, nb00, nb01, nb02, ne10, ne11, ne12, ne13, nb10, nb11, nb12, ne0, ne1, r2, r3 ); break; case GGML_TYPE_Q8_0: ggml_vk_mul_mat_q8_0( - seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, + ctx, seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, r2, r3 ); break; case GGML_TYPE_Q4_0: ggml_vk_mul_mat_q4_0( - seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, + ctx, seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, r2, r3 ); break; case GGML_TYPE_Q4_1: ggml_vk_mul_mat_q4_1( - seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, + ctx, seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, r2, r3 ); break; case GGML_TYPE_Q6_K: ggml_vk_mul_mat_q6_k( - seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, + ctx, seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, ne10, ne0, ne1, ne01, ne11, ne12, ne02 ); break; @@ -1672,15 +1756,15 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml case GGML_OP_GET_ROWS: { if (src0t == GGML_TYPE_F32) { - ggml_vk_get_rows_f32(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1)); + ggml_vk_get_rows_f32(ctx, seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1)); } else if (src0t == GGML_TYPE_F16) { - ggml_vk_get_rows_f16(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1)); + ggml_vk_get_rows_f16(ctx, seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1)); } else if (src0t == GGML_TYPE_Q4_0) { - ggml_vk_get_rows_q4_0(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1)); + ggml_vk_get_rows_q4_0(ctx, seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1)); } else if (src0t == GGML_TYPE_Q4_1) { - ggml_vk_get_rows_q4_1(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1)); + ggml_vk_get_rows_q4_1(ctx, seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1)); } else if (src0t == GGML_TYPE_Q6_K) { - ggml_vk_get_rows_q6_k(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1)); + ggml_vk_get_rows_q6_k(ctx, seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1)); } else { fprintf(stderr, "%s: %s: Unsupported quantization: %u\n", __func__, ggml_op_name(dst->op), src0t); goto not_implemented; @@ -1711,7 +1795,7 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float)); memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float)); ggml_vk_rope( - seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, src0t, n_dims, mode, n_ctx_orig, + ctx, seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, src0t, n_dims, mode, n_ctx_orig, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, nb0, nb1, nb2, nb3 ); @@ -1724,16 +1808,16 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml case GGML_TYPE_F32: { switch (dstt) { - case GGML_TYPE_F16: ggml_vk_cpy_f32_f16(seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break; - case GGML_TYPE_F32: ggml_vk_cpy_f32_f32(seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break; + case GGML_TYPE_F16: ggml_vk_cpy_f32_f16(ctx, seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break; + case GGML_TYPE_F32: ggml_vk_cpy_f32_f32(ctx, seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break; default: goto not_implemented; } } break; case GGML_TYPE_F16: { switch (dstt) { - case GGML_TYPE_F16: ggml_vk_cpy_f16_f16(seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break; - case GGML_TYPE_F32: ggml_vk_cpy_f16_f32(seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break; + case GGML_TYPE_F16: ggml_vk_cpy_f16_f16(ctx, seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break; + case GGML_TYPE_F32: ggml_vk_cpy_f16_f32(ctx, seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break; default: goto not_implemented; } break; default: goto not_implemented; @@ -1781,90 +1865,54 @@ kp::TensorT::dataType() // backend interface -struct ggml_backend_kompute_buffer_type_context { - int device; - int device_ref = 0; - uint64_t buffer_alignment; - uint64_t max_alloc; - std::string name; - - ggml_backend_kompute_buffer_type_context(int device, uint64_t buffer_alignment, uint64_t max_alloc) - : device(device), buffer_alignment(buffer_alignment), max_alloc(max_alloc), name(ggml_kompute_format_name(device)) {} -}; - -static void ggml_backend_kompute_device_ref(ggml_backend_buffer_type_t buft) { - auto * ctx = static_cast(buft->context); - - if (!ctx->device_ref) { - komputeManager()->initializeDevice( - ctx->device, {}, { - "VK_KHR_shader_float16_int8", "VK_KHR_8bit_storage", - "VK_KHR_16bit_storage", "VK_KHR_shader_non_semantic_info" - } - ); - } - - assert(ggml_vk_has_device()); - ctx->device_ref++; -} - -static void ggml_backend_kompute_device_unref(ggml_backend_buffer_type_t buft) { - auto * ctx = static_cast(buft->context); - - assert(ctx->device_ref > 0); - - ctx->device_ref--; - - if (!ctx->device_ref) { - komputeManager.destroy(); - } -} - static const char * ggml_backend_kompute_buffer_get_name(ggml_backend_buffer_t buffer) { auto * ctx = static_cast(buffer->buft->context); return ctx->name.c_str(); } static void ggml_backend_kompute_buffer_free_buffer(ggml_backend_buffer_t buffer) { - auto * memory = (ggml_vk_memory *)buffer->context; - if (ggml_vk_has_device()) { - ggml_vk_free_memory(*memory); + auto * ctx = static_cast(buffer->context); + auto * backend_ctx = kompute_backend_context(buffer); + if (backend_ctx && ggml_vk_has_device(backend_ctx)) { + ggml_vk_free_memory(backend_ctx, ctx->memory); } - delete memory; + delete ctx; } static void * ggml_backend_kompute_buffer_get_base(ggml_backend_buffer_t buffer) { - return ((ggml_vk_memory *)buffer->context)->data; + auto * ctx = static_cast(buffer->context); + return ctx->memory.data; } static void ggml_backend_kompute_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) { - GGML_UNUSED(buffer); + auto * backend_ctx = kompute_backend_context(buffer); - const auto res = ggml_vk_get_tensor(tensor); + const auto res = ggml_vk_get_tensor(backend_ctx, tensor); GGML_ASSERT(res); memcpy((char *)tensor->data + offset, data, size); - komputeManager()->sequence()->eval({res}); + backend_ctx->manager.sequence()->eval({res}); } static void ggml_backend_kompute_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) { - GGML_UNUSED(buffer); + auto * backend_ctx = kompute_backend_context(buffer); - const auto res = ggml_vk_get_tensor(tensor); + const auto res = ggml_vk_get_tensor(backend_ctx, tensor); GGML_ASSERT(res); - komputeManager()->sequence()->eval({res}); + backend_ctx->manager.sequence()->eval({res}); memcpy(data, (const char *)tensor->data + offset, size); } static void ggml_backend_kompute_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { - auto * memory = (ggml_vk_memory *)buffer->context; - memset(memory->data, value, buffer->size); + auto * ctx = static_cast(buffer->context); + auto * backend_ctx = kompute_backend_context(buffer); + memset(ctx->memory.data, value, ctx->memory.size); - if (memory->stagingBuffer) - komputeManager()->sequence()->eval(memory->primaryBuffer, memory->stagingBuffer, memory->size); + if (ctx->memory.stagingBuffer) + backend_ctx->manager.sequence()->eval(ctx->memory.primaryBuffer, ctx->memory.stagingBuffer, ctx->memory.size); } static ggml_backend_buffer_i ggml_backend_kompute_buffer_i = { @@ -1887,8 +1935,9 @@ static const char * ggml_backend_kompute_buffer_type_get_name(ggml_backend_buffe } static ggml_backend_buffer_t ggml_backend_kompute_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { - ggml_backend_kompute_device_ref(buft); - auto * ctx = new ggml_vk_memory(ggml_vk_allocate(size)); + auto * backend_ctx = kompute_backend_context(buft); + auto * ctx = new ggml_backend_kompute_buffer_context; + ctx->memory = ggml_vk_allocate(backend_ctx, size); return ggml_backend_buffer_init(buft, ggml_backend_kompute_buffer_i, ctx, size); } @@ -1912,52 +1961,44 @@ static ggml_backend_buffer_type_i ggml_backend_kompute_buffer_type_interface = { }; ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device) { - static std::vector bufts = []() { - std::vector vec; - auto devices = ggml_vk_available_devices_internal(0); - vec.reserve(devices.size()); + auto * backend = komputeManager.create_backend(device); + auto * buft = &(static_cast(backend->context))->buft; - for (const auto & dev : devices) { - vec.push_back({ - /* .iface = */ ggml_backend_kompute_buffer_type_interface, - /* .context = */ new ggml_backend_kompute_buffer_type_context(dev.index, dev.bufferAlignment, dev.maxAlloc) - }); + if (!buft->context) { + auto devices = ggml_vk_available_devices_internal(0); + for (std::size_t i = 0; i < devices.size(); i++) { + if (device == devices[i].index) { + buft->context = new ggml_backend_kompute_buffer_type_context( + devices[i].index, + devices[i].bufferAlignment, + devices[i].maxAlloc); + buft->iface = ggml_backend_kompute_buffer_type_interface; + break; + } } - return vec; - }(); + } - auto it = std::find_if(bufts.begin(), bufts.end(), [device](const ggml_backend_buffer_type & t) { - return device == static_cast(t.context)->device; - }); - return it < bufts.end() ? &*it : nullptr; + return buft; } // backend static const char * ggml_backend_kompute_name(ggml_backend_t backend) { - auto * ctx = static_cast(backend->context); + auto * ctx = static_cast(backend->context); return ctx->name.c_str(); } static void ggml_backend_kompute_free(ggml_backend_t backend) { - auto * ctx = static_cast(backend->context); - - assert(ctx == s_kompute_context); - s_kompute_context = nullptr; - if (ctx != nullptr) { - delete ctx; - } - - delete backend; + komputeManager.destroy_backend(backend); } static ggml_backend_buffer_type_t ggml_backend_kompute_get_default_buffer_type(ggml_backend_t backend) { - auto * ctx = static_cast(backend->context); + auto * ctx = static_cast(backend->context); return ggml_backend_kompute_buffer_type(ctx->device); } static ggml_status ggml_backend_kompute_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) { - auto * ctx = static_cast(backend->context); + auto * ctx = static_cast(backend->context); ggml_vk_graph_compute(ctx, cgraph); return GGML_STATUS_SUCCESS; } @@ -1968,8 +2009,16 @@ static bool ggml_backend_kompute_supports_op(ggml_backend_t backend, const struc } static bool ggml_backend_kompute_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) { + auto *ctx = static_cast(backend->context); + return &ctx->buft == buft; +} + +static bool ggml_backend_kompute_offload_op(ggml_backend_t backend, const ggml_tensor * op) { GGML_UNUSED(backend); - return buft->iface.get_name == ggml_backend_kompute_buffer_type_get_name; + const int min_batch_size = 32; + + return (op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS) || + (op->ne[2] >= min_batch_size && op->op == GGML_OP_MUL_MAT_ID); } static struct ggml_backend_i kompute_backend_i = { @@ -1987,7 +2036,7 @@ static struct ggml_backend_i kompute_backend_i = { /* .graph_compute = */ ggml_backend_kompute_graph_compute, /* .supports_op = */ ggml_backend_kompute_supports_op, /* .supports_buft = */ ggml_backend_kompute_supports_buft, - /* .offload_op = */ NULL, + /* .offload_op = */ ggml_backend_kompute_offload_op, /* .event_new = */ NULL, /* .event_free = */ NULL, /* .event_record = */ NULL, @@ -2000,17 +2049,94 @@ static ggml_guid_t ggml_backend_kompute_guid() { return &guid; } -ggml_backend_t ggml_backend_kompute_init(int device) { - GGML_ASSERT(s_kompute_context == nullptr); - s_kompute_context = new ggml_kompute_context(device); - ggml_backend_t kompute_backend = new ggml_backend { + +kompute_manager::kompute_manager() : m_backends(GGML_KOMPUTE_MAX_DEVICES, nullptr) +{ + m_kp_manager = nullptr; +} + +kompute_manager::~kompute_manager() +{ + if (m_kp_manager) { + delete m_kp_manager; + m_kp_manager = nullptr; + } + + for (std::size_t i = 0; i < m_backends.size(); i++) { + destroy_backend(m_backends[i]); + } +} + +kp::Manager * kompute_manager::get_kp_manager(void) +{ + if (!m_kp_manager) + m_kp_manager = new kp::Manager; + + return m_kp_manager; +} + +ggml_backend_t kompute_manager::create_backend(int device) +{ + if (device < 0 || device >= GGML_KOMPUTE_MAX_DEVICES) + return nullptr; + + // already exist + ggml_backend_t backend = get_backend(device); + if (backend) + return backend; + + // create new one + auto *context = new ggml_backend_kompute_context(device); + context->manager.initializeDevice(device, {}, + { + "VK_KHR_shader_float16_int8", + "VK_KHR_8bit_storage", + "VK_KHR_16bit_storage", + "VK_KHR_shader_non_semantic_info" + }); + + backend = new ggml_backend { /* .guid = */ ggml_backend_kompute_guid(), /* .interface = */ kompute_backend_i, - /* .context = */ s_kompute_context, + /* .context = */ context, }; - return kompute_backend; + m_backends[device] = backend; + + std::cerr << "Kompute: Init device " << device << std::endl; + + return backend; +} + +void kompute_manager::destroy_backend(ggml_backend_t backend) +{ + if (!backend) + return; + + for (std::size_t i = 0; i < m_backends.size(); i++) { + if (backend == m_backends[i]) { + auto *context = static_cast(backend->context); + delete context; + delete backend; + m_backends[i] = nullptr; + break; + } + } +} + +ggml_backend_t kompute_manager::get_backend(int device) +{ + if (device >= 0 && static_cast(device) < m_backends.size()) + return m_backends[device]; + + return nullptr; +} + + + +ggml_backend_t ggml_backend_kompute_init(int device) { + return komputeManager.create_backend(device); } bool ggml_backend_is_kompute(ggml_backend_t backend) { diff --git a/src/llama.cpp b/src/llama.cpp index fe3c0db6f2931..28b2ad605e51a 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -2856,6 +2856,8 @@ static size_t llama_get_device_count(const llama_model & model) { count = ggml_backend_sycl_get_device_count(); #elif defined(GGML_USE_VULKAN) count = ggml_backend_vk_get_device_count(); +#elif defined(GGML_USE_KOMPUTE) + count = ggml_backend_kompute_get_device_count(); #elif defined(GGML_USE_CANN) return ggml_backend_cann_get_device_count(); #endif @@ -2952,6 +2954,11 @@ static size_t llama_get_device_memory(const llama_model & model, int device) { size_t free; ggml_backend_vk_get_device_memory(device, &free, &total); return free; +#elif defined(GGML_USE_KOMPUTE) + size_t total; + size_t free; + ggml_backend_kompute_get_device_memory(device, &free, &total); + return free; #elif defined(GGML_USE_CANN) size_t total; size_t free; @@ -16899,6 +16906,8 @@ size_t llama_max_devices(void) { return GGML_SYCL_MAX_DEVICES; #elif defined(GGML_USE_VULKAN) return GGML_VK_MAX_DEVICES; +#elif defined(GGML_USE_KOMPUTE) + return GGML_KOMPUTE_MAX_DEVICES; #elif defined(GGML_USE_CANN) return GGML_CANN_MAX_DEVICES; #else @@ -17234,13 +17243,35 @@ struct llama_context * llama_new_context_with_model( } #elif defined(GGML_USE_KOMPUTE) if (model->n_gpu_layers > 0) { - auto * backend = ggml_backend_kompute_init(model->main_gpu); - if (backend == nullptr) { - LLAMA_LOG_ERROR("%s: failed to initialize Kompute backend\n", __func__); + if (model->split_mode == LLAMA_SPLIT_MODE_NONE) { + auto * backend = ggml_backend_kompute_init(model->main_gpu); + if (!backend) { + LLAMA_LOG_ERROR("%s: failed to initialize Kompute backend\n", __func__); + llama_free(ctx); + return nullptr; + } + ctx->backends.push_back(backend); + } else if (model->split_mode == LLAMA_SPLIT_MODE_LAYER) { + size_t count = 0; + auto * devices =ggml_vk_available_devices(0, &count); + for (size_t i = 0; i < count; i++) { + LLAMA_LOG_INFO("Kompute: Found device #%d, %s, %s, max-alloc %ld, heap-size %lu\n", + devices[i].index, devices[i].vendor, devices[i].name, + devices[i].maxAlloc, devices[i].heapSize); + auto * backend = ggml_backend_kompute_init(devices[i].index); + if (!backend) { + LLAMA_LOG_ERROR("%s: failed to initialize Kompute backend\n", __func__); + llama_free(ctx); + return nullptr; + } + ctx->backends.push_back(backend); + } + free(devices); + } else { + LLAMA_LOG_ERROR("%s: Failed to init Kompute backend: split mode %d not supported\n", __func__, model->split_mode); llama_free(ctx); return nullptr; } - ctx->backends.push_back(backend); } #elif defined(GGML_USE_CANN) // with split_mode LLAMA_SPLIT_MODE_NONE or LLAMA_SPLIT_MODE_ROW, only the main GPU backend is used