From 3952a221af54b8a6549bc2bd4a7363ef7ad3081e Mon Sep 17 00:00:00 2001 From: Anthony Van de Gejuchte Date: Tue, 19 Nov 2024 23:18:17 +0100 Subject: [PATCH 1/5] Fix missing file renames in Makefile due to changes in commit ae8de6d50a (#10413) --- Makefile | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/Makefile b/Makefile index 95110d4eb..5c8994385 100644 --- a/Makefile +++ b/Makefile @@ -730,10 +730,10 @@ GLSLC_CMD = glslc _ggml_vk_genshaders_cmd = $(shell pwd)/vulkan-shaders-gen _ggml_vk_header = ggml/src/ggml-vulkan-shaders.hpp _ggml_vk_source = ggml/src/ggml-vulkan-shaders.cpp -_ggml_vk_input_dir = ggml/src/vulkan-shaders +_ggml_vk_input_dir = ggml/src/ggml-vulkan/vulkan-shaders _ggml_vk_shader_deps = $(echo $(_ggml_vk_input_dir)/*.comp) -ggml/src/ggml-vulkan.o: ggml/src/ggml-vulkan.cpp ggml/include/ggml-vulkan.h $(_ggml_vk_header) $(_ggml_vk_source) +ggml/src/ggml-vulkan.o: ggml/src/ggml-vulkan/ggml-vulkan.cpp ggml/include/ggml-vulkan.h $(_ggml_vk_header) $(_ggml_vk_source) $(CXX) $(CXXFLAGS) $(shell pkg-config --cflags vulkan) -c $< -o $@ $(_ggml_vk_header): $(_ggml_vk_source) @@ -745,8 +745,8 @@ $(_ggml_vk_source): $(_ggml_vk_shader_deps) vulkan-shaders-gen --target-hpp $(_ggml_vk_header) \ --target-cpp $(_ggml_vk_source) -vulkan-shaders-gen: ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp - $(CXX) $(CXXFLAGS) -o $@ $(LDFLAGS) ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp +vulkan-shaders-gen: ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp + $(CXX) $(CXXFLAGS) -o $@ $(LDFLAGS) ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp endif # GGML_VULKAN From ad21c9e1f14d82b8c15ae369a8839019e3d498b4 Mon Sep 17 00:00:00 2001 From: Neo Zhang Jianyu Date: Wed, 20 Nov 2024 13:54:25 +0800 Subject: [PATCH 2/5] update rel to 4040 (#10395) Co-authored-by: arthw <14088817+arthw@users.noreply.github.com> --- docs/backend/SYCL.md | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-) diff --git a/docs/backend/SYCL.md b/docs/backend/SYCL.md index e431f51f1..8d8312e91 100644 --- a/docs/backend/SYCL.md +++ b/docs/backend/SYCL.md @@ -34,9 +34,10 @@ The SYCL backend would be broken by some PRs due to no online CI. The following release is verified with good quality: -|Commit ID|Tag|Release|Verified Platform| -|-|-|-|-| -|fb76ec31a9914b7761c1727303ab30380fd4f05c|b3038 |[llama-b3038-bin-win-sycl-x64.zip](https://github.com/ggerganov/llama.cpp/releases/download/b3038/llama-b3038-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1
MTL Arc GPU/Windows 11/oneAPI 2024.1| +|Commit ID|Tag|Release|Verified Platform| Update date| +|-|-|-|-|-| +|3bcd40b3c593d14261fb2abfabad3c0fb5b9e318|b4040 |[llama-b4040-bin-win-sycl-x64.zip](https://github.com/ggerganov/llama.cpp/releases/download/b4040/llama-b4040-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1
MTL Arc GPU/Windows 11/oneAPI 2024.1| 2024-11-19| +|fb76ec31a9914b7761c1727303ab30380fd4f05c|b3038 |[llama-b3038-bin-win-sycl-x64.zip](https://github.com/ggerganov/llama.cpp/releases/download/b3038/llama-b3038-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1
MTL Arc GPU/Windows 11/oneAPI 2024.1|| ## News From 1bacb9f62514b520bdf74ed6feb46c80508dad38 Mon Sep 17 00:00:00 2001 From: Jeff Bolz Date: Wed, 20 Nov 2024 01:11:00 -0600 Subject: [PATCH 3/5] vulkan: further optimize mul_mat_vec using larger loads (#10387) * vulkan: Use pipeline_robustness to disable robustness in mul_mat_vec. Add some early returns for nonexistent rows in mul_mat_vec shaders. These can only be hit when dispatching a 2D grid of workgroups. Fix the logic for the 2D grid of workgroups to round up. Enable the pipeline robustness extension if it's available, and use it to disable robustness for these pipelines. The instructions to do the bounds checking contend for the same ALU resources as the bit twiddling dequant instructions. * vulkan: Add GLSL structure aliases for quant types to allow larger loads In Vulkan it's not possible to cast pointer types, so instead you have to declare an aliased binding for the memory with a different type. This commit adds aliases for the quant formats using 16b ints, and in a few places where the struct size is a multiple of 4 also using 32b ints. Currently only q4_k's aliases are used, but others will be used in subsequent commits. * vulkan: use larger loads in q5_k and q6_k shaders. Similar to the optimization I did in q4_k recently, this vectorizes some loads and reduces the number of bit twiddling instructions. * vulkan: use larger K step per iteration in mul_mat_vec. Add vec4 dequantization functions, and use them to do K=8 per iteration in mul_mat_vec. This uses 16b loads for the quant values and 128b loads for B which helps reduce the load on the memory system. The K_PER_ITER==2 logic is still there, just for F16/F32, and really only because they support unaligned sizes. Tweak the num_iters/unrolling logic to be simpler and catch a couple missed unrolling opportunities. --- ggml/src/ggml-vulkan/ggml-vulkan.cpp | 101 ++++++++----- .../vulkan-shaders/dequant_funcs.comp | 48 ++++++ .../vulkan-shaders/mul_mat_vec.comp | 82 ++++++++-- .../vulkan-shaders/mul_mat_vec_base.comp | 3 + .../vulkan-shaders/mul_mat_vec_q2_k.comp | 4 + .../vulkan-shaders/mul_mat_vec_q3_k.comp | 4 + .../vulkan-shaders/mul_mat_vec_q4_k.comp | 34 ++--- .../vulkan-shaders/mul_mat_vec_q5_k.comp | 140 ++++++++++++------ .../vulkan-shaders/mul_mat_vec_q6_k.comp | 69 ++++++--- .../src/ggml-vulkan/vulkan-shaders/types.comp | 113 +++++++++++++- .../vulkan-shaders/vulkan-shaders-gen.cpp | 6 +- 11 files changed, 457 insertions(+), 147 deletions(-) diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index 21fee2f3d..ca71da2f7 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -158,6 +158,7 @@ struct vk_device_struct { std::string name; uint64_t max_memory_allocation_size; bool fp16; + bool pipeline_robustness; vk::Device device; uint32_t vendor_id; vk_queue compute_queue; @@ -654,7 +655,7 @@ static uint32_t compile_count = 0; static std::mutex compile_count_mutex; static std::condition_variable compile_count_cond; -static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipeline, const std::string name, size_t spv_size, const void* spv_data, const std::string entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array wg_denoms, std::vector specialization_constants, uint32_t align) { +static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipeline, const std::string name, size_t spv_size, const void* spv_data, const std::string entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array wg_denoms, std::vector specialization_constants, uint32_t align, bool disable_robustness) { VK_LOG_DEBUG("ggml_vk_create_pipeline(" << device->name << ", " << name << ", " << entrypoint << ", " << parameter_count << ", " << push_constant_size << ", (" << wg_denoms[0] << "," << wg_denoms[1] << "," << wg_denoms[2] << "), specialization_constants, " << align << ")"); GGML_ASSERT(parameter_count > 0); GGML_ASSERT(wg_denoms[0] > 0 && wg_denoms[1] > 0 && wg_denoms[2] > 0); // NOLINT @@ -724,6 +725,15 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin vk::PipelineCreateFlags(), pipeline_shader_create_info, pipeline->layout); + + vk::PipelineRobustnessCreateInfoEXT rci; + + if (device->pipeline_robustness && disable_robustness) { + rci.storageBuffers = vk::PipelineRobustnessBufferBehaviorEXT::eDisabled; + rci.uniformBuffers = vk::PipelineRobustnessBufferBehaviorEXT::eDisabled; + compute_pipeline_create_info.setPNext(&rci); + } + pipeline->pipeline = device->device.createComputePipeline(VK_NULL_HANDLE, compute_pipeline_create_info).value; { @@ -1261,7 +1271,7 @@ static void ggml_vk_load_shaders(vk_device& device) { device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL] = std::make_shared(); std::vector> compiles; - auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const std::string &entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array wg_denoms, const std::vector& specialization_constants, uint32_t align) { + auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const std::string &entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array wg_denoms, const std::vector& specialization_constants, uint32_t align, bool disable_robustness = false) { { // wait until fewer than N compiles are in progress uint32_t N = std::max(1u, std::thread::hardware_concurrency()); @@ -1271,7 +1281,7 @@ static void ggml_vk_load_shaders(vk_device& device) { } compile_count++; } - compiles.push_back(std::async(ggml_vk_create_pipeline_func, std::ref(device), std::ref(pipeline), name, spv_size, spv_data, entrypoint, parameter_count, push_constant_size, wg_denoms, specialization_constants, align)); + compiles.push_back(std::async(ggml_vk_create_pipeline_func, std::ref(device), std::ref(pipeline), name, spv_size, spv_data, entrypoint, parameter_count, push_constant_size, wg_denoms, specialization_constants, align, disable_robustness)); }; if (device->fp16) { @@ -1370,45 +1380,45 @@ static void ggml_vk_load_shaders(vk_device& device) { // computing two rows per workgroup is a benefit for Q4_0 -> Q5_1, but not for Q8_0. ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f32_f32", mul_mat_vec_f32_f32_f32_len, mul_mat_vec_f32_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f32_f32", mul_mat_vec_f16_f32_f32_len, mul_mat_vec_f16_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f32_f32", mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f32_f32", mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f32_f32", mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f32_f32", mul_mat_vec_q5_1_f32_f32_len, mul_mat_vec_q5_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f32_f32", mul_mat_vec_q8_0_f32_f32_len, mul_mat_vec_q8_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f32_f32", mul_mat_vec_q2_k_f32_f32_len, mul_mat_vec_q2_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f32_f32", mul_mat_vec_q3_k_f32_f32_len, mul_mat_vec_q3_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f32_f32", mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f32_f32", mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f32_f32", mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f32_f32", mul_mat_vec_iq4_nl_f32_f32_len, mul_mat_vec_iq4_nl_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f32_f32", mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f32_f32", mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f32_f32", mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f32_f32", mul_mat_vec_q5_1_f32_f32_len, mul_mat_vec_q5_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f32_f32", mul_mat_vec_q8_0_f32_f32_len, mul_mat_vec_q8_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f32_f32", mul_mat_vec_q2_k_f32_f32_len, mul_mat_vec_q2_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f32_f32", mul_mat_vec_q3_k_f32_f32_len, mul_mat_vec_q3_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f32_f32", mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f32_f32", mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f32_f32", mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f32_f32", mul_mat_vec_iq4_nl_f32_f32_len, mul_mat_vec_iq4_nl_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f16_f32", mul_mat_vec_f32_f16_f32_len, mul_mat_vec_f32_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f16_f32", mul_mat_vec_f16_f16_f32_len, mul_mat_vec_f16_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f16_f32", mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f16_f32", mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f16_f32", mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f16_f32", mul_mat_vec_q5_1_f16_f32_len, mul_mat_vec_q5_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f16_f32", mul_mat_vec_q8_0_f16_f32_len, mul_mat_vec_q8_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f16_f32", mul_mat_vec_q2_k_f16_f32_len, mul_mat_vec_q2_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f16_f32", mul_mat_vec_q3_k_f16_f32_len, mul_mat_vec_q3_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f16_f32", mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f16_f32", mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f16_f32", mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f16_f32", mul_mat_vec_iq4_nl_f16_f32_len, mul_mat_vec_iq4_nl_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f16_f32", mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f16_f32", mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f16_f32", mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f16_f32", mul_mat_vec_q5_1_f16_f32_len, mul_mat_vec_q5_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f16_f32", mul_mat_vec_q8_0_f16_f32_len, mul_mat_vec_q8_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f16_f32", mul_mat_vec_q2_k_f16_f32_len, mul_mat_vec_q2_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f16_f32", mul_mat_vec_q3_k_f16_f32_len, mul_mat_vec_q3_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f16_f32", mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f16_f32", mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f16_f32", mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f16_f32", mul_mat_vec_iq4_nl_f16_f32_len, mul_mat_vec_iq4_nl_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32", mul_mat_vec_id_f32_f32_len, mul_mat_vec_id_f32_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F16 ], "mul_mat_vec_id_f16_f32", mul_mat_vec_id_f16_f32_len, mul_mat_vec_id_f16_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32", mul_mat_vec_id_q5_1_f32_len, mul_mat_vec_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32", mul_mat_vec_id_q8_0_f32_len, mul_mat_vec_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32", mul_mat_vec_id_q2_k_f32_len, mul_mat_vec_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32", mul_mat_vec_id_q3_k_f32_len, mul_mat_vec_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", mul_mat_vec_id_q4_k_f32_len, mul_mat_vec_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32", mul_mat_vec_id_q5_k_f32_len, mul_mat_vec_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_f32", mul_mat_vec_id_q6_k_f32_len, mul_mat_vec_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_id_iq4_nl_f32", mul_mat_vec_id_iq4_nl_f32_len, mul_mat_vec_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32", mul_mat_vec_id_q5_1_f32_len, mul_mat_vec_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32", mul_mat_vec_id_q8_0_f32_len, mul_mat_vec_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32", mul_mat_vec_id_q2_k_f32_len, mul_mat_vec_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32", mul_mat_vec_id_q3_k_f32_len, mul_mat_vec_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", mul_mat_vec_id_q4_k_f32_len, mul_mat_vec_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32", mul_mat_vec_id_q5_k_f32_len, mul_mat_vec_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_f32", mul_mat_vec_id_q6_k_f32_len, mul_mat_vec_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_id_iq4_nl_f32", mul_mat_vec_id_iq4_nl_f32_len, mul_mat_vec_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true); // dequant shaders ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_F32 ], "f32_to_f16", dequant_f32_len, dequant_f32_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1); @@ -1591,12 +1601,15 @@ static vk_device ggml_vk_get_device(size_t idx) { bool fp16_storage = false; bool fp16_compute = false; + bool pipeline_robustness = false; for (const auto& properties : ext_props) { if (strcmp("VK_KHR_16bit_storage", properties.extensionName) == 0) { fp16_storage = true; } else if (strcmp("VK_KHR_shader_float16_int8", properties.extensionName) == 0) { fp16_compute = true; + } else if (strcmp("VK_EXT_pipeline_robustness", properties.extensionName) == 0) { + pipeline_robustness = true; } } @@ -1642,10 +1655,22 @@ static vk_device ggml_vk_get_device(size_t idx) { vk12_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES; vk11_features.pNext = &vk12_features; + VkPhysicalDevicePipelineRobustnessFeaturesEXT pl_robustness_features; + pl_robustness_features.pNext = nullptr; + pl_robustness_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_ROBUSTNESS_FEATURES_EXT; + pl_robustness_features.pipelineRobustness = VK_FALSE; + + if (pipeline_robustness) { + vk12_features.pNext = &pl_robustness_features; + device_extensions.push_back("VK_EXT_pipeline_robustness"); + } + vkGetPhysicalDeviceFeatures2(device->physical_device, &device_features2); device->fp16 = device->fp16 && vk12_features.shaderFloat16; + device->pipeline_robustness = pl_robustness_features.pipelineRobustness; + if (!vk11_features.storageBuffer16BitAccess) { std::cerr << "ggml_vulkan: device " << GGML_VK_NAME << idx << " does not support 16-bit storage." << std::endl; throw std::runtime_error("Unsupported device"); @@ -3190,7 +3215,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context& if (ne01 > max_groups_x) { groups_z = 64; - groups_x /= groups_z; + groups_x = CEIL_DIV(groups_x, groups_z); } // compute @@ -3767,7 +3792,7 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte if (ne01 > max_groups_x) { groups_z = 64; - groups_x /= groups_z; + groups_x = CEIL_DIV(groups_x, groups_z); } // compute diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp index d5b989735..5fc1ba4ad 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp @@ -2,6 +2,15 @@ #extension GL_EXT_shader_explicit_arithmetic_types_int8 : require #endif +#include "types.comp" + +#if defined(A_TYPE_PACKED16) +layout (binding = 0) readonly buffer A_PACKED16 {A_TYPE_PACKED16 data_a_packed16[];}; +#endif +#if defined(A_TYPE_PACKED32) +layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32[];}; +#endif + #if defined(DATA_A_F32) vec2 dequantize(uint ib, uint iqs, uint a_offset) { return vec2(data_a[a_offset + ib], data_a[a_offset + ib + 1]); @@ -20,6 +29,11 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) { const uint vui = uint(data_a[a_offset + ib].qs[iqs]); return (vec2(vui & 0xF, vui >> 4) - 8.0f) * d; } +vec4 dequantize4(uint ib, uint iqs, uint a_offset) { + const float d = float(data_a_packed16[a_offset + ib].d); + const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]); + return (vec4(vui & 0xF, (vui >> 4) & 0xF, (vui >> 8) & 0xF, (vui >> 12) & 0xF) - 8.0f) * d; +} #endif #if defined(DATA_A_Q4_1) @@ -29,6 +43,12 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) { const uint vui = uint(data_a[a_offset + ib].qs[iqs]); return vec2(vui & 0xF, vui >> 4) * d + m; } +vec4 dequantize4(uint ib, uint iqs, uint a_offset) { + const float d = float(data_a_packed16[a_offset + ib].d); + const float m = float(data_a_packed16[a_offset + ib].m); + const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]); + return vec4(vui & 0xF, (vui >> 4) & 0xF, (vui >> 8) & 0xF, (vui >> 12) & 0xF) * d + m; +} #endif #if defined(DATA_A_Q5_0) @@ -39,6 +59,14 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) { const uint vui = uint(data_a[a_offset + ib].qs[iqs]); return (vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) - 16.0f) * d; } +vec4 dequantize4(uint ib, uint iqs, uint a_offset) { + const float d = float(data_a_packed16[a_offset + ib].d); + const uint uint_qh = uint(data_a_packed16[a_offset + ib].qh[1]) << 16 | data_a_packed16[a_offset + ib].qh[0]; + const ivec2 qh0 = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10); + const ivec2 qh1 = ivec2(((uint_qh >> (iqs + 1)) << 4) & 0x10, (uint_qh >> (iqs + 13)) & 0x10); + const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]); + return (vec4(((vui >> 0) & 0xF) | qh0.x, ((vui >> 4) & 0xF) | qh0.y, ((vui >> 8) & 0xF) | qh1.x, ((vui >> 12) & 0xF) | qh1.y) - 16.0f) * d; +} #endif #if defined(DATA_A_Q5_1) @@ -50,6 +78,15 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) { const uint vui = uint(data_a[a_offset + ib].qs[iqs]); return vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) * d + m; } +vec4 dequantize4(uint ib, uint iqs, uint a_offset) { + const float d = float(data_a_packed16[a_offset + ib].d); + const float m = float(data_a_packed16[a_offset + ib].m); + const uint uint_qh = data_a_packed16[a_offset + ib].qh; + const ivec2 qh0 = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10); + const ivec2 qh1 = ivec2(((uint_qh >> (iqs + 1)) << 4) & 0x10, (uint_qh >> (iqs + 13)) & 0x10); + const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]); + return vec4(((vui >> 0) & 0xF) | qh0.x, ((vui >> 4) & 0xF) | qh0.y, ((vui >> 8) & 0xF) | qh1.x, ((vui >> 12) & 0xF) | qh1.y) * d + m; +} #endif #if defined(DATA_A_Q8_0) @@ -57,6 +94,12 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) { const float d = float(data_a[a_offset + ib].d); return vec2(int(data_a[a_offset + ib].qs[iqs]), int(data_a[a_offset + ib].qs[iqs + 1])) * d; } +vec4 dequantize4(uint ib, uint iqs, uint a_offset) { + const float d = float(data_a_packed16[a_offset + ib].d); + uint32_t v0 = data_a_packed16[a_offset + ib].qs[iqs/2]; + uint32_t v1 = data_a_packed16[a_offset + ib].qs[iqs/2 + 1]; + return vec4(int8_t(v0 & 0xFF), int8_t((v0 >> 8) & 0xFF), int8_t(v1 & 0xFF), int8_t((v1 >> 8) & 0xFF)) * d; +} #endif #if defined(DATA_A_IQ4_NL) @@ -65,4 +108,9 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) { const uint vui = uint(data_a[a_offset + ib].qs[iqs]); return vec2(kvalues_iq4nl[vui & 0xF], kvalues_iq4nl[vui >> 4]) * d; } +vec4 dequantize4(uint ib, uint iqs, uint a_offset) { + const float d = float(data_a_packed16[a_offset + ib].d); + const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]); + return vec4(kvalues_iq4nl[vui & 0xF], kvalues_iq4nl[(vui >> 4) & 0xF], kvalues_iq4nl[(vui >> 8) & 0xF], kvalues_iq4nl[(vui >> 12) & 0xF]) * d; +} #endif diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp index 970aac6ef..00807a060 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp @@ -3,7 +3,7 @@ #ifdef FLOAT16 #extension GL_EXT_shader_explicit_arithmetic_types_float16 : require #endif -#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require +#extension GL_EXT_shader_explicit_arithmetic_types : require #include "mul_mat_vec_base.comp" @@ -12,16 +12,48 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; layout (constant_id = 0) const uint BLOCK_SIZE = 32; layout (constant_id = 1) const uint NUM_ROWS = 1; +#if !defined(DATA_A_F32) && !defined(DATA_A_F16) +#define K_PER_ITER 8 +#else +#define K_PER_ITER 2 +#endif + + uint a_offset, b_offset, d_offset, y_offset; shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE]; void iter(inout FLOAT_TYPE temp[NUM_ROWS], const uint first_row, const uint num_rows, const uint tid, const uint i, bool lastiter) { - const uint col = i*BLOCK_SIZE + 2*tid; + const uint col = i*BLOCK_SIZE + K_PER_ITER*tid; const uint iqs = (col%QUANT_K)/QUANT_R; // quant index const uint iybs = col - col%QUANT_K; // y block start index +#if K_PER_ITER == 8 +#if QUANT_R == 2 + B_TYPE_VEC4 bv02 = data_b_v4[(b_offset + iybs + iqs) / 4]; + B_TYPE_VEC4 bv13 = data_b_v4[(b_offset + iybs + iqs + y_offset) / 4]; + FLOAT_TYPE b0 = FLOAT_TYPE(bv02.x); + FLOAT_TYPE b1 = FLOAT_TYPE(bv13.x); + FLOAT_TYPE b2 = FLOAT_TYPE(bv02.y); + FLOAT_TYPE b3 = FLOAT_TYPE(bv13.y); + FLOAT_TYPE b4 = FLOAT_TYPE(bv02.z); + FLOAT_TYPE b5 = FLOAT_TYPE(bv13.z); + FLOAT_TYPE b6 = FLOAT_TYPE(bv02.w); + FLOAT_TYPE b7 = FLOAT_TYPE(bv13.w); +#else + B_TYPE_VEC4 bv0 = data_b_v4[(b_offset + iybs + iqs) / 4]; + B_TYPE_VEC4 bv1 = data_b_v4[(b_offset + iybs + iqs) / 4 + 1]; + FLOAT_TYPE b0 = FLOAT_TYPE(bv0.x); + FLOAT_TYPE b1 = FLOAT_TYPE(bv0.y); + FLOAT_TYPE b2 = FLOAT_TYPE(bv0.z); + FLOAT_TYPE b3 = FLOAT_TYPE(bv0.w); + FLOAT_TYPE b4 = FLOAT_TYPE(bv1.x); + FLOAT_TYPE b5 = FLOAT_TYPE(bv1.y); + FLOAT_TYPE b6 = FLOAT_TYPE(bv1.z); + FLOAT_TYPE b7 = FLOAT_TYPE(bv1.w); +#endif +#else // Check if the second of the pair of elements is OOB, and don't fetch B or // accumulate it. We still fetch a pair of elements for A, which is fine for // quantized formats since they'll be within the same block. We should @@ -34,9 +66,24 @@ void iter(inout FLOAT_TYPE temp[NUM_ROWS], const uint first_row, const uint num_ if (!OOB) { b1 = FLOAT_TYPE(data_b[b_offset + iybs + iqs + y_offset]); } +#endif [[unroll]] for (uint n = 0; n < num_rows; ++n) { const uint ib = ((first_row + n)*p.ncols + col)/QUANT_K; // block index +#if K_PER_ITER == 8 + const vec4 v = dequantize4(ib, iqs, a_offset); + const vec4 v2 = dequantize4(ib, iqs+(4/QUANT_R), a_offset); + + // matrix multiplication + temp[n] = fma(FLOAT_TYPE(v.x), b0, temp[n]); + temp[n] = fma(FLOAT_TYPE(v.y), b1, temp[n]); + temp[n] = fma(FLOAT_TYPE(v.z), b2, temp[n]); + temp[n] = fma(FLOAT_TYPE(v.w), b3, temp[n]); + temp[n] = fma(FLOAT_TYPE(v2.x), b4, temp[n]); + temp[n] = fma(FLOAT_TYPE(v2.y), b5, temp[n]); + temp[n] = fma(FLOAT_TYPE(v2.z), b6, temp[n]); + temp[n] = fma(FLOAT_TYPE(v2.w), b7, temp[n]); +#else const vec2 v = dequantize(ib, iqs, a_offset); // matrix multiplication @@ -44,6 +91,7 @@ void iter(inout FLOAT_TYPE temp[NUM_ROWS], const uint first_row, const uint num_ if (!OOB) { temp[n] = fma(FLOAT_TYPE(v.y), b1, temp[n]); } +#endif } } @@ -61,22 +109,33 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { temp[i] = FLOAT_TYPE(0); } - const int unroll_count = 8; - - const uint num_iters = (p.ncols >= 2*tid) ? ((p.ncols - 2*tid + BLOCK_SIZE - 1) / BLOCK_SIZE) : 0; - const uint unrolled_iters = num_iters & ~(2*unroll_count - 1); + uint num_iters = p.ncols / (K_PER_ITER * BLOCK_SIZE); + if (num_iters * K_PER_ITER * BLOCK_SIZE + K_PER_ITER*tid < p.ncols) { + num_iters++; + } + int unroll_count = 4; + uint unrolled_iters = num_iters & ~(unroll_count - 1); uint i = 0; while (i < unrolled_iters) { // Manually partially unroll the loop [[unroll]] for (uint k = 0; k < unroll_count; ++k) { - iter(temp, first_row, num_rows, tid, i, false); - i += 2; + iter(temp, first_row, num_rows, tid, i*K_PER_ITER, false); + i++; + } + } + unroll_count = 2; + unrolled_iters = num_iters & ~(unroll_count - 1); + while (i < unrolled_iters) { + // Manually partially unroll the loop + [[unroll]] for (uint k = 0; k < unroll_count; ++k) { + iter(temp, first_row, num_rows, tid, i*K_PER_ITER, false); + i++; } } while (i < num_iters) { - iter(temp, first_row, num_rows, tid, i, true); - i += 2; + iter(temp, first_row, num_rows, tid, i*K_PER_ITER, true); + i++; } // sum up partial sums and write back result @@ -106,6 +165,9 @@ void main() { if (first_row + NUM_ROWS <= p.stride_d) { compute_outputs(first_row, NUM_ROWS); } else { + if (first_row >= p.stride_d) { + return; + } compute_outputs(first_row, p.stride_d - first_row); } } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp index 5920bc936..8d0a57913 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp @@ -12,6 +12,9 @@ layout (binding = 0) readonly buffer A {A_TYPE data_a[];}; layout (binding = 1) readonly buffer B {B_TYPE data_b[];}; +layout (binding = 1) readonly buffer BV2 {B_TYPE_VEC2 data_b_v2[];}; +layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];}; + layout (binding = 2) writeonly buffer D {D_TYPE data_d[];}; #ifdef MUL_MAT_ID layout (binding = 3) readonly buffer IDS {int data_ids[];}; diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp index ec8eadcd5..e2625d32b 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp @@ -9,6 +9,10 @@ shared FLOAT_TYPE tmp[32]; void main() { const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z; + if (row >= p.stride_d) { + return; + } + uint a_offset, b_offset, d_offset; get_offsets(a_offset, b_offset, d_offset); diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp index 3ca4ad85a..a28804533 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp @@ -9,6 +9,10 @@ shared FLOAT_TYPE tmp[32]; void main() { const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z; + if (row >= p.stride_d) { + return; + } + uint a_offset, b_offset, d_offset; get_offsets(a_offset, b_offset, d_offset); diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp index b7c9b722d..5846f2e86 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp @@ -8,30 +8,14 @@ layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in; shared FLOAT_TYPE tmp[32]; -// Declare aliased versions of A and B bindings that can use 16b/32b loads for -// the quantized values, and vec4 loads for B. -struct block_q4_K_u32 -{ - f16vec2 d; - uint32_t scales[3*QUANT_K/64/4]; - uint32_t qs[QUANT_K/2/4]; -}; - -struct block_q4_K_u16 -{ - f16vec2 d; - uint16_t scales[3*QUANT_K/64/2]; - uint16_t qs[QUANT_K/2/2]; -}; - -layout (binding = 0) readonly buffer A_u32 {block_q4_K_u32 data_a_u32[];}; -layout (binding = 0) readonly buffer A_u16 {block_q4_K_u16 data_a_u16[];}; -layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];}; - // This shader assumes K_QUANTS_PER_ITERATION == 2 for alignment of loads void main() { const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z; + if (row >= p.stride_d) { + return; + } + uint a_offset, b_offset, d_offset; get_offsets(a_offset, b_offset, d_offset); @@ -64,9 +48,9 @@ void main() { const FLOAT_TYPE dall = FLOAT_TYPE(d.x); const FLOAT_TYPE dmin = FLOAT_TYPE(d.y); - uint32_t scale0_u32 = data_a_u16[ib0 + i].scales[v_im ]; - uint32_t scale4_u32 = data_a_u16[ib0 + i].scales[v_im + 2]; - uint32_t scale8_u32 = data_a_u16[ib0 + i].scales[v_im + 4]; + uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im ]; + uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2]; + uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4]; uvec4 scale0 = uvec4(unpack8(scale0_u32)); uvec4 scale4 = uvec4(unpack8(scale4_u32)); uvec4 scale8 = uvec4(unpack8(scale8_u32)); @@ -80,8 +64,8 @@ void main() { const uint32_t sc6 = (((scale8.x >> 4) & 0x0f) | ((scale4.x & 0xc0) >> 2)); const uint32_t sc7 = (((scale8.y >> 4) & 0x0f) | ((scale4.y & 0xc0) >> 2)); - uint32_t qs0_u32 = data_a_u32[ib0 + i].qs[q_offset / 4]; - uint32_t qs64_u32 = data_a_u32[ib0 + i].qs[q_offset / 4 + 16]; + uint32_t qs0_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4]; + uint32_t qs64_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4 + 16]; uint32_t qs0_u32_lo4 = qs0_u32 & 0x0F0F0F0F; uint32_t qs0_u32_hi4 = (qs0_u32 >> 4) & 0x0F0F0F0F; diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp index 2306785af..22a6bfae4 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp @@ -1,5 +1,7 @@ #version 450 +#extension GL_EXT_shader_explicit_arithmetic_types : require + #include "mul_mat_vec_base.comp" layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in; @@ -9,6 +11,10 @@ shared FLOAT_TYPE tmp[32]; void main() { const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z; + if (row >= p.stride_d) { + return; + } + uint a_offset, b_offset, d_offset; get_offsets(a_offset, b_offset, d_offset); @@ -31,70 +37,106 @@ void main() { const uint8_t hm1 = uint8_t(1 << (2*v_im)); const uint8_t hm2 = uint8_t(hm1 << 4); - tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp + FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += 2) { const uint y1_idx = i * QUANT_K + y_offset; const uint y2_idx = y1_idx + 128; - const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib0 + i].d.x); - const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib0 + i].d.y); + f16vec2 d = data_a[ib0 + i].d; + const FLOAT_TYPE dall = FLOAT_TYPE(d.x); + const FLOAT_TYPE dmin = FLOAT_TYPE(d.y); - const uint8_t sc0 = uint8_t( data_a[ib0 + i].scales[v_im * 2 ] & 0x3f); - const uint8_t sc1 = uint8_t( data_a[ib0 + i].scales[v_im * 2 + 1] & 0x3f); - const uint8_t sc2 = uint8_t( data_a[ib0 + i].scales[v_im * 2 + 4] & 0x3f); - const uint8_t sc3 = uint8_t( data_a[ib0 + i].scales[v_im * 2 + 5] & 0x3f); - const uint8_t sc4 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 8] & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 ] & 0xc0) >> 2)); - const uint8_t sc5 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 9] & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 1] & 0xc0) >> 2)); - const uint8_t sc6 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 8] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 4] & 0xc0) >> 2)); - const uint8_t sc7 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 9] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 5] & 0xc0) >> 2)); + uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im ]; + uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2]; + uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4]; + uvec4 scale0 = uvec4(unpack8(scale0_u32)); + uvec4 scale4 = uvec4(unpack8(scale4_u32)); + uvec4 scale8 = uvec4(unpack8(scale8_u32)); - const uint8_t q4_0 = uint8_t(data_a[ib0 + i].qs[q_offset ] & 0xf); - const uint8_t q4_1 = uint8_t(data_a[ib0 + i].qs[q_offset + 1] & 0xf); - const uint8_t q4_2 = uint8_t(data_a[ib0 + i].qs[q_offset + 16] & 0xf); - const uint8_t q4_3 = uint8_t(data_a[ib0 + i].qs[q_offset + 17] & 0xf); - const uint8_t q4_4 = uint8_t(data_a[ib0 + i].qs[q_offset ] >> 4); - const uint8_t q4_5 = uint8_t(data_a[ib0 + i].qs[q_offset + 1] >> 4); - const uint8_t q4_6 = uint8_t(data_a[ib0 + i].qs[q_offset + 16] >> 4); - const uint8_t q4_7 = uint8_t(data_a[ib0 + i].qs[q_offset + 17] >> 4); - const uint8_t q4_8 = uint8_t(data_a[ib0 + i].qs[q_offset + 64] & 0xf); - const uint8_t q4_9 = uint8_t(data_a[ib0 + i].qs[q_offset + 65] & 0xf); - const uint8_t q4_10 = uint8_t(data_a[ib0 + i].qs[q_offset + 80] & 0xf); - const uint8_t q4_11 = uint8_t(data_a[ib0 + i].qs[q_offset + 81] & 0xf); - const uint8_t q4_12 = uint8_t(data_a[ib0 + i].qs[q_offset + 64] >> 4); - const uint8_t q4_13 = uint8_t(data_a[ib0 + i].qs[q_offset + 65] >> 4); - const uint8_t q4_14 = uint8_t(data_a[ib0 + i].qs[q_offset + 80] >> 4); - const uint8_t q4_15 = uint8_t(data_a[ib0 + i].qs[q_offset + 81] >> 4); + const uint32_t sc0 = ( scale0.x & 0x3f); + const uint32_t sc1 = ( scale0.y & 0x3f); + const uint32_t sc2 = ( scale4.x & 0x3f); + const uint32_t sc3 = ( scale4.y & 0x3f); + const uint32_t sc4 = (( scale8.x & 0x0f) | ((scale0.x & 0xc0) >> 2)); + const uint32_t sc5 = (( scale8.y & 0x0f) | ((scale0.y & 0xc0) >> 2)); + const uint32_t sc6 = (((scale8.x >> 4) & 0x0f) | ((scale4.x & 0xc0) >> 2)); + const uint32_t sc7 = (((scale8.y >> 4) & 0x0f) | ((scale4.y & 0xc0) >> 2)); + + uint32_t qs0_16_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 8]) << 16); + uint32_t qs64_80_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 32]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 40]) << 16); + + uint32_t qs0_16_u32_lo4 = qs0_16_u32 & 0x0F0F0F0F; + uint32_t qs0_16_u32_hi4 = (qs0_16_u32 >> 4) & 0x0F0F0F0F; + uint32_t qs64_80_u32_lo4 = qs64_80_u32 & 0x0F0F0F0F; + uint32_t qs64_80_u32_hi4 = (qs64_80_u32 >> 4) & 0x0F0F0F0F; + + uvec4 qs0_16_lo4 = uvec4(unpack8(qs0_16_u32_lo4)); + uvec4 qs64_80_lo4 = uvec4(unpack8(qs64_80_u32_lo4)); + uvec4 qs0_16_hi4 = uvec4(unpack8(qs0_16_u32_hi4)); + uvec4 qs64_80_hi4 = uvec4(unpack8(qs64_80_u32_hi4)); + + const uint32_t q4_0 = qs0_16_lo4.x; + const uint32_t q4_1 = qs0_16_lo4.y; + const uint32_t q4_2 = qs0_16_lo4.z; + const uint32_t q4_3 = qs0_16_lo4.w; + const uint32_t q4_4 = qs0_16_hi4.x; + const uint32_t q4_5 = qs0_16_hi4.y; + const uint32_t q4_6 = qs0_16_hi4.z; + const uint32_t q4_7 = qs0_16_hi4.w; + const uint32_t q4_8 = qs64_80_lo4.x; + const uint32_t q4_9 = qs64_80_lo4.y; + const uint32_t q4_10 = qs64_80_lo4.z; + const uint32_t q4_11 = qs64_80_lo4.w; + const uint32_t q4_12 = qs64_80_hi4.x; + const uint32_t q4_13 = qs64_80_hi4.y; + const uint32_t q4_14 = qs64_80_hi4.z; + const uint32_t q4_15 = qs64_80_hi4.w; + + B_TYPE_VEC2 by10 = data_b_v2[(b_offset + y1_idx) / 2]; + B_TYPE_VEC2 by116 = data_b_v2[(b_offset + y1_idx) / 2 + 8]; + B_TYPE_VEC2 by132 = data_b_v2[(b_offset + y1_idx) / 2 + 16]; + B_TYPE_VEC2 by148 = data_b_v2[(b_offset + y1_idx) / 2 + 24]; + B_TYPE_VEC2 by20 = data_b_v2[(b_offset + y2_idx) / 2]; + B_TYPE_VEC2 by216 = data_b_v2[(b_offset + y2_idx) / 2 + 8]; + B_TYPE_VEC2 by232 = data_b_v2[(b_offset + y2_idx) / 2 + 16]; + B_TYPE_VEC2 by248 = data_b_v2[(b_offset + y2_idx) / 2 + 24]; + + uint32_t qh0 = data_a_packed16[ib0 + i].qh[l0 / 2]; + uint32_t qh1 = qh0 >> 8; + uint32_t qh16 = data_a_packed16[ib0 + i].qh[l0 / 2 + 8]; + uint32_t qh17 = qh16 >> 8; const FLOAT_TYPE sx = - fma(FLOAT_TYPE(data_b[b_offset + y1_idx ]), (q4_0 + (((data_a[ib0 + i].qh[l0 ] & hm1) != 0) ? 16 : 0)), - fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 1]), (q4_1 + (((data_a[ib0 + i].qh[l0 + 1] & hm1) != 0) ? 16 : 0)), - fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 16]), (q4_2 + (((data_a[ib0 + i].qh[l0 + 16] & hm1) != 0) ? 16 : 0)), - FLOAT_TYPE(data_b[b_offset + y1_idx + 17]) * (q4_3 + (((data_a[ib0 + i].qh[l0 + 17] & hm1) != 0) ? 16 : 0))))); + fma(FLOAT_TYPE(by10.x), (q4_0 + (((qh0 & hm1) != 0) ? 16 : 0)), + fma(FLOAT_TYPE(by10.y), (q4_1 + (((qh1 & hm1) != 0) ? 16 : 0)), + fma(FLOAT_TYPE(by116.x), (q4_2 + (((qh16 & hm1) != 0) ? 16 : 0)), + FLOAT_TYPE(by116.y) * (q4_3 + (((qh17 & hm1) != 0) ? 16 : 0))))); const FLOAT_TYPE sy = - fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), (q4_4 + (((data_a[ib0 + i].qh[l0 ] & (hm1 << 1)) != 0) ? 16 : 0)), - fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 33]), (q4_5 + (((data_a[ib0 + i].qh[l0 + 1] & (hm1 << 1)) != 0) ? 16 : 0)), - fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 48]), (q4_6 + (((data_a[ib0 + i].qh[l0 + 16] & (hm1 << 1)) != 0) ? 16 : 0)), - FLOAT_TYPE(data_b[b_offset + y1_idx + 49]) * (q4_7 + (((data_a[ib0 + i].qh[l0 + 17] & (hm1 << 1)) != 0) ? 16 : 0))))); + fma(FLOAT_TYPE(by132.x), (q4_4 + (((qh0 & (hm1 << 1)) != 0) ? 16 : 0)), + fma(FLOAT_TYPE(by132.y), (q4_5 + (((qh1 & (hm1 << 1)) != 0) ? 16 : 0)), + fma(FLOAT_TYPE(by148.x), (q4_6 + (((qh16 & (hm1 << 1)) != 0) ? 16 : 0)), + FLOAT_TYPE(by148.y) * (q4_7 + (((qh17 & (hm1 << 1)) != 0) ? 16 : 0))))); const FLOAT_TYPE sz = - fma(FLOAT_TYPE(data_b[b_offset + y2_idx ]), (q4_8 + (((data_a[ib0 + i].qh[l0 ] & hm2) != 0) ? 16 : 0)), - fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 1]), (q4_9 + (((data_a[ib0 + i].qh[l0 + 1] & hm2) != 0) ? 16 : 0)), - fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 16]), (q4_10 + (((data_a[ib0 + i].qh[l0 + 16] & hm2) != 0) ? 16 : 0)), - FLOAT_TYPE(data_b[b_offset + y2_idx + 17]) * (q4_11 + (((data_a[ib0 + i].qh[l0 + 17] & hm2) != 0) ? 16 : 0))))); + fma(FLOAT_TYPE(by20.x), (q4_8 + (((qh0 & hm2) != 0) ? 16 : 0)), + fma(FLOAT_TYPE(by20.y), (q4_9 + (((qh1 & hm2) != 0) ? 16 : 0)), + fma(FLOAT_TYPE(by216.x), (q4_10 + (((qh16 & hm2) != 0) ? 16 : 0)), + FLOAT_TYPE(by216.y) * (q4_11 + (((qh17 & hm2) != 0) ? 16 : 0))))); const FLOAT_TYPE sw = - fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), (q4_12 + (((data_a[ib0 + i].qh[l0 ] & (hm2 << 1)) != 0) ? 16 : 0)), - fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 33]), (q4_13 + (((data_a[ib0 + i].qh[l0 + 1] & (hm2 << 1)) != 0) ? 16 : 0)), - fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 48]), (q4_14 + (((data_a[ib0 + i].qh[l0 + 16] & (hm2 << 1)) != 0) ? 16 : 0)), - FLOAT_TYPE(data_b[b_offset + y2_idx + 49]) * (q4_15 + (((data_a[ib0 + i].qh[l0 + 17] & (hm2 << 1)) != 0) ? 16 : 0))))); + fma(FLOAT_TYPE(by232.x), (q4_12 + (((qh0 & (hm2 << 1)) != 0) ? 16 : 0)), + fma(FLOAT_TYPE(by232.y), (q4_13 + (((qh1 & (hm2 << 1)) != 0) ? 16 : 0)), + fma(FLOAT_TYPE(by248.x), (q4_14 + (((qh16 & (hm2 << 1)) != 0) ? 16 : 0)), + FLOAT_TYPE(by248.y) * (q4_15 + (((qh17 & (hm2 << 1)) != 0) ? 16 : 0))))); const FLOAT_TYPE smin = - fma(FLOAT_TYPE(data_b[b_offset + y1_idx ]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 1 ]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 16]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 17]), sc2, - fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 48]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 49]), sc3, - fma(FLOAT_TYPE(data_b[b_offset + y2_idx ]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 1 ]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 16]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 17]), sc6, - (FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 48]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 49])) * sc7))); - const uint tmp_idx = 16 * ix + tid; - tmp[tmp_idx] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, tmp[tmp_idx])); + fma(FLOAT_TYPE(by10.x) + FLOAT_TYPE(by10.y) + FLOAT_TYPE(by116.x) + FLOAT_TYPE(by116.y), sc2, + fma(FLOAT_TYPE(by132.x) + FLOAT_TYPE(by132.y) + FLOAT_TYPE(by148.x) + FLOAT_TYPE(by148.y), sc3, + fma(FLOAT_TYPE(by20.x) + FLOAT_TYPE(by20.y) + FLOAT_TYPE(by216.x) + FLOAT_TYPE(by216.y), sc6, + (FLOAT_TYPE(by232.x) + FLOAT_TYPE(by232.y) + FLOAT_TYPE(by248.x) + FLOAT_TYPE(by248.y)) * sc7))); + temp = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp)); } + tmp[gl_LocalInvocationID.x] = temp; + // sum up partial sums and write back result barrier(); [[unroll]] for (uint s = 16; s > 0; s >>= 1) { diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp index 95c286eeb..0b392d68d 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp @@ -1,5 +1,7 @@ #version 450 +#extension GL_EXT_shader_explicit_arithmetic_types : require + #include "mul_mat_vec_base.comp" layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in; @@ -9,6 +11,10 @@ shared FLOAT_TYPE tmp[32]; void main() { const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z; + if (row >= p.stride_d) { + return; + } + uint a_offset, b_offset, d_offset; get_offsets(a_offset, b_offset, d_offset); @@ -36,41 +42,66 @@ void main() { const uint s_offset = 8*v_im + is; const uint y_offset = 128*v_im + l0; - tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp + FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { const uint y_idx = i * QUANT_K + y_offset; const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d); -#if K_QUANTS_PER_ITERATION == 1 - const uint tmp_idx = 16 * ix + tid; - tmp[tmp_idx] = fma(FLOAT_TYPE(data_b[b_offset + y_idx + 0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 0] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 0] & 0x03) << 4)) - 32), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + 16]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 1]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x03) << 4)) - 32), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + 32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 0] & 0x0c) << 2)) - 32), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + 48]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 3]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x0c) << 2)) - 32), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + 64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 0] >> 4) | ((data_a[ib0 + i].qh[qh_offset + 0] & 0x30) >> 0)) - 32), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + 80]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 5]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16] >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x30) >> 0)) - 32), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + 96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32] >> 4) | ((data_a[ib0 + i].qh[qh_offset + 0] & 0xc0) >> 2)) - 32), - fma(FLOAT_TYPE(data_b[b_offset + y_idx +112]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 7]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48] >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0xc0) >> 2)) - 32), tmp[tmp_idx])))))))); -#else + FLOAT_TYPE scales[4]; + scales[0] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]); + scales[1] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]); + scales[2] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]); + scales[3] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]); + + uint32_t ql0_u32 = uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 1]) << 16); + uint32_t ql32_u32 = uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 16]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 17]) << 16); + + uint32_t ql0_u32_lo4 = ql0_u32 & 0x0F0F0F0F; + uint32_t ql0_u32_hi4 = (ql0_u32 >> 4) & 0x0F0F0F0F; + uint32_t ql32_u32_lo4 = ql32_u32 & 0x0F0F0F0F; + uint32_t ql32_u32_hi4 = (ql32_u32 >> 4) & 0x0F0F0F0F; + + uint32_t qh_u32 = uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2 + 1]) << 16); + uint32_t qh0_u32 = (qh_u32 & 0x03030303) << 4; + uint32_t qh2_u32 = (qh_u32 & 0x0C0C0C0C) << 2; + uint32_t qh4_u32 = (qh_u32 & 0x30303030) << 0; + uint32_t qh6_u32 = (qh_u32 & 0xC0C0C0C0) >> 2; + + uint32_t q0_u32 = ql0_u32_lo4 | qh0_u32; + uint32_t q1_u32 = ql32_u32_lo4 | qh2_u32; + uint32_t q2_u32 = ql0_u32_hi4 | qh4_u32; + uint32_t q3_u32 = ql32_u32_hi4 | qh6_u32; + + uvec4 q0 = uvec4(unpack8(q0_u32)); + uvec4 q1 = uvec4(unpack8(q1_u32)); + uvec4 q2 = uvec4(unpack8(q2_u32)); + uvec4 q3 = uvec4(unpack8(q3_u32)); + + B_TYPE_VEC4 by0 = data_b_v4[(b_offset + y_idx) / 4]; + B_TYPE_VEC4 by32 = data_b_v4[(b_offset + y_idx) / 4 + 8]; + B_TYPE_VEC4 by64 = data_b_v4[(b_offset + y_idx) / 4 + 16]; + B_TYPE_VEC4 by96 = data_b_v4[(b_offset + y_idx) / 4 + 24]; + FLOAT_TYPE sum = FLOAT_TYPE(0.0); [[unroll]] for (int l = 0; l < 4; ++l) { - sum = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+ 0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 0) & 3) << 4)) - 32), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 2) & 3) << 4)) - 32), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0] >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 4) & 3) << 4)) - 32), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32] >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 6) & 3) << 4)) - 32), sum)))); + sum = fma(FLOAT_TYPE(by0[l]) * scales[0], FLOAT_TYPE(int8_t(q0[l]) - 32), + fma(FLOAT_TYPE(by32[l]) * scales[1], FLOAT_TYPE(int8_t(q1[l]) - 32), + fma(FLOAT_TYPE(by64[l]) * scales[2], FLOAT_TYPE(int8_t(q2[l]) - 32), + fma(FLOAT_TYPE(by96[l]) * scales[3], FLOAT_TYPE(int8_t(q3[l]) - 32), sum)))); } - tmp[16 * ix + tid] += sum; -#endif + temp += sum * d; } + tmp[gl_LocalInvocationID.x] = temp; + // sum up partial sums and write back result barrier(); [[unroll]] for (uint s = 16; s > 0; s >>= 1) { if (tid < s) { tmp[tid] += tmp[tid + s]; - } + } barrier(); } if (tid == 0) { diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/types.comp b/ggml/src/ggml-vulkan/vulkan-shaders/types.comp index 21dce72fc..7a34820bc 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/types.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/types.comp @@ -1,6 +1,8 @@ -#if !defined(DATA_A_F32) && !defined(DATA_A_F16) -#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require -#endif + +#if !defined(GGML_TYPES_COMP) +#define GGML_TYPES_COMP + +#extension GL_EXT_shader_explicit_arithmetic_types : require #if defined(DATA_A_F32) #define QUANT_K 1 @@ -38,8 +40,14 @@ struct block_q4_0 float16_t d; uint8_t qs[16]; }; +struct block_q4_0_packed16 +{ + float16_t d; + uint16_t qs[16/2]; +}; #define A_TYPE block_q4_0 +#define A_TYPE_PACKED16 block_q4_0_packed16 #endif #if defined(DATA_A_Q4_1) @@ -54,7 +62,15 @@ struct block_q4_1 uint8_t qs[16]; }; +struct block_q4_1_packed16 +{ + float16_t d; + float16_t m; + uint16_t qs[16/2]; +}; + #define A_TYPE block_q4_1 +#define A_TYPE_PACKED16 block_q4_1_packed16 #endif #if defined(DATA_A_Q5_0) @@ -70,7 +86,15 @@ struct block_q5_0 uint8_t qs[16]; }; +struct block_q5_0_packed16 +{ + float16_t d; + uint16_t qh[2]; + uint16_t qs[16/2]; +}; + #define A_TYPE block_q5_0 +#define A_TYPE_PACKED16 block_q5_0_packed16 #endif #if defined(DATA_A_Q5_1) @@ -87,7 +111,16 @@ struct block_q5_1 uint8_t qs[16]; }; +struct block_q5_1_packed16 +{ + float16_t d; + float16_t m; + uint qh; + uint16_t qs[16/2]; +}; + #define A_TYPE block_q5_1 +#define A_TYPE_PACKED16 block_q5_1_packed16 #endif #if defined(DATA_A_Q8_0) @@ -100,8 +133,14 @@ struct block_q8_0 float16_t d; int8_t qs[32]; }; +struct block_q8_0_packed16 +{ + float16_t d; + uint16_t qs[32/2]; +}; #define A_TYPE block_q8_0 +#define A_TYPE_PACKED16 block_q8_0_packed16 #endif // K-quants @@ -116,7 +155,23 @@ struct block_q2_K f16vec2 d; }; +struct block_q2_K_packed16 +{ + uint16_t scales[QUANT_K/16/2]; + uint16_t qs[QUANT_K/4/2]; + f16vec2 d; +}; + +struct block_q2_K_packed32 +{ + uint32_t scales[QUANT_K/16/4]; + uint32_t qs[QUANT_K/4/4]; + f16vec2 d; +}; + #define A_TYPE block_q2_K +#define A_TYPE_PACKED16 block_q2_K_packed16 +#define A_TYPE_PACKED32 block_q2_K_packed32 #endif #if defined(DATA_A_Q3_K) @@ -131,7 +186,16 @@ struct block_q3_K float16_t d; }; +struct block_q3_K_packed16 +{ + uint16_t hmask[QUANT_K/8/2]; + uint16_t qs[QUANT_K/4/2]; + uint16_t scales[12/2]; + float16_t d; +}; + #define A_TYPE block_q3_K +#define A_TYPE_PACKED16 block_q3_K_packed16 #endif #if defined(DATA_A_Q4_K) @@ -145,7 +209,23 @@ struct block_q4_K uint8_t qs[QUANT_K/2]; }; +struct block_q4_K_packed16 +{ + f16vec2 d; + uint16_t scales[3*QUANT_K/64/2]; + uint16_t qs[QUANT_K/2/2]; +}; + +struct block_q4_K_packed32 +{ + f16vec2 d; + uint32_t scales[3*QUANT_K/64/4]; + uint32_t qs[QUANT_K/2/4]; +}; + #define A_TYPE block_q4_K +#define A_TYPE_PACKED16 block_q4_K_packed16 +#define A_TYPE_PACKED32 block_q4_K_packed32 #endif #if defined(DATA_A_Q5_K) @@ -160,7 +240,16 @@ struct block_q5_K uint8_t qs[QUANT_K/2]; }; +struct block_q5_K_packed16 +{ + f16vec2 d; + uint16_t scales[12/2]; + uint16_t qh[QUANT_K/8/2]; + uint16_t qs[QUANT_K/2/2]; +}; + #define A_TYPE block_q5_K +#define A_TYPE_PACKED16 block_q5_K_packed16 #endif #if defined(DATA_A_Q6_K) @@ -175,7 +264,16 @@ struct block_q6_K float16_t d; }; +struct block_q6_K_packed16 +{ + uint16_t ql[QUANT_K/2/2]; + uint16_t qh[QUANT_K/4/2]; + int8_t scales[QUANT_K/16]; + float16_t d; +}; + #define A_TYPE block_q6_K +#define A_TYPE_PACKED16 block_q6_K_packed16 #endif // IQuants @@ -191,10 +289,19 @@ struct block_iq4_nl uint8_t qs[QUANT_K/2]; }; +struct block_iq4_nl_packed16 +{ + float16_t d; + uint16_t qs[QUANT_K/2/2]; +}; + #define A_TYPE block_iq4_nl +#define A_TYPE_PACKED16 block_iq4_nl_packed16 const int8_t kvalues_iq4nl[16] = { int8_t(-127), int8_t(-104), int8_t(-83), int8_t(-65), int8_t(-49), int8_t(-35), int8_t(-22), int8_t(-10), int8_t(1), int8_t(13), int8_t(25), int8_t(38), int8_t(53), int8_t(69), int8_t(89), int8_t(113) }; #endif + +#endif // !defined(GGML_TYPES_COMP) diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp index fe3e4cb39..f75310955 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp @@ -317,10 +317,10 @@ void process_shaders() { std::string data_a_key = "DATA_A_" + to_uppercase(tname); std::string shader = (string_ends_with(tname, "_k")) ? "mul_mat_vec_" + tname + ".comp" : "mul_mat_vec.comp"; - string_to_spv("mul_mat_vec_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}})); - string_to_spv("mul_mat_vec_" + tname + "_f16_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPE_VEC4", "f16vec4"}, {"D_TYPE", "float"}})); + string_to_spv("mul_mat_vec_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}})); + string_to_spv("mul_mat_vec_" + tname + "_f16_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPE_VEC2", "f16vec2"}, {"B_TYPE_VEC4", "f16vec4"}, {"D_TYPE", "float"}})); - string_to_spv("mul_mat_vec_id_" + tname + "_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}})); + string_to_spv("mul_mat_vec_id_" + tname + "_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}})); // Dequant shaders if (tname != "f16") { From 8fd4b7fa29c3061b2e02e897d818dfcbc593430a Mon Sep 17 00:00:00 2001 From: Jeff Bolz Date: Wed, 20 Nov 2024 01:40:18 -0600 Subject: [PATCH 4/5] vulkan: copy iq4_nl LUT into shared memory (#10409) --- .../ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp | 2 ++ .../ggml-vulkan/vulkan-shaders/get_rows_quant.comp | 4 ++++ .../src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp | 4 ++++ ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp | 4 ++++ ggml/src/ggml-vulkan/vulkan-shaders/types.comp | 13 ++++++++++++- .../vulkan-shaders/vulkan-shaders-gen.cpp | 6 +++--- 6 files changed, 29 insertions(+), 4 deletions(-) diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp index 34ef3da30..8de14fc03 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp @@ -10,6 +10,8 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];}; void main() { const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64; + init_iq4nl_shmem(); + const uint tid = gl_LocalInvocationID.x % 64; const uint il = tid/32; const uint ir = tid%32; diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp b/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp index 8d30b63c1..7f608315b 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp @@ -12,6 +12,10 @@ void main() { const uint i11 = (gl_GlobalInvocationID.z)/p.ne12; const uint i12 = (gl_GlobalInvocationID.z)%p.ne12; +#if defined(DATA_A_IQ4_NL) + init_iq4nl_shmem(); +#endif + if (i00 >= p.ne00) { return; } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp index 00807a060..2d5b8e466 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp @@ -161,6 +161,10 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { void main() { const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z); +#if defined(DATA_A_IQ4_NL) + init_iq4nl_shmem(); +#endif + // do NUM_ROWS at a time, unless there aren't enough remaining rows if (first_row + NUM_ROWS <= p.stride_d) { compute_outputs(first_row, NUM_ROWS); diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp index fffdd1818..2ff5c4305 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp @@ -75,6 +75,10 @@ shared u16vec2 row_ids[3072]; #endif void main() { +#if defined(DATA_A_IQ4_NL) + init_iq4nl_shmem(); +#endif + #ifdef MUL_MAT_ID const uint expert_idx = gl_GlobalInvocationID.z; #else diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/types.comp b/ggml/src/ggml-vulkan/vulkan-shaders/types.comp index 7a34820bc..bc28e0ab8 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/types.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/types.comp @@ -298,10 +298,21 @@ struct block_iq4_nl_packed16 #define A_TYPE block_iq4_nl #define A_TYPE_PACKED16 block_iq4_nl_packed16 -const int8_t kvalues_iq4nl[16] = { +const int8_t kvalues_iq4nl_const[16] = { int8_t(-127), int8_t(-104), int8_t(-83), int8_t(-65), int8_t(-49), int8_t(-35), int8_t(-22), int8_t(-10), int8_t(1), int8_t(13), int8_t(25), int8_t(38), int8_t(53), int8_t(69), int8_t(89), int8_t(113) }; + +shared FLOAT_TYPE kvalues_iq4nl[16]; + +void init_iq4nl_shmem() +{ + // copy the table into shared memory and sync + if (gl_LocalInvocationIndex.x < 16) { + kvalues_iq4nl[gl_LocalInvocationIndex.x] = FLOAT_TYPE(kvalues_iq4nl_const[gl_LocalInvocationIndex.x]); + } + barrier(); +} #endif #endif // !defined(GGML_TYPES_COMP) diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp index f75310955..6bbe8e96e 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp @@ -331,11 +331,11 @@ void process_shaders() { shader = (tname == "f32" || tname == "f16") ? "get_rows.comp" : "get_rows_quant.comp"; if (tname == "f16") { - string_to_spv("get_rows_" + tname, shader, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}}); + string_to_spv("get_rows_" + tname, shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}})); } else { - string_to_spv("get_rows_" + tname, shader, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}}); + string_to_spv("get_rows_" + tname, shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}})); } - string_to_spv("get_rows_" + tname + "_f32", shader, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float"}}); + string_to_spv("get_rows_" + tname + "_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float"}})); } } From fab5d30ff6729ff6ff615c41e8c0215d6bc30393 Mon Sep 17 00:00:00 2001 From: Diego Devesa Date: Wed, 20 Nov 2024 12:57:53 +0100 Subject: [PATCH 5/5] llama : add .clang-format file (#10415) --- .clang-format | 161 ++++++ examples/llama-bench/llama-bench.cpp | 700 +++++++++++++++------------ 2 files changed, 550 insertions(+), 311 deletions(-) create mode 100644 .clang-format diff --git a/.clang-format b/.clang-format new file mode 100644 index 000000000..45232b80e --- /dev/null +++ b/.clang-format @@ -0,0 +1,161 @@ +--- +Language: Cpp +AlignAfterOpenBracket: Align +AlignArrayOfStructures: Left +AlignConsecutiveAssignments: AcrossComments +AlignConsecutiveBitFields: AcrossComments +AlignConsecutiveDeclarations: AcrossComments +AlignConsecutiveMacros: AcrossComments +# AlignConsecutiveShortCaseStatements: AcrossComments +AlignEscapedNewlines: Left # LeftWithLastLine +AlignOperands: Align +AlignTrailingComments: + Kind: Always + OverEmptyLines: 1 +AllowAllArgumentsOnNextLine: true +AllowAllParametersOfDeclarationOnNextLine: false +# AllowBreakBeforeNoexceptSpecifier: OnlyWithParen +AllowShortBlocksOnASingleLine: Never +AllowShortCaseLabelsOnASingleLine: false +AllowShortFunctionsOnASingleLine: Inline +AllowShortIfStatementsOnASingleLine: Never +AllowShortLambdasOnASingleLine: Inline +AllowShortLoopsOnASingleLine: false +AlwaysBreakBeforeMultilineStrings: true +BinPackArguments: true +BinPackParameters: true # OnePerLine +BitFieldColonSpacing: Both +BreakBeforeBraces: Custom # Attach +BraceWrapping: + AfterCaseLabel: true + AfterClass: false + AfterControlStatement: false + AfterEnum: false + AfterFunction: false + AfterNamespace: false + AfterObjCDeclaration: false + AfterStruct: false + AfterUnion: false + AfterExternBlock: false + BeforeCatch: false + BeforeElse: false + BeforeLambdaBody: false + BeforeWhile: false + IndentBraces: false + SplitEmptyFunction: false + SplitEmptyRecord: false + SplitEmptyNamespace: false +# BreakAdjacentStringLiterals: true +BreakAfterAttributes: Never +BreakBeforeBinaryOperators: None +BreakBeforeInlineASMColon: OnlyMultiline +BreakBeforeTernaryOperators: false +# BreakBinaryOperations: Never +BreakConstructorInitializers: AfterColon +# BreakFunctionDefinitionParameters: false +BreakInheritanceList: AfterComma +BreakStringLiterals: true +# BreakTemplateDeclarations: Yes +ColumnLimit: 120 +CommentPragmas: '^ IWYU pragma:' +CompactNamespaces: false +ConstructorInitializerIndentWidth: 4 +ContinuationIndentWidth: 4 +Cpp11BracedListStyle: false +DerivePointerAlignment: false +DisableFormat: false +EmptyLineBeforeAccessModifier: Leave +EmptyLineAfterAccessModifier: Never +ExperimentalAutoDetectBinPacking: false +FixNamespaceComments: true +IncludeBlocks: Regroup +IncludeCategories: + - Regex: '^<.*\.h>' + Priority: 1 + SortPriority: 0 + - Regex: '^<.*' + Priority: 2 + SortPriority: 0 + - Regex: '.*' + Priority: 3 + SortPriority: 0 +IncludeIsMainRegex: '([-_](test|unittest))?$' +IncludeIsMainSourceRegex: '' +IndentAccessModifiers: false +IndentCaseBlocks: true +IndentCaseLabels: true +IndentExternBlock: NoIndent +IndentGotoLabels: false +IndentPPDirectives: AfterHash +IndentWidth: 4 +IndentWrappedFunctionNames: false +InsertBraces: true # NOTE: may lead to incorrect formatting +InsertNewlineAtEOF: true +JavaScriptQuotes: Leave +JavaScriptWrapImports: true +KeepEmptyLinesAtTheStartOfBlocks: false +LambdaBodyIndentation: Signature +LineEnding: LF +MacroBlockBegin: '' +MacroBlockEnd: '' +MaxEmptyLinesToKeep: 1 +NamespaceIndentation: None +ObjCBinPackProtocolList: Auto +ObjCBlockIndentWidth: 4 +ObjCSpaceAfterProperty: true +ObjCSpaceBeforeProtocolList: true +PPIndentWidth: -1 +PackConstructorInitializers: CurrentLine +PenaltyBreakAssignment: 2 +PenaltyBreakBeforeFirstCallParameter: 1 +PenaltyBreakComment: 300 +PenaltyBreakFirstLessLess: 120 +PenaltyBreakString: 1000 +PenaltyBreakTemplateDeclaration: 10 +PenaltyExcessCharacter: 1000000 +PenaltyReturnTypeOnItsOwnLine: 200 +PointerAlignment: Middle +QualifierAlignment: Left +#QualifierOrder: ['static', 'inline', 'friend', 'constexpr', 'const', 'volatile', 'type', 'restrict'] +RawStringFormats: + - Language: Cpp + Delimiters: + - cc + - CC + - cpp + - Cpp + - CPP + - 'c++' + - 'C++' + CanonicalDelimiter: '' +ReferenceAlignment: Middle +ReflowComments: false # IndentOnly +SeparateDefinitionBlocks: Always +SortIncludes: CaseInsensitive +SortUsingDeclarations: LexicographicNumeric +SpaceAfterCStyleCast: true +SpaceAfterLogicalNot: false +SpaceAfterTemplateKeyword: true +SpaceBeforeAssignmentOperators: true +SpaceBeforeCpp11BracedList: false +SpaceBeforeCtorInitializerColon: true +SpaceBeforeInheritanceColon: true +SpaceBeforeParens: ControlStatements +SpaceBeforeRangeBasedForLoopColon: true +SpaceInEmptyBlock: false +SpaceInEmptyParentheses: false +SpacesBeforeTrailingComments: 2 +SpacesInAngles: Never +SpacesInContainerLiterals: true +SpacesInLineCommentPrefix: + Minimum: 1 + Maximum: -1 +SpacesInParentheses: false +SpacesInSquareBrackets: false +SpaceBeforeSquareBrackets: false +Standard: c++17 +TabWidth: 4 +UseTab: Never +WhitespaceSensitiveMacros: ['STRINGIZE'] +... + diff --git a/examples/llama-bench/llama-bench.cpp b/examples/llama-bench/llama-bench.cpp index 8f4e0e206..3dc84a75c 100644 --- a/examples/llama-bench/llama-bench.cpp +++ b/examples/llama-bench/llama-bench.cpp @@ -6,28 +6,28 @@ #include #include #include +#include #include #include -#include #include #include #include #include #include #include -#include #include +#include +#include "common.h" #include "ggml.h" #include "llama.h" -#include "common.h" #ifdef _WIN32 -#define WIN32_LEAN_AND_MEAN -#ifndef NOMINMAX -# define NOMINMAX -#endif -#include +# define WIN32_LEAN_AND_MEAN +# ifndef NOMINMAX +# define NOMINMAX +# endif +# include #endif // utils @@ -36,8 +36,7 @@ static uint64_t get_time_ns() { return std::chrono::nanoseconds(clock::now().time_since_epoch()).count(); } -template -static std::string join(const std::vector & values, const std::string & delim) { +template static std::string join(const std::vector & values, const std::string & delim) { std::ostringstream str; for (size_t i = 0; i < values.size(); i++) { str << values[i]; @@ -48,38 +47,35 @@ static std::string join(const std::vector & values, const std::string & delim return str.str(); } -template -static std::vector transform_to_str(const std::vector & values, F f) { +template static std::vector transform_to_str(const std::vector & values, F f) { std::vector str_values; std::transform(values.begin(), values.end(), std::back_inserter(str_values), f); return str_values; } -template -static T avg(const std::vector & v) { +template static T avg(const std::vector & v) { if (v.empty()) { return 0; } T sum = std::accumulate(v.begin(), v.end(), T(0)); - return sum / (T)v.size(); + return sum / (T) v.size(); } -template -static T stdev(const std::vector & v) { +template static T stdev(const std::vector & v) { if (v.size() <= 1) { return 0; } - T mean = avg(v); + T mean = avg(v); T sq_sum = std::inner_product(v.begin(), v.end(), v.begin(), T(0)); - T stdev = std::sqrt(sq_sum / (T)(v.size() - 1) - mean * mean * (T)v.size() / (T)(v.size() - 1)); + T stdev = std::sqrt(sq_sum / (T) (v.size() - 1) - mean * mean * (T) v.size() / (T) (v.size() - 1)); return stdev; } static std::string get_cpu_info() { std::vector cpu_list; for (size_t i = 0; i < ggml_backend_dev_count(); i++) { - auto * dev = ggml_backend_dev_get(i); - auto dev_type = ggml_backend_dev_type(dev); + auto * dev = ggml_backend_dev_get(i); + auto dev_type = ggml_backend_dev_type(dev); if (dev_type == GGML_BACKEND_DEVICE_TYPE_CPU || dev_type == GGML_BACKEND_DEVICE_TYPE_ACCEL) { cpu_list.push_back(ggml_backend_dev_description(dev)); } @@ -90,8 +86,8 @@ static std::string get_cpu_info() { static std::string get_gpu_info() { std::vector gpu_list; for (size_t i = 0; i < ggml_backend_dev_count(); i++) { - auto * dev = ggml_backend_dev_get(i); - auto dev_type = ggml_backend_dev_type(dev); + auto * dev = ggml_backend_dev_get(i); + auto dev_type = ggml_backend_dev_type(dev); if (dev_type == GGML_BACKEND_DEVICE_TYPE_GPU) { gpu_list.push_back(ggml_backend_dev_description(dev)); } @@ -100,17 +96,24 @@ static std::string get_gpu_info() { } // command line params -enum output_formats {NONE, CSV, JSON, JSONL, MARKDOWN, SQL}; +enum output_formats { NONE, CSV, JSON, JSONL, MARKDOWN, SQL }; static const char * output_format_str(output_formats format) { switch (format) { - case NONE: return "none"; - case CSV: return "csv"; - case JSON: return "json"; - case JSONL: return "jsonl"; - case MARKDOWN: return "md"; - case SQL: return "sql"; - default: GGML_ABORT("invalid output format"); + case NONE: + return "none"; + case CSV: + return "csv"; + case JSON: + return "json"; + case JSONL: + return "jsonl"; + case MARKDOWN: + return "md"; + case SQL: + return "sql"; + default: + GGML_ABORT("invalid output format"); } } @@ -135,10 +138,14 @@ static bool output_format_from_str(const std::string & s, output_formats & forma static const char * split_mode_str(llama_split_mode mode) { switch (mode) { - case LLAMA_SPLIT_MODE_NONE: return "none"; - case LLAMA_SPLIT_MODE_LAYER: return "layer"; - case LLAMA_SPLIT_MODE_ROW: return "row"; - default: GGML_ABORT("invalid split mode"); + case LLAMA_SPLIT_MODE_NONE: + return "none"; + case LLAMA_SPLIT_MODE_LAYER: + return "layer"; + case LLAMA_SPLIT_MODE_ROW: + return "row"; + default: + GGML_ABORT("invalid split mode"); } } @@ -149,59 +156,59 @@ static std::string pair_str(const std::pair & p) { } struct cmd_params { - std::vector model; - std::vector n_prompt; - std::vector n_gen; + std::vector model; + std::vector n_prompt; + std::vector n_gen; std::vector> n_pg; - std::vector n_batch; - std::vector n_ubatch; - std::vector type_k; - std::vector type_v; - std::vector n_threads; - std::vector cpu_mask; - std::vector cpu_strict; - std::vector poll; - std::vector n_gpu_layers; - std::vector rpc_servers; - std::vector split_mode; - std::vector main_gpu; - std::vector no_kv_offload; - std::vector flash_attn; - std::vector> tensor_split; - std::vector use_mmap; - std::vector embeddings; - ggml_numa_strategy numa; - int reps; - ggml_sched_priority prio; - int delay; - bool verbose; - bool progress; - output_formats output_format; - output_formats output_format_stderr; + std::vector n_batch; + std::vector n_ubatch; + std::vector type_k; + std::vector type_v; + std::vector n_threads; + std::vector cpu_mask; + std::vector cpu_strict; + std::vector poll; + std::vector n_gpu_layers; + std::vector rpc_servers; + std::vector split_mode; + std::vector main_gpu; + std::vector no_kv_offload; + std::vector flash_attn; + std::vector> tensor_split; + std::vector use_mmap; + std::vector embeddings; + ggml_numa_strategy numa; + int reps; + ggml_sched_priority prio; + int delay; + bool verbose; + bool progress; + output_formats output_format; + output_formats output_format_stderr; }; static const cmd_params cmd_params_defaults = { - /* model */ {"models/7B/ggml-model-q4_0.gguf"}, - /* n_prompt */ {512}, - /* n_gen */ {128}, + /* model */ { "models/7B/ggml-model-q4_0.gguf" }, + /* n_prompt */ { 512 }, + /* n_gen */ { 128 }, /* n_pg */ {}, - /* n_batch */ {2048}, - /* n_ubatch */ {512}, - /* type_k */ {GGML_TYPE_F16}, - /* type_v */ {GGML_TYPE_F16}, - /* n_threads */ {cpu_get_num_math()}, - /* cpu_mask */ {"0x0"}, - /* cpu_strict */ {false}, - /* poll */ {50}, - /* n_gpu_layers */ {99}, - /* rpc_servers */ {""}, - /* split_mode */ {LLAMA_SPLIT_MODE_LAYER}, - /* main_gpu */ {0}, - /* no_kv_offload */ {false}, - /* flash_attn */ {false}, - /* tensor_split */ {std::vector(llama_max_devices(), 0.0f)}, - /* use_mmap */ {true}, - /* embeddings */ {false}, + /* n_batch */ { 2048 }, + /* n_ubatch */ { 512 }, + /* type_k */ { GGML_TYPE_F16 }, + /* type_v */ { GGML_TYPE_F16 }, + /* n_threads */ { cpu_get_num_math() }, + /* cpu_mask */ { "0x0" }, + /* cpu_strict */ { false }, + /* poll */ { 50 }, + /* n_gpu_layers */ { 99 }, + /* rpc_servers */ { "" }, + /* split_mode */ { LLAMA_SPLIT_MODE_LAYER }, + /* main_gpu */ { 0 }, + /* no_kv_offload */ { false }, + /* flash_attn */ { false }, + /* tensor_split */ { std::vector(llama_max_devices(), 0.0f) }, + /* use_mmap */ { true }, + /* embeddings */ { false }, /* numa */ GGML_NUMA_STRATEGY_DISABLED, /* reps */ 5, /* prio */ GGML_SCHED_PRIO_NORMAL, @@ -218,38 +225,59 @@ static void print_usage(int /* argc */, char ** argv) { printf("options:\n"); printf(" -h, --help\n"); printf(" -m, --model (default: %s)\n", join(cmd_params_defaults.model, ",").c_str()); - printf(" -p, --n-prompt (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str()); + printf(" -p, --n-prompt (default: %s)\n", + join(cmd_params_defaults.n_prompt, ",").c_str()); printf(" -n, --n-gen (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str()); - printf(" -pg (default: %s)\n", join(transform_to_str(cmd_params_defaults.n_pg, pair_str), ",").c_str()); - printf(" -b, --batch-size (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str()); - printf(" -ub, --ubatch-size (default: %s)\n", join(cmd_params_defaults.n_ubatch, ",").c_str()); - printf(" -ctk, --cache-type-k (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str()); - printf(" -ctv, --cache-type-v (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str()); - printf(" -t, --threads (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str()); - printf(" -C, --cpu-mask (default: %s)\n", join(cmd_params_defaults.cpu_mask, ",").c_str()); - printf(" --cpu-strict <0|1> (default: %s)\n", join(cmd_params_defaults.cpu_strict, ",").c_str()); + printf(" -pg (default: %s)\n", + join(transform_to_str(cmd_params_defaults.n_pg, pair_str), ",").c_str()); + printf(" -b, --batch-size (default: %s)\n", + join(cmd_params_defaults.n_batch, ",").c_str()); + printf(" -ub, --ubatch-size (default: %s)\n", + join(cmd_params_defaults.n_ubatch, ",").c_str()); + printf(" -ctk, --cache-type-k (default: %s)\n", + join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str()); + printf(" -ctv, --cache-type-v (default: %s)\n", + join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str()); + printf(" -t, --threads (default: %s)\n", + join(cmd_params_defaults.n_threads, ",").c_str()); + printf(" -C, --cpu-mask (default: %s)\n", + join(cmd_params_defaults.cpu_mask, ",").c_str()); + printf(" --cpu-strict <0|1> (default: %s)\n", + join(cmd_params_defaults.cpu_strict, ",").c_str()); printf(" --poll <0...100> (default: %s)\n", join(cmd_params_defaults.poll, ",").c_str()); - printf(" -ngl, --n-gpu-layers (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str()); + printf(" -ngl, --n-gpu-layers (default: %s)\n", + join(cmd_params_defaults.n_gpu_layers, ",").c_str()); if (llama_supports_rpc()) { - printf(" -rpc, --rpc (default: %s)\n", join(cmd_params_defaults.rpc_servers, ",").c_str()); + printf(" -rpc, --rpc (default: %s)\n", + join(cmd_params_defaults.rpc_servers, ",").c_str()); } - printf(" -sm, --split-mode (default: %s)\n", join(transform_to_str(cmd_params_defaults.split_mode, split_mode_str), ",").c_str()); - printf(" -mg, --main-gpu (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str()); - printf(" -nkvo, --no-kv-offload <0|1> (default: %s)\n", join(cmd_params_defaults.no_kv_offload, ",").c_str()); - printf(" -fa, --flash-attn <0|1> (default: %s)\n", join(cmd_params_defaults.flash_attn, ",").c_str()); - printf(" -mmp, --mmap <0|1> (default: %s)\n", join(cmd_params_defaults.use_mmap, ",").c_str()); + printf(" -sm, --split-mode (default: %s)\n", + join(transform_to_str(cmd_params_defaults.split_mode, split_mode_str), ",").c_str()); + printf(" -mg, --main-gpu (default: %s)\n", + join(cmd_params_defaults.main_gpu, ",").c_str()); + printf(" -nkvo, --no-kv-offload <0|1> (default: %s)\n", + join(cmd_params_defaults.no_kv_offload, ",").c_str()); + printf(" -fa, --flash-attn <0|1> (default: %s)\n", + join(cmd_params_defaults.flash_attn, ",").c_str()); + printf(" -mmp, --mmap <0|1> (default: %s)\n", + join(cmd_params_defaults.use_mmap, ",").c_str()); printf(" --numa (default: disabled)\n"); - printf(" -embd, --embeddings <0|1> (default: %s)\n", join(cmd_params_defaults.embeddings, ",").c_str()); + printf(" -embd, --embeddings <0|1> (default: %s)\n", + join(cmd_params_defaults.embeddings, ",").c_str()); printf(" -ts, --tensor-split (default: 0)\n"); printf(" -r, --repetitions (default: %d)\n", cmd_params_defaults.reps); printf(" --prio <0|1|2|3> (default: %d)\n", cmd_params_defaults.prio); printf(" --delay <0...N> (seconds) (default: %d)\n", cmd_params_defaults.delay); - printf(" -o, --output (default: %s)\n", output_format_str(cmd_params_defaults.output_format)); - printf(" -oe, --output-err (default: %s)\n", output_format_str(cmd_params_defaults.output_format_stderr)); + printf(" -o, --output (default: %s)\n", + output_format_str(cmd_params_defaults.output_format)); + printf(" -oe, --output-err (default: %s)\n", + output_format_str(cmd_params_defaults.output_format_stderr)); printf(" -v, --verbose (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0"); printf(" --progress (default: %s)\n", cmd_params_defaults.progress ? "1" : "0"); printf("\n"); - printf("Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.\n"); + printf( + "Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter " + "multiple times.\n"); } static ggml_type ggml_type_from_name(const std::string & s) { @@ -281,22 +309,21 @@ static ggml_type ggml_type_from_name(const std::string & s) { return GGML_TYPE_COUNT; } - static cmd_params parse_cmd_params(int argc, char ** argv) { - cmd_params params; - std::string arg; - bool invalid_param = false; - const std::string arg_prefix = "--"; - const char split_delim = ','; + cmd_params params; + std::string arg; + bool invalid_param = false; + const std::string arg_prefix = "--"; + const char split_delim = ','; - params.verbose = cmd_params_defaults.verbose; - params.output_format = cmd_params_defaults.output_format; + params.verbose = cmd_params_defaults.verbose; + params.output_format = cmd_params_defaults.output_format; params.output_format_stderr = cmd_params_defaults.output_format_stderr; - params.reps = cmd_params_defaults.reps; - params.numa = cmd_params_defaults.numa; - params.prio = cmd_params_defaults.prio; - params.delay = cmd_params_defaults.delay; - params.progress = cmd_params_defaults.progress; + params.reps = cmd_params_defaults.reps; + params.numa = cmd_params_defaults.numa; + params.prio = cmd_params_defaults.prio; + params.delay = cmd_params_defaults.delay; + params.progress = cmd_params_defaults.progress; for (int i = 1; i < argc; i++) { arg = argv[i]; @@ -338,7 +365,7 @@ static cmd_params parse_cmd_params(int argc, char ** argv) { invalid_param = true; break; } - params.n_pg.push_back({std::stoi(p[0]), std::stoi(p[1])}); + params.n_pg.push_back({ std::stoi(p[0]), std::stoi(p[1]) }); } else if (arg == "-b" || arg == "--batch-size") { if (++i >= argc) { invalid_param = true; @@ -358,7 +385,7 @@ static cmd_params parse_cmd_params(int argc, char ** argv) { invalid_param = true; break; } - auto p = string_split(argv[i], split_delim); + auto p = string_split(argv[i], split_delim); std::vector types; for (const auto & t : p) { ggml_type gt = ggml_type_from_name(t); @@ -377,7 +404,7 @@ static cmd_params parse_cmd_params(int argc, char ** argv) { invalid_param = true; break; } - auto p = string_split(argv[i], split_delim); + auto p = string_split(argv[i], split_delim); std::vector types; for (const auto & t : p) { ggml_type gt = ggml_type_from_name(t); @@ -437,7 +464,7 @@ static cmd_params parse_cmd_params(int argc, char ** argv) { invalid_param = true; break; } - auto p = string_split(argv[i], split_delim); + auto p = string_split(argv[i], split_delim); std::vector modes; for (const auto & m : p) { llama_split_mode mode; @@ -476,10 +503,16 @@ static cmd_params parse_cmd_params(int argc, char ** argv) { break; } else { std::string value(argv[i]); - /**/ if (value == "distribute" || value == "" ) { params.numa = GGML_NUMA_STRATEGY_DISTRIBUTE; } - else if (value == "isolate") { params.numa = GGML_NUMA_STRATEGY_ISOLATE; } - else if (value == "numactl") { params.numa = GGML_NUMA_STRATEGY_NUMACTL; } - else { invalid_param = true; break; } + /**/ if (value == "distribute" || value == "") { + params.numa = GGML_NUMA_STRATEGY_DISTRIBUTE; + } else if (value == "isolate") { + params.numa = GGML_NUMA_STRATEGY_ISOLATE; + } else if (value == "numactl") { + params.numa = GGML_NUMA_STRATEGY_NUMACTL; + } else { + invalid_param = true; + break; + } } } else if (arg == "-fa" || arg == "--flash-attn") { if (++i >= argc) { @@ -509,9 +542,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) { } for (auto ts : string_split(argv[i], split_delim)) { // split string by ; and / - const std::regex regex{R"([;/]+)"}; - std::sregex_token_iterator it{ts.begin(), ts.end(), regex, -1}; - std::vector split_arg{it, {}}; + const std::regex regex{ R"([;/]+)" }; + std::sregex_token_iterator it{ ts.begin(), ts.end(), regex, -1 }; + std::vector split_arg{ it, {} }; GGML_ASSERT(split_arg.size() <= llama_max_devices()); std::vector tensor_split(llama_max_devices()); @@ -570,52 +603,94 @@ static cmd_params parse_cmd_params(int argc, char ** argv) { } // set defaults - if (params.model.empty()) { params.model = cmd_params_defaults.model; } - if (params.n_prompt.empty()) { params.n_prompt = cmd_params_defaults.n_prompt; } - if (params.n_gen.empty()) { params.n_gen = cmd_params_defaults.n_gen; } - if (params.n_pg.empty()) { params.n_pg = cmd_params_defaults.n_pg; } - if (params.n_batch.empty()) { params.n_batch = cmd_params_defaults.n_batch; } - if (params.n_ubatch.empty()) { params.n_ubatch = cmd_params_defaults.n_ubatch; } - if (params.type_k.empty()) { params.type_k = cmd_params_defaults.type_k; } - if (params.type_v.empty()) { params.type_v = cmd_params_defaults.type_v; } - if (params.n_gpu_layers.empty()) { params.n_gpu_layers = cmd_params_defaults.n_gpu_layers; } - if (params.rpc_servers.empty()) { params.rpc_servers = cmd_params_defaults.rpc_servers; } - if (params.split_mode.empty()) { params.split_mode = cmd_params_defaults.split_mode; } - if (params.main_gpu.empty()) { params.main_gpu = cmd_params_defaults.main_gpu; } - if (params.no_kv_offload.empty()){ params.no_kv_offload = cmd_params_defaults.no_kv_offload; } - if (params.flash_attn.empty()) { params.flash_attn = cmd_params_defaults.flash_attn; } - if (params.tensor_split.empty()) { params.tensor_split = cmd_params_defaults.tensor_split; } - if (params.use_mmap.empty()) { params.use_mmap = cmd_params_defaults.use_mmap; } - if (params.embeddings.empty()) { params.embeddings = cmd_params_defaults.embeddings; } - if (params.n_threads.empty()) { params.n_threads = cmd_params_defaults.n_threads; } - if (params.cpu_mask.empty()) { params.cpu_mask = cmd_params_defaults.cpu_mask; } - if (params.cpu_strict.empty()) { params.cpu_strict = cmd_params_defaults.cpu_strict; } - if (params.poll.empty()) { params.poll = cmd_params_defaults.poll; } + if (params.model.empty()) { + params.model = cmd_params_defaults.model; + } + if (params.n_prompt.empty()) { + params.n_prompt = cmd_params_defaults.n_prompt; + } + if (params.n_gen.empty()) { + params.n_gen = cmd_params_defaults.n_gen; + } + if (params.n_pg.empty()) { + params.n_pg = cmd_params_defaults.n_pg; + } + if (params.n_batch.empty()) { + params.n_batch = cmd_params_defaults.n_batch; + } + if (params.n_ubatch.empty()) { + params.n_ubatch = cmd_params_defaults.n_ubatch; + } + if (params.type_k.empty()) { + params.type_k = cmd_params_defaults.type_k; + } + if (params.type_v.empty()) { + params.type_v = cmd_params_defaults.type_v; + } + if (params.n_gpu_layers.empty()) { + params.n_gpu_layers = cmd_params_defaults.n_gpu_layers; + } + if (params.rpc_servers.empty()) { + params.rpc_servers = cmd_params_defaults.rpc_servers; + } + if (params.split_mode.empty()) { + params.split_mode = cmd_params_defaults.split_mode; + } + if (params.main_gpu.empty()) { + params.main_gpu = cmd_params_defaults.main_gpu; + } + if (params.no_kv_offload.empty()) { + params.no_kv_offload = cmd_params_defaults.no_kv_offload; + } + if (params.flash_attn.empty()) { + params.flash_attn = cmd_params_defaults.flash_attn; + } + if (params.tensor_split.empty()) { + params.tensor_split = cmd_params_defaults.tensor_split; + } + if (params.use_mmap.empty()) { + params.use_mmap = cmd_params_defaults.use_mmap; + } + if (params.embeddings.empty()) { + params.embeddings = cmd_params_defaults.embeddings; + } + if (params.n_threads.empty()) { + params.n_threads = cmd_params_defaults.n_threads; + } + if (params.cpu_mask.empty()) { + params.cpu_mask = cmd_params_defaults.cpu_mask; + } + if (params.cpu_strict.empty()) { + params.cpu_strict = cmd_params_defaults.cpu_strict; + } + if (params.poll.empty()) { + params.poll = cmd_params_defaults.poll; + } return params; } struct cmd_params_instance { - std::string model; - int n_prompt; - int n_gen; - int n_batch; - int n_ubatch; - ggml_type type_k; - ggml_type type_v; - int n_threads; - std::string cpu_mask; - bool cpu_strict; - int poll; - int n_gpu_layers; - std::string rpc_servers; - llama_split_mode split_mode; - int main_gpu; - bool no_kv_offload; - bool flash_attn; + std::string model; + int n_prompt; + int n_gen; + int n_batch; + int n_ubatch; + ggml_type type_k; + ggml_type type_v; + int n_threads; + std::string cpu_mask; + bool cpu_strict; + int poll; + int n_gpu_layers; + std::string rpc_servers; + llama_split_mode split_mode; + int main_gpu; + bool no_kv_offload; + bool flash_attn; std::vector tensor_split; - bool use_mmap; - bool embeddings; + bool use_mmap; + bool embeddings; llama_model_params to_llama_mparams() const { llama_model_params mparams = llama_model_default_params(); @@ -624,35 +699,31 @@ struct cmd_params_instance { if (!rpc_servers.empty()) { mparams.rpc_servers = rpc_servers.c_str(); } - mparams.split_mode = split_mode; - mparams.main_gpu = main_gpu; + mparams.split_mode = split_mode; + mparams.main_gpu = main_gpu; mparams.tensor_split = tensor_split.data(); - mparams.use_mmap = use_mmap; + mparams.use_mmap = use_mmap; return mparams; } bool equal_mparams(const cmd_params_instance & other) const { - return model == other.model && - n_gpu_layers == other.n_gpu_layers && - rpc_servers == other.rpc_servers && - split_mode == other.split_mode && - main_gpu == other.main_gpu && - use_mmap == other.use_mmap && + return model == other.model && n_gpu_layers == other.n_gpu_layers && rpc_servers == other.rpc_servers && + split_mode == other.split_mode && main_gpu == other.main_gpu && use_mmap == other.use_mmap && tensor_split == other.tensor_split; } llama_context_params to_llama_cparams() const { llama_context_params cparams = llama_context_default_params(); - cparams.n_ctx = n_prompt + n_gen; - cparams.n_batch = n_batch; - cparams.n_ubatch = n_ubatch; - cparams.type_k = type_k; - cparams.type_v = type_v; + cparams.n_ctx = n_prompt + n_gen; + cparams.n_batch = n_batch; + cparams.n_ubatch = n_ubatch; + cparams.type_k = type_k; + cparams.type_v = type_v; cparams.offload_kqv = !no_kv_offload; - cparams.flash_attn = flash_attn; - cparams.embeddings = embeddings; + cparams.flash_attn = flash_attn; + cparams.embeddings = embeddings; return cparams; } @@ -662,6 +733,7 @@ static std::vector get_cmd_params_instances(const cmd_param std::vector instances; // this ordering minimizes the number of times that each model needs to be reloaded + // clang-format off for (const auto & m : params.model) for (const auto & nl : params.n_gpu_layers) for (const auto & rpc : params.rpc_servers) @@ -767,100 +839,94 @@ static std::vector get_cmd_params_instances(const cmd_param instances.push_back(instance); } } + // clang-format on return instances; } struct test { static const std::string build_commit; - static const int build_number; + static const int build_number; static const std::string cpu_info; static const std::string gpu_info; - std::string model_filename; - std::string model_type; - uint64_t model_size; - uint64_t model_n_params; - int n_batch; - int n_ubatch; - int n_threads; - std::string cpu_mask; - bool cpu_strict; - int poll; - ggml_type type_k; - ggml_type type_v; - int n_gpu_layers; - llama_split_mode split_mode; - int main_gpu; - bool no_kv_offload; - bool flash_attn; - std::vector tensor_split; - bool use_mmap; - bool embeddings; - int n_prompt; - int n_gen; - std::string test_time; - std::vector samples_ns; + std::string model_filename; + std::string model_type; + uint64_t model_size; + uint64_t model_n_params; + int n_batch; + int n_ubatch; + int n_threads; + std::string cpu_mask; + bool cpu_strict; + int poll; + ggml_type type_k; + ggml_type type_v; + int n_gpu_layers; + llama_split_mode split_mode; + int main_gpu; + bool no_kv_offload; + bool flash_attn; + std::vector tensor_split; + bool use_mmap; + bool embeddings; + int n_prompt; + int n_gen; + std::string test_time; + std::vector samples_ns; test(const cmd_params_instance & inst, const llama_model * lmodel, const llama_context * ctx) { model_filename = inst.model; char buf[128]; llama_model_desc(lmodel, buf, sizeof(buf)); - model_type = buf; - model_size = llama_model_size(lmodel); + model_type = buf; + model_size = llama_model_size(lmodel); model_n_params = llama_model_n_params(lmodel); - n_batch = inst.n_batch; - n_ubatch = inst.n_ubatch; - n_threads = inst.n_threads; - cpu_mask = inst.cpu_mask; - cpu_strict = inst.cpu_strict; - poll = inst.poll; - type_k = inst.type_k; - type_v = inst.type_v; - n_gpu_layers = inst.n_gpu_layers; - split_mode = inst.split_mode; - main_gpu = inst.main_gpu; - no_kv_offload = inst.no_kv_offload; - flash_attn = inst.flash_attn; - tensor_split = inst.tensor_split; - use_mmap = inst.use_mmap; - embeddings = inst.embeddings; - n_prompt = inst.n_prompt; - n_gen = inst.n_gen; + n_batch = inst.n_batch; + n_ubatch = inst.n_ubatch; + n_threads = inst.n_threads; + cpu_mask = inst.cpu_mask; + cpu_strict = inst.cpu_strict; + poll = inst.poll; + type_k = inst.type_k; + type_v = inst.type_v; + n_gpu_layers = inst.n_gpu_layers; + split_mode = inst.split_mode; + main_gpu = inst.main_gpu; + no_kv_offload = inst.no_kv_offload; + flash_attn = inst.flash_attn; + tensor_split = inst.tensor_split; + use_mmap = inst.use_mmap; + embeddings = inst.embeddings; + n_prompt = inst.n_prompt; + n_gen = inst.n_gen; // RFC 3339 date-time format - time_t t = time(NULL); + time_t t = time(NULL); std::strftime(buf, sizeof(buf), "%FT%TZ", gmtime(&t)); test_time = buf; (void) ctx; } - uint64_t avg_ns() const { - return ::avg(samples_ns); - } + uint64_t avg_ns() const { return ::avg(samples_ns); } - uint64_t stdev_ns() const { - return ::stdev(samples_ns); - } + uint64_t stdev_ns() const { return ::stdev(samples_ns); } std::vector get_ts() const { - int n_tokens = n_prompt + n_gen; + int n_tokens = n_prompt + n_gen; std::vector ts; - std::transform(samples_ns.begin(), samples_ns.end(), std::back_inserter(ts), [n_tokens](uint64_t t) { return 1e9 * n_tokens / t; }); + std::transform(samples_ns.begin(), samples_ns.end(), std::back_inserter(ts), + [n_tokens](uint64_t t) { return 1e9 * n_tokens / t; }); return ts; } - double avg_ts() const { - return ::avg(get_ts()); - } + double avg_ts() const { return ::avg(get_ts()); } - double stdev_ts() const { - return ::stdev(get_ts()); - } + double stdev_ts() const { return ::stdev(get_ts()); } static std::string get_backend() { std::vector backends; for (size_t i = 0; i < ggml_backend_reg_count(); i++) { - auto * reg = ggml_backend_reg_get(i); + auto * reg = ggml_backend_reg_get(i); std::string name = ggml_backend_reg_name(reg); if (name != "CPU") { backends.push_back(ggml_backend_reg_name(reg)); @@ -871,36 +937,27 @@ struct test { static const std::vector & get_fields() { static const std::vector fields = { - "build_commit", "build_number", - "cpu_info", "gpu_info", "backends", - "model_filename", "model_type", "model_size", "model_n_params", - "n_batch", "n_ubatch", - "n_threads", "cpu_mask", "cpu_strict", "poll", - "type_k", "type_v", - "n_gpu_layers", "split_mode", - "main_gpu", "no_kv_offload", "flash_attn", - "tensor_split", "use_mmap", "embeddings", - "n_prompt", "n_gen", "test_time", - "avg_ns", "stddev_ns", - "avg_ts", "stddev_ts", + "build_commit", "build_number", "cpu_info", "gpu_info", "backends", "model_filename", + "model_type", "model_size", "model_n_params", "n_batch", "n_ubatch", "n_threads", + "cpu_mask", "cpu_strict", "poll", "type_k", "type_v", "n_gpu_layers", + "split_mode", "main_gpu", "no_kv_offload", "flash_attn", "tensor_split", "use_mmap", + "embeddings", "n_prompt", "n_gen", "test_time", "avg_ns", "stddev_ns", + "avg_ts", "stddev_ts", }; return fields; } - enum field_type {STRING, BOOL, INT, FLOAT}; + enum field_type { STRING, BOOL, INT, FLOAT }; static field_type get_field_type(const std::string & field) { - if (field == "build_number" || field == "n_batch" || field == "n_ubatch" || - field == "n_threads" || field == "poll" || - field == "model_size" || field == "model_n_params" || - field == "n_gpu_layers" || field == "main_gpu" || - field == "n_prompt" || field == "n_gen" || - field == "avg_ns" || field == "stddev_ns") { + if (field == "build_number" || field == "n_batch" || field == "n_ubatch" || field == "n_threads" || + field == "poll" || field == "model_size" || field == "model_n_params" || field == "n_gpu_layers" || + field == "main_gpu" || field == "n_prompt" || field == "n_gen" || field == "avg_ns" || + field == "stddev_ns") { return INT; } - if (field == "f16_kv" || field == "no_kv_offload" || - field == "cpu_strict" || - field == "flash_attn" || field == "use_mmap" || field == "embeddings") { + if (field == "f16_kv" || field == "no_kv_offload" || field == "cpu_strict" || field == "flash_attn" || + field == "use_mmap" || field == "embeddings") { return BOOL; } if (field == "avg_ts" || field == "stddev_ts") { @@ -911,7 +968,7 @@ struct test { std::vector get_values() const { std::string tensor_split_str; - int max_nonzero = 0; + int max_nonzero = 0; for (size_t i = 0; i < llama_max_devices(); i++) { if (tensor_split[i] > 0) { max_nonzero = i; @@ -925,29 +982,47 @@ struct test { tensor_split_str += "/"; } } - std::vector values = { - build_commit, std::to_string(build_number), - cpu_info, gpu_info, get_backend(), - model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params), - std::to_string(n_batch), std::to_string(n_ubatch), - std::to_string(n_threads), cpu_mask, std::to_string(cpu_strict), std::to_string(poll), - ggml_type_name(type_k), ggml_type_name(type_v), - std::to_string(n_gpu_layers), split_mode_str(split_mode), - std::to_string(main_gpu), std::to_string(no_kv_offload), std::to_string(flash_attn), - tensor_split_str, std::to_string(use_mmap), std::to_string(embeddings), - std::to_string(n_prompt), std::to_string(n_gen), test_time, - std::to_string(avg_ns()), std::to_string(stdev_ns()), - std::to_string(avg_ts()), std::to_string(stdev_ts()) - }; + std::vector values = { build_commit, + std::to_string(build_number), + cpu_info, + gpu_info, + get_backend(), + model_filename, + model_type, + std::to_string(model_size), + std::to_string(model_n_params), + std::to_string(n_batch), + std::to_string(n_ubatch), + std::to_string(n_threads), + cpu_mask, + std::to_string(cpu_strict), + std::to_string(poll), + ggml_type_name(type_k), + ggml_type_name(type_v), + std::to_string(n_gpu_layers), + split_mode_str(split_mode), + std::to_string(main_gpu), + std::to_string(no_kv_offload), + std::to_string(flash_attn), + tensor_split_str, + std::to_string(use_mmap), + std::to_string(embeddings), + std::to_string(n_prompt), + std::to_string(n_gen), + test_time, + std::to_string(avg_ns()), + std::to_string(stdev_ns()), + std::to_string(avg_ts()), + std::to_string(stdev_ts()) }; return values; } std::map get_map() const { std::map map; - auto fields = get_fields(); - auto values = get_values(); - std::transform(fields.begin(), fields.end(), values.begin(), - std::inserter(map, map.end()), std::make_pair); + auto fields = get_fields(); + auto values = get_values(); + std::transform(fields.begin(), fields.end(), values.begin(), std::inserter(map, map.end()), + std::make_pair); return map; } }; @@ -961,9 +1036,12 @@ struct printer { virtual ~printer() {} FILE * fout; + virtual void print_header(const cmd_params & params) { (void) params; } + virtual void print_test(const test & t) = 0; - virtual void print_footer() { } + + virtual void print_footer() {} }; struct csv_printer : public printer { @@ -979,7 +1057,7 @@ struct csv_printer : public printer { return escaped; } - void print_header(const cmd_params & params) override { + void print_header(const cmd_params & params) override { std::vector fields = test::get_fields(); fprintf(fout, "%s\n", join(fields, ",").c_str()); (void) params; @@ -992,7 +1070,6 @@ struct csv_printer : public printer { } }; - static std::string escape_json(const std::string & value) { std::string escaped; for (auto c : value) { @@ -1000,7 +1077,7 @@ static std::string escape_json(const std::string & value) { escaped += "\\\""; } else if (c == '\\') { escaped += "\\\\"; - } else if (c <= 0x1f) { + } else if (c <= 0x1f) { char buf[8]; snprintf(buf, sizeof(buf), "\\u%04x", c); escaped += buf; @@ -1033,7 +1110,8 @@ struct json_printer : public printer { void print_fields(const std::vector & fields, const std::vector & values) { assert(fields.size() == values.size()); for (size_t i = 0; i < fields.size(); i++) { - fprintf(fout, " \"%s\": %s,\n", fields.at(i).c_str(), format_json_value(fields.at(i), values.at(i)).c_str()); + fprintf(fout, " \"%s\": %s,\n", fields.at(i).c_str(), + format_json_value(fields.at(i), values.at(i)).c_str()); } } @@ -1051,12 +1129,9 @@ struct json_printer : public printer { fflush(fout); } - void print_footer() override { - fprintf(fout, "\n]\n"); - } + void print_footer() override { fprintf(fout, "\n]\n"); } }; - struct jsonl_printer : public printer { void print_fields(const std::vector & fields, const std::vector & values) { assert(fields.size() == values.size()); @@ -1116,7 +1191,7 @@ struct markdown_printer : public printer { return 13; } - int width = std::max((int)field.length(), 10); + int width = std::max((int) field.length(), 10); if (test::get_field_type(field) == test::STRING) { return -width; @@ -1230,18 +1305,18 @@ struct markdown_printer : public printer { fprintf(fout, "|"); for (const auto & field : fields) { std::string value; - char buf[128]; + char buf[128]; if (field == "model") { value = t.model_type; } else if (field == "size") { - if (t.model_size < 1024*1024*1024) { + if (t.model_size < 1024 * 1024 * 1024) { snprintf(buf, sizeof(buf), "%.2f MiB", t.model_size / 1024.0 / 1024.0); } else { snprintf(buf, sizeof(buf), "%.2f GiB", t.model_size / 1024.0 / 1024.0 / 1024.0); } value = buf; } else if (field == "params") { - if (t.model_n_params < 1000*1000*1000) { + if (t.model_n_params < 1000 * 1000 * 1000) { snprintf(buf, sizeof(buf), "%.2f M", t.model_n_params / 1e6); } else { snprintf(buf, sizeof(buf), "%.2f B", t.model_n_params / 1e9); @@ -1303,7 +1378,8 @@ struct sql_printer : public printer { std::vector fields = test::get_fields(); fprintf(fout, "CREATE TABLE IF NOT EXISTS test (\n"); for (size_t i = 0; i < fields.size(); i++) { - fprintf(fout, " %s %s%s\n", fields.at(i).c_str(), get_sql_field_type(fields.at(i)).c_str(), i < fields.size() - 1 ? "," : ""); + fprintf(fout, " %s %s%s\n", fields.at(i).c_str(), get_sql_field_type(fields.at(i)).c_str(), + i < fields.size() - 1 ? "," : ""); } fprintf(fout, ");\n"); fprintf(fout, "\n"); @@ -1324,8 +1400,8 @@ struct sql_printer : public printer { static void test_prompt(llama_context * ctx, int n_prompt, int n_batch, int n_threads) { llama_set_n_threads(ctx, n_threads, n_threads); - const llama_model * model = llama_get_model(ctx); - const int32_t n_vocab = llama_n_vocab(model); + const llama_model * model = llama_get_model(ctx); + const int32_t n_vocab = llama_n_vocab(model); std::vector tokens(n_batch); @@ -1333,7 +1409,7 @@ static void test_prompt(llama_context * ctx, int n_prompt, int n_batch, int n_th while (n_processed < n_prompt) { int n_tokens = std::min(n_prompt - n_processed, n_batch); - tokens[0] = n_processed == 0 && llama_add_bos_token(model) ? llama_token_bos(model) : std::rand() % n_vocab; + tokens[0] = n_processed == 0 && llama_add_bos_token(model) ? llama_token_bos(model) : std::rand() % n_vocab; for (int i = 1; i < n_tokens; i++) { tokens[i] = std::rand() % n_vocab; } @@ -1347,8 +1423,8 @@ static void test_prompt(llama_context * ctx, int n_prompt, int n_batch, int n_th static void test_gen(llama_context * ctx, int n_gen, int n_threads) { llama_set_n_threads(ctx, n_threads, n_threads); - const llama_model * model = llama_get_model(ctx); - const int32_t n_vocab = llama_n_vocab(model); + const llama_model * model = llama_get_model(ctx); + const int32_t n_vocab = llama_n_vocab(model); llama_token token = llama_add_bos_token(model) ? llama_token_bos(model) : std::rand() % n_vocab; @@ -1411,7 +1487,7 @@ int main(int argc, char ** argv) { set_process_priority(params.prio); // initialize printer - std::unique_ptr p = create_printer(params.output_format); + std::unique_ptr p = create_printer(params.output_format); std::unique_ptr p_err = create_printer(params.output_format_stderr); if (p) { @@ -1426,13 +1502,13 @@ int main(int argc, char ** argv) { std::vector params_instances = get_cmd_params_instances(params); - llama_model * lmodel = nullptr; + llama_model * lmodel = nullptr; const cmd_params_instance * prev_inst = nullptr; - int params_idx = 0; + int params_idx = 0; auto params_count = params_instances.size(); for (const auto & inst : params_instances) { - params_idx ++; + params_idx++; if (params.progress) { fprintf(stderr, "llama-bench: benchmark %d/%ld: starting\n", params_idx, params_count); } @@ -1475,7 +1551,7 @@ int main(int argc, char ** argv) { tpp.poll = t.poll; tpp.prio = params.prio; - struct ggml_threadpool* threadpool = ggml_threadpool_new(&tpp); + struct ggml_threadpool * threadpool = ggml_threadpool_new(&tpp); if (!threadpool) { fprintf(stderr, "%s: threadpool create failed : n_threads %d\n", __func__, tpp.n_threads); exit(1); @@ -1505,13 +1581,15 @@ int main(int argc, char ** argv) { if (t.n_prompt > 0) { if (params.progress) { - fprintf(stderr, "llama-bench: benchmark %d/%ld: prompt run %d/%d\n", params_idx, params_count, i + 1, params.reps); + fprintf(stderr, "llama-bench: benchmark %d/%ld: prompt run %d/%d\n", params_idx, params_count, + i + 1, params.reps); } test_prompt(ctx, t.n_prompt, t.n_batch, t.n_threads); } if (t.n_gen > 0) { if (params.progress) { - fprintf(stderr, "llama-bench: benchmark %d/%ld: generation run %d/%d\n", params_idx, params_count, i + 1, params.reps); + fprintf(stderr, "llama-bench: benchmark %d/%ld: generation run %d/%d\n", params_idx, params_count, + i + 1, params.reps); } test_gen(ctx, t.n_gen, t.n_threads); }