diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index ff53bdfbe171c..7cb9fd99c46ae 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -5,8 +5,14 @@ #include "ggml-cpu.h" #endif +// See https://github.com/KhronosGroup/Vulkan-Hpp?tab=readme-ov-file#extensions--per-device-function-pointers +#define VULKAN_HPP_DISPATCH_LOADER_DYNAMIC 1 + #include +// See https://github.com/KhronosGroup/Vulkan-Hpp?tab=readme-ov-file#extensions--per-device-function-pointers +VULKAN_HPP_DEFAULT_DISPATCH_LOADER_DYNAMIC_STORAGE + #include #include #include @@ -29,6 +35,7 @@ #include "ggml-vulkan-shaders.hpp" +#define ROUNDUP_POW2(M, N) (((M) + (N) - 1) & ~((N) - 1)) #define CEIL_DIV(M, N) (((M) + (N)-1) / (N)) #define VK_VENDOR_ID_AMD 0x1002 @@ -89,6 +96,9 @@ struct vk_pipeline_struct { bool needed {}; // set to true when the shader has been compiled bool compiled {}; + // number of registers used, extracted from pipeline executable properties + uint32_t register_count {}; + std::vector specialization_constants; }; typedef std::shared_ptr vk_pipeline; @@ -183,6 +193,8 @@ struct vk_device_struct { uint32_t coopmat_k; bool coopmat2; + bool pipeline_executable_properties_support {}; + size_t idx; bool mul_mat_l[GGML_TYPE_COUNT]; @@ -368,6 +380,7 @@ struct vk_mat_mat_push_constants { uint32_t batch_stride_a; uint32_t batch_stride_b; uint32_t batch_stride_d; uint32_t k_split; uint32_t ne02; uint32_t ne12; uint32_t broadcast2; uint32_t broadcast3; + uint32_t padded_N; }; struct vk_mat_vec_push_constants { uint32_t ncols; uint32_t stride_a; uint32_t stride_b; uint32_t stride_d; @@ -380,6 +393,7 @@ struct vk_mat_mat_id_push_constants { uint32_t stride_a; uint32_t stride_b; uint32_t stride_d; uint32_t batch_stride_a; uint32_t batch_stride_b; uint32_t batch_stride_d; uint32_t nei0; uint32_t nei1; uint32_t nbi1; uint32_t ne11; + uint32_t padded_N; }; struct vk_mat_vec_id_push_constants { uint32_t ncols; uint32_t stride_a; uint32_t stride_b; uint32_t stride_d; @@ -890,6 +904,20 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin } pipeline->compiled = true; + if (device->pipeline_executable_properties_support) { + vk::PipelineExecutableInfoKHR executableInfo; + executableInfo.pipeline = pipeline->pipeline; + + auto statistics = device->device.getPipelineExecutableStatisticsKHR(executableInfo); + for (auto & s : statistics) { + VK_LOG_DEBUG(pipeline->name << " " << s.name << ": " << s.value.u64); + // "Register Count" is reported by NVIDIA drivers. + if (strcmp(s.name, "Register Count") == 0) { + pipeline->register_count = (uint32_t)s.value.u64; + } + } + } + { std::lock_guard guard(device->mutex); device->pipelines.insert({ pipeline->name, pipeline }); @@ -1466,36 +1494,36 @@ static void ggml_vk_load_shaders(vk_device& device) { uint32_t l_align, m_align, s_align; if (device->coopmat2) { // spec constants and tile sizes for non-quant matmul/matmul_id - l_warptile = { 256, 128, 256, 64 }; - m_warptile = { 256, 128, 128, 64 }; - s_warptile = { 128, 64, 64, 64 }; + l_warptile = { 256, 128, 256, 64, 1 }; + m_warptile = { 256, 128, 128, 64, 0 }; + s_warptile = { 128, 64, 64, 64, 0 }; l_wg_denoms = {128, 256, 1 }; m_wg_denoms = {128, 128, 1 }; s_wg_denoms = { 64, 64, 1 }; // spec constants and tile sizes for quant matmul (non-Qi_K) - l_warptile_mmq = { 256, 128, 256, 64 }; - m_warptile_mmq = { 256, 128, 128, 64 }; - s_warptile_mmq = { 256, 128, 128, 64 }; + l_warptile_mmq = { 256, 128, 256, 64, 1 }; + m_warptile_mmq = { 256, 128, 128, 64, 1 }; + s_warptile_mmq = { 256, 32, 64, 128, 0 }; l_mmq_wg_denoms = { 128, 256, 1 }; m_mmq_wg_denoms = { 128, 128, 1 }; - s_mmq_wg_denoms = { 128, 128, 1 }; + s_mmq_wg_denoms = { 32, 64, 1 }; // spec constants and tile sizes for quant matmul (Qi_K) - l_warptile_mmq_k = { 256, 128, 512, 16 }; - m_warptile_mmq_k = { 256, 128, 256, 16 }; - s_warptile_mmq_k = { 256, 32, 128, 64 }; - l_mmq_wg_denoms_k = { 128, 512, 1 }; - m_mmq_wg_denoms_k = { 128, 256, 1 }; - s_mmq_wg_denoms_k = { 32, 128, 1 }; + l_warptile_mmq_k = { 256, 64, 128, 64, 1 }; + m_warptile_mmq_k = { 256, 32, 64, 64, 0 }; + s_warptile_mmq_k = { 256, 32, 32, 128, 0 }; + l_mmq_wg_denoms_k = { 64, 128, 1 }; + m_mmq_wg_denoms_k = { 32, 64, 1 }; + s_mmq_wg_denoms_k = { 32, 32, 1 }; // spec constants and tile sizes for quant matmul_id - l_warptile_mmqid = { 256, 128, 128, 16 }; - m_warptile_mmqid = { 256, 128, 64, 16 }; - s_warptile_mmqid = { 256, 64, 64, 16 }; - l_mmqid_wg_denoms = { 128, 128, 1 }; + l_warptile_mmqid = { 256, 128, 64, 16, 0 }; + m_warptile_mmqid = { 256, 128, 64, 16, 0 }; + s_warptile_mmqid = { 256, 128, 64, 16, 0 }; + l_mmqid_wg_denoms = { 128, 64, 1 }; m_mmqid_wg_denoms = { 128, 64, 1 }; - s_mmqid_wg_denoms = { 64, 64, 1 }; + s_mmqid_wg_denoms = { 128, 64, 1 }; l_align = 128; m_align = 64; @@ -1578,6 +1606,7 @@ static void ggml_vk_load_shaders(vk_device& device) { pipeline->push_constant_size = push_constant_size; pipeline->wg_denoms = wg_denoms; pipeline->align = align; + pipeline->specialization_constants = specialization_constants; } if (!pipeline->needed || pipeline->compiled) { @@ -2286,6 +2315,7 @@ static vk_device ggml_vk_get_device(size_t idx) { bool amd_shader_core_properties2 = false; bool pipeline_robustness = false; bool coopmat2_support = false; + bool pipeline_executable_properties_support = false; device->coopmat_support = false; // Check if maintenance4 is supported @@ -2313,6 +2343,8 @@ static vk_device ggml_vk_get_device(size_t idx) { } else if (strcmp("VK_NV_cooperative_matrix2", properties.extensionName) == 0 && !getenv("GGML_VK_DISABLE_COOPMAT2")) { coopmat2_support = true; + } else if (strcmp("VK_KHR_pipeline_executable_properties", properties.extensionName) == 0) { + pipeline_executable_properties_support = true; } } @@ -2497,8 +2529,18 @@ static vk_device ggml_vk_get_device(size_t idx) { device_extensions.push_back("VK_KHR_maintenance4"); } + VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR pep_features {}; + pep_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR; + if (pipeline_executable_properties_support) { + last_struct->pNext = (VkBaseOutStructure *)&pep_features; + last_struct = (VkBaseOutStructure *)&pep_features; + device_extensions.push_back("VK_KHR_pipeline_executable_properties"); + } + vkGetPhysicalDeviceFeatures2(device->physical_device, &device_features2); + device->pipeline_executable_properties_support = pipeline_executable_properties_support; + device->fp16 = device->fp16 && vk12_features.shaderFloat16; device->pipeline_robustness = pl_robustness_features.pipelineRobustness; @@ -2873,6 +2915,9 @@ static void ggml_vk_instance_init() { } VK_LOG_DEBUG("ggml_vk_instance_init()"); + // See https://github.com/KhronosGroup/Vulkan-Hpp?tab=readme-ov-file#extensions--per-device-function-pointers + VULKAN_HPP_DEFAULT_DISPATCHER.init(vkGetInstanceProcAddr); + uint32_t api_version = vk::enumerateInstanceVersion(); if (api_version < VK_API_VERSION_1_2) { @@ -2925,6 +2970,9 @@ static void ggml_vk_instance_init() { vk_instance.instance = vk::createInstance(instance_create_info); vk_instance_initialized = true; + // See https://github.com/KhronosGroup/Vulkan-Hpp?tab=readme-ov-file#extensions--per-device-function-pointers + VULKAN_HPP_DEFAULT_DISPATCHER.init(vk_instance.instance); + size_t num_available_devices = vk_instance.instance.enumeratePhysicalDevices().size(); // Emulate behavior of CUDA_VISIBLE_DEVICES for Vulkan @@ -3829,12 +3877,21 @@ static uint32_t ggml_vk_guess_split_k(ggml_backend_vk_context * ctx, int m, int VK_LOG_DEBUG("ggml_vk_guess_split_k(" << m << ", " << n << ", " << k << ")"); uint32_t split_k = 1; - if (ctx->device->shader_core_count != 0 && m >= (int)pipeline->wg_denoms[0] && n >= (int)pipeline->wg_denoms[1]) { - // If k is 'large' and the SMs will fill less than halfway, use split_k. + if (ctx->device->shader_core_count != 0) { uint32_t m_tiles = CEIL_DIV(m, pipeline->wg_denoms[0]); uint32_t n_tiles = CEIL_DIV(n, pipeline->wg_denoms[1]); - if (k >= 2048 && m_tiles * n_tiles < ctx->device->shader_core_count / 2) { - split_k = ctx->device->shader_core_count / (m_tiles * n_tiles); + uint32_t occupancy_factor = 1; + // Estimate how many workgroups can fit on an SM at a time. + // Other factors like shared memory could affect this, and aren't taken into account. + if (ctx->device->vendor_id == VK_VENDOR_ID_NVIDIA && pipeline->register_count > 0) { + uint32_t block_size = pipeline->specialization_constants[0]; + assert(block_size > 0); + occupancy_factor = 65536 / (block_size * pipeline->register_count); + } + // The extra factor of 4 is to try to run up to 4x as many workgroups as can fit, + // to prefer shorter shaders that will be less prone to tail effects + if (k >= 2048 && m_tiles * n_tiles < ctx->device->shader_core_count * occupancy_factor * 4) { + split_k = occupancy_factor * 4 * ctx->device->shader_core_count / (m_tiles * n_tiles); // Clamp to 2 or 4 split_k = std::min(split_k, 4u); if (split_k == 3) { @@ -3850,10 +3907,14 @@ static vk_pipeline ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx, VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline(" << m << ", " << n << ", " << aligned << ", " << ggml_type_name(src0_type) << ")"); if (ctx->device->coopmat2) { - if ((ctx->device->mul_mat_l[src0_type] && (m % mmp->l->wg_denoms[0]) == 0 && (n % mmp->l->wg_denoms[1]) == 0) || (!ctx->device->mul_mat_m[src0_type] && !ctx->device->mul_mat_s[src0_type])) { + // Use large shader when the N dimension is greater than the medium shader's tile size + uint32_t crossover_large = mmp->m->wg_denoms[1]; + if ((ctx->device->mul_mat_l[src0_type] && (n > crossover_large)) || (!ctx->device->mul_mat_m[src0_type] && !ctx->device->mul_mat_s[src0_type])) { return aligned ? mmp->a_l : mmp->l; } - if ((ctx->device->mul_mat_m[src0_type] && (m % mmp->m->wg_denoms[0]) == 0 && (n % mmp->m->wg_denoms[1]) == 0) || !ctx->device->mul_mat_s[src0_type]) { + // Use medium shader when the N dimension is greater than the small shader's tile size + uint32_t crossover_medium = mmp->s->wg_denoms[1]; + if ((ctx->device->mul_mat_m[src0_type] && (n > crossover_medium)) || !ctx->device->mul_mat_s[src0_type]) { return aligned ? mmp->a_m : mmp->m; } return aligned ? mmp->a_s : mmp->s; @@ -3878,18 +3939,19 @@ static void ggml_vk_matmul( vk_subbuffer&& a, vk_subbuffer&& b, vk_subbuffer&& d, vk_subbuffer&& split_k_buffer, uint32_t m, uint32_t n, uint32_t k, uint32_t stride_a, uint32_t stride_b, uint32_t stride_d, uint32_t batch_stride_a, uint32_t batch_stride_b, uint32_t batch_stride_d, - uint32_t split_k, uint32_t batch, uint32_t ne02, uint32_t ne12, uint32_t broadcast2, uint32_t broadcast3) { + uint32_t split_k, uint32_t batch, uint32_t ne02, uint32_t ne12, uint32_t broadcast2, uint32_t broadcast3, + uint32_t padded_n) { VK_LOG_DEBUG("ggml_vk_matmul(a: (" << a.buffer->buffer << ", " << a.offset << ", " << a.size << "), b: (" << b.buffer->buffer << ", " << b.offset << ", " << b.size << "), d: (" << d.buffer->buffer << ", " << d.offset << ", " << d.size << "), split_k: (" << (split_k_buffer.buffer != nullptr ? split_k_buffer.buffer->buffer : VK_NULL_HANDLE) << ", " << split_k_buffer.offset << ", " << split_k_buffer.size << "), m: " << m << ", n: " << n << ", k: " << k << ", stride_a: " << stride_a << ", stride_b: " << stride_b << ", stride_d: " << stride_d << ", batch_stride_a: " << batch_stride_a << ", batch_stride_b: " << batch_stride_b << ", batch_stride_d: " << batch_stride_d << ", split_k: " << split_k << ", batch: " << batch << ", ne02: " << ne02 << ", ne12: " << ne12 << ", broadcast2: " << broadcast2 << ", broadcast3: " << broadcast3 << ")"); ggml_vk_sync_buffers(subctx); if (split_k == 1) { - const vk_mat_mat_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, k, ne02, ne12, broadcast2, broadcast3 }; + const vk_mat_mat_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, k, ne02, ne12, broadcast2, broadcast3, padded_n }; ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d }, sizeof(vk_mat_mat_push_constants), &pc, { m, n, batch }); return; } GGML_ASSERT(batch_stride_d == m * n); - const vk_mat_mat_push_constants pc1 = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, CEIL_DIV(k, split_k), ne02, ne12, broadcast2, broadcast3 }; + const vk_mat_mat_push_constants pc1 = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, CEIL_DIV(k, split_k), ne02, ne12, broadcast2, broadcast3, padded_n }; // Make sure enough workgroups get assigned for split k to work ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, split_k_buffer }, sizeof(vk_mat_mat_push_constants), &pc1, { (CEIL_DIV(m, pipeline->wg_denoms[0]) * pipeline->wg_denoms[0]) * split_k, n, batch }); ggml_vk_sync_buffers(subctx); @@ -3898,13 +3960,17 @@ static void ggml_vk_matmul( } static vk_pipeline ggml_vk_guess_matmul_id_pipeline(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, bool aligned, ggml_type src0_type) { - VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline(" << m << ", " << n << ", " << aligned << ", " << ggml_type_name(src0_type) << ")"); + VK_LOG_DEBUG("ggml_vk_guess_matmul_id_pipeline(" << m << ", " << n << ", " << aligned << ", " << ggml_type_name(src0_type) << ")"); if (ctx->device->coopmat2) { - if ((ctx->device->mul_mat_id_l[src0_type] && (m % mmp->l->wg_denoms[0]) == 0 && (n % mmp->l->wg_denoms[1]) == 0) || (!ctx->device->mul_mat_id_m[src0_type] && !ctx->device->mul_mat_id_s[src0_type])) { + // Use large shader when the N dimension is greater than the medium shader's tile size + uint32_t crossover_large = mmp->m->wg_denoms[1]; + if ((ctx->device->mul_mat_id_l[src0_type] && (n > crossover_large)) || (!ctx->device->mul_mat_id_m[src0_type] && !ctx->device->mul_mat_id_s[src0_type])) { return aligned ? mmp->a_l : mmp->l; } - if ((ctx->device->mul_mat_id_m[src0_type] && (m % mmp->m->wg_denoms[0]) == 0 && (n % mmp->m->wg_denoms[1]) == 0) || !ctx->device->mul_mat_id_s[src0_type]) { + // Use medium shader when the N dimension is greater than the small shader's tile size + uint32_t crossover_medium = mmp->s->wg_denoms[1]; + if ((ctx->device->mul_mat_id_m[src0_type] && (n > crossover_medium)) || !ctx->device->mul_mat_id_s[src0_type]) { return aligned ? mmp->a_m : mmp->m; } return aligned ? mmp->a_s : mmp->s; @@ -3929,14 +3995,15 @@ static void ggml_vk_matmul_id( vk_subbuffer&& a, vk_subbuffer&& b, vk_subbuffer&& d, vk_subbuffer&& ids, uint32_t m, uint32_t n, uint32_t k, uint32_t stride_a, uint32_t stride_b, uint32_t stride_d, uint32_t batch_stride_a, uint32_t batch_stride_b, uint32_t batch_stride_d, - uint32_t n_as, uint32_t nei0, uint32_t nei1, uint32_t nbi1, uint32_t ne11) { + uint32_t n_as, uint32_t nei0, uint32_t nei1, uint32_t nbi1, uint32_t ne11, + uint32_t padded_n) { VK_LOG_DEBUG("ggml_vk_matmul_id(a: (" << a.buffer->buffer << ", " << a.offset << ", " << a.size << "), b: (" << b.buffer->buffer << ", " << b.offset << ", " << b.size << "), d: (" << d.buffer->buffer << ", " << d.offset << ", " << d.size << "), ids: (" << ids.buffer->buffer << ", " << ids.offset << ", " << ids.size << "), " << "m: " << m << ", n: " << n << ", k: " << k << ", stride_a: " << stride_a << ", stride_b: " << stride_b << ", stride_d: " << stride_d << ", " << "batch_stride_a: " << batch_stride_a << ", batch_stride_b: " << batch_stride_b << ", batch_stride_d: " << batch_stride_d << ", " << "n_as: " << n_as << ", nei0: " << nei0 << ", nei1: " << nei1 << ", nbi1: " << nbi1 << ", ne11: " << ne11 << ")"); ggml_vk_sync_buffers(subctx); const vk_mat_mat_id_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, - nei0, nei1, nbi1, ne11 }; + nei0, nei1, nbi1, ne11, padded_n }; ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d, ids }, sizeof(vk_mat_mat_id_push_constants), &pc, { m, nei1, n_as }); } @@ -4098,16 +4165,18 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub // Not implemented GGML_ASSERT(y_non_contig || !qy_needs_dequant); // NOLINT - const int x_ne = ne01 * ne00; - const int y_ne = ne11 * ne10; - const int d_ne = ne11 * ne01; - const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, ne11, qx_needs_dequant ? GGML_TYPE_F16 : src0->type)); const bool aligned = ne10 == kpad && ne01 > 8 && ne11 > 8; vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, ne11, aligned, qx_needs_dequant ? GGML_TYPE_F16 : src0->type); - const uint32_t split_k = ggml_vk_guess_split_k(ctx, ne01, ne11, ne10, pipeline); + // Reserve extra storage in the N dimension for the Y matrix, so we can avoid bounds-checking + uint32_t padded_n = qy_needs_dequant ? ROUNDUP_POW2(ne11, pipeline->wg_denoms[1]) :ne11; + const int x_ne = ne01 * ne00; + const int y_ne = padded_n * ne10; + const int d_ne = ne11 * ne01; + + uint32_t split_k = ggml_vk_guess_split_k(ctx, ne01, ne11, ne10, pipeline); const uint64_t qx_sz = ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type); const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type); @@ -4131,10 +4200,10 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr); // NOLINT GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr); // NOLINT + const uint64_t split_k_size = split_k > 1 ? d_sz * ne12 * ne13 * split_k : 0; if (dryrun) { const uint64_t x_sz_upd = x_sz * ne02 * ne03; const uint64_t y_sz_upd = y_sz * ne12 * ne13; - const uint64_t split_k_size = split_k > 1 ? d_sz * ne12 * ne13 * split_k : 0; if ( (qx_needs_dequant && x_sz_upd > ctx->device->max_memory_allocation_size) || (qy_needs_dequant && y_sz_upd > ctx->device->max_memory_allocation_size) || @@ -4159,11 +4228,19 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub if (qy_needs_dequant) { ggml_pipeline_request_descriptor_sets(ctx->device, to_fp16_vk_1, 1); } - if (split_k > 1) { + // ggml_vk_guess_split_k may make a different determination after the pipeline + // is compiled (based on register count), so prepare for split_k just in case. + if (split_k > 1 || !pipeline->compiled) { ggml_pipeline_request_descriptor_sets(ctx->device, ctx->device->pipeline_matmul_split_k_reduce, 1); } return; } + // ggml_vk_guess_split_k may make a different determination after the pipeline + // is compiled (based on register count). Fallback to no split_k if we didn't + // reserve enough memory. + if (split_k_size > ctx->prealloc_size_split_k) { + split_k = 1; + } vk_buffer d_D = dst_buf_ctx->dev_buffer; const uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs; @@ -4229,7 +4306,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub { d_D, d_buf_offset, d_sz * ne12 * ne13 }, { ctx->prealloc_split_k, 0, d_sz * ne12 * ne13 * split_k }, ne01, ne11, ne10, ne10, ne10, ne01, stride_batch_x, stride_batch_y, ne20*ne21, - split_k, ne12*ne13, ne02, ne12, r2, r3 + split_k, ne12*ne13, ne02, ne12, r2, r3, padded_n ); // NOLINT } @@ -4680,15 +4757,17 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& // Not implemented GGML_ASSERT(y_non_contig || !qy_needs_dequant); // NOLINT - const uint64_t x_ne = ne01 * ne00; - const uint64_t y_ne = ne11 * ne10; - const uint64_t d_ne = ne21 * ne20; - const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_id_pipeline_align(ctx, mmp, ne01, nei1, qx_needs_dequant ? GGML_TYPE_F16 : src0->type)); const bool aligned = ne10 == kpad && ne01 > 8 && nei1 > 8; vk_pipeline pipeline = ggml_vk_guess_matmul_id_pipeline(ctx, mmp, ne01, nei1, aligned, qx_needs_dequant ? GGML_TYPE_F16 : src0->type); + // Reserve extra storage in the N dimension for the Y matrix, so we can avoid bounds-checking + uint32_t padded_n = qy_needs_dequant ? ROUNDUP_POW2(ne11, pipeline->wg_denoms[1]) :ne11; + const uint64_t x_ne = ne01 * ne00; + const uint64_t y_ne = padded_n * ne10; + const uint64_t d_ne = ne21 * ne20; + const uint64_t qx_sz = ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type); const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type); const uint64_t x_sz = !qx_needs_dequant ? qx_sz : sizeof(ggml_fp16_t) * x_ne; @@ -4807,7 +4886,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& { d_D, d_buf_offset, d_sz * ne22 * ne23 }, { d_ids, ids_buf_offset, ids_sz }, ne01, ne21, ne10, ne10, ne10, ne01, stride_batch_x, stride_batch_y, ne20*ne21, - n_as, nei0, nei1, nbi1 / ggml_type_size(ids->type), ne11 + n_as, nei0, nei1, nbi1 / ggml_type_size(ids->type), ne11, padded_n ); // NOLINT } @@ -6767,7 +6846,7 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t ctx, subctx, p, ggml_vk_subbuffer(d_X), ggml_vk_subbuffer(d_Y), ggml_vk_subbuffer(d_D), ggml_vk_subbuffer(ctx->prealloc_split_k), m, n, k, k, k, m, k*m, k*n, m*n, - split_k, batch, batch, batch, 1, 1 + split_k, batch, batch, batch, 1, 1, n ); } ggml_vk_ctx_end(subctx); @@ -7112,7 +7191,7 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m, ctx, subctx, p, ggml_vk_subbuffer(qx_buf), ggml_vk_subbuffer(y_buf), ggml_vk_subbuffer(d_buf), ggml_vk_subbuffer(ctx->prealloc_split_k), m, n, k, k, k, m, k*m, k*n, m*n, - split_k, batch, batch, batch, 1, 1 + split_k, batch, batch, batch, 1, 1, n ); } ggml_vk_ctx_end(subctx); diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp index 4ccbe613af2ce..8efe4653ffe75 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp @@ -178,7 +178,7 @@ float16_t dequantFuncQ4_K(const in decodeBufQ4_K bl, const in uint blockCoords[2 uvec4 v = bl128.block.q4k[0]; - const f16vec2 loadd = unpackFloat2x16(v.x); + const vec2 loadd = vec2(unpackFloat2x16(v.x)); uint32_t sc; uint32_t mbyte; @@ -199,15 +199,15 @@ float16_t dequantFuncQ4_K(const in decodeBufQ4_K bl, const in uint blockCoords[2 sc &= 0x3F; mbyte &= 0x3F; - const float16_t d = loadd.x * float16_t(sc); - const float16_t m = loadd.y * float16_t(mbyte); + const float d = loadd.x * float(sc); + const float m = loadd.y * float(mbyte); uint qs = uint32_t(bl16.block.qs[((idx & 0xC0) >> 2) + ((idx & 0x1E) >> 1)]); qs = (qs >> (b * 4 + 8 * (idx & 1))) & 0xF; - float16_t ret = d * float16_t(qs) - m; + float ret = d * float(qs) - m; - return ret; + return float16_t(ret); } layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ5_K { diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp index 66dd2c860d82d..7649febb07193 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp @@ -23,6 +23,10 @@ layout (constant_id = 1) const uint BM = 64; layout (constant_id = 2) const uint BN = 64; layout (constant_id = 3) const uint BK = 16; // Assumed to be 32 if working with a quant +layout (constant_id = 4) const bool enable_smaller_matrices = false; +const uint BNover2 = enable_smaller_matrices ? (BN / 2) : BN; +const uint BNover4 = enable_smaller_matrices ? (BN / 4) : BN; + layout (push_constant) uniform parameter { uint M; @@ -48,6 +52,8 @@ layout (push_constant) uniform parameter uint broadcast2; uint broadcast3; #endif + // N dimension for the B matrix can be >= p.N + uint padded_N; } p; @@ -166,15 +172,13 @@ void main() { const uint end_k = min(p.K, (ik + 1) * p.k_split); #endif - coopmat sum; - sum = coopmat(0.0); - #ifdef MUL_MAT_ID uint pos_a = (expert_idx * p.batch_stride_a) / QUANT_K; uint pos_b = 0; #else uint pos_a = (batch_idx_a * p.batch_stride_a) / QUANT_K; uint pos_b = batch_idx * p.batch_stride_b; + uint pos_d = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * gl_NumWorkGroups.z; #endif uint stride_a = p.stride_a / QUANT_K; @@ -195,6 +199,7 @@ void main() { tensorLayoutNV<2> tensorLayoutB = createTensorLayoutNV(2); tensorLayoutNV<2, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutBClamp = createTensorLayoutNV(2, gl_CooperativeMatrixClampModeConstantNV); tensorLayoutNV<2, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutD = createTensorLayoutNV(2, gl_CooperativeMatrixClampModeConstantNV); + tensorLayoutD = setTensorLayoutStrideNV(tensorLayoutD, p.stride_d, 1); #if QUANT_K > 1 tensorLayoutA = setTensorLayoutBlockSizeNV(tensorLayoutA, 1, QUANT_K); @@ -202,18 +207,19 @@ void main() { #endif // Use end_k rather than p.K as the dimension because that's what - // we need to bound check against when using split_k + // we need to bound check against when using split_k. + // Bounds check B against padded_N, but bounds check D against N. tensorLayoutA = setTensorLayoutDimensionNV(tensorLayoutA, p.M, end_k); - tensorLayoutB = setTensorLayoutDimensionNV(tensorLayoutB, p.N, end_k); + tensorLayoutB = setTensorLayoutDimensionNV(tensorLayoutB, p.padded_N, end_k); tensorLayoutD = setTensorLayoutDimensionNV(tensorLayoutD, p.N, p.M); tensorLayoutAClamp = setTensorLayoutDimensionNV(tensorLayoutAClamp, p.M, end_k); - tensorLayoutBClamp = setTensorLayoutDimensionNV(tensorLayoutBClamp, p.N, end_k); + tensorLayoutBClamp = setTensorLayoutDimensionNV(tensorLayoutBClamp, p.padded_N, end_k); tensorViewNV<2, false, 1, 0> tensorViewTranspose = createTensorViewNV(2, false, 1, 0); #if !defined(MUL_MAT_ID) // Detect a fast path where all loads are entirely in bounds and no clamping is required - if ((ir + 1) * BM <= p.M && (ic + 1) * BN <= p.N && (start_k % BK) == 0 && (end_k % BK) == 0 && + if ((ir + 1) * BM <= p.M && (ic + 1) * BN <= p.padded_N && (start_k % BK) == 0 && (end_k % BK) == 0 && #if QUANT_K == 1 (stride_a % 8) == 0 && #endif @@ -229,16 +235,54 @@ void main() { tensorLayoutB = setTensorLayoutStrideNV(tensorLayoutB, stride_b, 1); uint k_iters = (end_k - start_k + BK - 1) / BK; + if (enable_smaller_matrices && ic * BN + BNover4 >= p.N) { + coopmat sum = coopmat(0.0); + for (uint block_k = start_k, i = 0; i < k_iters; block_k += BK, ++i) { - for (uint block_k = start_k, i = 0; i < k_iters; block_k += BK, ++i) { + coopmat mat_a; + coopmat mat_b; - coopmat mat_a; - coopmat mat_b; + coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); + coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover4, block_k, BK), tensorViewTranspose); + + sum = coopMatMulAdd(mat_a, mat_b, sum); + } + coopmat mat_d = coopmat(sum); + + coopMatStoreTensorNV(mat_d, data_d, pos_d, sliceTensorLayoutNV(tensorLayoutD, ic * BN, BNover4, ir * BM, BM), tensorViewTranspose); + return; + } else if (enable_smaller_matrices && ic * BN + BNover2 >= p.N) { + coopmat sum = coopmat(0.0); + for (uint block_k = start_k, i = 0; i < k_iters; block_k += BK, ++i) { + + coopmat mat_a; + coopmat mat_b; + + coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); + coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover2, block_k, BK), tensorViewTranspose); + + sum = coopMatMulAdd(mat_a, mat_b, sum); + } + coopmat mat_d = coopmat(sum); + + coopMatStoreTensorNV(mat_d, data_d, pos_d, sliceTensorLayoutNV(tensorLayoutD, ic * BN, BNover2, ir * BM, BM), tensorViewTranspose); + return; + } else { + coopmat sum = coopmat(0.0); + for (uint block_k = start_k, i = 0; i < k_iters; block_k += BK, ++i) { + + coopmat mat_a; + coopmat mat_b; - coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); - coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose); + coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); + coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose); + + sum = coopMatMulAdd(mat_a, mat_b, sum); + } + coopmat mat_d = coopmat(sum); - sum = coopMatMulAdd(mat_a, mat_b, sum); + coopMatStoreTensorNV(mat_d, data_d, pos_d, sliceTensorLayoutNV(tensorLayoutD, ic * BN, BN, ir * BM, BM), tensorViewTranspose); + return; } } else #endif // !defined(MUL_MAT_ID) @@ -251,6 +295,9 @@ void main() { tensorLayoutBClamp = setTensorLayoutStrideNV(tensorLayoutBClamp, stride_b, 1); + coopmat sum; + sum = coopmat(0.0); + [[dont_unroll]] for (uint block_k = start_k; block_k < end_k; block_k += BK) { @@ -263,7 +310,7 @@ void main() { #ifdef MUL_MAT_ID bool unclampedB = true; #else - bool unclampedB = (ic + 1) * BN <= p.N && block_k + BK <= end_k && (block_k % 8) == 0; + bool unclampedB = (ic + 1) * BN <= p.padded_N && block_k + BK <= end_k && (block_k % 8) == 0; #endif if (unclampedA && unclampedB) { coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, (block_k & ~7), BK) DECODEFUNCA); @@ -293,19 +340,16 @@ void main() { sum = coopMatMulAdd(mat_a, mat_b, sum); } } - } - // Convert from ACC_TYPE to D_TYPE - coopmat mat_d; - mat_d = coopmat(sum); + // Convert from ACC_TYPE to D_TYPE + coopmat mat_d; + mat_d = coopmat(sum); #ifdef MUL_MAT_ID - // Call callback to store each element, remapping row through shared memory - coopMatPerElementNV(mat_d, mat_d, perElemOpD, ir, ic); + // Call callback to store each element, remapping row through shared memory + coopMatPerElementNV(mat_d, mat_d, perElemOpD, ir, ic); #else - tensorLayoutD = setTensorLayoutStrideNV(tensorLayoutD, p.stride_d, 1); - - uint pos_d = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * gl_NumWorkGroups.z; - coopMatStoreTensorNV(mat_d, data_d, pos_d, sliceTensorLayoutNV(tensorLayoutD, ic * BN, BN, ir * BM, BM), tensorViewTranspose); + coopMatStoreTensorNV(mat_d, data_d, pos_d, sliceTensorLayoutNV(tensorLayoutD, ic * BN, BN, ir * BM, BM), tensorViewTranspose); #endif + } }