From 4e642bc3b4a74a8f756110cba32e0ee7c70345bd Mon Sep 17 00:00:00 2001 From: Jeff Bolz Date: Mon, 1 Sep 2025 17:05:26 -0500 Subject: [PATCH 1/2] vulkan: don't use std::string in load_shaders, to improve compile time --- ggml/src/ggml-vulkan/ggml-vulkan.cpp | 22 +++++++--------------- 1 file changed, 7 insertions(+), 15 deletions(-) diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index f7812ab37c865..9fd8a6421b52a 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -2340,7 +2340,7 @@ static void ggml_vk_load_shaders(vk_device& device) { } 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, + auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const char *name, size_t spv_size, const void* spv_data, const char *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, bool require_full_subgroups = false, uint32_t required_subgroup_size = 0) { @@ -3112,9 +3112,9 @@ static void ggml_vk_load_shaders(vk_device& device) { for (uint32_t i = 0; i < p021_max_gqa_ratio; ++i) { if (device->subgroup_arithmetic && device->subgroup_require_full_support) { - ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_p021_f16_f32[i], "mul_mat_vec_p021_f16_f32"+std::to_string(i+1), mul_mat_vec_p021_f16_f32_subgroup_add_len, mul_mat_vec_p021_f16_f32_subgroup_add_data, "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true, true); + ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_p021_f16_f32[i], "mul_mat_vec_p021_f16_f32", mul_mat_vec_p021_f16_f32_subgroup_add_len, mul_mat_vec_p021_f16_f32_subgroup_add_data, "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true, true); } else { - ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_p021_f16_f32[i], "mul_mat_vec_p021_f16_f32"+std::to_string(i+1), mul_mat_vec_p021_f16_f32_len, mul_mat_vec_p021_f16_f32_data, "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_p021_f16_f32[i], "mul_mat_vec_p021_f16_f32", mul_mat_vec_p021_f16_f32_len, mul_mat_vec_p021_f16_f32_data, "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true); } } ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", 3, 12 * sizeof(uint32_t), {1, 1, 1}, {}, 1); @@ -3187,19 +3187,11 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_Q8_0], "cpy_q8_0_f32", cpy_q8_0_f32_len, cpy_q8_0_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q8_0), 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_IQ4_NL], "cpy_iq4_nl_f32", cpy_iq4_nl_f32_len, cpy_iq4_nl_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_IQ4_NL), 1, 1}, {}, 1); - auto get_suffix = [](bool src0_f16, bool src1_f16, bool dst_f16) { - std::string s; - s += std::string(src0_f16 ? "_f16" : "_f32"); - s += std::string(src1_f16 ? "_f16" : "_f32"); - s += std::string(dst_f16 ? "_f16" : "_f32"); - return s; - }; - bool rte = device->float_controls_rte_fp16; #define CREATE_BINARY(name, namemod, spec, bindings) \ for (int s0 : {0,1}) for (int s1 : {0,1}) for (int d : {0,1}) \ ggml_vk_create_pipeline(device, device->pipeline_ ## name ## namemod[s0][s1][d], \ - #name + get_suffix(s0, s1, d) + #namemod, name ## _len[s0][s1][d][rte], name ## _data[s0][s1][d][rte], \ + #name #namemod, name ## _len[s0][s1][d][rte], name ## _data[s0][s1][d][rte], \ "main", (bindings), sizeof(vk_op_binary_push_constants), {512, 1, 1}, spec, 1); CREATE_BINARY(add, , {0}, 4) @@ -3216,8 +3208,8 @@ static void ggml_vk_load_shaders(vk_device& device) { if (device->multi_add) { for (uint32_t i = 0; i < MAX_FUSED_ADDS; ++i) { - ggml_vk_create_pipeline(device, device->pipeline_multi_add[i], "multi_add_f32_" + std::to_string(i+1), multi_add_f32_len, multi_add_f32_data, "main", MAX_PARAMETER_COUNT, sizeof(vk_op_multi_add_push_constants), {512, 1, 1}, {i+2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_multi_add_rms[i], "multi_add_rms_f32_" + std::to_string(i+1), multi_add_rms_f32_len, multi_add_rms_f32_data, "main", MAX_PARAMETER_COUNT, sizeof(vk_op_multi_add_push_constants), {512, 1, 1}, {i+2}, 1); + ggml_vk_create_pipeline(device, device->pipeline_multi_add[i], "multi_add_f32_" , multi_add_f32_len, multi_add_f32_data, "main", MAX_PARAMETER_COUNT, sizeof(vk_op_multi_add_push_constants), {512, 1, 1}, {i+2}, 1); + ggml_vk_create_pipeline(device, device->pipeline_multi_add_rms[i], "multi_add_rms_f32_" , multi_add_rms_f32_len, multi_add_rms_f32_data, "main", MAX_PARAMETER_COUNT, sizeof(vk_op_multi_add_push_constants), {512, 1, 1}, {i+2}, 1); } } @@ -3309,7 +3301,7 @@ static void ggml_vk_load_shaders(vk_device& device) { } for (uint32_t i = 0; i < num_argsort_pipelines; ++i) { - ggml_vk_create_pipeline(device, device->pipeline_argsort_f32[i], "argsort_f32_"+std::to_string(i), argsort_f32_len, argsort_f32_data, "main", 2, sizeof(vk_op_argsort_push_constants), {1u<pipeline_argsort_f32[i], "argsort_f32_", argsort_f32_len, argsort_f32_data, "main", 2, sizeof(vk_op_argsort_push_constants), {1u<pipeline_argmax_f32, "argmax_f32", argmax_f32_len, argmax_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); From 94bf9bc54e4e3d9a0513d170a68dba87613de4e0 Mon Sep 17 00:00:00 2001 From: Jeff Bolz Date: Tue, 2 Sep 2025 07:43:35 -0500 Subject: [PATCH 2/2] keep the string version for those calls that use it --- ggml/src/ggml-vulkan/ggml-vulkan.cpp | 30 +++++++++++++++++++++------- 1 file changed, 23 insertions(+), 7 deletions(-) diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index 9fd8a6421b52a..6e02bae34667e 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -2377,6 +2377,14 @@ static void ggml_vk_load_shaders(vk_device& device) { parameter_count, wg_denoms, specialization_constants, disable_robustness, require_full_subgroups, required_subgroup_size)); }; + auto const &ggml_vk_create_pipeline2 = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const char *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, bool require_full_subgroups = false, uint32_t required_subgroup_size = 0) { + return ggml_vk_create_pipeline(device, pipeline, name.c_str(), spv_size, spv_data, entrypoint, + parameter_count, push_constant_size, wg_denoms, specialization_constants, + align, disable_robustness, require_full_subgroups, required_subgroup_size); + }; + auto const &fa_wg_denoms = [&](FaCodePath path, uint32_t hsk, uint32_t hsv, uint32_t clamp, ggml_type type, bool small_rows) -> std::array { return {fa_rows_cols(path, hsk, hsv, clamp, type, small_rows)[0], 1, 1}; }; @@ -3112,9 +3120,9 @@ static void ggml_vk_load_shaders(vk_device& device) { for (uint32_t i = 0; i < p021_max_gqa_ratio; ++i) { if (device->subgroup_arithmetic && device->subgroup_require_full_support) { - ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_p021_f16_f32[i], "mul_mat_vec_p021_f16_f32", mul_mat_vec_p021_f16_f32_subgroup_add_len, mul_mat_vec_p021_f16_f32_subgroup_add_data, "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true, true); + ggml_vk_create_pipeline2(device, device->pipeline_mul_mat_vec_p021_f16_f32[i], "mul_mat_vec_p021_f16_f32"+std::to_string(i+1), mul_mat_vec_p021_f16_f32_subgroup_add_len, mul_mat_vec_p021_f16_f32_subgroup_add_data, "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true, true); } else { - ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_p021_f16_f32[i], "mul_mat_vec_p021_f16_f32", mul_mat_vec_p021_f16_f32_len, mul_mat_vec_p021_f16_f32_data, "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true); + ggml_vk_create_pipeline2(device, device->pipeline_mul_mat_vec_p021_f16_f32[i], "mul_mat_vec_p021_f16_f32"+std::to_string(i+1), mul_mat_vec_p021_f16_f32_len, mul_mat_vec_p021_f16_f32_data, "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true); } } ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", 3, 12 * sizeof(uint32_t), {1, 1, 1}, {}, 1); @@ -3187,11 +3195,19 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_Q8_0], "cpy_q8_0_f32", cpy_q8_0_f32_len, cpy_q8_0_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q8_0), 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_IQ4_NL], "cpy_iq4_nl_f32", cpy_iq4_nl_f32_len, cpy_iq4_nl_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_IQ4_NL), 1, 1}, {}, 1); + auto get_suffix = [](bool src0_f16, bool src1_f16, bool dst_f16) { + std::string s; + s += std::string(src0_f16 ? "_f16" : "_f32"); + s += std::string(src1_f16 ? "_f16" : "_f32"); + s += std::string(dst_f16 ? "_f16" : "_f32"); + return s; + }; + bool rte = device->float_controls_rte_fp16; #define CREATE_BINARY(name, namemod, spec, bindings) \ for (int s0 : {0,1}) for (int s1 : {0,1}) for (int d : {0,1}) \ - ggml_vk_create_pipeline(device, device->pipeline_ ## name ## namemod[s0][s1][d], \ - #name #namemod, name ## _len[s0][s1][d][rte], name ## _data[s0][s1][d][rte], \ + ggml_vk_create_pipeline2(device, device->pipeline_ ## name ## namemod[s0][s1][d], \ + #name + get_suffix(s0, s1, d) + #namemod, name ## _len[s0][s1][d][rte], name ## _data[s0][s1][d][rte], \ "main", (bindings), sizeof(vk_op_binary_push_constants), {512, 1, 1}, spec, 1); CREATE_BINARY(add, , {0}, 4) @@ -3208,8 +3224,8 @@ static void ggml_vk_load_shaders(vk_device& device) { if (device->multi_add) { for (uint32_t i = 0; i < MAX_FUSED_ADDS; ++i) { - ggml_vk_create_pipeline(device, device->pipeline_multi_add[i], "multi_add_f32_" , multi_add_f32_len, multi_add_f32_data, "main", MAX_PARAMETER_COUNT, sizeof(vk_op_multi_add_push_constants), {512, 1, 1}, {i+2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_multi_add_rms[i], "multi_add_rms_f32_" , multi_add_rms_f32_len, multi_add_rms_f32_data, "main", MAX_PARAMETER_COUNT, sizeof(vk_op_multi_add_push_constants), {512, 1, 1}, {i+2}, 1); + ggml_vk_create_pipeline2(device, device->pipeline_multi_add[i], "multi_add_f32_" + std::to_string(i+1), multi_add_f32_len, multi_add_f32_data, "main", MAX_PARAMETER_COUNT, sizeof(vk_op_multi_add_push_constants), {512, 1, 1}, {i+2}, 1); + ggml_vk_create_pipeline2(device, device->pipeline_multi_add_rms[i], "multi_add_rms_f32_" + std::to_string(i+1), multi_add_rms_f32_len, multi_add_rms_f32_data, "main", MAX_PARAMETER_COUNT, sizeof(vk_op_multi_add_push_constants), {512, 1, 1}, {i+2}, 1); } } @@ -3301,7 +3317,7 @@ static void ggml_vk_load_shaders(vk_device& device) { } for (uint32_t i = 0; i < num_argsort_pipelines; ++i) { - ggml_vk_create_pipeline(device, device->pipeline_argsort_f32[i], "argsort_f32_", argsort_f32_len, argsort_f32_data, "main", 2, sizeof(vk_op_argsort_push_constants), {1u<pipeline_argsort_f32[i], "argsort_f32_"+std::to_string(i), argsort_f32_len, argsort_f32_data, "main", 2, sizeof(vk_op_argsort_push_constants), {1u<pipeline_argmax_f32, "argmax_f32", argmax_f32_len, argmax_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);