From 01012d65115f83c4d474c1a695b66cbb3f38c91a Mon Sep 17 00:00:00 2001 From: Ruben Ortlam Date: Thu, 22 Oct 2026 07:25:38 +0200 Subject: [PATCH] add f8 copy pipelines --- ggml/src/ggml-vulkan/ggml-vulkan-types.h | 4 ++-- ggml/src/ggml-vulkan/ggml-vulkan.cpp | 22 +++++++++++++++---- .../vulkan-shaders/contig_copy.comp | 10 +++++++-- ggml/src/ggml-vulkan/vulkan-shaders/copy.comp | 6 ++++- .../vulkan-shaders/vulkan-shaders-gen.cpp | 4 ++++ 5 files changed, 37 insertions(+), 9 deletions(-) diff --git a/ggml/src/ggml-vulkan/ggml-vulkan-types.h b/ggml/src/ggml-vulkan/ggml-vulkan-types.h index e8b986dddd..c5330312a6 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan-types.h +++ b/ggml/src/ggml-vulkan/ggml-vulkan-types.h @@ -838,8 +838,8 @@ struct vk_device_struct { vk_pipeline pipeline_roll_f32; vk_pipeline pipeline_repeat_i32, pipeline_repeat_back_f32; vk_pipeline pipeline_repeat_i16; - vk_pipeline pipeline_cpy_32, pipeline_cpy_f32_f16, pipeline_cpy_16, pipeline_cpy_f16_f32, pipeline_cpy_f32_bf16, pipeline_cpy_bf16_f32, pipeline_cpy_bf16_f16, pipeline_cpy_f32_i32, pipeline_cpy_i32_f32; - vk_pipeline pipeline_contig_cpy_32, pipeline_contig_cpy_f32_f16, pipeline_contig_cpy_16, pipeline_contig_cpy_f16_f32, pipeline_contig_cpy_f32_bf16, pipeline_contig_cpy_bf16_f32, pipeline_contig_cpy_bf16_f16, pipeline_contig_cpy_f32_i32, pipeline_contig_cpy_i32_f32; + vk_pipeline pipeline_cpy_32, pipeline_cpy_f32_f16, pipeline_cpy_16, pipeline_cpy_f16_f32, pipeline_cpy_f32_bf16, pipeline_cpy_bf16_f32, pipeline_cpy_bf16_f16, pipeline_cpy_8, pipeline_cpy_f8_e4m3_f16, pipeline_cpy_f32_i32, pipeline_cpy_i32_f32; + vk_pipeline pipeline_contig_cpy_32, pipeline_contig_cpy_f32_f16, pipeline_contig_cpy_16, pipeline_contig_cpy_f16_f32, pipeline_contig_cpy_f32_bf16, pipeline_contig_cpy_bf16_f32, pipeline_contig_cpy_bf16_f16, pipeline_contig_cpy_8, pipeline_contig_cpy_f8_e4m3_f16, pipeline_contig_cpy_f32_i32, pipeline_contig_cpy_i32_f32; vk_pipeline pipeline_cpy_f32_quant[GGML_TYPE_COUNT]; vk_pipeline pipeline_cpy_quant_f32[GGML_TYPE_COUNT]; vk_pipeline pipeline_cpy_transpose_16, pipeline_cpy_transpose_32; diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index ad98685de9..73eb213b9b 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -3331,6 +3331,8 @@ void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) { ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_bf16,"cpy_f32_bf16",cpy_f32_bf16_len,cpy_f32_bf16_data,"main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_cpy_bf16_f32,"cpy_bf16_f32",cpy_bf16_f32_len,cpy_bf16_f32_data,"main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_cpy_bf16_f16,"cpy_bf16_f16",cpy_bf16_f16_len,cpy_bf16_f16_data,"main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_8, "cpy_8", cpy_8_len, cpy_8_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f8_e4m3_f16, "cpy_f8_e4m3_f16", cpy_f8_e4m3_f16_len, cpy_f8_e4m3_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_cpy_i32_f32, "cpy_i32_f32", cpy_i32_f32_len, cpy_i32_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_i32, "cpy_f32_i32", cpy_f32_i32_len, cpy_f32_i32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); @@ -3341,6 +3343,8 @@ void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) { ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_bf16,"contig_cpy_f32_bf16",contig_cpy_f32_bf16_len,contig_cpy_f32_bf16_data,"main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_bf16_f32,"contig_cpy_bf16_f32",contig_cpy_bf16_f32_len,contig_cpy_bf16_f32_data,"main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_bf16_f16,"contig_cpy_bf16_f16",contig_cpy_bf16_f16_len,contig_cpy_bf16_f16_data,"main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_8, "contig_cpy_8", contig_cpy_8_len, contig_cpy_8_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f8_e4m3_f16, "contig_cpy_f8_e4m3_f16", contig_cpy_f8_e4m3_f16_len, contig_cpy_f8_e4m3_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_i32_f32, "contig_cpy_i32_f32", contig_cpy_i32_f32_len, contig_cpy_i32_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_i32, "contig_cpy_f32_i32", contig_cpy_f32_i32_len, contig_cpy_f32_i32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); @@ -6239,6 +6243,13 @@ vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, const ggml_t return ctx->device->pipeline_cpy_bf16_f16; } } + if (src->type == GGML_TYPE_F8_E4M3 && to == GGML_TYPE_F16) { + if (contig) { + return ctx->device->pipeline_contig_cpy_f8_e4m3_f16; + } else { + return ctx->device->pipeline_cpy_f8_e4m3_f16; + } + } if (src->type == GGML_TYPE_F32 && to == GGML_TYPE_I32) { if (contig) { return ctx->device->pipeline_contig_cpy_f32_i32; @@ -6285,6 +6296,9 @@ vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, const ggml_t } } + if (src->type == to && !ggml_is_quantized(to) && ggml_type_size(to) == 1) { + return contig ? ctx->device->pipeline_contig_cpy_8 : ctx->device->pipeline_cpy_8; + } if (src->type == to) { // Copy two or four bytes at a time, depending on block size. // For quantized types, we scale by block size/type size. But @@ -6425,7 +6439,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << ggml_type_name(src1->type) << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3]; std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << ggml_type_name(dst->type) << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3]; std::cerr << "))"); - GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16); // NOLINT + GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16 || src0->type == GGML_TYPE_F8_E4M3); // NOLINT GGML_ASSERT(ggml_vk_dim01_contiguous(src1) || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16); // NOLINT const uint64_t ne00 = src0->ne[0]; @@ -6787,7 +6801,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context& std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3]; std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3]; std::cerr << ")),)"); - GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16); // NOLINT + GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16 || src0->type == GGML_TYPE_F8_E4M3); // NOLINT GGML_ASSERT(ggml_vk_dim01_contiguous(src1) || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16); // NOLINT const uint64_t ne00 = src0->ne[0]; @@ -7862,7 +7876,7 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte std::cerr << "), (" << ids << ", name=" << ids->name << ", type=" << ids->type << ", ne0=" << ids->ne[0] << ", ne1=" << ids->ne[1] << ", ne2=" << ids->ne[2] << ", ne3=" << ids->ne[3] << ", nb0=" << ids->nb[0] << ", nb1=" << ids->nb[1] << ", nb2=" << ids->nb[2] << ", nb3=" << ids->nb[3]; std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3]; std::cerr << "))"); - GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16); // NOLINT + GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16 || src0->type == GGML_TYPE_F8_E4M3); // NOLINT GGML_ASSERT(ggml_vk_dim01_contiguous(src1) || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16); // NOLINT GGML_ASSERT(ids->type == GGML_TYPE_I32); @@ -15759,7 +15773,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm if (op->op == GGML_OP_MUL_MAT_ID && op->src[2]->ne[3] != op->src[1]->ne[3]) { return false; } - if (!(ggml_vk_dim01_contiguous(op->src[0]) || op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_BF16) || + if (!(ggml_vk_dim01_contiguous(op->src[0]) || op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_BF16 || op->src[0]->type == GGML_TYPE_F8_E4M3) || !(ggml_vk_dim01_contiguous(op->src[1]) || op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_F16)) { return false; } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp b/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp index b3b182fb08..db017cd04a 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp @@ -1,5 +1,7 @@ #version 450 +#extension GL_EXT_shader_8bit_storage : require + #include "types.glsl" #include "generic_unary_head.glsl" @@ -19,7 +21,9 @@ void main() { if (idx + (num_iter-1)*num_threads < p.ne) { [[unroll]] for (uint i = 0; i < num_iter; ++i) { -#if defined(DATA_A_BF16) +#if defined(DATA_A_F8_E4M3) + data_d[get_doffset() + idx] = D_TYPE(e4m3_to_fp32(data_a[get_aoffset() + idx])); +#elif defined(DATA_A_BF16) data_d[get_doffset() + idx] = D_TYPE(bf16_to_fp32(uint32_t(data_a[get_aoffset() + idx]))); #elif defined(DATA_D_BF16) float f = float(data_a[get_aoffset() + idx]); @@ -37,7 +41,9 @@ void main() { continue; } -#if defined(DATA_A_BF16) +#if defined(DATA_A_F8_E4M3) + data_d[get_doffset() + idx] = D_TYPE(e4m3_to_fp32(data_a[get_aoffset() + idx])); +#elif defined(DATA_A_BF16) data_d[get_doffset() + idx] = D_TYPE(bf16_to_fp32(uint32_t(data_a[get_aoffset() + idx]))); #elif defined(DATA_D_BF16) float f = float(data_a[get_aoffset() + idx]); diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp b/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp index d55e13253a..c2f4c02864 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp @@ -1,5 +1,7 @@ #version 450 +#extension GL_EXT_shader_8bit_storage : require + #include "types.glsl" #include "generic_unary_head.glsl" @@ -12,7 +14,9 @@ void main() { return; } -#if defined(DATA_A_BF16) +#if defined(DATA_A_F8_E4M3) + data_d[get_doffset() + dst_idx(idx)] = D_TYPE(e4m3_to_fp32(data_a[get_aoffset() + src0_idx(idx)])); +#elif defined(DATA_A_BF16) data_d[get_doffset() + dst_idx(idx)] = D_TYPE(bf16_to_fp32(uint32_t(data_a[get_aoffset() + src0_idx(idx)]))); #elif defined(DATA_D_BF16) float f = float(data_a[get_aoffset() + src0_idx(idx)]); 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 fefa174f14..756566698f 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp @@ -885,6 +885,8 @@ void process_shaders() { string_to_spv("cpy_f32_bf16","copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "uint16_t"}, {"DATA_D_BF16", "1"}}); string_to_spv("cpy_bf16_f32","copy.comp", {{"A_TYPE", "uint16_t"}, {"D_TYPE", "float"}, {"DATA_A_BF16", "1"}}); string_to_spv("cpy_bf16_f16","copy.comp", {{"A_TYPE", "uint16_t"}, {"D_TYPE", "float16_t"}, {"DATA_A_BF16", "1"}}); + string_to_spv("cpy_8", "copy.comp", {{"A_TYPE", "uint8_t"}, {"D_TYPE", "uint8_t"}}); + string_to_spv("cpy_f8_e4m3_f16", "copy.comp", {{"A_TYPE", "uint8_t"}, {"D_TYPE", "float16_t"}, {"DATA_A_F8_E4M3", "1"}}); string_to_spv("contig_cpy_32", "contig_copy.comp", {{"A_TYPE", "uint"}, {"D_TYPE", "uint"}}); string_to_spv("contig_cpy_f32_i32", "contig_copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "int"}}); string_to_spv("contig_cpy_i32_f32", "contig_copy.comp", {{"A_TYPE", "int"}, {"D_TYPE", "float"}}); @@ -894,6 +896,8 @@ void process_shaders() { string_to_spv("contig_cpy_f32_bf16","contig_copy.comp",{{"A_TYPE", "float"}, {"D_TYPE", "uint16_t"}, {"DATA_D_BF16", "1"}}); string_to_spv("contig_cpy_bf16_f32","contig_copy.comp",{{"A_TYPE", "uint16_t"}, {"D_TYPE", "float"}, {"DATA_A_BF16", "1"}}); string_to_spv("contig_cpy_bf16_f16","contig_copy.comp",{{"A_TYPE", "uint16_t"}, {"D_TYPE", "float16_t"}, {"DATA_A_BF16", "1"}}); + string_to_spv("contig_cpy_8", "contig_copy.comp", {{"A_TYPE", "uint8_t"}, {"D_TYPE", "uint8_t"}}); + string_to_spv("contig_cpy_f8_e4m3_f16", "contig_copy.comp", {{"A_TYPE", "uint8_t"}, {"D_TYPE", "float16_t"}, {"DATA_A_F8_E4M3", "1"}}); string_to_spv("cpy_f32_i32", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "int"}}); string_to_spv("cpy_i32_f32", "copy.comp", {{"A_TYPE", "int"}, {"D_TYPE", "float"}});