CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-27 08:21:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

Vulkan: Add Integer Dot Product mul_mat_vec shader for legacy quants (#14903)

Browse Source 
* vulkan: Add Integer Dot Product mul_mat_vec shader for legacy quants

* vulkan: use subgroup operations for quantize_q8_1 shader

* vulkan: add q8_1_x4 type with 128-bit alignment, use in mul_mat_vecq shader

* vulkan: use q8_1_x4 blocks in mul_mmq shader

* vulkan: do 8 calculations per invocation instead of 32 in mul_mat_vecq, similar to mul_mat_vec

* vulkan: tune mul_mat_vecq performance for Intel

* vulkan: fix quantizing issue when tensor is not divisible by 128

* vulkan: adapt integer dot mmv to mmv small m optimization (#15355)

* vulkan: allow all subgroup modes for mmv and mmvq

* vulkan: use prealloc intermediate reuse for mmvq path

* vulkan: tune mmvq for Intel, AMD GCN and Nvidia RTX 3090

* vulkan: adapt mmv quantize_y path to conditional sync logic

* vulkan: disable q8_0 mmvq on Nvidia

* vulkan: enable q8_0 on Nvidia pre-turing

* fix prealloc sync condition

* fix llvmpipe subgroup 8 issue
This commit is contained in:
Ruben Ortlam
2025-09-01 16:19:07 +02:00
committed by GitHub
parent 77dee9de97
commit 02c1813517
8 changed files with 559 additions and 115 deletions
Download Patch File Download Diff File Expand all files Collapse all files

353
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View File

@@ -360,6 +360,13 @@ struct vk_fa_pipeline_state {
}
};
enum shader_reduction_mode {
SHADER_REDUCTION_MODE_SHMEM,
SHADER_REDUCTION_MODE_HYBRID,
SHADER_REDUCTION_MODE_SUBGROUP,
SHADER_REDUCTION_MODE_COUNT,
};
static constexpr uint32_t num_argsort_pipelines = 11;
static constexpr uint32_t max_argsort_cols = 1 << (num_argsort_pipelines-1);
@@ -386,15 +393,18 @@ struct vk_device_struct {
bool uma;
bool prefer_host_memory;
bool float_controls_rte_fp16;
bool subgroup_add;
bool subgroup_arithmetic;
bool subgroup_shuffle;
bool subgroup_ballot;
bool subgroup_clustered;
bool multi_add;
bool add_rms_fusion;
uint32_t partials_binding_alignment;
bool integer_dot_product;
// 0: default, 1: force mmvq, -1: disable mmvq
int32_t mmvq_mode;
bool subgroup_size_control;
uint32_t subgroup_min_size;
@@ -452,12 +462,15 @@ struct vk_device_struct {
vk_pipeline pipeline_matmul_split_k_reduce;
vk_pipeline pipeline_quantize_q8_1;
vk_pipeline pipeline_quantize_q8_1_x4;
vk_pipeline pipeline_dequant[GGML_TYPE_COUNT];
vk_pipeline pipeline_dequant_mul_mat_vec_f32_f32[DMMV_WG_SIZE_COUNT][GGML_TYPE_COUNT][mul_mat_vec_max_cols];
vk_pipeline pipeline_dequant_mul_mat_vec_f16_f32[DMMV_WG_SIZE_COUNT][GGML_TYPE_COUNT][mul_mat_vec_max_cols];
vk_pipeline pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_COUNT];
vk_pipeline pipeline_dequant_mul_mat_vec_q8_1_f32[DMMV_WG_SIZE_COUNT][GGML_TYPE_COUNT][mul_mat_vec_max_cols];
vk_pipeline pipeline_mul_mat_vec_p021_f16_f32[p021_max_gqa_ratio];
vk_pipeline pipeline_mul_mat_vec_nc_f16_f32;
vk_pipeline pipeline_get_rows[GGML_TYPE_COUNT];
@@ -2911,60 +2924,91 @@ static void ggml_vk_load_shaders(vk_device& device) {
rm_stdq = 2;
uint32_t rm_iq = 2 * rm_kq;
for (uint32_t w = 0; w < DMMV_WG_SIZE_COUNT; ++w) {
uint32_t wg_size_subgroup16 = (w == DMMV_WG_SIZE_SUBGROUP) ? subgroup_size_16 : (subgroup_size_16 * 4);
uint32_t wg_size_subgroup = (w == DMMV_WG_SIZE_SUBGROUP) ? device->subgroup_size : (device->subgroup_size * 4);
const bool use_subgroups = device->subgroup_arithmetic && device->architecture != vk_device_architecture::AMD_GCN;
// Ensure a subgroup size >= 16 is available
const bool use_subgroups16 = use_subgroups &&
(!device->subgroup_size_control && device->subgroup_size >= 16 ||
device->subgroup_size_control && device->subgroup_min_size <= 16 && device->subgroup_max_size >= 16);
const bool s = device->subgroup_add && device->architecture != vk_device_architecture::AMD_GCN;
const uint32_t subgroup_size = (device->vendor_id == VK_VENDOR_ID_INTEL && device->subgroup_size_control && device->subgroup_min_size <= 16 && device->subgroup_max_size >= 16) ? 16 : device->subgroup_size;
const uint32_t subgroup_size16 = std::max(subgroup_size, 16u);
const uint32_t force_subgroup_size = use_subgroups ? subgroup_size : 0;
const uint32_t force_subgroup_size16 = use_subgroups16 ? subgroup_size16 : 0;
for (uint32_t w = 0; w < DMMV_WG_SIZE_COUNT; ++w) {
const uint32_t wg_size_subgroup = (w == DMMV_WG_SIZE_SUBGROUP) ? subgroup_size : (subgroup_size * 4);
const uint32_t wg_size_subgroup16 = (w == DMMV_WG_SIZE_SUBGROUP) ? subgroup_size16 : (subgroup_size16 * 4);
const shader_reduction_mode reduc = (use_subgroups && w == DMMV_WG_SIZE_SUBGROUP) ? SHADER_REDUCTION_MODE_SUBGROUP :
(use_subgroups && w == DMMV_WG_SIZE_LARGE) ? SHADER_REDUCTION_MODE_HYBRID :
SHADER_REDUCTION_MODE_SHMEM;
const shader_reduction_mode reduc16 = (use_subgroups16 && w == DMMV_WG_SIZE_SUBGROUP) ? SHADER_REDUCTION_MODE_SUBGROUP :
(use_subgroups16 && w == DMMV_WG_SIZE_LARGE) ? SHADER_REDUCTION_MODE_HYBRID :
SHADER_REDUCTION_MODE_SHMEM;
for (uint32_t i = 0; i < mul_mat_vec_max_cols; ++i) {
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f32_f32", arr_dmmv_f32_f32_f32_len[s], arr_dmmv_f32_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f32_f32", arr_dmmv_f16_f32_f32_len[s], arr_dmmv_f16_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f32_f32", arr_dmmv_bf16_f32_f32_len[s], arr_dmmv_bf16_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f32_f32", arr_dmmv_q4_0_f32_f32_len[s], arr_dmmv_q4_0_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f32_f32", arr_dmmv_q4_1_f32_f32_len[s], arr_dmmv_q4_1_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f32_f32", arr_dmmv_q5_0_f32_f32_len[s], arr_dmmv_q5_0_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_f32_f32", arr_dmmv_q5_1_f32_f32_len[s], arr_dmmv_q5_1_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f32_f32", arr_dmmv_q8_0_f32_f32_len[s], arr_dmmv_q8_0_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {wg_size_subgroup, 1*rm_stdq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f32_f32", arr_dmmv_q2_k_f32_f32_len[s], arr_dmmv_q2_k_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f32_f32", arr_dmmv_q3_k_f32_f32_len[s], arr_dmmv_q3_k_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f32_f32", arr_dmmv_q4_k_f32_f32_len[s], arr_dmmv_q4_k_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f32_f32", arr_dmmv_q5_k_f32_f32_len[s], arr_dmmv_q5_k_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f32_f32", arr_dmmv_q6_k_f32_f32_len[s], arr_dmmv_q6_k_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ1_S][i], "mul_mat_vec_iq1_s_f32_f32", arr_dmmv_iq1_s_f32_f32_len[s], arr_dmmv_iq1_s_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ1_M][i], "mul_mat_vec_iq1_m_f32_f32", arr_dmmv_iq1_m_f32_f32_len[s], arr_dmmv_iq1_m_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f32_f32", arr_dmmv_iq2_xxs_f32_f32_len[s], arr_dmmv_iq2_xxs_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ2_XS][i], "mul_mat_vec_iq2_xs_f32_f32", arr_dmmv_iq2_xs_f32_f32_len[s], arr_dmmv_iq2_xs_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ2_S][i], "mul_mat_vec_iq2_s_f32_f32", arr_dmmv_iq2_s_f32_f32_len[s], arr_dmmv_iq2_s_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f32_f32", arr_dmmv_iq3_xxs_f32_f32_len[s], arr_dmmv_iq3_xxs_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ3_S][i], "mul_mat_vec_iq3_s_f32_f32", arr_dmmv_iq3_s_f32_f32_len[s], arr_dmmv_iq3_s_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ4_XS][i], "mul_mat_vec_iq4_xs_f32_f32", arr_dmmv_iq4_xs_f32_f32_len[s], arr_dmmv_iq4_xs_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ4_NL][i], "mul_mat_vec_iq4_nl_f32_f32", arr_dmmv_iq4_nl_f32_f32_len[s], arr_dmmv_iq4_nl_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_MXFP4][i], "mul_mat_vec_mxfp4_f32_f32", arr_dmmv_mxfp4_f32_f32_len[s], arr_dmmv_mxfp4_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f32_f32", arr_dmmv_f32_f32_f32_len[reduc], arr_dmmv_f32_f32_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1, false, use_subgroups, force_subgroup_size);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f32_f32", arr_dmmv_f16_f32_f32_len[reduc], arr_dmmv_f16_f32_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1, false, use_subgroups, force_subgroup_size);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f32_f32", arr_dmmv_bf16_f32_f32_len[reduc], arr_dmmv_bf16_f32_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1, false, use_subgroups, force_subgroup_size);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f32_f32", arr_dmmv_q4_0_f32_f32_len[reduc], arr_dmmv_q4_0_f32_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f32_f32", arr_dmmv_q4_1_f32_f32_len[reduc], arr_dmmv_q4_1_f32_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f32_f32", arr_dmmv_q5_0_f32_f32_len[reduc], arr_dmmv_q5_0_f32_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_f32_f32", arr_dmmv_q5_1_f32_f32_len[reduc], arr_dmmv_q5_1_f32_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f32_f32", arr_dmmv_q8_0_f32_f32_len[reduc], arr_dmmv_q8_0_f32_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {wg_size_subgroup, 1*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f32_f32", arr_dmmv_q2_k_f32_f32_len[reduc16], arr_dmmv_q2_k_f32_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f32_f32", arr_dmmv_q3_k_f32_f32_len[reduc16], arr_dmmv_q3_k_f32_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f32_f32", arr_dmmv_q4_k_f32_f32_len[reduc16], arr_dmmv_q4_k_f32_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f32_f32", arr_dmmv_q5_k_f32_f32_len[reduc16], arr_dmmv_q5_k_f32_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f32_f32", arr_dmmv_q6_k_f32_f32_len[reduc16], arr_dmmv_q6_k_f32_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ1_S][i], "mul_mat_vec_iq1_s_f32_f32", arr_dmmv_iq1_s_f32_f32_len[reduc16], arr_dmmv_iq1_s_f32_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ1_M][i], "mul_mat_vec_iq1_m_f32_f32", arr_dmmv_iq1_m_f32_f32_len[reduc16], arr_dmmv_iq1_m_f32_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f32_f32", arr_dmmv_iq2_xxs_f32_f32_len[reduc16], arr_dmmv_iq2_xxs_f32_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ2_XS][i], "mul_mat_vec_iq2_xs_f32_f32", arr_dmmv_iq2_xs_f32_f32_len[reduc16], arr_dmmv_iq2_xs_f32_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ2_S][i], "mul_mat_vec_iq2_s_f32_f32", arr_dmmv_iq2_s_f32_f32_len[reduc16], arr_dmmv_iq2_s_f32_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f32_f32", arr_dmmv_iq3_xxs_f32_f32_len[reduc16], arr_dmmv_iq3_xxs_f32_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ3_S][i], "mul_mat_vec_iq3_s_f32_f32", arr_dmmv_iq3_s_f32_f32_len[reduc16], arr_dmmv_iq3_s_f32_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ4_XS][i], "mul_mat_vec_iq4_xs_f32_f32", arr_dmmv_iq4_xs_f32_f32_len[reduc16], arr_dmmv_iq4_xs_f32_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ4_NL][i], "mul_mat_vec_iq4_nl_f32_f32", arr_dmmv_iq4_nl_f32_f32_len[reduc16], arr_dmmv_iq4_nl_f32_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_MXFP4][i], "mul_mat_vec_mxfp4_f32_f32", arr_dmmv_mxfp4_f32_f32_len[reduc16], arr_dmmv_mxfp4_f32_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f16_f32", arr_dmmv_f32_f16_f32_len[s], arr_dmmv_f32_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f16_f32", arr_dmmv_f16_f16_f32_len[s], arr_dmmv_f16_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f16_f32", arr_dmmv_bf16_f16_f32_len[s], arr_dmmv_bf16_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f16_f32", arr_dmmv_q4_0_f16_f32_len[s], arr_dmmv_q4_0_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f16_f32", arr_dmmv_q4_1_f16_f32_len[s], arr_dmmv_q4_1_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f16_f32", arr_dmmv_q5_0_f16_f32_len[s], arr_dmmv_q5_0_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_f16_f32", arr_dmmv_q5_1_f16_f32_len[s], arr_dmmv_q5_1_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f16_f32", arr_dmmv_q8_0_f16_f32_len[s], arr_dmmv_q8_0_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {wg_size_subgroup, 1*rm_stdq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f16_f32", arr_dmmv_q2_k_f16_f32_len[s], arr_dmmv_q2_k_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f16_f32", arr_dmmv_q3_k_f16_f32_len[s], arr_dmmv_q3_k_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f16_f32", arr_dmmv_q4_k_f16_f32_len[s], arr_dmmv_q4_k_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f16_f32", arr_dmmv_q5_k_f16_f32_len[s], arr_dmmv_q5_k_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f16_f32", arr_dmmv_q6_k_f16_f32_len[s], arr_dmmv_q6_k_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ1_S][i], "mul_mat_vec_iq1_s_f16_f32", arr_dmmv_iq1_s_f16_f32_len[s], arr_dmmv_iq1_s_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ1_M][i], "mul_mat_vec_iq1_m_f16_f32", arr_dmmv_iq1_m_f16_f32_len[s], arr_dmmv_iq1_m_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f16_f32", arr_dmmv_iq2_xxs_f16_f32_len[s], arr_dmmv_iq2_xxs_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ2_XS][i], "mul_mat_vec_iq2_xs_f16_f32", arr_dmmv_iq2_xs_f16_f32_len[s], arr_dmmv_iq2_xs_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ2_S][i], "mul_mat_vec_iq2_s_f16_f32", arr_dmmv_iq2_s_f16_f32_len[s], arr_dmmv_iq2_s_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f16_f32", arr_dmmv_iq3_xxs_f16_f32_len[s], arr_dmmv_iq3_xxs_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ3_S][i], "mul_mat_vec_iq3_s_f16_f32", arr_dmmv_iq3_s_f16_f32_len[s], arr_dmmv_iq3_s_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ4_XS][i], "mul_mat_vec_iq4_xs_f16_f32", arr_dmmv_iq4_xs_f16_f32_len[s], arr_dmmv_iq4_xs_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ4_NL][i], "mul_mat_vec_iq4_nl_f16_f32", arr_dmmv_iq4_nl_f16_f32_len[s], arr_dmmv_iq4_nl_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_MXFP4][i], "mul_mat_vec_mxfp4_f16_f32", arr_dmmv_mxfp4_f16_f32_len[s], arr_dmmv_mxfp4_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f16_f32", arr_dmmv_f32_f16_f32_len[reduc], arr_dmmv_f32_f16_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1, false, use_subgroups, force_subgroup_size);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f16_f32", arr_dmmv_f16_f16_f32_len[reduc], arr_dmmv_f16_f16_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1, false, use_subgroups, force_subgroup_size);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f16_f32", arr_dmmv_bf16_f16_f32_len[reduc], arr_dmmv_bf16_f16_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1, false, use_subgroups, force_subgroup_size);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f16_f32", arr_dmmv_q4_0_f16_f32_len[reduc], arr_dmmv_q4_0_f16_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f16_f32", arr_dmmv_q4_1_f16_f32_len[reduc], arr_dmmv_q4_1_f16_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f16_f32", arr_dmmv_q5_0_f16_f32_len[reduc], arr_dmmv_q5_0_f16_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_f16_f32", arr_dmmv_q5_1_f16_f32_len[reduc], arr_dmmv_q5_1_f16_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f16_f32", arr_dmmv_q8_0_f16_f32_len[reduc], arr_dmmv_q8_0_f16_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {wg_size_subgroup, 1*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f16_f32", arr_dmmv_q2_k_f16_f32_len[reduc16], arr_dmmv_q2_k_f16_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f16_f32", arr_dmmv_q3_k_f16_f32_len[reduc16], arr_dmmv_q3_k_f16_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f16_f32", arr_dmmv_q4_k_f16_f32_len[reduc16], arr_dmmv_q4_k_f16_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f16_f32", arr_dmmv_q5_k_f16_f32_len[reduc16], arr_dmmv_q5_k_f16_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f16_f32", arr_dmmv_q6_k_f16_f32_len[reduc16], arr_dmmv_q6_k_f16_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ1_S][i], "mul_mat_vec_iq1_s_f16_f32", arr_dmmv_iq1_s_f16_f32_len[reduc16], arr_dmmv_iq1_s_f16_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ1_M][i], "mul_mat_vec_iq1_m_f16_f32", arr_dmmv_iq1_m_f16_f32_len[reduc16], arr_dmmv_iq1_m_f16_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f16_f32", arr_dmmv_iq2_xxs_f16_f32_len[reduc16], arr_dmmv_iq2_xxs_f16_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ2_XS][i], "mul_mat_vec_iq2_xs_f16_f32", arr_dmmv_iq2_xs_f16_f32_len[reduc16], arr_dmmv_iq2_xs_f16_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ2_S][i], "mul_mat_vec_iq2_s_f16_f32", arr_dmmv_iq2_s_f16_f32_len[reduc16], arr_dmmv_iq2_s_f16_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f16_f32", arr_dmmv_iq3_xxs_f16_f32_len[reduc16], arr_dmmv_iq3_xxs_f16_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ3_S][i], "mul_mat_vec_iq3_s_f16_f32", arr_dmmv_iq3_s_f16_f32_len[reduc16], arr_dmmv_iq3_s_f16_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ4_XS][i], "mul_mat_vec_iq4_xs_f16_f32", arr_dmmv_iq4_xs_f16_f32_len[reduc16], arr_dmmv_iq4_xs_f16_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ4_NL][i], "mul_mat_vec_iq4_nl_f16_f32", arr_dmmv_iq4_nl_f16_f32_len[reduc16], arr_dmmv_iq4_nl_f16_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_MXFP4][i], "mul_mat_vec_mxfp4_f16_f32", arr_dmmv_mxfp4_f16_f32_len[reduc16], arr_dmmv_mxfp4_f16_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
if (device->integer_dot_product) {
const uint32_t subgroup_size_int = (device->vendor_id == VK_VENDOR_ID_INTEL && device->subgroup_size_control) ? device->subgroup_min_size : device->subgroup_size;
const uint32_t wg_size_subgroup_int = (w == DMMV_WG_SIZE_SUBGROUP) ? subgroup_size_int : (subgroup_size_int * 4);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[w][GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_q8_1_f32", arr_dmmv_q4_0_q8_1_f32_len[reduc], arr_dmmv_q4_0_q8_1_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup_int, 2*rm_stdq, i+1}, 1, true, use_subgroups, subgroup_size_int);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[w][GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_q8_1_f32", arr_dmmv_q4_1_q8_1_f32_len[reduc], arr_dmmv_q4_1_q8_1_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup_int, 2*rm_stdq, i+1}, 1, true, use_subgroups, subgroup_size_int);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[w][GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_q8_1_f32", arr_dmmv_q5_0_q8_1_f32_len[reduc], arr_dmmv_q5_0_q8_1_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup_int, 2*rm_stdq, i+1}, 1, true, use_subgroups, subgroup_size_int);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[w][GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_q8_1_f32", arr_dmmv_q5_1_q8_1_f32_len[reduc], arr_dmmv_q5_1_q8_1_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup_int, 2*rm_stdq, i+1}, 1, true, use_subgroups, subgroup_size_int);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[w][GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_q8_1_f32", arr_dmmv_q8_0_q8_1_f32_len[reduc], arr_dmmv_q8_0_q8_1_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {wg_size_subgroup_int, 1*rm_stdq, i+1}, 1, true, use_subgroups, subgroup_size_int);
}
#endif // GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT
}
}
@@ -3056,10 +3100,17 @@ static void ggml_vk_load_shaders(vk_device& device) {
ggml_vk_create_pipeline(device, device->pipeline_matmul_split_k_reduce, "split_k_reduce", split_k_reduce_len, split_k_reduce_data, "main", 2, 2 * sizeof(uint32_t), {256 * 4, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_flash_attn_split_k_reduce, "fa_split_k_reduce", fa_split_k_reduce_len, fa_split_k_reduce_data, "main", 3, 5 * sizeof(uint32_t), {1, device->subgroup_size, 1}, {device->subgroup_size}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_quantize_q8_1, "quantize_q8_1", quantize_q8_1_len, quantize_q8_1_data, "main", 2, 1 * sizeof(uint32_t), {32 * device->subgroup_size / 8, 1, 1}, { device->subgroup_size }, 1);
if (device->subgroup_clustered && device->subgroup_require_full_support) {
ggml_vk_create_pipeline(device, device->pipeline_quantize_q8_1, "quantize_q8_1", quantize_q8_1_subgroup_len, quantize_q8_1_subgroup_data, "main", 2, 1 * sizeof(uint32_t), {32 * device->subgroup_size / 8, 1, 1}, { device->subgroup_size }, 1, true, true);
ggml_vk_create_pipeline(device, device->pipeline_quantize_q8_1_x4, "quantize_q8_1_x4", quantize_q8_1_x4_subgroup_len, quantize_q8_1_x4_subgroup_data, "main", 2, 1 * sizeof(uint32_t), {32 * device->subgroup_size / 8, 1, 1}, { device->subgroup_size }, 1, true, true);
} else {
ggml_vk_create_pipeline(device, device->pipeline_quantize_q8_1, "quantize_q8_1", quantize_q8_1_len, quantize_q8_1_data, "main", 2, 1 * sizeof(uint32_t), {32 * device->subgroup_size / 8, 1, 1}, { device->subgroup_size }, 1);
ggml_vk_create_pipeline(device, device->pipeline_quantize_q8_1_x4, "quantize_q8_1_x4", quantize_q8_1_x4_len, quantize_q8_1_x4_data, "main", 2, 1 * sizeof(uint32_t), {32 * device->subgroup_size / 8, 1, 1}, { device->subgroup_size }, 1);
}
for (uint32_t i = 0; i < p021_max_gqa_ratio; ++i) {
if (device->subgroup_add && device->subgroup_require_full_support) {
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);
} 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);
@@ -3578,11 +3629,12 @@ static vk_device ggml_vk_get_device(size_t idx) {
}
device->float_controls_rte_fp16 = vk12_props.shaderRoundingModeRTEFloat16;
device->subgroup_add = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
(vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eArithmetic);
device->subgroup_arithmetic = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
(vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eArithmetic);
device->subgroup_shuffle = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
(vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eShuffle);
device->subgroup_clustered = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
(vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eClustered);
device->subgroup_ballot = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
(vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eBallot);
@@ -4038,11 +4090,18 @@ static vk_device ggml_vk_get_device(size_t idx) {
device->disable_fusion = getenv("GGML_VK_DISABLE_FUSION") != nullptr;
device->add_rms_fusion = !device->disable_fusion &&
device->subgroup_add &&
device->subgroup_arithmetic &&
device->vendor_id != VK_VENDOR_ID_INTEL;
device->partials_binding_alignment =
std::max(4u, (uint32_t)device->properties.limits.minStorageBufferOffsetAlignment);
device->mmvq_mode = 0;
if (getenv("GGML_VK_DISABLE_MMVQ")) {
device->mmvq_mode = -1;
} else if (getenv("GGML_VK_FORCE_MMVQ")) {
device->mmvq_mode = 1;
}
return device;
}
@@ -4556,9 +4615,22 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte
static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type, uint32_t num_cols, uint32_t m, uint32_t k) {
VK_LOG_DEBUG("ggml_vk_get_dequantize_mul_mat_vec()");
GGML_ASSERT(b_type == GGML_TYPE_F32 || b_type == GGML_TYPE_F16);
GGML_ASSERT(b_type == GGML_TYPE_F32 || b_type == GGML_TYPE_F16 || b_type == GGML_TYPE_Q8_1);
GGML_ASSERT(num_cols >= 1 && num_cols <= mul_mat_vec_max_cols);
if (b_type == GGML_TYPE_Q8_1) {
switch (a_type) {
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q5_0:
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
break;
default:
return nullptr;
}
}
switch (a_type) {
case GGML_TYPE_F32:
case GGML_TYPE_F16:
@@ -4605,6 +4677,13 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context *
}
}
if (b_type == GGML_TYPE_Q8_1) {
if (ctx->device->vendor_id == VK_VENDOR_ID_INTEL) {
dmmv_wg = DMMV_WG_SIZE_SUBGROUP;
}
return ctx->device->pipeline_dequant_mul_mat_vec_q8_1_f32[dmmv_wg][a_type][num_cols-1];
}
return b_type == GGML_TYPE_F32 ? ctx->device->pipeline_dequant_mul_mat_vec_f32_f32[dmmv_wg][a_type][num_cols-1] : ctx->device->pipeline_dequant_mul_mat_vec_f16_f32[dmmv_wg][a_type][num_cols-1];
}
@@ -4674,7 +4753,7 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_co
}
static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec_id(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type) {
VK_LOG_DEBUG("ggml_vk_get_dequantize_mul_mat_vec()");
VK_LOG_DEBUG("ggml_vk_get_dequantize_mul_mat_vec_id()");
GGML_ASSERT(b_type == GGML_TYPE_F32);
switch (a_type) {
@@ -5587,20 +5666,20 @@ static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context&
ggml_vk_sync_buffers(ctx, subctx);
}
static vk_pipeline ggml_vk_get_quantize_pipeline(ggml_backend_vk_context * ctx, ggml_type type) {
static vk_pipeline ggml_vk_get_quantize_pipeline(ggml_backend_vk_context * ctx, ggml_type type, bool use_x4_blocks) {
switch(type) {
case GGML_TYPE_Q8_1:
return ctx->device->pipeline_quantize_q8_1;
return use_x4_blocks ? ctx->device->pipeline_quantize_q8_1_x4 : ctx->device->pipeline_quantize_q8_1;
default:
std::cerr << "Missing quantize pipeline for type: " << ggml_type_name(type) << std::endl;
GGML_ABORT("fatal error");
}
}
static void ggml_vk_quantize_q8_1(ggml_backend_vk_context * ctx, vk_context& subctx, vk_subbuffer&& in, vk_subbuffer&& out, uint32_t ne) {
static void ggml_vk_quantize_q8_1(ggml_backend_vk_context * ctx, vk_context& subctx, vk_subbuffer&& in, vk_subbuffer&& out, uint32_t ne, bool use_x4_blocks = false) {
VK_LOG_DEBUG("ggml_vk_quantize_q8_1(" << "buffer in size=" << in.buffer->size << ", buffer out size=" << out.buffer->size << ", " << ne << ")");
vk_pipeline pipeline = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1);
vk_pipeline pipeline = use_x4_blocks ? ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1, true) : ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1, false);
ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, std::array<uint32_t, 1>{ne}, { ne, 1, 1 });
ggml_vk_sync_buffers(ctx, subctx);
@@ -5720,12 +5799,15 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr); // NOLINT
if (quantize_y) {
to_q8_1 = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1);
to_q8_1 = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1, true);
}
if (dryrun) {
const uint64_t x_sz_upd = x_sz * ne02 * ne03;
const uint64_t y_sz_upd = y_sz * ne12 * ne13;
uint64_t y_sz_upd = y_sz * ne12 * ne13;
if (quantize_y) {
y_sz_upd = CEIL_DIV(y_sz_upd, 144) * 144;
}
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) ||
@@ -5791,7 +5873,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
GGML_ASSERT(d_Y->size >= y_sz * ne12 * ne13);
} else if (quantize_y) {
d_Y = ctx->prealloc_y;
GGML_ASSERT(d_Y->size >= y_ne * ggml_type_size(GGML_TYPE_Q8_1) / ggml_blck_size(GGML_TYPE_Q8_1));
GGML_ASSERT(d_Y->size >= CEIL_DIV(y_sz * ne12 * ne13, 144) * 144);
} else {
d_Y = d_Qy;
y_buf_offset = qy_buf_offset;
@@ -5828,7 +5910,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
if (ctx->prealloc_y_need_sync) {
ggml_vk_sync_buffers(ctx, subctx);
}
ggml_vk_quantize_q8_1(ctx, subctx, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE }, y_ne * ne12 * ne13);
ggml_vk_quantize_q8_1(ctx, subctx, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE }, y_ne * ne12 * ne13, true);
ctx->prealloc_y_last_pipeline_used = to_q8_1.get();
ctx->prealloc_y_last_tensor_used = src1;
}
@@ -5845,10 +5927,15 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
stride_batch_y = src1->nb[0] / ggml_type_size(src1->type);
}
uint32_t y_sz_total = y_sz * ne12 * ne13;
if (quantize_y) {
y_sz_total = CEIL_DIV(y_sz_total, 144) * 144;
}
// compute
ggml_vk_matmul(
ctx, subctx, pipeline,
{ d_X, x_buf_offset, x_sz * ne02 * ne03 }, { d_Y, y_buf_offset, y_sz * ne12 * ne13 },
{ d_X, x_buf_offset, x_sz * ne02 * ne03 }, { d_Y, y_buf_offset, y_sz_total },
{ 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,
@@ -5863,6 +5950,51 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
}
}
// Device tuning
static bool ggml_vk_should_use_mmvq(const vk_device& device, uint32_t m, uint32_t n, uint32_t k, ggml_type src0_type) {
if (device->mmvq_mode == 1) {
return true;
} else if (device->mmvq_mode == -1) {
return false;
}
// MMVQ is generally good for batches
if (n > 1) {
return true;
}
switch (device->vendor_id) {
case VK_VENDOR_ID_NVIDIA:
switch (src0_type) {
case GGML_TYPE_Q8_0:
return device->architecture == vk_device_architecture::NVIDIA_PRE_TURING;
default:
return true;
}
case VK_VENDOR_ID_AMD:
switch (src0_type) {
case GGML_TYPE_Q8_0:
return device->architecture == vk_device_architecture::AMD_GCN;
default:
return true;
}
case VK_VENDOR_ID_INTEL:
switch (src0_type) {
// From tests on A770 Linux, may need more tuning
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q5_1:
return false;
default:
return true;
}
default:
return true;
}
GGML_UNUSED(m);
GGML_UNUSED(k);
}
static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
VK_LOG_DEBUG("ggml_vk_mul_mat_vec_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
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];
@@ -5917,22 +6049,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
const bool y_non_contig = !ggml_vk_dim01_contiguous(src1);
const bool f16_f32_kernel = src1->type == GGML_TYPE_F32;
const bool qx_needs_dequant = x_non_contig;
const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !f16_f32_kernel) || y_non_contig;
// 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 = ne11 * ne01;
const uint64_t qx_sz = ggml_vk_align_size(ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type), ctx->device->properties.limits.minStorageBufferOffsetAlignment);
const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type);
const uint64_t x_sz = x_non_contig ? ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment) : qx_sz;
const uint64_t y_sz = f16_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne;
const uint64_t d_sz = sizeof(float) * d_ne;
bool quantize_y = ctx->device->integer_dot_product && src1->type == GGML_TYPE_F32 && ggml_is_contiguous(src1) && (ne11 * ne10) % 4 == 0 && ggml_vk_should_use_mmvq(ctx->device, ne01, ne11, ne10, src0->type);
vk_pipeline to_fp16_vk_0 = nullptr;
vk_pipeline to_fp16_vk_1 = nullptr;
@@ -5944,14 +6061,47 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
} else {
to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type);
}
vk_pipeline dmmv = ggml_vk_get_dequantize_mul_mat_vec(ctx, src0->type, src1->type, ne11, ne20, ne00);
// Check for mmq first
vk_pipeline dmmv = quantize_y ? ggml_vk_get_dequantize_mul_mat_vec(ctx, src0->type, GGML_TYPE_Q8_1, ne11, ne20, ne00) : nullptr;
vk_pipeline to_q8_1 = nullptr;
if (dmmv == nullptr) {
// Fall back to f16 dequant mul mat
dmmv = ggml_vk_get_dequantize_mul_mat_vec(ctx, src0->type, src1->type, ne11, ne20, ne00);
quantize_y = false;
}
if (quantize_y) {
to_q8_1 = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1, true);
}
const bool qx_needs_dequant = x_non_contig;
const bool qy_needs_dequant = !quantize_y && ((src1->type != GGML_TYPE_F16 && !f16_f32_kernel) || y_non_contig);
// Not implemented
GGML_ASSERT(y_non_contig || !qy_needs_dequant); // NOLINT
GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr); // NOLINT
GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr); // NOLINT
GGML_ASSERT(dmmv != nullptr);
const uint64_t x_ne = ne01 * ne00;
const uint64_t y_ne = ne11 * ne10;
const uint64_t d_ne = ne11 * ne01;
const uint64_t qx_sz = ggml_vk_align_size(ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type), ctx->device->properties.limits.minStorageBufferOffsetAlignment);
const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type);
const uint64_t x_sz = x_non_contig ? ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment) : qx_sz;
const uint64_t y_sz = quantize_y ? (y_ne * ggml_type_size(GGML_TYPE_Q8_1) / ggml_blck_size(GGML_TYPE_Q8_1)) : (f16_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne);
const uint64_t d_sz = sizeof(float) * d_ne;
if (dryrun) {
const uint64_t x_sz_upd = x_sz * ne02 * ne03;
const uint64_t y_sz_upd = y_sz * ne12 * ne13;
uint64_t y_sz_upd = y_sz * ne12 * ne13;
if (quantize_y) {
y_sz_upd = CEIL_DIV(y_sz_upd, 144) * 144;
}
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)) {
@@ -5960,7 +6110,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
if (qx_needs_dequant && ctx->prealloc_size_x < x_sz_upd) {
ctx->prealloc_size_x = x_sz_upd;
}
if (qy_needs_dequant && ctx->prealloc_size_y < y_sz_upd) {
if ((qy_needs_dequant || quantize_y) && ctx->prealloc_size_y < y_sz_upd) {
ctx->prealloc_size_y = y_sz_upd;
}
@@ -5971,6 +6121,9 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
if (qy_needs_dequant) {
ggml_pipeline_request_descriptor_sets(ctx, to_fp16_vk_1, 1);
}
if (quantize_y) {
ggml_pipeline_request_descriptor_sets(ctx, to_q8_1, 1);
}
ggml_pipeline_request_descriptor_sets(ctx, dmmv, 1);
return;
}
@@ -6001,6 +6154,9 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
}
if (qy_needs_dequant) {
d_Y = ctx->prealloc_y;
} else if (quantize_y) {
d_Y = ctx->prealloc_y;
GGML_ASSERT(d_Y->size >= CEIL_DIV(y_sz * ne12 * ne13, 144) * 144);
} else {
d_Y = d_Qy;
y_buf_offset = qy_buf_offset;
@@ -6011,9 +6167,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
if (ctx->prealloc_x_need_sync) {
ggml_vk_sync_buffers(ctx, subctx);
}
}
if (x_non_contig) {
GGML_ASSERT(x_sz == ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment));
ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, { d_Qx, qx_buf_offset, VK_WHOLE_SIZE }, { d_X, 0, VK_WHOLE_SIZE });
}
@@ -6029,6 +6183,17 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
ctx->prealloc_y_last_tensor_used = src1;
}
}
if (quantize_y) {
if (ctx->prealloc_y_last_pipeline_used != to_q8_1.get() ||
ctx->prealloc_y_last_tensor_used != src1) {
if (ctx->prealloc_y_need_sync) {
ggml_vk_sync_buffers(ctx, subctx);
}
ggml_vk_quantize_q8_1(ctx, subctx, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE }, y_ne * ne12 * ne13, true);
ctx->prealloc_y_last_pipeline_used = to_q8_1.get();
ctx->prealloc_y_last_tensor_used = src1;
}
}
// For batch_n, the A matrix is the same for each batch, and B/D use the row stride as the batch stride
uint32_t stride_batch_x = batch_n ? 0 : ne00*ne01;
@@ -6053,6 +6218,12 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
groups_x = CEIL_DIV(groups_x, groups_z);
}
// TODO: Clean up this whole sz * ne_2 * ne_3 thing, it hasn't been necessary for a long time
uint32_t y_sz_total = y_sz * ne12 * ne13;
if (quantize_y) {
y_sz_total = CEIL_DIV(y_sz_total, 144) * 144;
}
// compute
const vk_mat_vec_push_constants pc = {
(uint32_t)ne00, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne01,
@@ -6060,13 +6231,13 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
(uint32_t)ne02, (uint32_t)ne12, (uint32_t)r2, (uint32_t)r3,
};
ggml_vk_dispatch_pipeline(ctx, subctx, dmmv,
{ vk_subbuffer{ d_X, x_buf_offset, x_sz * ne02 * ne03 }, vk_subbuffer{ d_Y, y_buf_offset, y_sz * ne12 * ne13 }, vk_subbuffer{ d_D, d_buf_offset, d_sz * ne22 * ne23} },
{ vk_subbuffer{ d_X, x_buf_offset, x_sz * ne02 * ne03 }, vk_subbuffer{ d_Y, y_buf_offset, y_sz_total }, vk_subbuffer{ d_D, d_buf_offset, d_sz * ne22 * ne23} },
pc, { groups_x, (uint32_t)(ne12 * ne13), groups_z });
if (x_non_contig) {
ctx->prealloc_x_need_sync = true;
}
if (y_non_contig) {
if (y_non_contig || quantize_y) {
ctx->prealloc_y_need_sync = true;
}
}

30
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp
View File

@@ -1,7 +1,8 @@
#extension GL_EXT_control_flow_attributes : enable
#extension GL_EXT_shader_16bit_storage : require
#extension GL_EXT_shader_8bit_storage : require
#if USE_SUBGROUP_ADD
#if USE_SUBGROUP_ADD || USE_SUBGROUP_ADD_NO_SHMEM
#extension GL_KHR_shader_subgroup_basic : require
#extension GL_KHR_shader_subgroup_arithmetic : require
#endif
@@ -12,10 +13,19 @@
#include "types.comp"
#ifndef MMQ
layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
#else
layout (binding = 0) readonly buffer A {A_TYPE_PACKED16 data_a[];};
#endif
layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
#ifdef B_TYPE_VEC2
layout (binding = 1) readonly buffer BV2 {B_TYPE_VEC2 data_b_v2[];};
#endif
#ifdef B_TYPE_VEC4
layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];};
#endif
layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
#ifdef MUL_MAT_ID
@@ -92,6 +102,23 @@ layout (constant_id = 0) const uint BLOCK_SIZE = 32;
layout (constant_id = 1) const uint NUM_ROWS = 1;
layout (constant_id = 2) const uint NUM_COLS = 1;
#ifdef USE_SUBGROUP_ADD_NO_SHMEM
void reduce_result(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offset, const in uint32_t first_row, const in uint32_t num_rows, const in uint32_t tid) {
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
temp[j][n] = subgroupAdd(temp[j][n]);
}
}
if (tid == 0) {
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
data_d[j*p.batch_stride_d + d_offset + first_row + n] = D_TYPE(temp[j][n]);
}
}
}
}
#else
shared FLOAT_TYPE tmpsh[NUM_COLS][NUM_ROWS][BLOCK_SIZE];
void reduce_result(FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offset, const in uint32_t first_row, const in uint32_t num_rows, const in uint32_t tid) {
@@ -152,3 +179,4 @@ void reduce_result(FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offs
}
#endif
}
#endif

140
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq.comp Normal file
View File

@@ -0,0 +1,140 @@
#version 450
#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
#extension GL_EXT_integer_dot_product : require
#define MMQ
#define B_TYPE block_q8_1_x4
#include "mul_mat_vec_base.comp"
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
#define K_PER_ITER 8
#include "mul_mmq_funcs.comp"
uint a_offset, b_offset, d_offset;
int32_t cache_b_qs[2];
vec2 cache_b_ds;
void iter(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const uint first_row, const uint num_rows, const uint tid, const uint i) {
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
const uint col = i*BLOCK_SIZE + tid*K_PER_ITER;
// Preload data_b block
const uint b_block_idx = (j*p.batch_stride_b + col) / QUANT_K_Q8_1 + b_offset;
const uint b_qs_idx = tid % 4;
const uint b_block_idx_outer = b_block_idx / 4;
const uint b_block_idx_inner = b_block_idx % 4;
cache_b_ds = vec2(data_b[b_block_idx_outer].ds[b_block_idx_inner]);
#if QUANT_R == 2
cache_b_qs[0] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx];
cache_b_qs[1] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx + 4];
#else
cache_b_qs[0] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx * 2];
cache_b_qs[1] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx * 2 + 1];
#endif
uint ibi = first_row*p.ncols;
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
const uint a_block_idx = (ibi + col)/QUANT_K + a_offset;
ibi += p.ncols;
int32_t q_sum = 0;
#if QUANT_R == 2
const i32vec2 data_a_qs = repack(a_block_idx, b_qs_idx);
q_sum += dotPacked4x8EXT(data_a_qs.x,
cache_b_qs[0]);
q_sum += dotPacked4x8EXT(data_a_qs.y,
cache_b_qs[1]);
#else
int32_t data_a_qs = repack(a_block_idx, b_qs_idx * 2);
q_sum += dotPacked4x8EXT(data_a_qs,
cache_b_qs[0]);
data_a_qs = repack(a_block_idx, b_qs_idx * 2 + 1);
q_sum += dotPacked4x8EXT(data_a_qs,
cache_b_qs[1]);
#endif
#if QUANT_AUXF == 1
temp[j][n] += mul_q8_1(q_sum, get_d(a_block_idx), cache_b_ds, 4);
#else
temp[j][n] += mul_q8_1(q_sum, get_dm(a_block_idx), cache_b_ds, 4);
#endif
}
}
}
void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
const uint tid = gl_LocalInvocationID.x;
get_offsets(a_offset, b_offset, d_offset);
a_offset /= QUANT_K;
b_offset /= QUANT_K_Q8_1;
FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
temp[j][n] = FLOAT_TYPE(0.0f);
}
}
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*K_PER_ITER);
i++;
}
}
unroll_count = 2;
unrolled_iters = num_iters & ~(unroll_count - 1);
#if K_PER_ITER == 2
if ((p.ncols & 1) != 0 &&
unrolled_iters == num_iters &&
unrolled_iters > 0) {
unrolled_iters -= unroll_count;
}
#endif
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);
i++;
}
}
while (i < num_iters) {
iter(temp, first_row, num_rows, tid, i*K_PER_ITER);
i++;
}
reduce_result(temp, d_offset, first_row, num_rows, tid);
}
void main() {
const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
// 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);
} else {
if (first_row >= p.stride_d) {
return;
}
compute_outputs(first_row, p.stride_d - first_row);
}
}

17
ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp
View File

@@ -28,7 +28,7 @@ layout (binding = 0) readonly buffer A {A_TYPE_PACKED16 data_a[];};
#if defined(A_TYPE_PACKED32)
layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32[];};
#endif
layout (binding = 1) readonly buffer B {block_q8_1_packed32 data_b[];};
layout (binding = 1) readonly buffer B {block_q8_1_x4_packed128 data_b[];};
layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
#ifdef MUL_MAT_ID
@@ -98,7 +98,7 @@ shared FLOAT_TYPE_VEC2 buf_b_ds[BN];
#endif
#define LOAD_VEC_A (4 * QUANT_R)
#define LOAD_VEC_B 4
#define LOAD_VEC_B 16
#ifdef MUL_MAT_ID
shared u16vec2 row_ids[4096];
@@ -270,15 +270,22 @@ void main() {
const uint iqs = idx & 0x7;
#else
const uint ib = pos_b_ib + (loadc_b + l) * p.stride_b / BK;
const uint ib_outer = ib / 4;
const uint ib_inner = ib % 4;
const uint iqs = loadr_b;
#endif
const uint buf_ib = loadc_b + l;
if (iqs == 0) {
buf_b_ds[buf_ib] = FLOAT_TYPE_VEC2(data_b[ib].ds);
buf_b_ds[buf_ib] = FLOAT_TYPE_VEC2(data_b[ib_outer].ds[ib_inner]);
}
buf_b_qs[buf_ib * SHMEM_STRIDE + iqs] = data_b[ib].qs[iqs];
const ivec4 values = data_b[ib_outer].qs[ib_inner * 2 + iqs];
buf_b_qs[buf_ib * SHMEM_STRIDE + iqs * 4 ] = values.x;
buf_b_qs[buf_ib * SHMEM_STRIDE + iqs * 4 + 1] = values.y;
buf_b_qs[buf_ib * SHMEM_STRIDE + iqs * 4 + 2] = values.z;
buf_b_qs[buf_ib * SHMEM_STRIDE + iqs * 4 + 3] = values.w;
}
barrier();
@@ -349,7 +356,7 @@ void main() {
cache_b_qs[cc * (BK / 4) + idx_k]);
}
sums[sums_idx] += mul_q8_1(q_sum, cache_a_dm[cache_a_idx], cache_b_ds[cc]);
sums[sums_idx] += mul_q8_1(q_sum, cache_a_dm[cache_a_idx], cache_b_ds[cc], 1);
}
}
}

18
ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.comp
View File

@@ -16,8 +16,8 @@ i32vec2 repack(uint ib, uint iqs) {
(vui >> 4) & 0x0F0F0F0F);
}
ACC_TYPE mul_q8_1(int32_t q_sum, float da, vec2 dsb) {
return ACC_TYPE(da * (float(q_sum) * dsb.x - 8.0f * dsb.y));
ACC_TYPE mul_q8_1(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
return ACC_TYPE(da * (float(q_sum) * dsb.x - (8 / sum_divisor) * dsb.y));
}
#endif
@@ -29,8 +29,8 @@ i32vec2 repack(uint ib, uint iqs) {
(vui >> 4) & 0x0F0F0F0F);
}
ACC_TYPE mul_q8_1(int32_t q_sum, vec2 dma, vec2 dsb) {
return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y);
ACC_TYPE mul_q8_1(const int32_t q_sum, const vec2 dma, const vec2 dsb, const int32_t sum_divisor) {
return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y / sum_divisor);
}
#endif
@@ -50,8 +50,8 @@ i32vec2 repack(uint ib, uint iqs) {
return i32vec2(v0, v1);
}
ACC_TYPE mul_q8_1(int32_t q_sum, float da, vec2 dsb) {
return ACC_TYPE(da * (float(q_sum) * dsb.x - 16.0f * dsb.y));
ACC_TYPE mul_q8_1(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
return ACC_TYPE(da * (float(q_sum) * dsb.x - (16 / sum_divisor) * dsb.y));
}
#endif
@@ -69,8 +69,8 @@ i32vec2 repack(uint ib, uint iqs) {
return i32vec2(v0, v1);
}
ACC_TYPE mul_q8_1(int32_t q_sum, vec2 dma, vec2 dsb) {
return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y);
ACC_TYPE mul_q8_1(const int32_t q_sum, const vec2 dma, const vec2 dsb, const int32_t sum_divisor) {
return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y / sum_divisor);
}
#endif
@@ -81,7 +81,7 @@ int32_t repack(uint ib, uint iqs) {
data_a[ib].qs[iqs * 2 + 1]));
}
ACC_TYPE mul_q8_1(int32_t q_sum, float da, vec2 dsb) {
ACC_TYPE mul_q8_1(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
return ACC_TYPE(float(q_sum) * da * dsb.x);
}
#endif

56
ggml/src/ggml-vulkan/vulkan-shaders/quantize_q8_1.comp
View File

@@ -3,6 +3,15 @@
#extension GL_EXT_control_flow_attributes : require
#extension GL_EXT_shader_16bit_storage : require
#ifdef USE_SUBGROUPS
#extension GL_KHR_shader_subgroup_basic : require
#extension GL_KHR_shader_subgroup_clustered : require
#define INVOCATION_ID gl_SubgroupInvocationID.x
#else
#define INVOCATION_ID gl_LocalInvocationID.x
#endif
layout (push_constant) uniform parameter
{
uint ne;
@@ -14,13 +23,19 @@ layout(constant_id = 0) const uint GROUP_SIZE = 32;
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {vec4 data_a[];};
#ifndef QBLOCK_X4
layout (binding = 1) writeonly buffer D {block_q8_1_packed32 data_b[];};
#else
layout (binding = 1) writeonly buffer D {block_q8_1_x4 data_b[];};
#endif
#ifndef USE_SUBGROUPS
shared float shmem[GROUP_SIZE];
#endif
void quantize() {
const uint wgid = gl_WorkGroupID.x;
const uint tid = gl_LocalInvocationID.x;
const uint tid = INVOCATION_ID;
// Each thread handles a vec4, so 8 threads handle a block
const uint blocks_per_group = GROUP_SIZE / 8;
@@ -30,9 +45,19 @@ void quantize() {
const uint ib = wgid * blocks_per_group + block_in_wg;
const uint iqs = tid % 8;
#ifndef QBLOCK_X4
if (ib >= gl_NumWorkGroups.x * blocks_per_group) {
return;
}
#else
const uint ibx4_outer = ib / 4;
const uint ibx4_inner = ib % 4;
const uint required_x4_blocks = (p.ne + 127) / 128;
if (ibx4_outer >= required_x4_blocks) {
return;
}
#endif
const uint a_idx = ib * 8 + iqs;
@@ -40,7 +65,9 @@ void quantize() {
const vec4 abs_vals = abs(vals);
// Find absolute max for each block
shmem[tid] = max(max(abs_vals.x, abs_vals.y), max(abs_vals.z, abs_vals.w));
const float thread_max = max(max(abs_vals.x, abs_vals.y), max(abs_vals.z, abs_vals.w));
#ifndef USE_SUBGROUPS
shmem[tid] = thread_max;
barrier();
[[unroll]] for (uint s = 4; s > 0; s >>= 1) {
if (iqs < s) {
@@ -50,14 +77,28 @@ void quantize() {
}
const float amax = shmem[block_in_wg * 8];
#else
const float amax = subgroupClusteredMax(thread_max, 8);
#endif
const float d = amax / 127.0;
const float d_inv = d != 0.0 ? 1.0 / d : 0.0;
vals = round(vals * d_inv);
#ifndef QBLOCK_X4
data_b[ib].qs[iqs] = pack32(i8vec4(round(vals)));
#else
data_b[ibx4_outer].qs[ibx4_inner * 8 + iqs] = pack32(i8vec4(round(vals)));
#endif
#ifndef USE_SUBGROUPS
barrier();
#endif
// Calculate the sum for each block
shmem[tid] = vals.x + vals.y + vals.z + vals.w;
const float thread_sum = vals.x + vals.y + vals.z + vals.w;
#ifndef USE_SUBGROUPS
shmem[tid] = thread_sum;
barrier();
[[unroll]] for (uint s = 4; s > 0; s >>= 1) {
if (iqs < s) {
@@ -65,10 +106,19 @@ void quantize() {
}
barrier();
}
#else
const float sum = subgroupClusteredAdd(thread_sum, 8);
#endif
if (iqs == 0) {
#ifndef USE_SUBGROUPS
const float sum = shmem[tid];
#endif
#ifndef QBLOCK_X4
data_b[ib].ds = f16vec2(vec2(d, sum * d));
#else
data_b[ibx4_outer].ds[ibx4_inner] = f16vec2(vec2(d, sum * d));
#endif
}
}

12
ggml/src/ggml-vulkan/vulkan-shaders/types.comp
View File

@@ -207,6 +207,18 @@ struct block_q8_1_packed32
int32_t qs[8];
};
// 4 blocks in one to allow 16-byte/128-bit alignment and loads
struct block_q8_1_x4
{
f16vec2 ds[4];
int32_t qs[32];
};
struct block_q8_1_x4_packed128
{
f16vec2 ds[4];
ivec4 qs[8];
};
// K-quants
#define QUANT_K_Q2_K 256

48
ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
View File

@@ -206,6 +206,22 @@ bool string_ends_with(const std::string& str, const std::string& suffix) {
return std::equal(suffix.rbegin(), suffix.rend(), str.rbegin());
}
bool is_quantized_type(const std::string& type_name) {
return type_name != "f32" && type_name != "f16" && type_name != "bf16";
}
bool is_legacy_quant(const std::string& type_name) {
return type_name == "q4_0" || type_name == "q4_1" || type_name == "q5_0" || type_name == "q5_1" || type_name == "q8_0";
}
bool is_k_quant(const std::string& type_name) {
return string_ends_with(type_name, "_k");
}
bool is_iq_quant(const std::string& type_name) {
return string_starts_with(type_name, "iq");
}
static const char path_separator = '/';
std::string join_paths(const std::string& path1, const std::string& path2) {
@@ -402,7 +418,7 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
}
#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
if (!coopmat && !coopmat2 && matmul_id_type == MatMulIdType::NONE && (tname == "q4_0" || tname == "q4_1" || tname == "q5_0" || tname == "q5_1" || tname == "q8_0")) {
if (!coopmat && !coopmat2 && matmul_id_type == MatMulIdType::NONE && is_legacy_quant(tname)) {
string_to_spv(shader_name + "_" + tname + "_q8_1", "mul_mmq.comp", merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"D_TYPE", "float"},}), fp16, coopmat, coopmat2, f16acc);
}
#endif
@@ -495,8 +511,20 @@ void process_shaders() {
string_to_spv("mul_mat_vec_" + tname + "_f32_f32_subgroup", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
string_to_spv("mul_mat_vec_" + tname + "_f16_f32_subgroup", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPE_VEC2", "f16vec2"}, {"B_TYPE_VEC4", "f16vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
string_to_spv("mul_mat_vec_" + tname + "_f32_f32_subgroup_no_shmem", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
string_to_spv("mul_mat_vec_" + tname + "_f16_f32_subgroup_no_shmem", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPE_VEC2", "f16vec2"}, {"B_TYPE_VEC4", "f16vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
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"}}));
// mul mat vec with integer dot product
#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
if (is_legacy_quant(tname)) {
string_to_spv("mul_mat_vec_" + tname + "_q8_1_f32", "mul_mat_vecq.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}, {"ACC_TYPE", "float"}}));
string_to_spv("mul_mat_vec_" + tname + "_q8_1_f32_subgroup", "mul_mat_vecq.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}, {"ACC_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
string_to_spv("mul_mat_vec_" + tname + "_q8_1_f32_subgroup_no_shmem", "mul_mat_vecq.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}, {"ACC_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
}
#endif
// Dequant shaders
if (tname != "f16" && tname != "bf16") {
string_to_spv("dequant_" + tname, "dequant_" + tname + ".comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float16_t"}}));
@@ -579,7 +607,12 @@ void process_shaders() {
string_to_spv("split_k_reduce", "mul_mat_split_k_reduce.comp", {});
string_to_spv("fa_split_k_reduce", "flash_attn_split_k_reduce.comp", {});
string_to_spv("quantize_q8_1", "quantize_q8_1.comp", {});
string_to_spv("quantize_q8_1_subgroup", "quantize_q8_1.comp", {{"USE_SUBGROUPS", "1"}});
string_to_spv("quantize_q8_1_x4", "quantize_q8_1.comp", {{"QBLOCK_X4", "1"}});
string_to_spv("quantize_q8_1_x4_subgroup", "quantize_q8_1.comp", {{"QBLOCK_X4", "1"}, {"USE_SUBGROUPS", "1"}});
string_to_spv("mul_f32", "mul.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
@@ -821,12 +854,15 @@ void write_output_files() {
fputs(len.c_str(), src);
}
for (const std::string& btype : {"f16", "f32"}) {
for (const std::string& btype : {"f16", "f32", "q8_1"}) {
for (const auto& tname : type_names) {
fprintf(hdr, "extern unsigned char *arr_dmmv_%s_%s_f32_data[2];\n", tname.c_str(), btype.c_str());
fprintf(hdr, "extern uint64_t arr_dmmv_%s_%s_f32_len[2];\n", tname.c_str(), btype.c_str());
std::string data = "unsigned char *arr_dmmv_" + tname + "_" + btype + "_f32_data[2] = {mul_mat_vec_" + tname + "_" + btype + "_f32_data, mul_mat_vec_" + tname + "_" + btype + "_f32_subgroup_data};\n";
std::string len = "uint64_t arr_dmmv_" + tname + "_" + btype + "_f32_len[2] = {mul_mat_vec_" + tname + "_" + btype + "_f32_len, mul_mat_vec_" + tname + "_" + btype + "_f32_subgroup_len};\n";
if (btype == "q8_1" && !is_legacy_quant(tname)) {
continue;
}
fprintf(hdr, "extern unsigned char *arr_dmmv_%s_%s_f32_data[3];\n", tname.c_str(), btype.c_str());
fprintf(hdr, "extern uint64_t arr_dmmv_%s_%s_f32_len[3];\n", tname.c_str(), btype.c_str());
std::string data = "unsigned char *arr_dmmv_" + tname + "_" + btype + "_f32_data[3] = {mul_mat_vec_" + tname + "_" + btype + "_f32_data, mul_mat_vec_" + tname + "_" + btype + "_f32_subgroup_data, mul_mat_vec_" + tname + "_" + btype + "_f32_subgroup_no_shmem_data};\n";
std::string len = "uint64_t arr_dmmv_" + tname + "_" + btype + "_f32_len[3] = {mul_mat_vec_" + tname + "_" + btype + "_f32_len, mul_mat_vec_" + tname + "_" + btype + "_f32_subgroup_len, mul_mat_vec_" + tname + "_" + btype + "_f32_subgroup_no_shmem_len};\n";
fputs(data.c_str(), src);
fputs(len.c_str(), src);
}