happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

WIP to update subgroup size per kernel

This commit is contained in:
Nakasaka, Masato 2025-12-04 00:38:25 -08:00
parent 710f84862f
commit 36b976f3d4
1 changed files with 52 additions and 58 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

110
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View File

@ -2720,24 +2720,18 @@ static void ggml_vk_load_shaders(vk_device& device) {
VK_LOG_DEBUG("ggml_vk_load_shaders(" << device->name << ")"); VK_LOG_DEBUG("ggml_vk_load_shaders(" << device->name << ")");
std::lock_guard<std::recursive_mutex> guard(device->mutex); std::lock_guard<std::recursive_mutex> guard(device->mutex);
// For Intel we want to use a non-default subgroup_size as the base block size
// Other architectures will use the default subgroup size as the base block size
const uint32_t target_subgroup_size = (device->vendor_id == VK_VENDOR_ID_INTEL) ? get_subgroup_size("", device->architecture) : device->subgroup_size;
// We need to force a certain subgroup size for some kernels
// since the kernels expect the block sizes to match or are multiples of subgroup size
const uint32_t forced_subgroup_size = (device->vendor_id == VK_VENDOR_ID_INTEL) ? target_subgroup_size : 0;
// some shaders have a minimum subgroup size // some shaders have a minimum subgroup size
const uint32_t subgroup_size_8 = std::max(target_subgroup_size, 8u); const uint32_t subgroup_size_8 = std::max(device->subgroup_size, 8u);
const uint32_t subgroup_size_16 = std::max(target_subgroup_size, 16u); const uint32_t subgroup_size_16 = std::max(device->subgroup_size, 16u);
const uint32_t subgroup_size_32 = std::max(target_subgroup_size, 32u); const uint32_t subgroup_size_32 = std::max(device->subgroup_size, 32u);
const uint32_t mul_mat_subgroup_size = (device->vendor_id == VK_VENDOR_ID_INTEL && device->subgroup_size_control) ? device->subgroup_min_size : target_subgroup_size; const uint32_t mul_mat_subgroup_size = (device->vendor_id == VK_VENDOR_ID_INTEL && device->subgroup_size_control) ? device->subgroup_min_size : device->subgroup_size;
const uint32_t mul_mat_subgroup_size_8 = std::max(mul_mat_subgroup_size, 8u); const uint32_t mul_mat_subgroup_size_8 = std::max(mul_mat_subgroup_size, 8u);
const uint32_t mul_mat_subgroup_size_16 = std::max(mul_mat_subgroup_size, 16u); const uint32_t mul_mat_subgroup_size_16 = std::max(mul_mat_subgroup_size, 16u);
const uint32_t mul_mat_subgroup_size_32 = std::max(mul_mat_subgroup_size, 32u); const uint32_t mul_mat_subgroup_size_32 = std::max(mul_mat_subgroup_size, 32u);
const bool subgroup_min_size_16 = (!device->subgroup_size_control && target_subgroup_size >= 16) || const bool subgroup_min_size_16 = (!device->subgroup_size_control && device->subgroup_size >= 16) ||
(device->subgroup_size_control && device->subgroup_max_size >= 16); (device->subgroup_size_control && device->subgroup_max_size >= 16);
// mulmat // mulmat
@ -2782,9 +2776,9 @@ static void ggml_vk_load_shaders(vk_device& device) {
s_mmq_wg_denoms_k = { 32, 64, 1 }; s_mmq_wg_denoms_k = { 32, 64, 1 };
// spec constants and tile sizes for quant matmul_id // spec constants and tile sizes for quant matmul_id
l_warptile_mmqid = { 256, 128, 128, 16, 1, target_subgroup_size }; l_warptile_mmqid = { 256, 128, 128, 16, 1, device->subgroup_size };
m_warptile_mmqid = { 256, 128, 64, 16, 0, target_subgroup_size }; m_warptile_mmqid = { 256, 128, 64, 16, 0, device->subgroup_size };
s_warptile_mmqid = { 256, 128, 64, 16, 0, target_subgroup_size }; s_warptile_mmqid = { 256, 128, 64, 16, 0, device->subgroup_size };
l_mmqid_wg_denoms = { 128, 128, 1 }; l_mmqid_wg_denoms = { 128, 128, 1 };
m_mmqid_wg_denoms = { 128, 64, 1 }; m_mmqid_wg_denoms = { 128, 64, 1 };
s_mmqid_wg_denoms = { 128, 64, 1 }; s_mmqid_wg_denoms = { 128, 64, 1 };
@ -2901,7 +2895,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, const std::vector<uint32_t>& specialization_constants, uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, const std::vector<uint32_t>& specialization_constants,
uint32_t align, bool disable_robustness = false, bool require_full_subgroups = false, uint32_t required_subgroup_size = 0) { uint32_t align, bool disable_robustness = false, bool require_full_subgroups = false, uint32_t required_subgroup_size = 0) {
if (!require_full_subgroups && required_subgroup_size == 0) { if (required_subgroup_size == 0) {
required_subgroup_size = get_subgroup_size(name, device->architecture); required_subgroup_size = get_subgroup_size(name, device->architecture);
} }
@ -2963,7 +2957,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
// D_split can't be larger than a subgroup because we use subgroupShuffle to reduce it. // D_split can't be larger than a subgroup because we use subgroupShuffle to reduce it.
// D_split can't be larger than the LSB of D divided by 4 due to vectorization in the shader. // D_split can't be larger than the LSB of D divided by 4 due to vectorization in the shader.
const uint32_t D_lsb = D ^ (D & (D-1)); const uint32_t D_lsb = D ^ (D & (D-1));
uint32_t D_split = std::min(std::min(target_subgroup_size, 8u), D_lsb / 4); uint32_t D_split = std::min(std::min(device->subgroup_size, 8u), D_lsb / 4);
return {wg_size, rows_cols[0], rows_cols[1], hsk, hsv, clamp, D_split}; return {wg_size, rows_cols[0], rows_cols[1], hsk, hsv, clamp, D_split};
}; };
@ -3100,17 +3094,17 @@ static void ggml_vk_load_shaders(vk_device& device) {
// Create 6 variants, {s,m,l}x{unaligned,aligned} // Create 6 variants, {s,m,l}x{unaligned,aligned}
#define CREATE_MM(TYPE, PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \ #define CREATE_MM(TYPE, PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
if (device->mul_mat ## ID ## _l[TYPE]) \ if (device->mul_mat ## ID ## _l[TYPE]) \
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, false, true, forced_subgroup_size); \ ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, false, true); \
if (device->mul_mat ## ID ## _m[TYPE]) \ if (device->mul_mat ## ID ## _m[TYPE]) \
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1, false, true, forced_subgroup_size); \ ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1, false, true); \
if (device->mul_mat ## ID ## _s[TYPE]) \ if (device->mul_mat ## ID ## _s[TYPE]) \
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1, false, true, forced_subgroup_size); \ ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1, false, true); \
if (device->mul_mat ## ID ## _l[TYPE]) \ if (device->mul_mat ## ID ## _l[TYPE]) \
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _cm1_len, NAMELC ## _aligned ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align, false, true, forced_subgroup_size); \ ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _cm1_len, NAMELC ## _aligned ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align, false, true); \
if (device->mul_mat ## ID ## _m[TYPE]) \ if (device->mul_mat ## ID ## _m[TYPE]) \
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _cm1_len, NAMELC ## _aligned ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align, false, true, forced_subgroup_size); \ ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _cm1_len, NAMELC ## _aligned ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align, false, true); \
if (device->mul_mat ## ID ## _s[TYPE]) \ if (device->mul_mat ## ID ## _s[TYPE]) \
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _cm1_len, NAMELC ## _aligned ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align, false, true, forced_subgroup_size); \ ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _cm1_len, NAMELC ## _aligned ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align, false, true); \
// Create 2 variants, {f16,f32} accumulator // Create 2 variants, {f16,f32} accumulator
#define CREATE_MM2(TYPE, PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \ #define CREATE_MM2(TYPE, PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
@ -3544,7 +3538,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
// Ensure a subgroup size >= 16 is available // Ensure a subgroup size >= 16 is available
const bool use_subgroups16 = use_subgroups && subgroup_min_size_16; const bool use_subgroups16 = use_subgroups && subgroup_min_size_16;
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 : target_subgroup_size; 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 subgroup_size16 = std::max(subgroup_size, 16u);
const uint32_t force_subgroup_size = use_subgroups ? subgroup_size : 0; const uint32_t force_subgroup_size = use_subgroups ? subgroup_size : 0;
@ -3615,7 +3609,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT) #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
if (device->integer_dot_product) { 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 : target_subgroup_size; 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); 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", mul_mat_vec_num_bindings, 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_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", mul_mat_vec_num_bindings, 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);
@ -3628,14 +3622,14 @@ static void ggml_vk_load_shaders(vk_device& device) {
} }
} }
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32", mul_mat_vec_id_f32_f32_len, mul_mat_vec_id_f32_f32_data, "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {target_subgroup_size, 2}, 1); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32", mul_mat_vec_id_f32_f32_len, mul_mat_vec_id_f32_f32_data, "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F16 ], "mul_mat_vec_id_f16_f32", mul_mat_vec_id_f16_f32_len, mul_mat_vec_id_f16_f32_data, "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {target_subgroup_size, 2}, 1); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F16 ], "mul_mat_vec_id_f16_f32", mul_mat_vec_id_f16_f32_len, mul_mat_vec_id_f16_f32_data, "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_BF16], "mul_mat_vec_id_bf16_f32", mul_mat_vec_id_bf16_f32_len, mul_mat_vec_id_bf16_f32_data, "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {target_subgroup_size, 2}, 1); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_BF16], "mul_mat_vec_id_bf16_f32", mul_mat_vec_id_bf16_f32_len, mul_mat_vec_id_bf16_f32_data, "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {target_subgroup_size, 2*rm_stdq}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {target_subgroup_size, 2*rm_stdq}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {target_subgroup_size, 2*rm_stdq}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32", mul_mat_vec_id_q5_1_f32_len, mul_mat_vec_id_q5_1_f32_data, "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {target_subgroup_size, 2*rm_stdq}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32", mul_mat_vec_id_q5_1_f32_len, mul_mat_vec_id_q5_1_f32_data, "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32", mul_mat_vec_id_q8_0_f32_len, mul_mat_vec_id_q8_0_f32_data, "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {1*rm_stdq, 1, 1}, {target_subgroup_size, 1*rm_stdq}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32", mul_mat_vec_id_q8_0_f32_len, mul_mat_vec_id_q8_0_f32_data, "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {1*rm_stdq, 1, 1}, {device->subgroup_size, 1*rm_stdq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32", mul_mat_vec_id_q2_k_f32_len, mul_mat_vec_id_q2_k_f32_data, "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32", mul_mat_vec_id_q2_k_f32_len, mul_mat_vec_id_q2_k_f32_data, "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32", mul_mat_vec_id_q3_k_f32_len, mul_mat_vec_id_q3_k_f32_data, "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32", mul_mat_vec_id_q3_k_f32_len, mul_mat_vec_id_q3_k_f32_data, "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", mul_mat_vec_id_q4_k_f32_len, mul_mat_vec_id_q4_k_f32_data, "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", mul_mat_vec_id_q4_k_f32_len, mul_mat_vec_id_q4_k_f32_data, "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
@ -3725,19 +3719,19 @@ static void ggml_vk_load_shaders(vk_device& device) {
ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_MXFP4], "get_rows_mxfp4_f32", get_rows_mxfp4_f32_len, get_rows_mxfp4_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_MXFP4], "get_rows_mxfp4_f32", get_rows_mxfp4_f32_len, get_rows_mxfp4_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
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_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, target_subgroup_size, 1}, {target_subgroup_size}, 1, true); 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);
if (device->subgroup_clustered && device->subgroup_require_full_support) { if (device->subgroup_clustered && device->subgroup_require_full_support) {
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 * target_subgroup_size / 8, 1, 1}, { target_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 { } else {
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 * target_subgroup_size / 8, 1, 1}, { target_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) { for (uint32_t i = 0; i < p021_max_gqa_ratio; ++i) {
if (device->subgroup_arithmetic && device->subgroup_require_full_support) { if (device->subgroup_arithmetic && device->subgroup_require_full_support) {
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", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_p021_push_constants), {1, 1, 1}, {target_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", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_p021_push_constants), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true, true);
} else { } else {
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", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_p021_push_constants), {1, 1, 1}, {target_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", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_p021_push_constants), {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", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_nc_push_constants), {1, 1, 1}, {}, 1); 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", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_nc_push_constants), {1, 1, 1}, {}, 1);
@ -3956,11 +3950,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
ggml_vk_create_pipeline(device, device->pipeline_diag_mask_inf_f32, "diag_mask_inf_f32", diag_mask_inf_f32_len, diag_mask_inf_f32_data, "main", 2, sizeof(vk_op_diag_mask_push_constants), {1, 512, 1}, {}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_diag_mask_inf_f32, "diag_mask_inf_f32", diag_mask_inf_f32_len, diag_mask_inf_f32_data, "main", 2, sizeof(vk_op_diag_mask_push_constants), {1, 512, 1}, {}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32, "soft_max_f32", soft_max_f32_len, soft_max_f32_data, "main", 4, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { target_subgroup_size }, 1); ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32, "soft_max_f32", soft_max_f32_len, soft_max_f32_data, "main", 4, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_wg512, "soft_max_f32_wg512", soft_max_f32_len, soft_max_f32_data, "main", 4, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 512 }, 1); ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_wg512, "soft_max_f32_wg512", soft_max_f32_len, soft_max_f32_data, "main", 4, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 512 }, 1);
ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16, "soft_max_f32_f16", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 4, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { target_subgroup_size }, 1); ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16, "soft_max_f32_f16", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 4, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16_wg512, "soft_max_f32_f16_wg512", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 4, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 512 }, 1); ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16_wg512, "soft_max_f32_f16_wg512", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 4, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 512 }, 1);
ggml_vk_create_pipeline(device, device->pipeline_soft_max_back_f32, "soft_max_back_f32", soft_max_back_f32_len, soft_max_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, { target_subgroup_size }, 1, true); ggml_vk_create_pipeline(device, device->pipeline_soft_max_back_f32, "soft_max_back_f32", soft_max_back_f32_len, soft_max_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f32, "rope_norm_f32", rope_norm_f32_len, rope_norm_f32_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f32, "rope_norm_f32", rope_norm_f32_len, rope_norm_f32_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f32, "rope_neox_f32", rope_neox_f32_len, rope_neox_f32_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f32, "rope_neox_f32", rope_neox_f32_len, rope_neox_f32_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
@ -3997,22 +3991,22 @@ static void ggml_vk_load_shaders(vk_device& device) {
ggml_vk_create_pipeline2(device, device->pipeline_argsort_large_f32[i], "argsort_large_f32_"+std::to_string(i), argsort_large_f32_len, argsort_large_f32_data, "main", 3, sizeof(vk_op_argsort_push_constants), {BLOCK_SIZE * WG_UNROLL_FACTOR, 1, 1}, {BLOCK_SIZE, WG_UNROLL_FACTOR}, 1, true); ggml_vk_create_pipeline2(device, device->pipeline_argsort_large_f32[i], "argsort_large_f32_"+std::to_string(i), argsort_large_f32_len, argsort_large_f32_data, "main", 3, sizeof(vk_op_argsort_push_constants), {BLOCK_SIZE * WG_UNROLL_FACTOR, 1, 1}, {BLOCK_SIZE, WG_UNROLL_FACTOR}, 1, true);
} }
ggml_vk_create_pipeline(device, device->pipeline_argmax_f32, "argmax_f32", argmax_f32_len, argmax_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, { target_subgroup_size }, 1); ggml_vk_create_pipeline(device, device->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);
ggml_vk_create_pipeline(device, device->pipeline_sum_rows_f32, "sum_rows_f32", sum_rows_f32_len, sum_rows_f32_data, "main", 2, sizeof(vk_op_sum_rows_push_constants), {1, 1, 1}, { target_subgroup_size }, 1); ggml_vk_create_pipeline(device, device->pipeline_sum_rows_f32, "sum_rows_f32", sum_rows_f32_len, sum_rows_f32_data, "main", 2, sizeof(vk_op_sum_rows_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
ggml_vk_create_pipeline(device, device->pipeline_cumsum_f32, "cumsum_f32", cumsum_f32_len, cumsum_f32_data, "main", 2, sizeof(vk_op_sum_rows_push_constants), {1, 1, 1}, { 128, target_subgroup_size }, 1, true, true, target_subgroup_size); ggml_vk_create_pipeline(device, device->pipeline_cumsum_f32, "cumsum_f32", cumsum_f32_len, cumsum_f32_data, "main", 2, sizeof(vk_op_sum_rows_push_constants), {1, 1, 1}, { 128, device->subgroup_size }, 1, true, true, device->subgroup_size);
ggml_vk_create_pipeline(device, device->pipeline_count_equal_i32, "count_equal_i32", count_equal_i32_len, count_equal_i32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, { target_subgroup_size }, 1); ggml_vk_create_pipeline(device, device->pipeline_count_equal_i32, "count_equal_i32", count_equal_i32_len, count_equal_i32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, { device->subgroup_size }, 1);
#define IM2COL(bda) \ #define IM2COL(bda) \
ggml_vk_create_pipeline(device, device->pipeline_im2col_f32, "im2col_f32", im2col_f32 ## bda ## _len, im2col_f32 ## bda ## _data, "main", 2, sizeof(vk_op_im2col_push_constants), {512, 1, 1}, { target_subgroup_size }, 1, true); \ ggml_vk_create_pipeline(device, device->pipeline_im2col_f32, "im2col_f32", im2col_f32 ## bda ## _len, im2col_f32 ## bda ## _data, "main", 2, sizeof(vk_op_im2col_push_constants), {512, 1, 1}, { device->subgroup_size }, 1, true); \
ggml_vk_create_pipeline(device, device->pipeline_im2col_3d_f32, "im2col_3d_f32", im2col_3d_f32 ## bda ## _len, im2col_3d_f32 ## bda ## _data, "main", 2, sizeof(vk_op_im2col_3d_push_constants), {512, 1, 1}, { 512 }, 1, true); \ ggml_vk_create_pipeline(device, device->pipeline_im2col_3d_f32, "im2col_3d_f32", im2col_3d_f32 ## bda ## _len, im2col_3d_f32 ## bda ## _data, "main", 2, sizeof(vk_op_im2col_3d_push_constants), {512, 1, 1}, { 512 }, 1, true); \
if (device->float_controls_rte_fp16) { \ if (device->float_controls_rte_fp16) { \
ggml_vk_create_pipeline(device, device->pipeline_im2col_f32_f16, "im2col_f32_f16", im2col_f32_f16_rte ## bda ## _len, im2col_f32_f16_rte ## bda ## _data, "main", 2, sizeof(vk_op_im2col_push_constants), {512, 1, 1}, { target_subgroup_size }, 1, true); \ ggml_vk_create_pipeline(device, device->pipeline_im2col_f32_f16, "im2col_f32_f16", im2col_f32_f16_rte ## bda ## _len, im2col_f32_f16_rte ## bda ## _data, "main", 2, sizeof(vk_op_im2col_push_constants), {512, 1, 1}, { device->subgroup_size }, 1, true); \
ggml_vk_create_pipeline(device, device->pipeline_im2col_3d_f32_f16, "im2col_3d_f32_f16", im2col_3d_f32_f16_rte ## bda ## _len, im2col_3d_f32_f16_rte ## bda ## _data, "main", 2, sizeof(vk_op_im2col_3d_push_constants), {512, 1, 1}, { 512 }, 1, true); \ ggml_vk_create_pipeline(device, device->pipeline_im2col_3d_f32_f16, "im2col_3d_f32_f16", im2col_3d_f32_f16_rte ## bda ## _len, im2col_3d_f32_f16_rte ## bda ## _data, "main", 2, sizeof(vk_op_im2col_3d_push_constants), {512, 1, 1}, { 512 }, 1, true); \
} else { \ } else { \
ggml_vk_create_pipeline(device, device->pipeline_im2col_f32_f16, "im2col_f32_f16", im2col_f32_f16 ## bda ## _len, im2col_f32_f16 ## bda ## _data, "main", 2, sizeof(vk_op_im2col_push_constants), {512, 1, 1}, { target_subgroup_size }, 1, true); \ ggml_vk_create_pipeline(device, device->pipeline_im2col_f32_f16, "im2col_f32_f16", im2col_f32_f16 ## bda ## _len, im2col_f32_f16 ## bda ## _data, "main", 2, sizeof(vk_op_im2col_push_constants), {512, 1, 1}, { device->subgroup_size }, 1, true); \
ggml_vk_create_pipeline(device, device->pipeline_im2col_3d_f32_f16, "im2col_3d_f32_f16", im2col_3d_f32_f16 ## bda ## _len, im2col_3d_f32_f16 ## bda ## _data, "main", 2, sizeof(vk_op_im2col_3d_push_constants), {512, 1, 1}, { 512 }, 1, true); \ ggml_vk_create_pipeline(device, device->pipeline_im2col_3d_f32_f16, "im2col_3d_f32_f16", im2col_3d_f32_f16 ## bda ## _len, im2col_3d_f32_f16 ## bda ## _data, "main", 2, sizeof(vk_op_im2col_3d_push_constants), {512, 1, 1}, { 512 }, 1, true); \
} }
if (device->shader_int64 && device->buffer_device_address) { if (device->shader_int64 && device->buffer_device_address) {
@ -4027,16 +4021,16 @@ static void ggml_vk_load_shaders(vk_device& device) {
ggml_vk_create_pipeline(device, device->pipeline_pool2d_f32, "pool2d_f32", pool2d_f32_len, pool2d_f32_data, "main", 2, sizeof(vk_op_pool2d_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_pool2d_f32, "pool2d_f32", pool2d_f32_len, pool2d_f32_data, "main", 2, sizeof(vk_op_pool2d_push_constants), {512, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_rwkv_wkv6_f32, "rwkv_wkv6_f32", rwkv_wkv6_f32_len, rwkv_wkv6_f32_data, "main", 7, sizeof(vk_op_rwkv_wkv6_push_constants), {1, 1, 1}, {target_subgroup_size}, 1); ggml_vk_create_pipeline(device, device->pipeline_rwkv_wkv6_f32, "rwkv_wkv6_f32", rwkv_wkv6_f32_len, rwkv_wkv6_f32_data, "main", 7, sizeof(vk_op_rwkv_wkv6_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
ggml_vk_create_pipeline(device, device->pipeline_rwkv_wkv7_f32, "rwkv_wkv7_f32", rwkv_wkv7_f32_len, rwkv_wkv7_f32_data, "main", 8, sizeof(vk_op_rwkv_wkv7_push_constants), {1, 1, 1}, {target_subgroup_size}, 1); ggml_vk_create_pipeline(device, device->pipeline_rwkv_wkv7_f32, "rwkv_wkv7_f32", rwkv_wkv7_f32_len, rwkv_wkv7_f32_data, "main", 8, sizeof(vk_op_rwkv_wkv7_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
if (device->subgroup_arithmetic && device->subgroup_require_full_support) { if (device->subgroup_arithmetic && device->subgroup_require_full_support) {
ggml_vk_create_pipeline(device, device->pipeline_ssm_scan_f32_d128, "ssm_scan_128_f32", ssm_scan_subgroup_f32_len, ssm_scan_subgroup_f32_data, "main", 8, sizeof(vk_op_ssm_scan_push_constants), {1, 1, 1}, {128, target_subgroup_size, 16}, 1, true, true, target_subgroup_size); ggml_vk_create_pipeline(device, device->pipeline_ssm_scan_f32_d128, "ssm_scan_128_f32", ssm_scan_subgroup_f32_len, ssm_scan_subgroup_f32_data, "main", 8, sizeof(vk_op_ssm_scan_push_constants), {1, 1, 1}, {128, device->subgroup_size, 16}, 1, true, true, device->subgroup_size);
ggml_vk_create_pipeline(device, device->pipeline_ssm_scan_f32_d256, "ssm_scan_256_f32", ssm_scan_subgroup_f32_len, ssm_scan_subgroup_f32_data, "main", 8, sizeof(vk_op_ssm_scan_push_constants), {1, 1, 1}, {256, target_subgroup_size, 16}, 1, true, true, target_subgroup_size); ggml_vk_create_pipeline(device, device->pipeline_ssm_scan_f32_d256, "ssm_scan_256_f32", ssm_scan_subgroup_f32_len, ssm_scan_subgroup_f32_data, "main", 8, sizeof(vk_op_ssm_scan_push_constants), {1, 1, 1}, {256, device->subgroup_size, 16}, 1, true, true, device->subgroup_size);
} else { } else {
ggml_vk_create_pipeline(device, device->pipeline_ssm_scan_f32_d128, "ssm_scan_128_f32", ssm_scan_f32_len, ssm_scan_f32_data, "main", 8, sizeof(vk_op_ssm_scan_push_constants), {1, 1, 1}, {128, target_subgroup_size, 16}, 1, true, true); ggml_vk_create_pipeline(device, device->pipeline_ssm_scan_f32_d128, "ssm_scan_128_f32", ssm_scan_f32_len, ssm_scan_f32_data, "main", 8, sizeof(vk_op_ssm_scan_push_constants), {1, 1, 1}, {128, device->subgroup_size, 16}, 1, true, true);
ggml_vk_create_pipeline(device, device->pipeline_ssm_scan_f32_d256, "ssm_scan_256_f32", ssm_scan_f32_len, ssm_scan_f32_data, "main", 8, sizeof(vk_op_ssm_scan_push_constants), {1, 1, 1}, {256, target_subgroup_size, 16}, 1, true, true); ggml_vk_create_pipeline(device, device->pipeline_ssm_scan_f32_d256, "ssm_scan_256_f32", ssm_scan_f32_len, ssm_scan_f32_data, "main", 8, sizeof(vk_op_ssm_scan_push_constants), {1, 1, 1}, {256, device->subgroup_size, 16}, 1, true, true);
} }
ggml_vk_create_pipeline(device, device->pipeline_ssm_conv_f32, "ssm_conv_f32", ssm_conv_f32_len, ssm_conv_f32_data, "main", 3, sizeof(vk_op_ssm_conv_push_constants), {32, 1, 1}, {32}, 1); ggml_vk_create_pipeline(device, device->pipeline_ssm_conv_f32, "ssm_conv_f32", ssm_conv_f32_len, ssm_conv_f32_data, "main", 3, sizeof(vk_op_ssm_conv_push_constants), {32, 1, 1}, {32}, 1);
@ -4104,7 +4098,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
allow_collectives_amd) { allow_collectives_amd) {
use_collectives = 1; use_collectives = 1;
conv2d_BS_CRS = std::min( conv2d_BS_CRS = std::min(
target_subgroup_size, device->subgroup_size,
conv2d_BS_CRS); // CRS block size should be capped at subgroup size for correctness when shuffle is used. conv2d_BS_CRS); // CRS block size should be capped at subgroup size for correctness when shuffle is used.
} }
@ -4113,7 +4107,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
if (device->properties.limits.maxComputeSharedMemorySize < conv2d_shmem_req) { if (device->properties.limits.maxComputeSharedMemorySize < conv2d_shmem_req) {
conv2d_BS_CRS = 8; conv2d_BS_CRS = 8;
if (use_collectives) { if (use_collectives) {
conv2d_BS_CRS = std::min(target_subgroup_size, conv2d_BS_CRS); conv2d_BS_CRS = std::min(device->subgroup_size, conv2d_BS_CRS);
} }
} }
@ -4164,9 +4158,9 @@ static void ggml_vk_load_shaders(vk_device& device) {
ggml_vk_create_pipeline(device, device->pipeline_conv2d_dw_cwhn_f16_f32, "conv2d_dw_cwhn_f16_f32", conv2d_dw_cwhn_f16_f32_len, conv2d_dw_cwhn_f16_f32_data, "main", 3, sizeof(vk_op_conv2d_dw_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_conv2d_dw_cwhn_f16_f32, "conv2d_dw_cwhn_f16_f32", conv2d_dw_cwhn_f16_f32_len, conv2d_dw_cwhn_f16_f32_data, "main", 3, sizeof(vk_op_conv2d_dw_push_constants), {512, 1, 1}, {}, 1);
for (uint32_t i = 0; i < num_topk_moe_pipelines; ++i) { for (uint32_t i = 0; i < num_topk_moe_pipelines; ++i) {
ggml_vk_create_pipeline2(device, device->pipeline_topk_moe[i][TOPK_MOE_EARLY_SOFTMAX], "topk_moe_f32_early_softmax_"+std::to_string(i), topk_moe_f32_len, topk_moe_f32_data, "main", 3, sizeof(vk_op_topk_moe_push_constants), {1, 1, 1}, {target_subgroup_size, 1u<<i, 0, 0}, 1, true, true, forced_subgroup_size); ggml_vk_create_pipeline2(device, device->pipeline_topk_moe[i][TOPK_MOE_EARLY_SOFTMAX], "topk_moe_f32_early_softmax_"+std::to_string(i), topk_moe_f32_len, topk_moe_f32_data, "main", 3, sizeof(vk_op_topk_moe_push_constants), {1, 1, 1}, {device->subgroup_size, 1u<<i, 0, 0}, 1, true, true, device->subgroup_size);
ggml_vk_create_pipeline2(device, device->pipeline_topk_moe[i][TOPK_MOE_EARLY_SOFTMAX_NORM], "topk_moe_f32_early_softmax_norm"+std::to_string(i), topk_moe_f32_len, topk_moe_f32_data, "main", 3, sizeof(vk_op_topk_moe_push_constants), {1, 1, 1}, {target_subgroup_size, 1u<<i, 1, 0}, 1, true, true, forced_subgroup_size); ggml_vk_create_pipeline2(device, device->pipeline_topk_moe[i][TOPK_MOE_EARLY_SOFTMAX_NORM], "topk_moe_f32_early_softmax_norm"+std::to_string(i), topk_moe_f32_len, topk_moe_f32_data, "main", 3, sizeof(vk_op_topk_moe_push_constants), {1, 1, 1}, {device->subgroup_size, 1u<<i, 1, 0}, 1, true, true, device->subgroup_size);
ggml_vk_create_pipeline2(device, device->pipeline_topk_moe[i][TOPK_MOE_LATE_SOFTMAX], "topk_moe_f32_late_softmax"+std::to_string(i), topk_moe_f32_len, topk_moe_f32_data, "main", 3, sizeof(vk_op_topk_moe_push_constants), {1, 1, 1}, {target_subgroup_size, 1u<<i, 0, 1}, 1, true, true, forced_subgroup_size); ggml_vk_create_pipeline2(device, device->pipeline_topk_moe[i][TOPK_MOE_LATE_SOFTMAX], "topk_moe_f32_late_softmax"+std::to_string(i), topk_moe_f32_len, topk_moe_f32_data, "main", 3, sizeof(vk_op_topk_moe_push_constants), {1, 1, 1}, {device->subgroup_size, 1u<<i, 0, 1}, 1, true, true, device->subgroup_size);
} }
for (auto &c : compiles) { for (auto &c : compiles) {