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

metal : extend mat-mat multiplication support (#16225)

Browse Source 
* metal : support mul_mm with src1->type == GGML_TYPE_F16

* metal : support mul_mm_id with src1->type == GGML_TYPE_F16

[no ci]

* metal : mul_mm support ne00 % 32 != 0

* metal : support mul_mm_id with ne00 % 32 != 0

* cont : remove unnecessary unrolls

* cont : simplify data loading

* metal : optimize mul_mm when output bounds checks are not needed
This commit is contained in:
Georgi Gerganov
2025-09-28 09:34:44 +03:00
committed by GitHub
parent 3b53634fe3
commit 6a2c6145a0
7 changed files with 247 additions and 120 deletions
Download Patch File Download Diff File Expand all files Collapse all files

39
ggml/src/ggml-metal/ggml-metal-device.cpp
View File

@@ -438,21 +438,35 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mv_ext(ggml_metal_libr
return res;
}
ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mm(ggml_metal_library_t lib, ggml_type tsrc0, ggml_type tsrc1) {
ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mm(ggml_metal_library_t lib, const ggml_tensor * op) {
char base[256];
char name[256];
const ggml_type tsrc0 = op->src[0]->type;
const ggml_type tsrc1 = op->src[1]->type;
const bool bc_inp = op->src[0]->ne[0] % 32 != 0;
const bool bc_out = op->ne[0] % 64 != 0 || op->ne[1] % 32 != 0;
snprintf(base, 256, "kernel_mul_mm_%s_%s", ggml_type_name(tsrc0), ggml_type_name(tsrc1));
snprintf(name, 256, "%s", base);
snprintf(name, 256, "%s_bci=%d_bco=%d", base, bc_inp, bc_out);
ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
if (res) {
return res;
}
res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
ggml_metal_cv_t cv = ggml_metal_cv_init();
ggml_metal_pipeline_set_smem(res, 8192);
ggml_metal_cv_set_bool(cv, bc_inp, FC_MUL_MM + 0);
ggml_metal_cv_set_bool(cv, bc_out, FC_MUL_MM + 1);
res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
ggml_metal_cv_free(cv);
// when the output size is not multiple of 64x32, we need extra smem to prevent out-of-bounds writes
ggml_metal_pipeline_set_smem(res, bc_out ? 8192 : 4096 + 2048);
return res;
}
@@ -659,19 +673,30 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mm_id_map0(ggml_metal_
return res;
}
ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mm_id(ggml_metal_library_t lib, ggml_type tsrc0, ggml_type tsrc1) {
ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mm_id(ggml_metal_library_t lib, const ggml_tensor * op) {
char base[256];
char name[256];
const ggml_type tsrc0 = op->src[0]->type;
const ggml_type tsrc1 = op->src[1]->type;
const bool bc_inp = op->src[0]->ne[0] % 32 != 0;
snprintf(base, 256, "kernel_mul_mm_id_%s_%s", ggml_type_name(tsrc0), ggml_type_name(tsrc1));
snprintf(name, 256, "%s", base);
snprintf(name, 256, "%s_bci=%d", base, bc_inp);
ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
if (res) {
return res;
}
res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
ggml_metal_cv_t cv = ggml_metal_cv_init();
ggml_metal_cv_set_bool(cv, bc_inp, FC_MUL_MM + 0);
res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
ggml_metal_cv_free(cv);
ggml_metal_pipeline_set_smem(res, 8192);

4
ggml/src/ggml-metal/ggml-metal-device.h
View File

@@ -115,10 +115,10 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_ssm_conv (ggml_me
ggml_metal_pipeline_t ggml_metal_library_get_pipeline_ssm_scan (ggml_metal_library_t lib, const struct ggml_tensor * op);
ggml_metal_pipeline_t ggml_metal_library_get_pipeline_rwkv (ggml_metal_library_t lib, const struct ggml_tensor * op);
ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mv_ext (ggml_metal_library_t lib, enum ggml_type tsrc0, enum ggml_type tsrc1, int nsg, int nxpsg, int r1ptg);
ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mm (ggml_metal_library_t lib, enum ggml_type tsrc0, enum ggml_type tsrc1);
ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mm (ggml_metal_library_t lib, const struct ggml_tensor * op);
ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mv (ggml_metal_library_t lib, const struct ggml_tensor * op);
ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mm_id_map0 (ggml_metal_library_t lib, int ne02, int ne20);
ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mm_id (ggml_metal_library_t lib, enum ggml_type tsrc0, enum ggml_type tsrc1);
ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mm_id (ggml_metal_library_t lib, const struct ggml_tensor * op);
ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mv_id (ggml_metal_library_t lib, const struct ggml_tensor * op);
ggml_metal_pipeline_t ggml_metal_library_get_pipeline_argmax (ggml_metal_library_t lib, const struct ggml_tensor * op);
ggml_metal_pipeline_t ggml_metal_library_get_pipeline_argsort (ggml_metal_library_t lib, const struct ggml_tensor * op);

3
ggml/src/ggml-metal/ggml-metal-device.m
View File

@@ -717,8 +717,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
return true;
case GGML_OP_MUL_MAT:
case GGML_OP_MUL_MAT_ID:
return has_simdgroup_reduction &&
(op->src[0]->type != GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_F32);
return has_simdgroup_reduction;
case GGML_OP_CPY:
case GGML_OP_DUP:
case GGML_OP_CONT:

1
ggml/src/ggml-metal/ggml-metal-impl.h
View File

@@ -76,6 +76,7 @@
#define FC_FLASH_ATTN_EXT_VEC 200
#define FC_FLASH_ATTN_EXT_VEC_REDUCE 300
#define FC_MUL_MV 400
#define FC_MUL_MM 500
// kernel argument structs
//

40
ggml/src/ggml-metal/ggml-metal-ops.cpp
View File

@@ -1520,22 +1520,20 @@ int ggml_metal_op_mul_mat(ggml_metal_op_t ctx, int idx) {
!ggml_is_transposed(op->src[1]) &&
// for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
// AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
props_dev->has_simdgroup_mm &&
op->src[1]->type == GGML_TYPE_F32 &&
ne00 % 32 == 0 && ne00 >= 64 &&
props_dev->has_simdgroup_mm && ne00 >= 64 &&
(ne11 > ne11_mm_min || (ggml_is_quantized(op->src[0]->type) && ne12 > 1))) {
//printf("matrix: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
// some Metal matrix data types require aligned pointers
// ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5)
switch (op->src[0]->type) {
case GGML_TYPE_F32: GGML_ASSERT(nb01 % 16 == 0); break;
case GGML_TYPE_F16: GGML_ASSERT(nb01 % 8 == 0); break;
case GGML_TYPE_BF16: GGML_ASSERT(nb01 % 8 == 0); break;
default: break;
}
//switch (op->src[0]->type) {
// case GGML_TYPE_F32: GGML_ASSERT(nb01 % 16 == 0); break;
// case GGML_TYPE_F16: GGML_ASSERT(nb01 % 8 == 0); break;
// case GGML_TYPE_BF16: GGML_ASSERT(nb01 % 8 == 0); break;
// default: break;
//}
ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_mul_mm(lib, op->src[0]->type, op->src[1]->type);
ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_mul_mm(lib, op);
ggml_metal_kargs_mul_mm args = {
/*.ne00 =*/ ne00,
@@ -1655,8 +1653,6 @@ int ggml_metal_op_mul_mat_id(ggml_metal_op_t ctx, int idx) {
GGML_ASSERT(!ggml_is_transposed(op->src[0]));
GGML_ASSERT(!ggml_is_transposed(op->src[1]));
GGML_ASSERT(op->src[1]->type == GGML_TYPE_F32);
GGML_ASSERT(ne03 == 1);
GGML_ASSERT(ne13 == 1);
@@ -1674,19 +1670,15 @@ int ggml_metal_op_mul_mat_id(ggml_metal_op_t ctx, int idx) {
// ne21 = n_rows (batch size)
const int ne21_mm_id_min = 32;
if (props_dev->has_simdgroup_mm &&
ne00 % 32 == 0 && ne00 >= 64 &&
(ne21 >= ne21_mm_id_min)) {
GGML_ASSERT(ne00 % 4 == 0);
if (props_dev->has_simdgroup_mm && ne00 >= 64 && (ne21 >= ne21_mm_id_min)) {
// some Metal matrix data types require aligned pointers
// ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5)
switch (op->src[0]->type) {
case GGML_TYPE_F32: GGML_ASSERT(nb01 % 16 == 0); break;
case GGML_TYPE_F16: GGML_ASSERT(nb01 % 8 == 0); break;
case GGML_TYPE_BF16: GGML_ASSERT(nb01 % 8 == 0); break;
default: break;
}
//switch (op->src[0]->type) {
// case GGML_TYPE_F32: GGML_ASSERT(nb01 % 16 == 0); break;
// case GGML_TYPE_F16: GGML_ASSERT(nb01 % 8 == 0); break;
// case GGML_TYPE_BF16: GGML_ASSERT(nb01 % 8 == 0); break;
// default: break;
//}
// extra buffers for intermediate id mapping
ggml_metal_buffer_id bid_tpe = bid_dst;
@@ -1730,7 +1722,7 @@ int ggml_metal_op_mul_mat_id(ggml_metal_op_t ctx, int idx) {
ggml_metal_op_concurrency_reset(ctx);
{
ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_mul_mm_id(lib, op->src[0]->type, GGML_TYPE_F16);
ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_mul_mm_id(lib, op);
ggml_metal_kargs_mul_mm_id args = {
/*.ne00 =*/ ne00,

264
ggml/src/ggml-metal/ggml-metal.metal
View File

@@ -33,6 +33,7 @@ using namespace metal;
#if defined(GGML_METAL_HAS_BF16)
typedef matrix<bfloat, 4, 4> bfloat4x4;
typedef matrix<bfloat, 2, 4> bfloat2x4;
#endif
constexpr constant static float kvalues_iq4nl_f[16] = {
@@ -7856,6 +7857,9 @@ kernel void kernel_set_rows_f(
}
}
constant bool FC_mul_mm_bc_inp [[function_constant(FC_MUL_MM + 0)]];
constant bool FC_mul_mm_bc_out [[function_constant(FC_MUL_MM + 1)]];
#define BLOCK_SIZE_M 64 // 8 simdgroup matrices from matrix A
#define BLOCK_SIZE_N 32 // 4 simdgroup matrices from matrix B
#define BLOCK_SIZE_K 32
@@ -7868,7 +7872,7 @@ kernel void kernel_set_rows_f(
#define SG_MAT_ROW 8
// each block_q contains 16*nl weights
template<typename T, typename T4x4, typename simdgroup_T8x8, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread T4x4 &)>
template<typename S0, typename S0_4x4, typename S0_8x8, typename S1, typename S1_2x4, typename S1_8x8, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread S0_4x4 &), typename T0, typename T0_4x4, typename T1, typename T1_2x4>
kernel void kernel_mul_mm(
constant ggml_metal_kargs_mul_mm & args,
device const char * src0,
@@ -7879,8 +7883,8 @@ kernel void kernel_mul_mm(
ushort tiitg[[thread_index_in_threadgroup]],
ushort sgitg[[simdgroup_index_in_threadgroup]]) {
threadgroup T * sa = (threadgroup T *)(shmem);
threadgroup float * sb = (threadgroup float *)(shmem + 4096);
threadgroup S0 * sa = (threadgroup S0 *)(shmem);
threadgroup S1 * sb = (threadgroup S1 *)(shmem + 4096);
const int r0 = tgpig.y;
const int r1 = tgpig.x;
@@ -7894,8 +7898,9 @@ kernel void kernel_mul_mm(
const short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
const short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
simdgroup_T8x8 ma[4];
simdgroup_float8x8 mb[2];
S0_8x8 ma[4];
S1_8x8 mb[2];
simdgroup_float8x8 mc[8];
for (short i = 0; i < 8; i++){
@@ -7913,27 +7918,45 @@ kernel void kernel_mul_mm(
device const block_q * x = (device const block_q *)(src0
+ args.nb01*(r0*BLOCK_SIZE_M + thread_row) + offset0) + offset1;
device const float * y = (device const float *)(src1
const short iy = (BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL));
device const T1 * y = (device const T1 *)(src1
+ args.nb13*i13
+ args.nb12*i12
+ args.nb11*(r1*BLOCK_SIZE_N + thread_col)
+ args.nb10*(BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL)));
+ args.nb10*iy);
for (int loop_k = 0; loop_k < args.ne00; loop_k += BLOCK_SIZE_K) {
// load data and store to threadgroup memory
T4x4 temp_a;
dequantize_func(x, il, temp_a);
if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
threadgroup_barrier(mem_flags::mem_threadgroup);
threadgroup_barrier(mem_flags::mem_threadgroup);
// no need for dequantization
for (short i = 0; i < 16; i++) {
*(sa + SG_MAT_SIZE * ((tiitg/THREAD_PER_ROW/8) \
+ (tiitg%THREAD_PER_ROW)*16 + (i/8)*8) \
+ (tiitg/THREAD_PER_ROW)%8 + (i&7)*8) = loop_k + 16*il + i < args.ne00 ? ((device T0 *) x)[i] : 0;
}
} else {
S0_4x4 temp_a;
dequantize_func(x, il, temp_a);
#pragma unroll(16)
for (short i = 0; i < 16; i++) {
*(sa + SG_MAT_SIZE * ((tiitg/THREAD_PER_ROW/8) \
+ (tiitg%THREAD_PER_ROW)*16 + (i/8)*8) \
+ (tiitg/THREAD_PER_ROW)%8 + (i&7)*8) = temp_a[i/4][i%4];
threadgroup_barrier(mem_flags::mem_threadgroup);
FOR_UNROLL (short i = 0; i < 16; i++) {
*(sa + SG_MAT_SIZE * ((tiitg/THREAD_PER_ROW/8) \
+ (tiitg%THREAD_PER_ROW)*16 + (i/8)*8) \
+ (tiitg/THREAD_PER_ROW)%8 + (i&7)*8) = temp_a[i/4][i%4];
}
}
*(threadgroup float2x4 *)(sb + 32*8*(tiitg%THREAD_PER_COL) + 8*(tiitg/THREAD_PER_COL)) = *((device float2x4 *) y);
if (FC_mul_mm_bc_inp) {
for (short i = 0; i < 8; ++i) {
sb[32*8*(tiitg%THREAD_PER_COL) + 8*(tiitg/THREAD_PER_COL) + i] = loop_k + iy + i < args.ne00 ? (S1) ((device T1 *) y)[i] : 0;
}
} else {
*(threadgroup S1_2x4 *)(sb + 32*8*(tiitg%THREAD_PER_COL) + 8*(tiitg/THREAD_PER_COL)) = (S1_2x4)(*((device T1_2x4 *) y));
}
il = (il + 2 < nl) ? il + 2 : il % 2;
x = (il < 2) ? x + (2 + nl - 1)/nl : x;
@@ -7942,8 +7965,8 @@ kernel void kernel_mul_mm(
threadgroup_barrier(mem_flags::mem_threadgroup);
// load matrices from threadgroup memory and conduct outer products
threadgroup const T * lsma = (sa + THREAD_MAT_M*SG_MAT_SIZE*(sgitg%2));
threadgroup const float * lsmb = (sb + THREAD_MAT_N*SG_MAT_SIZE*(sgitg/2));
threadgroup const S0 * lsma = (sa + THREAD_MAT_M*SG_MAT_SIZE*(sgitg%2));
threadgroup const S1 * lsmb = (sb + THREAD_MAT_N*SG_MAT_SIZE*(sgitg/2));
#pragma unroll(4)
for (short ik = 0; ik < BLOCK_SIZE_K/8; ik++) {
@@ -7971,7 +7994,8 @@ kernel void kernel_mul_mm(
}
}
if ((r0 + 1) * BLOCK_SIZE_M <= args.ne0 && (r1 + 1) * BLOCK_SIZE_N <= args.ne1) {
if (!FC_mul_mm_bc_out || ((r0 + 1) * BLOCK_SIZE_M <= args.ne0 && (r1 + 1) * BLOCK_SIZE_N <= args.ne1)) {
// if no bounds checks on the output are needed, we can directly write to device memory
device float * C = (device float *) dst +
(BLOCK_SIZE_M * r0 + 32*(sgitg & 1)) + \
(BLOCK_SIZE_N * r1 + 16*(sgitg >> 1)) * args.ne0 + im*args.ne1*args.ne0;
@@ -8076,7 +8100,7 @@ template [[host_name("kernel_mul_mm_id_map0_ne20_8" )]] kernel kernel_mul_mm_id_
template [[host_name("kernel_mul_mm_id_map0_ne20_10")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<10>;
template [[host_name("kernel_mul_mm_id_map0_ne20_16")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<16>;
template<typename T, typename T4x4, typename simdgroup_T8x8, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread T4x4 &)>
template<typename S0, typename S0_4x4, typename S0_8x8, typename S1, typename S1_2x4, typename S1_8x8, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread S0_4x4 &), typename T0, typename T0_4x4, typename T1, typename T1_2x4>
kernel void kernel_mul_mm_id(
constant ggml_metal_kargs_mul_mm_id & args,
device const char * src0,
@@ -8090,8 +8114,8 @@ kernel void kernel_mul_mm_id(
ushort tiisg[[thread_index_in_simdgroup]],
ushort sgitg[[simdgroup_index_in_threadgroup]]) {
threadgroup T * sa = (threadgroup T *)(shmem);
threadgroup half * sb = (threadgroup half *)(shmem + 4096);
threadgroup S0 * sa = (threadgroup S0 *)(shmem);
threadgroup S1 * sb = (threadgroup S1 *)(shmem + 4096);
const int r0 = tgpig.y;
const int r1 = tgpig.x;
@@ -8114,8 +8138,9 @@ kernel void kernel_mul_mm_id(
const short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
const short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
simdgroup_T8x8 ma[4];
simdgroup_half8x8 mb[2];
S0_8x8 ma[4];
S1_8x8 mb[2];
simdgroup_float8x8 mc[8];
for (short i = 0; i < 8; i++){
@@ -8136,27 +8161,45 @@ kernel void kernel_mul_mm_id(
device const block_q * x = (device const block_q *)(src0
+ args.nb01*(r0*BLOCK_SIZE_M + thread_row) + offset0) + offset1;
device const float * y = (device const float *)(src1
const short iy = (BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL));
device const T1 * y = (device const T1 *)(src1
+ args.nb13*i13
+ args.nb12*i12
+ args.nb11*i11
+ args.nb10*(BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL)));
+ args.nb10*iy);
for (int loop_k = 0; loop_k < args.ne00; loop_k += BLOCK_SIZE_K) {
// load data and store to threadgroup memory
T4x4 temp_a;
dequantize_func(x, il, temp_a);
if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
threadgroup_barrier(mem_flags::mem_threadgroup);
threadgroup_barrier(mem_flags::mem_threadgroup);
// no need for dequantization
for (short i = 0; i < 16; i++) {
*(sa + SG_MAT_SIZE * ((tiitg/THREAD_PER_ROW/8) \
+ (tiitg%THREAD_PER_ROW)*16 + (i/8)*8) \
+ (tiitg/THREAD_PER_ROW)%8 + (i&7)*8) = loop_k + 16*il + i < args.ne00 ? ((device T0 *) x)[i] : 0;
}
} else {
S0_4x4 temp_a;
dequantize_func(x, il, temp_a);
#pragma unroll(16)
for (short i = 0; i < 16; i++) {
*(sa + SG_MAT_SIZE * ((tiitg/THREAD_PER_ROW/8) \
+ (tiitg%THREAD_PER_ROW)*16 + (i/8)*8) \
+ (tiitg/THREAD_PER_ROW)%8 + (i&7)*8) = temp_a[i/4][i%4];
threadgroup_barrier(mem_flags::mem_threadgroup);
FOR_UNROLL (short i = 0; i < 16; i++) {
*(sa + SG_MAT_SIZE * ((tiitg/THREAD_PER_ROW/8) \
+ (tiitg%THREAD_PER_ROW)*16 + (i/8)*8) \
+ (tiitg/THREAD_PER_ROW)%8 + (i&7)*8) = temp_a[i/4][i%4];
}
}
*(threadgroup half2x4 *)(sb + 32*8*(tiitg%THREAD_PER_COL) + 8*(tiitg/THREAD_PER_COL)) = (half2x4)(*((device float2x4 *) y));
if (FC_mul_mm_bc_inp) {
for (short i = 0; i < 8; ++i) {
sb[32*8*(tiitg%THREAD_PER_COL) + 8*(tiitg/THREAD_PER_COL) + i] = loop_k + iy + i < args.ne00 ? (S1) ((device T1 *) y)[i] : 0;
}
} else {
*(threadgroup S1_2x4 *)(sb + 32*8*(tiitg%THREAD_PER_COL) + 8*(tiitg/THREAD_PER_COL)) = (S1_2x4)(*((device T1_2x4 *) y));
}
il = (il + 2 < nl) ? il + 2 : il % 2;
x = (il < 2) ? x + (2 + nl - 1)/nl : x;
@@ -8165,8 +8208,8 @@ kernel void kernel_mul_mm_id(
threadgroup_barrier(mem_flags::mem_threadgroup);
// load matrices from threadgroup memory and conduct outer products
threadgroup const T * lsma = (sa + THREAD_MAT_M*SG_MAT_SIZE*(sgitg%2));
threadgroup const half * lsmb = (sb + THREAD_MAT_N*SG_MAT_SIZE*(sgitg/2));
threadgroup const S0 * lsma = (sa + THREAD_MAT_M*SG_MAT_SIZE*(sgitg%2));
threadgroup const S1 * lsmb = (sb + THREAD_MAT_N*SG_MAT_SIZE*(sgitg/2));
#pragma unroll(4)
for (short ik = 0; ik < BLOCK_SIZE_K/8; ik++) {
@@ -8299,66 +8342,117 @@ template [[host_name("kernel_set_rows_iq4_nl_i32")]] kernel set_rows_q32_t kerne
// matrix-matrix multiplication
//
typedef decltype(kernel_mul_mm<half, half4x4, simdgroup_half8x8, float4x4, 1, dequantize_f32>) mul_mm_t;
typedef decltype(kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, float4x4, 1, dequantize_f32, float, float4x4, float, float2x4>) mul_mm_t;
template [[host_name("kernel_mul_mm_f32_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, float4x4, 1, dequantize_f32>;
template [[host_name("kernel_mul_mm_f16_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half4x4, 1, dequantize_f16>;
template [[host_name("kernel_mul_mm_f32_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, float4x4, 1, dequantize_f32, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_f16_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, half4x4, 1, dequantize_f16, half, half4x4, float, float2x4>;
#if defined(GGML_METAL_HAS_BF16)
template [[host_name("kernel_mul_mm_bf16_f32")]] kernel mul_mm_t kernel_mul_mm<bfloat, bfloat4x4, simdgroup_bfloat8x8, bfloat4x4, 1, dequantize_bf16>;
template [[host_name("kernel_mul_mm_bf16_f32")]] kernel mul_mm_t kernel_mul_mm<bfloat, bfloat4x4, simdgroup_bfloat8x8, bfloat, bfloat2x4, simdgroup_bfloat8x8, bfloat4x4, 1, dequantize_bf16, bfloat, bfloat4x4, float, float2x4>;
#endif
template [[host_name("kernel_mul_mm_q4_0_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, block_q4_0, 2, dequantize_q4_0>;
template [[host_name("kernel_mul_mm_q4_1_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, block_q4_1, 2, dequantize_q4_1>;
template [[host_name("kernel_mul_mm_q5_0_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, block_q5_0, 2, dequantize_q5_0>;
template [[host_name("kernel_mul_mm_q5_1_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, block_q5_1, 2, dequantize_q5_1>;
template [[host_name("kernel_mul_mm_q8_0_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, block_q8_0, 2, dequantize_q8_0>;
template [[host_name("kernel_mul_mm_mxfp4_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, block_mxfp4, 2, dequantize_mxfp4>;
template [[host_name("kernel_mul_mm_q2_K_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, block_q2_K, QK_NL, dequantize_q2_K>;
template [[host_name("kernel_mul_mm_q3_K_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, block_q3_K, QK_NL, dequantize_q3_K>;
template [[host_name("kernel_mul_mm_q4_K_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, block_q4_K, QK_NL, dequantize_q4_K>;
template [[host_name("kernel_mul_mm_q5_K_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, block_q5_K, QK_NL, dequantize_q5_K>;
template [[host_name("kernel_mul_mm_q6_K_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, block_q6_K, QK_NL, dequantize_q6_K>;
template [[host_name("kernel_mul_mm_iq2_xxs_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
template [[host_name("kernel_mul_mm_iq2_xs_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, block_iq2_xs, QK_NL, dequantize_iq2_xs>;
template [[host_name("kernel_mul_mm_iq3_xxs_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
template [[host_name("kernel_mul_mm_iq3_s_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, block_iq3_s, QK_NL, dequantize_iq3_s>;
template [[host_name("kernel_mul_mm_iq2_s_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, block_iq2_s, QK_NL, dequantize_iq2_s>;
template [[host_name("kernel_mul_mm_iq1_s_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, block_iq1_s, QK_NL, dequantize_iq1_s>;
template [[host_name("kernel_mul_mm_iq1_m_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, block_iq1_m, QK_NL, dequantize_iq1_m>;
template [[host_name("kernel_mul_mm_iq4_nl_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, block_iq4_nl, 2, dequantize_iq4_nl>;
template [[host_name("kernel_mul_mm_iq4_xs_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, block_iq4_xs, QK_NL, dequantize_iq4_xs>;
template [[host_name("kernel_mul_mm_q4_0_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q4_0, 2, dequantize_q4_0, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_q4_1_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q4_1, 2, dequantize_q4_1, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_q5_0_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q5_0, 2, dequantize_q5_0, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_q5_1_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q5_1, 2, dequantize_q5_1, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_q8_0_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q8_0, 2, dequantize_q8_0, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_mxfp4_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_mxfp4, 2, dequantize_mxfp4, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_q2_K_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q2_K, QK_NL, dequantize_q2_K, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_q3_K_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q3_K, QK_NL, dequantize_q3_K, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_q4_K_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q4_K, QK_NL, dequantize_q4_K, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_q5_K_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q5_K, QK_NL, dequantize_q5_K, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_q6_K_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q6_K, QK_NL, dequantize_q6_K, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_iq2_xxs_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_iq2_xs_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq2_xs, QK_NL, dequantize_iq2_xs, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_iq3_xxs_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_iq3_s_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq3_s, QK_NL, dequantize_iq3_s, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_iq2_s_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq2_s, QK_NL, dequantize_iq2_s, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_iq1_s_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq1_s, QK_NL, dequantize_iq1_s, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_iq1_m_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq1_m, QK_NL, dequantize_iq1_m, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_iq4_nl_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq4_nl, 2, dequantize_iq4_nl, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_iq4_xs_f32")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq4_xs, QK_NL, dequantize_iq4_xs, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_f32_f16")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, float4x4, 1, dequantize_f32, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_f16_f16")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, half4x4, 1, dequantize_f16, half, half4x4, half, half2x4>;
#if defined(GGML_METAL_HAS_BF16)
template [[host_name("kernel_mul_mm_bf16_f16")]] kernel mul_mm_t kernel_mul_mm<bfloat, bfloat4x4, simdgroup_bfloat8x8, half, half2x4, simdgroup_half8x8, bfloat4x4, 1, dequantize_bf16, bfloat, bfloat4x4, half, half2x4>;
#endif
template [[host_name("kernel_mul_mm_q4_0_f16")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q4_0, 2, dequantize_q4_0, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_q4_1_f16")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q4_1, 2, dequantize_q4_1, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_q5_0_f16")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q5_0, 2, dequantize_q5_0, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_q5_1_f16")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q5_1, 2, dequantize_q5_1, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_q8_0_f16")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q8_0, 2, dequantize_q8_0, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_mxfp4_f16")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_mxfp4, 2, dequantize_mxfp4, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_q2_K_f16")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q2_K, QK_NL, dequantize_q2_K, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_q3_K_f16")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q3_K, QK_NL, dequantize_q3_K, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_q4_K_f16")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q4_K, QK_NL, dequantize_q4_K, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_q5_K_f16")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q5_K, QK_NL, dequantize_q5_K, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_q6_K_f16")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q6_K, QK_NL, dequantize_q6_K, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_iq2_xxs_f16")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_iq2_xs_f16")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq2_xs, QK_NL, dequantize_iq2_xs, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_iq3_xxs_f16")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_iq3_s_f16")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq3_s, QK_NL, dequantize_iq3_s, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_iq2_s_f16")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq2_s, QK_NL, dequantize_iq2_s, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_iq1_s_f16")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq1_s, QK_NL, dequantize_iq1_s, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_iq1_m_f16")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq1_m, QK_NL, dequantize_iq1_m, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_iq4_nl_f16")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq4_nl, 2, dequantize_iq4_nl, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_iq4_xs_f16")]] kernel mul_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq4_xs, QK_NL, dequantize_iq4_xs, float, float4x4, half, half2x4>;
//
// indirect matrix-matrix multiplication
//
typedef decltype(kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, float4x4, 1, dequantize_f32>) mul_mm_id;
typedef decltype(kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, float4x4, 1, dequantize_f32, float, float4x4, float, float2x4>) mul_mm_id;
template [[host_name("kernel_mul_mm_id_f32_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, float4x4, 1, dequantize_f32>;
template [[host_name("kernel_mul_mm_id_f16_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half4x4, 1, dequantize_f16>;
template [[host_name("kernel_mul_mm_id_f32_f32")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, float4x4, 1, dequantize_f32, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_f16_f32")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, half4x4, 1, dequantize_f16, half, half4x4, float, float2x4>;
#if defined(GGML_METAL_HAS_BF16)
template [[host_name("kernel_mul_mm_id_bf16_f16")]] kernel mul_mm_id kernel_mul_mm_id<bfloat, bfloat4x4, simdgroup_bfloat8x8, bfloat4x4, 1, dequantize_bf16>;
template [[host_name("kernel_mul_mm_id_bf16_f32")]] kernel mul_mm_id kernel_mul_mm_id<bfloat, bfloat4x4, simdgroup_bfloat8x8, bfloat, bfloat2x4, simdgroup_bfloat8x8, bfloat4x4, 1, dequantize_bf16, bfloat, bfloat4x4, float, float2x4>;
#endif
template [[host_name("kernel_mul_mm_id_q4_0_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, block_q4_0, 2, dequantize_q4_0>;
template [[host_name("kernel_mul_mm_id_q4_1_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, block_q4_1, 2, dequantize_q4_1>;
template [[host_name("kernel_mul_mm_id_q5_0_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, block_q5_0, 2, dequantize_q5_0>;
template [[host_name("kernel_mul_mm_id_q5_1_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, block_q5_1, 2, dequantize_q5_1>;
template [[host_name("kernel_mul_mm_id_q8_0_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, block_q8_0, 2, dequantize_q8_0>;
template [[host_name("kernel_mul_mm_id_mxfp4_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, block_mxfp4, 2, dequantize_mxfp4>;
template [[host_name("kernel_mul_mm_id_q2_K_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, block_q2_K, QK_NL, dequantize_q2_K>;
template [[host_name("kernel_mul_mm_id_q3_K_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, block_q3_K, QK_NL, dequantize_q3_K>;
template [[host_name("kernel_mul_mm_id_q4_K_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, block_q4_K, QK_NL, dequantize_q4_K>;
template [[host_name("kernel_mul_mm_id_q5_K_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, block_q5_K, QK_NL, dequantize_q5_K>;
template [[host_name("kernel_mul_mm_id_q6_K_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, block_q6_K, QK_NL, dequantize_q6_K>;
template [[host_name("kernel_mul_mm_id_iq2_xxs_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
template [[host_name("kernel_mul_mm_id_iq2_xs_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, block_iq2_xs, QK_NL, dequantize_iq2_xs>;
template [[host_name("kernel_mul_mm_id_iq3_xxs_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
template [[host_name("kernel_mul_mm_id_iq3_s_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, block_iq3_s, QK_NL, dequantize_iq3_s>;
template [[host_name("kernel_mul_mm_id_iq2_s_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, block_iq2_s, QK_NL, dequantize_iq2_s>;
template [[host_name("kernel_mul_mm_id_iq1_s_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, block_iq1_s, QK_NL, dequantize_iq1_s>;
template [[host_name("kernel_mul_mm_id_iq1_m_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, block_iq1_m, QK_NL, dequantize_iq1_m>;
template [[host_name("kernel_mul_mm_id_iq4_nl_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, block_iq4_nl, 2, dequantize_iq4_nl>;
template [[host_name("kernel_mul_mm_id_iq4_xs_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, block_iq4_xs, QK_NL, dequantize_iq4_xs>;
template [[host_name("kernel_mul_mm_id_q4_0_f32")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q4_0, 2, dequantize_q4_0, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_q4_1_f32")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q4_1, 2, dequantize_q4_1, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_q5_0_f32")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q5_0, 2, dequantize_q5_0, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_q5_1_f32")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q5_1, 2, dequantize_q5_1, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_q8_0_f32")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q8_0, 2, dequantize_q8_0, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_mxfp4_f32")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_mxfp4, 2, dequantize_mxfp4, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_q2_K_f32")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q2_K, QK_NL, dequantize_q2_K, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_q3_K_f32")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q3_K, QK_NL, dequantize_q3_K, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_q4_K_f32")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q4_K, QK_NL, dequantize_q4_K, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_q5_K_f32")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q5_K, QK_NL, dequantize_q5_K, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_q6_K_f32")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q6_K, QK_NL, dequantize_q6_K, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_iq2_xxs_f32")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_iq2_xs_f32")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq2_xs, QK_NL, dequantize_iq2_xs, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_iq3_xxs_f32")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_iq3_s_f32")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq3_s, QK_NL, dequantize_iq3_s, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_iq2_s_f32")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq2_s, QK_NL, dequantize_iq2_s, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_iq1_s_f32")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq1_s, QK_NL, dequantize_iq1_s, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_iq1_m_f32")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq1_m, QK_NL, dequantize_iq1_m, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_iq4_nl_f32")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq4_nl, 2, dequantize_iq4_nl, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq4_xs, QK_NL, dequantize_iq4_xs, float, float4x4, float, float2x4>;
template [[host_name("kernel_mul_mm_id_f32_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, float4x4, 1, dequantize_f32, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_id_f16_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, half4x4, 1, dequantize_f16, half, half4x4, half, half2x4>;
#if defined(GGML_METAL_HAS_BF16)
template [[host_name("kernel_mul_mm_id_bf16_f16")]] kernel mul_mm_id kernel_mul_mm_id<bfloat, bfloat4x4, simdgroup_bfloat8x8, half, half2x4, simdgroup_half8x8, bfloat4x4, 1, dequantize_bf16, bfloat, bfloat4x4, half, half2x4>;
#endif
template [[host_name("kernel_mul_mm_id_q4_0_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q4_0, 2, dequantize_q4_0, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_id_q4_1_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q4_1, 2, dequantize_q4_1, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_id_q5_0_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q5_0, 2, dequantize_q5_0, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_id_q5_1_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q5_1, 2, dequantize_q5_1, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_id_q8_0_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q8_0, 2, dequantize_q8_0, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_id_mxfp4_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_mxfp4, 2, dequantize_mxfp4, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_id_q2_K_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q2_K, QK_NL, dequantize_q2_K, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_id_q3_K_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q3_K, QK_NL, dequantize_q3_K, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_id_q4_K_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q4_K, QK_NL, dequantize_q4_K, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_id_q5_K_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q5_K, QK_NL, dequantize_q5_K, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_id_q6_K_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_q6_K, QK_NL, dequantize_q6_K, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_id_iq2_xxs_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_id_iq2_xs_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq2_xs, QK_NL, dequantize_iq2_xs, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_id_iq3_xxs_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_id_iq3_s_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq3_s, QK_NL, dequantize_iq3_s, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_id_iq2_s_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq2_s, QK_NL, dequantize_iq2_s, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_id_iq1_s_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq1_s, QK_NL, dequantize_iq1_s, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_id_iq1_m_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq1_m, QK_NL, dequantize_iq1_m, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_id_iq4_nl_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq4_nl, 2, dequantize_iq4_nl, float, float4x4, half, half2x4>;
template [[host_name("kernel_mul_mm_id_iq4_xs_f16")]] kernel mul_mm_id kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, block_iq4_xs, QK_NL, dequantize_iq4_xs, float, float4x4, half, half2x4>;
//
// matrix-vector multiplication