ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-03-11 06:20:09 +00:00
Code Issues Packages Projects Releases Wiki Activity

iq2_tn: Metal

Browse Source 
For TriLM-3.9B on a 30-core M2-Max we get PP-512 = 890 t/s,
TG-128 = 98.5 t/s.
This commit is contained in:
Iwan Kawrakow
2024-08-06 07:39:39 +02:00
parent e528505fc8
commit 5d02f7f4a5
2 changed files with 170 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

29
ggml/src/ggml-metal.m
View File

@@ -88,6 +88,7 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_M,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_BN,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_BN,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_TN,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_K,
@@ -122,6 +123,7 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_BN_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_BN_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_TN_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_K_F32,
@@ -152,6 +154,7 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_M_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_BN_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_BN_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_TN_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_K_F32,
@@ -179,6 +182,7 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_BN_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_BN_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_TN_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_K_F32,
@@ -206,6 +210,7 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_BN_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_BN_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_TN_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_K_F32,
@@ -577,6 +582,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_M, get_rows_iq1_m, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_BN, get_rows_iq1_bn, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_BN, get_rows_iq2_bn, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_TN, get_rows_iq2_tn, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL, get_rows_iq4_nl, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS, get_rows_iq4_xs, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_K, get_rows_iq2_k, true);
@@ -611,6 +617,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32, mul_mv_iq1_m_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_BN_F32, mul_mv_iq1_bn_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_BN_F32, mul_mv_iq2_bn_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_TN_F32, mul_mv_iq2_tn_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32, mul_mv_iq4_nl_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32, mul_mv_iq4_xs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_K_F32, mul_mv_iq2_k_f32, ctx->support_simdgroup_reduction);
@@ -641,6 +648,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_M_F32, mul_mv_id_iq1_m_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_BN_F32, mul_mv_id_iq1_bn_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_BN_F32, mul_mv_id_iq2_bn_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_TN_F32, mul_mv_id_iq2_tn_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32, mul_mv_id_iq4_nl_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32, mul_mv_id_iq4_xs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_K_F32, mul_mv_id_iq2_k_f32, ctx->support_simdgroup_reduction);
@@ -668,6 +676,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32, mul_mm_iq1_m_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_BN_F32, mul_mm_iq1_bn_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_BN_F32, mul_mm_iq2_bn_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_TN_F32, mul_mm_iq2_tn_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32, mul_mm_iq4_nl_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32, mul_mm_iq4_xs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_K_F32, mul_mm_iq2_k_f32, ctx->support_simdgroup_mm);
@@ -695,6 +704,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32, mul_mm_id_iq1_m_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_BN_F32, mul_mm_id_iq1_bn_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_BN_F32, mul_mm_id_iq2_bn_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_TN_F32, mul_mm_id_iq2_tn_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32, mul_mm_id_iq4_nl_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32, mul_mm_id_iq4_xs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_K_F32, mul_mm_id_iq2_k_f32, ctx->support_simdgroup_mm);
@@ -1728,6 +1738,7 @@ static enum ggml_status ggml_metal_graph_compute(
case GGML_TYPE_IQ1_M: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32 ].pipeline; break;
case GGML_TYPE_IQ1_BN: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_BN_F32 ].pipeline; break;
case GGML_TYPE_IQ2_BN: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_BN_F32 ].pipeline; break;
case GGML_TYPE_IQ2_TN: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_TN_F32 ].pipeline; break;
case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32 ].pipeline; break;
case GGML_TYPE_IQ4_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32 ].pipeline; break;
case GGML_TYPE_IQ2_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_K_F32 ].pipeline; break;
@@ -1904,6 +1915,12 @@ static enum ggml_status ggml_metal_graph_compute(
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_BN_F32].pipeline;
} break;
case GGML_TYPE_IQ2_TN:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_TN_F32].pipeline;
} break;
case GGML_TYPE_IQ4_NL:
{
nth0 = 4;
@@ -1972,7 +1989,7 @@ static enum ggml_status ggml_metal_graph_compute(
src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 || src0t == GGML_TYPE_Q2_K ||
src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S||
src0t == GGML_TYPE_IQ1_BN|| src0t == GGML_TYPE_IQ2_BN|| src0t == GGML_TYPE_IQ2_K||
src0t == GGML_TYPE_IQ3_K) {
src0t == GGML_TYPE_IQ3_K || src0t == GGML_TYPE_IQ2_TN) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
@@ -2074,6 +2091,7 @@ static enum ggml_status ggml_metal_graph_compute(
case GGML_TYPE_IQ1_M: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32 ].pipeline; break;
case GGML_TYPE_IQ1_BN: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_BN_F32 ].pipeline; break;
case GGML_TYPE_IQ2_BN: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_BN_F32 ].pipeline; break;
case GGML_TYPE_IQ2_TN: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_TN_F32 ].pipeline; break;
case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32 ].pipeline; break;
case GGML_TYPE_IQ4_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32 ].pipeline; break;
case GGML_TYPE_IQ2_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_K_F32 ].pipeline; break;
@@ -2244,6 +2262,12 @@ static enum ggml_status ggml_metal_graph_compute(
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_BN_F32].pipeline;
} break;
case GGML_TYPE_IQ2_TN:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_TN_F32].pipeline;
} break;
case GGML_TYPE_IQ4_NL:
{
nth0 = 4;
@@ -2323,7 +2347,7 @@ static enum ggml_status ggml_metal_graph_compute(
src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 || src0t == GGML_TYPE_Q2_K ||
src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S||
src0t == GGML_TYPE_IQ1_BN|| src0t == GGML_TYPE_IQ2_BN|| src0t == GGML_TYPE_IQ2_K||
src0t == GGML_TYPE_IQ3_K) {
src0t == GGML_TYPE_IQ3_K || src0t == GGML_TYPE_IQ2_TN) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
@@ -2384,6 +2408,7 @@ static enum ggml_status ggml_metal_graph_compute(
case GGML_TYPE_IQ1_M: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_M ].pipeline; break;
case GGML_TYPE_IQ1_BN: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_BN ].pipeline; break;
case GGML_TYPE_IQ2_BN: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_BN ].pipeline; break;
case GGML_TYPE_IQ2_TN: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_TN ].pipeline; break;
case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL ].pipeline; break;
case GGML_TYPE_IQ4_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS ].pipeline; break;
case GGML_TYPE_IQ2_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_K ].pipeline; break;

144
ggml/src/ggml-metal.metal
View File

@@ -3330,6 +3330,129 @@ kernel void kernel_mul_mv_q2_K_f32(
kernel_mul_mv_q2_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
}
void kernel_mul_mv_iq2_tn_f32_impl(
device const void * src0,
device const float * src1,
device float * dst,
int64_t ne00,
int64_t ne01,
int64_t ne02,
int64_t ne10,
int64_t ne12,
int64_t ne0,
int64_t ne1,
uint r2,
uint r3,
threadgroup int8_t * shared_values,
uint3 tgpig,
uint tiisg,
uint sgitg) {
const int nb = ne00/QK_K;
const int r0 = tgpig.x;
const int r1 = tgpig.y;
const int im = tgpig.z;
const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
const int ib_row = first_row * nb;
const uint i12 = im%ne12;
const uint i13 = im/ne12;
const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
device const block_iq2_tn * x = (device const block_iq2_tn *) src0 + ib_row + offset0;
device const float * y = (device const float *) src1 + r1*ne10 + im*ne00*ne1;
float yl[32];
float sumf[N_DST]={0.f}, all_sum;
const int step = sizeof(block_iq2_tn) * nb / 2;
const int ix = tiisg/8; // 0...3
const int it = tiisg%8; // 0...7
const int iq = it/4; // 0 or 1
const int ir = it%4; // 0...3
const int is = (8*ir)/16;// 0 or 1
device const float * y4 = y + ix * QK_K + 128 * iq + 8 * ir;
for (int ib = ix; ib < nb; ib += 4) {
float sumy = 0.f;
for (int i = 0; i < 8; ++i) {
yl[i+ 0] = y4[i+ 0]; sumy += yl[i+ 0];
yl[i+ 8] = y4[i+32]; sumy += yl[i+ 8];
yl[i+16] = y4[i+64]; sumy += yl[i+16];
yl[i+24] = y4[i+96]; sumy += yl[i+24];
}
device const half * dh = &x[ib].d;
device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir;
for (int row = 0; row < N_DST; row++) {
float4 acc1 = {0.f, 0.f, 0.f, 0.f};
float4 acc2 = {0.f, 0.f, 0.f, 0.f};
for (int i = 0; i < 8; i += 2) {
acc1[0] += yl[i+ 0] * (qs[i/2] & 0x0003);
acc2[0] += yl[i+ 1] * (qs[i/2] & 0x0300);
acc1[1] += yl[i+ 8] * (qs[i/2] & 0x000c);
acc2[1] += yl[i+ 9] * (qs[i/2] & 0x0c00);
acc1[2] += yl[i+16] * (qs[i/2] & 0x0030);
acc2[2] += yl[i+17] * (qs[i/2] & 0x3000);
acc1[3] += yl[i+24] * (qs[i/2] & 0x00c0);
acc2[3] += yl[i+25] * (qs[i/2] & 0xc000);
}
float dall = dh[0];
sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc2[0]) * 1.f/ 1.f +
(acc1[1] + 1.f/256.f * acc2[1]) * 1.f/ 4.f +
(acc1[2] + 1.f/256.f * acc2[2]) * 1.f/16.f +
(acc1[3] + 1.f/256.f * acc2[3]) * 1.f/64.f - sumy);
qs += step;
dh += step;
}
y4 += 4 * QK_K;
}
for (int row = 0; row < N_DST; ++row) {
all_sum = simd_sum(sumf[row]);
if (tiisg == 0) {
dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
}
}
}
[[host_name("kernel_mul_mv_iq2_tn_f32")]]
kernel void kernel_mul_mv_iq2_tn_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
kernel_mul_mv_iq2_tn_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
}
void kernel_mul_mv_q3_K_f32_impl(
device const void * src0,
device const float * src1,
@@ -6009,7 +6132,7 @@ void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg
}
template <typename type4x4>
void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg) {
void dequantize_q2_K(device const block_q2_K * xb, short il, thread type4x4 & reg) {
const float d = xb->d;
const float min = xb->dmin;
device const uint8_t * q = (device const uint8_t *)xb->qs;
@@ -6027,6 +6150,21 @@ void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg
}
}
template <typename type4x4>
void dequantize_iq2_tn(device const block_iq2_tn * xb, short il, thread type4x4 & reg) {
const half d = xb->d;
device const uint8_t * q = (device const uint8_t *)xb->qs + 32*(il/8) + 16*(il&1);
il = (il/2)%4;
half coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
uchar mask = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3);
const half dl = d * coef;
for (int i = 0; i < 16; ++i) {
reg[i/4][i%4] = dl * (q[i] & mask) - d;
}
}
template <typename type4x4>
void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg) {
const half d_all = xb->d;
@@ -6892,6 +7030,7 @@ template [[host_name("kernel_get_rows_q5_0")]] kernel get_rows_q_t kernel_get
template [[host_name("kernel_get_rows_q5_1")]] kernel get_rows_q_t kernel_get_rows_q<block_q5_1, 2, dequantize_q5_1>;
template [[host_name("kernel_get_rows_q8_0")]] kernel get_rows_q_t kernel_get_rows_q<block_q8_0, 2, dequantize_q8_0>;
template [[host_name("kernel_get_rows_q2_K")]] kernel get_rows_q_t kernel_get_rows_q<block_q2_K, QK_NL, dequantize_q2_K>;
template [[host_name("kernel_get_rows_iq2_tn")]] kernel get_rows_q_t kernel_get_rows_q<block_iq2_tn, QK_NL, dequantize_iq2_tn>;
template [[host_name("kernel_get_rows_q3_K")]] kernel get_rows_q_t kernel_get_rows_q<block_q3_K, QK_NL, dequantize_q3_K>;
template [[host_name("kernel_get_rows_q4_K")]] kernel get_rows_q_t kernel_get_rows_q<block_q4_K, QK_NL, dequantize_q4_K>;
template [[host_name("kernel_get_rows_q5_K")]] kernel get_rows_q_t kernel_get_rows_q<block_q5_K, QK_NL, dequantize_q5_K>;
@@ -6926,6 +7065,7 @@ template [[host_name("kernel_mul_mm_q5_0_f32")]] kernel mat_mm_t kernel_mul_m
template [[host_name("kernel_mul_mm_q5_1_f32")]] kernel mat_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 mat_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, block_q8_0, 2, dequantize_q8_0>;
template [[host_name("kernel_mul_mm_q2_K_f32")]] kernel mat_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, block_q2_K, QK_NL, dequantize_q2_K>;
template [[host_name("kernel_mul_mm_iq2_tn_f32")]] kernel mat_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, block_iq2_tn, QK_NL, dequantize_iq2_tn>;
template [[host_name("kernel_mul_mm_q3_K_f32")]] kernel mat_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 mat_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 mat_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, block_q5_K, QK_NL, dequantize_q5_K>;
@@ -6960,6 +7100,7 @@ template [[host_name("kernel_mul_mm_id_q5_0_f32")]] kernel mat_mm_id_t kernel
template [[host_name("kernel_mul_mm_id_q5_1_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q5_1, 2, dequantize_q5_1>;
template [[host_name("kernel_mul_mm_id_q8_0_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q8_0, 2, dequantize_q8_0>;
template [[host_name("kernel_mul_mm_id_q2_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q2_K, QK_NL, dequantize_q2_K>;
template [[host_name("kernel_mul_mm_id_iq2_tn_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_tn, QK_NL, dequantize_iq2_tn>;
template [[host_name("kernel_mul_mm_id_q3_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q3_K, QK_NL, dequantize_q3_K>;
template [[host_name("kernel_mul_mm_id_q4_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q4_K, QK_NL, dequantize_q4_K>;
template [[host_name("kernel_mul_mm_id_q5_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q5_K, QK_NL, dequantize_q5_K>;
@@ -7175,6 +7316,7 @@ template [[host_name("kernel_mul_mv_id_q4_1_f32")]] kernel kernel_mul_mv_id_t
template [[host_name("kernel_mul_mv_id_q5_0_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<mul_vec_q_n_f32_impl<block_q5_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>>>;
template [[host_name("kernel_mul_mv_id_q5_1_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<mul_vec_q_n_f32_impl<block_q5_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>>>;
template [[host_name("kernel_mul_mv_id_q2_K_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q2_K_f32_impl>>;
template [[host_name("kernel_mul_mv_id_iq2_tn_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq2_tn_f32_impl>>;
template [[host_name("kernel_mul_mv_id_q3_K_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q3_K_f32_impl>>;
template [[host_name("kernel_mul_mv_id_q4_K_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q4_K_f32_impl>>;
template [[host_name("kernel_mul_mv_id_q5_K_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q5_K_f32_impl>>;