Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

opencl: add q6_K gemm and gemv kernels for Adreno (#20089)

Browse Source 
* opencl: add q6_K noshuffle kernels, initial q6_K gemv, some host code

* opencl: add q6_K transpose

* opencl: fix cvt kernel name

* opencl: add call to q6_K gemv

* opencl: fix q6_K scale transpose

* opencl: fix loading for gemv q6_K, refactor

* opencl: fix transpose_8_buf kernel assignment, refactor

* opencl: refactor q6_K transpose

* opencl: add gemm_noshuffle_q6_k_f32

* opencl: fix qh loading

* opencl: refactor q6_K gemv host side, release bufs and imgs

* opencl: refactor

* opencl: fix q6_K dequant and scale selection

* opencl: workaround compiler bug, fix dump_tensor

* opencl: refactor q6_K convert kernels

* opencl: unpack transformed q6_K in get_tensor

* opencl: refactor, handle non-uniform workgroups

* opencl: support non-vector subgroup bcast
This commit is contained in:
lhez
2026-03-23 12:44:18 -07:00
committed by GitHub
parent 39bf0d3c6a
commit 1772701f99
5 changed files with 918 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-opencl/kernels/cvt.cl
+126 -2
View File
@@ -486,8 +486,13 @@ kernel void kernel_convert_block_q6_K(
global uchar * dst_ql,
global uchar * dst_qh,
global char * dst_s,
global half * dst_d
global half * dst_d,
uchar mask_lsb_8,
ulong n_blk
) {
if (get_global_id(0) >= n_blk) {
return;
}
global struct block_q6_K * b = (global struct block_q6_K *) src0 + get_global_id(0);
global uchar * ql = (global uchar *) dst_ql + QK_K/2*get_global_id(0);
global uchar * qh = (global uchar *) dst_qh + QK_K/4*get_global_id(0);
@@ -514,8 +519,13 @@ kernel void kernel_restore_block_q6_K(
global uchar * dst_qh,
global char * dst_s,
global half * dst_d,
global struct block_q6_K * dst
global struct block_q6_K * dst,
uchar mask_lsb_8,
ulong n_blk
) {
if (get_global_id(0) >= n_blk) {
return;
}
global struct block_q6_K * b = (global struct block_q6_K *) dst + get_global_id(0);
global uchar * ql = (global uchar *) dst_ql + QK_K/2*get_global_id(0);
global uchar * qh = (global uchar *) dst_qh + QK_K/4*get_global_id(0);
@@ -534,3 +544,117 @@ kernel void kernel_restore_block_q6_K(
b->scales[i] = s[i];
}
}
kernel void kernel_convert_block_q6_K_noshuffle(
global struct block_q6_K * src0,
global uchar * dst_ql,
global uchar * dst_qh,
global char * dst_s,
global half * dst_d,
uchar mask_lsb_8,
ulong n_blk
) {
if (get_global_id(0) >= n_blk) {
return;
}
global struct block_q6_K * b = (global struct block_q6_K *) src0 + get_global_id(0);
global uchar * ql = (global uchar *) dst_ql + QK_K/2*get_global_id(0);
global uchar * qh = (global uchar *) dst_qh + QK_K/4*get_global_id(0);
global char * s = (global char *) dst_s + QK_K/16*get_global_id(0);
global half * d = (global half *) dst_d + get_global_id(0);
*d = b->d;
for (int i = 0; i < QK_K/2/4; ++i) {
uchar x0 = b->ql[i*2 + 0] & mask_lsb_8;
uchar x1 = b->ql[i*2 + 1] & mask_lsb_8;
ql[i + 0] = (x0 & 0x0F) | ((x1 & 0x0F) << 4);
ql[i + 32] = ((x0 & 0xF0) >> 4) | (x1 & 0xF0);
uchar x2 = b->ql[i*2 + 0 + 64] & mask_lsb_8;
uchar x3 = b->ql[i*2 + 1 + 64] & mask_lsb_8;
ql[i + 64] = (x2 & 0x0F) | ((x3 & 0x0F) << 4);
ql[i + 96] = ((x2 & 0xF0) >> 4) | (x3 & 0xF0);
}
for (int i = 0; i < QK_K/4/8; ++i) {
uchar x0 = b->qh[i*4 + 0] & mask_lsb_8;
uchar x1 = b->qh[i*4 + 1] & mask_lsb_8;
uchar x2 = b->qh[i*4 + 2] & mask_lsb_8;
uchar x3 = b->qh[i*4 + 3] & mask_lsb_8;
qh[i + 0] = (x0 & 0x03) | ((x1 & 0x03) << 2) | ((x2 & 0x03) << 4) | ((x3 & 0x03) << 6);
qh[i + 8] = ((x0 & 0x0C) >> 2) | (x1 & 0x0C) | ((x2 & 0x0C) << 2) | ((x3 & 0x0C) << 4);
qh[i + 16] = ((x0 & 0x30) >> 4) | ((x1 & 0x30) >> 2) | (x2 & 0x30) | ((x3 & 0x30) << 2);
qh[i + 24] = ((x0 & 0xC0) >> 6) | ((x1 & 0xC0) >> 4) | ((x2 & 0xC0) >> 2) | (x3 & 0xC0);
uchar x4 = b->qh[i*4 + 0 + 32] & mask_lsb_8;
uchar x5 = b->qh[i*4 + 1 + 32] & mask_lsb_8;
uchar x6 = b->qh[i*4 + 2 + 32] & mask_lsb_8;
uchar x7 = b->qh[i*4 + 3 + 32] & mask_lsb_8;
qh[i + 32] = (x4 & 0x03) | ((x5 & 0x03) << 2) | ((x6 & 0x03) << 4) | ((x7 & 0x03) << 6);
qh[i + 40] = ((x4 & 0x0C) >> 2) | (x5 & 0x0C) | ((x6 & 0x0C) << 2) | ((x7 & 0x0C) << 4);
qh[i + 48] = ((x4 & 0x30) >> 4) | ((x5 & 0x30) >> 2) | (x6 & 0x30) | ((x7 & 0x30) << 2);
qh[i + 56] = ((x4 & 0xC0) >> 6) | ((x5 & 0xC0) >> 4) | ((x6 & 0xC0) >> 2) | (x7 & 0xC0);
}
for (int i = 0; i < QK_K/16; ++i) {
s[i] = b->scales[i];
}
}
kernel void kernel_restore_block_q6_K_noshuffle(
global uchar * src_ql,
global uchar * src_qh,
global char * src_s,
global half * src_d,
global struct block_q6_K * dst,
uchar mask_lsb_8,
ulong n_blk
) {
if (get_global_id(0) >= n_blk) {
return;
}
global struct block_q6_K * b = (global struct block_q6_K *) dst + get_global_id(0);
global uchar * ql = (global uchar *) src_ql + QK_K/2*get_global_id(0);
global uchar * qh = (global uchar *) src_qh + QK_K/4*get_global_id(0);
global char * s = (global char *) src_s + QK_K/16*get_global_id(0);
global half * d = (global half *) src_d + get_global_id(0);
b->d = *d;
for (int i = 0; i < QK_K/2/4; ++i) {
uchar x0 = ql[i + 0] & mask_lsb_8;
uchar x1 = ql[i + 32] & mask_lsb_8;
b->ql[i*2 + 0] = (x0 & 0x0F) | ((x1 & 0x0F) << 4);
b->ql[i*2 + 1] = ((x0 & 0xF0) >> 4) | (x1 & 0xF0);
uchar x2 = ql[i + 64] & mask_lsb_8;
uchar x3 = ql[i + 96] & mask_lsb_8;
b->ql[i*2 + 0 + 64] = (x2 & 0x0F) | ((x3 & 0x0F) << 4);
b->ql[i*2 + 1 + 64] = ((x2 & 0xF0) >> 4) | (x3 & 0xF0);
}
for (int i = 0; i < QK_K/4/8; ++i) {
uchar x0 = qh[i + 0] & mask_lsb_8;
uchar x1 = qh[i + 8] & mask_lsb_8;
uchar x2 = qh[i + 16] & mask_lsb_8;
uchar x3 = qh[i + 24] & mask_lsb_8;
b->qh[i*4 + 0] = (x0 & 0x03) | ((x1 & 0x03) << 2) | ((x2 & 0x03) << 4) | ((x3 & 0x03) << 6);
b->qh[i*4 + 1] = ((x0 & 0x0C) >> 2) | (x1 & 0x0C) | ((x2 & 0x0C) << 2) | ((x3 & 0x0C) << 4);
b->qh[i*4 + 2] = ((x0 & 0x30) >> 4) | ((x1 & 0x30) >> 2) | (x2 & 0x30) | ((x3 & 0x30) << 2);
b->qh[i*4 + 3] = ((x0 & 0xC0) >> 6) | ((x1 & 0xC0) >> 4) | ((x2 & 0xC0) >> 2) | (x3 & 0xC0);
uchar x4 = qh[i + 0 + 32] & mask_lsb_8;
uchar x5 = qh[i + 8 + 32] & mask_lsb_8;
uchar x6 = qh[i + 16 + 32] & mask_lsb_8;
uchar x7 = qh[i + 24 + 32] & mask_lsb_8;
b->qh[i*4 + 0 + 32] = (x4 & 0x03) | ((x5 & 0x03) << 2) | ((x6 & 0x03) << 4) | ((x7 & 0x03) << 6);
b->qh[i*4 + 1 + 32] = ((x4 & 0x0C) >> 2) | (x5 & 0x0C) | ((x6 & 0x0C) << 2) | ((x7 & 0x0C) << 4);
b->qh[i*4 + 2 + 32] = ((x4 & 0x30) >> 4) | ((x5 & 0x30) >> 2) | (x6 & 0x30) | ((x7 & 0x30) << 2);
b->qh[i*4 + 3 + 32] = ((x4 & 0xC0) >> 6) | ((x5 & 0xC0) >> 4) | ((x6 & 0xC0) >> 2) | (x7 & 0xC0);
}
for (int i = 0; i < QK_K/16; ++i) {
b->scales[i] = s[i];
}
}
ggml/src/ggml-opencl/kernels/gemm_noshuffle_q6_k_f32.cl
+140
View File
@@ -0,0 +1,140 @@
#pragma OPENCL EXTENSION cl_khr_fp16 : enable
#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
#ifdef cl_qcom_reqd_sub_group_size
#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
#define ADRENO_GPU 1
#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
#endif
#ifdef ADRENO_GPU
REQD_SUBGROUP_SIZE_128
#endif
kernel void kernel_gemm_noshuffle_q6_K_f32(
global const ushort * src0_ql,
global const uchar * src0_qh,
global const ushort * src0_s,
global const half * src0_d,
read_only image1d_buffer_t src1,
global float * dst,
ulong offsetd,
int m,
int n,
int k,
int n_no_padding,
ushort mask_f000,
uchar mask_c0
) {
dst = (global float *)( (global char *)dst + offsetd );
int m_4 = m >> 2;
int n_4 = n >> 2;
int gy = get_global_id(0); // n
int gx = get_global_id(1); // m
int gx_2 = gx << 2;
half8 c0 = 0, c1 = 0, c2 = 0, c3 = 0;
half8 B;
half4 dequantized_weights;
global const ushort * ptr_ql = src0_ql + gx_2;
global const uchar * ptr_qh = src0_qh + gx_2;
global const ushort * ptr_s = src0_s + gx_2;
global const half * ptr_d = src0_d + gx_2;
for (int i = 0; i < k; i += 4) {
// load 4x elements (ushort) of ql on M, each ushort contains 4 weights
// 4x ushort correspons to 4 rows on M
ushort4 bits4 = vload4(0, ptr_ql + (i/4)*m); // ql packed in 4s in ushort
uchar4 bits2 = vload4(0, ptr_qh + (i/4)*m); // qh packed in 4s in uchar
// load 4 consecutive scales
char8 scale_s_8 = as_char8(vload4(0, ptr_s + (i/16/2)*m)); // 1 char scale every 16 elements, packed in 2s
char4 scale_s = ((i/16) % 2) == 0 ? scale_s_8.s0246 : scale_s_8.s1357; // transposed as ushort, 2 blocks
half4 scale_d = vload4(0, ptr_d + (i/256)*m); // 1 half scale every 256 elements
// j=0
// load 2x 4 elements of activations on N, corresponding to 8 rows on N
B.s0123 = read_imageh(src1, gy*2 + (i + 0)*n_4 + 0);
B.s4567 = read_imageh(src1, gy*2 + (i + 0)*n_4 + 1);
dequantized_weights.s0 = (convert_half((bits4.s0 & 0x000F) | ((bits2.s0 & 0x03) << 4)) - 32.f) * scale_s.s0 * scale_d.s0;
dequantized_weights.s1 = (convert_half((bits4.s1 & 0x000F) | ((bits2.s1 & 0x03) << 4)) - 32.f) * scale_s.s1 * scale_d.s1;
dequantized_weights.s2 = (convert_half((bits4.s2 & 0x000F) | ((bits2.s2 & 0x03) << 4)) - 32.f) * scale_s.s2 * scale_d.s2;
dequantized_weights.s3 = (convert_half((bits4.s3 & 0x000F) | ((bits2.s3 & 0x03) << 4)) - 32.f) * scale_s.s3 * scale_d.s3;
c0 += B * dequantized_weights.s0;
c1 += B * dequantized_weights.s1;
c2 += B * dequantized_weights.s2;
c3 += B * dequantized_weights.s3;
// j=1
B.s0123 = read_imageh(src1, gy*2 + (i + 1)*n_4 + 0);
B.s4567 = read_imageh(src1, gy*2 + (i + 1)*n_4 + 1);
dequantized_weights.s0 = (convert_half((((bits4.s0 & 0x00F0) >> 4) | ((bits2.s0 & 0x0C) << 2))) - 32.f) * scale_s.s0 * scale_d.s0;
dequantized_weights.s1 = (convert_half((((bits4.s1 & 0x00F0) >> 4) | ((bits2.s1 & 0x0C) << 2))) - 32.f) * scale_s.s1 * scale_d.s1;
dequantized_weights.s2 = (convert_half((((bits4.s2 & 0x00F0) >> 4) | ((bits2.s2 & 0x0C) << 2))) - 32.f) * scale_s.s2 * scale_d.s2;
dequantized_weights.s3 = (convert_half((((bits4.s3 & 0x00F0) >> 4) | ((bits2.s3 & 0x0C) << 2))) - 32.f) * scale_s.s3 * scale_d.s3;
c0 += B * dequantized_weights.s0;
c1 += B * dequantized_weights.s1;
c2 += B * dequantized_weights.s2;
c3 += B * dequantized_weights.s3;
// j=2
B.s0123 = read_imageh(src1, gy*2 + (i + 2)*n_4 + 0);
B.s4567 = read_imageh(src1, gy*2 + (i + 2)*n_4 + 1);
dequantized_weights.s0 = (convert_half((((bits4.s0 & 0x0F00) >> 8) | (bits2.s0 & 0x30))) - 32.f) * scale_s.s0 * scale_d.s0;
dequantized_weights.s1 = (convert_half((((bits4.s1 & 0x0F00) >> 8) | (bits2.s1 & 0x30))) - 32.f) * scale_s.s1 * scale_d.s1;
dequantized_weights.s2 = (convert_half((((bits4.s2 & 0x0F00) >> 8) | (bits2.s2 & 0x30))) - 32.f) * scale_s.s2 * scale_d.s2;
dequantized_weights.s3 = (convert_half((((bits4.s3 & 0x0F00) >> 8) | (bits2.s3 & 0x30))) - 32.f) * scale_s.s3 * scale_d.s3;
c0 += B * dequantized_weights.s0;
c1 += B * dequantized_weights.s1;
c2 += B * dequantized_weights.s2;
c3 += B * dequantized_weights.s3;
// j=3
B.s0123 = read_imageh(src1, gy*2 + (i + 3)*n_4 + 0);
B.s4567 = read_imageh(src1, gy*2 + (i + 3)*n_4 + 1);
dequantized_weights.s0 = (convert_half((((bits4.s0 & mask_f000) >> 12) | ((bits2.s0 & mask_c0) >> 2))) - 32.f) * scale_s.s0 * scale_d.s0;
dequantized_weights.s1 = (convert_half((((bits4.s1 & mask_f000) >> 12) | ((bits2.s1 & mask_c0) >> 2))) - 32.f) * scale_s.s1 * scale_d.s1;
dequantized_weights.s2 = (convert_half((((bits4.s2 & mask_f000) >> 12) | ((bits2.s2 & mask_c0) >> 2))) - 32.f) * scale_s.s2 * scale_d.s2;
dequantized_weights.s3 = (convert_half((((bits4.s3 & mask_f000) >> 12) | ((bits2.s3 & mask_c0) >> 2))) - 32.f) * scale_s.s3 * scale_d.s3;
c0 += B * dequantized_weights.s0;
c1 += B * dequantized_weights.s1;
c2 += B * dequantized_weights.s2;
c3 += B * dequantized_weights.s3;
}
int idx = (gy<<3)*m + (gx<<2);
if(idx+3 < m*n_no_padding){
vstore4((float4)(c0.s0, c1.s0, c2.s0, c3.s0), 0, dst + idx);
idx += m;
}
if(idx+3 < m*n_no_padding){
vstore4((float4)(c0.s1, c1.s1, c2.s1, c3.s1), 0, dst + idx);
idx += m;
}
if(idx+3 < m*n_no_padding){
vstore4((float4)(c0.s2, c1.s2, c2.s2, c3.s2), 0, dst + idx);
idx += m;
}
if(idx+3 < m*n_no_padding){
vstore4((float4)(c0.s3, c1.s3, c2.s3, c3.s3), 0, dst + idx);
idx += m;
}
if(idx+3 < m*n_no_padding){
vstore4((float4)(c0.s4, c1.s4, c2.s4, c3.s4), 0, dst + idx);
idx += m;
}
if(idx+3 < m*n_no_padding){
vstore4((float4)(c0.s5, c1.s5, c2.s5, c3.s5), 0, dst + idx);
idx += m;
}
if(idx+3 < m*n_no_padding){
vstore4((float4)(c0.s6, c1.s6, c2.s6, c3.s6), 0, dst + idx);
idx += m;
}
if(idx+3 < m*n_no_padding){
vstore4((float4)(c0.s7, c1.s7, c2.s7, c3.s7), 0, dst + idx);
}
}
ggml/src/ggml-opencl/kernels/gemv_noshuffle_q6_k_f32.cl
+293
View File
@@ -0,0 +1,293 @@
#pragma OPENCL EXTENSION cl_khr_fp16 : enable
#pragma OPENCL EXTENSION cl_khr_subgroups : enable
#ifdef cl_intel_required_subgroup_size
#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
#define INTEL_GPU 1
#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
#elif defined(cl_qcom_reqd_sub_group_size)
#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
#define ADRENO_GPU 1
#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half")))
#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
#endif
#define NSUBGROUPS 4
#define SUBGROUP_SIZE 64
#define dequantize_block_acc_bcast_8_hi(total_sum, bits4, bits2, scale_d, scale_s, y) \
float8 shared_y; \
shared_y = sub_group_broadcast(y, 0); \
total_sum.s0 += ((float)(((bits4.s0 & 0x000F) ) | ((bits2.s0 & 0x03) << 4)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s0; \
total_sum.s0 += ((float)(((bits4.s0 & 0x00F0) >> 4) | ((bits2.s0 & 0x0C) << 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s1; \
total_sum.s0 += ((float)(((bits4.s0 & 0x0F00) >> 8) | ((bits2.s0 & 0x30) )) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s2; \
total_sum.s0 += ((float)(((bits4.s0 & 0xF000) >> 12) | ((bits2.s0 & 0xC0) >> 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s3; \
total_sum.s0 += ((float)(((bits4.s2 & 0x000F) ) | ((bits2.s2 & 0x03) << 4)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s4; \
total_sum.s0 += ((float)(((bits4.s2 & 0x00F0) >> 4) | ((bits2.s2 & 0x0C) << 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s5; \
total_sum.s0 += ((float)(((bits4.s2 & 0x0F00) >> 8) | ((bits2.s2 & 0x30) )) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s6; \
total_sum.s0 += ((float)(((bits4.s2 & 0xF000) >> 12) | ((bits2.s2 & 0xC0) >> 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s7; \
total_sum.s1 += ((float)(((bits4.s1 & 0x000F) ) | ((bits2.s1 & 0x03) << 4)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s0; \
total_sum.s1 += ((float)(((bits4.s1 & 0x00F0) >> 4) | ((bits2.s1 & 0x0C) << 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s1; \
total_sum.s1 += ((float)(((bits4.s1 & 0x0F00) >> 8) | ((bits2.s1 & 0x30) )) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s2; \
total_sum.s1 += ((float)(((bits4.s1 & 0xF000) >> 12) | ((bits2.s1 & 0xC0) >> 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s3; \
total_sum.s1 += ((float)(((bits4.s3 & 0x000F) ) | ((bits2.s3 & 0x03) << 4)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s4; \
total_sum.s1 += ((float)(((bits4.s3 & 0x00F0) >> 4) | ((bits2.s3 & 0x0C) << 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s5; \
total_sum.s1 += ((float)(((bits4.s3 & 0x0F00) >> 8) | ((bits2.s3 & 0x30) )) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s6; \
total_sum.s1 += ((float)(((bits4.s3 & 0xF000) >> 12) | ((bits2.s3 & 0xC0) >> 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s7; \
shared_y = sub_group_broadcast(y, 1); \
total_sum.s0 += ((float)(((bits4.s4 & 0x000F) ) | ((bits2.s4 & 0x03) << 4)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s0; \
total_sum.s0 += ((float)(((bits4.s4 & 0x00F0) >> 4) | ((bits2.s4 & 0x0C) << 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s1; \
total_sum.s0 += ((float)(((bits4.s4 & 0x0F00) >> 8) | ((bits2.s4 & 0x30) )) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s2; \
total_sum.s0 += ((float)(((bits4.s4 & 0xF000) >> 12) | ((bits2.s4 & 0xC0) >> 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s3; \
total_sum.s0 += ((float)(((bits4.s6 & 0x000F) ) | ((bits2.s6 & 0x03) << 4)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s4; \
total_sum.s0 += ((float)(((bits4.s6 & 0x00F0) >> 4) | ((bits2.s6 & 0x0C) << 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s5; \
total_sum.s0 += ((float)(((bits4.s6 & 0x0F00) >> 8) | ((bits2.s6 & 0x30) )) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s6; \
total_sum.s0 += ((float)(((bits4.s6 & 0xF000) >> 12) | ((bits2.s6 & 0xC0) >> 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y.s7; \
total_sum.s1 += ((float)(((bits4.s5 & 0x000F) ) | ((bits2.s5 & 0x03) << 4)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s0; \
total_sum.s1 += ((float)(((bits4.s5 & 0x00F0) >> 4) | ((bits2.s5 & 0x0C) << 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s1; \
total_sum.s1 += ((float)(((bits4.s5 & 0x0F00) >> 8) | ((bits2.s5 & 0x30) )) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s2; \
total_sum.s1 += ((float)(((bits4.s5 & 0xF000) >> 12) | ((bits2.s5 & 0xC0) >> 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s3; \
total_sum.s1 += ((float)(((bits4.s7 & 0x000F) ) | ((bits2.s7 & 0x03) << 4)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s4; \
total_sum.s1 += ((float)(((bits4.s7 & 0x00F0) >> 4) | ((bits2.s7 & 0x0C) << 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s5; \
total_sum.s1 += ((float)(((bits4.s7 & 0x0F00) >> 8) | ((bits2.s7 & 0x30) )) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s6; \
total_sum.s1 += ((float)(((bits4.s7 & 0xF000) >> 12) | ((bits2.s7 & 0xC0) >> 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y.s7; \
#define dequantize_block_acc_bcast_8_lo(total_sum, bits4, bits2, scale_d, scale_s, y) \
shared_y = sub_group_broadcast(y, 2); \
total_sum.s0 += ((float)(((bits4.s0 & 0x000F) ) | ((bits2.s0 & 0x03) << 4)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s0; \
total_sum.s0 += ((float)(((bits4.s0 & 0x00F0) >> 4) | ((bits2.s0 & 0x0C) << 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s1; \
total_sum.s0 += ((float)(((bits4.s0 & 0x0F00) >> 8) | ((bits2.s0 & 0x30) )) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s2; \
total_sum.s0 += ((float)(((bits4.s0 & 0xF000) >> 12) | ((bits2.s0 & 0xC0) >> 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s3; \
total_sum.s0 += ((float)(((bits4.s2 & 0x000F) ) | ((bits2.s2 & 0x03) << 4)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s4; \
total_sum.s0 += ((float)(((bits4.s2 & 0x00F0) >> 4) | ((bits2.s2 & 0x0C) << 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s5; \
total_sum.s0 += ((float)(((bits4.s2 & 0x0F00) >> 8) | ((bits2.s2 & 0x30) )) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s6; \
total_sum.s0 += ((float)(((bits4.s2 & 0xF000) >> 12) | ((bits2.s2 & 0xC0) >> 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s7; \
total_sum.s1 += ((float)(((bits4.s1 & 0x000F) ) | ((bits2.s1 & 0x03) << 4)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s0; \
total_sum.s1 += ((float)(((bits4.s1 & 0x00F0) >> 4) | ((bits2.s1 & 0x0C) << 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s1; \
total_sum.s1 += ((float)(((bits4.s1 & 0x0F00) >> 8) | ((bits2.s1 & 0x30) )) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s2; \
total_sum.s1 += ((float)(((bits4.s1 & 0xF000) >> 12) | ((bits2.s1 & 0xC0) >> 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s3; \
total_sum.s1 += ((float)(((bits4.s3 & 0x000F) ) | ((bits2.s3 & 0x03) << 4)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s4; \
total_sum.s1 += ((float)(((bits4.s3 & 0x00F0) >> 4) | ((bits2.s3 & 0x0C) << 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s5; \
total_sum.s1 += ((float)(((bits4.s3 & 0x0F00) >> 8) | ((bits2.s3 & 0x30) )) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s6; \
total_sum.s1 += ((float)(((bits4.s3 & 0xF000) >> 12) | ((bits2.s3 & 0xC0) >> 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s7; \
shared_y = sub_group_broadcast(y, 3); \
total_sum.s0 += ((float)(((bits4.s4 & 0x000F) ) | ((bits2.s4 & 0x03) << 4)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s0; \
total_sum.s0 += ((float)(((bits4.s4 & 0x00F0) >> 4) | ((bits2.s4 & 0x0C) << 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s1; \
total_sum.s0 += ((float)(((bits4.s4 & 0x0F00) >> 8) | ((bits2.s4 & 0x30) )) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s2; \
total_sum.s0 += ((float)(((bits4.s4 & 0xF000) >> 12) | ((bits2.s4 & 0xC0) >> 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s3; \
total_sum.s0 += ((float)(((bits4.s6 & 0x000F) ) | ((bits2.s6 & 0x03) << 4)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s4; \
total_sum.s0 += ((float)(((bits4.s6 & 0x00F0) >> 4) | ((bits2.s6 & 0x0C) << 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s5; \
total_sum.s0 += ((float)(((bits4.s6 & 0x0F00) >> 8) | ((bits2.s6 & 0x30) )) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s6; \
total_sum.s0 += ((float)(((bits4.s6 & 0xF000) >> 12) | ((bits2.s6 & 0xC0) >> 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y.s7; \
total_sum.s1 += ((float)(((bits4.s5 & 0x000F) ) | ((bits2.s5 & 0x03) << 4)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s0; \
total_sum.s1 += ((float)(((bits4.s5 & 0x00F0) >> 4) | ((bits2.s5 & 0x0C) << 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s1; \
total_sum.s1 += ((float)(((bits4.s5 & 0x0F00) >> 8) | ((bits2.s5 & 0x30) )) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s2; \
total_sum.s1 += ((float)(((bits4.s5 & 0xF000) >> 12) | ((bits2.s5 & 0xC0) >> 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s3; \
total_sum.s1 += ((float)(((bits4.s7 & 0x000F) ) | ((bits2.s7 & 0x03) << 4)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s4; \
total_sum.s1 += ((float)(((bits4.s7 & 0x00F0) >> 4) | ((bits2.s7 & 0x0C) << 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s5; \
total_sum.s1 += ((float)(((bits4.s7 & 0x0F00) >> 8) | ((bits2.s7 & 0x30) )) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s6; \
total_sum.s1 += ((float)(((bits4.s7 & 0xF000) >> 12) | ((bits2.s7 & 0xC0) >> 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y.s7; \
#define dequantize_block_acc_bcast_1_hi(total_sum, bits4, bits2, scale_d, scale_s, y) \
float shared_y; \
shared_y = sub_group_broadcast(y.s0, 0); \
total_sum.s0 += ((float)(((bits4.s0 & 0x000F) ) | ((bits2.s0 & 0x03) << 4)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s1 & 0x000F) ) | ((bits2.s1 & 0x03) << 4)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
shared_y = sub_group_broadcast(y.s1, 0); \
total_sum.s0 += ((float)(((bits4.s0 & 0x00F0) >> 4) | ((bits2.s0 & 0x0C) << 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s1 & 0x00F0) >> 4) | ((bits2.s1 & 0x0C) << 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
shared_y = sub_group_broadcast(y.s2, 0); \
total_sum.s0 += ((float)(((bits4.s0 & 0x0F00) >> 8) | ((bits2.s0 & 0x30) )) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s1 & 0x0F00) >> 8) | ((bits2.s1 & 0x30) )) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
shared_y = sub_group_broadcast(y.s3, 0); \
total_sum.s0 += ((float)(((bits4.s0 & 0xF000) >> 12) | ((bits2.s0 & 0xC0) >> 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s1 & 0xF000) >> 12) | ((bits2.s1 & 0xC0) >> 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
shared_y = sub_group_broadcast(y.s4, 0); \
total_sum.s0 += ((float)(((bits4.s2 & 0x000F) ) | ((bits2.s2 & 0x03) << 4)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s3 & 0x000F) ) | ((bits2.s3 & 0x03) << 4)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
shared_y = sub_group_broadcast(y.s5, 0); \
total_sum.s0 += ((float)(((bits4.s2 & 0x00F0) >> 4) | ((bits2.s2 & 0x0C) << 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s3 & 0x00F0) >> 4) | ((bits2.s3 & 0x0C) << 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
shared_y = sub_group_broadcast(y.s6, 0); \
total_sum.s0 += ((float)(((bits4.s2 & 0x0F00) >> 8) | ((bits2.s2 & 0x30) )) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s3 & 0x0F00) >> 8) | ((bits2.s3 & 0x30) )) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
shared_y = sub_group_broadcast(y.s7, 0); \
total_sum.s0 += ((float)(((bits4.s2 & 0xF000) >> 12) | ((bits2.s2 & 0xC0) >> 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s3 & 0xF000) >> 12) | ((bits2.s3 & 0xC0) >> 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
shared_y = sub_group_broadcast(y.s0, 1); \
total_sum.s0 += ((float)(((bits4.s4 & 0x000F) ) | ((bits2.s4 & 0x03) << 4)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s5 & 0x000F) ) | ((bits2.s5 & 0x03) << 4)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
shared_y = sub_group_broadcast(y.s1, 1); \
total_sum.s0 += ((float)(((bits4.s4 & 0x00F0) >> 4) | ((bits2.s4 & 0x0C) << 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s5 & 0x00F0) >> 4) | ((bits2.s5 & 0x0C) << 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
shared_y = sub_group_broadcast(y.s2, 1); \
total_sum.s0 += ((float)(((bits4.s4 & 0x0F00) >> 8) | ((bits2.s4 & 0x30) )) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s5 & 0x0F00) >> 8) | ((bits2.s5 & 0x30) )) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
shared_y = sub_group_broadcast(y.s3, 1); \
total_sum.s0 += ((float)(((bits4.s4 & 0xF000) >> 12) | ((bits2.s4 & 0xC0) >> 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s5 & 0xF000) >> 12) | ((bits2.s5 & 0xC0) >> 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
shared_y = sub_group_broadcast(y.s4, 1); \
total_sum.s0 += ((float)(((bits4.s6 & 0x000F) ) | ((bits2.s6 & 0x03) << 4)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s7 & 0x000F) ) | ((bits2.s7 & 0x03) << 4)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
shared_y = sub_group_broadcast(y.s5, 1); \
total_sum.s0 += ((float)(((bits4.s6 & 0x00F0) >> 4) | ((bits2.s6 & 0x0C) << 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s7 & 0x00F0) >> 4) | ((bits2.s7 & 0x0C) << 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
shared_y = sub_group_broadcast(y.s6, 1); \
total_sum.s0 += ((float)(((bits4.s6 & 0x0F00) >> 8) | ((bits2.s6 & 0x30) )) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s7 & 0x0F00) >> 8) | ((bits2.s7 & 0x30) )) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
shared_y = sub_group_broadcast(y.s7, 1); \
total_sum.s0 += ((float)(((bits4.s6 & 0xF000) >> 12) | ((bits2.s6 & 0xC0) >> 2)) - 32.f) * scale_s.s0 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s7 & 0xF000) >> 12) | ((bits2.s7 & 0xC0) >> 2)) - 32.f) * scale_s.s2 * scale_d.s1 * shared_y; \
#define dequantize_block_acc_bcast_1_lo(total_sum, bits4, bits2, scale_d, scale_s, y) \
shared_y = sub_group_broadcast(y.s0, 2); \
total_sum.s0 += ((float)(((bits4.s0 & 0x000F) ) | ((bits2.s0 & 0x03) << 4)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s1 & 0x000F) ) | ((bits2.s1 & 0x03) << 4)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
shared_y = sub_group_broadcast(y.s1, 2); \
total_sum.s0 += ((float)(((bits4.s0 & 0x00F0) >> 4) | ((bits2.s0 & 0x0C) << 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s1 & 0x00F0) >> 4) | ((bits2.s1 & 0x0C) << 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
shared_y = sub_group_broadcast(y.s2, 2); \
total_sum.s0 += ((float)(((bits4.s0 & 0x0F00) >> 8) | ((bits2.s0 & 0x30) )) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s1 & 0x0F00) >> 8) | ((bits2.s1 & 0x30) )) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
shared_y = sub_group_broadcast(y.s3, 2); \
total_sum.s0 += ((float)(((bits4.s0 & 0xF000) >> 12) | ((bits2.s0 & 0xC0) >> 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s1 & 0xF000) >> 12) | ((bits2.s1 & 0xC0) >> 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
shared_y = sub_group_broadcast(y.s4, 2); \
total_sum.s0 += ((float)(((bits4.s2 & 0x000F) ) | ((bits2.s2 & 0x03) << 4)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s3 & 0x000F) ) | ((bits2.s3 & 0x03) << 4)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
shared_y = sub_group_broadcast(y.s5, 2); \
total_sum.s0 += ((float)(((bits4.s2 & 0x00F0) >> 4) | ((bits2.s2 & 0x0C) << 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s3 & 0x00F0) >> 4) | ((bits2.s3 & 0x0C) << 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
shared_y = sub_group_broadcast(y.s6, 2); \
total_sum.s0 += ((float)(((bits4.s2 & 0x0F00) >> 8) | ((bits2.s2 & 0x30) )) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s3 & 0x0F00) >> 8) | ((bits2.s3 & 0x30) )) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
shared_y = sub_group_broadcast(y.s7, 2); \
total_sum.s0 += ((float)(((bits4.s2 & 0xF000) >> 12) | ((bits2.s2 & 0xC0) >> 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s3 & 0xF000) >> 12) | ((bits2.s3 & 0xC0) >> 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
shared_y = sub_group_broadcast(y.s0, 3); \
total_sum.s0 += ((float)(((bits4.s4 & 0x000F) ) | ((bits2.s4 & 0x03) << 4)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s5 & 0x000F) ) | ((bits2.s5 & 0x03) << 4)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
shared_y = sub_group_broadcast(y.s1, 3); \
total_sum.s0 += ((float)(((bits4.s4 & 0x00F0) >> 4) | ((bits2.s4 & 0x0C) << 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s5 & 0x00F0) >> 4) | ((bits2.s5 & 0x0C) << 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
shared_y = sub_group_broadcast(y.s2, 3); \
total_sum.s0 += ((float)(((bits4.s4 & 0x0F00) >> 8) | ((bits2.s4 & 0x30) )) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s5 & 0x0F00) >> 8) | ((bits2.s5 & 0x30) )) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
shared_y = sub_group_broadcast(y.s3, 3); \
total_sum.s0 += ((float)(((bits4.s4 & 0xF000) >> 12) | ((bits2.s4 & 0xC0) >> 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s5 & 0xF000) >> 12) | ((bits2.s5 & 0xC0) >> 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
shared_y = sub_group_broadcast(y.s4, 3); \
total_sum.s0 += ((float)(((bits4.s6 & 0x000F) ) | ((bits2.s6 & 0x03) << 4)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s7 & 0x000F) ) | ((bits2.s7 & 0x03) << 4)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
shared_y = sub_group_broadcast(y.s5, 3); \
total_sum.s0 += ((float)(((bits4.s6 & 0x00F0) >> 4) | ((bits2.s6 & 0x0C) << 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s7 & 0x00F0) >> 4) | ((bits2.s7 & 0x0C) << 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
shared_y = sub_group_broadcast(y.s6, 3); \
total_sum.s0 += ((float)(((bits4.s6 & 0x0F00) >> 8) | ((bits2.s6 & 0x30) )) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s7 & 0x0F00) >> 8) | ((bits2.s7 & 0x30) )) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
shared_y = sub_group_broadcast(y.s7, 3); \
total_sum.s0 += ((float)(((bits4.s6 & 0xF000) >> 12) | ((bits2.s6 & 0xC0) >> 2)) - 32.f) * scale_s.s1 * scale_d.s0 * shared_y; \
total_sum.s1 += ((float)(((bits4.s7 & 0xF000) >> 12) | ((bits2.s7 & 0xC0) >> 2)) - 32.f) * scale_s.s3 * scale_d.s1 * shared_y; \
#if defined(ADRENO_GPU)
REQD_SUBGROUP_SIZE_64
#endif
kernel void kernel_gemv_noshuffle_q6_K_f32(
read_only image1d_buffer_t src0_ql,
read_only image1d_buffer_t src0_qh,
global half2 * src0_s,
global half2 * src0_d,
read_only image1d_buffer_t src1,
global float * dst,
ulong offsetd,
int ne00,
int ne01
) {
int grp = get_local_id(1);
int gid = get_global_id(0);
ushort slid = get_sub_group_local_id();
int nb = ne00 / 32;
uint4 reg_a_l;
ushort4 reg_a_h;
half2 reg_d;
char4 reg_s;
float8 reg_b;
float2 total_sum = 0.0f;
int line_stride_a = ne01 / 2;
int block_stride_a = NSUBGROUPS * ne01;
for (int k = grp; k < nb; k += NSUBGROUPS) {
reg_d = src0_d[gid + k/8 * line_stride_a];
reg_s = as_char4(src0_s[gid + k * line_stride_a]);
if (slid < 4) {
reg_b.s0123 = read_imagef(src1, 0 + slid*2 + k*8);
reg_b.s4567 = read_imagef(src1, 1 + slid*2 + k*8);
}
reg_a_l.s0 = read_imageui(src0_ql, gid + k*block_stride_a + line_stride_a*0).x;
reg_a_l.s1 = read_imageui(src0_ql, gid + k*block_stride_a + line_stride_a*1).x;
reg_a_l.s2 = read_imageui(src0_ql, gid + k*block_stride_a + line_stride_a*2).x;
reg_a_l.s3 = read_imageui(src0_ql, gid + k*block_stride_a + line_stride_a*3).x;
reg_a_h.s0 = as_ushort(read_imageh(src0_qh, gid + k*block_stride_a + line_stride_a*0).x);
reg_a_h.s1 = as_ushort(read_imageh(src0_qh, gid + k*block_stride_a + line_stride_a*1).x);
reg_a_h.s2 = as_ushort(read_imageh(src0_qh, gid + k*block_stride_a + line_stride_a*2).x);
reg_a_h.s3 = as_ushort(read_imageh(src0_qh, gid + k*block_stride_a + line_stride_a*3).x);
#ifdef VECTOR_SUB_GROUP_BROADCAT
dequantize_block_acc_bcast_8_hi(total_sum, as_ushort8(reg_a_l), as_uchar8(reg_a_h), reg_d, reg_s, reg_b);
#else
dequantize_block_acc_bcast_1_hi(total_sum, as_ushort8(reg_a_l), as_uchar8(reg_a_h), reg_d, reg_s, reg_b);
#endif // VECTOR_SUB_GROUP_BROADCAT
reg_a_l.s0 = read_imageui(src0_ql, gid + k*block_stride_a + line_stride_a*4).x;
reg_a_l.s1 = read_imageui(src0_ql, gid + k*block_stride_a + line_stride_a*5).x;
reg_a_l.s2 = read_imageui(src0_ql, gid + k*block_stride_a + line_stride_a*6).x;
reg_a_l.s3 = read_imageui(src0_ql, gid + k*block_stride_a + line_stride_a*7).x;
reg_a_h.s0 = as_ushort(read_imageh(src0_qh, gid + k*block_stride_a + line_stride_a*4).x);
reg_a_h.s1 = as_ushort(read_imageh(src0_qh, gid + k*block_stride_a + line_stride_a*5).x);
reg_a_h.s2 = as_ushort(read_imageh(src0_qh, gid + k*block_stride_a + line_stride_a*6).x);
reg_a_h.s3 = as_ushort(read_imageh(src0_qh, gid + k*block_stride_a + line_stride_a*7).x);
#ifdef VECTOR_SUB_GROUP_BROADCAT
dequantize_block_acc_bcast_8_lo(total_sum, as_ushort8(reg_a_l), as_uchar8(reg_a_h), reg_d, reg_s, reg_b);
#else
dequantize_block_acc_bcast_1_lo(total_sum, as_ushort8(reg_a_l), as_uchar8(reg_a_h), reg_d, reg_s, reg_b);
#endif // VECTOR_SUB_GROUP_BROADCAT
}
local float2 reduce_lm[SUBGROUP_SIZE * 3];
if (grp == 1) {
reduce_lm[SUBGROUP_SIZE*0 + slid] = total_sum;
}
if (grp == 2) {
reduce_lm[SUBGROUP_SIZE*1 + slid] = total_sum;
}
if (grp == 3) {
reduce_lm[SUBGROUP_SIZE*2 + slid] = total_sum;
}
barrier(CLK_LOCAL_MEM_FENCE);
if (grp == 0) {
total_sum += reduce_lm[SUBGROUP_SIZE*0 + slid];
}
if (grp == 0) {
total_sum += reduce_lm[SUBGROUP_SIZE*1 + slid];
}
if (grp == 0) {
total_sum += reduce_lm[SUBGROUP_SIZE*2 + slid];
}
if (grp == 0) {
dst = (global float*)((global char*)dst + offsetd);
vstore2(total_sum, 0, &(dst[gid * 2]));
}
}