opencl: add q5_0/q5_1 gemm and gemv kernels for Adreno (#24319)
* opencl: add q5_0 adreno support * opencl: add q5_1 adreno support * opencl: cosmetic fix --------- Co-authored-by: Li He <lih@qti.qualcomm.com>
This commit is contained in:
shaofeiqi
@@ -584,6 +584,60 @@ kernel void kernel_restore_block_q5_0(
|
||||
}
|
||||
}
|
||||
|
||||
kernel void kernel_convert_block_q5_0_noshuffle(
|
||||
global struct block_q5_0 * src0,
|
||||
global uchar * dst_q,
|
||||
global uint * dst_qh,
|
||||
global half * dst_d
|
||||
) {
|
||||
global struct block_q5_0 * b = (global struct block_q5_0 *) src0 + get_global_id(0);
|
||||
global uchar * q = (global uchar *) dst_q + QK5_0/2*get_global_id(0);
|
||||
global uint * qh = (global uint *) dst_qh + get_global_id(0);
|
||||
global half * d = (global half *) dst_d + get_global_id(0);
|
||||
|
||||
*d = b->d;
|
||||
*qh = *((global uint *)(b->qh));
|
||||
|
||||
for (int i = 0; i < QK5_0/4; ++i) {
|
||||
uchar x0 = b->qs[2*i + 0];
|
||||
uchar x1 = b->qs[2*i + 1];
|
||||
|
||||
q[i + 0 ] = convert_uchar(x0 & 0x0F) | convert_uchar((x1 & 0x0F) << 4);
|
||||
q[i + QK5_0/4] = convert_uchar((x0 & 0xF0) >> 4) | convert_uchar(x1 & 0xF0);
|
||||
|
||||
#ifdef ADRENO_GPU
|
||||
if (get_global_id(0) == 65536*4096) {
|
||||
printf("%04x - %02x\n", *(global ushort*)d, ((x0 & 0xF0) >> 4) | (x1 & 0xF0));
|
||||
}
|
||||
#endif
|
||||
}
|
||||
}
|
||||
|
||||
kernel void kernel_restore_block_q5_0_noshuffle(
|
||||
global uchar * src_q,
|
||||
global uint * src_qh,
|
||||
global half * src_d,
|
||||
global struct block_q5_0 * dst,
|
||||
uchar mask_0F,
|
||||
uchar mask_F0
|
||||
) {
|
||||
global struct block_q5_0 * b = (global struct block_q5_0 *) dst + get_global_id(0);
|
||||
global uchar * q = (global uchar *) src_q + QK5_0/2*get_global_id(0);
|
||||
global uint * qh = (global uint *) src_qh + get_global_id(0);
|
||||
global half * d = (global half *) src_d + get_global_id(0);
|
||||
|
||||
b->d = *d;
|
||||
*((global uint *)(b->qh)) = *qh;
|
||||
|
||||
for (int i = 0; i < QK5_0/4; ++i) {
|
||||
uchar x0 = q[i + 0 ];
|
||||
uchar x1 = q[i + QK5_0/4];
|
||||
|
||||
b->qs[2*i + 0] = convert_uchar((x0 & mask_0F) | ((x1 & mask_0F) << 4));
|
||||
b->qs[2*i + 1] = convert_uchar(((x0 & mask_F0) >> 4) | (x1 & mask_F0));
|
||||
}
|
||||
}
|
||||
|
||||
kernel void kernel_convert_block_q5_0_trans4_ns(
|
||||
__global struct block_q5_0 * src0,
|
||||
__global uint * dst_qs,
|
||||
@@ -736,6 +790,66 @@ kernel void kernel_restore_block_q5_1(
|
||||
}
|
||||
}
|
||||
|
||||
kernel void kernel_convert_block_q5_1_noshuffle(
|
||||
global struct block_q5_1 * src0,
|
||||
global uchar * dst_q,
|
||||
global uint * dst_qh,
|
||||
global half * dst_d,
|
||||
global half * dst_m
|
||||
) {
|
||||
global struct block_q5_1 * b = (global struct block_q5_1 *) src0 + get_global_id(0);
|
||||
global uchar * q = (global uchar *) dst_q + QK5_1/2*get_global_id(0);
|
||||
global uint * qh = (global uint *) dst_qh + get_global_id(0);
|
||||
global half * d = (global half *) dst_d + get_global_id(0);
|
||||
global half * m = (global half *) dst_m + get_global_id(0);
|
||||
|
||||
*d = b->d;
|
||||
*m = b->m;
|
||||
*qh = *((global uint *)(b->qh));
|
||||
|
||||
for (int i = 0; i < QK5_1/4; ++i) {
|
||||
uchar x0 = b->qs[2*i + 0];
|
||||
uchar x1 = b->qs[2*i + 1];
|
||||
|
||||
q[i + 0 ] = convert_uchar(x0 & 0x0F) | convert_uchar((x1 & 0x0F) << 4);
|
||||
q[i + QK5_1/4] = convert_uchar((x0 & 0xF0) >> 4) | convert_uchar(x1 & 0xF0);
|
||||
|
||||
#ifdef ADRENO_GPU
|
||||
if (get_global_id(0) == 65536*4096) {
|
||||
printf("%04x - %02x\n", *(global ushort*)d, ((x0 & 0xF0) >> 4) | (x1 & 0xF0));
|
||||
}
|
||||
#endif
|
||||
}
|
||||
}
|
||||
|
||||
kernel void kernel_restore_block_q5_1_noshuffle(
|
||||
global uchar * src_q,
|
||||
global uint * src_qh,
|
||||
global half * src_d,
|
||||
global half * src_m,
|
||||
global struct block_q5_1 * dst,
|
||||
uchar mask_0F,
|
||||
uchar mask_F0
|
||||
) {
|
||||
global struct block_q5_1 * b = (global struct block_q5_1 *) dst + get_global_id(0);
|
||||
global uchar * q = (global uchar *) src_q + QK5_1/2*get_global_id(0);
|
||||
global uint * qh = (global uint *) src_qh + get_global_id(0);
|
||||
global half * d = (global half *) src_d + get_global_id(0);
|
||||
global half * m = (global half *) src_m + get_global_id(0);
|
||||
|
||||
b->d = *d;
|
||||
b->m = *m;
|
||||
*((global uint *)(b->qh)) = *qh;
|
||||
|
||||
for (int i = 0; i < QK5_1/4; ++i) {
|
||||
uchar x0 = q[i + 0 ];
|
||||
uchar x1 = q[i + QK5_1/4];
|
||||
|
||||
b->qs[2*i + 0] = convert_uchar((x0 & mask_0F) | ((x1 & mask_0F) << 4));
|
||||
b->qs[2*i + 1] = convert_uchar(((x0 & mask_F0) >> 4) | (x1 & mask_F0));
|
||||
}
|
||||
}
|
||||
|
||||
kernel void kernel_convert_block_q5_1_trans4_ns(
|
||||
__global struct block_q5_1 * src0,
|
||||
__global uint * dst_qs,
|
||||
|
||||
@@ -0,0 +1,131 @@
|
||||
#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_q5_0_f32(
|
||||
global const ushort * src0_qs, // quantized A
|
||||
global const uchar * src0_qh, // 5th bits
|
||||
global const half * src0_d, // A scales
|
||||
__read_only image1d_buffer_t src1, // B (1d image)
|
||||
global float * dst, // C
|
||||
int m, // M
|
||||
int n, // N with padding
|
||||
int k, // K
|
||||
int n_no_padding // N without padding
|
||||
) {
|
||||
|
||||
int n_4 = n >> 2;
|
||||
|
||||
int gy = get_global_id(0);
|
||||
int gx = get_global_id(1);
|
||||
int gx_2 = gx << 2;
|
||||
|
||||
half8 c0 = 0, c1 = 0, c2 = 0, c3 = 0;
|
||||
half8 B;
|
||||
half4 dequantized_weights;
|
||||
|
||||
global const ushort * weight_ptr = src0_qs + gx_2;
|
||||
global const uchar * qh_ptr = src0_qh + gx_2;
|
||||
global const half * scale_ptr = src0_d + gx_2;
|
||||
|
||||
for (int i = 0; i < k; i += 4) {
|
||||
|
||||
B.s0123 = read_imageh(src1, gy*2 + i*n_4);
|
||||
B.s4567 = read_imageh(src1, gy*2 + i*n_4 + 1);
|
||||
|
||||
ushort4 bits4 = vload4(0, weight_ptr + (i >> 2)*m);
|
||||
uchar4 bits1 = vload4(0, qh_ptr + (i >> 3)*m);
|
||||
uchar4 qh = bits1 >> (uchar4)(i & 4);
|
||||
|
||||
half4 scale = vload4(0, scale_ptr + (i >> 5)*m);
|
||||
|
||||
// j=0
|
||||
dequantized_weights.s0 = (convert_half((bits4.s0 & 0x000F) | ((qh.s0 & 0x01) << 4)) - 16.0h) * scale.s0;
|
||||
dequantized_weights.s1 = (convert_half((bits4.s1 & 0x000F) | ((qh.s1 & 0x01) << 4)) - 16.0h) * scale.s1;
|
||||
dequantized_weights.s2 = (convert_half((bits4.s2 & 0x000F) | ((qh.s2 & 0x01) << 4)) - 16.0h) * scale.s2;
|
||||
dequantized_weights.s3 = (convert_half((bits4.s3 & 0x000F) | ((qh.s3 & 0x01) << 4)) - 16.0h) * scale.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);
|
||||
B.s4567 = read_imageh(src1, gy*2 + (i+1)*n_4 + 1);
|
||||
dequantized_weights.s0 = (convert_half(((bits4.s0 & 0x00F0) >> 4) | ((qh.s0 & 0x02) << 3)) - 16.0h) * scale.s0;
|
||||
dequantized_weights.s1 = (convert_half(((bits4.s1 & 0x00F0) >> 4) | ((qh.s1 & 0x02) << 3)) - 16.0h) * scale.s1;
|
||||
dequantized_weights.s2 = (convert_half(((bits4.s2 & 0x00F0) >> 4) | ((qh.s2 & 0x02) << 3)) - 16.0h) * scale.s2;
|
||||
dequantized_weights.s3 = (convert_half(((bits4.s3 & 0x00F0) >> 4) | ((qh.s3 & 0x02) << 3)) - 16.0h) * scale.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);
|
||||
B.s4567 = read_imageh(src1, gy*2 + (i+2)*n_4 + 1);
|
||||
dequantized_weights.s0 = (convert_half(((bits4.s0 & 0x0F00) >> 8) | ((qh.s0 & 0x04) << 2)) - 16.0h) * scale.s0;
|
||||
dequantized_weights.s1 = (convert_half(((bits4.s1 & 0x0F00) >> 8) | ((qh.s1 & 0x04) << 2)) - 16.0h) * scale.s1;
|
||||
dequantized_weights.s2 = (convert_half(((bits4.s2 & 0x0F00) >> 8) | ((qh.s2 & 0x04) << 2)) - 16.0h) * scale.s2;
|
||||
dequantized_weights.s3 = (convert_half(((bits4.s3 & 0x0F00) >> 8) | ((qh.s3 & 0x04) << 2)) - 16.0h) * scale.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);
|
||||
B.s4567 = read_imageh(src1, gy*2 + (i+3)*n_4 + 1);
|
||||
dequantized_weights.s0 = (convert_half(((bits4.s0 & 0xF000) >> 12) | ((qh.s0 & 0x08) << 1)) - 16.0h) * scale.s0;
|
||||
dequantized_weights.s1 = (convert_half(((bits4.s1 & 0xF000) >> 12) | ((qh.s1 & 0x08) << 1)) - 16.0h) * scale.s1;
|
||||
dequantized_weights.s2 = (convert_half(((bits4.s2 & 0xF000) >> 12) | ((qh.s2 & 0x08) << 1)) - 16.0h) * scale.s2;
|
||||
dequantized_weights.s3 = (convert_half(((bits4.s3 & 0xF000) >> 12) | ((qh.s3 & 0x08) << 1)) - 16.0h) * scale.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);
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,134 @@
|
||||
#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_q5_1_f32(
|
||||
global const ushort * src0_qs, // quantized A
|
||||
global const uchar * src0_qh, // 5th bits
|
||||
global const half * src0_d, // A scales
|
||||
global const half * src0_m, // A mins
|
||||
__read_only image1d_buffer_t src1, // B (1d image)
|
||||
global float * dst, // C
|
||||
int m, // M
|
||||
int n, // N with padding
|
||||
int k, // K
|
||||
int n_no_padding // N without padding
|
||||
) {
|
||||
|
||||
int n_4 = n >> 2;
|
||||
|
||||
int gy = get_global_id(0);
|
||||
int gx = get_global_id(1);
|
||||
int gx_2 = gx << 2;
|
||||
|
||||
half8 c0 = 0, c1 = 0, c2 = 0, c3 = 0;
|
||||
half8 B;
|
||||
half4 dequantized_weights;
|
||||
|
||||
global const ushort * weight_ptr = src0_qs + gx_2;
|
||||
global const uchar * qh_ptr = src0_qh + gx_2;
|
||||
global const half * scale_ptr = src0_d + gx_2;
|
||||
global const half * min_ptr = src0_m + gx_2;
|
||||
|
||||
for (int i = 0; i < k; i += 4) {
|
||||
|
||||
B.s0123 = read_imageh(src1, gy*2 + i*n_4);
|
||||
B.s4567 = read_imageh(src1, gy*2 + i*n_4 + 1);
|
||||
|
||||
ushort4 bits4 = vload4(0, weight_ptr + (i >> 2)*m);
|
||||
uchar4 bits1 = vload4(0, qh_ptr + (i >> 3)*m);
|
||||
uchar4 qh = bits1 >> (uchar4)(i & 4);
|
||||
|
||||
half4 scale = vload4(0, scale_ptr + (i >> 5)*m);
|
||||
half4 minv = vload4(0, min_ptr + (i >> 5)*m);
|
||||
|
||||
// j=0
|
||||
dequantized_weights.s0 = convert_half((bits4.s0 & 0x000F) | ((qh.s0 & 0x01) << 4)) * scale.s0 + minv.s0;
|
||||
dequantized_weights.s1 = convert_half((bits4.s1 & 0x000F) | ((qh.s1 & 0x01) << 4)) * scale.s1 + minv.s1;
|
||||
dequantized_weights.s2 = convert_half((bits4.s2 & 0x000F) | ((qh.s2 & 0x01) << 4)) * scale.s2 + minv.s2;
|
||||
dequantized_weights.s3 = convert_half((bits4.s3 & 0x000F) | ((qh.s3 & 0x01) << 4)) * scale.s3 + minv.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);
|
||||
B.s4567 = read_imageh(src1, gy*2 + (i+1)*n_4 + 1);
|
||||
dequantized_weights.s0 = convert_half(((bits4.s0 & 0x00F0) >> 4) | ((qh.s0 & 0x02) << 3)) * scale.s0 + minv.s0;
|
||||
dequantized_weights.s1 = convert_half(((bits4.s1 & 0x00F0) >> 4) | ((qh.s1 & 0x02) << 3)) * scale.s1 + minv.s1;
|
||||
dequantized_weights.s2 = convert_half(((bits4.s2 & 0x00F0) >> 4) | ((qh.s2 & 0x02) << 3)) * scale.s2 + minv.s2;
|
||||
dequantized_weights.s3 = convert_half(((bits4.s3 & 0x00F0) >> 4) | ((qh.s3 & 0x02) << 3)) * scale.s3 + minv.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);
|
||||
B.s4567 = read_imageh(src1, gy*2 + (i+2)*n_4 + 1);
|
||||
dequantized_weights.s0 = convert_half(((bits4.s0 & 0x0F00) >> 8) | ((qh.s0 & 0x04) << 2)) * scale.s0 + minv.s0;
|
||||
dequantized_weights.s1 = convert_half(((bits4.s1 & 0x0F00) >> 8) | ((qh.s1 & 0x04) << 2)) * scale.s1 + minv.s1;
|
||||
dequantized_weights.s2 = convert_half(((bits4.s2 & 0x0F00) >> 8) | ((qh.s2 & 0x04) << 2)) * scale.s2 + minv.s2;
|
||||
dequantized_weights.s3 = convert_half(((bits4.s3 & 0x0F00) >> 8) | ((qh.s3 & 0x04) << 2)) * scale.s3 + minv.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);
|
||||
B.s4567 = read_imageh(src1, gy*2 + (i+3)*n_4 + 1);
|
||||
dequantized_weights.s0 = convert_half(((bits4.s0 & 0xF000) >> 12) | ((qh.s0 & 0x08) << 1)) * scale.s0 + minv.s0;
|
||||
dequantized_weights.s1 = convert_half(((bits4.s1 & 0xF000) >> 12) | ((qh.s1 & 0x08) << 1)) * scale.s1 + minv.s1;
|
||||
dequantized_weights.s2 = convert_half(((bits4.s2 & 0xF000) >> 12) | ((qh.s2 & 0x08) << 1)) * scale.s2 + minv.s2;
|
||||
dequantized_weights.s3 = convert_half(((bits4.s3 & 0xF000) >> 12) | ((qh.s3 & 0x08) << 1)) * scale.s3 + minv.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);
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,291 @@
|
||||
#pragma OPENCL EXTENSION cl_khr_fp16 : enable
|
||||
#pragma OPENCL EXTENSION cl_khr_subgroups : 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_64 __attribute__((qcom_reqd_sub_group_size("half")))
|
||||
#endif
|
||||
|
||||
#define QK5_0 32
|
||||
#define NSUBGROUPS 4
|
||||
#define SUBGROUP_SIZE 64
|
||||
|
||||
#define dequantizeBlockAccum_ns_q5_0_sgbroadcast_1_hi(total_sums, bits4, bits1, scale, y) \
|
||||
float shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s0, 0); \
|
||||
total_sums.s0 += (((bits4.s0 & 0x000F) | (((bits1.s0 ) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += (((bits4.s1 & 0x000F) | (((bits1.s4 ) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s1, 0); \
|
||||
total_sums.s0 += ((((bits4.s0 & 0x00F0) >> 4) | (((bits1.s0 >> 1) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s1 & 0x00F0) >> 4) | (((bits1.s4 >> 1) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s2, 0); \
|
||||
total_sums.s0 += ((((bits4.s0 & 0x0F00) >> 8) | (((bits1.s0 >> 2) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s1 & 0x0F00) >> 8) | (((bits1.s4 >> 2) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s3, 0); \
|
||||
total_sums.s0 += ((((bits4.s0 & 0xF000) >> 12) | (((bits1.s0 >> 3) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s1 & 0xF000) >> 12) | (((bits1.s4 >> 3) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s4, 0); \
|
||||
total_sums.s0 += (((bits4.s2 & 0x000F) | (((bits1.s0 >> 4) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += (((bits4.s3 & 0x000F) | (((bits1.s4 >> 4) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s5, 0); \
|
||||
total_sums.s0 += ((((bits4.s2 & 0x00F0) >> 4) | (((bits1.s0 >> 5) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s3 & 0x00F0) >> 4) | (((bits1.s4 >> 5) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s6, 0); \
|
||||
total_sums.s0 += ((((bits4.s2 & 0x0F00) >> 8) | (((bits1.s0 >> 6) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s3 & 0x0F00) >> 8) | (((bits1.s4 >> 6) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s7, 0); \
|
||||
total_sums.s0 += ((((bits4.s2 & 0xF000) >> 12) | (((bits1.s0 >> 7) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s3 & 0xF000) >> 12) | (((bits1.s4 >> 7) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s0, 1); \
|
||||
total_sums.s0 += (((bits4.s4 & 0x000F) | (((bits1.s1 ) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += (((bits4.s5 & 0x000F) | (((bits1.s5 ) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s1, 1); \
|
||||
total_sums.s0 += ((((bits4.s4 & 0x00F0) >> 4) | (((bits1.s1 >> 1) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s5 & 0x00F0) >> 4) | (((bits1.s5 >> 1) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s2, 1); \
|
||||
total_sums.s0 += ((((bits4.s4 & 0x0F00) >> 8) | (((bits1.s1 >> 2) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s5 & 0x0F00) >> 8) | (((bits1.s5 >> 2) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s3, 1); \
|
||||
total_sums.s0 += ((((bits4.s4 & 0xF000) >> 12) | (((bits1.s1 >> 3) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s5 & 0xF000) >> 12) | (((bits1.s5 >> 3) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s4, 1); \
|
||||
total_sums.s0 += (((bits4.s6 & 0x000F) | (((bits1.s1 >> 4) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += (((bits4.s7 & 0x000F) | (((bits1.s5 >> 4) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s5, 1); \
|
||||
total_sums.s0 += ((((bits4.s6 & 0x00F0) >> 4) | (((bits1.s1 >> 5) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s7 & 0x00F0) >> 4) | (((bits1.s5 >> 5) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s6, 1); \
|
||||
total_sums.s0 += ((((bits4.s6 & 0x0F00) >> 8) | (((bits1.s1 >> 6) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s7 & 0x0F00) >> 8) | (((bits1.s5 >> 6) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s7, 1); \
|
||||
total_sums.s0 += ((((bits4.s6 & 0xF000) >> 12) | (((bits1.s1 >> 7) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s7 & 0xF000) >> 12) | (((bits1.s5 >> 7) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
|
||||
|
||||
#define dequantizeBlockAccum_ns_q5_0_sgbroadcast_1_lo(total_sums, bits4, bits1, scale, y) \
|
||||
shared_y = sub_group_broadcast(y.s0, 2); \
|
||||
total_sums.s0 += (((bits4.s0 & 0x000F) | (((bits1.s2 ) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += (((bits4.s1 & 0x000F) | (((bits1.s6 ) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s1, 2); \
|
||||
total_sums.s0 += ((((bits4.s0 & 0x00F0) >> 4) | (((bits1.s2 >> 1) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s1 & 0x00F0) >> 4) | (((bits1.s6 >> 1) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s2, 2); \
|
||||
total_sums.s0 += ((((bits4.s0 & 0x0F00) >> 8) | (((bits1.s2 >> 2) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s1 & 0x0F00) >> 8) | (((bits1.s6 >> 2) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s3, 2); \
|
||||
total_sums.s0 += ((((bits4.s0 & 0xF000) >> 12) | (((bits1.s2 >> 3) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s1 & 0xF000) >> 12) | (((bits1.s6 >> 3) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s4, 2); \
|
||||
total_sums.s0 += (((bits4.s2 & 0x000F) | (((bits1.s2 >> 4) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += (((bits4.s3 & 0x000F) | (((bits1.s6 >> 4) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s5, 2); \
|
||||
total_sums.s0 += ((((bits4.s2 & 0x00F0) >> 4) | (((bits1.s2 >> 5) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s3 & 0x00F0) >> 4) | (((bits1.s6 >> 5) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s6, 2); \
|
||||
total_sums.s0 += ((((bits4.s2 & 0x0F00) >> 8) | (((bits1.s2 >> 6) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s3 & 0x0F00) >> 8) | (((bits1.s6 >> 6) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s7, 2); \
|
||||
total_sums.s0 += ((((bits4.s2 & 0xF000) >> 12) | (((bits1.s2 >> 7) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s3 & 0xF000) >> 12) | (((bits1.s6 >> 7) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s0, 3); \
|
||||
total_sums.s0 += (((bits4.s4 & 0x000F) | (((bits1.s3 ) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += (((bits4.s5 & 0x000F) | (((bits1.s7 ) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s1, 3); \
|
||||
total_sums.s0 += ((((bits4.s4 & 0x00F0) >> 4) | (((bits1.s3 >> 1) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s5 & 0x00F0) >> 4) | (((bits1.s7 >> 1) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s2, 3); \
|
||||
total_sums.s0 += ((((bits4.s4 & 0x0F00) >> 8) | (((bits1.s3 >> 2) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s5 & 0x0F00) >> 8) | (((bits1.s7 >> 2) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s3, 3); \
|
||||
total_sums.s0 += ((((bits4.s4 & 0xF000) >> 12) | (((bits1.s3 >> 3) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s5 & 0xF000) >> 12) | (((bits1.s7 >> 3) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s4, 3); \
|
||||
total_sums.s0 += (((bits4.s6 & 0x000F) | (((bits1.s3 >> 4) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += (((bits4.s7 & 0x000F) | (((bits1.s7 >> 4) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s5, 3); \
|
||||
total_sums.s0 += ((((bits4.s6 & 0x00F0) >> 4) | (((bits1.s3 >> 5) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s7 & 0x00F0) >> 4) | (((bits1.s7 >> 5) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s6, 3); \
|
||||
total_sums.s0 += ((((bits4.s6 & 0x0F00) >> 8) | (((bits1.s3 >> 6) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s7 & 0x0F00) >> 8) | (((bits1.s7 >> 6) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s7, 3); \
|
||||
total_sums.s0 += ((((bits4.s6 & 0xF000) >> 12) | (((bits1.s3 >> 7) & 0x01) << 4)) - 16) * scale.s0 * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s7 & 0xF000) >> 12) | (((bits1.s7 >> 7) & 0x01) << 4)) - 16) * scale.s1 * shared_y; \
|
||||
|
||||
|
||||
#define dequantizeBlockAccum_ns_q5_0_sgbroadcast_8_hi(total_sums, bits4, bits1, scale, y) \
|
||||
float8 shared_y; \
|
||||
shared_y = sub_group_broadcast(y, 0); \
|
||||
total_sums.s0 += (((bits4.s0 & 0x000F) | (((bits1.s0 ) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s0; \
|
||||
total_sums.s0 += ((((bits4.s0 & 0x00F0) >> 4) | (((bits1.s0 >> 1) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s1; \
|
||||
total_sums.s0 += ((((bits4.s0 & 0x0F00) >> 8) | (((bits1.s0 >> 2) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s2; \
|
||||
total_sums.s0 += ((((bits4.s0 & 0xF000) >> 12) | (((bits1.s0 >> 3) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s3; \
|
||||
total_sums.s0 += (((bits4.s2 & 0x000F) | (((bits1.s0 >> 4) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s4; \
|
||||
total_sums.s0 += ((((bits4.s2 & 0x00F0) >> 4) | (((bits1.s0 >> 5) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s5; \
|
||||
total_sums.s0 += ((((bits4.s2 & 0x0F00) >> 8) | (((bits1.s0 >> 6) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s6; \
|
||||
total_sums.s0 += ((((bits4.s2 & 0xF000) >> 12) | (((bits1.s0 >> 7) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s7; \
|
||||
total_sums.s1 += (((bits4.s1 & 0x000F) | (((bits1.s4 ) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s0; \
|
||||
total_sums.s1 += ((((bits4.s1 & 0x00F0) >> 4) | (((bits1.s4 >> 1) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s1; \
|
||||
total_sums.s1 += ((((bits4.s1 & 0x0F00) >> 8) | (((bits1.s4 >> 2) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s2; \
|
||||
total_sums.s1 += ((((bits4.s1 & 0xF000) >> 12) | (((bits1.s4 >> 3) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s3; \
|
||||
total_sums.s1 += (((bits4.s3 & 0x000F) | (((bits1.s4 >> 4) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s4; \
|
||||
total_sums.s1 += ((((bits4.s3 & 0x00F0) >> 4) | (((bits1.s4 >> 5) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s5; \
|
||||
total_sums.s1 += ((((bits4.s3 & 0x0F00) >> 8) | (((bits1.s4 >> 6) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s6; \
|
||||
total_sums.s1 += ((((bits4.s3 & 0xF000) >> 12) | (((bits1.s4 >> 7) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s7; \
|
||||
shared_y = sub_group_broadcast(y, 1); \
|
||||
total_sums.s0 += (((bits4.s4 & 0x000F) | (((bits1.s1 ) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s0; \
|
||||
total_sums.s0 += ((((bits4.s4 & 0x00F0) >> 4) | (((bits1.s1 >> 1) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s1; \
|
||||
total_sums.s0 += ((((bits4.s4 & 0x0F00) >> 8) | (((bits1.s1 >> 2) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s2; \
|
||||
total_sums.s0 += ((((bits4.s4 & 0xF000) >> 12) | (((bits1.s1 >> 3) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s3; \
|
||||
total_sums.s0 += (((bits4.s6 & 0x000F) | (((bits1.s1 >> 4) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s4; \
|
||||
total_sums.s0 += ((((bits4.s6 & 0x00F0) >> 4) | (((bits1.s1 >> 5) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s5; \
|
||||
total_sums.s0 += ((((bits4.s6 & 0x0F00) >> 8) | (((bits1.s1 >> 6) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s6; \
|
||||
total_sums.s0 += ((((bits4.s6 & 0xF000) >> 12) | (((bits1.s1 >> 7) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s7; \
|
||||
total_sums.s1 += (((bits4.s5 & 0x000F) | (((bits1.s5 ) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s0; \
|
||||
total_sums.s1 += ((((bits4.s5 & 0x00F0) >> 4) | (((bits1.s5 >> 1) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s1; \
|
||||
total_sums.s1 += ((((bits4.s5 & 0x0F00) >> 8) | (((bits1.s5 >> 2) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s2; \
|
||||
total_sums.s1 += ((((bits4.s5 & 0xF000) >> 12) | (((bits1.s5 >> 3) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s3; \
|
||||
total_sums.s1 += (((bits4.s7 & 0x000F) | (((bits1.s5 >> 4) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s4; \
|
||||
total_sums.s1 += ((((bits4.s7 & 0x00F0) >> 4) | (((bits1.s5 >> 5) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s5; \
|
||||
total_sums.s1 += ((((bits4.s7 & 0x0F00) >> 8) | (((bits1.s5 >> 6) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s6; \
|
||||
total_sums.s1 += ((((bits4.s7 & 0xF000) >> 12) | (((bits1.s5 >> 7) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s7; \
|
||||
|
||||
|
||||
#define dequantizeBlockAccum_ns_q5_0_sgbroadcast_8_lo(total_sums, bits4, bits1, scale, y) \
|
||||
shared_y = sub_group_broadcast(y, 2); \
|
||||
total_sums.s0 += (((bits4.s0 & 0x000F) | (((bits1.s2 ) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s0; \
|
||||
total_sums.s0 += ((((bits4.s0 & 0x00F0) >> 4) | (((bits1.s2 >> 1) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s1; \
|
||||
total_sums.s0 += ((((bits4.s0 & 0x0F00) >> 8) | (((bits1.s2 >> 2) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s2; \
|
||||
total_sums.s0 += ((((bits4.s0 & 0xF000) >> 12) | (((bits1.s2 >> 3) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s3; \
|
||||
total_sums.s0 += (((bits4.s2 & 0x000F) | (((bits1.s2 >> 4) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s4; \
|
||||
total_sums.s0 += ((((bits4.s2 & 0x00F0) >> 4) | (((bits1.s2 >> 5) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s5; \
|
||||
total_sums.s0 += ((((bits4.s2 & 0x0F00) >> 8) | (((bits1.s2 >> 6) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s6; \
|
||||
total_sums.s0 += ((((bits4.s2 & 0xF000) >> 12) | (((bits1.s2 >> 7) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s7; \
|
||||
total_sums.s1 += (((bits4.s1 & 0x000F) | (((bits1.s6 ) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s0; \
|
||||
total_sums.s1 += ((((bits4.s1 & 0x00F0) >> 4) | (((bits1.s6 >> 1) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s1; \
|
||||
total_sums.s1 += ((((bits4.s1 & 0x0F00) >> 8) | (((bits1.s6 >> 2) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s2; \
|
||||
total_sums.s1 += ((((bits4.s1 & 0xF000) >> 12) | (((bits1.s6 >> 3) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s3; \
|
||||
total_sums.s1 += (((bits4.s3 & 0x000F) | (((bits1.s6 >> 4) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s4; \
|
||||
total_sums.s1 += ((((bits4.s3 & 0x00F0) >> 4) | (((bits1.s6 >> 5) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s5; \
|
||||
total_sums.s1 += ((((bits4.s3 & 0x0F00) >> 8) | (((bits1.s6 >> 6) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s6; \
|
||||
total_sums.s1 += ((((bits4.s3 & 0xF000) >> 12) | (((bits1.s6 >> 7) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s7; \
|
||||
shared_y = sub_group_broadcast(y, 3); \
|
||||
total_sums.s0 += (((bits4.s4 & 0x000F) | (((bits1.s3 ) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s0; \
|
||||
total_sums.s0 += ((((bits4.s4 & 0x00F0) >> 4) | (((bits1.s3 >> 1) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s1; \
|
||||
total_sums.s0 += ((((bits4.s4 & 0x0F00) >> 8) | (((bits1.s3 >> 2) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s2; \
|
||||
total_sums.s0 += ((((bits4.s4 & 0xF000) >> 12) | (((bits1.s3 >> 3) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s3; \
|
||||
total_sums.s0 += (((bits4.s6 & 0x000F) | (((bits1.s3 >> 4) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s4; \
|
||||
total_sums.s0 += ((((bits4.s6 & 0x00F0) >> 4) | (((bits1.s3 >> 5) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s5; \
|
||||
total_sums.s0 += ((((bits4.s6 & 0x0F00) >> 8) | (((bits1.s3 >> 6) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s6; \
|
||||
total_sums.s0 += ((((bits4.s6 & 0xF000) >> 12) | (((bits1.s3 >> 7) & 0x01) << 4)) - 16) * scale.s0 * shared_y.s7; \
|
||||
total_sums.s1 += (((bits4.s5 & 0x000F) | (((bits1.s7 ) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s0; \
|
||||
total_sums.s1 += ((((bits4.s5 & 0x00F0) >> 4) | (((bits1.s7 >> 1) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s1; \
|
||||
total_sums.s1 += ((((bits4.s5 & 0x0F00) >> 8) | (((bits1.s7 >> 2) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s2; \
|
||||
total_sums.s1 += ((((bits4.s5 & 0xF000) >> 12) | (((bits1.s7 >> 3) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s3; \
|
||||
total_sums.s1 += (((bits4.s7 & 0x000F) | (((bits1.s7 >> 4) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s4; \
|
||||
total_sums.s1 += ((((bits4.s7 & 0x00F0) >> 4) | (((bits1.s7 >> 5) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s5; \
|
||||
total_sums.s1 += ((((bits4.s7 & 0x0F00) >> 8) | (((bits1.s7 >> 6) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s6; \
|
||||
total_sums.s1 += ((((bits4.s7 & 0xF000) >> 12) | (((bits1.s7 >> 7) & 0x01) << 4)) - 16) * scale.s1 * shared_y.s7; \
|
||||
|
||||
#ifdef ADRENO_GPU
|
||||
REQD_SUBGROUP_SIZE_64
|
||||
#endif
|
||||
__kernel void kernel_gemv_noshuffle_q5_0_f32(
|
||||
__read_only image1d_buffer_t src0_qs, // quantized A
|
||||
global ushort * src0_qh, // 5th bits
|
||||
global half2 * src0_d, // A scales
|
||||
__read_only image1d_buffer_t src1, // B activations
|
||||
global float * dst,
|
||||
ulong offsetd,
|
||||
int ne00, // K
|
||||
int ne01) // M
|
||||
{
|
||||
uint groupId = get_local_id(1);
|
||||
uint gid = get_global_id(0);
|
||||
ushort slid = get_sub_group_local_id();
|
||||
|
||||
uint K = ne00;
|
||||
uint M = ne01;
|
||||
|
||||
uint LINE_STRIDE_A = M / 2;
|
||||
uint BLOCK_STRIDE_A = NSUBGROUPS * M;
|
||||
|
||||
private uint4 regA;
|
||||
private half2 regS;
|
||||
private float8 regB;
|
||||
|
||||
private float2 totalSum = (float2)(0.0f);
|
||||
|
||||
for (uint k = groupId; k < (K / QK5_0); k += NSUBGROUPS) {
|
||||
regS = src0_d[gid + k * LINE_STRIDE_A];
|
||||
|
||||
ushort4 qh_raw;
|
||||
qh_raw.s0 = src0_qh[gid + (4*k + 0) * LINE_STRIDE_A];
|
||||
qh_raw.s1 = src0_qh[gid + (4*k + 1) * LINE_STRIDE_A];
|
||||
qh_raw.s2 = src0_qh[gid + (4*k + 2) * LINE_STRIDE_A];
|
||||
qh_raw.s3 = src0_qh[gid + (4*k + 3) * LINE_STRIDE_A];
|
||||
|
||||
uchar8 raw = as_uchar8(qh_raw);
|
||||
uchar8 qh_bytes = (uchar8)(raw.s0, raw.s2, raw.s4, raw.s6,
|
||||
raw.s1, raw.s3, raw.s5, raw.s7);
|
||||
|
||||
// Load activations
|
||||
if (slid < 4) {
|
||||
regB.s0123 = read_imagef(src1, (slid * 2 + k * 8));
|
||||
regB.s4567 = read_imagef(src1, (1 + slid * 2 + k * 8));
|
||||
}
|
||||
|
||||
regA.s0 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x;
|
||||
regA.s1 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x;
|
||||
regA.s2 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x;
|
||||
regA.s3 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x;
|
||||
|
||||
#ifdef VECTOR_SUB_GROUP_BROADCAST
|
||||
dequantizeBlockAccum_ns_q5_0_sgbroadcast_8_hi(totalSum, as_ushort8(regA), qh_bytes, regS, regB);
|
||||
#else
|
||||
dequantizeBlockAccum_ns_q5_0_sgbroadcast_1_hi(totalSum, as_ushort8(regA), qh_bytes, regS, regB);
|
||||
#endif // VECTOR_SUB_GROUP_BROADCAST
|
||||
|
||||
regA.s0 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x;
|
||||
regA.s1 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x;
|
||||
regA.s2 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
|
||||
regA.s3 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;
|
||||
#ifdef VECTOR_SUB_GROUP_BROADCAST
|
||||
dequantizeBlockAccum_ns_q5_0_sgbroadcast_8_lo(totalSum, as_ushort8(regA), qh_bytes, regS, regB);
|
||||
#else
|
||||
dequantizeBlockAccum_ns_q5_0_sgbroadcast_1_lo(totalSum, as_ushort8(regA), qh_bytes, regS, regB);
|
||||
#endif // VECTOR_SUB_GROUP_BROADCAST
|
||||
}
|
||||
|
||||
// reduction in local memory, assumes #wave=4
|
||||
local float2 reduceLM[SUBGROUP_SIZE * 3];
|
||||
if (groupId == 1) {
|
||||
reduceLM[SUBGROUP_SIZE * 0 + slid] = totalSum;
|
||||
}
|
||||
if (groupId == 2) {
|
||||
reduceLM[SUBGROUP_SIZE * 1 + slid] = totalSum;
|
||||
}
|
||||
if (groupId == 3) {
|
||||
reduceLM[SUBGROUP_SIZE * 2 + slid] = totalSum;
|
||||
}
|
||||
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
if (groupId == 0) {
|
||||
totalSum += reduceLM[SUBGROUP_SIZE * 0 + slid];
|
||||
}
|
||||
if (groupId == 0) {
|
||||
totalSum += reduceLM[SUBGROUP_SIZE * 1 + slid];
|
||||
}
|
||||
if (groupId == 0) {
|
||||
totalSum += reduceLM[SUBGROUP_SIZE * 2 + slid];
|
||||
}
|
||||
|
||||
// 2 outputs per fiber in wave 0
|
||||
if (groupId == 0) {
|
||||
dst = (global float*)((global char*)dst + offsetd);
|
||||
vstore2(totalSum, 0, &(dst[gid * 2]));
|
||||
}
|
||||
|
||||
}
|
||||
@@ -0,0 +1,294 @@
|
||||
#pragma OPENCL EXTENSION cl_khr_fp16 : enable
|
||||
#pragma OPENCL EXTENSION cl_khr_subgroups : 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_64 __attribute__((qcom_reqd_sub_group_size("half")))
|
||||
#endif
|
||||
|
||||
#define QK5_1 32
|
||||
#define NSUBGROUPS 4
|
||||
#define SUBGROUP_SIZE 64
|
||||
|
||||
#define dequantizeBlockAccum_ns_q5_1_sgbroadcast_1_hi(total_sums, bits4, bits1, scale, minv, y) \
|
||||
float shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s0, 0); \
|
||||
total_sums.s0 += (((bits4.s0 & 0x000F) | (((bits1.s0 ) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += (((bits4.s1 & 0x000F) | (((bits1.s4 ) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s1, 0); \
|
||||
total_sums.s0 += ((((bits4.s0 & 0x00F0) >> 4) | (((bits1.s0 >> 1) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s1 & 0x00F0) >> 4) | (((bits1.s4 >> 1) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s2, 0); \
|
||||
total_sums.s0 += ((((bits4.s0 & 0x0F00) >> 8) | (((bits1.s0 >> 2) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s1 & 0x0F00) >> 8) | (((bits1.s4 >> 2) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s3, 0); \
|
||||
total_sums.s0 += ((((bits4.s0 & 0xF000) >> 12) | (((bits1.s0 >> 3) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s1 & 0xF000) >> 12) | (((bits1.s4 >> 3) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s4, 0); \
|
||||
total_sums.s0 += (((bits4.s2 & 0x000F) | (((bits1.s0 >> 4) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += (((bits4.s3 & 0x000F) | (((bits1.s4 >> 4) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s5, 0); \
|
||||
total_sums.s0 += ((((bits4.s2 & 0x00F0) >> 4) | (((bits1.s0 >> 5) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s3 & 0x00F0) >> 4) | (((bits1.s4 >> 5) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s6, 0); \
|
||||
total_sums.s0 += ((((bits4.s2 & 0x0F00) >> 8) | (((bits1.s0 >> 6) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s3 & 0x0F00) >> 8) | (((bits1.s4 >> 6) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s7, 0); \
|
||||
total_sums.s0 += ((((bits4.s2 & 0xF000) >> 12) | (((bits1.s0 >> 7) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s3 & 0xF000) >> 12) | (((bits1.s4 >> 7) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s0, 1); \
|
||||
total_sums.s0 += (((bits4.s4 & 0x000F) | (((bits1.s1 ) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += (((bits4.s5 & 0x000F) | (((bits1.s5 ) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s1, 1); \
|
||||
total_sums.s0 += ((((bits4.s4 & 0x00F0) >> 4) | (((bits1.s1 >> 1) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s5 & 0x00F0) >> 4) | (((bits1.s5 >> 1) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s2, 1); \
|
||||
total_sums.s0 += ((((bits4.s4 & 0x0F00) >> 8) | (((bits1.s1 >> 2) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s5 & 0x0F00) >> 8) | (((bits1.s5 >> 2) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s3, 1); \
|
||||
total_sums.s0 += ((((bits4.s4 & 0xF000) >> 12) | (((bits1.s1 >> 3) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s5 & 0xF000) >> 12) | (((bits1.s5 >> 3) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s4, 1); \
|
||||
total_sums.s0 += (((bits4.s6 & 0x000F) | (((bits1.s1 >> 4) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += (((bits4.s7 & 0x000F) | (((bits1.s5 >> 4) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s5, 1); \
|
||||
total_sums.s0 += ((((bits4.s6 & 0x00F0) >> 4) | (((bits1.s1 >> 5) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s7 & 0x00F0) >> 4) | (((bits1.s5 >> 5) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s6, 1); \
|
||||
total_sums.s0 += ((((bits4.s6 & 0x0F00) >> 8) | (((bits1.s1 >> 6) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s7 & 0x0F00) >> 8) | (((bits1.s5 >> 6) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s7, 1); \
|
||||
total_sums.s0 += ((((bits4.s6 & 0xF000) >> 12) | (((bits1.s1 >> 7) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s7 & 0xF000) >> 12) | (((bits1.s5 >> 7) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
|
||||
|
||||
#define dequantizeBlockAccum_ns_q5_1_sgbroadcast_1_lo(total_sums, bits4, bits1, scale, minv, y) \
|
||||
shared_y = sub_group_broadcast(y.s0, 2); \
|
||||
total_sums.s0 += (((bits4.s0 & 0x000F) | (((bits1.s2 ) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += (((bits4.s1 & 0x000F) | (((bits1.s6 ) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s1, 2); \
|
||||
total_sums.s0 += ((((bits4.s0 & 0x00F0) >> 4) | (((bits1.s2 >> 1) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s1 & 0x00F0) >> 4) | (((bits1.s6 >> 1) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s2, 2); \
|
||||
total_sums.s0 += ((((bits4.s0 & 0x0F00) >> 8) | (((bits1.s2 >> 2) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s1 & 0x0F00) >> 8) | (((bits1.s6 >> 2) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s3, 2); \
|
||||
total_sums.s0 += ((((bits4.s0 & 0xF000) >> 12) | (((bits1.s2 >> 3) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s1 & 0xF000) >> 12) | (((bits1.s6 >> 3) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s4, 2); \
|
||||
total_sums.s0 += (((bits4.s2 & 0x000F) | (((bits1.s2 >> 4) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += (((bits4.s3 & 0x000F) | (((bits1.s6 >> 4) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s5, 2); \
|
||||
total_sums.s0 += ((((bits4.s2 & 0x00F0) >> 4) | (((bits1.s2 >> 5) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s3 & 0x00F0) >> 4) | (((bits1.s6 >> 5) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s6, 2); \
|
||||
total_sums.s0 += ((((bits4.s2 & 0x0F00) >> 8) | (((bits1.s2 >> 6) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s3 & 0x0F00) >> 8) | (((bits1.s6 >> 6) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s7, 2); \
|
||||
total_sums.s0 += ((((bits4.s2 & 0xF000) >> 12) | (((bits1.s2 >> 7) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s3 & 0xF000) >> 12) | (((bits1.s6 >> 7) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s0, 3); \
|
||||
total_sums.s0 += (((bits4.s4 & 0x000F) | (((bits1.s3 ) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += (((bits4.s5 & 0x000F) | (((bits1.s7 ) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s1, 3); \
|
||||
total_sums.s0 += ((((bits4.s4 & 0x00F0) >> 4) | (((bits1.s3 >> 1) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s5 & 0x00F0) >> 4) | (((bits1.s7 >> 1) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s2, 3); \
|
||||
total_sums.s0 += ((((bits4.s4 & 0x0F00) >> 8) | (((bits1.s3 >> 2) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s5 & 0x0F00) >> 8) | (((bits1.s7 >> 2) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s3, 3); \
|
||||
total_sums.s0 += ((((bits4.s4 & 0xF000) >> 12) | (((bits1.s3 >> 3) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s5 & 0xF000) >> 12) | (((bits1.s7 >> 3) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s4, 3); \
|
||||
total_sums.s0 += (((bits4.s6 & 0x000F) | (((bits1.s3 >> 4) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += (((bits4.s7 & 0x000F) | (((bits1.s7 >> 4) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s5, 3); \
|
||||
total_sums.s0 += ((((bits4.s6 & 0x00F0) >> 4) | (((bits1.s3 >> 5) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s7 & 0x00F0) >> 4) | (((bits1.s7 >> 5) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s6, 3); \
|
||||
total_sums.s0 += ((((bits4.s6 & 0x0F00) >> 8) | (((bits1.s3 >> 6) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s7 & 0x0F00) >> 8) | (((bits1.s7 >> 6) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
shared_y = sub_group_broadcast(y.s7, 3); \
|
||||
total_sums.s0 += ((((bits4.s6 & 0xF000) >> 12) | (((bits1.s3 >> 7) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y; \
|
||||
total_sums.s1 += ((((bits4.s7 & 0xF000) >> 12) | (((bits1.s7 >> 7) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y; \
|
||||
|
||||
|
||||
#define dequantizeBlockAccum_ns_q5_1_sgbroadcast_8_hi(total_sums, bits4, bits1, scale, minv, y) \
|
||||
float8 shared_y; \
|
||||
shared_y = sub_group_broadcast(y, 0); \
|
||||
total_sums.s0 += (((bits4.s0 & 0x000F) | (((bits1.s0 ) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s0; \
|
||||
total_sums.s0 += ((((bits4.s0 & 0x00F0) >> 4) | (((bits1.s0 >> 1) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s1; \
|
||||
total_sums.s0 += ((((bits4.s0 & 0x0F00) >> 8) | (((bits1.s0 >> 2) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s2; \
|
||||
total_sums.s0 += ((((bits4.s0 & 0xF000) >> 12) | (((bits1.s0 >> 3) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s3; \
|
||||
total_sums.s0 += (((bits4.s2 & 0x000F) | (((bits1.s0 >> 4) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s4; \
|
||||
total_sums.s0 += ((((bits4.s2 & 0x00F0) >> 4) | (((bits1.s0 >> 5) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s5; \
|
||||
total_sums.s0 += ((((bits4.s2 & 0x0F00) >> 8) | (((bits1.s0 >> 6) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s6; \
|
||||
total_sums.s0 += ((((bits4.s2 & 0xF000) >> 12) | (((bits1.s0 >> 7) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s7; \
|
||||
total_sums.s1 += (((bits4.s1 & 0x000F) | (((bits1.s4 ) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s0; \
|
||||
total_sums.s1 += ((((bits4.s1 & 0x00F0) >> 4) | (((bits1.s4 >> 1) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s1; \
|
||||
total_sums.s1 += ((((bits4.s1 & 0x0F00) >> 8) | (((bits1.s4 >> 2) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s2; \
|
||||
total_sums.s1 += ((((bits4.s1 & 0xF000) >> 12) | (((bits1.s4 >> 3) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s3; \
|
||||
total_sums.s1 += (((bits4.s3 & 0x000F) | (((bits1.s4 >> 4) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s4; \
|
||||
total_sums.s1 += ((((bits4.s3 & 0x00F0) >> 4) | (((bits1.s4 >> 5) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s5; \
|
||||
total_sums.s1 += ((((bits4.s3 & 0x0F00) >> 8) | (((bits1.s4 >> 6) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s6; \
|
||||
total_sums.s1 += ((((bits4.s3 & 0xF000) >> 12) | (((bits1.s4 >> 7) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s7; \
|
||||
shared_y = sub_group_broadcast(y, 1); \
|
||||
total_sums.s0 += (((bits4.s4 & 0x000F) | (((bits1.s1 ) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s0; \
|
||||
total_sums.s0 += ((((bits4.s4 & 0x00F0) >> 4) | (((bits1.s1 >> 1) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s1; \
|
||||
total_sums.s0 += ((((bits4.s4 & 0x0F00) >> 8) | (((bits1.s1 >> 2) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s2; \
|
||||
total_sums.s0 += ((((bits4.s4 & 0xF000) >> 12) | (((bits1.s1 >> 3) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s3; \
|
||||
total_sums.s0 += (((bits4.s6 & 0x000F) | (((bits1.s1 >> 4) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s4; \
|
||||
total_sums.s0 += ((((bits4.s6 & 0x00F0) >> 4) | (((bits1.s1 >> 5) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s5; \
|
||||
total_sums.s0 += ((((bits4.s6 & 0x0F00) >> 8) | (((bits1.s1 >> 6) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s6; \
|
||||
total_sums.s0 += ((((bits4.s6 & 0xF000) >> 12) | (((bits1.s1 >> 7) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s7; \
|
||||
total_sums.s1 += (((bits4.s5 & 0x000F) | (((bits1.s5 ) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s0; \
|
||||
total_sums.s1 += ((((bits4.s5 & 0x00F0) >> 4) | (((bits1.s5 >> 1) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s1; \
|
||||
total_sums.s1 += ((((bits4.s5 & 0x0F00) >> 8) | (((bits1.s5 >> 2) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s2; \
|
||||
total_sums.s1 += ((((bits4.s5 & 0xF000) >> 12) | (((bits1.s5 >> 3) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s3; \
|
||||
total_sums.s1 += (((bits4.s7 & 0x000F) | (((bits1.s5 >> 4) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s4; \
|
||||
total_sums.s1 += ((((bits4.s7 & 0x00F0) >> 4) | (((bits1.s5 >> 5) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s5; \
|
||||
total_sums.s1 += ((((bits4.s7 & 0x0F00) >> 8) | (((bits1.s5 >> 6) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s6; \
|
||||
total_sums.s1 += ((((bits4.s7 & 0xF000) >> 12) | (((bits1.s5 >> 7) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s7; \
|
||||
|
||||
|
||||
#define dequantizeBlockAccum_ns_q5_1_sgbroadcast_8_lo(total_sums, bits4, bits1, scale, minv, y) \
|
||||
shared_y = sub_group_broadcast(y, 2); \
|
||||
total_sums.s0 += (((bits4.s0 & 0x000F) | (((bits1.s2 ) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s0; \
|
||||
total_sums.s0 += ((((bits4.s0 & 0x00F0) >> 4) | (((bits1.s2 >> 1) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s1; \
|
||||
total_sums.s0 += ((((bits4.s0 & 0x0F00) >> 8) | (((bits1.s2 >> 2) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s2; \
|
||||
total_sums.s0 += ((((bits4.s0 & 0xF000) >> 12) | (((bits1.s2 >> 3) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s3; \
|
||||
total_sums.s0 += (((bits4.s2 & 0x000F) | (((bits1.s2 >> 4) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s4; \
|
||||
total_sums.s0 += ((((bits4.s2 & 0x00F0) >> 4) | (((bits1.s2 >> 5) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s5; \
|
||||
total_sums.s0 += ((((bits4.s2 & 0x0F00) >> 8) | (((bits1.s2 >> 6) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s6; \
|
||||
total_sums.s0 += ((((bits4.s2 & 0xF000) >> 12) | (((bits1.s2 >> 7) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s7; \
|
||||
total_sums.s1 += (((bits4.s1 & 0x000F) | (((bits1.s6 ) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s0; \
|
||||
total_sums.s1 += ((((bits4.s1 & 0x00F0) >> 4) | (((bits1.s6 >> 1) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s1; \
|
||||
total_sums.s1 += ((((bits4.s1 & 0x0F00) >> 8) | (((bits1.s6 >> 2) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s2; \
|
||||
total_sums.s1 += ((((bits4.s1 & 0xF000) >> 12) | (((bits1.s6 >> 3) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s3; \
|
||||
total_sums.s1 += (((bits4.s3 & 0x000F) | (((bits1.s6 >> 4) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s4; \
|
||||
total_sums.s1 += ((((bits4.s3 & 0x00F0) >> 4) | (((bits1.s6 >> 5) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s5; \
|
||||
total_sums.s1 += ((((bits4.s3 & 0x0F00) >> 8) | (((bits1.s6 >> 6) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s6; \
|
||||
total_sums.s1 += ((((bits4.s3 & 0xF000) >> 12) | (((bits1.s6 >> 7) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s7; \
|
||||
shared_y = sub_group_broadcast(y, 3); \
|
||||
total_sums.s0 += (((bits4.s4 & 0x000F) | (((bits1.s3 ) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s0; \
|
||||
total_sums.s0 += ((((bits4.s4 & 0x00F0) >> 4) | (((bits1.s3 >> 1) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s1; \
|
||||
total_sums.s0 += ((((bits4.s4 & 0x0F00) >> 8) | (((bits1.s3 >> 2) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s2; \
|
||||
total_sums.s0 += ((((bits4.s4 & 0xF000) >> 12) | (((bits1.s3 >> 3) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s3; \
|
||||
total_sums.s0 += (((bits4.s6 & 0x000F) | (((bits1.s3 >> 4) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s4; \
|
||||
total_sums.s0 += ((((bits4.s6 & 0x00F0) >> 4) | (((bits1.s3 >> 5) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s5; \
|
||||
total_sums.s0 += ((((bits4.s6 & 0x0F00) >> 8) | (((bits1.s3 >> 6) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s6; \
|
||||
total_sums.s0 += ((((bits4.s6 & 0xF000) >> 12) | (((bits1.s3 >> 7) & 0x01) << 4)) * scale.s0 + minv.s0) * shared_y.s7; \
|
||||
total_sums.s1 += (((bits4.s5 & 0x000F) | (((bits1.s7 ) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s0; \
|
||||
total_sums.s1 += ((((bits4.s5 & 0x00F0) >> 4) | (((bits1.s7 >> 1) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s1; \
|
||||
total_sums.s1 += ((((bits4.s5 & 0x0F00) >> 8) | (((bits1.s7 >> 2) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s2; \
|
||||
total_sums.s1 += ((((bits4.s5 & 0xF000) >> 12) | (((bits1.s7 >> 3) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s3; \
|
||||
total_sums.s1 += (((bits4.s7 & 0x000F) | (((bits1.s7 >> 4) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s4; \
|
||||
total_sums.s1 += ((((bits4.s7 & 0x00F0) >> 4) | (((bits1.s7 >> 5) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s5; \
|
||||
total_sums.s1 += ((((bits4.s7 & 0x0F00) >> 8) | (((bits1.s7 >> 6) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s6; \
|
||||
total_sums.s1 += ((((bits4.s7 & 0xF000) >> 12) | (((bits1.s7 >> 7) & 0x01) << 4)) * scale.s1 + minv.s1) * shared_y.s7; \
|
||||
|
||||
#ifdef ADRENO_GPU
|
||||
REQD_SUBGROUP_SIZE_64
|
||||
#endif
|
||||
__kernel void kernel_gemv_noshuffle_q5_1_f32(
|
||||
__read_only image1d_buffer_t src0_qs, // quantized A
|
||||
global ushort * src0_qh, // 5th bits
|
||||
global half2 * src0_d, // A scales
|
||||
global half2 * src0_m, // A mins
|
||||
__read_only image1d_buffer_t src1, // B activations
|
||||
global float * dst,
|
||||
ulong offsetd,
|
||||
int ne00, // K
|
||||
int ne01) // M
|
||||
{
|
||||
uint groupId = get_local_id(1);
|
||||
uint gid = get_global_id(0);
|
||||
ushort slid = get_sub_group_local_id();
|
||||
|
||||
uint K = ne00;
|
||||
uint M = ne01;
|
||||
|
||||
uint LINE_STRIDE_A = M / 2;
|
||||
uint BLOCK_STRIDE_A = NSUBGROUPS * M;
|
||||
|
||||
__private uint4 regA;
|
||||
__private half2 regS;
|
||||
__private half2 regM;
|
||||
__private float8 regB;
|
||||
|
||||
__private float2 totalSum = (float2)(0.0f);
|
||||
|
||||
for (uint k = groupId; k < (K / QK5_1); k += NSUBGROUPS) {
|
||||
regS = src0_d[gid + k * LINE_STRIDE_A];
|
||||
regM = src0_m[gid + k * LINE_STRIDE_A];
|
||||
|
||||
ushort4 qh_raw;
|
||||
qh_raw.s0 = src0_qh[gid + (4*k + 0) * LINE_STRIDE_A];
|
||||
qh_raw.s1 = src0_qh[gid + (4*k + 1) * LINE_STRIDE_A];
|
||||
qh_raw.s2 = src0_qh[gid + (4*k + 2) * LINE_STRIDE_A];
|
||||
qh_raw.s3 = src0_qh[gid + (4*k + 3) * LINE_STRIDE_A];
|
||||
|
||||
uchar8 raw = as_uchar8(qh_raw);
|
||||
uchar8 qh_bytes = (uchar8)(raw.s0, raw.s2, raw.s4, raw.s6,
|
||||
raw.s1, raw.s3, raw.s5, raw.s7);
|
||||
|
||||
// Load activations
|
||||
if (slid < 4) {
|
||||
regB.s0123 = read_imagef(src1, (slid * 2 + k * 8));
|
||||
regB.s4567 = read_imagef(src1, (1 + slid * 2 + k * 8));
|
||||
}
|
||||
|
||||
regA.s0 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x;
|
||||
regA.s1 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x;
|
||||
regA.s2 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x;
|
||||
regA.s3 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x;
|
||||
|
||||
#ifdef VECTOR_SUB_GROUP_BROADCAST
|
||||
dequantizeBlockAccum_ns_q5_1_sgbroadcast_8_hi(totalSum, as_ushort8(regA), qh_bytes, regS, regM, regB);
|
||||
#else
|
||||
dequantizeBlockAccum_ns_q5_1_sgbroadcast_1_hi(totalSum, as_ushort8(regA), qh_bytes, regS, regM, regB);
|
||||
#endif // VECTOR_SUB_GROUP_BROADCAST
|
||||
|
||||
regA.s0 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x;
|
||||
regA.s1 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x;
|
||||
regA.s2 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
|
||||
regA.s3 = read_imageui(src0_qs, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;
|
||||
#ifdef VECTOR_SUB_GROUP_BROADCAST
|
||||
dequantizeBlockAccum_ns_q5_1_sgbroadcast_8_lo(totalSum, as_ushort8(regA), qh_bytes, regS, regM, regB);
|
||||
#else
|
||||
dequantizeBlockAccum_ns_q5_1_sgbroadcast_1_lo(totalSum, as_ushort8(regA), qh_bytes, regS, regM, regB);
|
||||
#endif // VECTOR_SUB_GROUP_BROADCAST
|
||||
}
|
||||
|
||||
// reduction in local memory, assumes #wave=4
|
||||
local float2 reduceLM[SUBGROUP_SIZE * 3];
|
||||
if (groupId == 1) {
|
||||
reduceLM[SUBGROUP_SIZE * 0 + slid] = totalSum;
|
||||
}
|
||||
if (groupId == 2) {
|
||||
reduceLM[SUBGROUP_SIZE * 1 + slid] = totalSum;
|
||||
}
|
||||
if (groupId == 3) {
|
||||
reduceLM[SUBGROUP_SIZE * 2 + slid] = totalSum;
|
||||
}
|
||||
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
if (groupId == 0) {
|
||||
totalSum += reduceLM[SUBGROUP_SIZE * 0 + slid];
|
||||
}
|
||||
if (groupId == 0) {
|
||||
totalSum += reduceLM[SUBGROUP_SIZE * 1 + slid];
|
||||
}
|
||||
if (groupId == 0) {
|
||||
totalSum += reduceLM[SUBGROUP_SIZE * 2 + slid];
|
||||
}
|
||||
|
||||
// 2 outputs per fiber in wave 0
|
||||
if (groupId == 0) {
|
||||
dst = (global float*)((global char*)dst + offsetd);
|
||||
vstore2(totalSum, 0, &(dst[gid * 2]));
|
||||
}
|
||||
|
||||
}
|
||||
Reference in New Issue
Block a user