vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

vulkan: use larger loads in q5_k and q6_k shaders.

Browse source 
Similar to the optimization I did in q4_k recently, this vectorizes some loads
and reduces the number of bit twiddling instructions.
This commit is contained in:
Jeff Bolz 2024-11-17 23:39:31 -06:00
parent 6c3ad9342d
commit 55f477b114
2 changed files with 132 additions and 67 deletions
Download patch file Download diff file Expand all files Collapse all files

136
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp
View file

@ -1,5 +1,7 @@
#version 450
#extension GL_EXT_shader_explicit_arithmetic_types : require
#include "mul_mat_vec_base.comp"
layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
@ -35,70 +37,106 @@ void main() {
const uint8_t hm1 = uint8_t(1 << (2*v_im));
const uint8_t hm2 = uint8_t(hm1 << 4);
tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += 2) {
const uint y1_idx = i * QUANT_K + y_offset;
const uint y2_idx = y1_idx + 128;
const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib0 + i].d.x);
const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib0 + i].d.y);
f16vec2 d = data_a[ib0 + i].d;
const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
const uint8_t sc0 = uint8_t( data_a[ib0 + i].scales[v_im * 2 ] & 0x3f);
const uint8_t sc1 = uint8_t( data_a[ib0 + i].scales[v_im * 2 + 1] & 0x3f);
const uint8_t sc2 = uint8_t( data_a[ib0 + i].scales[v_im * 2 + 4] & 0x3f);
const uint8_t sc3 = uint8_t( data_a[ib0 + i].scales[v_im * 2 + 5] & 0x3f);
const uint8_t sc4 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 8] & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 ] & 0xc0) >> 2));
const uint8_t sc5 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 9] & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 1] & 0xc0) >> 2));
const uint8_t sc6 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 8] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 4] & 0xc0) >> 2));
const uint8_t sc7 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 9] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 5] & 0xc0) >> 2));
uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im ];
uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4];
uvec4 scale0 = uvec4(unpack8(scale0_u32));
uvec4 scale4 = uvec4(unpack8(scale4_u32));
uvec4 scale8 = uvec4(unpack8(scale8_u32));
const uint8_t q4_0 = uint8_t(data_a[ib0 + i].qs[q_offset ] & 0xf);
const uint8_t q4_1 = uint8_t(data_a[ib0 + i].qs[q_offset + 1] & 0xf);
const uint8_t q4_2 = uint8_t(data_a[ib0 + i].qs[q_offset + 16] & 0xf);
const uint8_t q4_3 = uint8_t(data_a[ib0 + i].qs[q_offset + 17] & 0xf);
const uint8_t q4_4 = uint8_t(data_a[ib0 + i].qs[q_offset ] >> 4);
const uint8_t q4_5 = uint8_t(data_a[ib0 + i].qs[q_offset + 1] >> 4);
const uint8_t q4_6 = uint8_t(data_a[ib0 + i].qs[q_offset + 16] >> 4);
const uint8_t q4_7 = uint8_t(data_a[ib0 + i].qs[q_offset + 17] >> 4);
const uint8_t q4_8 = uint8_t(data_a[ib0 + i].qs[q_offset + 64] & 0xf);
const uint8_t q4_9 = uint8_t(data_a[ib0 + i].qs[q_offset + 65] & 0xf);
const uint8_t q4_10 = uint8_t(data_a[ib0 + i].qs[q_offset + 80] & 0xf);
const uint8_t q4_11 = uint8_t(data_a[ib0 + i].qs[q_offset + 81] & 0xf);
const uint8_t q4_12 = uint8_t(data_a[ib0 + i].qs[q_offset + 64] >> 4);
const uint8_t q4_13 = uint8_t(data_a[ib0 + i].qs[q_offset + 65] >> 4);
const uint8_t q4_14 = uint8_t(data_a[ib0 + i].qs[q_offset + 80] >> 4);
const uint8_t q4_15 = uint8_t(data_a[ib0 + i].qs[q_offset + 81] >> 4);
const uint32_t sc0 = ( scale0.x & 0x3f);
const uint32_t sc1 = ( scale0.y & 0x3f);
const uint32_t sc2 = ( scale4.x & 0x3f);
const uint32_t sc3 = ( scale4.y & 0x3f);
const uint32_t sc4 = (( scale8.x & 0x0f) | ((scale0.x & 0xc0) >> 2));
const uint32_t sc5 = (( scale8.y & 0x0f) | ((scale0.y & 0xc0) >> 2));
const uint32_t sc6 = (((scale8.x >> 4) & 0x0f) | ((scale4.x & 0xc0) >> 2));
const uint32_t sc7 = (((scale8.y >> 4) & 0x0f) | ((scale4.y & 0xc0) >> 2));
uint32_t qs0_16_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 8]) << 16);
uint32_t qs64_80_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 32]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 40]) << 16);
uint32_t qs0_16_u32_lo4 = qs0_16_u32 & 0x0F0F0F0F;
uint32_t qs0_16_u32_hi4 = (qs0_16_u32 >> 4) & 0x0F0F0F0F;
uint32_t qs64_80_u32_lo4 = qs64_80_u32 & 0x0F0F0F0F;
uint32_t qs64_80_u32_hi4 = (qs64_80_u32 >> 4) & 0x0F0F0F0F;
uvec4 qs0_16_lo4 = uvec4(unpack8(qs0_16_u32_lo4));
uvec4 qs64_80_lo4 = uvec4(unpack8(qs64_80_u32_lo4));
uvec4 qs0_16_hi4 = uvec4(unpack8(qs0_16_u32_hi4));
uvec4 qs64_80_hi4 = uvec4(unpack8(qs64_80_u32_hi4));
const uint32_t q4_0 = qs0_16_lo4.x;
const uint32_t q4_1 = qs0_16_lo4.y;
const uint32_t q4_2 = qs0_16_lo4.z;
const uint32_t q4_3 = qs0_16_lo4.w;
const uint32_t q4_4 = qs0_16_hi4.x;
const uint32_t q4_5 = qs0_16_hi4.y;
const uint32_t q4_6 = qs0_16_hi4.z;
const uint32_t q4_7 = qs0_16_hi4.w;
const uint32_t q4_8 = qs64_80_lo4.x;
const uint32_t q4_9 = qs64_80_lo4.y;
const uint32_t q4_10 = qs64_80_lo4.z;
const uint32_t q4_11 = qs64_80_lo4.w;
const uint32_t q4_12 = qs64_80_hi4.x;
const uint32_t q4_13 = qs64_80_hi4.y;
const uint32_t q4_14 = qs64_80_hi4.z;
const uint32_t q4_15 = qs64_80_hi4.w;
B_TYPE_VEC2 by10 = data_b_v2[(b_offset + y1_idx) / 2];
B_TYPE_VEC2 by116 = data_b_v2[(b_offset + y1_idx) / 2 + 8];
B_TYPE_VEC2 by132 = data_b_v2[(b_offset + y1_idx) / 2 + 16];
B_TYPE_VEC2 by148 = data_b_v2[(b_offset + y1_idx) / 2 + 24];
B_TYPE_VEC2 by20 = data_b_v2[(b_offset + y2_idx) / 2];
B_TYPE_VEC2 by216 = data_b_v2[(b_offset + y2_idx) / 2 + 8];
B_TYPE_VEC2 by232 = data_b_v2[(b_offset + y2_idx) / 2 + 16];
B_TYPE_VEC2 by248 = data_b_v2[(b_offset + y2_idx) / 2 + 24];
uint32_t qh0 = data_a_packed16[ib0 + i].qh[l0 / 2];
uint32_t qh1 = qh0 >> 8;
uint32_t qh16 = data_a_packed16[ib0 + i].qh[l0 / 2 + 8];
uint32_t qh17 = qh16 >> 8;
const FLOAT_TYPE sx =
fma(FLOAT_TYPE(data_b[b_offset + y1_idx ]), (q4_0 + (((data_a[ib0 + i].qh[l0 ] & hm1) != 0) ? 16 : 0)),
fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 1]), (q4_1 + (((data_a[ib0 + i].qh[l0 + 1] & hm1) != 0) ? 16 : 0)),
fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 16]), (q4_2 + (((data_a[ib0 + i].qh[l0 + 16] & hm1) != 0) ? 16 : 0)),
FLOAT_TYPE(data_b[b_offset + y1_idx + 17]) * (q4_3 + (((data_a[ib0 + i].qh[l0 + 17] & hm1) != 0) ? 16 : 0)))));
fma(FLOAT_TYPE(by10.x), (q4_0 + (((qh0 & hm1) != 0) ? 16 : 0)),
fma(FLOAT_TYPE(by10.y), (q4_1 + (((qh1 & hm1) != 0) ? 16 : 0)),
fma(FLOAT_TYPE(by116.x), (q4_2 + (((qh16 & hm1) != 0) ? 16 : 0)),
FLOAT_TYPE(by116.y) * (q4_3 + (((qh17 & hm1) != 0) ? 16 : 0)))));
const FLOAT_TYPE sy =
fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), (q4_4 + (((data_a[ib0 + i].qh[l0 ] & (hm1 << 1)) != 0) ? 16 : 0)),
fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 33]), (q4_5 + (((data_a[ib0 + i].qh[l0 + 1] & (hm1 << 1)) != 0) ? 16 : 0)),
fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 48]), (q4_6 + (((data_a[ib0 + i].qh[l0 + 16] & (hm1 << 1)) != 0) ? 16 : 0)),
FLOAT_TYPE(data_b[b_offset + y1_idx + 49]) * (q4_7 + (((data_a[ib0 + i].qh[l0 + 17] & (hm1 << 1)) != 0) ? 16 : 0)))));
fma(FLOAT_TYPE(by132.x), (q4_4 + (((qh0 & (hm1 << 1)) != 0) ? 16 : 0)),
fma(FLOAT_TYPE(by132.y), (q4_5 + (((qh1 & (hm1 << 1)) != 0) ? 16 : 0)),
fma(FLOAT_TYPE(by148.x), (q4_6 + (((qh16 & (hm1 << 1)) != 0) ? 16 : 0)),
FLOAT_TYPE(by148.y) * (q4_7 + (((qh17 & (hm1 << 1)) != 0) ? 16 : 0)))));
const FLOAT_TYPE sz =
fma(FLOAT_TYPE(data_b[b_offset + y2_idx ]), (q4_8 + (((data_a[ib0 + i].qh[l0 ] & hm2) != 0) ? 16 : 0)),
fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 1]), (q4_9 + (((data_a[ib0 + i].qh[l0 + 1] & hm2) != 0) ? 16 : 0)),
fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 16]), (q4_10 + (((data_a[ib0 + i].qh[l0 + 16] & hm2) != 0) ? 16 : 0)),
FLOAT_TYPE(data_b[b_offset + y2_idx + 17]) * (q4_11 + (((data_a[ib0 + i].qh[l0 + 17] & hm2) != 0) ? 16 : 0)))));
fma(FLOAT_TYPE(by20.x), (q4_8 + (((qh0 & hm2) != 0) ? 16 : 0)),
fma(FLOAT_TYPE(by20.y), (q4_9 + (((qh1 & hm2) != 0) ? 16 : 0)),
fma(FLOAT_TYPE(by216.x), (q4_10 + (((qh16 & hm2) != 0) ? 16 : 0)),
FLOAT_TYPE(by216.y) * (q4_11 + (((qh17 & hm2) != 0) ? 16 : 0)))));
const FLOAT_TYPE sw =
fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), (q4_12 + (((data_a[ib0 + i].qh[l0 ] & (hm2 << 1)) != 0) ? 16 : 0)),
fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 33]), (q4_13 + (((data_a[ib0 + i].qh[l0 + 1] & (hm2 << 1)) != 0) ? 16 : 0)),
fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 48]), (q4_14 + (((data_a[ib0 + i].qh[l0 + 16] & (hm2 << 1)) != 0) ? 16 : 0)),
FLOAT_TYPE(data_b[b_offset + y2_idx + 49]) * (q4_15 + (((data_a[ib0 + i].qh[l0 + 17] & (hm2 << 1)) != 0) ? 16 : 0)))));
fma(FLOAT_TYPE(by232.x), (q4_12 + (((qh0 & (hm2 << 1)) != 0) ? 16 : 0)),
fma(FLOAT_TYPE(by232.y), (q4_13 + (((qh1 & (hm2 << 1)) != 0) ? 16 : 0)),
fma(FLOAT_TYPE(by248.x), (q4_14 + (((qh16 & (hm2 << 1)) != 0) ? 16 : 0)),
FLOAT_TYPE(by248.y) * (q4_15 + (((qh17 & (hm2 << 1)) != 0) ? 16 : 0)))));
const FLOAT_TYPE smin =
fma(FLOAT_TYPE(data_b[b_offset + y1_idx ]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 1 ]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 16]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 17]), sc2,
fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 48]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 49]), sc3,
fma(FLOAT_TYPE(data_b[b_offset + y2_idx ]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 1 ]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 16]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 17]), sc6,
(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 48]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 49])) * sc7)));
const uint tmp_idx = 16 * ix + tid;
tmp[tmp_idx] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, tmp[tmp_idx]));
fma(FLOAT_TYPE(by10.x) + FLOAT_TYPE(by10.y) + FLOAT_TYPE(by116.x) + FLOAT_TYPE(by116.y), sc2,
fma(FLOAT_TYPE(by132.x) + FLOAT_TYPE(by132.y) + FLOAT_TYPE(by148.x) + FLOAT_TYPE(by148.y), sc3,
fma(FLOAT_TYPE(by20.x) + FLOAT_TYPE(by20.y) + FLOAT_TYPE(by216.x) + FLOAT_TYPE(by216.y), sc6,
(FLOAT_TYPE(by232.x) + FLOAT_TYPE(by232.y) + FLOAT_TYPE(by248.x) + FLOAT_TYPE(by248.y)) * sc7)));
temp = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp));
}
tmp[gl_LocalInvocationID.x] = temp;
// sum up partial sums and write back result
barrier();
[[unroll]] for (uint s = 16; s > 0; s >>= 1) {

63
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp
View file

@ -1,5 +1,7 @@
#version 450
#extension GL_EXT_shader_explicit_arithmetic_types : require
#include "mul_mat_vec_base.comp"
layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
@ -40,35 +42,60 @@ void main() {
const uint s_offset = 8*v_im + is;
const uint y_offset = 128*v_im + l0;
tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
const uint y_idx = i * QUANT_K + y_offset;
const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
#if K_QUANTS_PER_ITERATION == 1
const uint tmp_idx = 16 * ix + tid;
tmp[tmp_idx] = fma(FLOAT_TYPE(data_b[b_offset + y_idx + 0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 0] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 0] & 0x03) << 4)) - 32),
fma(FLOAT_TYPE(data_b[b_offset + y_idx + 16]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 1]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x03) << 4)) - 32),
fma(FLOAT_TYPE(data_b[b_offset + y_idx + 32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 0] & 0x0c) << 2)) - 32),
fma(FLOAT_TYPE(data_b[b_offset + y_idx + 48]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 3]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x0c) << 2)) - 32),
fma(FLOAT_TYPE(data_b[b_offset + y_idx + 64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 0] >> 4) | ((data_a[ib0 + i].qh[qh_offset + 0] & 0x30) >> 0)) - 32),
fma(FLOAT_TYPE(data_b[b_offset + y_idx + 80]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 5]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16] >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x30) >> 0)) - 32),
fma(FLOAT_TYPE(data_b[b_offset + y_idx + 96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32] >> 4) | ((data_a[ib0 + i].qh[qh_offset + 0] & 0xc0) >> 2)) - 32),
fma(FLOAT_TYPE(data_b[b_offset + y_idx +112]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 7]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48] >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0xc0) >> 2)) - 32), tmp[tmp_idx]))))))));
#else
FLOAT_TYPE scales[4];
scales[0] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]);
scales[1] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]);
scales[2] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]);
scales[3] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]);
uint32_t ql0_u32 = uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 1]) << 16);
uint32_t ql32_u32 = uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 16]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 17]) << 16);
uint32_t ql0_u32_lo4 = ql0_u32 & 0x0F0F0F0F;
uint32_t ql0_u32_hi4 = (ql0_u32 >> 4) & 0x0F0F0F0F;
uint32_t ql32_u32_lo4 = ql32_u32 & 0x0F0F0F0F;
uint32_t ql32_u32_hi4 = (ql32_u32 >> 4) & 0x0F0F0F0F;
uint32_t qh_u32 = uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2 + 1]) << 16);
uint32_t qh0_u32 = (qh_u32 & 0x03030303) << 4;
uint32_t qh2_u32 = (qh_u32 & 0x0C0C0C0C) << 2;
uint32_t qh4_u32 = (qh_u32 & 0x30303030) << 0;
uint32_t qh6_u32 = (qh_u32 & 0xC0C0C0C0) >> 2;
uint32_t q0_u32 = ql0_u32_lo4 | qh0_u32;
uint32_t q1_u32 = ql32_u32_lo4 | qh2_u32;
uint32_t q2_u32 = ql0_u32_hi4 | qh4_u32;
uint32_t q3_u32 = ql32_u32_hi4 | qh6_u32;
uvec4 q0 = uvec4(unpack8(q0_u32));
uvec4 q1 = uvec4(unpack8(q1_u32));
uvec4 q2 = uvec4(unpack8(q2_u32));
uvec4 q3 = uvec4(unpack8(q3_u32));
B_TYPE_VEC4 by0 = data_b_v4[(b_offset + y_idx) / 4];
B_TYPE_VEC4 by32 = data_b_v4[(b_offset + y_idx) / 4 + 8];
B_TYPE_VEC4 by64 = data_b_v4[(b_offset + y_idx) / 4 + 16];
B_TYPE_VEC4 by96 = data_b_v4[(b_offset + y_idx) / 4 + 24];
FLOAT_TYPE sum = FLOAT_TYPE(0.0);
[[unroll]] for (int l = 0; l < 4; ++l) {
sum = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+ 0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 0) & 3) << 4)) - 32),
fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 2) & 3) << 4)) - 32),
fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0] >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 4) & 3) << 4)) - 32),
fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32] >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 6) & 3) << 4)) - 32), sum))));
sum = fma(FLOAT_TYPE(by0[l]) * scales[0], FLOAT_TYPE(int8_t(q0[l]) - 32),
fma(FLOAT_TYPE(by32[l]) * scales[1], FLOAT_TYPE(int8_t(q1[l]) - 32),
fma(FLOAT_TYPE(by64[l]) * scales[2], FLOAT_TYPE(int8_t(q2[l]) - 32),
fma(FLOAT_TYPE(by96[l]) * scales[3], FLOAT_TYPE(int8_t(q3[l]) - 32), sum))));
}
tmp[16 * ix + tid] += sum;
#endif
temp += sum * d;
}
tmp[gl_LocalInvocationID.x] = temp;
// sum up partial sums and write back result
barrier();
[[unroll]] for (uint s = 16; s > 0; s >>= 1) {