KBaba7/llama.cpp

1#version 4502 3#include "dequant_head.comp"4 5layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;6 7layout (binding = 0) readonly buffer A {A_TYPE data_a[];};8layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};9 10void main() {11    [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {12        const uint ib = gl_WorkGroupID.x * 256 + wgy;13        if (ib >= p.M * p.K / QUANT_K) {14            return;15        }16 17        const uint tid = gl_LocalInvocationID.x;18        const uint il = tid / 8;19        const uint ir = tid % 8;20        const uint is = 2 * il;21        const uint n = 4;22 23        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib].d.x);24        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib].d.y);25 26        const uint y_idx = ib * QUANT_K + 64 * il + n * ir;27        const uint qs_idx = 32*il + n * ir;28 29        uint scidx0 = (is < 4) ? is : (is + 4);30        uint scidx1 = (is < 4) ? is : (is - 4);31        uint scidxmask1 = (is < 4) ? 0x30 : 0xC0;32        uint scidxshift1 = (is < 4) ? 0 : 2;33        uint mbidx0 = is + 4;34        uint mbidx1 = (is < 4) ? is + 4 : is;35        uint mbidxmask0 = (is < 4) ? 0xF : 0xF0;36        uint mbidxshift0 = (is < 4) ? 0 : 4;37        uint mbidxmask1 = (is < 4) ? 0x30 : 0xC0;38        uint mbidxshift1 = (is < 4) ? 0 : 2;39 40        uint8_t sc = uint8_t((data_a[ib].scales[scidx0] & 0xF) | ((data_a[ib].scales[scidx1] & scidxmask1) >> scidxshift1));41        uint8_t mbyte = uint8_t((data_a[ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0 | ((data_a[ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1));42 43        const FLOAT_TYPE d1 = dall * sc;44        const FLOAT_TYPE m1 = dmin * mbyte;45 46        scidx0 = (is < 4) ? is + 1 : (is + 5);47        scidx1 = (is < 4) ? is + 1 : (is - 3);48        scidxmask1 = (is < 4) ? 0x30 : 0xC0;49        scidxshift1 = (is < 4) ? 0 : 2;50        mbidx0 = is + 5;51        mbidx1 = (is < 4) ? is + 5 : is + 1;52        mbidxmask0 = (is < 4) ? 0xF : 0xF0;53        mbidxshift0 = (is < 4) ? 0 : 4;54        mbidxmask1 = (is < 4) ? 0x30 : 0xC0;55        mbidxshift1 = (is < 4) ? 0 : 2;56 57        sc = uint8_t((data_a[ib].scales[scidx0] & 0xF) | ((data_a[ib].scales[scidx1] & scidxmask1) >> scidxshift1));58        mbyte = uint8_t((data_a[ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0 | ((data_a[ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1));59 60        const FLOAT_TYPE d2 = dall * sc;61        const FLOAT_TYPE m2 = dmin * mbyte;62 63        [[unroll]] for (uint l = 0; l < n; ++l) {64            data_b[y_idx + l     ] = D_TYPE(d1 * FLOAT_TYPE(data_a[ib].qs[qs_idx + l] & 0xF) - m1);65            data_b[y_idx + l + 32] = D_TYPE(d2 * FLOAT_TYPE(data_a[ib].qs[qs_idx + l] >>  4) - m2);66        }67    }68}69