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

q2_k separate out

Browse source
This commit is contained in:
Eve 2025-01-09 21:41:50 -05:00
parent 973bc4069f
commit 6145fc79e5
2 changed files with 76 additions and 58 deletions
Download patch file Download diff file Expand all files Collapse all files

80
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp
View file

@ -7,44 +7,29 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
shared FLOAT_TYPE sccache[BLOCK_SIZE/16][2][16];
void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
uint a_offset, b_offset, d_offset;
get_offsets(a_offset, b_offset, d_offset);
const uint num_blocks_per_row = p.ncols / QUANT_K;
// 16 threads are used to process each block
const uint it_size = gl_WorkGroupSize.x/16;
const uint tid = gl_LocalInvocationID.x;
const uint itid = tid%16; // 0...15
const uint ix = tid/16;
const uint itid8 = itid%8;
const uint v_im = itid/8; // 0 or 1. 0 computes 0..., 1 computes 128...
const uint v_in = itid - 8*v_im; // 0...7
const uint l0 = 2*v_in; // 0...15
const uint q_offset = 32*v_im + l0;
const uint y_offset = 128*v_im + l0;
FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
temp[j][i] = FLOAT_TYPE(0);
}
}
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint itid8, const uint v_im, const uint ix, const uint q_offset, const uint y_offset, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows, const bool all_threads) {
const uint y_idx = i * QUANT_K + y_offset;
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
sccache[ix][0][itid] = FLOAT_TYPE(bitfieldExtract(uint(data_a[ib0 + i].scales[itid8]), int(v_im*4), 4)); // lower 8 bytes
sccache[ix][1][itid] = FLOAT_TYPE(bitfieldExtract(uint(data_a[ib0 + i].scales[itid8+8]), int(v_im*4), 4)); // upper 8 bytes
if (!all_threads) { // when we don't have enough blocks to use all threads
if (i < num_blocks_per_row) {
sccache[ix][0][itid] = FLOAT_TYPE((data_a[ib0 + i].scales[itid8] >> v_im*4) & 0xF); // lower 8 bytes
sccache[ix][1][itid] = FLOAT_TYPE((data_a[ib0 + i].scales[itid8+8] >> v_im*4) & 0xF); // upper 8 bytes
}
barrier();
if (i >= num_blocks_per_row)
continue;
} else {
sccache[ix][0][itid] = FLOAT_TYPE((data_a[ib0 + i].scales[itid8] >> v_im*4) & 0xF);
sccache[ix][1][itid] = FLOAT_TYPE((data_a[ib0 + i].scales[itid8+8] >> v_im*4) & 0xF);
barrier();
}
const uint32_t qs_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);
const vec4 qs_u32_0 = vec4(unpack8(qs_u32 & 0x03030303));
const vec4 qs_u32_2 = vec4(unpack8((qs_u32 >> 2) & 0x03030303));
@ -52,8 +37,8 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
const vec4 qs_u32_6 = vec4(unpack8((qs_u32 >> 6) & 0x03030303));
const f16vec2 d = data_a[ib0 + i].d;
const FLOAT_TYPE dall = d.x;
const FLOAT_TYPE dmin = d.y;
const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
B_TYPE_VEC2 b0 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2];
@ -90,6 +75,39 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
}
}
void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
uint a_offset, b_offset, d_offset;
get_offsets(a_offset, b_offset, d_offset);
const uint num_blocks_per_row = p.ncols / QUANT_K;
// 16 threads are used to process each block
const uint it_size = gl_WorkGroupSize.x/16;
const uint tid = gl_LocalInvocationID.x;
const uint itid = tid%16; // 0...15
const uint ix = tid/16;
const uint itid8 = itid%8;
const uint v_im = itid/8; // 0 or 1. 0 computes 0..., 1 computes 128...
const uint v_in = itid - 8*v_im; // 0...7
const uint l0 = 2*v_in; // 0...15
const uint q_offset = 32*v_im + l0;
const uint y_offset = 128*v_im + l0;
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
temp[j][i] = FLOAT_TYPE(0);
}
}
const uint nbr_par_th = num_blocks_per_row%it_size;
const uint nbr_all_th = num_blocks_per_row - nbr_par_th;
uint i0 = 0;
[[unroll]] for (; i0 < nbr_all_th; i0 += it_size)
calc_superblock(a_offset, b_offset, itid, itid8, v_im, ix, q_offset, y_offset, i0 + ix, num_blocks_per_row, first_row, num_rows, true);
calc_superblock(a_offset, b_offset, itid, itid8, v_im, ix, q_offset, y_offset, i0 + ix, num_blocks_per_row, first_row, num_rows, false);
reduce_result(temp, d_offset, first_row, num_rows, tid);
}

4
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp
View file

@ -16,7 +16,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint ix, co
if (!all_threads) { // when we don't have enough blocks to use all threads
if (i < num_blocks_per_row)
sccache[ix][v_im][itid8] = FLOAT_TYPE(int8_t(((data_a[ib0+i].scales[itid8] >> v_im4) & 0xF) | ((data_a[ib0+i].scales[itid8%4+8] >> s_shift & 3) << 4)) - 32);
sccache[ix][v_im][itid8] = FLOAT_TYPE(int8_t(((data_a[ib0+i].scales[itid8] >> v_im4) & 0xF) | (((data_a[ib0+i].scales[itid8%4+8] >> s_shift) & 3) << 4)) - 32);
barrier();
if (i >= num_blocks_per_row)
@ -38,7 +38,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint ix, co
const vec4 qs_u32_6 = vec4(unpack8((qs_u32 >> 6) & 0x03030303));
if (all_threads) {
sccache[ix][v_im][itid8] = FLOAT_TYPE(int8_t(((data_a[ib0+i].scales[itid8] >> v_im4) & 0xF) | ((data_a[ib0+i].scales[itid8%4+8] >> s_shift & 3) << 4)) - 32);
sccache[ix][v_im][itid8] = FLOAT_TYPE(int8_t(((data_a[ib0+i].scales[itid8] >> v_im4) & 0xF) | (((data_a[ib0+i].scales[itid8%4+8] >> s_shift) & 3) << 4)) - 32);
barrier();
}