ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

vulkan: multi-row k quants (#10846)

Browse source 
* multi row k quant shaders!

* better row selection

* more row choices

* readjust row selection

* rm_kq=2 by default
This commit is contained in:
Eve 2024-12-26 10:54:44 -05:00 • committed by GitHub
parent d283d02bf2
commit d79d8f39b4
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
6 changed files with 472 additions and 367 deletions
Download patch file Download diff file Expand all files Collapse all files

174
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp
View file

@ -7,21 +7,15 @@
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout (constant_id = 0) const uint BLOCK_SIZE = 32;
layout (constant_id = 1) const uint NUM_ROWS = 1;
shared FLOAT_TYPE tmp[BLOCK_SIZE];
void main() {
const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
if (row >= p.stride_d) {
return;
}
shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE];
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;
const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
// 16 threads are used to process each block
const uint it_size = gl_WorkGroupSize.x/16;
@ -31,8 +25,8 @@ void main() {
const uint step = 4;
const uint il = itid/step; // 0...3
const uint ir = itid - step*il; // 0...7 or 0...3
const uint il = itid/step; // 0...3
const uint ir = itid - step*il; // 0...7 or 0...3
const uint n = 4;
const uint v_im = il / 2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
@ -42,90 +36,116 @@ void main() {
const uint q_offset = 32*v_im + l0;
const uint y_offset = 64*v_im + l0;
FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
FLOAT_TYPE temp[NUM_ROWS];
[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
temp[i] = FLOAT_TYPE(0);
}
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
const uint y1_idx = i * QUANT_K + y_offset;
const uint y2_idx = y1_idx + 128;
f16vec2 d = data_a[ib0 + i].d;
const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
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 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_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4];
uint32_t qs64_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4 + 16];
uint32_t qs0_u32_lo4 = qs0_u32 & 0x0F0F0F0F;
uint32_t qs0_u32_hi4 = (qs0_u32 >> 4) & 0x0F0F0F0F;
uint32_t qs64_u32_lo4 = qs64_u32 & 0x0F0F0F0F;
uint32_t qs64_u32_hi4 = (qs64_u32 >> 4) & 0x0F0F0F0F;
uvec4 qs0_lo4 = uvec4(unpack8(qs0_u32_lo4));
uvec4 qs64_lo4 = uvec4(unpack8(qs64_u32_lo4));
uvec4 qs0_hi4 = uvec4(unpack8(qs0_u32_hi4));
uvec4 qs64_hi4 = uvec4(unpack8(qs64_u32_hi4));
const uint32_t q4_0 = qs0_lo4.x;
const uint32_t q4_1 = qs0_lo4.y;
const uint32_t q4_2 = qs0_lo4.z;
const uint32_t q4_3 = qs0_lo4.w;
const uint32_t q4_4 = qs0_hi4.x;
const uint32_t q4_5 = qs0_hi4.y;
const uint32_t q4_6 = qs0_hi4.z;
const uint32_t q4_7 = qs0_hi4.w;
const uint32_t q4_8 = qs64_lo4.x;
const uint32_t q4_9 = qs64_lo4.y;
const uint32_t q4_10 = qs64_lo4.z;
const uint32_t q4_11 = qs64_lo4.w;
const uint32_t q4_12 = qs64_hi4.x;
const uint32_t q4_13 = qs64_hi4.y;
const uint32_t q4_14 = qs64_hi4.z;
const uint32_t q4_15 = qs64_hi4.w;
B_TYPE_VEC4 by10 = data_b_v4[(b_offset + y1_idx) / 4];
B_TYPE_VEC4 by132 = data_b_v4[(b_offset + y1_idx) / 4 + 8];
B_TYPE_VEC4 by20 = data_b_v4[(b_offset + y2_idx) / 4];
B_TYPE_VEC4 by232 = data_b_v4[(b_offset + y2_idx) / 4 + 8];
const FLOAT_TYPE sx = fma(FLOAT_TYPE(by10.x), q4_0, fma(FLOAT_TYPE(by10.y), q4_1, fma(FLOAT_TYPE(by10.z), q4_2, FLOAT_TYPE(by10.w) * q4_3)));
const FLOAT_TYPE sy = fma(FLOAT_TYPE(by132.x), q4_4, fma(FLOAT_TYPE(by132.y), q4_5, fma(FLOAT_TYPE(by132.z), q4_6, FLOAT_TYPE(by132.w) * q4_7)));
const FLOAT_TYPE sz = fma(FLOAT_TYPE(by20.x), q4_8, fma(FLOAT_TYPE(by20.y), q4_9, fma(FLOAT_TYPE(by20.z), q4_10, FLOAT_TYPE(by20.w) * q4_11)));
const FLOAT_TYPE sw = fma(FLOAT_TYPE(by232.x), q4_12, fma(FLOAT_TYPE(by232.y), q4_13, fma(FLOAT_TYPE(by232.z), q4_14, FLOAT_TYPE(by232.w) * q4_15)));
const FLOAT_TYPE smin =
fma(FLOAT_TYPE(by10.x), sc2, fma(FLOAT_TYPE(by132.x), sc3, fma(FLOAT_TYPE(by20.x), sc6, fma(FLOAT_TYPE(by232.x), sc7,
fma(FLOAT_TYPE(by10.y), sc2, fma(FLOAT_TYPE(by132.y), sc3, fma(FLOAT_TYPE(by20.y), sc6, fma(FLOAT_TYPE(by232.y), sc7,
fma(FLOAT_TYPE(by10.z), sc2, fma(FLOAT_TYPE(by132.z), sc3, fma(FLOAT_TYPE(by20.z), sc6, fma(FLOAT_TYPE(by232.z), sc7,
fma(FLOAT_TYPE(by10.w), sc2, fma(FLOAT_TYPE(by132.w), sc3, fma(FLOAT_TYPE(by20.w), sc6, FLOAT_TYPE(by232.w) * sc7)))))))))))))));
temp = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp));
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
f16vec2 d = data_a[ib0 + i].d;
const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
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 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_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4];
uint32_t qs64_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4 + 16];
uint32_t qs0_u32_lo4 = qs0_u32 & 0x0F0F0F0F;
uint32_t qs0_u32_hi4 = (qs0_u32 >> 4) & 0x0F0F0F0F;
uint32_t qs64_u32_lo4 = qs64_u32 & 0x0F0F0F0F;
uint32_t qs64_u32_hi4 = (qs64_u32 >> 4) & 0x0F0F0F0F;
uvec4 qs0_lo4 = uvec4(unpack8(qs0_u32_lo4));
uvec4 qs64_lo4 = uvec4(unpack8(qs64_u32_lo4));
uvec4 qs0_hi4 = uvec4(unpack8(qs0_u32_hi4));
uvec4 qs64_hi4 = uvec4(unpack8(qs64_u32_hi4));
const uint32_t q4_0 = qs0_lo4.x;
const uint32_t q4_1 = qs0_lo4.y;
const uint32_t q4_2 = qs0_lo4.z;
const uint32_t q4_3 = qs0_lo4.w;
const uint32_t q4_4 = qs0_hi4.x;
const uint32_t q4_5 = qs0_hi4.y;
const uint32_t q4_6 = qs0_hi4.z;
const uint32_t q4_7 = qs0_hi4.w;
const uint32_t q4_8 = qs64_lo4.x;
const uint32_t q4_9 = qs64_lo4.y;
const uint32_t q4_10 = qs64_lo4.z;
const uint32_t q4_11 = qs64_lo4.w;
const uint32_t q4_12 = qs64_hi4.x;
const uint32_t q4_13 = qs64_hi4.y;
const uint32_t q4_14 = qs64_hi4.z;
const uint32_t q4_15 = qs64_hi4.w;
const FLOAT_TYPE sx = fma(FLOAT_TYPE(by10.x), q4_0, fma(FLOAT_TYPE(by10.y), q4_1, fma(FLOAT_TYPE(by10.z), q4_2, FLOAT_TYPE(by10.w) * q4_3)));
const FLOAT_TYPE sy = fma(FLOAT_TYPE(by132.x), q4_4, fma(FLOAT_TYPE(by132.y), q4_5, fma(FLOAT_TYPE(by132.z), q4_6, FLOAT_TYPE(by132.w) * q4_7)));
const FLOAT_TYPE sz = fma(FLOAT_TYPE(by20.x), q4_8, fma(FLOAT_TYPE(by20.y), q4_9, fma(FLOAT_TYPE(by20.z), q4_10, FLOAT_TYPE(by20.w) * q4_11)));
const FLOAT_TYPE sw = fma(FLOAT_TYPE(by232.x), q4_12, fma(FLOAT_TYPE(by232.y), q4_13, fma(FLOAT_TYPE(by232.z), q4_14, FLOAT_TYPE(by232.w) * q4_15)));
const FLOAT_TYPE smin =
fma(FLOAT_TYPE(by10.x), sc2, fma(FLOAT_TYPE(by132.x), sc3, fma(FLOAT_TYPE(by20.x), sc6, fma(FLOAT_TYPE(by232.x), sc7,
fma(FLOAT_TYPE(by10.y), sc2, fma(FLOAT_TYPE(by132.y), sc3, fma(FLOAT_TYPE(by20.y), sc6, fma(FLOAT_TYPE(by232.y), sc7,
fma(FLOAT_TYPE(by10.z), sc2, fma(FLOAT_TYPE(by132.z), sc3, fma(FLOAT_TYPE(by20.z), sc6, fma(FLOAT_TYPE(by232.z), sc7,
fma(FLOAT_TYPE(by10.w), sc2, fma(FLOAT_TYPE(by132.w), sc3, fma(FLOAT_TYPE(by20.w), sc6, FLOAT_TYPE(by232.w) * sc7)))))))))))))));
temp[n] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp[n]));
}
}
tmp[gl_LocalInvocationID.x] = temp;
// sum up partial sums and write back result
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] = temp[n];
}
barrier();
[[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
[[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] += tmpsh[n][tid + s];
}
}
barrier();
}
if (tid == 0) {
data_d[d_offset + row] = D_TYPE(tmp[0]);
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
data_d[d_offset + first_row + n] = D_TYPE(tmpsh[n][0]);
}
}
}
void main() {
const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
// do NUM_ROWS at a time, unless there aren't enough remaining rows
if (first_row + NUM_ROWS <= p.stride_d) {
compute_outputs(first_row, NUM_ROWS);
} else {
if (first_row >= p.stride_d) {
return;
}
compute_outputs(first_row, p.stride_d - first_row);
}
}