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

vulkan: optimize mul_mat for small values of N (#10991)

Browse source 
Make the mul_mat_vec shaders support N>1 (as a spec constant, NUM_COLS) where
the batch_strides are overloaded to hold the row strides. Put the loads from the
B matrix in the innermost loop because it should cache better.

Share some code for reducing the result values to memory in mul_mat_vec_base.
This commit is contained in:
Jeff Bolz 2024-12-30 11:27:11 -06:00 • committed by GitHub
parent c250ecb315
commit 716bd6dec3
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
9 changed files with 288 additions and 349 deletions
Download patch file Download diff file Expand all files Collapse all files

58
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp
View file

@ -6,11 +6,6 @@
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 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);
@ -36,20 +31,17 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
const uint s_offset = 8*v_im + is;
const uint y_offset = 128*v_im + l0;
FLOAT_TYPE temp[NUM_ROWS];
FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
temp[i] = FLOAT_TYPE(0);
[[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) {
const uint y_idx = i * QUANT_K + y_offset;
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];
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
@ -84,35 +76,25 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
uvec4 q2 = uvec4(unpack8(q2_u32));
uvec4 q3 = uvec4(unpack8(q3_u32));
FLOAT_TYPE sum = FLOAT_TYPE(0.0);
[[unroll]] for (int l = 0; l < 4; ++l) {
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))));
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
B_TYPE_VEC4 by0 = data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4];
B_TYPE_VEC4 by32 = data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 8];
B_TYPE_VEC4 by64 = data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 16];
B_TYPE_VEC4 by96 = data_b_v4[(j*p.batch_stride_b + 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(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))));
}
temp[j][n] += sum * d;
}
temp[n] += sum * d;
}
}
// 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 = BLOCK_SIZE/2; s > 0; s >>= 1) {
if (tid < s) {
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] += tmpsh[n][tid + s];
}
}
barrier();
}
if (tid == 0) {
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
data_d[d_offset + first_row + n] = D_TYPE(tmpsh[n][0]);
}
}
reduce_result(temp, d_offset, first_row, num_rows, tid);
}
void main() {