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

test

Browse source
This commit is contained in:
Eve 2024-11-21 11:27:18 -05:00
parent 3c3543ab56
commit be5295bd51
1 changed files with 7 additions and 14 deletions
Download patch file Download diff file Expand all files Collapse all files

21
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp
View file

@ -1,5 +1,7 @@
#version 450 #version 450
#extension GL_KHR_shader_subgroup_arithmetic: enable
#ifdef FLOAT16 #ifdef FLOAT16
#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require #extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
#endif #endif
@ -10,8 +12,6 @@ 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 = 0) const uint BLOCK_SIZE = 32;
shared FLOAT_TYPE tmp[BLOCK_SIZE];
void main() { void main() {
const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z; const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
const uint tid = gl_LocalInvocationID.x; const uint tid = gl_LocalInvocationID.x;
@ -28,7 +28,7 @@ void main() {
const uint y_offset = QUANT_R == 1 ? 1 : QUANT_K/2; const uint y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;
tmp[tid] = FLOAT_TYPE(0.0f); FLOAT_TYPE tmp = FLOAT_TYPE(0.0f);
[[unroll]] for (uint i = 0; i < p.ncols/block_size; i += 2) { [[unroll]] for (uint i = 0; i < p.ncols/block_size; i += 2) {
const uint col = i*block_size + 2*tid; const uint col = i*block_size + 2*tid;
@ -39,18 +39,11 @@ void main() {
vec2 v = dequantize(ib, iqs, a_offset / QUANT_K); vec2 v = dequantize(ib, iqs, a_offset / QUANT_K);
// matrix multiplication // matrix multiplication
tmp[tid] = fma(FLOAT_TYPE(v.x), FLOAT_TYPE(data_b[b_offset + iybs + iqs]), fma(FLOAT_TYPE(v.y), FLOAT_TYPE(data_b[b_offset + iybs + iqs + y_offset]), tmp[tid])); tmp = fma(FLOAT_TYPE(v.x), FLOAT_TYPE(data_b[b_offset + iybs + iqs]), fma(FLOAT_TYPE(v.y), FLOAT_TYPE(data_b[b_offset + iybs + iqs + y_offset]), tmp));
} }
// sum up partial sums and write back result // sum up partial sums and write back result
barrier(); tmp = subgroupAdd(tmp);
[[unroll]] for (uint s = block_size/2; s > 0; s >>= 1) { if (tid == 0)
if (tid < s) { data_d[d_offset + row] = D_TYPE(tmp);
tmp[tid] += tmp[tid + s];
}
barrier();
}
if (tid == 0) {
data_d[d_offset + row] = D_TYPE(tmp[0]);
}
} }