mirror of
https://github.com/ggml-org/llama.cpp.git
synced 2026-11-11 14:22:09 +01:00
AclyandJeff Bolz
e29acf74fe
* vulkan (DRAFT): split shader generation by GLSL source file, to improve incremental build times * support dep-files so shaders are recompiled if their included files change * rename shader files which are used as "headers" to use .glsl extension * move glslc extension detection shaders to separate folders * the above is to prevent them from getting glob'd with the actual compute shaders that need to be compiled * vulkan : only write embedded shader .hpp/.cpp when they change * avoid recompiling ggml-vulkan.cpp when editing shaders * pass single --source argument instead of --input-dir & --filter to shader gen * check for source file match earlier * fix hang in vulkan-shaders-gen when there are compilation errors * early out did not decrement compile_count * clean up * fix glslc integer dot product test * unconditionally write the embedded shader cpp output * replace output filepath in generated dep-files to match output in CMakeLists --------- Co-authored-by: Jeff Bolz <jbolz@nvidia.com>
36 lines
1.1 KiB
Plaintext
36 lines
1.1 KiB
Plaintext
#version 450
|
|
|
|
#include "dequant_head.glsl"
|
|
|
|
layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
|
|
|
|
layout (binding = 0) readonly buffer A {block_iq1_s data_a[];};
|
|
layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
|
|
|
|
void main() {
|
|
// Each thread handles 1 subblock (32 values with 2 scales)
|
|
const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8;
|
|
|
|
init_iq_shmem(gl_WorkGroupSize);
|
|
|
|
if (ib >= p.nel / 256) {
|
|
return;
|
|
}
|
|
|
|
const uint ib32 = gl_LocalInvocationID.x % 8;
|
|
const uint b_idx = 256 * ib + 32 * ib32;
|
|
|
|
uint qh = data_a[ib].qh[ib32];
|
|
const float d = float(data_a[ib].d);
|
|
const float dl = d * float(2 * bitfieldExtract(qh, 12, 3) + 1);
|
|
const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
|
|
[[unroll]] for (uint l = 0; l < 4; ++l) {
|
|
const uint qs = data_a[ib].qs[4 * ib32 + l];
|
|
const uint hi = bitfieldExtract(qh, 3 * int(l), 3);
|
|
const int16_t grid = int16_t(iq1s_grid[qs | (hi << 8)]);
|
|
[[unroll]] for (int j = 0; j < 8; ++j) {
|
|
data_b[b_idx + 8 * l + j] = D_TYPE(dl * (bitfieldExtract(grid, 2*j, 2) + delta));
|
|
}
|
|
}
|
|
}
|