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

go all the way

Browse source
This commit is contained in:
Eve 2024-12-27 15:07:34 -05:00
parent a56504fd7b
commit f833b75175
3 changed files with 8 additions and 19 deletions
Download patch file Download diff file Expand all files Collapse all files

3
ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp
View file

@ -7,9 +7,6 @@
#if defined(A_TYPE_PACKED16)
layout (binding = 0) readonly buffer A_PACKED16 {A_TYPE_PACKED16 data_a_packed16[];};
#endif
#if defined(A_TYPE_PACKED16_FLAT)
layout (binding = 0) readonly buffer A_PACKED16_FLAT {A_TYPE_PACKED16_FLAT data_a_packed16_flat[];};
#endif
#if defined(A_TYPE_PACKED32)
layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32[];};
#endif

16
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp
View file

@ -10,7 +10,7 @@ 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];
shared block_q6_K_packed16_flat blkcache[BLOCK_SIZE/16];
shared block_q6_K_packed16 blkcache[BLOCK_SIZE/16];
void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
uint a_offset, b_offset, d_offset;
@ -22,7 +22,7 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
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 ix = tid/16;
const uint v_im = itid/8; // 0 or 1. 0 computes 0..., 1 computes 128...
const uint v_in = itid - 8*v_im; // 0...15 or 0...7
@ -58,24 +58,24 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
// cache full superblock into shared memory with coalesced reads
// we assume 64 threads here!
[[unroll]] for (int l = 0; (l < 4) && (i0 + l < num_blocks_per_row); ++l) {
blkcache[l].blkd[tid] = data_a_packed16_flat[ib0 + i0 + l].blkd[tid];
// we read beyond the struct size but it looks like vulkan doesn't care? o_O
blkcache[l].ql[tid] = data_a_packed16[ib0 + i0 + l].ql[tid];
// hacky method of reading beyond ql and the block struct size but it looks like vulkan doesn't care? o_O
// this assumes that the struct is packed in continous 16 bit blocks to work
blkcache[l].blkd[64 + tid] = data_a_packed16_flat[ib0 + i0 + l].blkd[64 + tid];
blkcache[l].ql[64 + tid] = data_a_packed16[ib0 + i0 + l].ql[64 + tid];
}
barrier();
if (i >= num_blocks_per_row)
continue;
uint32_t ql0_u32 = uint32_t(blkcache[ix].blkd[ql_offset / 2]) | (uint32_t(blkcache[ix].blkd[ql_offset / 2 + 1]) << 16);
uint32_t ql32_u32 = uint32_t(blkcache[ix].blkd[ql_offset / 2 + 16]) | (uint32_t(blkcache[ix].blkd[ql_offset / 2 + 17]) << 16);
uint32_t ql0_u32 = uint32_t(blkcache[ix].ql[ql_offset / 2]) | (uint32_t(blkcache[ix].ql[ql_offset / 2 + 1]) << 16);
uint32_t ql32_u32 = uint32_t(blkcache[ix].ql[ql_offset / 2 + 16]) | (uint32_t(blkcache[ix].ql[ql_offset / 2 + 17]) << 16);
uint32_t ql0_u32_lo4 = ql0_u32 & 0x0F0F0F0F;
uint32_t ql0_u32_hi4 = (ql0_u32 >> 4) & 0x0F0F0F0F;
uint32_t ql32_u32_lo4 = ql32_u32 & 0x0F0F0F0F;
uint32_t ql32_u32_hi4 = (ql32_u32 >> 4) & 0x0F0F0F0F;
uint32_t qh_u32 = uint32_t(blkcache[ix].blkd[64 + qh_offset / 2]) | (uint32_t(blkcache[ix].blkd[64 + qh_offset / 2 + 1]) << 16);
uint32_t qh_u32 = uint32_t(blkcache[ix].qh[qh_offset / 2]) | (uint32_t(blkcache[ix].qh[qh_offset / 2 + 1]) << 16);
uint32_t qh0_u32 = (qh_u32 & 0x03030303) << 4;
uint32_t qh2_u32 = (qh_u32 & 0x0C0C0C0C) << 2;
uint32_t qh4_u32 = (qh_u32 & 0x30303030);

8
ggml/src/ggml-vulkan/vulkan-shaders/types.comp
View file

@ -273,18 +273,10 @@ struct block_q6_K_packed16
float16_t d;
};
struct block_q6_K_packed16_flat
{
uint16_t blkd[96];
int8_t scales[16];
float16_t d;
};
#if defined(DATA_A_Q6_K)
#define QUANT_K QUANT_K_Q6_K
#define A_TYPE block_q6_K
#define A_TYPE_PACKED16 block_q6_K_packed16
#define A_TYPE_PACKED16_FLAT block_q6_K_packed16_flat
#endif
// IQuants