@@ -10,6 +10,7 @@ layout (constant_id = 0) const uint BLOCK_SIZE = 32;
1010layout (constant_id = 1) const uint NUM_ROWS = 1;
1111
1212shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE];
13+ shared FLOAT_TYPE sccache[BLOCK_SIZE/16][16];
1314shared block_q6_K_packed16 blkcache[BLOCK_SIZE/16];
1415
1516void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
@@ -57,12 +58,13 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
5758
5859 // cache full superblock into shared memory with coalesced reads
5960 // we assume 64 threads here!
60- [[unroll]] for (int l = 0; (l < 4) && (i0 + l < num_blocks_per_row); ++l) {
61- blkcache[l].ql[tid] = data_a_packed16[ib0 + i0 + l].ql[tid];
62- // hacky method of reading beyond ql and the block struct size but it looks like vulkan doesn't care? o_O
63- // this assumes that the struct is packed in continous 16 bit blocks to work
64- blkcache[l ].ql[64 + tid ] = data_a_packed16[ib0 + i0 + l ].ql[64 + tid ];
61+ //
62+ // hacky method of reading beyond ql and the block struct size but it looks like vulkan doesn't care? o_O
63+ // this assumes that the struct is packed in continous 16 bit blocks to work
64+ [[unroll]] for (int l = 0; l < 7; ++l) {
65+ blkcache[0 ].ql[tid + 64*l ] = data_a_packed16[ib0 + i0].ql[tid + 64*l ];
6566 }
67+ sccache[ix][itid] = FLOAT_TYPE(blkcache[ix].scales[itid]);
6668 barrier();
6769 if (i >= num_blocks_per_row)
6870 continue;
@@ -100,7 +102,7 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
100102 }
101103
102104 [[unroll]] for (int l = 0; l < 4; ++l)
103- sum[l] *= FLOAT_TYPE(blkcache [ix].scales [s_offset + l*2]) ;
105+ sum[l] *= sccache [ix][s_offset + l*2];
104106 temp[n] += (sum[0] + sum[1] + sum[2] + sum[3]) * FLOAT_TYPE(blkcache[ix].d);
105107 }
106108 }
0 commit comments