Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

vulkan: unify type macros to use Vx instead of _VECx (#21605)

Browse Source
This commit is contained in:
Ruben Ortlam
2026-04-09 07:31:51 +02:00
committed by GitHub
parent 4293919068
commit 8a132faaa0
10 changed files with 167 additions and 167 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl
+8 -8
View File
@@ -21,7 +21,7 @@ void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
buf_a[buf_ib].qs[iqs] = data_a_packed32[ib].qs[iqs];
if (iqs == 0) {
buf_a[buf_ib].dm = FLOAT_TYPE_VEC2(data_a_packed32[ib].dm);
buf_a[buf_ib].dm = FLOAT_TYPEV2(data_a_packed32[ib].dm);
}
#endif
}
@@ -72,7 +72,7 @@ void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
buf_a[buf_ib].qs[iqs] = data_a_packed32[ib].qs[iqs];
if (iqs == 0) {
buf_a[buf_ib].dm = FLOAT_TYPE_VEC2(data_a_packed32[ib].dm);
buf_a[buf_ib].dm = FLOAT_TYPEV2(data_a_packed32[ib].dm);
buf_a[buf_ib].qh = data_a_packed32[ib].qh;
}
#endif
@@ -203,7 +203,7 @@ void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
buf_a[buf_ib].qs[iqs] = vals0 | (vals1 << 2) | (vals2 << 4) | (vals3 << 6);
if (iqs == 0) {
buf_a[buf_ib].dm = FLOAT_TYPE_VEC2(data_a_packed32[ib_k].dm);
buf_a[buf_ib].dm = FLOAT_TYPEV2(data_a_packed32[ib_k].dm);
buf_a[buf_ib].scales = unpack8(uint32_t(data_a_packed16[ib_k].scales[iqs_k / 8])).xy; // vec4 used due to #12147
}
}
@@ -264,7 +264,7 @@ void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
const i8vec2 scales = i8vec2(unpack8(uint32_t(((data_a_packed16[ib_k].scales[(is % 8 ) / 2] >> (4 * (is / 8))) & 0x0F0F) |
(((data_a_packed16[ib_k].scales[(8 + (is % 4)) / 2] >> (2 * (is / 4))) & 0x0303) << 4))).xy); // vec4 used due to #12147
buf_a[buf_ib].d_scales = FLOAT_TYPE_VEC2(float(data_a_packed16[ib_k].d) * vec2(scales - 32));
buf_a[buf_ib].d_scales = FLOAT_TYPEV2(float(data_a_packed16[ib_k].d) * vec2(scales - 32));
}
}
@@ -334,7 +334,7 @@ void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
(data_a[ib_k].scales[is+4] >> 4) | ((data_a[ib_k].scales[is ] & 0xC0) >> 2));
}
buf_a[buf_ib].dm = FLOAT_TYPE_VEC2(vec2(data_a_packed32[ib_k].dm) * vec2(scale_dm));
buf_a[buf_ib].dm = FLOAT_TYPEV2(vec2(data_a_packed32[ib_k].dm) * vec2(scale_dm));
}
}
@@ -385,7 +385,7 @@ void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
const uint is = iqs_k / 4;
const i8vec2 scales = unpack8(int32_t(data_a_packed16[ib_k].scales[is / 2])).xy;
buf_a[buf_ib].d_scales = FLOAT_TYPE_VEC2(float(data_a_packed16[ib_k].d) * vec2(scales));
buf_a[buf_ib].d_scales = FLOAT_TYPEV2(float(data_a_packed16[ib_k].d) * vec2(scales));
}
}
@@ -426,7 +426,7 @@ void block_b_to_shmem(const uint buf_ib, const uint ib, const uint iqs, const bo
const uint ib_inner = ib % 4;
if (iqs == 0) {
buf_b[buf_ib].ds = FLOAT_TYPE_VEC2(data_b[ib_outer].ds[ib_inner]);
buf_b[buf_ib].ds = FLOAT_TYPEV2(data_b[ib_outer].ds[ib_inner]);
}
const ivec4 values = data_b[ib_outer].qs[ib_inner * 2 + iqs];
@@ -436,7 +436,7 @@ void block_b_to_shmem(const uint buf_ib, const uint ib, const uint iqs, const bo
buf_b[buf_ib].qs[iqs * 4 + 3] = values.w;
} else {
if (iqs == 0) {
buf_b[buf_ib].ds = FLOAT_TYPE_VEC2(0.0f);
buf_b[buf_ib].ds = FLOAT_TYPEV2(0.0f);
}
buf_b[buf_ib].qs[iqs * 4 ] = 0;