vulkan: Replace 16-bit unpack8 calls to work around legacy Windows AMD driver bug (ggml-org#17285)

Author: 0cc4m

ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl

@@ -300,7 +300,7 @@ void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
 
     if (iqs == 0) {
         buf_a[buf_ib].dm = FLOAT_TYPE_VEC2(data_a_packed32[ib_k].dm);
-        buf_a[buf_ib].scales = unpack8(data_a_packed16[ib_k].scales[iqs_k / 8]);
+        buf_a[buf_ib].scales = unpack8(uint32_t(data_a_packed16[ib_k].scales[iqs_k / 8])).xy; // vec4 used due to #12147
     }
 }
 
@@ -345,21 +345,22 @@ void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
 
     // Repack 2x4 quants into one int
     // Add the 3rd bit instead of subtracting it to allow packing the quants
-    const i8vec2 vals00 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx * 2        ] >> qs_shift) & uint16_t(0x0303))) |
-                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[hm_idx * 2    ] >> hm_shift) & uint16_t(0x0101)) << 2));
-    const i8vec2 vals01 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx * 2 + 1    ] >> qs_shift) & uint16_t(0x0303))) |
-                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[hm_idx * 2 + 1] >> hm_shift) & uint16_t(0x0101)) << 2));
-    const i8vec2 vals10 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx * 2 + 2    ] >> qs_shift) & uint16_t(0x0303))) |
-                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[hm_idx * 2 + 2] >> hm_shift) & uint16_t(0x0101)) << 2));
-    const i8vec2 vals11 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx * 2 + 3    ] >> qs_shift) & uint16_t(0x0303))) |
-                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[hm_idx * 2 + 3] >> hm_shift) & uint16_t(0x0101)) << 2));
+    // vec4 for unpack8 used due to #12147
+    const i8vec2 vals00 = unpack8(int32_t(int16_t((data_a_packed16[ib_k].qs[qs_idx * 2        ] >> qs_shift) & uint16_t(0x0303)))).xy |
+                          unpack8(int32_t(int16_t(((data_a_packed16[ib_k].hmask[hm_idx * 2    ] >> hm_shift) & uint16_t(0x0101))) << 2)).xy;
+    const i8vec2 vals01 = unpack8(int32_t(int16_t((data_a_packed16[ib_k].qs[qs_idx * 2 + 1    ] >> qs_shift) & uint16_t(0x0303)))).xy |
+                          unpack8(int32_t(int16_t(((data_a_packed16[ib_k].hmask[hm_idx * 2 + 1] >> hm_shift) & uint16_t(0x0101))) << 2)).xy;
+    const i8vec2 vals10 = unpack8(int32_t(int16_t((data_a_packed16[ib_k].qs[qs_idx * 2 + 2    ] >> qs_shift) & uint16_t(0x0303)))).xy |
+                          unpack8(int32_t(int16_t(((data_a_packed16[ib_k].hmask[hm_idx * 2 + 2] >> hm_shift) & uint16_t(0x0101))) << 2)).xy;
+    const i8vec2 vals11 = unpack8(int32_t(int16_t((data_a_packed16[ib_k].qs[qs_idx * 2 + 3    ] >> qs_shift) & uint16_t(0x0303)))).xy |
+                          unpack8(int32_t(int16_t(((data_a_packed16[ib_k].hmask[hm_idx * 2 + 3] >> hm_shift) & uint16_t(0x0101))) << 2)).xy;
     buf_a[buf_ib].qs[iqs] = pack32(u8vec4(vals00.x, vals00.y, vals01.x, vals01.y)) |
                            (pack32(u8vec4(vals10.x, vals10.y, vals11.x, vals11.y)) << 4);
 
     if (iqs == 0) {
         const uint is = iqs_k / 4;
-        const i8vec2 scales = i8vec2(unpack8(((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)));
+        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(data_a_packed16[ib_k].d) * FLOAT_TYPE_VEC2(scales - 32);
     }
@@ -516,15 +517,15 @@ void block_a_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
     const uint qh_idx = (iqs_k / 32) * 8 + iqs;
     const uint qh_shift = ((iqs_k % 32) / 8) * 2;
 
-    const i8vec2 vals00 = (unpack8(int16_t((data_a_packed16[ib_k].ql[ql_idx * 2    ] >> ql_shift) & uint16_t(0x0F0F))) |
-                          unpack8(int16_t(((data_a_packed16[ib_k].qh[qh_idx * 2    ] >> qh_shift) & uint16_t(0x0303)) << 4))) - int8_t(32);
-    const i8vec2 vals01 = (unpack8(int16_t((data_a_packed16[ib_k].ql[ql_idx * 2 + 1] >> ql_shift) & uint16_t(0x0F0F))) |
-                          unpack8(int16_t(((data_a_packed16[ib_k].qh[qh_idx * 2 + 1] >> qh_shift) & uint16_t(0x0303)) << 4))) - int8_t(32);
+    const i8vec2 vals00 = (unpack8(int32_t((data_a_packed16[ib_k].ql[ql_idx * 2    ] >> ql_shift) & uint16_t(0x0F0F))).xy |
+                          unpack8(int32_t(((data_a_packed16[ib_k].qh[qh_idx * 2    ] >> qh_shift) & uint16_t(0x0303)) << 4)).xy) - int8_t(32);
+    const i8vec2 vals01 = (unpack8(int32_t((data_a_packed16[ib_k].ql[ql_idx * 2 + 1] >> ql_shift) & uint16_t(0x0F0F))).xy |
+                          unpack8(int32_t(((data_a_packed16[ib_k].qh[qh_idx * 2 + 1] >> qh_shift) & uint16_t(0x0303)) << 4)).xy) - int8_t(32);
     buf_a[buf_ib].qs[iqs] = pack32(i8vec4(vals00.x, vals00.y, vals01.x, vals01.y));
 
     if (iqs == 0) {
         const uint is = iqs_k / 4;
-        const i8vec2 scales = unpack8(data_a_packed16[ib_k].scales[is / 2]);
+        const i8vec2 scales = unpack8(int32_t(data_a_packed16[ib_k].scales[is / 2])).xy;
 
         buf_a[buf_ib].d_scales = FLOAT_TYPE(data_a_packed16[ib_k].d) * FLOAT_TYPE_VEC2(scales);
     }