vulkan: work around Intel fp16 bug in mmq (llama/18814)

diff --git a/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl b/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl
index 7f32dadf17d8496e691749ebbcee5c29f84ef73a..9c297d1c60d49dc67834549abd542c71e994a47f 100644 (file)
--- a/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl
+++ b/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl
@@ -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(data_a_packed16[ib_k].d) * FLOAT_TYPE_VEC2(scales - 32);
+        buf_a[buf_ib].d_scales = FLOAT_TYPE_VEC2(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(data_a_packed32[ib_k].dm) * FLOAT_TYPE_VEC2(scale_dm);
+        buf_a[buf_ib].dm = FLOAT_TYPE_VEC2(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(data_a_packed16[ib_k].d) * FLOAT_TYPE_VEC2(scales);
+        buf_a[buf_ib].d_scales = FLOAT_TYPE_VEC2(float(data_a_packed16[ib_k].d) * vec2(scales));
     }
 }