diff --git a/src/gpu.cpp b/src/gpu.cpp
index 979e950eb..b59a315f3 100644
--- a/src/gpu.cpp
+++ b/src/gpu.cpp
@@ -1518,6 +1518,12 @@ int create_gpu_instance()
             gpu_info.support_fp16_arithmetic = true;
         }
 
+        if (physicalDeviceProperties.vendorID == 0x5143 && !gpu_info.support_fp16_storage)
+        {
+            // fp16 arithmetic yields wrong result on old adreno drivers :(
+            gpu_info.support_fp16_arithmetic = false;
+        }
+
         if (gpu_info.support_cooperative_matrix)
         {
             // query supported cooperative matrix types and operations
@@ -3354,7 +3360,7 @@ int compile_spirv_module(const char* comp_data, int comp_data_size, const Option
         custom_defines.push_back(std::make_pair("afpmat4", "mat4"));
     }
 
-    if (opt.use_fp16_arithmetic)
+    if (opt.use_fp16_storage && opt.use_fp16_arithmetic)
     {
         custom_defines.push_back(std::make_pair("lfp", "float16_t"));
         custom_defines.push_back(std::make_pair("lfpvec4", "f16vec4"));
@@ -3384,7 +3390,7 @@ int compile_spirv_module(const char* comp_data, int comp_data_size, const Option
         custom_defines.push_back(std::make_pair("sfp2lfpvec4(v)", "v"));
 
         custom_defines.push_back(std::make_pair("lfp2afp(v)", "float16_t(v)"));
-        custom_defines.push_back(std::make_pair("lfp2afpvec4(v)", "f16vec4(vec4(unpackHalf2x16(v.x),unpackHalf2x16(v.y)))"));
+        custom_defines.push_back(std::make_pair("lfp2afpvec4(v)", "f16vec4(unpackFloat2x16(v.x),unpackFloat2x16(v.y))"));
     }
     else if (opt.use_fp16_storage)
     {
@@ -3439,20 +3445,20 @@ int compile_spirv_module(const char* comp_data, int comp_data_size, const Option
         custom_defines.push_back(std::make_pair("buffer_ld1(buf,i)", "float16_t(buf[i])"));
         custom_defines.push_back(std::make_pair("buffer_st1(buf,i,v)", "{buf[i]=float(v);}"));
         custom_defines.push_back(std::make_pair("buffer_cp1(buf,i,sbuf,si)", "{buf[i]=sbuf[si];}"));
-        custom_defines.push_back(std::make_pair("buffer_cp1to4(buf,i,sbuf,si4)", "{buf[i]=uvec2(packHalf2x16(vec2(f16vec2(sbuf[si4.r],sbuf[si4.g]))),packHalf2x16(vec2(f16vec2(sbuf[si4.b],sbuf[si4.a]))));}"));
-        custom_defines.push_back(std::make_pair("buffer_cp1to8(buf,i,sbuf,si4,sii4)", "{buf[i]=uvec4(packHalf2x16(vec2(f16vec2(sbuf[si4.r],sbuf[si4.g]))),packHalf2x16(vec2(f16vec2(sbuf[si4.b],sbuf[si4.a]))),packHalf2x16(vec2(f16vec2(sbuf[sii4.r],sbuf[sii4.g]))),packHalf2x16(vec2(f16vec2(sbuf[sii4.b],sbuf[sii4.a]))));}"));
-        custom_defines.push_back(std::make_pair("buffer_ld2(buf,i)", "f16vec2(unpackHalf2x16(buf[i]))"));
-        custom_defines.push_back(std::make_pair("buffer_st2(buf,i,v)", "{buf[i]=packHalf2x16(vec2(v))}"));
+        custom_defines.push_back(std::make_pair("buffer_cp1to4(buf,i,sbuf,si4)", "{buf[i]=uvec2(packFloat2x16(f16vec2(sbuf[si4.r],sbuf[si4.g])),packFloat2x16(f16vec2(sbuf[si4.b],sbuf[si4.a])));}"));
+        custom_defines.push_back(std::make_pair("buffer_cp1to8(buf,i,sbuf,si4,sii4)", "{buf[i]=uvec4(packFloat2x16(f16vec2(sbuf[si4.r],sbuf[si4.g])),packFloat2x16(f16vec2(sbuf[si4.b],sbuf[si4.a])),packFloat2x16(f16vec2(sbuf[sii4.r],sbuf[sii4.g])),packFloat2x16(f16vec2(sbuf[sii4.b],sbuf[sii4.a])));}"));
+        custom_defines.push_back(std::make_pair("buffer_ld2(buf,i)", "unpackFloat2x16(buf[i])"));
+        custom_defines.push_back(std::make_pair("buffer_st2(buf,i,v)", "{buf[i]=packFloat2x16(v)}"));
         custom_defines.push_back(std::make_pair("buffer_cp2(buf,i,sbuf,si)", "{buf[i]=sbuf[si];}"));
-        custom_defines.push_back(std::make_pair("buffer_ld4(buf,i)", "f16vec4(vec4(unpackHalf2x16(buf[i].x),unpackHalf2x16(buf[i].y)))"));
-        custom_defines.push_back(std::make_pair("buffer_st4(buf,i,v)", "{buf[i]=uvec2(packHalf2x16(vec2(v.rg)),packHalf2x16(vec2(v.ba)));}"));
+        custom_defines.push_back(std::make_pair("buffer_ld4(buf,i)", "f16vec4(unpackFloat2x16(buf[i].x),unpackFloat2x16(buf[i].y))"));
+        custom_defines.push_back(std::make_pair("buffer_st4(buf,i,v)", "{buf[i]=uvec2(packFloat2x16(v.rg),packFloat2x16(v.ba));}"));
         custom_defines.push_back(std::make_pair("buffer_cp4(buf,i,sbuf,si)", "{buf[i]=sbuf[si];}"));
-        custom_defines.push_back(std::make_pair("buffer_cp4to1(buf,i4,sbuf,si)", "{uvec2 _v=sbuf[si]; vec2 _v0=unpackHalf2x16(_v.x);vec2 _v1=unpackHalf2x16(_v.y); buf[i4.r]=_v0.r;buf[i4.g]=_v0.g;buf[i4.b]=_v1.r;buf[i4.a]=_v1.g;}"));
+        custom_defines.push_back(std::make_pair("buffer_cp4to1(buf,i4,sbuf,si)", "{uvec2 _v=sbuf[si]; f16vec2 _v0=unpackFloat2x16(_v.x);f16vec2 _v1=unpackFloat2x16(_v.y); buf[i4.r]=_v0.r;buf[i4.g]=_v0.g;buf[i4.b]=_v1.r;buf[i4.a]=_v1.g;}"));
         custom_defines.push_back(std::make_pair("buffer_cp4to8(buf,i,sbuf,si2)", "{buf[i]=uvec4(sbuf[si2.r],sbuf[si2.g]);}"));
-        custom_defines.push_back(std::make_pair("buffer_ld8(buf,i)", "f16mat2x4(f16vec4(vec4(unpackHalf2x16(buf[i].r),unpackHalf2x16(buf[i].g))),f16vec4(vec4(unpackHalf2x16(buf[i].b),unpackHalf2x16(buf[i].a))))"));
-        custom_defines.push_back(std::make_pair("buffer_st8(buf,i,v)", "{buf[i]=uvec4(uvec2(packHalf2x16(vec2(v[0].rg)),packHalf2x16(vec2(v[0].ba))),uvec2(packHalf2x16(vec2(v[1].rg)),packHalf2x16(vec2(v[1].ba))));}"));
+        custom_defines.push_back(std::make_pair("buffer_ld8(buf,i)", "f16mat2x4(f16vec4(unpackFloat2x16(buf[i].r),unpackFloat2x16(buf[i].g)),f16vec4(unpackFloat2x16(buf[i].b),unpackFloat2x16(buf[i].a)))"));
+        custom_defines.push_back(std::make_pair("buffer_st8(buf,i,v)", "{buf[i]=uvec4(uvec2(packFloat2x16(v[0].rg),packFloat2x16(v[0].ba)),uvec2(packFloat2x16(v[1].rg),packFloat2x16(v[1].ba)));}"));
         custom_defines.push_back(std::make_pair("buffer_cp8(buf,i,sbuf,si)", "{buf[i]=sbuf[si];}"));
-        custom_defines.push_back(std::make_pair("buffer_cp8to1(buf,i4,ii4,sbuf,si)", "{uvec4 _v=sbuf[si]; vec2 _v0=unpackHalf2x16(_v.r);vec2 _v1=unpackHalf2x16(_v.g);vec2 _v2=unpackHalf2x16(_v.b);vec2 _v3=unpackHalf2x16(_v.a); buf[i4.r]=_v0.r;buf[i4.g]=_v0.g;buf[i4.b]=_v1.r;buf[i4.a]=_v1.g; buf[ii4.r]=_v2.r;buf[ii4.g]=_v2.g;buf[ii4.b]=_v3.r;buf[ii4.a]=_v3.g;}"));
+        custom_defines.push_back(std::make_pair("buffer_cp8to1(buf,i4,ii4,sbuf,si)", "{uvec4 _v=sbuf[si]; f16vec2 _v0=unpackFloat2x16(_v.r);f16vec2 _v1=unpackFloat2x16(_v.g);f16vec2 _v2=unpackFloat2x16(_v.b);f16vec2 _v3=unpackFloat2x16(_v.a); buf[i4.r]=_v0.r;buf[i4.g]=_v0.g;buf[i4.b]=_v1.r;buf[i4.a]=_v1.g; buf[ii4.r]=_v2.r;buf[ii4.g]=_v2.g;buf[ii4.b]=_v3.r;buf[ii4.a]=_v3.g;}"));
         custom_defines.push_back(std::make_pair("buffer_cp8to4(buf,i2,sbuf,si)", "{uvec4 _v=sbuf[si]; buf[i2.r]=_v.rg;buf[i2.g]=_v.ba;}"));
     }
     else if (opt.use_fp16_storage)