cooperative matrix shader failed on 8 elite gen5

[nihui]

samsung galaxy s26
oneui 8.5
android 16
system vulkan driver 512.842.19

revert changes introduced in Tencent/ncnn#6719

./vkpeak fp16-matrix
[0 Adreno (TM) 840]  queueC=1[1]  queueT=0[4]  rebar=1  r-score=21
[0 Adreno (TM) 840]  fp16-p/s/u/a=1/1/0/1  int8-p/s/u/a=1/1/0/1  bf16-p/s=1/0
[0 Adreno (TM) 840]  subgroup=64(64~64)  ops=1/1/1/1/1/1/1/1/1/1
[0 Adreno (TM) 840]  fp16-cm=64x64x16/64x32x16/64x16x16  int8-cm=64x64x32/64x32x32/64x16x32  bf16-cm=0  fp8-cm=0
device       = Adreno (TM) 840
driver       = Qualcomm Technologies Inc. Adreno Vulkan Driver / Driver Build: bacff47b1e, I498f01ff61, 1763652273
Date: 11/20/25
Compiler Version: E031.50.19.13
Driver Branch: 

Layout[] Node:   %144 = call reassoc nnan ninf nsz arcp contract <64 x half> @llvm.qgpu.coopmat.layout.annotate.v64f16(i64 8, i32 64, i32 16, i32 undef, <64 x half> %77) nounwind: 
    adjacent Node: 
         %.03 = phi <64 x half> [ %77, %.lr.ph ], [ %149, %146 ]
         %145 = call reassoc nnan ninf nsz arcp contract <64 x half> @llvm.qgpu.coopmat.layout.annotate.v64f16(i64 9, i32 16, i32 64, i32 undef, <64 x half> %143) nounwind


vkCreateComputePipelines failed -13
new_pipeline failed
Layout[] Node:   %144 = call reassoc nnan ninf nsz arcp contract <64 x half> @llvm.qgpu.coopmat.layout.annotate.v64f16(i64 8, i32 64, i32 16, i32 undef, <64 x half> %77) nounwind: 
    adjacent Node: 
         %.05 = phi <64 x half> [ %77, %.lr.ph ], [ %191, %146 ]
         %.014 = phi <64 x half> [ %143, %.lr.ph ], [ %194, %146 ]
         %145 = call reassoc nnan ninf nsz arcp contract <64 x half> @llvm.qgpu.coopmat.layout.annotate.v64f16(i64 9, i32 16, i32 64, i32 undef, <64 x half> %143) nounwind


vkCreateComputePipelines failed -13
new_pipeline failed

can be reproduced with a more simple shader

#version 450

#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
#extension GL_KHR_memory_scope_semantics: require
#extension GL_EXT_shader_explicit_arithmetic_types: require
#extension GL_KHR_cooperative_matrix: require

layout (constant_id = 0) const int loop = 1;
layout (constant_id = 1) const int M = 1;
layout (constant_id = 2) const int N = 1;
layout (constant_id = 3) const int K = 1;
layout (constant_id = 4) const int SCOPE = 3;

layout (binding = 0) writeonly buffer c_blob { float c_blob_data[]; };

void main()
{
    const uint gx = gl_GlobalInvocationID.x;
    const uint lx = gl_LocalInvocationID.x;

    coopmat<float16_t, SCOPE, M, K, gl_MatrixUseA> a = coopmat<float16_t, SCOPE, M, K, gl_MatrixUseA>(float(gx));
    coopmat<float16_t, SCOPE, K, N, gl_MatrixUseB> b = coopmat<float16_t, SCOPE, K, N, gl_MatrixUseB>(float(lx));

    coopmat<float16_t, SCOPE, M, N, gl_MatrixUseAccumulator> c = coopmat<float16_t, SCOPE, M, N, gl_MatrixUseAccumulator>(float(gx));

    for (int i = 0; i < loop; i++)
    {
        c = coopMatMulAdd(a, b, c);
    }

    coopMatStore(c, c_blob_data, 0, 0, gl_CooperativeMatrixLayoutRowMajor);
}