#version 450 core
#extension GL_KHR_memory_scope_semantics : enable
#extension GL_NV_cooperative_vector : enable
#extension GL_EXT_shader_explicit_arithmetic_types : enable
#extension GL_EXT_buffer_reference : enable

layout (local_size_x = 64, local_size_y = 1, local_size_z = 1) in;

layout(set = 0, binding = 0) buffer MatrixBuf {
    float16_t matrixData[];
} matrixBuf;

void main()
{
    coopvecNV<float16_t, 8> v0;
    coopvecNV<float32_t, 5> v1;
    vec4 v2;

    coopVecOuterProductAccumulateNV(v0, v1, matrixBuf.matrixData, 100, 128, gl_CooperativeVectorMatrixLayoutInferencingOptimalNV, gl_ComponentTypeFloat16NV);
    coopVecOuterProductAccumulateNV(v0, v2, matrixBuf.matrixData, 100, 128, gl_CooperativeVectorMatrixLayoutInferencingOptimalNV, gl_ComponentTypeFloat16NV);
    coopVecReduceSumAccumulateNV(v2, matrixBuf.matrixData, 100);
}