88       v128_t va4 = wasm_v128_load(a4);  in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()  local
99       vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c0);  in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
104       vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c0);  in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
110       va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);  in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
119       vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c1);  in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
124       vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c1);  in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
130       va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);  in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
139       vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c2);  in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
144       vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c2);  in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
150       va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);  in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
159       vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123c3);  in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
164       vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567c3);  in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
179       v128_t va4 = wasm_v128_load(a4);  in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()  local
191 …vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c0, vz…  in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
196 …vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c0, vz…  in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
202       va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);  in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
211 …vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c1, vz…  in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
216 …vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c1, vz…  in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
222       va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);  in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
231 …vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c2, vz…  in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
236 …vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c2, vz…  in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
242       va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);  in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
251 …vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, wasm_v128_andnot(va4, wasm_f32x4_eq(vb0123c3, vz…  in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()
256 …vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, wasm_v128_andnot(va4, wasm_f32x4_eq(vb4567c3, vz…  in xnn_f32_gemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma()