MOVQ	CX, 0(SP)
	MOVQ	DX, 8(SP)

	LEAQ	ret+0(FP), DX
	MOVQ	-8(DX), CX
	MOVQ	CX, m_vdsoPC(BX)
	MOVQ	DX, m_vdsoSP(BX)

	CMPQ	R14, m_curg(BX)	// Only switch if on curg.
	JNE	noswitch

	MOVQ	m_g0(BX), DX
	MOVQ	(g_sched+gobuf_sp)(DX), SP	// Set SP to g0 stack

noswitch:
	SUBQ	$16, SP		// Space for results
	ANDQ	$~15, SP	// Align for C code

	MOVL	$1, DI // CLOCK_MONOTONIC
	LEAQ	0(SP), SI

TEXT runtime·mstart_stub(SB),NOSPLIT,$144
	// R0 points to the m.
	// We are already on m's g0 stack.

	// Save callee-save registers.
	SAVE_R19_TO_R28(8)
	SAVE_F8_TO_F15(88)

	MOVD    m_g0(R0), g
	BL	runtime·save_g(SB)

	BL	runtime·mstart(SB)

	// Restore callee-save registers.
	RESTORE_R19_TO_R28(8)
	RESTORE_F8_TO_F15(88)

	// Go is all done with this OS thread.

	// We are already on m's g0 stack.

	// Transition from C ABI to Go ABI.
	PUSH_REGS_HOST_TO_ABI0()

	// Load g and save to TLS entry.
	// See cmd/link/internal/ld/sym.go:computeTLSOffset.
	MOVQ	m_g0(DI), DX // g
	MOVQ	DX, -8(FS)

	CALL	runtime·mstart(SB)

	POP_REGS_HOST_TO_ABI0()

	// Go is all done with this OS thread.
	// Tell pthread everything is ok (we never join with this thread, so

	MOVF	F19, (16*8)(X2)
	MOVF	F20, (17*8)(X2)
	MOVF	F21, (18*8)(X2)
	MOVF	F22, (19*8)(X2)
	MOVF	F23, (20*8)(X2)
	MOVF	F24, (21*8)(X2)
	MOVF	F25, (22*8)(X2)
	MOVF	F26, (23*8)(X2)
	MOVF	F27, (24*8)(X2)

	MOV	m_g0(X10), g
	CALL	runtime·save_g(SB)

	CALL	runtime·mstart(SB)

	// Restore callee-save registers.
	MOV	(1*8)(X2), X8
	MOV	(2*8)(X2), X9
	MOV	(3*8)(X2), X18
	MOV	(4*8)(X2), X19
	MOV	(5*8)(X2), X20

TEXT runtime·mstart_stub(SB),NOSPLIT,$160
	// R0 points to the m.
	// We are already on m's g0 stack.

	// Save callee-save registers.
	SAVE_R19_TO_R28(8)
	SAVE_F8_TO_F15(88)

	MOVD	m_g0(R0), g
	BL	·save_g(SB)

	BL	runtime·mstart(SB)

	// Restore callee-save registers.
	RESTORE_R19_TO_R28(8)
	RESTORE_F8_TO_F15(88)

	// Go is all done with this OS thread.

TEXT runtime·mstart_stub(SB),NOSPLIT|NOFRAME,$0
	// DI points to the m.
	// We are already on m's g0 stack.

	// Transition from C ABI to Go ABI.
	PUSH_REGS_HOST_TO_ABI0()

	MOVQ	m_g0(DI), DX // g

	// Initialize TLS entry.
	// See cmd/link/internal/ld/sym.go:computeTLSOffset.
	MOVQ	DX, 0x30(GS)

	CALL	runtime·mstart(SB)

	POP_REGS_HOST_TO_ABI0()

// Note: called with the C calling convention.
TEXT runtime·mstart_stub(SB),NOSPLIT,$0
	// R0 points to the m.
	// We are already on m's g0 stack.

	// Save callee-save registers.
	MOVM.DB.W [R4-R11], (R13)

	MOVW	m_g0(R0), g
	BL	runtime·save_g(SB)

	BL	runtime·mstart(SB)

	// Restore callee-save registers.
	MOVM.IA.W (R13), [R4-R11]

	// Go is all done with this OS thread.

	// We are already on m's g0 stack.

	// Save callee-save registers.
	MOVL	BX, bx-4(SP)
	MOVL	BP, bp-8(SP)
	MOVL	SI, si-12(SP)
	MOVL	DI, di-16(SP)

	MOVL	32(SP), AX	// m
	MOVL	m_g0(AX), DX
	get_tls(CX)
	MOVL	DX, g(CX)

	CALL	runtime·mstart(SB)

	// Restore callee-save registers.
	MOVL	di-16(SP), DI
	MOVL	si-12(SP), SI
	MOVL	bp-8(SP),  BP
	MOVL	bx-4(SP),  BX

	EXPAND(T_0, T_1, M_0, M_1, M_2, M_3, M_4)

	// add 2¹²⁸ to each message block value
	VLEIB $4, $1, M_4
	VLEIB $12, $1, M_4

multiply:
	// accumulate the incoming message
	VAG H_0, M_0, M_0
	VAG H_3, M_3, M_3
	VAG H_1, M_1, M_1
	VAG H_4, M_4, M_4
	VAG H_2, M_2, M_2

	// multiply the accumulator by the key coefficient

        batch_matmul.getAsymmetricQuantizeInputsAttr());
    return success();
  }

 private:
  // Checks that tensor `a` has shape of [M, N] and `b` has
  // [M_0, M_1, ..., 1, N], where `M = M_0 * M_1 * ...`.
  bool InsertOneInSecondInnermostDim(TensorType a, TensorType b) const {
    if (!a.hasStaticShape() || !b.hasStaticShape()) return false;
    if (a.getRank() != 2 || b.getRank() < 2) return false;