#define REALNAME bli_cgemm_armv7a_ker_2x2 #define STACKSIZE 256 #define K r0 #define PTR_ALPHA r1 #define OLD_A r2 #define OLD_B r3 #define PTR_BETA [fp, #0 ] #define OLD_C [fp, #4 ] #define OLD_RSC [fp, #8 ] #define OLD_CSC [fp, #12 ] #define AUX [fp, #16 ] /****************************************************** * [fp, #-128] - [fp, #-64] is reserved * for store and restore of floating point * register *******************************************************/ #define L r2 #define AO r5 #define BO r6 #define CO1 r7 #define CO2 r8 #define A_PRE 96 #define B_PRE 96 #define C_PRE 0 /************************************************************************************** * Macro definitions **************************************************************************************/ #define FMAC_BR fnmacs #define FMAC_BI fmacs #define NN 1 #if defined(NN) || defined(NT) || defined(TN) || defined(TT) #define FADD_R fsubs #define FADD_I fadds #define FMAC_R1 fnmacs #define FMAC_R2 fnmacs #define FMAC_I1 fmacs #define FMAC_I2 fnmacs #elif defined(CN) || defined(CT) #define FADD_R fadds #define FADD_I fsubs #define FMAC_R1 fmacs #define FMAC_R2 fmacs #define FMAC_I1 fnmacs #define FMAC_I2 fmacs #elif defined(NC) || defined(TC) #define FADD_R fadds #define FADD_I fsubs #define FMAC_R1 fmacs #define FMAC_R2 fnmacs #define FMAC_I1 fmacs #define FMAC_I2 fmacs #else #define FADD_R fsubs #define FADD_I fadds #define FMAC_R1 fnmacs #define FMAC_R2 fmacs #define FMAC_I1 fnmacs #define FMAC_I2 fnmacs #endif .macro INIT2x2 vsub.f32 s16 , s16 , s16 vmov.f32 s17, s16 vmov.f32 s18, s16 vmov.f32 s19, s16 vmov.f32 s20, s16 vmov.f32 s21, s16 vmov.f32 s22, s16 vmov.f32 s23, s16 vmov.f32 s24, s16 vmov.f32 s25, s16 vmov.f32 s26, s16 vmov.f32 s27, s16 vmov.f32 s28, s16 vmov.f32 s29, s16 vmov.f32 s30, s16 vmov.f32 s31, s16 .endm .macro KERNEL2x2_I pld [ AO , #A_PRE ] pld [ BO , #B_PRE ] flds s0 , [ AO ] flds s1 , [ AO, #4 ] flds s8 , [ BO ] flds s9 , [ BO, #4 ] fmuls s16 , s0, s8 flds s2 , [ AO, #8 ] fmuls s24 , s1, s9 flds s3 , [ AO, #12 ] fmuls s17 , s0, s9 flds s10, [ BO, #8 ] fmuls s25 , s1, s8 flds s11, [ BO, #12 ] fmuls s18 , s2, s8 add BO , BO, #16 fmuls s26 , s3, s9 add AO , AO, #16 fmuls s19 , s2, s9 pld [ BO , #B_PRE ] fmuls s27 , s3, s8 pld [ AO , #A_PRE ] fmuls s20 , s0, s10 flds s4 , [ AO, #0 ] fmuls s28 , s1, s11 flds s5 , [ AO, #4 ] fmuls s21 , s0, s11 flds s12, [ BO ] fmuls s29 , s1, s10 flds s13, [ BO, #4 ] fmuls s22 , s2, s10 flds s6 , [ AO, #8 ] fmuls s30 , s3, s11 flds s7 , [ AO, #12 ] fmuls s23 , s2, s11 flds s14, [ BO, #8 ] fmuls s31 , s3, s10 flds s15, [ BO, #12 ] add BO , BO, #16 add AO , AO, #16 .endm .macro KERNEL2x2_M1 pld [ AO , #A_PRE ] fmacs s16 , s0, s8 pld [ BO , #B_PRE ] fmacs s24 , s1, s9 flds s4 , [ AO, #0 ] fmacs s17 , s0, s9 flds s5 , [ AO, #4 ] fmacs s25 , s1, s8 flds s12, [ BO ] fmacs s18 , s2, s8 flds s13, [ BO, #4 ] fmacs s26 , s3, s9 flds s6 , [ AO, #8 ] fmacs s19 , s2, s9 flds s7 , [ AO, #12 ] fmacs s27 , s3, s8 fmacs s20 , s0, s10 flds s14, [ BO, #8 ] fmacs s28 , s1, s11 fmacs s21 , s0, s11 flds s15, [ BO, #12 ] fmacs s29 , s1, s10 fmacs s22 , s2, s10 add BO , BO, #16 fmacs s30 , s3, s11 fmacs s23 , s2, s11 add AO , AO, #16 fmacs s31 , s3, s10 .endm .macro KERNEL2x2_M2 fmacs s16 , s4, s12 fmacs s24 , s5, s13 flds s0 , [ AO, #0 ] fmacs s17 , s4, s13 flds s1 , [ AO, #4 ] fmacs s25 , s5, s12 fmacs s18 , s6, s12 flds s8 , [ BO ] fmacs s26 , s7, s13 flds s9 , [ BO, #4 ] fmacs s19 , s6, s13 fmacs s27 , s7, s12 flds s2 , [ AO, #8 ] fmacs s20 , s4, s14 flds s3 , [ AO, #12 ] fmacs s28 , s5, s15 fmacs s21 , s4, s15 flds s10, [ BO, #8 ] fmacs s29 , s5, s14 flds s11, [ BO, #12 ] fmacs s22 , s6, s14 fmacs s30 , s7, s15 add BO , BO, #16 fmacs s23 , s6, s15 add AO , AO, #16 fmacs s31 , s7, s14 .endm .macro KERNEL2x2_E fmacs s16 , s4, s12 fmacs s24 , s5, s13 fmacs s17 , s4, s13 fmacs s25 , s5, s12 fmacs s18 , s6, s12 fmacs s26 , s7, s13 fmacs s19 , s6, s13 fmacs s27 , s7, s12 fmacs s20 , s4, s14 fmacs s28 , s5, s15 fmacs s21 , s4, s15 fmacs s29 , s5, s14 fmacs s22 , s6, s14 fmacs s30 , s7, s15 fmacs s23 , s6, s15 fmacs s31 , s7, s14 .endm .macro KERNEL2x2_SUB flds s0 , [ AO ] flds s1 , [ AO, #4 ] flds s8 , [ BO ] flds s9 , [ BO, #4 ] fmacs s16 , s0, s8 flds s2 , [ AO, #8 ] fmacs s24 , s1, s9 flds s3 , [ AO, #12 ] fmacs s17 , s0, s9 flds s10, [ BO, #8 ] fmacs s25 , s1, s8 flds s11, [ BO, #12 ] fmacs s18 , s2, s8 fmacs s26 , s3, s9 fmacs s19 , s2, s9 fmacs s27 , s3, s8 fmacs s20 , s0, s10 fmacs s28 , s1, s11 fmacs s21 , s0, s11 fmacs s29 , s1, s10 fmacs s22 , s2, s10 add BO , BO, #16 fmacs s30 , s3, s11 fmacs s23 , s2, s11 add AO , AO, #16 fmacs s31 , s3, s10 .endm .macro SAVE2x2 ldr r3, OLD_RSC // Row stride size lsl r3, r3, #3 // multiply with size of complex float flds s0, [ PTR_ALPHA ] // load real part of alpha flds s1, [ PTR_ALPHA, #4 ] // load imag part of alpha ldr r4, PTR_BETA flds s2, [ r4 ] // load real part of beta flds s3, [ r4, #4 ] // load imag part of beta // Add/Sub the real and the imag parts FADD_R s16, s24 , s16 FADD_I s17, s25 , s17 FADD_R s18, s26 , s18 FADD_I s19, s27 , s19 FADD_R s20, s28 , s20 FADD_I s21, s29 , s21 FADD_R s22, s30 , s22 FADD_I s23, s31 , s23 mov r4, CO1 // save pointer fldmias CO1, { s4 - s5 } // read real and imag part from C add CO1, CO1, r3 mov r2, CO2 // save pointer fldmias CO2, { s8 - s9 } // read real and imag part from C add CO2, CO2, r3 fmuls s24, s4, s2 // multiply Beta-real with C-real fmuls s25, s5, s2 // multiply Beta-real with C-imag fmuls s28, s8, s2 // multiply Beta-real with C-real fmuls s29, s9, s2 // multiply Beta-real with C-imag FMAC_BR s24, s3, s5 // multiply beta-imag with C-imag and add FMAC_BI s25, s3, s4 // multiply beta-imag with C-real and add FMAC_BR s28, s3, s9 // multiply beta-imag with C-imag and add FMAC_BI s29, s3, s8 // multiply beta-imag with C-real and add FMAC_R1 s24 , s0 , s16 FMAC_I1 s25 , s0 , s17 FMAC_R2 s24 , s1 , s17 FMAC_I2 s25 , s1 , s16 FMAC_R1 s28 , s0 , s20 FMAC_I1 s29 , s0 , s21 FMAC_R2 s28 , s1 , s21 FMAC_I2 s29 , s1 , s20 fldmias CO1, { s4 - s5 } // read real and imag part from C fldmias CO2, { s8 - s9 } // read real and imag part from C fmuls s26, s4, s2 // multiply Beta-real with C-real fmuls s27, s5, s2 // multiply Beta-real with C-imag fmuls s30, s8, s2 // multiply Beta-real with C-real fmuls s31, s9, s2 // multiply Beta-real with C-imag FMAC_BR s26, s3, s5 // multiply beta-imag with C-imag and add FMAC_BI s27, s3, s4 // multiply beta-imag with C-real and add FMAC_BR s30, s3, s9 // multiply beta-imag with C-imag and add FMAC_BI s31, s3, s8 // multiply beta-imag with C-real and add FMAC_R1 s26 , s0 , s18 FMAC_I1 s27 , s0 , s19 FMAC_R2 s26 , s1 , s19 FMAC_I2 s27 , s1 , s18 FMAC_R1 s30, s0 , s22 FMAC_I1 s31, s0 , s23 FMAC_R2 s30, s1 , s23 FMAC_I2 s31, s1 , s22 mov CO1, r4 // restore pointer mov CO2, r2 // restore pointer fstmias CO1, { s24 - s25 } fstmias CO2, { s28 - s29 } add CO1, CO1, r3 add CO2, CO2, r3 fstmias CO1, { s26 - s27 } fstmias CO2, { s30 - s31 } .endm /************************************************************************************** * End of macro definitions **************************************************************************************/ .arm .global REALNAME .func REALNAME REALNAME: push {r4 - r9, fp} // save register add fp, sp, #28 // add number of saved register multiplied by size of int sub sp, sp, #STACKSIZE // reserve stack mov AO, OLD_A // pointer matrix A mov BO, OLD_B // pointer matrix B sub r3, fp, #128 vstm r3, { s8 - s31} // store floating point registers ldr r2, OLD_C // pointer matrix C ldr r3, OLD_CSC // Col stride size of C lsl r3, r3, #3 // multiply with size of complex float mov CO1, r2 // first line of C add CO2, CO1, r3 // second line of C pld [ CO1, #C_PRE ] // prefetch the lines of C pld [ CO2, #C_PRE ] // prefetch the lines of C cgemm_kernel_L2_M2_20: asrs L , K, #3 // L = K / 8 cmp L , #2 blt cgemm_kernel_L2_M2_32 KERNEL2x2_I KERNEL2x2_M2 KERNEL2x2_M1 KERNEL2x2_M2 KERNEL2x2_M1 KERNEL2x2_M2 KERNEL2x2_M1 KERNEL2x2_M2 subs L, L, #2 ble cgemm_kernel_L2_M2_22a .align 5 cgemm_kernel_L2_M2_22: KERNEL2x2_M1 KERNEL2x2_M2 KERNEL2x2_M1 KERNEL2x2_M2 KERNEL2x2_M1 KERNEL2x2_M2 KERNEL2x2_M1 KERNEL2x2_M2 subs L, L, #1 bgt cgemm_kernel_L2_M2_22 cgemm_kernel_L2_M2_22a: KERNEL2x2_M1 KERNEL2x2_M2 KERNEL2x2_M1 KERNEL2x2_M2 KERNEL2x2_M1 KERNEL2x2_M2 KERNEL2x2_M1 KERNEL2x2_E b cgemm_kernel_L2_M2_44 cgemm_kernel_L2_M2_32: tst L, #1 ble cgemm_kernel_L2_M2_40 KERNEL2x2_I KERNEL2x2_M2 KERNEL2x2_M1 KERNEL2x2_M2 KERNEL2x2_M1 KERNEL2x2_M2 KERNEL2x2_M1 KERNEL2x2_E b cgemm_kernel_L2_M2_44 cgemm_kernel_L2_M2_40: INIT2x2 cgemm_kernel_L2_M2_44: ands L , K, #7 // L = K % 8 ble cgemm_kernel_L2_M2_100 cgemm_kernel_L2_M2_46: KERNEL2x2_SUB subs L, L, #1 bne cgemm_kernel_L2_M2_46 cgemm_kernel_L2_M2_100: SAVE2x2 cgemm_kernel_L999: sub r3, fp, #128 vldm r3, { s8 - s31} // restore floating point registers sub sp, fp, #28 pop {r4 - r9, fp} bx lr