From e79712d96941c099d2f5e4b11544b2a20d97fbdf Mon Sep 17 00:00:00 2001 From: AbdelRauf Date: Fri, 30 Aug 2019 02:52:04 +0000 Subject: [PATCH] cgemv using vec_vsx_ld instead of letting gcc to decide --- kernel/power/cgemv_n.c | 97 +++++++++++++++++++++++------------------- kernel/power/cgemv_t.c | 97 +++++++++++++++++++++++++++--------------- 2 files changed, 115 insertions(+), 79 deletions(-) diff --git a/kernel/power/cgemv_n.c b/kernel/power/cgemv_n.c index 6a195d6d1..e85517ffa 100644 --- a/kernel/power/cgemv_n.c +++ b/kernel/power/cgemv_n.c @@ -62,23 +62,24 @@ static void cgemv_kernel_4x4(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOA register __vector float vx3_r = {x[6], -x[6],x[6], -x[6]}; register __vector float vx3_i = {x[7], x[7],x[7], x[7]}; #endif - register __vector float *vy = (__vector float *) y; + register __vector float *vptr_y = (__vector float *) y; register __vector float *vptr_a0 = (__vector float *) a0; register __vector float *vptr_a1 = (__vector float *) a1; register __vector float *vptr_a2 = (__vector float *) a2; register __vector float *vptr_a3 = (__vector float *) a3; BLASLONG i = 0; - for (;i< n / 2; i+=2) { - register __vector float vy_0 = vy[i]; - register __vector float vy_1 = vy[i + 1]; - register __vector float va0 = vptr_a0[i]; - register __vector float va1 = vptr_a1[i]; - register __vector float va2 = vptr_a2[i]; - register __vector float va3 = vptr_a3[i]; - register __vector float va0_1 = vptr_a0[i + 1]; - register __vector float va1_1 = vptr_a1[i + 1]; - register __vector float va2_1 = vptr_a2[i + 1]; - register __vector float va3_1 = vptr_a3[i + 1]; + BLASLONG i2=16; + for (;i< n * 8; i+=32,i2+=32) { + register __vector float vy_0 = vec_vsx_ld(i,vptr_y); + register __vector float vy_1 = vec_vsx_ld(i2,vptr_y); + register __vector float va0 = vec_vsx_ld(i,vptr_a0); + register __vector float va1 = vec_vsx_ld(i, vptr_a1); + register __vector float va2 = vec_vsx_ld(i ,vptr_a2); + register __vector float va3 = vec_vsx_ld(i ,vptr_a3); + register __vector float va0_1 = vec_vsx_ld(i2 ,vptr_a0); + register __vector float va1_1 = vec_vsx_ld(i2 ,vptr_a1); + register __vector float va2_1 = vec_vsx_ld(i2 ,vptr_a2); + register __vector float va3_1 = vec_vsx_ld(i2 ,vptr_a3); vy_0 += va0*vx0_r + va1*vx1_r + va2*vx2_r + va3*vx3_r; vy_1 += va0_1*vx0_r + va1_1*vx1_r + va2_1*vx2_r + va3_1*vx3_r; @@ -93,8 +94,8 @@ static void cgemv_kernel_4x4(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOA vy_0 += va0*vx0_i + va1*vx1_i + va2*vx2_i + va3*vx3_i; vy_1 += va0_1*vx0_i + va1_1*vx1_i + va2_1*vx2_i + va3_1*vx3_i; - vy[i] = vy_0; - vy[i + 1] = vy_1; + vec_vsx_st(vy_0 ,i, vptr_y); + vec_vsx_st(vy_1,i2,vptr_y); } } @@ -118,17 +119,19 @@ static void cgemv_kernel_4x2(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOA register __vector float vx1_r = {x[2], -x[2],x[2], -x[2]}; register __vector float vx1_i = {x[3], x[3],x[3], x[3]}; #endif - register __vector float *vy = (__vector float *) y; + register __vector float *vptr_y = (__vector float *) y; register __vector float *vptr_a0 = (__vector float *) a0; register __vector float *vptr_a1 = (__vector float *) a1; - BLASLONG i = 0; - for (;i< n / 2; i+=2) { - register __vector float vy_0 = vy[i]; - register __vector float vy_1 = vy[i + 1]; - register __vector float va0 = vptr_a0[i]; - register __vector float va1 = vptr_a1[i]; - register __vector float va0_1 = vptr_a0[i + 1]; - register __vector float va1_1 = vptr_a1[i + 1]; + BLASLONG i = 0; + BLASLONG i2 = 16; + for (;i< n * 8; i+=32, i2+=32) { + register __vector float vy_0 = vec_vsx_ld(i,vptr_y); + register __vector float vy_1 = vec_vsx_ld(i2,vptr_y); + register __vector float va0 = vec_vsx_ld(i,vptr_a0); + register __vector float va1 = vec_vsx_ld(i, vptr_a1); + register __vector float va0_1 = vec_vsx_ld(i2 ,vptr_a0); + register __vector float va1_1 = vec_vsx_ld(i2 ,vptr_a1); + register __vector float va0x = vec_perm(va0, va0,swap_mask); register __vector float va0x_1 = vec_perm(va0_1, va0_1,swap_mask); register __vector float va1x = vec_perm(va1, va1,swap_mask); @@ -136,8 +139,8 @@ static void cgemv_kernel_4x2(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOA vy_0 += va0*vx0_r + va1*vx1_r + va0x*vx0_i + va1x*vx1_i; vy_1 += va0_1*vx0_r + va1_1*vx1_r + va0x_1*vx0_i + va1x_1*vx1_i; - vy[i] = vy_0; - vy[i + 1] = vy_1; + vec_vsx_st(vy_0 ,i, vptr_y); + vec_vsx_st(vy_1,i2,vptr_y); } } @@ -154,21 +157,23 @@ static void cgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y) { register __vector float vx0_r = {x[0], -x[0],x[0], -x[0]}; register __vector float vx0_i = {x[1], x[1],x[1], x[1]}; #endif - register __vector float *vy = (__vector float *) y; + register __vector float *vptr_y = (__vector float *) y; register __vector float *vptr_a0 = (__vector float *) ap; BLASLONG i = 0; - for (;i< n / 2; i+=2) { - register __vector float vy_0 = vy[i]; - register __vector float vy_1 = vy[i + 1]; - register __vector float va0 = vptr_a0[i]; - register __vector float va0_1 = vptr_a0[i + 1]; + BLASLONG i2 = 16; + for (;i< n * 8; i+=32, i2+=32) { + register __vector float vy_0 = vec_vsx_ld(i,vptr_y); + register __vector float vy_1 = vec_vsx_ld(i2,vptr_y); + register __vector float va0 = vec_vsx_ld(i,vptr_a0); + register __vector float va0_1 = vec_vsx_ld(i2 ,vptr_a0); + register __vector float va0x = vec_perm(va0, va0,swap_mask); register __vector float va0x_1 = vec_perm(va0_1, va0_1,swap_mask); vy_0 += va0*vx0_r + va0x*vx0_i; vy_1 += va0_1*vx0_r + va0x_1*vx0_i; - vy[i] = vy_0; - vy[i + 1] = vy_1; + vec_vsx_st(vy_0 ,i, vptr_y); + vec_vsx_st(vy_1,i2,vptr_y); } } @@ -213,20 +218,24 @@ static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest, FLOAT register __vector float *vptr_src = (__vector float *) src; register __vector float *vptr_y = (__vector float *) dest; - for (i = 0; i < n/2; i += 2 ){ - register __vector float vy_0 = vptr_y[i]; - register __vector float vy_1 = vptr_y[i +1]; + BLASLONG i2 = 16; + for (;i< n * 8; i+=32, i2+=32) { + register __vector float vy_0 = vec_vsx_ld(i,vptr_y); + register __vector float vy_1 = vec_vsx_ld(i2,vptr_y); - register __vector float vsrc = vptr_src[i]; - register __vector float vsrc_1 = vptr_src[i + 1]; - register __vector float vsrcx = vec_perm(vsrc, vsrc, swap_mask); - register __vector float vsrcx_1 = vec_perm(vsrc_1, vsrc_1, swap_mask); - vy_0 += vsrc*valpha_r + vsrcx*valpha_i; - vy_1 += vsrc_1*valpha_r + vsrcx_1*valpha_i; - vptr_y[i] = vy_0; - vptr_y[i+1 ] = vy_1; + register __vector float vsrc = vec_vsx_ld(i,vptr_src); + register __vector float vsrc_1 = vec_vsx_ld(i2,vptr_src); + + register __vector float vsrcx = vec_perm(vsrc, vsrc, swap_mask); + register __vector float vsrcx_1 = vec_perm(vsrc_1, vsrc_1, swap_mask); + + vy_0 += vsrc*valpha_r + vsrcx*valpha_i; + vy_1 += vsrc_1*valpha_r + vsrcx_1*valpha_i; + + vec_vsx_st(vy_0 ,i, vptr_y); + vec_vsx_st(vy_1,i2,vptr_y); } diff --git a/kernel/power/cgemv_t.c b/kernel/power/cgemv_t.c index 68bbdd60a..57eb066b0 100644 --- a/kernel/power/cgemv_t.c +++ b/kernel/power/cgemv_t.c @@ -32,7 +32,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. static const unsigned char __attribute__((aligned(16))) swap_mask_arr[]={ 4,5,6,7,0,1,2,3, 12,13,14,15, 8,9,10,11}; static void cgemv_kernel_4x4(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT alpha_r, FLOAT alpha_i) { - BLASLONG i; + FLOAT *a0, *a1, *a2, *a3; a0 = ap; a1 = ap + lda; @@ -48,26 +48,39 @@ static void cgemv_kernel_4x4(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOA register __vector float vtemp2_r = {0.0, 0.0,0.0,0.0}; register __vector float vtemp3_p = {0.0, 0.0,0.0,0.0}; register __vector float vtemp3_r = {0.0, 0.0,0.0,0.0}; - __vector float* va0 = (__vector float*) a0; - __vector float* va1 = (__vector float*) a1; - __vector float* va2 = (__vector float*) a2; - __vector float* va3 = (__vector float*) a3; + __vector float* vptr_a0 = (__vector float*) a0; + __vector float* vptr_a1 = (__vector float*) a1; + __vector float* vptr_a2 = (__vector float*) a2; + __vector float* vptr_a3 = (__vector float*) a3; __vector float* v_x = (__vector float*) x; - for (i = 0; i < n / 2; i+=2) { - register __vector float vx_0 = v_x[i]; - register __vector float vx_1 = v_x[i+1]; + BLASLONG i = 0; + BLASLONG i2 = 16; + for (;i< n * 8; i+=32, i2+=32) { + register __vector float vx_0 = vec_vsx_ld( i,v_x) ; + register __vector float vx_1 = vec_vsx_ld(i2, v_x); + register __vector float vxr_0 = vec_perm(vx_0, vx_0, swap_mask); register __vector float vxr_1 = vec_perm(vx_1, vx_1, swap_mask); - vtemp0_p += vx_0*va0[i] + vx_1*va0[i+1] ; - vtemp0_r += vxr_0*va0[i] + vxr_1*va0[i+1]; - vtemp1_p += vx_0*va1[i] + vx_1*va1[i+1]; - vtemp1_r += vxr_0*va1[i] + vxr_1*va1[i+1]; - vtemp2_p += vx_0*va2[i] + vx_1*va2[i+1]; - vtemp2_r += vxr_0*va2[i] + vxr_1*va2[i+1]; - vtemp3_p += vx_0*va3[i] + vx_1*va3[i+1]; - vtemp3_r += vxr_0*va3[i] + vxr_1*va3[i+1]; + register __vector float va0 = vec_vsx_ld(i,vptr_a0); + register __vector float va1 = vec_vsx_ld(i, vptr_a1); + register __vector float va2 = vec_vsx_ld(i ,vptr_a2); + register __vector float va3 = vec_vsx_ld(i ,vptr_a3); + register __vector float va0_1 = vec_vsx_ld(i2 ,vptr_a0); + register __vector float va1_1 = vec_vsx_ld(i2 ,vptr_a1); + register __vector float va2_1 = vec_vsx_ld(i2 ,vptr_a2); + register __vector float va3_1 = vec_vsx_ld(i2 ,vptr_a3); + + + vtemp0_p += vx_0*va0 + vx_1*va0_1 ; + vtemp0_r += vxr_0*va0 + vxr_1*va0_1; + vtemp1_p += vx_0*va1 + vx_1*va1_1; + vtemp1_r += vxr_0*va1 + vxr_1*va1_1; + vtemp2_p += vx_0*va2 + vx_1*va2_1; + vtemp2_r += vxr_0*va2 + vxr_1*va2_1; + vtemp3_p += vx_0*va3 + vx_1*va3_1; + vtemp3_r += vxr_0*va3 + vxr_1*va3_1; } @@ -128,7 +141,7 @@ static void cgemv_kernel_4x4(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOA static void cgemv_kernel_4x2(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT alpha_r, FLOAT alpha_i) { - BLASLONG i; + FLOAT *a0, *a1; a0 = ap; a1 = ap + lda; @@ -138,23 +151,33 @@ static void cgemv_kernel_4x2(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOA register __vector float vtemp0_r = {0.0, 0.0,0.0,0.0}; register __vector float vtemp1_p = {0.0, 0.0,0.0,0.0}; register __vector float vtemp1_r = {0.0, 0.0,0.0,0.0}; - __vector float* va0 = (__vector float*) a0; - __vector float* va1 = (__vector float*) a1; + + + __vector float* vptr_a0 = (__vector float*) a0; + __vector float* vptr_a1 = (__vector float*) a1; __vector float* v_x = (__vector float*) x; - for (i = 0; i < n / 2; i+=2) { - register __vector float vx_0 = v_x[i]; - register __vector float vx_1 = v_x[i+1]; + BLASLONG i = 0; + BLASLONG i2 = 16; + for (;i< n * 8; i+=32, i2+=32) { + register __vector float vx_0 = vec_vsx_ld( i,v_x) ; + register __vector float vx_1 = vec_vsx_ld(i2, v_x); + register __vector float vxr_0 = vec_perm(vx_0, vx_0, swap_mask); register __vector float vxr_1 = vec_perm(vx_1, vx_1, swap_mask); - vtemp0_p += vx_0*va0[i] + vx_1*va0[i+1] ; - vtemp0_r += vxr_0*va0[i] + vxr_1*va0[i+1]; - vtemp1_p += vx_0*va1[i] + vx_1*va1[i+1]; - vtemp1_r += vxr_0*va1[i] + vxr_1*va1[i+1]; + register __vector float va0 = vec_vsx_ld(i,vptr_a0); + register __vector float va1 = vec_vsx_ld(i, vptr_a1); + register __vector float va0_1 = vec_vsx_ld(i2 ,vptr_a0); + register __vector float va1_1 = vec_vsx_ld(i2 ,vptr_a1); + + + vtemp0_p += vx_0*va0 + vx_1*va0_1 ; + vtemp0_r += vxr_0*va0 + vxr_1*va0_1; + vtemp1_p += vx_0*va1 + vx_1*va1_1; + vtemp1_r += vxr_0*va1 + vxr_1*va1_1; } - #if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) ) register FLOAT temp_r0 = vtemp0_p[0] - vtemp0_p[1] + vtemp0_p[2] - vtemp0_p[3]; @@ -193,23 +216,27 @@ static void cgemv_kernel_4x2(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOA static void cgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT alpha_r, FLOAT alpha_i) { - BLASLONG i; + __vector unsigned char swap_mask = *((__vector unsigned char*)swap_mask_arr); //p for positive(real*real,image*image,real*real,image*image) r for image (real*image,image*real,real*image,image*real) register __vector float vtemp0_p = {0.0, 0.0,0.0,0.0}; register __vector float vtemp0_r = {0.0, 0.0,0.0,0.0}; - __vector float* va0 = (__vector float*) ap; + __vector float* vptr_a0 = (__vector float*) ap; __vector float* v_x = (__vector float*) x; - - for (i = 0; i < n / 2; i+=2) { - register __vector float vx_0 = v_x[i]; - register __vector float vx_1 = v_x[i+1]; + BLASLONG i = 0; + BLASLONG i2 = 16; + for (;i< n * 8; i+=32, i2+=32) { + register __vector float vx_0 = vec_vsx_ld( i,v_x) ; + register __vector float vx_1 = vec_vsx_ld(i2, v_x); + register __vector float vxr_0 = vec_perm(vx_0, vx_0, swap_mask); register __vector float vxr_1 = vec_perm(vx_1, vx_1, swap_mask); - vtemp0_p += vx_0*va0[i] + vx_1*va0[i+1] ; - vtemp0_r += vxr_0*va0[i] + vxr_1*va0[i+1]; + register __vector float va0 = vec_vsx_ld(i,vptr_a0); + register __vector float va0_1 = vec_vsx_ld(i2 ,vptr_a0); + vtemp0_p += vx_0*va0 + vx_1*va0_1 ; + vtemp0_r += vxr_0*va0 + vxr_1*va0_1; } #if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )