AVX2 STRSM kernel

kernel/x86_64/KERNEL.HASWELL

@@ -77,10 +77,10 @@ ZGEMMITCOPYOBJ =  zgemm_itcopy$(TSUFFIX).$(SUFFIX)
 ZGEMMONCOPYOBJ =  zgemm_oncopy$(TSUFFIX).$(SUFFIX)
 ZGEMMOTCOPYOBJ =  zgemm_otcopy$(TSUFFIX).$(SUFFIX)
 
-STRSMKERNEL_LN	=  ../generic/trsm_kernel_LN.c
-STRSMKERNEL_LT	=  ../generic/trsm_kernel_LT.c
-STRSMKERNEL_RN	=  ../generic/trsm_kernel_RN.c
-STRSMKERNEL_RT	=  ../generic/trsm_kernel_RT.c
+STRSMKERNEL_LN	=  strsm_kernel_8x4_haswell_LN.c
+STRSMKERNEL_LT	=  strsm_kernel_8x4_haswell_LT.c
+STRSMKERNEL_RN	=  strsm_kernel_8x4_haswell_RN.c
+STRSMKERNEL_RT	=  strsm_kernel_8x4_haswell_RT.c
 
 DTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c
 DTRSMKERNEL_LT	= ../generic/trsm_kernel_LT.c

kernel/x86_64/strsm_kernel_8x4_haswell_LN.c

@@ -0,0 +1,240 @@
+#include "common.h"
+#include <stdint.h>
+#include "strsm_kernel_8x4_haswell_L_common.h"
+
+#define SOLVE_LN_m1n4 \
+  "subq $4,%2; movq %2,%3;" GEMM_SUM_REORDER_1x4(4)\
+  SOLVE_m1n4(-4,4) SAVE_b_m1n4(-16,4)\
+  "movq %2,%3;" save_c_m1n4(4)
+
+#define SOLVE_LN_m1n8 \
+  "subq $4,%2; movq %2,%3;" GEMM_SUM_REORDER_1x4(4) GEMM_SUM_REORDER_1x4(5)\
+  SOLVE_m1n8(-4,4,5) SAVE_b_m1n8(-16,4,5)\
+  "movq %2,%3;" save_c_m1n4(4) save_c_m1n4(5)
+
+#define SOLVE_LN_m1n12 \
+  "subq $4,%2; movq %2,%3;" GEMM_SUM_REORDER_1x4(4) GEMM_SUM_REORDER_1x4(5) GEMM_SUM_REORDER_1x4(6)\
+  SOLVE_m1n12(-4,4,5,6) SAVE_b_m1n12(-16,4,5,6)\
+  "movq %2,%3;" save_c_m1n4(4) save_c_m1n4(5) save_c_m1n4(6)
+
+#define SOLVE_LN_m2n4 \
+  "subq $8,%2; movq %2,%3;" GEMM_SUM_REORDER_2x4(4,5,4)\
+  SOLVE_loup_m2n4(-8,4)\
+  SOLVE_up_m2n4(-16,4) SAVE_b_m2n4(-32,4)\
+  "movq %2,%3;" save_c_m2n4(4)
+
+#define SOLVE_LN_m2n8 \
+  "subq $8,%2; movq %2,%3;" GEMM_SUM_REORDER_2x4(4,5,4) GEMM_SUM_REORDER_2x4(6,7,5)\
+  SOLVE_loup_m2n8(-8,4,5)\
+  SOLVE_up_m2n8(-16,4,5) SAVE_b_m2n8(-32,4,5)\
+  "movq %2,%3;" save_c_m2n4(4) save_c_m2n4(5)
+
+#define SOLVE_LN_m2n12 \
+  "subq $8,%2; movq %2,%3;" GEMM_SUM_REORDER_2x4(4,5,4) GEMM_SUM_REORDER_2x4(6,7,5) GEMM_SUM_REORDER_2x4(8,9,6)\
+  SOLVE_loup_m2n12(-8,4,5,6)\
+  SOLVE_up_m2n12(-16,4,5,6) SAVE_b_m2n12(-32,4,5,6)\
+  "movq %2,%3;" save_c_m2n4(4) save_c_m2n4(5) save_c_m2n4(6)
+
+#define SOLVE_LN_m4n4 \
+  "subq $16,%2; movq %2,%3;" GEMM_SUM_REORDER_4x4(4,5,6,7,4,5)\
+\
+  SOLVE_loup_m2n4(-8,5) SUBTRACT_m2n4(-16,4)\
+  SOLVE_up_m2n4(-24,5) SUBTRACT_m2n4(-32,4) SAVE_b_m2n4(-32,5)\
+\
+  SOLVE_loup_m2n4(-48,4)\
+  SOLVE_up_m2n4(-64,4) SAVE_b_m2n4(-64,4)\
+\
+  "movq %2,%3;" save_c_m4n4(4,5)
+
+#define SOLVE_LN_m4n8 \
+  "subq $16,%2; movq %2,%3;" GEMM_SUM_REORDER_4x4(4,5,6,7,4,5) GEMM_SUM_REORDER_4x4(8,9,10,11,6,7)\
+\
+  SOLVE_loup_m2n8(-8,5,7) SUBTRACT_m2n8(-16,4,6)\
+  SOLVE_up_m2n8(-24,5,7) SUBTRACT_m2n8(-32,4,6) SAVE_b_m2n8(-32,5,7)\
+\
+  SOLVE_loup_m2n8(-48,4,6)\
+  SOLVE_up_m2n8(-64,4,6) SAVE_b_m2n8(-64,4,6)\
+\
+  "movq %2,%3;" save_c_m4n4(4,5) save_c_m4n4(6,7)
+
+#define SOLVE_LN_m4n12 \
+  "subq $16,%2; movq %2,%3;" GEMM_SUM_REORDER_4x4(4,5,6,7,4,5) GEMM_SUM_REORDER_4x4(8,9,10,11,6,7) GEMM_SUM_REORDER_4x4(12,13,14,15,8,9)\
+\
+  SOLVE_loup_m2n12(-8,5,7,9) SUBTRACT_m2n12(-16,4,6,8)\
+  SOLVE_up_m2n12(-24,5,7,9) SUBTRACT_m2n12(-32,4,6,8) SAVE_b_m2n12(-32,5,7,9)\
+\
+  SOLVE_loup_m2n12(-48,4,6,8)\
+  SOLVE_up_m2n12(-64,4,6,8) SAVE_b_m2n12(-64,4,6,8)\
+\
+  "movq %2,%3;" save_c_m4n4(4,5) save_c_m4n4(6,7) save_c_m4n4(8,9)
+
+#define SOLVE_LN_m8n4 \
+  "subq $32,%2; movq %2,%3;" GEMM_SUM_REORDER_8x4(4,5,6,7,-32)\
+\
+  SOLVE_loup_m2n4(-8,7) SUBTRACT_m2n4(-16,6) SUBTRACT_m2n4(-24,5) SUBTRACT_m2n4(-32,4)\
+  SOLVE_up_m2n4(-40,7) SUBTRACT_m2n4(-48,6) SUBTRACT_m2n4(-56,5) SUBTRACT_m2n4(-64,4) SAVE_b_m2n4(-32,7)\
+\
+  SOLVE_loup_m2n4(-80,6) SUBTRACT_m2n4(-88,5) SUBTRACT_m2n4(-96,4)\
+  SOLVE_up_m2n4(-112,6) SUBTRACT_m2n4(-120,5) SUBTRACT_m2n4(-128,4) SAVE_b_m2n4(-64,6)\
+\
+  SOLVE_loup_m2n4(-152,5) SUBTRACT_m2n4(-160,4)\
+  SOLVE_up_m2n4(-184,5) SUBTRACT_m2n4(-192,4) SAVE_b_m2n4(-96,5)\
+\
+  SOLVE_loup_m2n4(-224,4)\
+  SOLVE_up_m2n4(-256,4) SAVE_b_m2n4(-128,4)\
+\
+  "movq %2,%3;" save_c_m8n4(4,5,6,7)
+
+#define SOLVE_LN_m8n8 \
+  "subq $32,%2; movq %2,%3;" GEMM_SUM_REORDER_8x4(4,5,6,7,-32) GEMM_SUM_REORDER_8x4(8,9,10,11,-32)\
+\
+  SOLVE_loup_m2n8(-8,7,11) SUBTRACT_m2n8(-16,6,10) SUBTRACT_m2n8(-24,5,9) SUBTRACT_m2n8(-32,4,8)\
+  SOLVE_up_m2n8(-40,7,11) SUBTRACT_m2n8(-48,6,10) SUBTRACT_m2n8(-56,5,9) SUBTRACT_m2n8(-64,4,8) SAVE_b_m2n8(-32,7,11)\
+\
+  SOLVE_loup_m2n8(-80,6,10) SUBTRACT_m2n8(-88,5,9) SUBTRACT_m2n8(-96,4,8)\
+  SOLVE_up_m2n8(-112,6,10) SUBTRACT_m2n8(-120,5,9) SUBTRACT_m2n8(-128,4,8) SAVE_b_m2n8(-64,6,10)\
+\
+  SOLVE_loup_m2n8(-152,5,9) SUBTRACT_m2n8(-160,4,8)\
+  SOLVE_up_m2n8(-184,5,9) SUBTRACT_m2n8(-192,4,8) SAVE_b_m2n8(-96,5,9)\
+\
+  SOLVE_loup_m2n8(-224,4,8)\
+  SOLVE_up_m2n8(-256,4,8) SAVE_b_m2n8(-128,4,8)\
+\
+  "movq %2,%3;" save_c_m8n4(4,5,6,7) save_c_m8n4(8,9,10,11)
+
+#define SOLVE_LN_m8n12 \
+  "subq $32,%2; movq %2,%3;" GEMM_SUM_REORDER_8x4(4,5,6,7,-32) GEMM_SUM_REORDER_8x4(8,9,10,11,-32) GEMM_SUM_REORDER_8x4(12,13,14,15,-32)\
+\
+  SOLVE_loup_m2n12(-8,7,11,15) SUBTRACT_m2n12(-16,6,10,14) SUBTRACT_m2n12(-24,5,9,13) SUBTRACT_m2n12(-32,4,8,12)\
+  SOLVE_up_m2n12(-40,7,11,15) SUBTRACT_m2n12(-48,6,10,14) SUBTRACT_m2n12(-56,5,9,13) SUBTRACT_m2n12(-64,4,8,12) SAVE_b_m2n12(-32,7,11,15)\
+\
+  SOLVE_loup_m2n12(-80,6,10,14) SUBTRACT_m2n12(-88,5,9,13) SUBTRACT_m2n12(-96,4,8,12)\
+  SOLVE_up_m2n12(-112,6,10,14) SUBTRACT_m2n12(-120,5,9,13) SUBTRACT_m2n12(-128,4,8,12) SAVE_b_m2n12(-64,6,10,14)\
+\
+  SOLVE_loup_m2n12(-152,5,9,13) SUBTRACT_m2n12(-160,4,8,12)\
+  SOLVE_up_m2n12(-184,5,9,13) SUBTRACT_m2n12(-192,4,8,12) SAVE_b_m2n12(-96,5,9,13)\
+\
+  SOLVE_loup_m2n12(-224,4,8,12)\
+  SOLVE_up_m2n12(-256,4,8,12) SAVE_b_m2n12(-128,4,8,12)\
+\
+  "movq %2,%3;" save_c_m8n4(4,5,6,7) save_c_m8n4(8,9,10,11) save_c_m8n4(12,13,14,15)
+
+/* r13 = k-kk, r14 = b_tail, r15 = a_tail */
+
+#define GEMM_LN_SIMPLE(mdim,ndim) \
+  "movq %%r15,%0; negq %%r12; leaq (%%r15,%%r12,"#mdim"),%%r15; negq %%r12;"\
+  "movq %%r13,%5; addq $"#mdim",%%r13; movq %%r14,%1;" INIT_m##mdim##n##ndim\
+  "testq %5,%5; jz 2"#mdim""#ndim"2f;"\
+  "2"#mdim""#ndim"1:\n\t"\
+  "subq $16,%1; subq $"#mdim"*4,%0;" GEMM_KERNEL_k1m##mdim##n##ndim "decq %5; jnz 2"#mdim""#ndim"1b;"\
+  "2"#mdim""#ndim"2:\n\t"
+#define GEMM_LN_m8n4 GEMM_LN_SIMPLE(8,4)
+#define GEMM_LN_m8n8 GEMM_LN_SIMPLE(8,8)
+#define GEMM_LN_m8n12 \
+  "movq %%r15,%0; negq %%r12; leaq (%%r15,%%r12,8),%%r15; negq %%r12; movq %%r13,%5; addq $8,%%r13; movq %%r14,%1;" INIT_m8n12\
+  "cmpq $8,%5; jb 28122f;"\
+  "28121:\n\t"\
+  "prefetcht0 -384(%0); subq $32,%0; subq $16,%1;" GEMM_KERNEL_k1m8n12\
+                       "subq $32,%0; subq $16,%1;" GEMM_KERNEL_k1m8n12\
+  "prefetcht0 -384(%0); subq $32,%0; subq $16,%1;" GEMM_KERNEL_k1m8n12\
+                       "subq $32,%0; subq $16,%1;" GEMM_KERNEL_k1m8n12\
+  "prefetcht0 -384(%0); subq $32,%0; subq $16,%1;" GEMM_KERNEL_k1m8n12\
+                       "subq $32,%0; subq $16,%1;" GEMM_KERNEL_k1m8n12\
+  "prefetcht0 -384(%0); subq $32,%0; subq $16,%1;" GEMM_KERNEL_k1m8n12\
+                       "subq $32,%0; subq $16,%1;" GEMM_KERNEL_k1m8n12\
+  "subq $8,%5; cmpq $8,%5; jnb 28121b;"\
+  "28122:\n\t"\
+  "testq %5,%5; jz 28124f;"\
+  "28123:\n\t"\
+  "subq $32,%0; subq $16,%1;" GEMM_KERNEL_k1m8n12 "decq %5; jnz 28123b;"\
+  "28124:\n\t"
+#define GEMM_LN_m4n4 GEMM_LN_SIMPLE(4,4)
+#define GEMM_LN_m4n8 GEMM_LN_SIMPLE(4,8)
+#define GEMM_LN_m4n12 GEMM_LN_SIMPLE(4,12)
+#define GEMM_LN_m2n4 GEMM_LN_SIMPLE(2,4)
+#define GEMM_LN_m2n8 GEMM_LN_SIMPLE(2,8)
+#define GEMM_LN_m2n12 GEMM_LN_SIMPLE(2,12)
+#define GEMM_LN_m1n4 GEMM_LN_SIMPLE(1,4)
+#define GEMM_LN_m1n8 GEMM_LN_SIMPLE(1,8)
+#define GEMM_LN_m1n12 GEMM_LN_SIMPLE(1,12)
+
+#define COMPUTE(ndim) {\
+  c_ptr += M;\
+  __asm__ __volatile__(\
+    "movq %0,%%r15; movq %7,%%r13; movq %6,%%r12; salq $2,%%r12; leaq (%1,%%r12,4),%%r14; movq %10,%%r11;"\
+    "testq $1,%%r11; jz "#ndim"772f;"\
+    #ndim"771:\n\t"\
+    GEMM_LN_m1n##ndim SOLVE_LN_m1n##ndim "subq $1,%%r11;"\
+    #ndim"772:\n\t"\
+    "testq $2,%%r11; jz "#ndim"773f;"\
+    GEMM_LN_m2n##ndim SOLVE_LN_m2n##ndim "subq $2,%%r11;"\
+    #ndim"773:\n\t"\
+    "testq $4,%%r11; jz "#ndim"774f;"\
+    GEMM_LN_m4n##ndim SOLVE_LN_m4n##ndim "subq $4,%%r11;"\
+    #ndim"774:\n\t"\
+    "testq %%r11,%%r11; jz "#ndim"776f;"\
+    #ndim"775:\n\t"\
+    GEMM_LN_m8n##ndim SOLVE_LN_m8n##ndim "subq $8,%%r11; jnz "#ndim"775b;"\
+    #ndim"776:\n\t"\
+    "movq %%r15,%0; movq %%r14,%1; vzeroupper;"\
+  :"+r"(a_ptr),"+r"(b_ptr),"+r"(c_ptr),"+r"(c_tmp),"+r"(ldc_bytes),"+r"(k_cnt):"m"(K),"m"(kmkkinp),"m"(one[0]),"m"(zero[0]),"m"(M)\
+  :"r11","r12","r13","r14","r15","cc","memory",\
+  "xmm0","xmm1","xmm2","xmm3","xmm4","xmm5","xmm6","xmm7","xmm8","xmm9","xmm10","xmm11","xmm12","xmm13","xmm14","xmm15");\
+  a_ptr += M * K; b_ptr += (ndim-4) * K; c_ptr += ldc * ndim;\
+}
+static void solve_LN(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+  FLOAT a0, b0;
+  int i, j, k;
+  for (i=m-1;i>=0;i--) {
+    a0 = a[i*m+i]; //reciprocal of the original value
+    for (j=0;j<n;j++) {
+      b0 = c[j*ldc+i]*a0;
+      c[j*ldc+i] = b[i*n+j] = b0;
+      for (k=0;k<i;k++) c[j*ldc+k] -= b0*a[i*m+k];
+    }
+  }
+}
+static void COMPUTE_EDGE_1_nchunk(BLASLONG m, BLASLONG n, FLOAT *sa, FLOAT *sb, FLOAT *C, BLASLONG ldc, BLASLONG k, BLASLONG offset) {
+  BLASLONG m_count = m, kk = m+offset; FLOAT *a_ptr = sa+m*k, *c_ptr = C+m;
+  if(m_count&1){
+    a_ptr-=k; c_ptr--;
+    if(k-kk>0) GEMM_KERNEL_N(1,n,k-kk,-1.0,a_ptr+kk*1,sb+kk*n,c_ptr,ldc);
+    solve_LN(1,n,a_ptr+(kk-1)*1,sb+(kk-1)*n,c_ptr,ldc);
+    kk -= 1;
+    m_count--;
+  }
+  if(m_count&2){
+    a_ptr-=k*2; c_ptr-=2;
+    if(k-kk>0) GEMM_KERNEL_N(2,n,k-kk,-1.0,a_ptr+kk*2,sb+kk*n,c_ptr,ldc);
+    solve_LN(2,n,a_ptr+(kk-2)*2,sb+(kk-2)*n,c_ptr,ldc);
+    kk -= 2;
+    m_count-=2;
+  }
+  if(m_count&4){
+    a_ptr-=k*4; c_ptr-=4;
+    if(k-kk>0) GEMM_KERNEL_N(4,n,k-kk,-1.0,a_ptr+kk*4,sb+kk*n,c_ptr,ldc);
+    solve_LN(4,n,a_ptr+(kk-4)*4,sb+(kk-4)*n,c_ptr,ldc);
+    kk -= 4;
+    m_count-=4;
+  }
+  for(;m_count>7;m_count-=8){
+    a_ptr-=k*8; c_ptr-=8;
+    if(k-kk>0) GEMM_KERNEL_N(8,n,k-kk,-1.0,a_ptr+kk*8,sb+kk*n,c_ptr,ldc);
+    solve_LN(8,n,a_ptr+(kk-8)*8,sb+(kk-8)*n,c_ptr,ldc);
+    kk -= 8;
+  }
+}
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT dummy1, FLOAT *sa, FLOAT *sb, FLOAT *C, BLASLONG ldc, BLASLONG offset){
+  float *a_ptr = sa+m*k, *b_ptr = sb, *c_ptr = C, *c_tmp = C;
+  float one[8] = {1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0};
+  float zero[8] = {0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0};
+  uint64_t ldc_bytes = (uint64_t)ldc * sizeof(float), K = (uint64_t)k, M = (uint64_t)m, kmkkinp = (uint64_t)(k-m-offset), k_cnt = 0;
+  BLASLONG n_count = n;
+  for(;n_count>11;n_count-=12) COMPUTE(12)
+  for(;n_count>7;n_count-=8) COMPUTE(8)
+  for(;n_count>3;n_count-=4) COMPUTE(4)
+  for(;n_count>1;n_count-=2) { COMPUTE_EDGE_1_nchunk(m,2,sa,b_ptr,c_ptr,ldc,k,offset); b_ptr += 2*k; c_ptr += ldc*2;}
+  if(n_count>0) COMPUTE_EDGE_1_nchunk(m,1,sa,b_ptr,c_ptr,ldc,k,offset);
+  return 0;
+}
+

kernel/x86_64/strsm_kernel_8x4_haswell_LT.c

@@ -0,0 +1,228 @@
+#include "common.h"
+#include <stdint.h>
+#include "strsm_kernel_8x4_haswell_L_common.h"
+
+#define SOLVE_LT_m1n4 \
+  "movq %2,%3;" GEMM_SUM_REORDER_1x4(4)\
+  SOLVE_m1n4(0,4) SAVE_b_m1n4(0,4)\
+  "movq %2,%3; addq $4,%2;" save_c_m1n4(4)
+
+#define SOLVE_LT_m1n8 \
+  "movq %2,%3;" GEMM_SUM_REORDER_1x4(4) GEMM_SUM_REORDER_1x4(5)\
+  SOLVE_m1n8(0,4,5) SAVE_b_m1n8(0,4,5)\
+  "movq %2,%3; addq $4,%2;" save_c_m1n4(4) save_c_m1n4(5)
+
+#define SOLVE_LT_m1n12 \
+  "movq %2,%3;" GEMM_SUM_REORDER_1x4(4) GEMM_SUM_REORDER_1x4(5) GEMM_SUM_REORDER_1x4(6)\
+  SOLVE_m1n12(0,4,5,6) SAVE_b_m1n12(0,4,5,6)\
+  "movq %2,%3; addq $4,%2;" save_c_m1n4(4) save_c_m1n4(5) save_c_m1n4(6)
+
+#define SOLVE_LT_m2n4 \
+  "movq %2,%3;" GEMM_SUM_REORDER_2x4(4,5,4)\
+  SOLVE_uplo_m2n4(0,4)\
+  SOLVE_lo_m2n4(8,4) SAVE_b_m2n4(0,4)\
+  "movq %2,%3; addq $8,%2;" save_c_m2n4(4)
+
+#define SOLVE_LT_m2n8 \
+  "movq %2,%3;" GEMM_SUM_REORDER_2x4(4,5,4) GEMM_SUM_REORDER_2x4(6,7,5)\
+  SOLVE_uplo_m2n8(0,4,5)\
+  SOLVE_lo_m2n8(8,4,5) SAVE_b_m2n8(0,4,5)\
+  "movq %2,%3; addq $8,%2;" save_c_m2n4(4) save_c_m2n4(5)
+
+#define SOLVE_LT_m2n12 \
+  "movq %2,%3;" GEMM_SUM_REORDER_2x4(4,5,4) GEMM_SUM_REORDER_2x4(6,7,5) GEMM_SUM_REORDER_2x4(8,9,6)\
+  SOLVE_uplo_m2n12(0,4,5,6)\
+  SOLVE_lo_m2n12(8,4,5,6) SAVE_b_m2n12(0,4,5,6)\
+  "movq %2,%3; addq $8,%2;" save_c_m2n4(4) save_c_m2n4(5) save_c_m2n4(6)
+
+#define SOLVE_LT_m4n4 \
+  "movq %2,%3;" GEMM_SUM_REORDER_4x4(4,5,6,7,4,5)\
+\
+  SOLVE_uplo_m2n4(0,4) SUBTRACT_m2n4(8,5)\
+  SOLVE_lo_m2n4(16,4) SUBTRACT_m2n4(24,5) SAVE_b_m2n4(0,4)\
+\
+  SOLVE_uplo_m2n4(40,5)\
+  SOLVE_lo_m2n4(56,5) SAVE_b_m2n4(32,5)\
+\
+  "movq %2,%3; addq $16,%2;" save_c_m4n4(4,5)
+
+#define SOLVE_LT_m4n8 \
+  "movq %2,%3;" GEMM_SUM_REORDER_4x4(4,5,6,7,4,5) GEMM_SUM_REORDER_4x4(8,9,10,11,6,7)\
+\
+  SOLVE_uplo_m2n8(0,4,6) SUBTRACT_m2n8(8,5,7)\
+  SOLVE_lo_m2n8(16,4,6) SUBTRACT_m2n8(24,5,7) SAVE_b_m2n8(0,4,6)\
+\
+  SOLVE_uplo_m2n8(40,5,7)\
+  SOLVE_lo_m2n8(56,5,7) SAVE_b_m2n8(32,5,7)\
+\
+  "movq %2,%3; addq $16,%2;" save_c_m4n4(4,5) save_c_m4n4(6,7)
+
+#define SOLVE_LT_m4n12 \
+  "movq %2,%3;" GEMM_SUM_REORDER_4x4(4,5,6,7,4,5) GEMM_SUM_REORDER_4x4(8,9,10,11,6,7) GEMM_SUM_REORDER_4x4(12,13,14,15,8,9)\
+\
+  SOLVE_uplo_m2n12(0,4,6,8) SUBTRACT_m2n12(8,5,7,9)\
+  SOLVE_lo_m2n12(16,4,6,8) SUBTRACT_m2n12(24,5,7,9) SAVE_b_m2n12(0,4,6,8)\
+\
+  SOLVE_uplo_m2n12(40,5,7,9)\
+  SOLVE_lo_m2n12(56,5,7,9) SAVE_b_m2n12(32,5,7,9)\
+\
+  "movq %2,%3; addq $16,%2;" save_c_m4n4(4,5) save_c_m4n4(6,7) save_c_m4n4(8,9)
+
+#define SOLVE_LT_m8n4 \
+  "movq %2,%3;" GEMM_SUM_REORDER_8x4(4,5,6,7,63)\
+\
+  SOLVE_uplo_m2n4(0,4) SUBTRACT_m2n4(8,5) SUBTRACT_m2n4(16,6) SUBTRACT_m2n4(24,7)\
+  SOLVE_lo_m2n4(32,4) SUBTRACT_m2n4(40,5) SUBTRACT_m2n4(48,6) SUBTRACT_m2n4(56,7) SAVE_b_m2n4(0,4)\
+\
+  SOLVE_uplo_m2n4(72,5) SUBTRACT_m2n4(80,6) SUBTRACT_m2n4(88,7)\
+  SOLVE_lo_m2n4(104,5) SUBTRACT_m2n4(112,6) SUBTRACT_m2n4(120,7) SAVE_b_m2n4(32,5)\
+\
+  SOLVE_uplo_m2n4(144,6) SUBTRACT_m2n4(152,7)\
+  SOLVE_lo_m2n4(176,6) SUBTRACT_m2n4(184,7) SAVE_b_m2n4(64,6)\
+\
+  SOLVE_uplo_m2n4(216,7)\
+  SOLVE_lo_m2n4(248,7) SAVE_b_m2n4(96,7)\
+\
+  "movq %2,%3; addq $32,%2;" save_c_m8n4(4,5,6,7)
+
+#define SOLVE_LT_m8n8 \
+  "movq %2,%3;" GEMM_SUM_REORDER_8x4(4,5,6,7,63) GEMM_SUM_REORDER_8x4(8,9,10,11,63)\
+\
+  SOLVE_uplo_m2n8(0,4,8) SUBTRACT_m2n8(8,5,9) SUBTRACT_m2n8(16,6,10) SUBTRACT_m2n8(24,7,11)\
+  SOLVE_lo_m2n8(32,4,8) SUBTRACT_m2n8(40,5,9) SUBTRACT_m2n8(48,6,10) SUBTRACT_m2n8(56,7,11) SAVE_b_m2n8(0,4,8)\
+\
+  SOLVE_uplo_m2n8(72,5,9) SUBTRACT_m2n8(80,6,10) SUBTRACT_m2n8(88,7,11)\
+  SOLVE_lo_m2n8(104,5,9) SUBTRACT_m2n8(112,6,10) SUBTRACT_m2n8(120,7,11) SAVE_b_m2n8(32,5,9)\
+\
+  SOLVE_uplo_m2n8(144,6,10) SUBTRACT_m2n8(152,7,11)\
+  SOLVE_lo_m2n8(176,6,10) SUBTRACT_m2n8(184,7,11) SAVE_b_m2n8(64,6,10)\
+\
+  SOLVE_uplo_m2n8(216,7,11)\
+  SOLVE_lo_m2n8(248,7,11) SAVE_b_m2n8(96,7,11)\
+\
+  "movq %2,%3; addq $32,%2;" save_c_m8n4(4,5,6,7) save_c_m8n4(8,9,10,11)
+
+#define SOLVE_LT_m8n12 \
+  "movq %2,%3;" GEMM_SUM_REORDER_8x4(4,5,6,7,63) GEMM_SUM_REORDER_8x4(8,9,10,11,63) GEMM_SUM_REORDER_8x4(12,13,14,15,63)\
+\
+  SOLVE_uplo_m2n12(0,4,8,12) SUBTRACT_m2n12(8,5,9,13) SUBTRACT_m2n12(16,6,10,14) SUBTRACT_m2n12(24,7,11,15)\
+  SOLVE_lo_m2n12(32,4,8,12) SUBTRACT_m2n12(40,5,9,13) SUBTRACT_m2n12(48,6,10,14) SUBTRACT_m2n12(56,7,11,15) SAVE_b_m2n12(0,4,8,12)\
+\
+  SOLVE_uplo_m2n12(72,5,9,13) SUBTRACT_m2n12(80,6,10,14) SUBTRACT_m2n12(88,7,11,15)\
+  SOLVE_lo_m2n12(104,5,9,13) SUBTRACT_m2n12(112,6,10,14) SUBTRACT_m2n12(120,7,11,15) SAVE_b_m2n12(32,5,9,13)\
+\
+  SOLVE_uplo_m2n12(144,6,10,14) SUBTRACT_m2n12(152,7,11,15)\
+  SOLVE_lo_m2n12(176,6,10,14) SUBTRACT_m2n12(184,7,11,15) SAVE_b_m2n12(64,6,10,14)\
+\
+  SOLVE_uplo_m2n12(216,7,11,15)\
+  SOLVE_lo_m2n12(248,7,11,15) SAVE_b_m2n12(96,7,11,15)\
+\
+  "movq %2,%3; addq $32,%2;" save_c_m8n4(4,5,6,7) save_c_m8n4(8,9,10,11) save_c_m8n4(12,13,14,15)
+
+#define GEMM_LT_SIMPLE(mdim,ndim) \
+  "movq %%r15,%0; leaq (%%r15,%%r12,"#mdim"),%%r15; movq %%r13,%5; addq $"#mdim",%%r13; movq %%r14,%1;" INIT_m##mdim##n##ndim\
+  "testq %5,%5; jz 1"#mdim""#ndim"2f;"\
+  "1"#mdim""#ndim"1:\n\t"\
+  GEMM_KERNEL_k1m##mdim##n##ndim "addq $16,%1; addq $"#mdim"*4,%0; decq %5; jnz 1"#mdim""#ndim"1b;"\
+  "1"#mdim""#ndim"2:\n\t"
+#define GEMM_LT_m8n4 GEMM_LT_SIMPLE(8,4)
+#define GEMM_LT_m8n8 GEMM_LT_SIMPLE(8,8)
+#define GEMM_LT_m8n12 \
+  "movq %%r15,%0; leaq (%%r15,%%r12,8),%%r15; movq %%r13,%5; addq $8,%%r13; movq %%r14,%1;" INIT_m8n12\
+  "cmpq $8,%5; jb 18122f;"\
+  "18121:\n\t"\
+  GEMM_KERNEL_k1m8n12 "prefetcht0 384(%0); addq $32,%0; addq $16,%1;"\
+  GEMM_KERNEL_k1m8n12 "addq $32,%0; addq $16,%1;"\
+  GEMM_KERNEL_k1m8n12 "prefetcht0 384(%0); addq $32,%0; addq $16,%1;"\
+  GEMM_KERNEL_k1m8n12 "addq $32,%0; addq $16,%1;"\
+  GEMM_KERNEL_k1m8n12 "prefetcht0 384(%0); addq $32,%0; addq $16,%1;"\
+  GEMM_KERNEL_k1m8n12 "addq $32,%0; addq $16,%1;"\
+  GEMM_KERNEL_k1m8n12 "prefetcht0 384(%0); addq $32,%0; addq $16,%1;"\
+  GEMM_KERNEL_k1m8n12 "addq $32,%0; addq $16,%1;"\
+  "subq $8,%5; cmpq $8,%5; jnb 18121b;"\
+  "18122:\n\t"\
+  "testq %5,%5; jz 18124f;"\
+  "18123:\n\t"\
+  GEMM_KERNEL_k1m8n12 "addq $32,%0; addq $16,%1; decq %5; jnz 18123b;"\
+  "18124:\n\t"
+#define GEMM_LT_m4n4 GEMM_LT_SIMPLE(4,4)
+#define GEMM_LT_m4n8 GEMM_LT_SIMPLE(4,8)
+#define GEMM_LT_m4n12 GEMM_LT_SIMPLE(4,12)
+#define GEMM_LT_m2n4 GEMM_LT_SIMPLE(2,4)
+#define GEMM_LT_m2n8 GEMM_LT_SIMPLE(2,8)
+#define GEMM_LT_m2n12 GEMM_LT_SIMPLE(2,12)
+#define GEMM_LT_m1n4 GEMM_LT_SIMPLE(1,4)
+#define GEMM_LT_m1n8 GEMM_LT_SIMPLE(1,8)
+#define GEMM_LT_m1n12 GEMM_LT_SIMPLE(1,12)
+
+#define COMPUTE(ndim) {\
+  __asm__ __volatile__(\
+    "movq %0,%%r15; movq %1,%%r14; movq %7,%%r13; movq %6,%%r12; salq $2,%%r12; movq %10,%%r11;"\
+    "cmpq $8,%%r11; jb "#ndim"772f;"\
+    #ndim"771:\n\t"\
+    GEMM_LT_m8n##ndim SOLVE_LT_m8n##ndim "subq $8,%%r11; cmpq $8,%%r11; jnb "#ndim"771b;"\
+    #ndim"772:\n\t"\
+    "testq $4,%%r11; jz "#ndim"773f;"\
+    GEMM_LT_m4n##ndim SOLVE_LT_m4n##ndim "subq $4,%%r11;"\
+    #ndim"773:\n\t"\
+    "testq $2,%%r11; jz "#ndim"774f;"\
+    GEMM_LT_m2n##ndim SOLVE_LT_m2n##ndim "subq $2,%%r11;"\
+    #ndim"774:\n\t"\
+    "testq $1,%%r11; jz "#ndim"775f;"\
+    GEMM_LT_m1n##ndim SOLVE_LT_m1n##ndim "subq $1,%%r11;"\
+    #ndim"775:\n\t"\
+    "movq %%r15,%0; movq %%r14,%1; vzeroupper;"\
+  :"+r"(a_ptr),"+r"(b_ptr),"+r"(c_ptr),"+r"(c_tmp),"+r"(ldc_bytes),"+r"(k_cnt):"m"(K),"m"(OFF),"m"(one[0]),"m"(zero[0]),"m"(M)\
+  :"r11","r12","r13","r14","r15","cc","memory",\
+  "xmm0","xmm1","xmm2","xmm3","xmm4","xmm5","xmm6","xmm7","xmm8","xmm9","xmm10","xmm11","xmm12","xmm13","xmm14","xmm15");\
+  a_ptr -= M * K; b_ptr += ndim * K; c_ptr += ldc * ndim - M;\
+}
+static void solve_LT(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+  FLOAT a0, b0;
+  int i, j, k;
+  for (i=0;i<m;i++) {
+    a0 = a[i*m+i];
+    for (j=0;j<n;j++) {
+      b0 = c[j*ldc+i] * a0;
+      b[i*n+j] = c[j*ldc+i] = b0;
+      for (k=i+1;k<m;k++) c[j*ldc+k] -= b0 * a[i*m+k];
+    }
+  }
+}
+static void COMPUTE_EDGE_1_nchunk(BLASLONG m, BLASLONG n, FLOAT *sa, FLOAT *sb, FLOAT *C, BLASLONG ldc, BLASLONG k, BLASLONG offset) {
+  BLASLONG m_count = m, kk = offset; FLOAT *a_ptr = sa, *c_ptr = C;
+  for(;m_count>7;m_count-=8){
+    if(kk>0) GEMM_KERNEL_N(8,n,kk,-1.0,a_ptr,sb,c_ptr,ldc);
+    solve_LT(8,n,a_ptr+kk*8,sb+kk*n,c_ptr,ldc);
+    kk += 8; a_ptr += k * 8; c_ptr += 8;
+  }
+  for(;m_count>3;m_count-=4){
+    if(kk>0) GEMM_KERNEL_N(4,n,kk,-1.0,a_ptr,sb,c_ptr,ldc);
+    solve_LT(4,n,a_ptr+kk*4,sb+kk*n,c_ptr,ldc);
+    kk += 4; a_ptr += k * 4; c_ptr += 4;
+  }
+  for(;m_count>1;m_count-=2){
+    if(kk>0) GEMM_KERNEL_N(2,n,kk,-1.0,a_ptr,sb,c_ptr,ldc);
+    solve_LT(2,n,a_ptr+kk*2,sb+kk*n,c_ptr,ldc);
+    kk += 2; a_ptr += k * 2; c_ptr += 2;
+  }
+  if(m_count>0){
+    if(kk>0) GEMM_KERNEL_N(1,n,kk,-1.0,a_ptr,sb,c_ptr,ldc);
+    solve_LT(1,n,a_ptr+kk*1,sb+kk*n,c_ptr,ldc);
+    kk += 1; a_ptr += k * 1; c_ptr += 1;
+  }
+}
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT dummy1, FLOAT *sa, FLOAT *sb, FLOAT *C, BLASLONG ldc, BLASLONG offset){
+  float *a_ptr = sa, *b_ptr = sb, *c_ptr = C, *c_tmp = C;
+  float one[8] = {1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0};
+  float zero[8] = {0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0};
+  uint64_t ldc_bytes = (uint64_t)ldc * sizeof(float), K = (uint64_t)k, M = (uint64_t)m, OFF = (uint64_t)offset, k_cnt = 0;
+  BLASLONG n_count = n;
+  for(;n_count>11;n_count-=12) COMPUTE(12)
+  for(;n_count>7;n_count-=8) COMPUTE(8)
+  for(;n_count>3;n_count-=4) COMPUTE(4)
+  for(;n_count>1;n_count-=2) { COMPUTE_EDGE_1_nchunk(m,2,a_ptr,b_ptr,c_ptr,ldc,k,offset); b_ptr += 2*k; c_ptr += ldc*2;}
+  if(n_count>0) COMPUTE_EDGE_1_nchunk(m,1,a_ptr,b_ptr,c_ptr,ldc,k,offset);
+  return 0;
+}
+

kernel/x86_64/strsm_kernel_8x4_haswell_L_common.h

@@ -0,0 +1,187 @@
+/* r11 = m_counter, r12 = size_of_k_elements, r13 = kk, r14 = b_head, r15 = a_head */
+/* register i/o: %0 = a_ptr, %1 = b_ptr, %2 = c_ptr, %3 = c_tmp, %4 = ldc, %5 = k_counter */
+/* memory input: %6 = K, %7 = offset, %8 = {1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0}, %9 = {0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0}, %10 = M */
+
+#define init_m8n4(c1,c2,c3,c4)\
+  "vpxor %%ymm"#c1",%%ymm"#c1",%%ymm"#c1"; vpxor %%ymm"#c2",%%ymm"#c2",%%ymm"#c2"; vpxor %%ymm"#c3",%%ymm"#c3",%%ymm"#c3"; vpxor %%ymm"#c4",%%ymm"#c4",%%ymm"#c4";"
+#define INIT_m8n4 init_m8n4(4,5,6,7)
+#define INIT_m8n8 INIT_m8n4 init_m8n4(8,9,10,11)
+#define INIT_m8n12 INIT_m8n8 init_m8n4(12,13,14,15)
+
+#define init_m4n4(c1,c2,c3,c4)\
+  "vpxor %%xmm"#c1",%%xmm"#c1",%%xmm"#c1"; vpxor %%xmm"#c2",%%xmm"#c2",%%xmm"#c2"; vpxor %%xmm"#c3",%%xmm"#c3",%%xmm"#c3"; vpxor %%xmm"#c4",%%xmm"#c4",%%xmm"#c4";"
+#define INIT_m4n4 init_m4n4(4,5,6,7)
+#define INIT_m4n8 INIT_m4n4 init_m4n4(8,9,10,11)
+#define INIT_m4n12 INIT_m4n8 init_m4n4(12,13,14,15)
+
+#define init_m2n4(c1,c2)\
+  "vpxor %%xmm"#c1",%%xmm"#c1",%%xmm"#c1"; vpxor %%xmm"#c2",%%xmm"#c2",%%xmm"#c2";"
+#define INIT_m2n4 init_m2n4(4,5)
+#define INIT_m2n8 INIT_m2n4 init_m2n4(6,7)
+#define INIT_m2n12 INIT_m2n8 init_m2n4(8,9)
+
+#define init_m1n4(c1) "vpxor %%xmm"#c1",%%xmm"#c1",%%xmm"#c1";"
+#define INIT_m1n4 init_m1n4(4)
+#define INIT_m1n8 INIT_m1n4 init_m1n4(5)
+#define INIT_m1n12 INIT_m1n8 init_m1n4(6)
+
+#define GEMM_KERNEL_k1m8n4 \
+  "vmovsldup (%0),%%ymm1; vmovshdup (%0),%%ymm2;"\
+  "vbroadcastsd (%1),%%ymm3; vfnmadd231ps %%ymm3,%%ymm1,%%ymm4; vfnmadd231ps %%ymm3,%%ymm2,%%ymm5;"\
+  "vbroadcastsd 8(%1),%%ymm3; vfnmadd231ps %%ymm3,%%ymm1,%%ymm6; vfnmadd231ps %%ymm3,%%ymm2,%%ymm7;"
+#define GEMM_KERNEL_k1m8n8 GEMM_KERNEL_k1m8n4\
+  "vbroadcastsd (%1,%%r12,4),%%ymm3; vfnmadd231ps %%ymm3,%%ymm1,%%ymm8; vfnmadd231ps %%ymm3,%%ymm2,%%ymm9;"\
+  "vbroadcastsd 8(%1,%%r12,4),%%ymm3; vfnmadd231ps %%ymm3,%%ymm1,%%ymm10; vfnmadd231ps %%ymm3,%%ymm2,%%ymm11;"
+#define GEMM_KERNEL_k1m8n12 GEMM_KERNEL_k1m8n8\
+  "vbroadcastsd (%1,%%r12,8),%%ymm3; vfnmadd231ps %%ymm3,%%ymm1,%%ymm12; vfnmadd231ps %%ymm3,%%ymm2,%%ymm13;"\
+  "vbroadcastsd 8(%1,%%r12,8),%%ymm3; vfnmadd231ps %%ymm3,%%ymm1,%%ymm14; vfnmadd231ps %%ymm3,%%ymm2,%%ymm15;"
+
+#define GEMM_KERNEL_k1m4n4 \
+  "vmovsldup (%0),%%xmm1; vmovshdup (%0),%%xmm2;"\
+  "vmovddup (%1),%%xmm3; vfnmadd231ps %%xmm3,%%xmm1,%%xmm4; vfnmadd231ps %%xmm3,%%xmm2,%%xmm5;"\
+  "vmovddup 8(%1),%%xmm3; vfnmadd231ps %%xmm3,%%xmm1,%%xmm6; vfnmadd231ps %%xmm3,%%xmm2,%%xmm7;"
+#define GEMM_KERNEL_k1m4n8 GEMM_KERNEL_k1m4n4\
+  "vmovddup (%1,%%r12,4),%%xmm3; vfnmadd231ps %%xmm3,%%xmm1,%%xmm8; vfnmadd231ps %%xmm3,%%xmm2,%%xmm9;"\
+  "vmovddup 8(%1,%%r12,4),%%xmm3; vfnmadd231ps %%xmm3,%%xmm1,%%xmm10; vfnmadd231ps %%xmm3,%%xmm2,%%xmm11;"
+#define GEMM_KERNEL_k1m4n12 GEMM_KERNEL_k1m4n8\
+  "vmovddup (%1,%%r12,8),%%xmm3; vfnmadd231ps %%xmm3,%%xmm1,%%xmm12; vfnmadd231ps %%xmm3,%%xmm2,%%xmm13;"\
+  "vmovddup 8(%1,%%r12,8),%%xmm3; vfnmadd231ps %%xmm3,%%xmm1,%%xmm14; vfnmadd231ps %%xmm3,%%xmm2,%%xmm15;"
+
+#define GEMM_KERNEL_k1m2n4 \
+  "vbroadcastss (%0),%%xmm1; vbroadcastss 4(%0),%%xmm2;"\
+  "vmovups (%1),%%xmm3; vfnmadd231ps %%xmm3,%%xmm1,%%xmm4; vfnmadd231ps %%xmm3,%%xmm2,%%xmm5;"
+#define GEMM_KERNEL_k1m2n8 GEMM_KERNEL_k1m2n4\
+  "vmovups (%1,%%r12,4),%%xmm3; vfnmadd231ps %%xmm3,%%xmm1,%%xmm6; vfnmadd231ps %%xmm3,%%xmm2,%%xmm7;"
+#define GEMM_KERNEL_k1m2n12 GEMM_KERNEL_k1m2n8\
+  "vmovups (%1,%%r12,8),%%xmm3; vfnmadd231ps %%xmm3,%%xmm1,%%xmm8; vfnmadd231ps %%xmm3,%%xmm2,%%xmm9;"
+
+#define GEMM_KERNEL_k1m1n4 "vbroadcastss (%0),%%xmm1; vfnmadd231ps (%1),%%xmm1,%%xmm4;"
+#define GEMM_KERNEL_k1m1n8 GEMM_KERNEL_k1m1n4 "vfnmadd231ps (%1,%%r12,4),%%xmm1,%%xmm5;"
+#define GEMM_KERNEL_k1m1n12 GEMM_KERNEL_k1m1n8 "vfnmadd231ps (%1,%%r12,8),%%xmm1,%%xmm6;"
+
+#define GEMM_SUM_REORDER_8x4(c1,c2,c3,c4,prefpos)\
+  "vunpcklps %%ymm"#c2",%%ymm"#c1",%%ymm0; vunpckhps %%ymm"#c2",%%ymm"#c1",%%ymm1;"\
+  "vunpcklps %%ymm"#c4",%%ymm"#c3",%%ymm2; vunpckhps %%ymm"#c4",%%ymm"#c3",%%ymm3;"\
+  "vmovups (%3),%%ymm"#c1"; vmovups (%3,%4,1),%%ymm"#c2"; prefetcht1 "#prefpos"(%3); prefetcht1 "#prefpos"(%3,%4,1);"\
+  "vunpcklpd %%ymm"#c2",%%ymm"#c1",%%ymm"#c3"; vunpckhpd %%ymm"#c2",%%ymm"#c1",%%ymm"#c4";"\
+  "vaddps %%ymm0,%%ymm"#c3",%%ymm0; vaddps %%ymm1,%%ymm"#c4",%%ymm1;"\
+  "leaq (%3,%4,2),%3;"\
+  "vmovups (%3),%%ymm"#c1"; vmovups (%3,%4,1),%%ymm"#c2"; prefetcht1 "#prefpos"(%3); prefetcht1 "#prefpos"(%3,%4,1);"\
+  "vunpcklpd %%ymm"#c2",%%ymm"#c1",%%ymm"#c3"; vunpckhpd %%ymm"#c2",%%ymm"#c1",%%ymm"#c4";"\
+  "vaddps %%ymm2,%%ymm"#c3",%%ymm2; vaddps %%ymm3,%%ymm"#c4",%%ymm3;"\
+  "leaq (%3,%4,2),%3;"\
+  "vperm2f128 $2,%%ymm0,%%ymm2,%%ymm"#c1"; vperm2f128 $2,%%ymm1,%%ymm3,%%ymm"#c2";"\
+  "vperm2f128 $19,%%ymm0,%%ymm2,%%ymm"#c3"; vperm2f128 $19,%%ymm1,%%ymm3,%%ymm"#c4";"
+
+#define GEMM_SUM_REORDER_4x4(c1,c2,c3,c4,co1,co2)\
+  "vunpcklps %%xmm"#c2",%%xmm"#c1",%%xmm0; vunpckhps %%xmm"#c2",%%xmm"#c1",%%xmm1;"\
+  "vunpcklps %%xmm"#c4",%%xmm"#c3",%%xmm2; vunpckhps %%xmm"#c4",%%xmm"#c3",%%xmm3;"\
+  "vmovups (%3),%%xmm"#c1"; vmovups (%3,%4,1),%%xmm"#c2";"\
+  "vunpcklpd %%xmm"#c2",%%xmm"#c1",%%xmm"#c3"; vunpckhpd %%xmm"#c2",%%xmm"#c1",%%xmm"#c4";"\
+  "vaddps %%xmm0,%%xmm"#c3",%%xmm0; vaddps %%xmm1,%%xmm"#c4",%%xmm1;"\
+  "leaq (%3,%4,2),%3;"\
+  "vmovups (%3),%%xmm"#c1"; vmovups (%3,%4,1),%%xmm"#c2";"\
+  "vunpcklpd %%xmm"#c2",%%xmm"#c1",%%xmm"#c3"; vunpckhpd %%xmm"#c2",%%xmm"#c1",%%xmm"#c4";"\
+  "vaddps %%xmm2,%%xmm"#c3",%%xmm2; vaddps %%xmm3,%%xmm"#c4",%%xmm3;"\
+  "leaq (%3,%4,2),%3;"\
+  "vperm2f128 $2,%%ymm0,%%ymm2,%%ymm"#co1"; vperm2f128 $2,%%ymm1,%%ymm3,%%ymm"#co2";"
+
+#define GEMM_SUM_REORDER_2x4(c1,c2,co1)\
+  "vunpcklps %%xmm"#c2",%%xmm"#c1",%%xmm0; vunpckhps %%xmm"#c2",%%xmm"#c1",%%xmm1;"\
+  "vmovsd (%3),%%xmm2; vmovhpd (%3,%4,1),%%xmm2,%%xmm2; vaddps %%xmm0,%%xmm2,%%xmm0; leaq (%3,%4,2),%3;"\
+  "vmovsd (%3),%%xmm2; vmovhpd (%3,%4,1),%%xmm2,%%xmm2; vaddps %%xmm1,%%xmm2,%%xmm1; leaq (%3,%4,2),%3;"\
+  "vperm2f128 $2,%%ymm0,%%ymm1,%%ymm"#co1";"
+
+#define GEMM_SUM_REORDER_1x4(c1)\
+  "vmovss (%3),%%xmm1; vinsertps $16,(%3,%4,1),%%xmm1,%%xmm1; leaq (%3,%4,2),%3;"\
+  "vinsertps $32,(%3),%%xmm1,%%xmm1; vinsertps $48,(%3,%4,1),%%xmm1,%%xmm1; leaq (%3,%4,2),%3;"\
+  "vaddps %%xmm"#c1",%%xmm1,%%xmm"#c1";"
+
+#define save_c_m8n4(c1,c2,c3,c4)\
+  "vunpcklpd %%ymm"#c2",%%ymm"#c1",%%ymm0; vunpckhpd %%ymm"#c2",%%ymm"#c1",%%ymm1;"\
+  "vunpcklpd %%ymm"#c4",%%ymm"#c3",%%ymm2; vunpckhpd %%ymm"#c4",%%ymm"#c3",%%ymm3;"\
+  "vperm2f128 $2,%%ymm0,%%ymm2,%%ymm"#c1"; vperm2f128 $2,%%ymm1,%%ymm3,%%ymm"#c2";"\
+  "vmovups %%ymm"#c1",(%3); vmovups %%ymm"#c2",(%3,%4,1); leaq (%3,%4,2),%3;"\
+  "vperm2f128 $19,%%ymm0,%%ymm2,%%ymm"#c3"; vperm2f128 $19,%%ymm1,%%ymm3,%%ymm"#c4";"\
+  "vmovups %%ymm"#c3",(%3); vmovups %%ymm"#c4",(%3,%4,1); leaq (%3,%4,2),%3;"
+
+#define save_c_m4n4(c1,c2)\
+  "vunpcklpd %%ymm"#c2",%%ymm"#c1",%%ymm0; vunpckhpd %%ymm"#c2",%%ymm"#c1",%%ymm1;"\
+  "vmovups %%xmm0,(%3); vmovups %%xmm1,(%3,%4,1); leaq (%3,%4,2),%3;"\
+  "vextractf128 $1,%%ymm0,(%3); vextractf128 $1,%%ymm1,(%3,%4,1); leaq (%3,%4,2),%3;"
+
+#define save_c_m2n4(c1)\
+  "vextractf128 $1,%%ymm"#c1",%%xmm1; vmovsd %%xmm"#c1",(%3); vmovhpd %%xmm"#c1",(%3,%4,1); leaq (%3,%4,2),%3;"\
+  "vmovsd %%xmm1,(%3); vmovhpd %%xmm1,(%3,%4,1); leaq (%3,%4,2),%3;"
+
+#define save_c_m1n4(c1)\
+  "vmovss %%xmm"#c1",(%3); vextractps $1,%%xmm"#c1",(%3,%4,1); leaq (%3,%4,2),%3;"\
+  "vextractps $2,%%xmm"#c1",(%3); vextractps $3,%%xmm"#c1",(%3,%4,1); leaq (%3,%4,2),%3;"
+
+#define SOLVE_up_m2n4(a_off,c1)\
+  "vbroadcastsd "#a_off"(%0),%%ymm0; vblendps $170,%8,%%ymm0,%%ymm2;"\
+  "vmulps %%ymm2,%%ymm"#c1",%%ymm"#c1";"\
+  "vmovsldup %%ymm"#c1",%%ymm1;"
+
+#define SOLVE_up_m2n8(a_off,c1,c2)\
+  "vbroadcastsd "#a_off"(%0),%%ymm0; vblendps $170,%8,%%ymm0,%%ymm2;"\
+  "vmulps %%ymm2,%%ymm"#c1",%%ymm"#c1"; vmulps %%ymm2,%%ymm"#c2",%%ymm"#c2";"\
+  "vmovsldup %%ymm"#c1",%%ymm1; vmovsldup %%ymm"#c2",%%ymm2;"
+
+#define SOLVE_up_m2n12(a_off,c1,c2,c3)\
+  "vbroadcastsd "#a_off"(%0),%%ymm0; vblendps $170,%8,%%ymm0,%%ymm2;"\
+  "vmulps %%ymm2,%%ymm"#c1",%%ymm"#c1"; vmulps %%ymm2,%%ymm"#c2",%%ymm"#c2"; vmulps %%ymm2,%%ymm"#c3",%%ymm"#c3";"\
+  "vmovsldup %%ymm"#c1",%%ymm1; vmovsldup %%ymm"#c2",%%ymm2; vmovsldup %%ymm"#c3",%%ymm3;"
+
+#define SOLVE_uplo_m2n4(a_off,c1) SOLVE_up_m2n4(a_off,c1)\
+  "vblendps $85,%9,%%ymm0,%%ymm0; vfnmadd231ps %%ymm0,%%ymm1,%%ymm"#c1";"
+
+#define SOLVE_uplo_m2n8(a_off,c1,c2) SOLVE_up_m2n8(a_off,c1,c2)\
+  "vblendps $85,%9,%%ymm0,%%ymm0; vfnmadd231ps %%ymm0,%%ymm1,%%ymm"#c1"; vfnmadd231ps %%ymm0,%%ymm2,%%ymm"#c2";"
+
+#define SOLVE_uplo_m2n12(a_off,c1,c2,c3) SOLVE_up_m2n12(a_off,c1,c2,c3)\
+  "vblendps $85,%9,%%ymm0,%%ymm0; vfnmadd231ps %%ymm0,%%ymm1,%%ymm"#c1"; vfnmadd231ps %%ymm0,%%ymm2,%%ymm"#c2"; vfnmadd231ps %%ymm0,%%ymm3,%%ymm"#c3";"
+
+#define SOLVE_lo_m2n4(a_off,c1)\
+  "vbroadcastsd "#a_off"(%0),%%ymm0; vblendps $85,%8,%%ymm0,%%ymm2;"\
+  "vmulps %%ymm2,%%ymm"#c1",%%ymm"#c1";"\
+  "vmovshdup %%ymm"#c1",%%ymm1;"
+
+#define SOLVE_lo_m2n8(a_off,c1,c2)\
+  "vbroadcastsd "#a_off"(%0),%%ymm0; vblendps $85,%8,%%ymm0,%%ymm2;"\
+  "vmulps %%ymm2,%%ymm"#c1",%%ymm"#c1"; vmulps %%ymm2,%%ymm"#c2",%%ymm"#c2";"\
+  "vmovshdup %%ymm"#c1",%%ymm1; vmovshdup %%ymm"#c2",%%ymm2;"
+
+#define SOLVE_lo_m2n12(a_off,c1,c2,c3)\
+  "vbroadcastsd "#a_off"(%0),%%ymm0; vblendps $85,%8,%%ymm0,%%ymm2;"\
+  "vmulps %%ymm2,%%ymm"#c1",%%ymm"#c1"; vmulps %%ymm2,%%ymm"#c2",%%ymm"#c2"; vmulps %%ymm2,%%ymm"#c3",%%ymm"#c3";"\
+  "vmovshdup %%ymm"#c1",%%ymm1; vmovshdup %%ymm"#c2",%%ymm2; vmovshdup %%ymm"#c3",%%ymm3;"
+
+#define SOLVE_loup_m2n4(a_off,c1) SOLVE_lo_m2n4(a_off,c1)\
+  "vblendps $170,%9,%%ymm0,%%ymm0; vfnmadd231ps %%ymm0,%%ymm1,%%ymm"#c1";"
+
+#define SOLVE_loup_m2n8(a_off,c1,c2) SOLVE_lo_m2n8(a_off,c1,c2)\
+  "vblendps $170,%9,%%ymm0,%%ymm0; vfnmadd231ps %%ymm0,%%ymm1,%%ymm"#c1"; vfnmadd231ps %%ymm0,%%ymm2,%%ymm"#c2";"
+
+#define SOLVE_loup_m2n12(a_off,c1,c2,c3) SOLVE_lo_m2n12(a_off,c1,c2,c3)\
+  "vblendps $170,%9,%%ymm0,%%ymm0; vfnmadd231ps %%ymm0,%%ymm1,%%ymm"#c1"; vfnmadd231ps %%ymm0,%%ymm2,%%ymm"#c2"; vfnmadd231ps %%ymm0,%%ymm3,%%ymm"#c3";"
+
+#define SOLVE_m1n4(a_off,c1) "vbroadcastss "#a_off"(%0),%%xmm0; vmulps %%xmm0,%%xmm"#c1",%%xmm"#c1";"
+#define SOLVE_m1n8(a_off,c1,c2) SOLVE_m1n4(a_off,c1) "vmulps %%xmm0,%%xmm"#c2",%%xmm"#c2";"
+#define SOLVE_m1n12(a_off,c1,c2,c3) SOLVE_m1n8(a_off,c1,c2) "vmulps %%xmm0,%%xmm"#c3",%%xmm"#c3";"
+
+#define SUBTRACT_m2n4(a_off,c1) "vbroadcastsd "#a_off"(%0),%%ymm0; vfnmadd231ps %%ymm0,%%ymm1,%%ymm"#c1";"
+#define SUBTRACT_m2n8(a_off,c1,c2) SUBTRACT_m2n4(a_off,c1) "vfnmadd231ps %%ymm0,%%ymm2,%%ymm"#c2";"
+#define SUBTRACT_m2n12(a_off,c1,c2,c3) SUBTRACT_m2n8(a_off,c1,c2) "vfnmadd231ps %%ymm0,%%ymm3,%%ymm"#c3";"
+
+#define save_b_m2n4(c1,tmp,b_off,...)\
+  "vpermilps $216,%%ymm"#c1",%%ymm"#tmp"; vpermpd $216,%%ymm"#tmp",%%ymm"#tmp"; vmovups %%ymm"#tmp","#b_off"("#__VA_ARGS__");"
+
+#define SAVE_b_m2n4(b_off,c1) save_b_m2n4(c1,1,b_off,%1)
+#define SAVE_b_m2n8(b_off,c1,c2) SAVE_b_m2n4(b_off,c1) save_b_m2n4(c2,2,b_off,%1,%%r12,4)
+#define SAVE_b_m2n12(b_off,c1,c2,c3) SAVE_b_m2n8(b_off,c1,c2) save_b_m2n4(c3,3,b_off,%1,%%r12,8)
+
+#define SAVE_b_m1n4(b_off,c1) "vmovups %%xmm"#c1","#b_off"(%1);"
+#define SAVE_b_m1n8(b_off,c1,c2) SAVE_b_m1n4(b_off,c1) "vmovups %%xmm"#c2","#b_off"(%1,%%r12,4);"
+#define SAVE_b_m1n12(b_off,c1,c2,c3) SAVE_b_m1n8(b_off,c1,c2) "vmovups %%xmm"#c3","#b_off"(%1,%%r12,8);"
+

kernel/x86_64/strsm_kernel_8x4_haswell_RN.c

@@ -0,0 +1,279 @@
+#include "common.h"
+#include <stdint.h>
+#include "strsm_kernel_8x4_haswell_R_common.h"
+
+#define SOLVE_RN_m8n4 \
+  "movq %2,%3;" GEMM_SUM_REORDER_8x4(4,5,6,7,63) "movq %2,%3; addq $32,%2;"\
+  SOLVE_leri_m8n2(0,4,5,%1) SUBTRACT_m8n2(8,6,7,%1)\
+  SOLVE_ri_m8n2(16,4,5,%1) SUBTRACT_m8n2(24,6,7,%1)\
+  SAVE_SOLUTION_m8n2(4,5,0)\
+  SOLVE_leri_m8n2(40,6,7,%1)\
+  SOLVE_ri_m8n2(56,6,7,%1)\
+  SAVE_SOLUTION_m8n2(6,7,64)
+
+#define SOLVE_RN_m8n8 \
+  "movq %2,%3;" GEMM_SUM_REORDER_8x4(4,5,6,7,63) GEMM_SUM_REORDER_8x4(8,9,10,11,63) "movq %2,%3; addq $32,%2;"\
+  SOLVE_leri_m8n2(0,4,5,%1) SUBTRACT_m8n2(8,6,7,%1) SUBTRACT_m8n2(0,8,9,%1,%%r12,4) SUBTRACT_m8n2(8,10,11,%1,%%r12,4)\
+  SOLVE_ri_m8n2(16,4,5,%1) SUBTRACT_m8n2(24,6,7,%1) SUBTRACT_m8n2(16,8,9,%1,%%r12,4) SUBTRACT_m8n2(24,10,11,%1,%%r12,4)\
+  SAVE_SOLUTION_m8n2(4,5,0)\
+  SOLVE_leri_m8n2(40,6,7,%1) SUBTRACT_m8n2(32,8,9,%1,%%r12,4) SUBTRACT_m8n2(40,10,11,%1,%%r12,4)\
+  SOLVE_ri_m8n2(56,6,7,%1) SUBTRACT_m8n2(48,8,9,%1,%%r12,4) SUBTRACT_m8n2(56,10,11,%1,%%r12,4)\
+  SAVE_SOLUTION_m8n2(6,7,64)\
+  SOLVE_leri_m8n2(64,8,9,%1,%%r12,4) SUBTRACT_m8n2(72,10,11,%1,%%r12,4)\
+  SOLVE_ri_m8n2(80,8,9,%1,%%r12,4) SUBTRACT_m8n2(88,10,11,%1,%%r12,4)\
+  SAVE_SOLUTION_m8n2(8,9,128)\
+  SOLVE_leri_m8n2(104,10,11,%1,%%r12,4)\
+  SOLVE_ri_m8n2(120,10,11,%1,%%r12,4)\
+  SAVE_SOLUTION_m8n2(10,11,192)
+
+#define SOLVE_RN_m8n12 \
+  "movq %2,%3;" GEMM_SUM_REORDER_8x4(4,5,6,7,63) GEMM_SUM_REORDER_8x4(8,9,10,11,63) GEMM_SUM_REORDER_8x4(12,13,14,15,63) "movq %2,%3; addq $32,%2;"\
+  SOLVE_leri_m8n2(0,4,5,%1) SUBTRACT_m8n2(8,6,7,%1) SUBTRACT_m8n2(0,8,9,%1,%%r12,4) SUBTRACT_m8n2(8,10,11,%1,%%r12,4) SUBTRACT_m8n2(0,12,13,%1,%%r12,8) SUBTRACT_m8n2(8,14,15,%1,%%r12,8)\
+  SOLVE_ri_m8n2(16,4,5,%1) SUBTRACT_m8n2(24,6,7,%1) SUBTRACT_m8n2(16,8,9,%1,%%r12,4) SUBTRACT_m8n2(24,10,11,%1,%%r12,4) SUBTRACT_m8n2(16,12,13,%1,%%r12,8) SUBTRACT_m8n2(24,14,15,%1,%%r12,8)\
+  SAVE_SOLUTION_m8n2(4,5,0)\
+  SOLVE_leri_m8n2(40,6,7,%1) SUBTRACT_m8n2(32,8,9,%1,%%r12,4) SUBTRACT_m8n2(40,10,11,%1,%%r12,4) SUBTRACT_m8n2(32,12,13,%1,%%r12,8) SUBTRACT_m8n2(40,14,15,%1,%%r12,8)\
+  SOLVE_ri_m8n2(56,6,7,%1) SUBTRACT_m8n2(48,8,9,%1,%%r12,4) SUBTRACT_m8n2(56,10,11,%1,%%r12,4) SUBTRACT_m8n2(48,12,13,%1,%%r12,8) SUBTRACT_m8n2(56,14,15,%1,%%r12,8)\
+  SAVE_SOLUTION_m8n2(6,7,64)\
+  SOLVE_leri_m8n2(64,8,9,%1,%%r12,4) SUBTRACT_m8n2(72,10,11,%1,%%r12,4) SUBTRACT_m8n2(64,12,13,%1,%%r12,8) SUBTRACT_m8n2(72,14,15,%1,%%r12,8)\
+  SOLVE_ri_m8n2(80,8,9,%1,%%r12,4) SUBTRACT_m8n2(88,10,11,%1,%%r12,4) SUBTRACT_m8n2(80,12,13,%1,%%r12,8) SUBTRACT_m8n2(88,14,15,%1,%%r12,8)\
+  SAVE_SOLUTION_m8n2(8,9,128)\
+  SOLVE_leri_m8n2(104,10,11,%1,%%r12,4) SUBTRACT_m8n2(96,12,13,%1,%%r12,8) SUBTRACT_m8n2(104,14,15,%1,%%r12,8)\
+  SOLVE_ri_m8n2(120,10,11,%1,%%r12,4) SUBTRACT_m8n2(112,12,13,%1,%%r12,8) SUBTRACT_m8n2(120,14,15,%1,%%r12,8)\
+  SAVE_SOLUTION_m8n2(10,11,192)\
+  SOLVE_leri_m8n2(128,12,13,%1,%%r12,8) SUBTRACT_m8n2(136,14,15,%1,%%r12,8)\
+  SOLVE_ri_m8n2(144,12,13,%1,%%r12,8) SUBTRACT_m8n2(152,14,15,%1,%%r12,8)\
+  SAVE_SOLUTION_m8n2(12,13,256)\
+  SOLVE_leri_m8n2(168,14,15,%1,%%r12,8)\
+  SOLVE_ri_m8n2(184,14,15,%1,%%r12,8)\
+  SAVE_SOLUTION_m8n2(14,15,320)
+
+#define SOLVE_RN_m4n4 \
+  "movq %2,%3;" GEMM_SUM_REORDER_4x4(4,5,6,7,4,5) "movq %2,%3; addq $16,%2;"\
+  SOLVE_leri_m4n2(0,4,%1) SUBTRACT_m4n2(8,5,%1)\
+  SOLVE_ri_m4n2(16,4,%1) SUBTRACT_m4n2(24,5,%1)\
+  SAVE_SOLUTION_m4n2(4,0)\
+  SOLVE_leri_m4n2(40,5,%1)\
+  SOLVE_ri_m4n2(56,5,%1)\
+  SAVE_SOLUTION_m4n2(5,32)
+
+#define SOLVE_RN_m4n8 \
+  "movq %2,%3;" GEMM_SUM_REORDER_4x4(4,5,6,7,4,5) GEMM_SUM_REORDER_4x4(8,9,10,11,6,7) "movq %2,%3; addq $16,%2;"\
+  SOLVE_leri_m4n2(0,4,%1) SUBTRACT_m4n2(8,5,%1) SUBTRACT_m4n2(0,6,%1,%%r12,4) SUBTRACT_m4n2(8,7,%1,%%r12,4)\
+  SOLVE_ri_m4n2(16,4,%1) SUBTRACT_m4n2(24,5,%1) SUBTRACT_m4n2(16,6,%1,%%r12,4) SUBTRACT_m4n2(24,7,%1,%%r12,4)\
+  SAVE_SOLUTION_m4n2(4,0)\
+  SOLVE_leri_m4n2(40,5,%1) SUBTRACT_m4n2(32,6,%1,%%r12,4) SUBTRACT_m4n2(40,7,%1,%%r12,4)\
+  SOLVE_ri_m4n2(56,5,%1) SUBTRACT_m4n2(48,6,%1,%%r12,4) SUBTRACT_m4n2(56,7,%1,%%r12,4)\
+  SAVE_SOLUTION_m4n2(5,32)\
+  SOLVE_leri_m4n2(64,6,%1,%%r12,4) SUBTRACT_m4n2(72,7,%1,%%r12,4)\
+  SOLVE_ri_m4n2(80,6,%1,%%r12,4) SUBTRACT_m4n2(88,7,%1,%%r12,4)\
+  SAVE_SOLUTION_m4n2(6,64)\
+  SOLVE_leri_m4n2(104,7,%1,%%r12,4)\
+  SOLVE_ri_m4n2(120,7,%1,%%r12,4)\
+  SAVE_SOLUTION_m4n2(7,96)
+
+#define SOLVE_RN_m4n12 \
+  "movq %2,%3;" GEMM_SUM_REORDER_4x4(4,5,6,7,4,5) GEMM_SUM_REORDER_4x4(8,9,10,11,6,7) GEMM_SUM_REORDER_4x4(12,13,14,15,8,9) "movq %2,%3; addq $16,%2;"\
+  SOLVE_leri_m4n2(0,4,%1) SUBTRACT_m4n2(8,5,%1) SUBTRACT_m4n2(0,6,%1,%%r12,4) SUBTRACT_m4n2(8,7,%1,%%r12,4) SUBTRACT_m4n2(0,8,%1,%%r12,8) SUBTRACT_m4n2(8,9,%1,%%r12,8)\
+  SOLVE_ri_m4n2(16,4,%1) SUBTRACT_m4n2(24,5,%1) SUBTRACT_m4n2(16,6,%1,%%r12,4) SUBTRACT_m4n2(24,7,%1,%%r12,4) SUBTRACT_m4n2(16,8,%1,%%r12,8) SUBTRACT_m4n2(24,9,%1,%%r12,8)\
+  SAVE_SOLUTION_m4n2(4,0)\
+  SOLVE_leri_m4n2(40,5,%1) SUBTRACT_m4n2(32,6,%1,%%r12,4) SUBTRACT_m4n2(40,7,%1,%%r12,4) SUBTRACT_m4n2(32,8,%1,%%r12,8) SUBTRACT_m4n2(40,9,%1,%%r12,8)\
+  SOLVE_ri_m4n2(56,5,%1) SUBTRACT_m4n2(48,6,%1,%%r12,4) SUBTRACT_m4n2(56,7,%1,%%r12,4) SUBTRACT_m4n2(48,8,%1,%%r12,8) SUBTRACT_m4n2(56,9,%1,%%r12,8)\
+  SAVE_SOLUTION_m4n2(5,32)\
+  SOLVE_leri_m4n2(64,6,%1,%%r12,4) SUBTRACT_m4n2(72,7,%1,%%r12,4) SUBTRACT_m4n2(64,8,%1,%%r12,8) SUBTRACT_m4n2(72,9,%1,%%r12,8)\
+  SOLVE_ri_m4n2(80,6,%1,%%r12,4) SUBTRACT_m4n2(88,7,%1,%%r12,4) SUBTRACT_m4n2(80,8,%1,%%r12,8) SUBTRACT_m4n2(88,9,%1,%%r12,8)\
+  SAVE_SOLUTION_m4n2(6,64)\
+  SOLVE_leri_m4n2(104,7,%1,%%r12,4) SUBTRACT_m4n2(96,8,%1,%%r12,8) SUBTRACT_m4n2(104,9,%1,%%r12,8)\
+  SOLVE_ri_m4n2(120,7,%1,%%r12,4) SUBTRACT_m4n2(112,8,%1,%%r12,8) SUBTRACT_m4n2(120,9,%1,%%r12,8)\
+  SAVE_SOLUTION_m4n2(7,96)\
+  SOLVE_leri_m4n2(128,8,%1,%%r12,8) SUBTRACT_m4n2(136,9,%1,%%r12,8)\
+  SOLVE_ri_m4n2(144,8,%1,%%r12,8) SUBTRACT_m4n2(152,9,%1,%%r12,8)\
+  SAVE_SOLUTION_m4n2(8,128)\
+  SOLVE_leri_m4n2(168,9,%1,%%r12,8)\
+  SOLVE_ri_m4n2(184,9,%1,%%r12,8)\
+  SAVE_SOLUTION_m4n2(9,160)
+
+#define SOLVE_RN_m2n4 \
+  "movq %2,%3;" GEMM_SUM_REORDER_2x4(4,5) "movq %2,%3; addq $8,%2;"\
+  SOLVE_col1_ltor_m2n4(0,4,5,%1)\
+  SOLVE_col2_ltor_m2n4(16,4,5,%1)\
+  SOLVE_col3_ltor_m2n4(32,4,5,%1)\
+  SOLVE_col4_ltor_m2n4(48,4,5,%1)\
+  SAVE_SOLUTION_m2n4(4,5,0)
+
+#define SOLVE_RN_m2n8 \
+  "movq %2,%3;" GEMM_SUM_REORDER_2x4(4,5) GEMM_SUM_REORDER_2x4(6,7) "movq %2,%3; addq $8,%2;"\
+  SOLVE_col1_ltor_m2n4(0,4,5,%1) SUBTRACT_m2n4(0,6,7,%1,%%r12,4)\
+  SOLVE_col2_ltor_m2n4(16,4,5,%1) SUBTRACT_m2n4(16,6,7,%1,%%r12,4)\
+  SOLVE_col3_ltor_m2n4(32,4,5,%1) SUBTRACT_m2n4(32,6,7,%1,%%r12,4)\
+  SOLVE_col4_ltor_m2n4(48,4,5,%1) SUBTRACT_m2n4(48,6,7,%1,%%r12,4)\
+  SAVE_SOLUTION_m2n4(4,5,0)\
+  SOLVE_col1_ltor_m2n4(64,6,7,%1,%%r12,4)\
+  SOLVE_col2_ltor_m2n4(80,6,7,%1,%%r12,4)\
+  SOLVE_col3_ltor_m2n4(96,6,7,%1,%%r12,4)\
+  SOLVE_col4_ltor_m2n4(112,6,7,%1,%%r12,4)\
+  SAVE_SOLUTION_m2n4(6,7,32)
+
+#define SOLVE_RN_m2n12 \
+  "movq %2,%3;" GEMM_SUM_REORDER_2x4(4,5) GEMM_SUM_REORDER_2x4(6,7) GEMM_SUM_REORDER_2x4(8,9) "movq %2,%3; addq $8,%2;"\
+  SOLVE_col1_ltor_m2n4(0,4,5,%1) SUBTRACT_m2n4(0,6,7,%1,%%r12,4) SUBTRACT_m2n4(0,8,9,%1,%%r12,8)\
+  SOLVE_col2_ltor_m2n4(16,4,5,%1) SUBTRACT_m2n4(16,6,7,%1,%%r12,4) SUBTRACT_m2n4(16,8,9,%1,%%r12,8)\
+  SOLVE_col3_ltor_m2n4(32,4,5,%1) SUBTRACT_m2n4(32,6,7,%1,%%r12,4) SUBTRACT_m2n4(32,8,9,%1,%%r12,8)\
+  SOLVE_col4_ltor_m2n4(48,4,5,%1) SUBTRACT_m2n4(48,6,7,%1,%%r12,4) SUBTRACT_m2n4(48,8,9,%1,%%r12,8)\
+  SAVE_SOLUTION_m2n4(4,5,0)\
+  SOLVE_col1_ltor_m2n4(64,6,7,%1,%%r12,4) SUBTRACT_m2n4(64,8,9,%1,%%r12,8)\
+  SOLVE_col2_ltor_m2n4(80,6,7,%1,%%r12,4) SUBTRACT_m2n4(80,8,9,%1,%%r12,8)\
+  SOLVE_col3_ltor_m2n4(96,6,7,%1,%%r12,4) SUBTRACT_m2n4(96,8,9,%1,%%r12,8)\
+  SOLVE_col4_ltor_m2n4(112,6,7,%1,%%r12,4) SUBTRACT_m2n4(112,8,9,%1,%%r12,8)\
+  SAVE_SOLUTION_m2n4(6,7,32)\
+  SOLVE_col1_ltor_m2n4(128,8,9,%1,%%r12,8)\
+  SOLVE_col2_ltor_m2n4(144,8,9,%1,%%r12,8)\
+  SOLVE_col3_ltor_m2n4(160,8,9,%1,%%r12,8)\
+  SOLVE_col4_ltor_m2n4(176,8,9,%1,%%r12,8)\
+  SAVE_SOLUTION_m2n4(8,9,64)
+
+#define SOLVE_RN_m1n4 \
+  "movq %2,%3;" GEMM_SUM_REORDER_1x4(4) "movq %2,%3; addq $4,%2;"\
+  SOLVE_col1_ltor_m1n4(0,4,%1)\
+  SOLVE_col2_ltor_m1n4(16,4,%1)\
+  SOLVE_col3_ltor_m1n4(32,4,%1)\
+  SOLVE_col4_ltor_m1n4(48,4,%1)\
+  SAVE_SOLUTION_m1n4(4,0)
+
+#define SOLVE_RN_m1n8 \
+  "movq %2,%3;" GEMM_SUM_REORDER_1x4(4) GEMM_SUM_REORDER_1x4(5) "movq %2,%3; addq $4,%2;"\
+  SOLVE_col1_ltor_m1n4(0,4,%1) SUBTRACT_m1n4(0,5,%1,%%r12,4)\
+  SOLVE_col2_ltor_m1n4(16,4,%1) SUBTRACT_m1n4(16,5,%1,%%r12,4)\
+  SOLVE_col3_ltor_m1n4(32,4,%1) SUBTRACT_m1n4(32,5,%1,%%r12,4)\
+  SOLVE_col4_ltor_m1n4(48,4,%1) SUBTRACT_m1n4(48,5,%1,%%r12,4)\
+  SAVE_SOLUTION_m1n4(4,0)\
+  SOLVE_col1_ltor_m1n4(64,5,%1,%%r12,4)\
+  SOLVE_col2_ltor_m1n4(80,5,%1,%%r12,4)\
+  SOLVE_col3_ltor_m1n4(96,5,%1,%%r12,4)\
+  SOLVE_col4_ltor_m1n4(112,5,%1,%%r12,4)\
+  SAVE_SOLUTION_m1n4(5,16)
+
+#define SOLVE_RN_m1n12 \
+  "movq %2,%3;" GEMM_SUM_REORDER_1x4(4) GEMM_SUM_REORDER_1x4(5) GEMM_SUM_REORDER_1x4(6) "movq %2,%3; addq $4,%2;"\
+  SOLVE_col1_ltor_m1n4(0,4,%1) SUBTRACT_m1n4(0,5,%1,%%r12,4) SUBTRACT_m1n4(0,6,%1,%%r12,8)\
+  SOLVE_col2_ltor_m1n4(16,4,%1) SUBTRACT_m1n4(16,5,%1,%%r12,4) SUBTRACT_m1n4(16,6,%1,%%r12,8)\
+  SOLVE_col3_ltor_m1n4(32,4,%1) SUBTRACT_m1n4(32,5,%1,%%r12,4) SUBTRACT_m1n4(32,6,%1,%%r12,8)\
+  SOLVE_col4_ltor_m1n4(48,4,%1) SUBTRACT_m1n4(48,5,%1,%%r12,4) SUBTRACT_m1n4(48,6,%1,%%r12,8)\
+  SAVE_SOLUTION_m1n4(4,0)\
+  SOLVE_col1_ltor_m1n4(64,5,%1,%%r12,4) SUBTRACT_m1n4(64,6,%1,%%r12,8)\
+  SOLVE_col2_ltor_m1n4(80,5,%1,%%r12,4) SUBTRACT_m1n4(80,6,%1,%%r12,8)\
+  SOLVE_col3_ltor_m1n4(96,5,%1,%%r12,4) SUBTRACT_m1n4(96,6,%1,%%r12,8)\
+  SOLVE_col4_ltor_m1n4(112,5,%1,%%r12,4) SUBTRACT_m1n4(112,6,%1,%%r12,8)\
+  SAVE_SOLUTION_m1n4(5,16)\
+  SOLVE_col1_ltor_m1n4(128,6,%1,%%r12,8)\
+  SOLVE_col2_ltor_m1n4(144,6,%1,%%r12,8)\
+  SOLVE_col3_ltor_m1n4(160,6,%1,%%r12,8)\
+  SOLVE_col4_ltor_m1n4(176,6,%1,%%r12,8)\
+  SAVE_SOLUTION_m1n4(6,32)
+
+#define GEMM_RN_SIMPLE(mdim,ndim) \
+  "movq %%r15,%0; leaq (%%r15,%%r12,"#mdim"),%%r15; movq %%r13,%5; movq %%r14,%1;" INIT_m##mdim##n##ndim\
+  "testq %5,%5; jz 1"#mdim""#ndim"2f;"\
+  "1"#mdim""#ndim"1:\n\t"\
+  GEMM_KERNEL_k1m##mdim##n##ndim "addq $16,%1; addq $"#mdim"*4,%0; decq %5; jnz 1"#mdim""#ndim"1b;"\
+  "1"#mdim""#ndim"2:\n\t"
+#define GEMM_RN_m8n4 GEMM_RN_SIMPLE(8,4)
+#define GEMM_RN_m8n8 GEMM_RN_SIMPLE(8,8)
+#define GEMM_RN_m8n12 \
+  "movq %%r15,%0; leaq (%%r15,%%r12,8),%%r15; movq %%r13,%5; movq %%r14,%1;" INIT_m8n12\
+  "cmpq $8,%5; jb 18122f;"\
+  "18121:\n\t"\
+  GEMM_KERNEL_k1m8n12 "prefetcht0 384(%0); addq $32,%0; addq $16,%1;"\
+  GEMM_KERNEL_k1m8n12 "addq $32,%0; addq $16,%1;"\
+  GEMM_KERNEL_k1m8n12 "prefetcht0 384(%0); addq $32,%0; addq $16,%1;"\
+  GEMM_KERNEL_k1m8n12 "addq $32,%0; addq $16,%1;"\
+  GEMM_KERNEL_k1m8n12 "prefetcht0 384(%0); addq $32,%0; addq $16,%1;"\
+  GEMM_KERNEL_k1m8n12 "addq $32,%0; addq $16,%1;"\
+  GEMM_KERNEL_k1m8n12 "prefetcht0 384(%0); addq $32,%0; addq $16,%1;"\
+  GEMM_KERNEL_k1m8n12 "addq $32,%0; addq $16,%1;"\
+  "subq $8,%5; cmpq $8,%5; jnb 18121b;"\
+  "18122:\n\t"\
+  "testq %5,%5; jz 18124f;"\
+  "18123:\n\t"\
+  GEMM_KERNEL_k1m8n12 "addq $32,%0; addq $16,%1; decq %5; jnz 18123b;"\
+  "18124:\n\t"
+#define GEMM_RN_m4n4 GEMM_RN_SIMPLE(4,4)
+#define GEMM_RN_m4n8 GEMM_RN_SIMPLE(4,8)
+#define GEMM_RN_m4n12 GEMM_RN_SIMPLE(4,12)
+#define GEMM_RN_m2n4 GEMM_RN_SIMPLE(2,4)
+#define GEMM_RN_m2n8 GEMM_RN_SIMPLE(2,8)
+#define GEMM_RN_m2n12 GEMM_RN_SIMPLE(2,12)
+#define GEMM_RN_m1n4 GEMM_RN_SIMPLE(1,4)
+#define GEMM_RN_m1n8 GEMM_RN_SIMPLE(1,8)
+#define GEMM_RN_m1n12 GEMM_RN_SIMPLE(1,12)
+
+#define COMPUTE(ndim) {\
+  __asm__ __volatile__(\
+    "movq %0,%%r15; movq %1,%%r14; movq %7,%%r13; movq %6,%%r12; salq $2,%%r12; movq %10,%%r11;"\
+    "cmpq $8,%%r11; jb "#ndim"772f;"\
+    #ndim"771:\n\t"\
+    GEMM_RN_m8n##ndim SOLVE_RN_m8n##ndim "subq $8,%%r11; cmpq $8,%%r11; jnb "#ndim"771b;"\
+    #ndim"772:\n\t"\
+    "testq $4,%%r11; jz "#ndim"773f;"\
+    GEMM_RN_m4n##ndim SOLVE_RN_m4n##ndim "subq $4,%%r11;"\
+    #ndim"773:\n\t"\
+    "testq $2,%%r11; jz "#ndim"774f;"\
+    GEMM_RN_m2n##ndim SOLVE_RN_m2n##ndim "subq $2,%%r11;"\
+    #ndim"774:\n\t"\
+    "testq $1,%%r11; jz "#ndim"775f;"\
+    GEMM_RN_m1n##ndim SOLVE_RN_m1n##ndim "subq $1,%%r11;"\
+    #ndim"775:\n\t"\
+    "movq %%r15,%0; movq %%r14,%1; vzeroupper;"\
+  :"+r"(a_ptr),"+r"(b_ptr),"+r"(c_ptr),"+r"(c_tmp),"+r"(ldc_bytes),"+r"(k_cnt):"m"(K),"m"(OFF),"m"(one[0]),"m"(zero[0]),"m"(M)\
+  :"r11","r12","r13","r14","r15","cc","memory",\
+  "xmm0","xmm1","xmm2","xmm3","xmm4","xmm5","xmm6","xmm7","xmm8","xmm9","xmm10","xmm11","xmm12","xmm13","xmm14","xmm15");\
+  a_ptr -= M * K; b_ptr += ndim * K; c_ptr += ldc * ndim - M; OFF += ndim;\
+}
+
+static void solve_RN(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+  FLOAT a0, b0;
+  int i, j, k;
+  for (i=0; i<n; i++) {
+    b0 = b[i*n+i];
+    for (j=0; j<m; j++) {
+      a0 = c[i*ldc+j] * b0;
+      a[i*m+j] = c[i*ldc+j] = a0;
+      for (k=i+1; k<n; k++) c[k*ldc+j] -= a0 * b[i*n+k];
+    }
+  }
+}
+static void COMPUTE_EDGE_1_nchunk(BLASLONG m, BLASLONG n, FLOAT *sa, FLOAT *sb, FLOAT *C, BLASLONG ldc, BLASLONG k, BLASLONG offset) {
+  BLASLONG m_count = m, kk = offset; FLOAT *a_ptr = sa, *c_ptr = C;
+  for(;m_count>7;m_count-=8){
+    if(kk>0) GEMM_KERNEL_N(8,n,kk,-1.0,a_ptr,sb,c_ptr,ldc);
+    solve_RN(8,n,a_ptr+kk*8,sb+kk*n,c_ptr,ldc);
+    a_ptr += k * 8; c_ptr += 8;
+  }
+  for(;m_count>3;m_count-=4){
+    if(kk>0) GEMM_KERNEL_N(4,n,kk,-1.0,a_ptr,sb,c_ptr,ldc);
+    solve_RN(4,n,a_ptr+kk*4,sb+kk*n,c_ptr,ldc);
+    a_ptr += k * 4; c_ptr += 4;
+  }
+  for(;m_count>1;m_count-=2){
+    if(kk>0) GEMM_KERNEL_N(2,n,kk,-1.0,a_ptr,sb,c_ptr,ldc);
+    solve_RN(2,n,a_ptr+kk*2,sb+kk*n,c_ptr,ldc);
+    a_ptr += k * 2; c_ptr += 2;
+  }
+  if(m_count>0){
+    if(kk>0) GEMM_KERNEL_N(1,n,kk,-1.0,a_ptr,sb,c_ptr,ldc);
+    solve_RN(1,n,a_ptr+kk*1,sb+kk*n,c_ptr,ldc);
+    a_ptr += k * 1; c_ptr += 1;
+  }
+}
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT dummy1, FLOAT *sa, FLOAT *sb, FLOAT *C, BLASLONG ldc, BLASLONG offset){
+  float *a_ptr = sa, *b_ptr = sb, *c_ptr = C, *c_tmp = C;
+  float one[8] = {1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0};
+  float zero[8] = {0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0};
+  uint64_t ldc_bytes = (uint64_t)ldc * sizeof(float), K = (uint64_t)k, M = (uint64_t)m, OFF = (uint64_t)-offset, k_cnt = 0;
+  BLASLONG n_count = n;
+  for(;n_count>11;n_count-=12) COMPUTE(12)
+  for(;n_count>7;n_count-=8) COMPUTE(8)
+  for(;n_count>3;n_count-=4) COMPUTE(4)
+  for(;n_count>1;n_count-=2) { COMPUTE_EDGE_1_nchunk(m,2,a_ptr,b_ptr,c_ptr,ldc,k,OFF); b_ptr += 2*k; c_ptr += ldc*2; OFF+=2;}
+  if(n_count>0) COMPUTE_EDGE_1_nchunk(m,1,a_ptr,b_ptr,c_ptr,ldc,k,OFF);
+  return 0;
+}

kernel/x86_64/strsm_kernel_8x4_haswell_RT.c

@@ -0,0 +1,281 @@
+#include "common.h"
+#include <stdint.h>
+#include "strsm_kernel_8x4_haswell_R_common.h"
+
+#define SOLVE_RT_m8n4 \
+  "movq %2,%3;" GEMM_SUM_REORDER_8x4(4,5,6,7,63) "negq %4; leaq (%3,%4,2),%3; negq %4; addq $32,%2;"\
+  SOLVE_rile_m8n2(-8,6,7,%1) SUBTRACT_m8n2(-16,4,5,%1)\
+  SOLVE_le_m8n2(-24,6,7,%1) SUBTRACT_m8n2(-32,4,5,%1)\
+  SAVE_SOLUTION_m8n2(6,7,-64) "negq %4; leaq (%3,%4,4),%3; negq %4;"\
+  SOLVE_rile_m8n2(-48,4,5,%1)\
+  SOLVE_le_m8n2(-64,4,5,%1)\
+  SAVE_SOLUTION_m8n2(4,5,-128)
+
+#define SOLVE_RT_m8n8 \
+  "movq %2,%3;" GEMM_SUM_REORDER_8x4(4,5,6,7,63) GEMM_SUM_REORDER_8x4(8,9,10,11,63) "negq %4; leaq (%3,%4,2),%3; negq %4; addq $32,%2;"\
+  SOLVE_rile_m8n2(-8,10,11,%1,%%r12,4) SUBTRACT_m8n2(-16,8,9,%1,%%r12,4) SUBTRACT_m8n2(-8,6,7,%1) SUBTRACT_m8n2(-16,4,5,%1)\
+  SOLVE_le_m8n2(-24,10,11,%1,%%r12,4) SUBTRACT_m8n2(-32,8,9,%1,%%r12,4) SUBTRACT_m8n2(-24,6,7,%1) SUBTRACT_m8n2(-32,4,5,%1)\
+  SAVE_SOLUTION_m8n2(10,11,-64) "negq %4; leaq (%3,%4,4),%3; negq %4;"\
+  SOLVE_rile_m8n2(-48,8,9,%1,%%r12,4) SUBTRACT_m8n2(-40,6,7,%1) SUBTRACT_m8n2(-48,4,5,%1)\
+  SOLVE_le_m8n2(-64,8,9,%1,%%r12,4) SUBTRACT_m8n2(-56,6,7,%1) SUBTRACT_m8n2(-64,4,5,%1)\
+  SAVE_SOLUTION_m8n2(8,9,-128) "negq %4; leaq (%3,%4,4),%3; negq %4;"\
+  SOLVE_rile_m8n2(-72,6,7,%1) SUBTRACT_m8n2(-80,4,5,%1)\
+  SOLVE_le_m8n2(-88,6,7,%1) SUBTRACT_m8n2(-96,4,5,%1)\
+  SAVE_SOLUTION_m8n2(6,7,-192) "negq %4; leaq (%3,%4,4),%3; negq %4;"\
+  SOLVE_rile_m8n2(-112,4,5,%1)\
+  SOLVE_le_m8n2(-128,4,5,%1)\
+  SAVE_SOLUTION_m8n2(4,5,-256)
+
+#define SOLVE_RT_m8n12 \
+  "movq %2,%3;" GEMM_SUM_REORDER_8x4(4,5,6,7,63) GEMM_SUM_REORDER_8x4(8,9,10,11,63) GEMM_SUM_REORDER_8x4(12,13,14,15,63) "negq %4; leaq (%3,%4,2),%3; negq %4; addq $32,%2;"\
+  SOLVE_rile_m8n2(-8,14,15,%1,%%r12,8) SUBTRACT_m8n2(-16,12,13,%1,%%r12,8) SUBTRACT_m8n2(-8,10,11,%1,%%r12,4) SUBTRACT_m8n2(-16,8,9,%1,%%r12,4) SUBTRACT_m8n2(-8,6,7,%1) SUBTRACT_m8n2(-16,4,5,%1)\
+  SOLVE_le_m8n2(-24,14,15,%1,%%r12,8) SUBTRACT_m8n2(-32,12,13,%1,%%r12,8) SUBTRACT_m8n2(-24,10,11,%1,%%r12,4) SUBTRACT_m8n2(-32,8,9,%1,%%r12,4) SUBTRACT_m8n2(-24,6,7,%1) SUBTRACT_m8n2(-32,4,5,%1)\
+  SAVE_SOLUTION_m8n2(14,15,-64) "negq %4; leaq (%3,%4,4),%3; negq %4;"\
+  SOLVE_rile_m8n2(-48,12,13,%1,%%r12,8) SUBTRACT_m8n2(-40,10,11,%1,%%r12,4) SUBTRACT_m8n2(-48,8,9,%1,%%r12,4) SUBTRACT_m8n2(-40,6,7,%1) SUBTRACT_m8n2(-48,4,5,%1)\
+  SOLVE_le_m8n2(-64,12,13,%1,%%r12,8) SUBTRACT_m8n2(-56,10,11,%1,%%r12,4) SUBTRACT_m8n2(-64,8,9,%1,%%r12,4) SUBTRACT_m8n2(-56,6,7,%1) SUBTRACT_m8n2(-64,4,5,%1)\
+  SAVE_SOLUTION_m8n2(12,13,-128) "negq %4; leaq (%3,%4,4),%3; negq %4;"\
+  SOLVE_rile_m8n2(-72,10,11,%1,%%r12,4) SUBTRACT_m8n2(-80,8,9,%1,%%r12,4) SUBTRACT_m8n2(-72,6,7,%1) SUBTRACT_m8n2(-80,4,5,%1)\
+  SOLVE_le_m8n2(-88,10,11,%1,%%r12,4) SUBTRACT_m8n2(-96,8,9,%1,%%r12,4) SUBTRACT_m8n2(-88,6,7,%1) SUBTRACT_m8n2(-96,4,5,%1)\
+  SAVE_SOLUTION_m8n2(10,11,-192) "negq %4; leaq (%3,%4,4),%3; negq %4;"\
+  SOLVE_rile_m8n2(-112,8,9,%1,%%r12,4) SUBTRACT_m8n2(-104,6,7,%1) SUBTRACT_m8n2(-112,4,5,%1)\
+  SOLVE_le_m8n2(-128,8,9,%1,%%r12,4) SUBTRACT_m8n2(-120,6,7,%1) SUBTRACT_m8n2(-128,4,5,%1)\
+  SAVE_SOLUTION_m8n2(8,9,-256) "negq %4; leaq (%3,%4,4),%3; negq %4;"\
+  SOLVE_rile_m8n2(-136,6,7,%1) SUBTRACT_m8n2(-144,4,5,%1)\
+  SOLVE_le_m8n2(-152,6,7,%1) SUBTRACT_m8n2(-160,4,5,%1)\
+  SAVE_SOLUTION_m8n2(6,7,-320) "negq %4; leaq (%3,%4,4),%3; negq %4;"\
+  SOLVE_rile_m8n2(-176,4,5,%1)\
+  SOLVE_le_m8n2(-192,4,5,%1)\
+  SAVE_SOLUTION_m8n2(4,5,-384)
+
+#define SOLVE_RT_m4n4 \
+  "movq %2,%3;" GEMM_SUM_REORDER_4x4(4,5,6,7,4,5) "negq %4; leaq (%3,%4,2),%3; negq %4; addq $16,%2;"\
+  SOLVE_rile_m4n2(-8,5,%1) SUBTRACT_m4n2(-16,4,%1)\
+  SOLVE_le_m4n2(-24,5,%1) SUBTRACT_m4n2(-32,4,%1)\
+  SAVE_SOLUTION_m4n2(5,-32) "negq %4; leaq (%3,%4,4),%3; negq %4;"\
+  SOLVE_rile_m4n2(-48,4,%1)\
+  SOLVE_le_m4n2(-64,4,%1)\
+  SAVE_SOLUTION_m4n2(4,-64)
+
+#define SOLVE_RT_m4n8 \
+  "movq %2,%3;" GEMM_SUM_REORDER_4x4(4,5,6,7,4,5) GEMM_SUM_REORDER_4x4(8,9,10,11,6,7) "negq %4; leaq (%3,%4,2),%3; negq %4; addq $16,%2;"\
+  SOLVE_rile_m4n2(-8,7,%1,%%r12,4) SUBTRACT_m4n2(-16,6,%1,%%r12,4) SUBTRACT_m4n2(-8,5,%1) SUBTRACT_m4n2(-16,4,%1)\
+  SOLVE_le_m4n2(-24,7,%1,%%r12,4) SUBTRACT_m4n2(-32,6,%1,%%r12,4) SUBTRACT_m4n2(-24,5,%1) SUBTRACT_m4n2(-32,4,%1)\
+  SAVE_SOLUTION_m4n2(7,-32) "negq %4; leaq (%3,%4,4),%3; negq %4;"\
+  SOLVE_rile_m4n2(-48,6,%1,%%r12,4) SUBTRACT_m4n2(-40,5,%1) SUBTRACT_m4n2(-48,4,%1)\
+  SOLVE_le_m4n2(-64,6,%1,%%r12,4) SUBTRACT_m4n2(-56,5,%1) SUBTRACT_m4n2(-64,4,%1)\
+  SAVE_SOLUTION_m4n2(6,-64) "negq %4; leaq (%3,%4,4),%3; negq %4;"\
+  SOLVE_rile_m4n2(-72,5,%1) SUBTRACT_m4n2(-80,4,%1)\
+  SOLVE_le_m4n2(-88,5,%1) SUBTRACT_m4n2(-96,4,%1)\
+  SAVE_SOLUTION_m4n2(5,-96) "negq %4; leaq (%3,%4,4),%3; negq %4;"\
+  SOLVE_rile_m4n2(-112,4,%1)\
+  SOLVE_le_m4n2(-128,4,%1)\
+  SAVE_SOLUTION_m4n2(4,-128)
+
+#define SOLVE_RT_m4n12 \
+  "movq %2,%3;" GEMM_SUM_REORDER_4x4(4,5,6,7,4,5) GEMM_SUM_REORDER_4x4(8,9,10,11,6,7) GEMM_SUM_REORDER_4x4(12,13,14,15,8,9) "negq %4; leaq (%3,%4,2),%3; negq %4; addq $16,%2;"\
+  SOLVE_rile_m4n2(-8,9,%1,%%r12,8) SUBTRACT_m4n2(-16,8,%1,%%r12,8) SUBTRACT_m4n2(-8,7,%1,%%r12,4) SUBTRACT_m4n2(-16,6,%1,%%r12,4) SUBTRACT_m4n2(-8,5,%1) SUBTRACT_m4n2(-16,4,%1)\
+  SOLVE_le_m4n2(-24,9,%1,%%r12,8) SUBTRACT_m4n2(-32,8,%1,%%r12,8) SUBTRACT_m4n2(-24,7,%1,%%r12,4) SUBTRACT_m4n2(-32,6,%1,%%r12,4) SUBTRACT_m4n2(-24,5,%1) SUBTRACT_m4n2(-32,4,%1)\
+  SAVE_SOLUTION_m4n2(9,-32) "negq %4; leaq (%3,%4,4),%3; negq %4;"\
+  SOLVE_rile_m4n2(-48,8,%1,%%r12,8) SUBTRACT_m4n2(-40,7,%1,%%r12,4) SUBTRACT_m4n2(-48,6,%1,%%r12,4) SUBTRACT_m4n2(-40,5,%1) SUBTRACT_m4n2(-48,4,%1)\
+  SOLVE_le_m4n2(-64,8,%1,%%r12,8) SUBTRACT_m4n2(-56,7,%1,%%r12,4) SUBTRACT_m4n2(-64,6,%1,%%r12,4) SUBTRACT_m4n2(-56,5,%1) SUBTRACT_m4n2(-64,4,%1)\
+  SAVE_SOLUTION_m4n2(8,-64) "negq %4; leaq (%3,%4,4),%3; negq %4;"\
+  SOLVE_rile_m4n2(-72,7,%1,%%r12,4) SUBTRACT_m4n2(-80,6,%1,%%r12,4) SUBTRACT_m4n2(-72,5,%1) SUBTRACT_m4n2(-80,4,%1)\
+  SOLVE_le_m4n2(-88,7,%1,%%r12,4) SUBTRACT_m4n2(-96,6,%1,%%r12,4) SUBTRACT_m4n2(-88,5,%1) SUBTRACT_m4n2(-96,4,%1)\
+  SAVE_SOLUTION_m4n2(7,-96) "negq %4; leaq (%3,%4,4),%3; negq %4;"\
+  SOLVE_rile_m4n2(-112,6,%1,%%r12,4) SUBTRACT_m4n2(-104,5,%1) SUBTRACT_m4n2(-112,4,%1)\
+  SOLVE_le_m4n2(-128,6,%1,%%r12,4) SUBTRACT_m4n2(-120,5,%1) SUBTRACT_m4n2(-128,4,%1)\
+  SAVE_SOLUTION_m4n2(6,-128) "negq %4; leaq (%3,%4,4),%3; negq %4;"\
+  SOLVE_rile_m4n2(-136,5,%1) SUBTRACT_m4n2(-144,4,%1)\
+  SOLVE_le_m4n2(-152,5,%1) SUBTRACT_m4n2(-160,4,%1)\
+  SAVE_SOLUTION_m4n2(5,-160) "negq %4; leaq (%3,%4,4),%3; negq %4;"\
+  SOLVE_rile_m4n2(-176,4,%1)\
+  SOLVE_le_m4n2(-192,4,%1)\
+  SAVE_SOLUTION_m4n2(4,-192)
+
+#define SOLVE_RT_m2n4 \
+  "movq %2,%3;" GEMM_SUM_REORDER_2x4(4,5) "negq %4; leaq (%3,%4,4),%3; negq %4; addq $8,%2;"\
+  SOLVE_col4_rtol_m2n4(-16,4,5,%1)\
+  SOLVE_col3_rtol_m2n4(-32,4,5,%1)\
+  SOLVE_col2_rtol_m2n4(-48,4,5,%1)\
+  SOLVE_col1_rtol_m2n4(-64,4,5,%1)\
+  SAVE_SOLUTION_m2n4(4,5,-32)
+
+#define SOLVE_RT_m2n8 \
+  "movq %2,%3;" GEMM_SUM_REORDER_2x4(4,5) GEMM_SUM_REORDER_2x4(6,7) "negq %4; leaq (%3,%4,4),%3; negq %4; addq $8,%2;"\
+  SOLVE_col4_rtol_m2n4(-16,6,7,%1,%%r12,4) SUBTRACT_m2n4(-16,4,5,%1)\
+  SOLVE_col3_rtol_m2n4(-32,6,7,%1,%%r12,4) SUBTRACT_m2n4(-32,4,5,%1)\
+  SOLVE_col2_rtol_m2n4(-48,6,7,%1,%%r12,4) SUBTRACT_m2n4(-48,4,5,%1)\
+  SOLVE_col1_rtol_m2n4(-64,6,7,%1,%%r12,4) SUBTRACT_m2n4(-64,4,5,%1)\
+  SAVE_SOLUTION_m2n4(6,7,-32) "negq %4; leaq (%3,%4,8),%3; negq %4;"\
+  SOLVE_col4_rtol_m2n4(-80,4,5,%1)\
+  SOLVE_col3_rtol_m2n4(-96,4,5,%1)\
+  SOLVE_col2_rtol_m2n4(-112,4,5,%1)\
+  SOLVE_col1_rtol_m2n4(-128,4,5,%1)\
+  SAVE_SOLUTION_m2n4(4,5,-64)
+
+#define SOLVE_RT_m2n12 \
+  "movq %2,%3;" GEMM_SUM_REORDER_2x4(4,5) GEMM_SUM_REORDER_2x4(6,7) GEMM_SUM_REORDER_2x4(8,9) "negq %4; leaq (%3,%4,4),%3; negq %4; addq $8,%2;"\
+  SOLVE_col4_rtol_m2n4(-16,8,9,%1,%%r12,8) SUBTRACT_m2n4(-16,6,7,%1,%%r12,4) SUBTRACT_m2n4(-16,4,5,%1)\
+  SOLVE_col3_rtol_m2n4(-32,8,9,%1,%%r12,8) SUBTRACT_m2n4(-32,6,7,%1,%%r12,4) SUBTRACT_m2n4(-32,4,5,%1)\
+  SOLVE_col2_rtol_m2n4(-48,8,9,%1,%%r12,8) SUBTRACT_m2n4(-48,6,7,%1,%%r12,4) SUBTRACT_m2n4(-48,4,5,%1)\
+  SOLVE_col1_rtol_m2n4(-64,8,9,%1,%%r12,8) SUBTRACT_m2n4(-64,6,7,%1,%%r12,4) SUBTRACT_m2n4(-64,4,5,%1)\
+  SAVE_SOLUTION_m2n4(8,9,-32) "negq %4; leaq (%3,%4,8),%3; negq %4;"\
+  SOLVE_col4_rtol_m2n4(-80,6,7,%1,%%r12,4) SUBTRACT_m2n4(-80,4,5,%1)\
+  SOLVE_col3_rtol_m2n4(-96,6,7,%1,%%r12,4) SUBTRACT_m2n4(-96,4,5,%1)\
+  SOLVE_col2_rtol_m2n4(-112,6,7,%1,%%r12,4) SUBTRACT_m2n4(-112,4,5,%1)\
+  SOLVE_col1_rtol_m2n4(-128,6,7,%1,%%r12,4) SUBTRACT_m2n4(-128,4,5,%1)\
+  SAVE_SOLUTION_m2n4(6,7,-64) "negq %4; leaq (%3,%4,8),%3; negq %4;"\
+  SOLVE_col4_rtol_m2n4(-144,4,5,%1)\
+  SOLVE_col3_rtol_m2n4(-160,4,5,%1)\
+  SOLVE_col2_rtol_m2n4(-176,4,5,%1)\
+  SOLVE_col1_rtol_m2n4(-192,4,5,%1)\
+  SAVE_SOLUTION_m2n4(4,5,-96)
+
+#define SOLVE_RT_m1n4 \
+  "movq %2,%3;" GEMM_SUM_REORDER_1x4(4) "negq %4; leaq (%3,%4,4),%3; negq %4; addq $4,%2;"\
+  SOLVE_col4_rtol_m1n4(-16,4,%1)\
+  SOLVE_col3_rtol_m1n4(-32,4,%1)\
+  SOLVE_col2_rtol_m1n4(-48,4,%1)\
+  SOLVE_col1_rtol_m1n4(-64,4,%1)\
+  SAVE_SOLUTION_m1n4(4,-16)
+
+#define SOLVE_RT_m1n8 \
+  "movq %2,%3;" GEMM_SUM_REORDER_1x4(4) GEMM_SUM_REORDER_1x4(5) "negq %4; leaq (%3,%4,4),%3; negq %4; addq $4,%2;"\
+  SOLVE_col4_rtol_m1n4(-16,5,%1,%%r12,4) SUBTRACT_m1n4(-16,4,%1)\
+  SOLVE_col3_rtol_m1n4(-32,5,%1,%%r12,4) SUBTRACT_m1n4(-32,4,%1)\
+  SOLVE_col2_rtol_m1n4(-48,5,%1,%%r12,4) SUBTRACT_m1n4(-48,4,%1)\
+  SOLVE_col1_rtol_m1n4(-64,5,%1,%%r12,4) SUBTRACT_m1n4(-64,4,%1)\
+  SAVE_SOLUTION_m1n4(5,-16) "negq %4; leaq (%3,%4,8),%3; negq %4;"\
+  SOLVE_col4_rtol_m1n4(-80,4,%1)\
+  SOLVE_col3_rtol_m1n4(-96,4,%1)\
+  SOLVE_col2_rtol_m1n4(-112,4,%1)\
+  SOLVE_col1_rtol_m1n4(-128,4,%1)\
+  SAVE_SOLUTION_m1n4(4,-32)
+
+#define SOLVE_RT_m1n12 \
+  "movq %2,%3;" GEMM_SUM_REORDER_1x4(4) GEMM_SUM_REORDER_1x4(5) GEMM_SUM_REORDER_1x4(6) "negq %4; leaq (%3,%4,4),%3; negq %4; addq $4,%2;"\
+  SOLVE_col4_rtol_m1n4(-16,6,%1,%%r12,8) SUBTRACT_m1n4(-16,5,%1,%%r12,4) SUBTRACT_m1n4(-16,4,%1)\
+  SOLVE_col3_rtol_m1n4(-32,6,%1,%%r12,8) SUBTRACT_m1n4(-32,5,%1,%%r12,4) SUBTRACT_m1n4(-32,4,%1)\
+  SOLVE_col2_rtol_m1n4(-48,6,%1,%%r12,8) SUBTRACT_m1n4(-48,5,%1,%%r12,4) SUBTRACT_m1n4(-48,4,%1)\
+  SOLVE_col1_rtol_m1n4(-64,6,%1,%%r12,8) SUBTRACT_m1n4(-64,5,%1,%%r12,4) SUBTRACT_m1n4(-64,4,%1)\
+  SAVE_SOLUTION_m1n4(6,-16) "negq %4; leaq (%3,%4,8),%3; negq %4;"\
+  SOLVE_col4_rtol_m1n4(-80,5,%1,%%r12,4) SUBTRACT_m1n4(-80,4,%1)\
+  SOLVE_col3_rtol_m1n4(-96,5,%1,%%r12,4) SUBTRACT_m1n4(-96,4,%1)\
+  SOLVE_col2_rtol_m1n4(-112,5,%1,%%r12,4) SUBTRACT_m1n4(-112,4,%1)\
+  SOLVE_col1_rtol_m1n4(-128,5,%1,%%r12,4) SUBTRACT_m1n4(-128,4,%1)\
+  SAVE_SOLUTION_m1n4(5,-32) "negq %4; leaq (%3,%4,8),%3; negq %4;"\
+  SOLVE_col4_rtol_m1n4(-144,4,%1)\
+  SOLVE_col3_rtol_m1n4(-160,4,%1)\
+  SOLVE_col2_rtol_m1n4(-176,4,%1)\
+  SOLVE_col1_rtol_m1n4(-192,4,%1)\
+  SAVE_SOLUTION_m1n4(4,-48)
+
+/* r14 = b_tail, r15 = a_tail, r13 = k-kk */
+#define GEMM_RT_SIMPLE(mdim,ndim) \
+  "leaq (%%r15,%%r12,"#mdim"),%%r15; movq %%r15,%0; movq %%r13,%5; movq %%r14,%1;" INIT_m##mdim##n##ndim\
+  "testq %5,%5; jz 1"#mdim""#ndim"2f;"\
+  "1"#mdim""#ndim"1:\n\t"\
+  "subq $16,%1; subq $"#mdim"*4,%0;" GEMM_KERNEL_k1m##mdim##n##ndim "decq %5; jnz 1"#mdim""#ndim"1b;"\
+  "1"#mdim""#ndim"2:\n\t"
+#define GEMM_RT_m8n4 GEMM_RT_SIMPLE(8,4)
+#define GEMM_RT_m8n8 GEMM_RT_SIMPLE(8,8)
+#define GEMM_RT_m8n12 \
+  "leaq (%%r15,%%r12,8),%%r15; movq %%r15,%0; movq %%r13,%5; movq %%r14,%1;" INIT_m8n12\
+  "cmpq $8,%5; jb 18122f;"\
+  "18121:\n\t"\
+  "prefetcht0 -384(%0); subq $32,%0; subq $16,%1;" GEMM_KERNEL_k1m8n12\
+                       "subq $32,%0; subq $16,%1;" GEMM_KERNEL_k1m8n12\
+  "prefetcht0 -384(%0); subq $32,%0; subq $16,%1;" GEMM_KERNEL_k1m8n12\
+                       "subq $32,%0; subq $16,%1;" GEMM_KERNEL_k1m8n12\
+  "prefetcht0 -384(%0); subq $32,%0; subq $16,%1;" GEMM_KERNEL_k1m8n12\
+                       "subq $32,%0; subq $16,%1;" GEMM_KERNEL_k1m8n12\
+  "prefetcht0 -384(%0); subq $32,%0; subq $16,%1;" GEMM_KERNEL_k1m8n12\
+                       "subq $32,%0; subq $16,%1;" GEMM_KERNEL_k1m8n12\
+  "subq $8,%5; cmpq $8,%5; jnb 18121b;"\
+  "18122:\n\t"\
+  "testq %5,%5; jz 18124f;"\
+  "18123:\n\t"\
+  "subq $32,%0; subq $16,%1;" GEMM_KERNEL_k1m8n12 "decq %5; jnz 18123b;"\
+  "18124:\n\t"
+#define GEMM_RT_m4n4 GEMM_RT_SIMPLE(4,4)
+#define GEMM_RT_m4n8 GEMM_RT_SIMPLE(4,8)
+#define GEMM_RT_m4n12 GEMM_RT_SIMPLE(4,12)
+#define GEMM_RT_m2n4 GEMM_RT_SIMPLE(2,4)
+#define GEMM_RT_m2n8 GEMM_RT_SIMPLE(2,8)
+#define GEMM_RT_m2n12 GEMM_RT_SIMPLE(2,12)
+#define GEMM_RT_m1n4 GEMM_RT_SIMPLE(1,4)
+#define GEMM_RT_m1n8 GEMM_RT_SIMPLE(1,8)
+#define GEMM_RT_m1n12 GEMM_RT_SIMPLE(1,12)
+
+#define COMPUTE(ndim) {\
+  b_ptr -= (ndim-4)*K; c_ptr -= ndim * ldc;\
+  __asm__ __volatile__(\
+    "movq %0,%%r15; movq %6,%%r13; subq %7,%%r13; movq %6,%%r12; salq $2,%%r12; movq %1,%%r14; movq %10,%%r11;"\
+    "cmpq $8,%%r11; jb "#ndim"772f;"\
+    #ndim"771:\n\t"\
+    GEMM_RT_m8n##ndim SOLVE_RT_m8n##ndim "subq $8,%%r11; cmpq $8,%%r11; jnb "#ndim"771b;"\
+    #ndim"772:\n\t"\
+    "testq $4,%%r11; jz "#ndim"773f;"\
+    GEMM_RT_m4n##ndim SOLVE_RT_m4n##ndim "subq $4,%%r11;"\
+    #ndim"773:\n\t"\
+    "testq $2,%%r11; jz "#ndim"774f;"\
+    GEMM_RT_m2n##ndim SOLVE_RT_m2n##ndim "subq $2,%%r11;"\
+    #ndim"774:\n\t"\
+    "testq $1,%%r11; jz "#ndim"775f;"\
+    GEMM_RT_m1n##ndim SOLVE_RT_m1n##ndim "subq $1,%%r11;"\
+    #ndim"775:\n\t"\
+    "movq %%r15,%0; movq %%r14,%1; vzeroupper;"\
+  :"+r"(a_ptr),"+r"(b_ptr),"+r"(c_ptr),"+r"(c_tmp),"+r"(ldc_bytes),"+r"(k_cnt):"m"(K),"m"(OFF),"m"(one[0]),"m"(zero[0]),"m"(M)\
+  :"r11","r12","r13","r14","r15","cc","memory",\
+  "xmm0","xmm1","xmm2","xmm3","xmm4","xmm5","xmm6","xmm7","xmm8","xmm9","xmm10","xmm11","xmm12","xmm13","xmm14","xmm15");\
+  a_ptr -= M * K; b_ptr -= 4 * K; c_ptr -= M; OFF -= ndim;\
+}
+
+static void solve_RT(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc){
+  FLOAT a0, b0;
+  int i, j, k;
+  for (i=n-1;i>=0;i--) {
+    b0 = b[i*n+i];
+    for (j=0;j<m;j++) {
+      a0 = c[i*ldc+j] * b0;
+      a[i*m+j] = c[i*ldc+j] = a0;
+      for (k=0;k<i;k++) c[k*ldc+j] -= a0 * b[i*n+k];
+    }
+  }
+}
+static void COMPUTE_EDGE_1_nchunk(BLASLONG m, BLASLONG n, FLOAT *sa, FLOAT *sb, FLOAT *C, BLASLONG ldc, BLASLONG k, BLASLONG offset) {
+  BLASLONG m_count = m, kk = offset; FLOAT *a_ptr = sa, *c_ptr = C;
+  for(;m_count>7;m_count-=8){
+    if(k-kk>0) GEMM_KERNEL_N(8,n,k-kk,-1.0,a_ptr+kk*8,sb+kk*n,c_ptr,ldc);
+    solve_RT(8,n,a_ptr+(kk-n)*8,sb+(kk-n)*n,c_ptr,ldc);
+    a_ptr += k * 8; c_ptr += 8;
+  }
+  for(;m_count>3;m_count-=4){
+    if(k-kk>0) GEMM_KERNEL_N(4,n,k-kk,-1.0,a_ptr+kk*4,sb+kk*n,c_ptr,ldc);
+    solve_RT(4,n,a_ptr+(kk-n)*4,sb+(kk-n)*n,c_ptr,ldc);
+    a_ptr += k * 4; c_ptr += 4;
+  }
+  for(;m_count>1;m_count-=2){
+    if(k-kk>0) GEMM_KERNEL_N(2,n,k-kk,-1.0,a_ptr+kk*2,sb+kk*n,c_ptr,ldc);
+    solve_RT(2,n,a_ptr+(kk-n)*2,sb+(kk-n)*n,c_ptr,ldc);
+    a_ptr += k * 2; c_ptr += 2;
+  }
+  if(m_count>0){
+    if(k-kk>0) GEMM_KERNEL_N(1,n,k-kk,-1.0,a_ptr+kk*1,sb+kk*n,c_ptr,ldc);
+    solve_RT(1,n,a_ptr+(kk-n)*1,sb+(kk-n)*n,c_ptr,ldc);
+    a_ptr += k * 1; c_ptr += 1;
+  }
+}
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT dummy1, FLOAT *sa, FLOAT *sb, FLOAT *C, BLASLONG ldc, BLASLONG offset){
+  float *a_ptr = sa, *b_ptr = sb+n*k, *c_ptr = C+n*ldc, *c_tmp = C;
+  float one[8] = {1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0};
+  float zero[8] = {0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0};
+  uint64_t ldc_bytes = (uint64_t)ldc * sizeof(float), K = (uint64_t)k, M = (uint64_t)m, OFF = (uint64_t)(n-offset), k_cnt = 0;
+  BLASLONG n_count = n;
+  if(n&1){b_ptr-=k; c_ptr-=ldc; COMPUTE_EDGE_1_nchunk(m,1,a_ptr,b_ptr,c_ptr,ldc,k,OFF); OFF--; n_count--;}
+  if(n&2){b_ptr-=k*2; c_ptr-=ldc*2; COMPUTE_EDGE_1_nchunk(m,2,a_ptr,b_ptr,c_ptr,ldc,k,OFF); OFF-=2; n_count-=2;}
+  for(;n_count>11;n_count-=12) COMPUTE(12)
+  for(;n_count>7;n_count-=8) COMPUTE(8)
+  for(;n_count>3;n_count-=4) COMPUTE(4)
+  return 0;
+}

kernel/x86_64/strsm_kernel_8x4_haswell_R_common.h

@@ -0,0 +1,226 @@
+/* r11 = m_counter, r12 = size_of_k_elements, r13 = kk, r14 = b_head, r15 = a_head */
+/* register i/o: %0 = a_ptr, %1 = b_ptr, %2 = c_ptr, %3 = c_tmp, %4 = ldc, %5 = k_counter */
+/* memory input: %6 = K, %7 = offset, %8 = {1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0}, %9 = {0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0}, %10 = M */
+
+#define init_m8n4(c1,c2,c3,c4)\
+  "vpxor %%ymm"#c1",%%ymm"#c1",%%ymm"#c1"; vpxor %%ymm"#c2",%%ymm"#c2",%%ymm"#c2";"\
+  "vpxor %%ymm"#c3",%%ymm"#c3",%%ymm"#c3"; vpxor %%ymm"#c4",%%ymm"#c4",%%ymm"#c4";"
+#define INIT_m8n4 init_m8n4(4,5,6,7)
+#define INIT_m8n8 INIT_m8n4 init_m8n4(8,9,10,11)
+#define INIT_m8n12 INIT_m8n8 init_m8n4(12,13,14,15)
+
+#define init_m4n4(c1,c2,c3,c4)\
+  "vpxor %%xmm"#c1",%%xmm"#c1",%%xmm"#c1"; vpxor %%xmm"#c2",%%xmm"#c2",%%xmm"#c2";"\
+  "vpxor %%xmm"#c3",%%xmm"#c3",%%xmm"#c3"; vpxor %%xmm"#c4",%%xmm"#c4",%%xmm"#c4";"
+#define INIT_m4n4 init_m4n4(4,5,6,7)
+#define INIT_m4n8 INIT_m4n4 init_m4n4(8,9,10,11)
+#define INIT_m4n12 INIT_m4n8 init_m4n4(12,13,14,15)
+
+#define init_m2n4(c1,c2)\
+  "vpxor %%xmm"#c1",%%xmm"#c1",%%xmm"#c1"; vpxor %%xmm"#c2",%%xmm"#c2",%%xmm"#c2";"
+#define INIT_m2n4 init_m2n4(4,5)
+#define INIT_m2n8 INIT_m2n4 init_m2n4(6,7)
+#define INIT_m2n12 INIT_m2n8 init_m2n4(8,9)
+
+#define init_m1n4(c1) "vpxor %%xmm"#c1",%%xmm"#c1",%%xmm"#c1";"
+#define INIT_m1n4 init_m1n4(4)
+#define INIT_m1n8 INIT_m1n4 init_m1n4(5)
+#define INIT_m1n12 INIT_m1n8 init_m1n4(6)
+
+#define GEMM_KERNEL_k1m8n4 \
+  "vmovsldup (%0),%%ymm1; vmovshdup (%0),%%ymm2;"\
+  "vbroadcastsd (%1),%%ymm3; vfnmadd231ps %%ymm3,%%ymm1,%%ymm4; vfnmadd231ps %%ymm3,%%ymm2,%%ymm5;"\
+  "vbroadcastsd 8(%1),%%ymm3; vfnmadd231ps %%ymm3,%%ymm1,%%ymm6; vfnmadd231ps %%ymm3,%%ymm2,%%ymm7;"
+#define GEMM_KERNEL_k1m8n8 GEMM_KERNEL_k1m8n4\
+  "vbroadcastsd (%1,%%r12,4),%%ymm3; vfnmadd231ps %%ymm3,%%ymm1,%%ymm8; vfnmadd231ps %%ymm3,%%ymm2,%%ymm9;"\
+  "vbroadcastsd 8(%1,%%r12,4),%%ymm3; vfnmadd231ps %%ymm3,%%ymm1,%%ymm10; vfnmadd231ps %%ymm3,%%ymm2,%%ymm11;"
+#define GEMM_KERNEL_k1m8n12 GEMM_KERNEL_k1m8n8\
+  "vbroadcastsd (%1,%%r12,8),%%ymm3; vfnmadd231ps %%ymm3,%%ymm1,%%ymm12; vfnmadd231ps %%ymm3,%%ymm2,%%ymm13;"\
+  "vbroadcastsd 8(%1,%%r12,8),%%ymm3; vfnmadd231ps %%ymm3,%%ymm1,%%ymm14; vfnmadd231ps %%ymm3,%%ymm2,%%ymm15;"
+
+#define GEMM_KERNEL_k1m4n4 \
+  "vmovsldup (%0),%%xmm1; vmovshdup (%0),%%xmm2;"\
+  "vmovddup (%1),%%xmm3; vfnmadd231ps %%xmm3,%%xmm1,%%xmm4; vfnmadd231ps %%xmm3,%%xmm2,%%xmm5;"\
+  "vmovddup 8(%1),%%xmm3; vfnmadd231ps %%xmm3,%%xmm1,%%xmm6; vfnmadd231ps %%xmm3,%%xmm2,%%xmm7;"
+#define GEMM_KERNEL_k1m4n8 GEMM_KERNEL_k1m4n4\
+  "vmovddup (%1,%%r12,4),%%xmm3; vfnmadd231ps %%xmm3,%%xmm1,%%xmm8; vfnmadd231ps %%xmm3,%%xmm2,%%xmm9;"\
+  "vmovddup 8(%1,%%r12,4),%%xmm3; vfnmadd231ps %%xmm3,%%xmm1,%%xmm10; vfnmadd231ps %%xmm3,%%xmm2,%%xmm11;"
+#define GEMM_KERNEL_k1m4n12 GEMM_KERNEL_k1m4n8\
+  "vmovddup (%1,%%r12,8),%%xmm3; vfnmadd231ps %%xmm3,%%xmm1,%%xmm12; vfnmadd231ps %%xmm3,%%xmm2,%%xmm13;"\
+  "vmovddup 8(%1,%%r12,8),%%xmm3; vfnmadd231ps %%xmm3,%%xmm1,%%xmm14; vfnmadd231ps %%xmm3,%%xmm2,%%xmm15;"
+
+#define GEMM_KERNEL_k1m2n4 \
+  "vbroadcastss (%0),%%xmm1; vbroadcastss 4(%0),%%xmm2;"\
+  "vmovups (%1),%%xmm3; vfnmadd231ps %%xmm3,%%xmm1,%%xmm4; vfnmadd231ps %%xmm3,%%xmm2,%%xmm5;"
+#define GEMM_KERNEL_k1m2n8 GEMM_KERNEL_k1m2n4\
+  "vmovups (%1,%%r12,4),%%xmm3; vfnmadd231ps %%xmm3,%%xmm1,%%xmm6; vfnmadd231ps %%xmm3,%%xmm2,%%xmm7;"
+#define GEMM_KERNEL_k1m2n12 GEMM_KERNEL_k1m2n8\
+  "vmovups (%1,%%r12,8),%%xmm3; vfnmadd231ps %%xmm3,%%xmm1,%%xmm8; vfnmadd231ps %%xmm3,%%xmm2,%%xmm9;"
+
+#define GEMM_KERNEL_k1m1n4 "vbroadcastss (%0),%%xmm1; vfnmadd231ps (%1),%%xmm1,%%xmm4;"
+#define GEMM_KERNEL_k1m1n8 GEMM_KERNEL_k1m1n4 "vfnmadd231ps (%1,%%r12,4),%%xmm1,%%xmm5;"
+#define GEMM_KERNEL_k1m1n12 GEMM_KERNEL_k1m1n8 "vfnmadd231ps (%1,%%r12,8),%%xmm1,%%xmm6;"
+
+#define GEMM_SUM_REORDER_8x4(c1,c2,c3,c4,prefpos)\
+  "vmovups (%3),%%ymm0; vmovups (%3,%4,1),%%ymm1; prefetcht1 "#prefpos"(%3); prefetcht1 "#prefpos"(%3,%4,1); leaq (%3,%4,2),%3;"\
+  "vunpcklps %%ymm1,%%ymm0,%%ymm2; vunpckhps %%ymm1,%%ymm0,%%ymm3; vunpcklpd %%ymm3,%%ymm2,%%ymm0; vunpckhpd %%ymm3,%%ymm2,%%ymm1;"\
+  "vaddps %%ymm0,%%ymm"#c1",%%ymm"#c1"; vaddps %%ymm1,%%ymm"#c2",%%ymm"#c2";"\
+  "vmovups (%3),%%ymm0; vmovups (%3,%4,1),%%ymm1; prefetcht1 "#prefpos"(%3); prefetcht1 "#prefpos"(%3,%4,1); leaq (%3,%4,2),%3;"\
+  "vunpcklps %%ymm1,%%ymm0,%%ymm2; vunpckhps %%ymm1,%%ymm0,%%ymm3; vunpcklpd %%ymm3,%%ymm2,%%ymm0; vunpckhpd %%ymm3,%%ymm2,%%ymm1;"\
+  "vaddps %%ymm0,%%ymm"#c3",%%ymm"#c3"; vaddps %%ymm1,%%ymm"#c4",%%ymm"#c4";"
+
+#define GEMM_SUM_REORDER_4x4(c1,c2,c3,c4,co1,co2)\
+  "vmovups (%3),%%xmm0; vmovups (%3,%4,1),%%xmm1; leaq (%3,%4,2),%3;"\
+  "vunpcklps %%xmm1,%%xmm0,%%xmm2; vunpckhps %%xmm1,%%xmm0,%%xmm3;"\
+  "vunpcklpd %%xmm"#c2",%%xmm"#c1",%%xmm0; vunpckhpd %%xmm"#c2",%%xmm"#c1",%%xmm1;"\
+  "vaddps %%xmm0,%%xmm2,%%xmm"#c1"; vaddps %%xmm1,%%xmm3,%%xmm"#c2";"\
+  "vmovups (%3),%%xmm0; vmovups (%3,%4,1),%%xmm1; leaq (%3,%4,2),%3;"\
+  "vunpcklps %%xmm1,%%xmm0,%%xmm2; vunpckhps %%xmm1,%%xmm0,%%xmm3;"\
+  "vunpcklpd %%xmm"#c4",%%xmm"#c3",%%xmm0; vunpckhpd %%xmm"#c4",%%xmm"#c3",%%xmm1;"\
+  "vaddps %%xmm0,%%xmm2,%%xmm"#c3"; vaddps %%xmm1,%%xmm3,%%xmm"#c4";"\
+  "vperm2f128 $2,%%ymm"#c1",%%ymm"#c2",%%ymm"#co1"; vperm2f128 $2,%%ymm"#c3",%%ymm"#c4",%%ymm"#co2";"
+
+#define GEMM_SUM_REORDER_2x4(c1,c2)\
+  "vmovsd (%3),%%xmm0; vmovhpd (%3,%4,1),%%xmm0,%%xmm0; leaq (%3,%4,2),%3; vpermilps $216,%%xmm0,%%xmm0;"\
+  "vmovsd (%3),%%xmm1; vmovhpd (%3,%4,1),%%xmm1,%%xmm1; leaq (%3,%4,2),%3; vpermilps $216,%%xmm1,%%xmm1;"\
+  "vunpcklpd %%xmm1,%%xmm0,%%xmm2; vaddps %%xmm2,%%xmm"#c1",%%xmm"#c1";"\
+  "vunpckhpd %%xmm1,%%xmm0,%%xmm3; vaddps %%xmm3,%%xmm"#c2",%%xmm"#c2";"\
+
+#define GEMM_SUM_REORDER_1x4(c1)\
+  "vmovss (%3),%%xmm1; vinsertps $16,(%3,%4,1),%%xmm1,%%xmm1; leaq (%3,%4,2),%3;"\
+  "vinsertps $32,(%3),%%xmm1,%%xmm1; vinsertps $48,(%3,%4,1),%%xmm1,%%xmm1; leaq (%3,%4,2),%3;"\
+  "vaddps %%xmm"#c1",%%xmm1,%%xmm"#c1";"
+
+#define SOLVE_le_m4n2(b_off,c1,...)\
+  "vbroadcastsd "#b_off"("#__VA_ARGS__"),%%ymm0; vblendps $170,%8,%%ymm0,%%ymm2;"\
+  "vmulps %%ymm2,%%ymm"#c1",%%ymm"#c1";"\
+  "vmovsldup %%ymm"#c1",%%ymm1;"
+
+#define SOLVE_le_m8n2(b_off,c1,c2,...)\
+  "vbroadcastsd "#b_off"("#__VA_ARGS__"),%%ymm0; vblendps $170,%8,%%ymm0,%%ymm2;"\
+  "vmulps %%ymm2,%%ymm"#c1",%%ymm"#c1"; vmulps %%ymm2,%%ymm"#c2",%%ymm"#c2";"\
+  "vmovsldup %%ymm"#c1",%%ymm1; vmovsldup %%ymm"#c2",%%ymm2;"
+
+#define SOLVE_leri_m4n2(b_off,c1,...) SOLVE_le_m4n2(b_off,c1,__VA_ARGS__)\
+  "vblendps $85,%9,%%ymm0,%%ymm0; vfnmadd231ps %%ymm0,%%ymm1,%%ymm"#c1";"
+
+#define SOLVE_leri_m8n2(b_off,c1,c2,...) SOLVE_le_m8n2(b_off,c1,c2,__VA_ARGS__)\
+  "vblendps $85,%9,%%ymm0,%%ymm0; vfnmadd231ps %%ymm0,%%ymm1,%%ymm"#c1"; vfnmadd231ps %%ymm0,%%ymm2,%%ymm"#c2";"
+
+#define SOLVE_ri_m4n2(b_off,c1,...)\
+  "vbroadcastsd "#b_off"("#__VA_ARGS__"),%%ymm0; vblendps $85,%8,%%ymm0,%%ymm2;"\
+  "vmulps %%ymm2,%%ymm"#c1",%%ymm"#c1";"\
+  "vmovshdup %%ymm"#c1",%%ymm1;"
+
+#define SOLVE_ri_m8n2(b_off,c1,c2,...)\
+  "vbroadcastsd "#b_off"("#__VA_ARGS__"),%%ymm0; vblendps $85,%8,%%ymm0,%%ymm2;"\
+  "vmulps %%ymm2,%%ymm"#c1",%%ymm"#c1"; vmulps %%ymm2,%%ymm"#c2",%%ymm"#c2";"\
+  "vmovshdup %%ymm"#c1",%%ymm1; vmovshdup %%ymm"#c2",%%ymm2;"
+
+#define SOLVE_rile_m4n2(b_off,c1,...) SOLVE_ri_m4n2(b_off,c1,__VA_ARGS__)\
+  "vblendps $170,%9,%%ymm0,%%ymm0; vfnmadd231ps %%ymm0,%%ymm1,%%ymm"#c1";"
+
+#define SOLVE_rile_m8n2(b_off,c1,c2,...) SOLVE_ri_m8n2(b_off,c1,c2,__VA_ARGS__)\
+  "vblendps $170,%9,%%ymm0,%%ymm0; vfnmadd231ps %%ymm0,%%ymm1,%%ymm"#c1"; vfnmadd231ps %%ymm0,%%ymm2,%%ymm"#c2";"
+
+#define SOLVE_col1_rtol_m1n4(b_off,c1,...)\
+  "vmovups "#b_off"("#__VA_ARGS__"),%%xmm0; vblendps $14,%8,%%xmm0,%%xmm2;"\
+  "vmulps %%xmm2,%%xmm"#c1",%%xmm"#c1";"\
+  "vpermilps $0,%%xmm"#c1",%%xmm1;"
+
+#define SOLVE_col1_rtol_m2n4(b_off,c1,c2,...)\
+  "vmovups "#b_off"("#__VA_ARGS__"),%%xmm0; vblendps $14,%8,%%xmm0,%%xmm2;"\
+  "vmulps %%xmm2,%%xmm"#c1",%%xmm"#c1"; vmulps %%xmm2,%%xmm"#c2",%%xmm"#c2";"\
+  "vpermilps $0,%%xmm"#c1",%%xmm1; vpermilps $0,%%xmm"#c2",%%xmm2;"
+
+#define SOLVE_col1_ltor_m1n4(b_off,c1,...) SOLVE_col1_rtol_m1n4(b_off,c1,__VA_ARGS__)\
+  "vblendps $1,%9,%%xmm0,%%xmm0; vfnmadd231ps %%xmm0,%%xmm1,%%xmm"#c1";"
+
+#define SOLVE_col1_ltor_m2n4(b_off,c1,c2,...) SOLVE_col1_rtol_m2n4(b_off,c1,c2,__VA_ARGS__)\
+  "vblendps $1,%9,%%xmm0,%%xmm0; vfnmadd231ps %%xmm0,%%xmm1,%%xmm"#c1"; vfnmadd231ps %%xmm0,%%xmm2,%%xmm"#c2";"
+
+#define SOLVE_col2_mul_m1n4(b_off,c1,...)\
+  "vmovups "#b_off"("#__VA_ARGS__"),%%xmm0; vblendps $13,%8,%%xmm0,%%xmm2;"\
+  "vmulps %%xmm2,%%xmm"#c1",%%xmm"#c1";"\
+  "vpermilps $85,%%xmm"#c1",%%xmm1;"
+
+#define SOLVE_col2_mul_m2n4(b_off,c1,c2,...)\
+  "vmovups "#b_off"("#__VA_ARGS__"),%%xmm0; vblendps $13,%8,%%xmm0,%%xmm2;"\
+  "vmulps %%xmm2,%%xmm"#c1",%%xmm"#c1"; vmulps %%xmm2,%%xmm"#c2",%%xmm"#c2";"\
+  "vpermilps $85,%%xmm"#c1",%%xmm1; vpermilps $85,%%xmm"#c2",%%xmm2;"
+
+#define SOLVE_col2_rtol_m1n4(b_off,c1,...) SOLVE_col2_mul_m1n4(b_off,c1,__VA_ARGS__)\
+  "vblendps $14,%9,%%xmm0,%%xmm0; vfnmadd231ps %%xmm0,%%xmm1,%%xmm"#c1";"
+
+#define SOLVE_col2_rtol_m2n4(b_off,c1,c2,...) SOLVE_col2_mul_m2n4(b_off,c1,c2,__VA_ARGS__)\
+  "vblendps $14,%9,%%xmm0,%%xmm0; vfnmadd231ps %%xmm0,%%xmm1,%%xmm"#c1"; vfnmadd231ps %%xmm0,%%xmm2,%%xmm"#c2";"
+
+#define SOLVE_col2_ltor_m1n4(b_off,c1,...) SOLVE_col2_mul_m1n4(b_off,c1,__VA_ARGS__)\
+  "vblendps $3,%9,%%xmm0,%%xmm0; vfnmadd231ps %%xmm0,%%xmm1,%%xmm"#c1";"
+
+#define SOLVE_col2_ltor_m2n4(b_off,c1,c2,...) SOLVE_col2_mul_m2n4(b_off,c1,c2,__VA_ARGS__)\
+  "vblendps $3,%9,%%xmm0,%%xmm0; vfnmadd231ps %%xmm0,%%xmm1,%%xmm"#c1"; vfnmadd231ps %%xmm0,%%xmm2,%%xmm"#c2";"
+
+#define SOLVE_col3_mul_m1n4(b_off,c1,...)\
+  "vmovups "#b_off"("#__VA_ARGS__"),%%xmm0; vblendps $11,%8,%%xmm0,%%xmm2;"\
+  "vmulps %%xmm2,%%xmm"#c1",%%xmm"#c1";"\
+  "vpermilps $170,%%xmm"#c1",%%xmm1;"
+
+#define SOLVE_col3_mul_m2n4(b_off,c1,c2,...)\
+  "vmovups "#b_off"("#__VA_ARGS__"),%%xmm0; vblendps $11,%8,%%xmm0,%%xmm2;"\
+  "vmulps %%xmm2,%%xmm"#c1",%%xmm"#c1"; vmulps %%xmm2,%%xmm"#c2",%%xmm"#c2";"\
+  "vpermilps $170,%%xmm"#c1",%%xmm1; vpermilps $170,%%xmm"#c2",%%xmm2;"
+
+#define SOLVE_col3_rtol_m1n4(b_off,c1,...) SOLVE_col3_mul_m1n4(b_off,c1,__VA_ARGS__)\
+  "vblendps $12,%9,%%xmm0,%%xmm0; vfnmadd231ps %%xmm0,%%xmm1,%%xmm"#c1";"
+
+#define SOLVE_col3_rtol_m2n4(b_off,c1,c2,...) SOLVE_col3_mul_m2n4(b_off,c1,c2,__VA_ARGS__)\
+  "vblendps $12,%9,%%xmm0,%%xmm0; vfnmadd231ps %%xmm0,%%xmm1,%%xmm"#c1"; vfnmadd231ps %%xmm0,%%xmm2,%%xmm"#c2";"
+
+#define SOLVE_col3_ltor_m1n4(b_off,c1,...) SOLVE_col3_mul_m1n4(b_off,c1,__VA_ARGS__)\
+  "vblendps $7,%9,%%xmm0,%%xmm0; vfnmadd231ps %%xmm0,%%xmm1,%%xmm"#c1";"
+
+#define SOLVE_col3_ltor_m2n4(b_off,c1,c2,...) SOLVE_col3_mul_m2n4(b_off,c1,c2,__VA_ARGS__)\
+  "vblendps $7,%9,%%xmm0,%%xmm0; vfnmadd231ps %%xmm0,%%xmm1,%%xmm"#c1"; vfnmadd231ps %%xmm0,%%xmm2,%%xmm"#c2";"
+
+#define SOLVE_col4_ltor_m1n4(b_off,c1,...)\
+  "vmovups "#b_off"("#__VA_ARGS__"),%%xmm0; vblendps $7,%8,%%xmm0,%%xmm2;"\
+  "vmulps %%xmm2,%%xmm"#c1",%%xmm"#c1";"\
+  "vpermilps $255,%%xmm"#c1",%%xmm1;"
+
+#define SOLVE_col4_ltor_m2n4(b_off,c1,c2,...)\
+  "vmovups "#b_off"("#__VA_ARGS__"),%%xmm0; vblendps $7,%8,%%xmm0,%%xmm2;"\
+  "vmulps %%xmm2,%%xmm"#c1",%%xmm"#c1"; vmulps %%xmm2,%%xmm"#c2",%%xmm"#c2";"\
+  "vpermilps $255,%%xmm"#c1",%%xmm1; vpermilps $255,%%xmm"#c2",%%xmm2;"
+
+#define SOLVE_col4_rtol_m1n4(b_off,c1,...) SOLVE_col4_ltor_m1n4(b_off,c1,__VA_ARGS__)\
+  "vblendps $8,%9,%%xmm0,%%xmm0; vfnmadd231ps %%xmm0,%%xmm1,%%xmm"#c1";"
+
+#define SOLVE_col4_rtol_m2n4(b_off,c1,c2,...) SOLVE_col4_ltor_m2n4(b_off,c1,c2,__VA_ARGS__)\
+  "vblendps $8,%9,%%xmm0,%%xmm0; vfnmadd231ps %%xmm0,%%xmm1,%%xmm"#c1"; vfnmadd231ps %%xmm0,%%xmm2,%%xmm"#c2";"
+
+#define SUBTRACT_m4n2(b_off,c1,...) "vbroadcastsd "#b_off"("#__VA_ARGS__"),%%ymm0; vfnmadd231ps %%ymm0,%%ymm1,%%ymm"#c1";"
+
+#define SUBTRACT_m8n2(b_off,c1,c2,...) SUBTRACT_m4n2(b_off,c1,__VA_ARGS__) "vfnmadd231ps %%ymm0,%%ymm2,%%ymm"#c2";"
+
+#define SUBTRACT_m1n4(b_off,c1,...) "vmovups "#b_off"("#__VA_ARGS__"),%%xmm0; vfnmadd231ps %%xmm0,%%xmm1,%%xmm"#c1";"
+
+#define SUBTRACT_m2n4(b_off,c1,c2,...) SUBTRACT_m1n4(b_off,c1,__VA_ARGS__) "vfnmadd231ps %%xmm0,%%xmm2,%%xmm"#c2";"
+
+#define SAVE_SOLUTION_m8n2(c1,c2,a_off)\
+  "vunpcklps %%ymm"#c2",%%ymm"#c1",%%ymm0; vunpckhps %%ymm"#c2",%%ymm"#c1",%%ymm1;"\
+  "vunpcklpd %%ymm1,%%ymm0,%%ymm"#c1"; vunpckhpd %%ymm1,%%ymm0,%%ymm"#c2";"\
+  "vmovups %%ymm"#c1","#a_off"(%0); vmovups %%ymm"#c2","#a_off"+32(%0);"\
+  "vmovups %%ymm"#c1",(%3); vmovups %%ymm"#c2",(%3,%4,1); leaq (%3,%4,2),%3;"
+
+#define SAVE_SOLUTION_m4n2(c1,a_off)\
+  "vpermilps $216,%%ymm"#c1",%%ymm"#c1"; vpermpd $216,%%ymm"#c1",%%ymm"#c1";"\
+  "vmovups %%ymm"#c1","#a_off"(%0); vmovups %%xmm"#c1",(%3); vextractf128 $1,%%ymm"#c1",(%3,%4,1); leaq (%3,%4,2),%3;"
+
+#define SAVE_SOLUTION_m2n4(c1,c2,a_off)\
+  "vunpcklps %%xmm"#c2",%%xmm"#c1",%%xmm0; vmovups %%xmm0,"#a_off"(%0); vmovsd %%xmm0,(%3); vmovhpd %%xmm0,(%3,%4,1); leaq (%3,%4,2),%3;"\
+  "vunpckhps %%xmm"#c2",%%xmm"#c1",%%xmm0; vmovups %%xmm0,"#a_off"+16(%0); vmovsd %%xmm0,(%3); vmovhpd %%xmm0,(%3,%4,1); leaq (%3,%4,2),%3;"
+
+#define SAVE_SOLUTION_m1n4(c1,a_off)\
+  "vmovups %%xmm"#c1","#a_off"(%0); vmovss %%xmm"#c1",(%3); vextractps $1,%%xmm"#c1",(%3,%4,1); leaq (%3,%4,2),%3;"\
+  "vextractps $2,%%xmm"#c1",(%3); vextractps $3,%%xmm"#c1",(%3,%4,1); leaq (%3,%4,2),%3;"