99 lines · plain
1; RUN: opt %loadNPMPolly '-passes=polly-custom<flatten>' -polly-print-flatten-schedule -disable-output < %s | FileCheck %s2;3; dgemm kernel4; C := alpha*A*B + beta*C5; C[ni][nj]6; A[ni][nk]7; B[nk][nj]8 9target datalayout = "e-m:x-p:32:32-i64:64-f80:32-n8:16:32-a:0:32-S32"10 11define void @gemm(i32 %ni, i32 %nj, i32 %nk, double %alpha, double %beta, ptr noalias nonnull %C, ptr noalias nonnull %A, ptr noalias nonnull %B) {12entry:13 br label %ni.for14 15ni.for:16 %i = phi i32 [0, %entry], [%i.inc, %ni.inc]17 %i.cmp = icmp slt i32 %i, 318 br i1 %i.cmp, label %nj.for, label %ni.exit19 20 nj.for:21 %j = phi i32 [0, %ni.for], [%j.inc, %nj.inc]22 %j.cmp = icmp slt i32 %j, 723 br i1 %j.cmp, label %nj_beta, label %nj.exit24 25 nj_beta:26 %c_stride = mul nsw i32 %i, 3; %nj27 %c_idx_i = getelementptr inbounds double, ptr %C, i32 %c_stride28 %c_idx_ij = getelementptr inbounds double, ptr %c_idx_i, i32 %j29 30 ; C[i]ptr= beta31 %c = load double, ptr %c_idx_ij32 %c_beta = fmul double %c, %beta33 store double %c_beta, ptr %c_idx_ij34 35 br label %nk.for36 37 nk.for:38 %k = phi i32 [0, %nj_beta], [%k.inc, %nk.inc]39 %k.cmp = icmp slt i32 %k, 3 ; %nk40 br i1 %k.cmp, label %nk_alpha, label %nk.exit41 42 nk_alpha:43 %a_stride = mul nsw i32 %i, 3; %nk44 %a_idx_i = getelementptr inbounds double, ptr %A, i32 %a_stride45 %a_idx_ik = getelementptr inbounds double, ptr %a_idx_i, i32 %k46 47 %b_stride = mul nsw i32 %k, 3; %nj48 %b_idx_k = getelementptr inbounds double, ptr %B, i32 %b_stride49 %b_idx_kj = getelementptr inbounds double, ptr %b_idx_k, i32 %j50 51 ; C[i][j] += alpha * A[i]ptr B[k][j]52 %a = load double, ptr %a_idx_ik53 %b = load double, ptr %b_idx_kj54 %beta_c = load double, ptr %c_idx_ij55 56 %alpha_a = fmul double %a, %alpha57 %alpha_a_b = fmul double %alpha_a, %b58 %beta_c_alpha_a_b = fadd double %beta_c, %alpha_a_b59 60 store double %beta_c_alpha_a_b, ptr %c_idx_ij61 62 br label %nk.inc63 64 nk.inc:65 %k.inc = add nuw nsw i32 %k, 166 br label %nk.for67 68 nk.exit:69 ; store double %c, ptr %c_idx_ij70 br label %nj.inc71 72 nj.inc:73 %j.inc = add nuw nsw i32 %j, 174 br label %nj.for75 76 nj.exit:77 br label %ni.inc78 79ni.inc:80 %i.inc = add nuw nsw i32 %i, 181 br label %ni.for82 83ni.exit:84 br label %return85 86return:87 ret void88}89 90 91; CHECK: Schedule before flattening {92; CHECK-NEXT: { Stmt_nk_alpha[i0, i1, i2] -> [i0, i1, 1, i2] }93; CHECK-NEXT: { Stmt_nj_beta[i0, i1] -> [i0, i1, 0, 0] }94; CHECK-NEXT: }95; CHECK: Schedule after flattening {96; CHECK-NEXT: { Stmt_nj_beta[i0, i1] -> [28i0 + 4i1] }97; CHECK-NEXT: { Stmt_nk_alpha[i0, i1, i2] -> [1 + 28i0 + 4i1 + i2] }98; CHECK-NEXT: }99