/* BLIS An object-based framework for developing high-performance BLAS-like libraries. Copyright (C) 2014, The University of Texas at Austin Copyright (C) 2020, Advanced Micro Devices, Inc. Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met: - Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. - Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution. - Neither the name(s) of the copyright holder(s) nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission. THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */ #include "blis.h" #define BLIS_ASM_SYNTAX_ATT #include "bli_x86_asm_macros.h" // assumes beta.r, beta.i have been broadcast into ymm1, ymm2. // outputs to ymm0 #define CGEMM_INPUT_SCALE_CS_BETA_NZ \ vmovlpd(mem(rcx), xmm0, xmm0) \ vmovhpd(mem(rcx, rsi, 1), xmm0, xmm0) \ vmovlpd(mem(rcx, rsi, 2), xmm3, xmm3) \ vmovhpd(mem(rcx, r13, 1), xmm3, xmm3) \ vinsertf128(imm(1), xmm3, ymm0, ymm0) \ vpermilps(imm(0xb1), ymm0, ymm3) \ vmulps(ymm1, ymm0, ymm0) \ vmulps(ymm2, ymm3, ymm3) \ vaddsubps(ymm3, ymm0, ymm0) #define CGEMM_INPUT_SCALE_CS_BETA_NZ_128 \ vmovlpd(mem(rcx), xmm0, xmm0) \ vmovhpd(mem(rcx, rsi, 1), xmm0, xmm0) \ vpermilps(imm(0xb1), xmm0, xmm3) \ vmulps(xmm1, xmm0, xmm0) \ vmulps(xmm2, xmm3, xmm3) \ vaddsubps(xmm3, xmm0, xmm0) #define CGEMM_INPUT_SCALE_RS_BETA_NZ \ vmovups(mem(rcx), ymm0) \ vpermilps(imm(0xb1), ymm0, ymm3) \ vmulps(ymm1, ymm0, ymm0) \ vmulps(ymm2, ymm3, ymm3) \ vaddsubps(ymm3, ymm0, ymm0) #define CGEMM_OUTPUT_RS \ vmovups(ymm0, mem(rcx)) \ #define CGEMM_INPUT_SCALE_RS_BETA_NZ_NEXT \ vmovups(mem(rcx, rsi, 8), ymm0) \ vpermilps(imm(0xb1), ymm0, ymm3) \ vmulps(ymm1, ymm0, ymm0) \ vmulps(ymm2, ymm3, ymm3) \ vaddsubps(ymm3, ymm0, ymm0) #define CGEMM_OUTPUT_RS_NEXT \ vmovups(ymm0, mem(rcx, rsi, 8)) \ /* rrr: -------- ------ -------- -------- += ------ ... -------- -------- ------ -------- -------- ------ : rcr: -------- | | | | -------- -------- += | | | | ... -------- -------- | | | | -------- -------- | | | | : Assumptions: - B is row-stored; - A is row- or column-stored; - m0 and n0 are at most MR and NR, respectively. Therefore, this (r)ow-preferential kernel is well-suited for contiguous (v)ector loads on B and single-element broadcasts from A. NOTE: These kernels explicitly support column-oriented IO, implemented via an in-register transpose. And thus they also support the crr and ccr cases, though only crr is ever utilized (because ccr is handled by transposing the operation and executing rcr, which does not incur the cost of the in-register transpose). crr: | | | | | | | | ------ -------- | | | | | | | | += ------ ... -------- | | | | | | | | ------ -------- | | | | | | | | ------ : */ void bli_cgemmsup_rv_zen_asm_3x8m ( conj_t conja, conj_t conjb, dim_t m0, dim_t n0, dim_t k0, scomplex* restrict alpha, scomplex* restrict a, inc_t rs_a0, inc_t cs_a0, scomplex* restrict b, inc_t rs_b0, inc_t cs_b0, scomplex* restrict beta, scomplex* restrict c, inc_t rs_c0, inc_t cs_c0, auxinfo_t* data, cntx_t* cntx ) { uint64_t n_left = n0 % 8; // First check whether this is a edge case in the n dimension. If so, // dispatch other 3x?m kernels, as needed. if (n_left ) { scomplex* cij = c; scomplex* bj = b; scomplex* ai = a; if ( 4 <= n_left ) { const dim_t nr_cur = 4; bli_cgemmsup_rv_zen_asm_3x4m ( conja, conjb, m0, nr_cur, k0, alpha, ai, rs_a0, cs_a0, bj, rs_b0, cs_b0, beta, cij, rs_c0, cs_c0, data, cntx ); cij += nr_cur*cs_c0; bj += nr_cur*cs_b0; n_left -= nr_cur; } if ( 2 <= n_left ) { const dim_t nr_cur = 2; bli_cgemmsup_rv_zen_asm_3x2m ( conja, conjb, m0, nr_cur, k0, alpha, ai, rs_a0, cs_a0, bj, rs_b0, cs_b0, beta, cij, rs_c0, cs_c0, data, cntx ); cij += nr_cur*cs_c0; bj += nr_cur*cs_b0; n_left -= nr_cur; } if ( 1 == n_left ) { bli_cgemv_ex ( BLIS_NO_TRANSPOSE, conjb, m0, k0, alpha, ai, rs_a0, cs_a0, bj, rs_b0, beta, cij, rs_c0, cntx, NULL ); } return; } //void* a_next = bli_auxinfo_next_a( data ); //void* b_next = bli_auxinfo_next_b( data ); // Typecast local copies of integers in case dim_t and inc_t are a // different size than is expected by load instructions. uint64_t k_iter = k0 / 4; uint64_t k_left = k0 % 4; uint64_t m_iter = m0 / 3; uint64_t m_left = m0 % 3; uint64_t rs_a = rs_a0; uint64_t cs_a = cs_a0; uint64_t rs_b = rs_b0; uint64_t cs_b = cs_b0; uint64_t rs_c = rs_c0; uint64_t cs_c = cs_c0; if ( m_iter == 0 ) goto consider_edge_cases; // ------------------------------------------------------------------------- begin_asm() mov(var(a), r14) // load address of a. mov(var(rs_a), r8) // load rs_a mov(var(cs_a), r9) // load cs_a lea(mem(, r8, 8), r8) // rs_a *= sizeof(dt) lea(mem(, r9, 8), r9) // cs_a *= sizeof(dt) mov(var(rs_b), r10) // load rs_b lea(mem(, r10, 8), r10) // rs_b *= sizeof(dt) // NOTE: We cannot pre-load elements of a or b // because it could eventually, in the last // unrolled iter or the cleanup loop, result // in reading beyond the bounds allocated mem // (the likely result: a segmentation fault). mov(var(c), r12) // load address of c mov(var(rs_c), rdi) // load rs_c lea(mem(, rdi, 8), rdi) // rs_c *= sizeof(dt) // During preamble and loops: // r12 = rcx = c // r14 = rax = a // read rbx from var(b) near beginning of loop // r11 = m dim index ii mov(var(m_iter), r11) // ii = m_iter; label(.SLOOP3X8I) // LOOP OVER ii = [ m_iter ... 1 0 ] vzeroall() // zero all xmm/ymm registers. mov(var(b), rbx) // load address of b. //mov(r12, rcx) // reset rcx to current utile of c. mov(r14, rax) // reset rax to current upanel of a. cmp(imm(8), rdi) // set ZF if (8*rs_c) == 8. jz(.SCOLPFETCH) // jump to column storage case label(.SROWPFETCH) // row-stored pre-fetching on c // not used lea(mem(r12, rdi, 2), rdx) // lea(mem(rdx, rdi, 1), rdx) // rdx = c + 3*rs_c; jmp(.SPOSTPFETCH) // jump to end of pre-fetching c label(.SCOLPFETCH) // column-stored pre-fetching c mov(var(cs_c), rsi) // load cs_c to rsi (temporarily) lea(mem(, rsi, 4), rsi) // cs_c *= sizeof(dt) lea(mem(r12, rsi, 2), rdx) // lea(mem(rdx, rsi, 1), rdx) // rdx = c + 3*cs_c; label(.SPOSTPFETCH) // done prefetching c lea(mem(r9, r9, 2), rcx) // rcx = 3*cs_a; lea(mem(rax, r8, 4), rdx) // use rdx for pre-fetching lines lea(mem(rdx, r8, 2), rdx) // from next upanel of a. mov(var(k_iter), rsi) // i = k_iter; test(rsi, rsi) // check i via logical AND. je(.SCONSIDKLEFT) // if i == 0, jump to code that // contains the k_left loop. label(.SLOOPKITER) // MAIN LOOP // ---------------------------------- iteration 0 vmovups(mem(rbx, 0*32), ymm0) vmovups(mem(rbx, 1*32), ymm1) add(r10, rbx) // b += rs_b; vbroadcastss(mem(rax ), ymm2) vfmadd231ps(ymm0, ymm2, ymm4) vfmadd231ps(ymm1, ymm2, ymm5) vbroadcastss(mem(rax, r8, 1), ymm2) vfmadd231ps(ymm0, ymm2, ymm8) vfmadd231ps(ymm1, ymm2, ymm9) vbroadcastss(mem(rax, r8, 2), ymm2) vfmadd231ps(ymm0, ymm2, ymm12) vfmadd231ps(ymm1, ymm2, ymm13) vbroadcastss(mem(rax, 4), ymm3) vfmadd231ps(ymm0, ymm3, ymm6) vfmadd231ps(ymm1, ymm3, ymm7) vbroadcastss(mem(rax, r8, 1, 4), ymm3) vfmadd231ps(ymm0, ymm3, ymm10) vfmadd231ps(ymm1, ymm3, ymm11) vbroadcastss(mem(rax, r8, 2, 4), ymm3) vfmadd231ps(ymm0, ymm3, ymm14) vfmadd231ps(ymm1, ymm3, ymm15) add(r9, rax) // a += cs_a; // ---------------------------------- iteration 1 vmovups(mem(rbx, 0*32), ymm0) vmovups(mem(rbx, 1*32), ymm1) add(r10, rbx) // b += rs_b; vbroadcastss(mem(rax ), ymm2) vfmadd231ps(ymm0, ymm2, ymm4) vfmadd231ps(ymm1, ymm2, ymm5) vbroadcastss(mem(rax, r8, 1), ymm2) vfmadd231ps(ymm0, ymm2, ymm8) vfmadd231ps(ymm1, ymm2, ymm9) vbroadcastss(mem(rax, r8, 2), ymm2) vfmadd231ps(ymm0, ymm2, ymm12) vfmadd231ps(ymm1, ymm2, ymm13) vbroadcastss(mem(rax, 4), ymm3) vfmadd231ps(ymm0, ymm3, ymm6) vfmadd231ps(ymm1, ymm3, ymm7) vbroadcastss(mem(rax, r8, 1, 4), ymm3) vfmadd231ps(ymm0, ymm3, ymm10) vfmadd231ps(ymm1, ymm3, ymm11) vbroadcastss(mem(rax, r8, 2, 4), ymm3) vfmadd231ps(ymm0, ymm3, ymm14) vfmadd231ps(ymm1, ymm3, ymm15) add(r9, rax) // a += cs_a; // ---------------------------------- iteration 2 vmovups(mem(rbx, 0*32), ymm0) vmovups(mem(rbx, 1*32), ymm1) add(r10, rbx) // b += rs_b; vbroadcastss(mem(rax ), ymm2) vfmadd231ps(ymm0, ymm2, ymm4) vfmadd231ps(ymm1, ymm2, ymm5) vbroadcastss(mem(rax, r8, 1), ymm2) vfmadd231ps(ymm0, ymm2, ymm8) vfmadd231ps(ymm1, ymm2, ymm9) vbroadcastss(mem(rax, r8, 2), ymm2) vfmadd231ps(ymm0, ymm2, ymm12) vfmadd231ps(ymm1, ymm2, ymm13) vbroadcastss(mem(rax, 4), ymm3) vfmadd231ps(ymm0, ymm3, ymm6) vfmadd231ps(ymm1, ymm3, ymm7) vbroadcastss(mem(rax, r8, 1, 4), ymm3) vfmadd231ps(ymm0, ymm3, ymm10) vfmadd231ps(ymm1, ymm3, ymm11) vbroadcastss(mem(rax, r8, 2, 4), ymm3) vfmadd231ps(ymm0, ymm3, ymm14) vfmadd231ps(ymm1, ymm3, ymm15) add(r9, rax) // a += cs_a; // ---------------------------------- iteration 3 lea(mem(rdx, r9, 4), rdx) // a_prefetch += 4*cs_a; vmovups(mem(rbx, 0*32), ymm0) vmovups(mem(rbx, 1*32), ymm1) add(r10, rbx) // b += rs_b; vbroadcastss(mem(rax ), ymm2) vfmadd231ps(ymm0, ymm2, ymm4) vfmadd231ps(ymm1, ymm2, ymm5) vbroadcastss(mem(rax, r8, 1), ymm2) vfmadd231ps(ymm0, ymm2, ymm8) vfmadd231ps(ymm1, ymm2, ymm9) vbroadcastss(mem(rax, r8, 2), ymm2) vfmadd231ps(ymm0, ymm2, ymm12) vfmadd231ps(ymm1, ymm2, ymm13) vbroadcastss(mem(rax, 4), ymm3) vfmadd231ps(ymm0, ymm3, ymm6) vfmadd231ps(ymm1, ymm3, ymm7) vbroadcastss(mem(rax, r8, 1, 4), ymm3) vfmadd231ps(ymm0, ymm3, ymm10) vfmadd231ps(ymm1, ymm3, ymm11) vbroadcastss(mem(rax, r8, 2, 4), ymm3) vfmadd231ps(ymm0, ymm3, ymm14) vfmadd231ps(ymm1, ymm3, ymm15) add(r9, rax) // a += cs_a; dec(rsi) // i -= 1; jne(.SLOOPKITER) // iterate again if i != 0. label(.SCONSIDKLEFT) mov(var(k_left), rsi) // i = k_left; test(rsi, rsi) // check i via logical AND. je(.SPOSTACCUM) // if i == 0, we're done; jump to end. // else, we prepare to enter k_left loop. label(.SLOOPKLEFT) // EDGE LOOP vmovups(mem(rbx, 0*32), ymm0) vmovups(mem(rbx, 1*32), ymm1) add(r10, rbx) // b += rs_b; vbroadcastss(mem(rax ), ymm2) vfmadd231ps(ymm0, ymm2, ymm4) vfmadd231ps(ymm1, ymm2, ymm5) vbroadcastss(mem(rax, r8, 1), ymm2) vfmadd231ps(ymm0, ymm2, ymm8) vfmadd231ps(ymm1, ymm2, ymm9) vbroadcastss(mem(rax, r8, 2), ymm2) vfmadd231ps(ymm0, ymm2, ymm12) vfmadd231ps(ymm1, ymm2, ymm13) vbroadcastss(mem(rax, 4 ), ymm3) vfmadd231ps(ymm0, ymm3, ymm6) vfmadd231ps(ymm1, ymm3, ymm7) vbroadcastss(mem(rax, r8, 1, 4), ymm3) vfmadd231ps(ymm0, ymm3, ymm10) vfmadd231ps(ymm1, ymm3, ymm11) vbroadcastss(mem(rax, r8, 2, 4), ymm3) vfmadd231ps(ymm0, ymm3, ymm14) vfmadd231ps(ymm1, ymm3, ymm15) add(r9, rax) // a += cs_a; dec(rsi) // i -= 1; jne(.SLOOPKLEFT) // iterate again if i != 0. label(.SPOSTACCUM) mov(r12, rcx) // reset rcx to current utile of c. // permute even and odd elements // of ymm6/7, ymm10/11, ymm/14/15 vpermilps(imm(0xb1), ymm6, ymm6) vpermilps(imm(0xb1), ymm7, ymm7) vpermilps(imm(0xb1), ymm10, ymm10) vpermilps(imm(0xb1), ymm11, ymm11) vpermilps(imm(0xb1), ymm14, ymm14) vpermilps(imm(0xb1), ymm15, ymm15) // subtract/add even/odd elements vaddsubps(ymm6, ymm4, ymm4) vaddsubps(ymm7, ymm5, ymm5) vaddsubps(ymm10, ymm8, ymm8) vaddsubps(ymm11, ymm9, ymm9) vaddsubps(ymm14, ymm12, ymm12) vaddsubps(ymm15, ymm13, ymm13) /* (ar + ai) x AB */ mov(var(alpha), rax) // load address of alpha vbroadcastss(mem(rax), ymm0) // load alpha_r and duplicate vbroadcastss(mem(rax, 4), ymm1) // load alpha_i and duplicate vpermilps(imm(0xb1), ymm4, ymm3) vmulps(ymm0, ymm4, ymm4) vmulps(ymm1, ymm3, ymm3) vaddsubps(ymm3, ymm4, ymm4) vpermilps(imm(0xb1), ymm5, ymm3) vmulps(ymm0, ymm5, ymm5) vmulps(ymm1, ymm3, ymm3) vaddsubps(ymm3, ymm5, ymm5) vpermilps(imm(0xb1), ymm8, ymm3) vmulps(ymm0, ymm8, ymm8) vmulps(ymm1, ymm3, ymm3) vaddsubps(ymm3, ymm8, ymm8) vpermilps(imm(0xb1), ymm9, ymm3) vmulps(ymm0, ymm9, ymm9) vmulps(ymm1, ymm3, ymm3) vaddsubps(ymm3, ymm9, ymm9) vpermilps(imm(0xb1), ymm12, ymm3) vmulps(ymm0, ymm12, ymm12) vmulps(ymm1, ymm3, ymm3) vaddsubps(ymm3, ymm12, ymm12) vpermilps(imm(0xb1), ymm13, ymm3) vmulps(ymm0, ymm13, ymm13) vmulps(ymm1, ymm3, ymm3) vaddsubps(ymm3, ymm13, ymm13) /* (�r + �i)x C + ((ar + ai) x AB) */ mov(var(beta), rbx) // load address of beta vbroadcastss(mem(rbx), ymm1) // load beta_r and duplicate vbroadcastss(mem(rbx, 4), ymm2) // load beta_i and duplicate mov(var(cs_c), rsi) // load cs_c lea(mem(, rsi, 4), rsi) // rsi = cs_c * sizeof(dt) lea(mem(rcx, rdi, 4), rdx) // load address of c + 4*rs_c; lea(mem(rsi, rsi, 2), rax) // rax = 3*cs_c; // now avoid loading C if beta == 0 vxorps(ymm0, ymm0, ymm0) // set ymm0 to zero. vucomiss(xmm0, xmm1) // set ZF if beta_r == 0. sete(r13b) // r13b = ( ZF == 1 ? 1 : 0 ); vucomiss(xmm0, xmm2) // set ZF if beta_i == 0. sete(r15b) // r15b = ( ZF == 1 ? 1 : 0 ); and(r13b, r15b) // set ZF if r13b & r15b == 1. jne(.SBETAZERO) // if ZF = 1, jump to beta == 0 case lea(mem(r8, r8, 2), r13) // r13 = 3*rs_a cmp(imm(8), rdi) // set ZF if (8*rs_c) == 8. jz(.SCOLSTORED) // jump to column storage case label(.SROWSTORED) CGEMM_INPUT_SCALE_RS_BETA_NZ vaddps(ymm4, ymm0, ymm0) CGEMM_OUTPUT_RS CGEMM_INPUT_SCALE_RS_BETA_NZ_NEXT vaddps(ymm5, ymm0, ymm0) CGEMM_OUTPUT_RS_NEXT add(rdi, rcx) // rcx = c + 1*rs_c CGEMM_INPUT_SCALE_RS_BETA_NZ vaddps(ymm8, ymm0, ymm0) CGEMM_OUTPUT_RS CGEMM_INPUT_SCALE_RS_BETA_NZ_NEXT vaddps(ymm9, ymm0, ymm0) CGEMM_OUTPUT_RS_NEXT add(rdi, rcx) // rcx = c + 2*rs_c CGEMM_INPUT_SCALE_RS_BETA_NZ vaddps(ymm12, ymm0, ymm0) CGEMM_OUTPUT_RS CGEMM_INPUT_SCALE_RS_BETA_NZ_NEXT vaddps(ymm13, ymm0, ymm0) CGEMM_OUTPUT_RS_NEXT jmp(.SDONE) // jump to end. label(.SCOLSTORED) /*|----------------| |-------| | | | | | | 3x4 | 3x4 | | 4x3 | | | | |-------| |----------------| | | | 4x3 | |-------| */ mov(var(cs_c), rsi) // load cs_c lea(mem(, rsi, 8), rsi) // rsi = cs_c * sizeof(dt) lea(mem(rsi, rsi, 2), r13) // r13 = 3*rs_a CGEMM_INPUT_SCALE_CS_BETA_NZ vaddps(ymm4, ymm0, ymm4) add(rdi, rcx) CGEMM_INPUT_SCALE_CS_BETA_NZ vaddps(ymm8, ymm0, ymm8) add(rdi, rcx) CGEMM_INPUT_SCALE_CS_BETA_NZ vaddps(ymm12, ymm0, ymm12) lea(mem(r12, rsi, 4), rcx) CGEMM_INPUT_SCALE_CS_BETA_NZ vaddps(ymm5, ymm0, ymm5) add(rdi, rcx) CGEMM_INPUT_SCALE_CS_BETA_NZ vaddps(ymm9, ymm0, ymm9) add(rdi, rcx) CGEMM_INPUT_SCALE_CS_BETA_NZ vaddps(ymm13, ymm0, ymm13) mov(r12, rcx) // reset rcx to current utile of c. vunpcklpd(ymm8, ymm4, ymm0) //a0a1b0b1 a4a4b4b5 //gamma00-10 gamma02-12 vunpckhpd(ymm8, ymm4, ymm2) //a2a3b2b3 a6a7b6b7 //gamma01-11 gamma03-13 /******************Transpose top tile 4x3***************************/ vmovups(xmm0, mem(rcx)) // store (gamma00-10) vmovlpd(xmm12, mem(rcx, 16)) // store (gamma20) lea(mem(rcx, rsi, 1), rcx) vmovups(xmm2, mem(rcx)) // store (gamma01-11) vmovhpd(xmm12, mem(rcx, 16)) // store (gamma21) lea(mem(rcx, rsi, 1), rcx) vextractf128(imm(0x1), ymm0, xmm0) vextractf128(imm(0x1), ymm2, xmm2) vextractf128(imm(0x1), ymm12, xmm12) vmovups(xmm0, mem(rcx)) // store (gamma02-12) vmovlpd(xmm12, mem(rcx, 16)) // store (gamma22) lea(mem(rcx, rsi, 1), rcx) vmovups(xmm2, mem(rcx)) // store (gamma03-13) vmovhpd(xmm12, mem(rcx, 16)) // store (gamma33) lea(mem(rcx, rsi, 1), rcx) /******************Transpose bottom tile 4x3***************************/ vunpcklpd(ymm9, ymm5, ymm0) //a8a9b8b9 a12a13b12b13 //gamma04-14 gamma06-16 vunpckhpd(ymm9, ymm5, ymm2) //a10a11b10b11 a14a15b14b15 //gamma05-15 gamma07-17 vmovups(xmm0, mem(rcx)) // store (gamma04-14) vmovlpd(xmm13, mem(rcx, 16)) // store (gamma24) lea(mem(rcx, rsi, 1), rcx) vmovups(xmm2, mem(rcx)) // store (gamma05-15) vmovhpd(xmm13, mem(rcx, 16)) // store (gamma25) lea(mem(rcx, rsi, 1), rcx) vextractf128(imm(0x1), ymm0, xmm0) vextractf128(imm(0x1), ymm2, xmm2) vextractf128(imm(0x1), ymm13, xmm13) vmovups(xmm0, mem(rcx)) // store (gamma06-16) vmovlpd(xmm13, mem(rcx, 16)) // store (gamma26) lea(mem(rcx, rsi, 1), rcx) vmovups(xmm2, mem(rcx)) // store (gamma07-17) vmovhpd(xmm13, mem(rcx, 16)) // store (gamma27) jmp(.SDONE) // jump to end. label(.SBETAZERO) lea(mem(r8, r8, 2), r13) // r13 = 3*rs_a cmp(imm(8), rdi) // set ZF if (8*rs_c) == 8. jz(.SCOLSTORBZ) // jump to column storage case label(.SROWSTORBZ) vmovups(ymm4, mem(rcx)) vmovups(ymm5, mem(rcx, rsi, 8)) add(rdi, rcx) vmovups(ymm8, mem(rcx)) vmovups(ymm9, mem(rcx, rsi, 8)) add(rdi, rcx) vmovups(ymm12, mem(rcx)) vmovups(ymm13, mem(rcx, rsi, 8)) jmp(.SDONE) // jump to end. label(.SCOLSTORBZ) /****3x8 tile going to save into 8x3 tile in C*****/ mov(var(cs_c), rsi) // load cs_c lea(mem(, rsi, 8), rsi) // rsi = cs_c * sizeof(dt) vunpcklpd(ymm8, ymm4, ymm0) //a0a1b0b1 a4a4b4b5 vunpckhpd(ymm8, ymm4, ymm2) //a2a3b2b3 a6a7b6b7 /******************Transpose top tile 4x3***************************/ vmovups(xmm0, mem(rcx)) vmovlpd(xmm12, mem(rcx,16)) lea(mem(rcx, rsi, 1), rcx) vmovups(xmm2, mem(rcx)) vmovhpd(xmm12,mem(rcx,16)) lea(mem(rcx, rsi, 1), rcx) vextractf128(imm(0x1), ymm0, xmm0) vextractf128(imm(0x1), ymm2, xmm2) vextractf128(imm(0x1), ymm12, xmm12) vmovups(xmm0, mem(rcx)) vmovlpd(xmm12, mem(rcx, 16)) lea(mem(rcx, rsi, 1), rcx) vmovups(xmm2, mem(rcx)) vmovhpd(xmm12, mem(rcx, 16)) lea(mem(rcx, rsi, 1), rcx) /******************Transpose bottom tile 4x3***************************/ vunpcklpd(ymm9, ymm5, ymm0) //a8a9b8b9 a12a13b12b13 vunpckhpd(ymm9, ymm5, ymm2) //a10a11b10b11 a14a15b14b15 vmovups(xmm0, mem(rcx)) vmovlpd(xmm13, mem(rcx, 16)) lea(mem(rcx, rsi, 1), rcx) vmovups(xmm2, mem(rcx)) vmovhpd(xmm13, mem(rcx, 16)) lea(mem(rcx, rsi, 1), rcx) vextractf128(imm(0x1), ymm0, xmm0) vextractf128(imm(0x1), ymm2, xmm2) vextractf128(imm(0x1), ymm13, xmm13) vmovups(xmm0, mem(rcx)) vmovlpd(xmm13, mem(rcx, 16)) lea(mem(rcx, rsi, 1), rcx) vmovups(xmm2, mem(rcx)) vmovhpd(xmm13, mem(rcx, 16)) label(.SDONE) lea(mem(r12, rdi, 2), r12) lea(mem(r12, rdi, 1), r12) // c_ii = r12 += 3*rs_c lea(mem(r14, r8, 2), r14) lea(mem(r14, r8, 1), r14) //a_ii = r14 += 3*rs_a dec(r11) // ii -= 1; jne(.SLOOP3X8I) // iterate again if ii != 0. label(.SRETURN) end_asm( : // output operands (none) : // input operands [m_iter] "m" (m_iter), [k_iter] "m" (k_iter), [k_left] "m" (k_left), [a] "m" (a), [rs_a] "m" (rs_a), [cs_a] "m" (cs_a), [b] "m" (b), [rs_b] "m" (rs_b), [cs_b] "m" (cs_b), [alpha] "m" (alpha), [beta] "m" (beta), [c] "m" (c), [rs_c] "m" (rs_c), [cs_c] "m" (cs_c)/*, [a_next] "m" (a_next), [b_next] "m" (b_next)*/ : // register clobber list "rax", "rbx", "rcx", "rdx", "rsi", "rdi", "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7", "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13", "xmm14", "xmm15", "memory" ) consider_edge_cases: // Handle edge cases in the m dimension, if they exist. if ( m_left ) { const dim_t nr_cur = 8; const dim_t i_edge = m0 - ( dim_t )m_left; scomplex* cij = c + i_edge*rs_c; scomplex* ai = a + i_edge*rs_a; scomplex* bj = b; cgemmsup_ker_ft ker_fps[3] = { NULL, bli_cgemmsup_rv_zen_asm_1x8, bli_cgemmsup_rv_zen_asm_2x8, }; cgemmsup_ker_ft ker_fp = ker_fps[ m_left ]; ker_fp ( conja, conjb, m_left, nr_cur, k0, alpha, ai, rs_a0, cs_a0, bj, rs_b0, cs_b0, beta, cij, rs_c0, cs_c0, data, cntx ); return; } } void bli_cgemmsup_rv_zen_asm_3x4m ( conj_t conja, conj_t conjb, dim_t m0, dim_t n0, dim_t k0, scomplex* restrict alpha, scomplex* restrict a, inc_t rs_a0, inc_t cs_a0, scomplex* restrict b, inc_t rs_b0, inc_t cs_b0, scomplex* restrict beta, scomplex* restrict c, inc_t rs_c0, inc_t cs_c0, auxinfo_t* data, cntx_t* cntx ) { //void* a_next = bli_auxinfo_next_a( data ); //void* b_next = bli_auxinfo_next_b( data ); // Typecast local copies of integers in case dim_t and inc_t are a // different size than is expected by load instructions. uint64_t k_iter = k0 / 4; uint64_t k_left = k0 % 4; uint64_t m_iter = m0 / 3; uint64_t m_left = m0 % 3; uint64_t rs_a = rs_a0; uint64_t cs_a = cs_a0; uint64_t rs_b = rs_b0; uint64_t cs_b = cs_b0; uint64_t rs_c = rs_c0; uint64_t cs_c = cs_c0; if ( m_iter == 0 ) goto consider_edge_cases; // ------------------------------------------------------------------------- begin_asm() mov(var(a), r14) // load address of a. mov(var(rs_a), r8) // load rs_a mov(var(cs_a), r9) // load cs_a lea(mem(, r8, 8), r8) // rs_a *= sizeof(dt) lea(mem(, r9, 8), r9) // cs_a *= sizeof(dt) mov(var(rs_b), r10) // load rs_b lea(mem(, r10, 8), r10) // rs_b *= sizeof(dt) // NOTE: We cannot pre-load elements of a or b // because it could eventually, in the last // unrolled iter or the cleanup loop, result // in reading beyond the bounds allocated mem // (the likely result: a segmentation fault). mov(var(c), r12) // load address of c mov(var(rs_c), rdi) // load rs_c lea(mem(, rdi, 8), rdi) // rs_c *= sizeof(dt) // During preamble and loops: // r12 = rcx = c // r14 = rax = a // read rbx from var(b) near beginning of loop // r11 = m dim index ii mov(var(m_iter), r11) // ii = m_iter; label(.SLOOP3X4I) // LOOP OVER ii = [ m_iter ... 1 0 ] vzeroall() // zero all xmm/ymm registers. mov(var(b), rbx) // load address of b. //mov(r12, rcx) // reset rcx to current utile of c. mov(r14, rax) // reset rax to current upanel of a. cmp(imm(8), rdi) // set ZF if (8*rs_c) == 8. jz(.SCOLPFETCH) // jump to column storage case label(.SROWPFETCH) // row-stored pre-fetching on c // not used lea(mem(r12, rdi, 2), rdx) // lea(mem(rdx, rdi, 1), rdx) // rdx = c + 3*rs_c; jmp(.SPOSTPFETCH) // jump to end of pre-fetching c label(.SCOLPFETCH) // column-stored pre-fetching c mov(var(cs_c), rsi) // load cs_c to rsi (temporarily) lea(mem(, rsi, 4), rsi) // cs_c *= sizeof(dt) lea(mem(r12, rsi, 2), rdx) // lea(mem(rdx, rsi, 1), rdx) // rdx = c + 3*cs_c; label(.SPOSTPFETCH) // done prefetching c lea(mem(r9, r9, 2), rcx) // rcx = 3*cs_a; lea(mem(rax, r8, 4), rdx) // use rdx for pre-fetching lines lea(mem(rdx, r8, 2), rdx) // from next upanel of a. mov(var(k_iter), rsi) // i = k_iter; test(rsi, rsi) // check i via logical AND. je(.SCONSIDKLEFT) // if i == 0, jump to code that // contains the k_left loop. label(.SLOOPKITER) // MAIN LOOP // ---------------------------------- iteration 0 vmovups(mem(rbx, 0*32), ymm0) add(r10, rbx) // b += rs_b; vbroadcastss(mem(rax ), ymm2) vfmadd231ps(ymm0, ymm2, ymm4) vbroadcastss(mem(rax, r8, 1), ymm2) vfmadd231ps(ymm0, ymm2, ymm8) vbroadcastss(mem(rax, r8, 2), ymm2) vfmadd231ps(ymm0, ymm2, ymm12) vbroadcastss(mem(rax, 4), ymm3) vfmadd231ps(ymm0, ymm3, ymm6) vbroadcastss(mem(rax, r8, 1, 4), ymm3) vfmadd231ps(ymm0, ymm3, ymm10) vbroadcastss(mem(rax, r8, 2, 4), ymm3) vfmadd231ps(ymm0, ymm3, ymm14) add(r9, rax) // a += cs_a; // ---------------------------------- iteration 1 vmovups(mem(rbx, 0*32), ymm0) add(r10, rbx) // b += rs_b; vbroadcastss(mem(rax ), ymm2) vfmadd231ps(ymm0, ymm2, ymm4) vbroadcastss(mem(rax, r8, 1), ymm2) vfmadd231ps(ymm0, ymm2, ymm8) vbroadcastss(mem(rax, r8, 2), ymm2) vfmadd231ps(ymm0, ymm2, ymm12) vbroadcastss(mem(rax, 4), ymm3) vfmadd231ps(ymm0, ymm3, ymm6) vbroadcastss(mem(rax, r8, 1, 4), ymm3) vfmadd231ps(ymm0, ymm3, ymm10) vbroadcastss(mem(rax, r8, 2, 4), ymm3) vfmadd231ps(ymm0, ymm3, ymm14) add(r9, rax) // a += cs_a; // ---------------------------------- iteration 2 vmovups(mem(rbx, 0*32), ymm0) add(r10, rbx) // b += rs_b; vbroadcastss(mem(rax ), ymm2) vfmadd231ps(ymm0, ymm2, ymm4) vbroadcastss(mem(rax, r8, 1), ymm2) vfmadd231ps(ymm0, ymm2, ymm8) vbroadcastss(mem(rax, r8, 2), ymm2) vfmadd231ps(ymm0, ymm2, ymm12) vbroadcastss(mem(rax, 4), ymm3) vfmadd231ps(ymm0, ymm3, ymm6) vbroadcastss(mem(rax, r8, 1, 4), ymm3) vfmadd231ps(ymm0, ymm3, ymm10) vbroadcastss(mem(rax, r8, 2, 4), ymm3) vfmadd231ps(ymm0, ymm3, ymm14) add(r9, rax) // a += cs_a; // ---------------------------------- iteration 3 lea(mem(rdx, r9, 4), rdx) // a_prefetch += 4*cs_a; vmovups(mem(rbx, 0*32), ymm0) add(r10, rbx) // b += rs_b; vbroadcastss(mem(rax ), ymm2) vfmadd231ps(ymm0, ymm2, ymm4) vbroadcastss(mem(rax, r8, 1), ymm2) vfmadd231ps(ymm0, ymm2, ymm8) vbroadcastss(mem(rax, r8, 2), ymm2) vfmadd231ps(ymm0, ymm2, ymm12) vbroadcastss(mem(rax, 4), ymm3) vfmadd231ps(ymm0, ymm3, ymm6) vbroadcastss(mem(rax, r8, 1, 4), ymm3) vfmadd231ps(ymm0, ymm3, ymm10) vbroadcastss(mem(rax, r8, 2, 4), ymm3) vfmadd231ps(ymm0, ymm3, ymm14) add(r9, rax) // a += cs_a; dec(rsi) // i -= 1; jne(.SLOOPKITER) // iterate again if i != 0. label(.SCONSIDKLEFT) mov(var(k_left), rsi) // i = k_left; test(rsi, rsi) // check i via logical AND. je(.SPOSTACCUM) // if i == 0, we're done; jump to end. // else, we prepare to enter k_left loop. label(.SLOOPKLEFT) // EDGE LOOP vmovups(mem(rbx, 0*32), ymm0) add(r10, rbx) // b += rs_b; vbroadcastss(mem(rax ), ymm2) vfmadd231ps(ymm0, ymm2, ymm4) vbroadcastss(mem(rax, r8, 1), ymm2) vfmadd231ps(ymm0, ymm2, ymm8) vbroadcastss(mem(rax, r8, 2), ymm2) vfmadd231ps(ymm0, ymm2, ymm12) vbroadcastss(mem(rax, 4), ymm3) vfmadd231ps(ymm0, ymm3, ymm6) vbroadcastss(mem(rax, r8, 1, 4), ymm3) vfmadd231ps(ymm0, ymm3, ymm10) vbroadcastss(mem(rax, r8, 2, 4), ymm3) vfmadd231ps(ymm0, ymm3, ymm14) add(r9, rax) // a += cs_a; dec(rsi) // i -= 1; jne(.SLOOPKLEFT) // iterate again if i != 0. label(.SPOSTACCUM) mov(r12, rcx) // reset rcx to current utile of c. // permute even and odd elements // of ymm6/7, ymm10/11, ymm/14/15 vpermilps(imm(0xb1), ymm6, ymm6) vpermilps(imm(0xb1), ymm10, ymm10) vpermilps(imm(0xb1), ymm14, ymm14) // subtract/add even/odd elements vaddsubps(ymm6, ymm4, ymm4) vaddsubps(ymm10, ymm8, ymm8) vaddsubps(ymm14, ymm12, ymm12) /* (ar + ai) x AB */ mov(var(alpha), rax) // load address of alpha vbroadcastss(mem(rax), ymm0) // load alpha_r and duplicate vbroadcastss(mem(rax, 4), ymm1) // load alpha_i and duplicate vpermilps(imm(0xb1), ymm4, ymm3) vmulps(ymm0, ymm4, ymm4) vmulps(ymm1, ymm3, ymm3) vaddsubps(ymm3, ymm4, ymm4) vpermilps(imm(0xb1), ymm8, ymm3) vmulps(ymm0, ymm8, ymm8) vmulps(ymm1, ymm3, ymm3) vaddsubps(ymm3, ymm8, ymm8) vpermilps(imm(0xb1), ymm12, ymm3) vmulps(ymm0, ymm12, ymm12) vmulps(ymm1, ymm3, ymm3) vaddsubps(ymm3, ymm12, ymm12) /* (�r + �i)x C + ((ar + ai) x AB) */ mov(var(beta), rbx) // load address of beta vbroadcastss(mem(rbx), ymm1) // load beta_r and duplicate vbroadcastss(mem(rbx, 4), ymm2) // load beta_i and duplicate mov(var(cs_c), rsi) // load cs_c lea(mem(, rsi, 4), rsi) // rsi = cs_c * sizeof(dt) lea(mem(rcx, rdi, 4), rdx) // load address of c + 4*rs_c; lea(mem(rsi, rsi, 2), rax) // rax = 3*cs_c; // now avoid loading C if beta == 0 vxorps(ymm0, ymm0, ymm0) // set ymm0 to zero. vucomiss(xmm0, xmm1) // set ZF if beta_r == 0. sete(r13b) // r13b = ( ZF == 1 ? 1 : 0 ); vucomiss(xmm0, xmm2) // set ZF if beta_i == 0. sete(r15b) // r15b = ( ZF == 1 ? 1 : 0 ); and(r13b, r15b) // set ZF if r13b & r15b == 1. jne(.SBETAZERO) // if ZF = 1, jump to beta == 0 case lea(mem(r8, r8, 2), r13) // r13 = 3*rs_a cmp(imm(8), rdi) // set ZF if (8*rs_c) == 8. jz(.SCOLSTORED) // jump to column storage case label(.SROWSTORED) CGEMM_INPUT_SCALE_RS_BETA_NZ vaddps(ymm4, ymm0, ymm0) CGEMM_OUTPUT_RS add(rdi, rcx) // rcx = c + 1*rs_c CGEMM_INPUT_SCALE_RS_BETA_NZ vaddps(ymm8, ymm0, ymm0) CGEMM_OUTPUT_RS add(rdi, rcx) // rcx = c + 2*rs_c CGEMM_INPUT_SCALE_RS_BETA_NZ vaddps(ymm12, ymm0, ymm0) CGEMM_OUTPUT_RS jmp(.SDONE) // jump to end. label(.SCOLSTORED) /*|--------| |-------| | | | | | 3x4 | | 4x3 | |--------| |-------| */ mov(var(cs_c), rsi) // load cs_c lea(mem(, rsi, 8), rsi) // rsi = cs_c * sizeof(dt) lea(mem(rsi, rsi, 2), r13) // r13 = 3*rs_a CGEMM_INPUT_SCALE_CS_BETA_NZ vaddps(ymm4, ymm0, ymm4) add(rdi, rcx) CGEMM_INPUT_SCALE_CS_BETA_NZ vaddps(ymm8, ymm0, ymm8) add(rdi, rcx) CGEMM_INPUT_SCALE_CS_BETA_NZ vaddps(ymm12, ymm0, ymm12) mov(r12, rcx) // reset rcx to current utile of c. vunpcklpd(ymm8, ymm4, ymm0) //a0a1b0b1 a4a4b4b5 //gamma00-10 gamma02-12 vunpckhpd(ymm8, ymm4, ymm2) //a2a3b2b3 a6a7b6b7 //gamma01-11 gamma03-13 /******************Transpose tile 4x3***************************/ vmovups(xmm0, mem(rcx)) // store (gamma00-10) vmovlpd(xmm12, mem(rcx, 16)) // store (gamma20) lea(mem(rcx, rsi, 1), rcx) vmovups(xmm2, mem(rcx)) // store (gamma01-11) vmovhpd(xmm12, mem(rcx, 16)) // store (gamma21) lea(mem(rcx, rsi, 1), rcx) vextractf128(imm(0x1), ymm0, xmm0) vextractf128(imm(0x1), ymm2, xmm2) vextractf128(imm(0x1), ymm12, xmm12) vmovups(xmm0, mem(rcx)) // store (gamma02-12) vmovlpd(xmm12, mem(rcx, 16)) // store (gamma22) lea(mem(rcx, rsi, 1), rcx) vmovups(xmm2, mem(rcx)) // store (gamma03-13) vmovhpd(xmm12, mem(rcx, 16)) // store (gamma33) lea(mem(rcx, rsi, 1), rcx) jmp(.SDONE) // jump to end. label(.SBETAZERO) lea(mem(r8, r8, 2), r13) // r13 = 3*rs_a cmp(imm(8), rdi) // set ZF if (8*rs_c) == 8. jz(.SCOLSTORBZ) // jump to column storage case label(.SROWSTORBZ) vmovups(ymm4, mem(rcx)) add(rdi, rcx) vmovups(ymm8, mem(rcx)) add(rdi, rcx) vmovups(ymm12, mem(rcx)) jmp(.SDONE) // jump to end. label(.SCOLSTORBZ) /****3x4 tile going to save into 4x3 tile in C*****/ mov(var(cs_c), rsi) // load cs_c lea(mem(, rsi, 8), rsi) // rsi = cs_c * sizeof(dt) vunpcklpd(ymm8, ymm4, ymm0) //a0a1b0b1 a4a4b4b5 vunpckhpd(ymm8, ymm4, ymm2) //a2a3b2b3 a6a7b6b7 vmovups(xmm0, mem(rcx)) vmovlpd(xmm12, mem(rcx, 16)) lea(mem(rcx, rsi, 1), rcx) vmovups(xmm2, mem(rcx)) vmovhpd(xmm12, mem(rcx, 16)) lea(mem(rcx, rsi, 1), rcx) vextractf128(imm(0x1), ymm0, xmm0) vextractf128(imm(0x1), ymm2, xmm2) vextractf128(imm(0x1), ymm12, xmm12) vmovups(xmm0, mem(rcx)) vmovlpd(xmm12, mem(rcx, 16)) lea(mem(rcx, rsi, 1), rcx) vmovups(xmm2, mem(rcx)) vmovhpd(xmm12, mem(rcx, 16)) label(.SDONE) lea(mem(r12, rdi, 2), r12) lea(mem(r12, rdi, 1), r12) // c_ii = r12 += 3*rs_c lea(mem(r14, r8, 2), r14) lea(mem(r14, r8, 1), r14) //a_ii = r14 += 3*rs_a dec(r11) // ii -= 1; jne(.SLOOP3X4I) // iterate again if ii != 0. label(.SRETURN) end_asm( : // output operands (none) : // input operands [m_iter] "m" (m_iter), [k_iter] "m" (k_iter), [k_left] "m" (k_left), [a] "m" (a), [rs_a] "m" (rs_a), [cs_a] "m" (cs_a), [b] "m" (b), [rs_b] "m" (rs_b), [cs_b] "m" (cs_b), [alpha] "m" (alpha), [beta] "m" (beta), [c] "m" (c), [rs_c] "m" (rs_c), [cs_c] "m" (cs_c)/*, [a_next] "m" (a_next), [b_next] "m" (b_next)*/ : // register clobber list "rax", "rbx", "rcx", "rdx", "rsi", "rdi", "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7", "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13", "xmm14", "xmm15", "memory" ) consider_edge_cases: // Handle edge cases in the m dimension, if they exist. if ( m_left ) { const dim_t nr_cur = 4; const dim_t i_edge = m0 - ( dim_t )m_left; scomplex* cij = c + i_edge*rs_c; scomplex* ai = a + i_edge*rs_a; scomplex* bj = b; cgemmsup_ker_ft ker_fps[3] = { NULL, bli_cgemmsup_rv_zen_asm_1x4, bli_cgemmsup_rv_zen_asm_2x4, }; cgemmsup_ker_ft ker_fp = ker_fps[ m_left ]; ker_fp ( conja, conjb, m_left, nr_cur, k0, alpha, ai, rs_a0, cs_a0, bj, rs_b0, cs_b0, beta, cij, rs_c0, cs_c0, data, cntx ); return; } } void bli_cgemmsup_rv_zen_asm_3x2m ( conj_t conja, conj_t conjb, dim_t m0, dim_t n0, dim_t k0, scomplex* restrict alpha, scomplex* restrict a, inc_t rs_a0, inc_t cs_a0, scomplex* restrict b, inc_t rs_b0, inc_t cs_b0, scomplex* restrict beta, scomplex* restrict c, inc_t rs_c0, inc_t cs_c0, auxinfo_t* data, cntx_t* cntx ) { //void* a_next = bli_auxinfo_next_a( data ); //void* b_next = bli_auxinfo_next_b( data ); // Typecast local copies of integers in case dim_t and inc_t are a // different size than is expected by load instructions. uint64_t k_iter = k0 / 4; uint64_t k_left = k0 % 4; uint64_t m_iter = m0 / 3; uint64_t m_left = m0 % 3; uint64_t rs_a = rs_a0; uint64_t cs_a = cs_a0; uint64_t rs_b = rs_b0; uint64_t cs_b = cs_b0; uint64_t rs_c = rs_c0; uint64_t cs_c = cs_c0; if ( m_iter == 0 ) goto consider_edge_cases; // ------------------------------------------------------------------------- begin_asm() mov(var(a), r14) // load address of a. mov(var(rs_a), r8) // load rs_a mov(var(cs_a), r9) // load cs_a lea(mem(, r8, 8), r8) // rs_a *= sizeof(dt) lea(mem(, r9, 8), r9) // cs_a *= sizeof(dt) mov(var(rs_b), r10) // load rs_b lea(mem(, r10, 8), r10) // rs_b *= sizeof(dt) // NOTE: We cannot pre-load elements of a or b // because it could eventually, in the last // unrolled iter or the cleanup loop, result // in reading beyond the bounds allocated mem // (the likely result: a segmentation fault). mov(var(c), r12) // load address of c mov(var(rs_c), rdi) // load rs_c lea(mem(, rdi, 8), rdi) // rs_c *= sizeof(dt) // During preamble and loops: // r12 = rcx = c // r14 = rax = a // read rbx from var(b) near beginning of loop // r11 = m dim index ii mov(var(m_iter), r11) // ii = m_iter; label(.SLOOP3X2I) // LOOP OVER ii = [ m_iter ... 1 0 ] vzeroall() // zero all xmm/ymm registers. mov(var(b), rbx) // load address of b. //mov(r12, rcx) // reset rcx to current utile of c. mov(r14, rax) // reset rax to current upanel of a. cmp(imm(8), rdi) // set ZF if (8*rs_c) == 8. jz(.SCOLPFETCH) // jump to column storage case label(.SROWPFETCH) // row-stored pre-fetching on c // not used lea(mem(r12, rdi, 2), rdx) // lea(mem(rdx, rdi, 1), rdx) // rdx = c + 3*rs_c; jmp(.SPOSTPFETCH) // jump to end of pre-fetching c label(.SCOLPFETCH) // column-stored pre-fetching c mov(var(cs_c), rsi) // load cs_c to rsi (temporarily) lea(mem(, rsi, 4), rsi) // cs_c *= sizeof(dt) lea(mem(r12, rsi, 2), rdx) // lea(mem(rdx, rsi, 1), rdx) // rdx = c + 3*cs_c; label(.SPOSTPFETCH) // done prefetching c lea(mem(r9, r9, 2), rcx) // rcx = 3*cs_a; lea(mem(rax, r8, 4), rdx) // use rdx for pre-fetching lines lea(mem(rdx, r8, 2), rdx) // from next upanel of a. mov(var(k_iter), rsi) // i = k_iter; test(rsi, rsi) // check i via logical AND. je(.SCONSIDKLEFT) // if i == 0, jump to code that // contains the k_left loop. label(.SLOOPKITER) // MAIN LOOP // ---------------------------------- iteration 0 vmovups(mem(rbx, 0*32), xmm0) add(r10, rbx) // b += rs_b; vbroadcastss(mem(rax ), xmm2) vfmadd231ps(xmm0, xmm2, xmm4) vbroadcastss(mem(rax, r8, 1), xmm2) vfmadd231ps(xmm0, xmm2, xmm8) vbroadcastss(mem(rax, r8, 2), xmm2) vfmadd231ps(xmm0, xmm2, xmm12) vbroadcastss(mem(rax, 4), xmm3) vfmadd231ps(xmm0, xmm3, xmm6) vbroadcastss(mem(rax, r8, 1, 4), xmm3) vfmadd231ps(xmm0, xmm3, xmm10) vbroadcastss(mem(rax, r8, 2, 4), xmm3) vfmadd231ps(xmm0, xmm3, xmm14) add(r9, rax) // a += cs_a; // ---------------------------------- iteration 1 vmovups(mem(rbx, 0*32), xmm0) add(r10, rbx) // b += rs_b; vbroadcastss(mem(rax ), xmm2) vfmadd231ps(xmm0, xmm2, xmm4) vbroadcastss(mem(rax, r8, 1), xmm2) vfmadd231ps(xmm0, xmm2, xmm8) vbroadcastss(mem(rax, r8, 2), xmm2) vfmadd231ps(xmm0, xmm2, xmm12) vbroadcastss(mem(rax, 4), xmm3) vfmadd231ps(xmm0, xmm3, xmm6) vbroadcastss(mem(rax, r8, 1, 4), xmm3) vfmadd231ps(xmm0, xmm3, xmm10) vbroadcastss(mem(rax, r8, 2, 4), xmm3) vfmadd231ps(xmm0, xmm3, xmm14) add(r9, rax) // a += cs_a; // ---------------------------------- iteration 2 vmovups(mem(rbx, 0*32), xmm0) add(r10, rbx) // b += rs_b; vbroadcastss(mem(rax ), xmm2) vfmadd231ps(xmm0, xmm2, xmm4) vbroadcastss(mem(rax, r8, 1), xmm2) vfmadd231ps(xmm0, xmm2, xmm8) vbroadcastss(mem(rax, r8, 2), xmm2) vfmadd231ps(xmm0, xmm2, xmm12) vbroadcastss(mem(rax, 4), xmm3) vfmadd231ps(xmm0, xmm3, xmm6) vbroadcastss(mem(rax, r8, 1, 4), xmm3) vfmadd231ps(xmm0, xmm3, xmm10) vbroadcastss(mem(rax, r8, 2, 4), xmm3) vfmadd231ps(xmm0, xmm3, xmm14) add(r9, rax) // a += cs_a; // ---------------------------------- iteration 3 lea(mem(rdx, r9, 4), rdx) // a_prefetch += 4*cs_a; vmovups(mem(rbx, 0*32), xmm0) add(r10, rbx) // b += rs_b; vbroadcastss(mem(rax ), xmm2) vfmadd231ps(xmm0, xmm2, xmm4) vbroadcastss(mem(rax, r8, 1), xmm2) vfmadd231ps(xmm0, xmm2, xmm8) vbroadcastss(mem(rax, r8, 2), xmm2) vfmadd231ps(xmm0, xmm2, xmm12) vbroadcastss(mem(rax, 4), xmm3) vfmadd231ps(xmm0, xmm3, xmm6) vbroadcastss(mem(rax, r8, 1, 4), xmm3) vfmadd231ps(xmm0, xmm3, xmm10) vbroadcastss(mem(rax, r8, 2, 4), xmm3) vfmadd231ps(xmm0, xmm3, xmm14) add(r9, rax) // a += cs_a; dec(rsi) // i -= 1; jne(.SLOOPKITER) // iterate again if i != 0. label(.SCONSIDKLEFT) mov(var(k_left), rsi) // i = k_left; test(rsi, rsi) // check i via logical AND. je(.SPOSTACCUM) // if i == 0, we're done; jump to end. // else, we prepare to enter k_left loop. label(.SLOOPKLEFT) // EDGE LOOP vmovups(mem(rbx, 0*32), xmm0) add(r10, rbx) // b += rs_b; vbroadcastss(mem(rax ), xmm2) vfmadd231ps(xmm0, xmm2, xmm4) vbroadcastss(mem(rax, r8, 1), xmm2) vfmadd231ps(xmm0, xmm2, xmm8) vbroadcastss(mem(rax, r8, 2), xmm2) vfmadd231ps(xmm0, xmm2, xmm12) vbroadcastss(mem(rax, 4), xmm3) vfmadd231ps(xmm0, xmm3, xmm6) vbroadcastss(mem(rax, r8, 1, 4), xmm3) vfmadd231ps(xmm0, xmm3, xmm10) vbroadcastss(mem(rax, r8, 2, 4), xmm3) vfmadd231ps(xmm0, xmm3, xmm14) add(r9, rax) // a += cs_a; dec(rsi) // i -= 1; jne(.SLOOPKLEFT) // iterate again if i != 0. label(.SPOSTACCUM) mov(r12, rcx) // reset rcx to current utile of c. // permute even and odd elements // of xmm6/7, xmm10/11, xmm/14/15 vpermilps(imm(0xb1), xmm6, xmm6) vpermilps(imm(0xb1), xmm10, xmm10) vpermilps(imm(0xb1), xmm14, xmm14) // subtract/add even/odd elements vaddsubps(xmm6, xmm4, xmm4) vaddsubps(xmm10, xmm8, xmm8) vaddsubps(xmm14, xmm12, xmm12) /* (ar + ai) x AB */ mov(var(alpha), rax) // load address of alpha vbroadcastss(mem(rax), xmm0) // load alpha_r and duplicate vbroadcastss(mem(rax, 4), xmm1) // load alpha_i and duplicate vpermilps(imm(0xb1), xmm4, xmm3) vmulps(xmm0, xmm4, xmm4) vmulps(xmm1, xmm3, xmm3) vaddsubps(xmm3, xmm4, xmm4) vpermilps(imm(0xb1), xmm8, xmm3) vmulps(xmm0, xmm8, xmm8) vmulps(xmm1, xmm3, xmm3) vaddsubps(xmm3, xmm8, xmm8) vpermilps(imm(0xb1), xmm12, xmm3) vmulps(xmm0, xmm12, xmm12) vmulps(xmm1, xmm3, xmm3) vaddsubps(xmm3, xmm12, xmm12) /* (�r + �i)x C + ((ar + ai) x AB) */ mov(var(beta), rbx) // load address of beta vbroadcastss(mem(rbx), xmm1) // load beta_r and duplicate vbroadcastss(mem(rbx, 4), xmm2) // load beta_i and duplicate mov(var(cs_c), rsi) // load cs_c lea(mem(, rsi, 8), rsi) // rsi = cs_c * sizeof(dt) lea(mem(rcx, rdi, 2), rdx) // load address of c + 4*rs_c; lea(mem(rsi, rsi, 2), rax) // rax = 3*cs_c; // now avoid loading C if beta == 0 vxorps(xmm0, xmm0, xmm0) // set xmm0 to zero. vucomiss(xmm0, xmm1) // set ZF if beta_r == 0. sete(r13b) // r13b = ( ZF == 1 ? 1 : 0 ); vucomiss(xmm0, xmm2) // set ZF if beta_i == 0. sete(r15b) // r15b = ( ZF == 1 ? 1 : 0 ); and(r13b, r15b) // set ZF if r13b & r15b == 1. jne(.SBETAZERO) // if ZF = 1, jump to beta == 0 case lea(mem(r8, r8, 2), r13) // r13 = 3*rs_a cmp(imm(8), rdi) // set ZF if (8*rs_c) == 8. jz(.SCOLSTORED) // jump to column storage case label(.SROWSTORED) vmovlpd(mem(rcx), xmm0, xmm0) vmovhpd(mem(rcx, rsi, 1), xmm0, xmm0) vpermilps(imm(0xb1), xmm0, xmm3) vmulps(xmm1, xmm0, xmm0) vmulps(xmm2, xmm3, xmm3) vaddsubps(xmm3, xmm0, xmm0) vaddps(xmm4, xmm0, xmm0) vmovlpd(xmm0, mem(rcx)) vmovhpd(xmm0, mem(rcx, rsi, 1)) add(rdi, rcx) // rcx = c + 1*rs_c vmovlpd(mem(rcx), xmm0, xmm0) vmovhpd(mem(rcx, rsi, 1), xmm0, xmm0) vpermilps(imm(0xb1), xmm0, xmm3) vmulps(xmm1, xmm0, xmm0) vmulps(xmm2, xmm3, xmm3) vaddsubps(xmm3, xmm0, xmm0) vaddps(xmm8, xmm0, xmm0) vmovlpd(xmm0, mem(rcx)) vmovhpd(xmm0, mem(rcx, rsi, 1)) add(rdi, rcx) // rcx = c + 2*rs_c vmovlpd(mem(rcx), xmm0, xmm0) vmovhpd(mem(rcx, rsi, 1), xmm0, xmm0) vpermilps(imm(0xb1), xmm0, xmm3) vmulps(xmm1, xmm0, xmm0) vmulps(xmm2, xmm3, xmm3) vaddsubps(xmm3, xmm0, xmm0) vaddps(xmm12, xmm0, xmm0) vmovlpd(xmm0, mem(rcx)) vmovhpd(xmm0, mem(rcx, rsi, 1)) jmp(.SDONE) // jump to end. label(.SCOLSTORED) /*|--------| |-------| | | | | | 3x2 | | 2x3 | | | |-------| |--------| */ mov(var(cs_c), rsi) // load cs_c lea(mem(, rsi, 8), rsi) // rsi = cs_c * sizeof(dt) lea(mem(rsi, rsi, 2), r13) // r13 = 3*rs_a CGEMM_INPUT_SCALE_CS_BETA_NZ_128 vaddps(xmm4, xmm0, xmm4) add(rdi, rcx) CGEMM_INPUT_SCALE_CS_BETA_NZ_128 vaddps(xmm8, xmm0, xmm8) add(rdi, rcx) CGEMM_INPUT_SCALE_CS_BETA_NZ_128 vaddps(xmm12, xmm0, xmm12) mov(r12, rcx) // reset rcx to current utile of c. vunpcklpd(xmm8, xmm4, xmm0) //a0a1b0b1 a4a4b4b5 //gamma00-10 gamma02-02 vunpckhpd(xmm8, xmm4, xmm2) //a2a3b2b3 a6a7b6b7 //gamma01-11 gamma03-13 vmovups(xmm0, mem(rcx)) // store (gamma00-10) vmovlpd(xmm12, mem(rcx, 16)) // store (gamma20) lea(mem(rcx, rsi, 1), rcx) vmovups(xmm2, mem(rcx)) // store (gamma01-11) vmovhpd(xmm12, mem(rcx, 16)) // store (gamma21) jmp(.SDONE) // jump to end. label(.SBETAZERO) lea(mem(r8, r8, 2), r13) // r13 = 3*rs_a cmp(imm(8), rdi) // set ZF if (8*rs_c) == 8. jz(.SCOLSTORBZ) // jump to column storage case label(.SROWSTORBZ) vmovups(xmm4, mem(rcx)) add(rdi, rcx) vmovups(xmm8, mem(rcx)) add(rdi, rcx) vmovups(xmm12, mem(rcx)) jmp(.SDONE) // jump to end. label(.SCOLSTORBZ) /****3x2 tile going to save into 2x3 tile in C*****/ mov(var(cs_c), rsi) // load cs_c lea(mem(, rsi, 8), rsi) // rsi = cs_c * sizeof(dt) vunpcklpd(xmm8, xmm4, xmm0) //a0a1b0b1 a4a4b4b5 //gamma00-10 gamma02-02 vunpckhpd(xmm8, xmm4, xmm2) //a2a3b2b3 a6a7b6b7 //gamma01-11 gamma03-13 vmovups(xmm0, mem(rcx)) // store (gamma00-10) vmovlpd(xmm12, mem(rcx, 16)) // store (gamma20) lea(mem(rcx, rsi, 1), rcx) vmovups(xmm2, mem(rcx)) // store (gamma01-11) vmovhpd(xmm12, mem(rcx, 16)) // store (gamma21) label(.SDONE) lea(mem(r12, rdi, 2), r12) lea(mem(r12, rdi, 1), r12) // c_ii = r12 += 3*rs_c lea(mem(r14, r8, 2), r14) lea(mem(r14, r8, 1), r14) //a_ii = r14 += 3*rs_a dec(r11) // ii -= 1; jne(.SLOOP3X2I) // iterate again if ii != 0. label(.SRETURN) end_asm( : // output operands (none) : // input operands [m_iter] "m" (m_iter), [k_iter] "m" (k_iter), [k_left] "m" (k_left), [a] "m" (a), [rs_a] "m" (rs_a), [cs_a] "m" (cs_a), [b] "m" (b), [rs_b] "m" (rs_b), [cs_b] "m" (cs_b), [alpha] "m" (alpha), [beta] "m" (beta), [c] "m" (c), [rs_c] "m" (rs_c), [cs_c] "m" (cs_c)/*, [a_next] "m" (a_next), [b_next] "m" (b_next)*/ : // register clobber list "rax", "rbx", "rcx", "rdx", "rsi", "rdi", "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7", "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13", "xmm14", "xmm15", "memory" ) consider_edge_cases: // Handle edge cases in the m dimension, if they exist. if ( m_left ) { const dim_t nr_cur = 2; const dim_t i_edge = m0 - ( dim_t )m_left; scomplex* cij = c + i_edge*rs_c; scomplex* ai = a + i_edge*rs_a; scomplex* bj = b; cgemmsup_ker_ft ker_fps[3] = { NULL, bli_cgemmsup_rv_zen_asm_1x2, bli_cgemmsup_rv_zen_asm_2x2, }; cgemmsup_ker_ft ker_fp = ker_fps[ m_left ]; ker_fp ( conja, conjb, m_left, nr_cur, k0, alpha, ai, rs_a0, cs_a0, bj, rs_b0, cs_b0, beta, cij, rs_c0, cs_c0, data, cntx ); return; } }