diff --git a/driver/others/parameter.c b/driver/others/parameter.c index 36da13369..791e5dc27 100644 --- a/driver/others/parameter.c +++ b/driver/others/parameter.c @@ -524,6 +524,9 @@ void blas_set_parameter(void){ xgemm_p = ((xgemm_p + XGEMM_UNROLL_M - 1)/XGEMM_UNROLL_M) * XGEMM_UNROLL_M; #endif +#ifdef BUILD_BFLOAT16 + sbgemm_r = (((BUFFER_SIZE - ((SBGEMM_P * SBGEMM_Q * 4 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (SBGEMM_Q * 4)) - 15) & ~15; +#endif sgemm_r = (((BUFFER_SIZE - ((SGEMM_P * SGEMM_Q * 4 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (SGEMM_Q * 4)) - 15) & ~15; dgemm_r = (((BUFFER_SIZE - ((DGEMM_P * DGEMM_Q * 8 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (DGEMM_Q * 8)) - 15) & ~15; cgemm_r = (((BUFFER_SIZE - ((CGEMM_P * CGEMM_Q * 8 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (CGEMM_Q * 8)) - 15) & ~15; @@ -629,7 +632,9 @@ void blas_set_parameter(void){ xgemm_p = 16 * (size + 1); #endif +#ifdef BUILD_BFLOAT16 sbgemm_r = (((BUFFER_SIZE - ((SBGEMM_P * SBGEMM_Q * 4 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (SBGEMM_Q * 4)) - 15) & ~15; +#endif sgemm_r = (((BUFFER_SIZE - ((SGEMM_P * SGEMM_Q * 4 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (SGEMM_Q * 4)) - 15) & ~15; dgemm_r = (((BUFFER_SIZE - ((DGEMM_P * DGEMM_Q * 8 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (DGEMM_Q * 8)) - 15) & ~15; cgemm_r = (((BUFFER_SIZE - ((CGEMM_P * CGEMM_Q * 8 + GEMM_OFFSET_A + GEMM_ALIGN) & ~GEMM_ALIGN)) / (CGEMM_Q * 8)) - 15) & ~15; diff --git a/interface/gemm.c b/interface/gemm.c index 6dcc54041..71cc77a1b 100644 --- a/interface/gemm.c +++ b/interface/gemm.c @@ -326,7 +326,7 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANS PRINT_DEBUG_CNAME; -#if !defined(COMPLEX) && !defined(DOUBLE) && defined(USE_SGEMM_KERNEL_DIRECT) +#if !defined(COMPLEX) && !defined(DOUBLE) && !defined(BFLOAT16) && defined(USE_SGEMM_KERNEL_DIRECT) #ifdef DYNAMIC_ARCH if (support_avx512() ) #endif diff --git a/kernel/x86_64/KERNEL.COOPERLAKE b/kernel/x86_64/KERNEL.COOPERLAKE index 6272dd73d..dba94aea8 100644 --- a/kernel/x86_64/KERNEL.COOPERLAKE +++ b/kernel/x86_64/KERNEL.COOPERLAKE @@ -9,3 +9,14 @@ SBGEMM_SMALL_K_TN = sbgemm_small_kernel_tn_cooperlake.c SBGEMM_SMALL_K_B0_TN = sbgemm_small_kernel_tn_cooperlake.c SBGEMM_SMALL_K_TT = sbgemm_small_kernel_tt_cooperlake.c SBGEMM_SMALL_K_B0_TT = sbgemm_small_kernel_tt_cooperlake.c + +SBGEMM_BETA = sgemm_beta_skylakex.c +SBGEMMKERNEL = sbgemm_kernel_16x4_cooperlake.c +SBGEMMINCOPY = sbgemm_ncopy_16_cooperlake.c +SBGEMMITCOPY = sbgemm_tcopy_16_cooperlake.c +SBGEMMONCOPY = sbgemm_ncopy_4_cooperlake.c +SBGEMMOTCOPY = sbgemm_tcopy_4_cooperlake.c +SBGEMMINCOPYOBJ = sbgemm_incopy$(TSUFFIX).$(SUFFIX) +SBGEMMITCOPYOBJ = sbgemm_itcopy$(TSUFFIX).$(SUFFIX) +SBGEMMONCOPYOBJ = sbgemm_oncopy$(TSUFFIX).$(SUFFIX) +SBGEMMOTCOPYOBJ = sbgemm_otcopy$(TSUFFIX).$(SUFFIX) diff --git a/kernel/x86_64/sbgemm_kernel_16x4_cooperlake.c b/kernel/x86_64/sbgemm_kernel_16x4_cooperlake.c new file mode 100644 index 000000000..7af51b6d8 --- /dev/null +++ b/kernel/x86_64/sbgemm_kernel_16x4_cooperlake.c @@ -0,0 +1,499 @@ +/*************************************************************************** +Copyright (c) 2021, The OpenBLAS Project +All rights reserved. +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions are +met: +1. Redistributions of source code must retain the above copyright +notice, this list of conditions and the following disclaimer. +2. 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. +3. Neither the name of the OpenBLAS project 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 OPENBLAS PROJECT 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 +#include "common.h" + +#define VMOVLDUP(addr, zmm) asm("vmovsldup (%1), %0": "=v"(zmm): "r"(addr)) +#define VMOVHDUP(addr, zmm) asm("vmovshdup (%1), %0": "=v"(zmm): "r"(addr)) +#define BROADCAST64(base, step, n, offset, zmm) \ + if (n == 0) asm("vbroadcastsd %c2(%1), %0": "=v"(zmm): "r"(base), "n"(offset*2)); \ + else asm("vbroadcastsd %c4(%1, %2, %c3), %0": "=v"(zmm): "r"(base), "r"(step), "n"(n*2), "n"(offset*2)) + +#define DECLARE_A_PAIR(A) \ + __m512i A_lo_##A; __m512i A_hi_##A; + +#define LOAD_A_PAIR(A) \ + VMOVLDUP(ptr_a##A, A_lo_##A); \ + VMOVHDUP(ptr_a##A, A_hi_##A); + +#define MASK_LOAD_A_PAIR(A) { \ + __m512 tmp = _mm512_maskz_loadu_ps(mmask, ptr_a##A); \ + A_lo_##A = (__m512i) _mm512_moveldup_ps(tmp); \ + A_hi_##A = (__m512i) _mm512_movehdup_ps(tmp); \ +} + +#define LOAD_A_PAIR_TAIL(A) { \ + __m256i ymm = _mm256_loadu_si256((void *)ptr_a##A); \ + __m512 zmm = (__m512) _mm512_cvtepu16_epi32(ymm); \ + A_lo_##A = (__m512i) _mm512_moveldup_ps(zmm); \ + A_hi_##A = (__m512i) _mm512_movehdup_ps(zmm); \ +} + +#define MASK_LOAD_A_PAIR_TAIL(A) { \ + __m256i ymm = _mm256_maskz_loadu_epi16(mmask, ptr_a##A); \ + __m512 zmm = (__m512) _mm512_cvtepu16_epi32(ymm); \ + A_lo_##A = (__m512i) _mm512_moveldup_ps(zmm); \ + A_hi_##A = (__m512i) _mm512_movehdup_ps(zmm); \ +} + +#define DECLARE_B_PAIR() \ + __m512i B_lo; __m512i B_hi; + +#define PREFETCH_B_STEP 32 +#define PREFETCH_B(Bx, By) \ + if (By == 0) asm("prefetcht0 %c1(%0)": : "r"(ptr_b##Bx), "n"(PREFETCH_B_STEP * 2)); \ + else asm("prefetcht0 %c3(%0, %1, %c2)": : "r"(ptr_b##Bx), "r"(n_blksize), "n"(By*2), "n"(PREFETCH_B_STEP * 2)) + +#define BROADCAST_B_PAIR(Bx, By) \ + BROADCAST64(ptr_b##Bx, n_blksize, By, 0, B_lo); \ + BROADCAST64(ptr_b##Bx, n_blksize, By, 4, B_hi); + +#define MASK_BROADCAST_B_PAIR(Bx, x) {\ + __m128 xmm = _mm_maskz_loadu_ps(nmask, ptr_b##Bx); \ + B_lo = (__m512i) _mm512_broadcastsd_pd((__m128d) xmm); \ + B_hi = (__m512i) _mm512_broadcastsd_pd(_mm_permute_pd((__m128d) xmm, 0x1)); \ +} + +#define BROADCAST_B_PAIR_TAIL(Bx, By) {\ + __m128i xmm = (__m128i) _mm_load_sd((double *)(ptr_b##Bx + n_blksize * By)); \ + xmm = _mm_cvtepu16_epi32(xmm); \ + B_lo = _mm512_broadcast_i32x2(xmm); \ + B_hi = _mm512_broadcast_i32x2((__m128i) _mm_permute_pd((__m128d) xmm, 0x1)); \ +} + +#define MASK_BROADCAST_B_PAIR_TAIL(Bx, By) {\ + __m128i xmm = _mm_maskz_loadu_epi16(nmask, ptr_b##Bx + n_blksize * By); \ + xmm = _mm_cvtepu16_epi32(xmm); \ + B_lo = _mm512_broadcast_i32x2(xmm); \ + B_hi = _mm512_broadcast_i32x2((__m128i) _mm_permute_pd((__m128d) xmm, 0x1)); \ +} + +#define DECLARE_RESULT_4X(A, Bx, By) \ + __m512 result_00_##A##Bx##By = _mm512_setzero_ps(); \ + __m512 result_01_##A##Bx##By = _mm512_setzero_ps(); \ + __m512 result_10_##A##Bx##By = _mm512_setzero_ps(); \ + __m512 result_11_##A##Bx##By = _mm512_setzero_ps(); + +#define FMA(a, b, r) r = _mm512_dpbf16_ps(r, (__m512bh)a, (__m512bh)b) + +#define MATMUL_4X(A, Bx, By) \ + FMA(A_lo_##A, B_lo, result_00_##A##Bx##By); \ + FMA(A_hi_##A, B_lo, result_01_##A##Bx##By); \ + FMA(A_lo_##A, B_hi, result_10_##A##Bx##By); \ + FMA(A_hi_##A, B_hi, result_11_##A##Bx##By); + +#define _STORE_C_2nx16(addr, val0, val1) \ + asm("vfmadd213ps (%1), %2, %0": "+v"(val0) : "r"(addr), "v"(alpha_512)); \ + asm("vfmadd213ps (%1, %3, 4), %2, %0": "+v"(val1) : "r"(addr), "v"(alpha_512), "r"(ldc)); \ + asm("vmovups %0, (%1)": : "v"(val0), "r"(addr)); \ + asm("vmovups %0, (%1, %2, 4)": : "v"(val1), "r"(addr), "r"(ldc)) + +#define _MASK_STORE_C_2nx16(addr, val0, val1) \ + asm("vfmadd213ps (%1), %2, %0 %{%3%} ": "+v"(val0) : "r"(addr), "v"(alpha_512), "k"(mmask)); \ + asm("vfmadd213ps (%1, %3, 4), %2, %0 %{%4%}": "+v"(val1) : "r"(addr), "v"(alpha_512), "r"(ldc), "k"(mmask)); \ + asm("vmovups %0, (%1) %{%2%}": : "v"(val0), "r"(addr), "k"(mmask)); \ + asm("vmovups %0, (%1, %2, 4) %{%3%}": : "v"(val1), "r"(addr), "r"(ldc), "k"(mmask)) + +#define _REORDER_C_2X(result_0, result_1) { \ + __m512 tmp0, tmp1; \ + tmp0 = _mm512_unpacklo_ps(result_0, result_1); \ + tmp1 = _mm512_unpackhi_ps(result_0, result_1); \ + result_0 = (__m512) _mm512_unpacklo_pd((__m512d) tmp0, (__m512d) tmp1); \ + result_1 = (__m512) _mm512_unpackhi_pd((__m512d) tmp0, (__m512d) tmp1); \ +} + +#define _STORE_2X(ptr_c, result_0, result_1) {\ + _REORDER_C_2X(result_0, result_1) \ + _STORE_C_2nx16(ptr_c, result_0, result_1); \ + ptr_c += ldc * 2; \ +} + +#define _MASK_STORE_2X(ptr_c, result_0, result_1) {\ + _REORDER_C_2X(result_0, result_1) \ + _MASK_STORE_C_2nx16(ptr_c, result_0, result_1); \ + ptr_c += ldc * 2; \ +} + +#define STORE_4X(A, Bx, By) { \ + _STORE_2X(ptr_c##A, result_00_##A##Bx##By, result_01_##A##Bx##By); \ + _STORE_2X(ptr_c##A, result_10_##A##Bx##By, result_11_##A##Bx##By); \ +} + +#define MASK_STORE_4X(A, Bx, By) { \ + _MASK_STORE_2X(ptr_c##A, result_00_##A##Bx##By, result_01_##A##Bx##By); \ + _MASK_STORE_2X(ptr_c##A, result_10_##A##Bx##By, result_11_##A##Bx##By); \ +} + +#define _STORE_C_16(addr, val0) \ + asm("vfmadd213ps (%1), %2, %0": "+v"(val0) : "r"(addr), "v"(alpha_512)); \ + asm("vmovups %0, (%1)": : "v"(val0), "r"(addr)); + +#define _MASK_STORE_C_16(addr, val0) \ + asm("vfmadd213ps (%1), %2, %0 %{%3%} ": "+v"(val0) : "r"(addr), "v"(alpha_512), "k"(mmask)); \ + asm("vmovups %0, (%1) %{%2%}": : "v"(val0), "r"(addr), "k"(mmask)); + +#define N_STORE_4X(A, Bx, By) { \ + _REORDER_C_2X(result_00_##A##Bx##By, result_01_##A##Bx##By); \ + _REORDER_C_2X(result_10_##A##Bx##By, result_11_##A##Bx##By); \ + switch(n_count) { \ + case 3: _STORE_C_16(ptr_c + ldc * 2, result_10_##A##Bx##By); \ + case 2: _STORE_C_16(ptr_c + ldc * 1, result_01_##A##Bx##By); \ + case 1: _STORE_C_16(ptr_c + ldc * 0, result_00_##A##Bx##By); \ + } \ + ptr_c##A += ldc * n_count; \ +} + +#define N_MASK_STORE_4X(A, Bx, By) { \ + _REORDER_C_2X(result_00_##A##Bx##By, result_01_##A##Bx##By); \ + _REORDER_C_2X(result_10_##A##Bx##By, result_11_##A##Bx##By); \ + switch(n_count) { \ + case 3: _MASK_STORE_C_16(ptr_c + ldc * 2, result_10_##A##Bx##By); \ + case 2: _MASK_STORE_C_16(ptr_c + ldc * 1, result_01_##A##Bx##By); \ + case 1: _MASK_STORE_C_16(ptr_c + ldc * 0, result_00_##A##Bx##By); \ + } \ + ptr_c##A += ldc * n_count; \ +} + + +int CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, IFLOAT * A, IFLOAT * B, FLOAT * C, BLASLONG ldc) +{ + IFLOAT *ptr_a = A, *ptr_b = B; + IFLOAT *ptr_b0, *ptr_b1; + IFLOAT *ptr_a0, *ptr_a1; + FLOAT *ptr_c = C; + FLOAT *ptr_c0, *ptr_c1; + BLASLONG n_count = n; + BLASLONG m_count, k_count; + BLASLONG n_blksize = 4 * k; + BLASLONG cn_offset = 0; + __m512 alpha_512 = _mm512_broadcastss_ps(_mm_load_ss(&alpha)); + + for (; n_count > 23; n_count -= 24) { + IFLOAT *ptr_b00 = ptr_b; + IFLOAT *ptr_b10 = ptr_b + n_blksize * 3; + ptr_a0 = ptr_a; + ptr_c = C + cn_offset * ldc; + m_count = m; + for (; m_count > 15; m_count -= 16) { + ptr_b0 = ptr_b00; + ptr_b1 = ptr_b10; + DECLARE_A_PAIR(0); + DECLARE_B_PAIR(); + DECLARE_RESULT_4X(0, 0, 0); DECLARE_RESULT_4X(0, 0, 1); DECLARE_RESULT_4X(0, 0, 2); + DECLARE_RESULT_4X(0, 1, 0); DECLARE_RESULT_4X(0, 1, 1); DECLARE_RESULT_4X(0, 1, 2); + k_count = k; + for (; k_count > 3; k_count -=4) { + LOAD_A_PAIR(0); + _mm_prefetch(ptr_a0 + 128, _MM_HINT_T0); + ptr_a0 += 16 * 2; + BROADCAST_B_PAIR(0, 0); PREFETCH_B(0, 0); MATMUL_4X(0, 0, 0); + BROADCAST_B_PAIR(0, 1); PREFETCH_B(0, 1); MATMUL_4X(0, 0, 1); + BROADCAST_B_PAIR(0, 2); PREFETCH_B(0, 2); MATMUL_4X(0, 0, 2); + ptr_b0 += 4 * 2; + BROADCAST_B_PAIR(1, 0); PREFETCH_B(1, 0); MATMUL_4X(0, 1, 0); + BROADCAST_B_PAIR(1, 1); PREFETCH_B(1, 1); MATMUL_4X(0, 1, 1); + BROADCAST_B_PAIR(1, 2); PREFETCH_B(1, 2); MATMUL_4X(0, 1, 2); + ptr_b1 += 4 * 2; + + LOAD_A_PAIR(0); + _mm_prefetch(ptr_a0 + 128, _MM_HINT_T0); + ptr_a0 += 16 * 2; + BROADCAST_B_PAIR(0, 0); MATMUL_4X(0, 0, 0); + BROADCAST_B_PAIR(0, 1); MATMUL_4X(0, 0, 1); + BROADCAST_B_PAIR(0, 2); MATMUL_4X(0, 0, 2); + ptr_b0 += 4 * 2; + BROADCAST_B_PAIR(1, 0); MATMUL_4X(0, 1, 0); + BROADCAST_B_PAIR(1, 1); MATMUL_4X(0, 1, 1); + BROADCAST_B_PAIR(1, 2); MATMUL_4X(0, 1, 2); + ptr_b1 += 4 * 2; + } + for (; k_count > 1; k_count -=2) { + LOAD_A_PAIR(0); + ptr_a0 += 16 * 2; + BROADCAST_B_PAIR(0, 0); MATMUL_4X(0, 0, 0); + BROADCAST_B_PAIR(0, 1); MATMUL_4X(0, 0, 1); + BROADCAST_B_PAIR(0, 2); MATMUL_4X(0, 0, 2); + ptr_b0 += 4 * 2; + BROADCAST_B_PAIR(1, 0); MATMUL_4X(0, 1, 0); + BROADCAST_B_PAIR(1, 1); MATMUL_4X(0, 1, 1); + BROADCAST_B_PAIR(1, 2); MATMUL_4X(0, 1, 2); + ptr_b1 += 4 * 2; + } + if (k_count > 0) { + LOAD_A_PAIR_TAIL(0); + ptr_a0 += 16; + BROADCAST_B_PAIR_TAIL(0, 0); MATMUL_4X(0, 0, 0); + BROADCAST_B_PAIR_TAIL(0, 1); MATMUL_4X(0, 0, 1); + BROADCAST_B_PAIR_TAIL(0, 2); MATMUL_4X(0, 0, 2); + ptr_b0 += 4; + BROADCAST_B_PAIR_TAIL(1, 0); MATMUL_4X(0, 1, 0); + BROADCAST_B_PAIR_TAIL(1, 1); MATMUL_4X(0, 1, 1); + BROADCAST_B_PAIR_TAIL(1, 2); MATMUL_4X(0, 1, 2); + ptr_b1 += 4; + } + ptr_c0 = ptr_c; + STORE_4X(0, 0, 0); STORE_4X(0, 0, 1); STORE_4X(0, 0, 2); + STORE_4X(0, 1, 0); STORE_4X(0, 1, 1); STORE_4X(0, 1, 2); + ptr_c += 16; + } + if (m_count > 0) { + __mmask16 mmask = (1UL << m_count) - 1; + ptr_b0 = ptr_b00; + ptr_b1 = ptr_b10; + DECLARE_A_PAIR(0); + DECLARE_B_PAIR(); + DECLARE_RESULT_4X(0, 0, 0); DECLARE_RESULT_4X(0, 0, 1); DECLARE_RESULT_4X(0, 0, 2); + DECLARE_RESULT_4X(0, 1, 0); DECLARE_RESULT_4X(0, 1, 1); DECLARE_RESULT_4X(0, 1, 2); + for (k_count = k; k_count > 1; k_count -=2) { + MASK_LOAD_A_PAIR(0); + ptr_a0 += m_count * 2; + BROADCAST_B_PAIR(0, 0); MATMUL_4X(0, 0, 0); + BROADCAST_B_PAIR(0, 1); MATMUL_4X(0, 0, 1); + BROADCAST_B_PAIR(0, 2); MATMUL_4X(0, 0, 2); + ptr_b0 += 4 * 2; + BROADCAST_B_PAIR(1, 0); MATMUL_4X(0, 1, 0); + BROADCAST_B_PAIR(1, 1); MATMUL_4X(0, 1, 1); + BROADCAST_B_PAIR(1, 2); MATMUL_4X(0, 1, 2); + ptr_b1 += 4 * 2; + } + if (k_count > 0) { + MASK_LOAD_A_PAIR_TAIL(0); + ptr_a0 += m_count; + BROADCAST_B_PAIR_TAIL(0, 0); MATMUL_4X(0, 0, 0); + BROADCAST_B_PAIR_TAIL(0, 1); MATMUL_4X(0, 0, 1); + BROADCAST_B_PAIR_TAIL(0, 2); MATMUL_4X(0, 0, 2); + ptr_b0 += 4; + BROADCAST_B_PAIR_TAIL(1, 0); MATMUL_4X(0, 1, 0); + BROADCAST_B_PAIR_TAIL(1, 1); MATMUL_4X(0, 1, 1); + BROADCAST_B_PAIR_TAIL(1, 2); MATMUL_4X(0, 1, 2); + ptr_b1 += 4; + } + ptr_c0 = ptr_c; + MASK_STORE_4X(0, 0, 0); MASK_STORE_4X(0, 0, 1); MASK_STORE_4X(0, 0, 2); + MASK_STORE_4X(0, 1, 0); MASK_STORE_4X(0, 1, 1); MASK_STORE_4X(0, 1, 2); + ptr_c += m_count; + } + ptr_b += 24 * k; + cn_offset += 24; + } + for (; n_count > 11; n_count -= 12) { + IFLOAT *ptr_b00 = ptr_b; + ptr_a0 = ptr_a; + ptr_a1 = ptr_a + 16 * k; + ptr_c = C + cn_offset * ldc; + m_count = m; + for (; m_count > 31; m_count -= 32) { + ptr_b0 = ptr_b00; + DECLARE_A_PAIR(0); DECLARE_A_PAIR(1); + DECLARE_B_PAIR(); + DECLARE_RESULT_4X(0, 0, 0); DECLARE_RESULT_4X(0, 0, 1); DECLARE_RESULT_4X(0, 0, 2); + DECLARE_RESULT_4X(1, 0, 0); DECLARE_RESULT_4X(1, 0, 1); DECLARE_RESULT_4X(1, 0, 2); + for (k_count = k; k_count > 1; k_count -=2) { + LOAD_A_PAIR(0); LOAD_A_PAIR(1); + ptr_a0 += 16 * 2; + ptr_a1 += 16 * 2; + BROADCAST_B_PAIR(0, 0); MATMUL_4X(0, 0, 0); MATMUL_4X(1, 0, 0); + BROADCAST_B_PAIR(0, 1); MATMUL_4X(0, 0, 1); MATMUL_4X(1, 0, 1); + BROADCAST_B_PAIR(0, 2); MATMUL_4X(0, 0, 2); MATMUL_4X(1, 0, 2); + ptr_b0 += 4 * 2; + } + if (k_count > 0) { + LOAD_A_PAIR_TAIL(0); LOAD_A_PAIR_TAIL(1); + ptr_a0 += 16; + ptr_a1 += 16; + BROADCAST_B_PAIR_TAIL(0, 0); MATMUL_4X(0, 0, 0); MATMUL_4X(1, 0, 0); + BROADCAST_B_PAIR_TAIL(0, 1); MATMUL_4X(0, 0, 1); MATMUL_4X(1, 0, 1); + BROADCAST_B_PAIR_TAIL(0, 2); MATMUL_4X(0, 0, 2); MATMUL_4X(1, 0, 2); + ptr_b0 += 4; + } + ptr_c0 = ptr_c; + ptr_c1 = ptr_c + 16; + STORE_4X(0, 0, 0); STORE_4X(1, 0, 0); + STORE_4X(0, 0, 1); STORE_4X(1, 0, 1); + STORE_4X(0, 0, 2); STORE_4X(1, 0, 2); + ptr_c += 16 * 2; + ptr_a0 = ptr_a1; + ptr_a1 = ptr_a0 + 16 * k; + } + for (; m_count > 15; m_count -= 16) { + ptr_b0 = ptr_b00; + DECLARE_A_PAIR(0); + DECLARE_B_PAIR(); + DECLARE_RESULT_4X(0, 0, 0); DECLARE_RESULT_4X(0, 0, 1); DECLARE_RESULT_4X(0, 0, 2); + for (k_count = k; k_count > 1; k_count -=2) { + LOAD_A_PAIR(0); + ptr_a0 += 16 * 2; + BROADCAST_B_PAIR(0, 0); MATMUL_4X(0, 0, 0); + BROADCAST_B_PAIR(0, 1); MATMUL_4X(0, 0, 1); + BROADCAST_B_PAIR(0, 2); MATMUL_4X(0, 0, 2); + ptr_b0 += 4 * 2; + } + if (k_count > 0) { + LOAD_A_PAIR_TAIL(0); + ptr_a0 += 16; + BROADCAST_B_PAIR_TAIL(0, 0); MATMUL_4X(0, 0, 0); + BROADCAST_B_PAIR_TAIL(0, 1); MATMUL_4X(0, 0, 1); + BROADCAST_B_PAIR_TAIL(0, 2); MATMUL_4X(0, 0, 2); + ptr_b0 += 4; + } + ptr_c0 = ptr_c; + STORE_4X(0, 0, 0); STORE_4X(0, 0, 1); STORE_4X(0, 0, 2); + ptr_c += 16; + } + if (m_count > 0) { + __mmask16 mmask = (1UL << m_count) - 1; + ptr_b0 = ptr_b00; + DECLARE_A_PAIR(0); + DECLARE_B_PAIR(); + DECLARE_RESULT_4X(0, 0, 0); DECLARE_RESULT_4X(0, 0, 1); DECLARE_RESULT_4X(0, 0, 2); + for (k_count = k; k_count > 1; k_count -=2) { + MASK_LOAD_A_PAIR(0); + ptr_a0 += m_count * 2; + BROADCAST_B_PAIR(0, 0); MATMUL_4X(0, 0, 0); + BROADCAST_B_PAIR(0, 1); MATMUL_4X(0, 0, 1); + BROADCAST_B_PAIR(0, 2); MATMUL_4X(0, 0, 2); + ptr_b0 += 4 * 2; + } + if (k_count > 0) { + MASK_LOAD_A_PAIR_TAIL(0); + ptr_a0 += m_count; + BROADCAST_B_PAIR_TAIL(0, 0); MATMUL_4X(0, 0, 0); + BROADCAST_B_PAIR_TAIL(0, 1); MATMUL_4X(0, 0, 1); + BROADCAST_B_PAIR_TAIL(0, 2); MATMUL_4X(0, 0, 2); + ptr_b0 += 4; + } + ptr_c0 = ptr_c; + MASK_STORE_4X(0, 0, 0); MASK_STORE_4X(0, 0, 1); MASK_STORE_4X(0, 0, 2); + ptr_c += m_count; + } + ptr_b += 12 * k; + cn_offset += 12; + } + for (; n_count > 3; n_count -= 4) { + IFLOAT *ptr_b00 = ptr_b; + ptr_a0 = ptr_a; + ptr_c = C + cn_offset * ldc; + m_count = m; + for (; m_count > 15; m_count -= 16) { + ptr_b0 = ptr_b00; + DECLARE_A_PAIR(0); + DECLARE_B_PAIR(); + DECLARE_RESULT_4X(0, 0, 0); + for (k_count = k; k_count > 1; k_count -=2) { + LOAD_A_PAIR(0); + BROADCAST_B_PAIR(0, 0); MATMUL_4X(0, 0, 0); + ptr_b0 += 4 * 2; + ptr_a0 += 16 * 2; + } + if (k_count > 0) { + LOAD_A_PAIR_TAIL(0); + BROADCAST_B_PAIR_TAIL(0, 0); MATMUL_4X(0, 0, 0); + ptr_b0 += 4; + ptr_a0 += 16; + } + ptr_c0 = ptr_c; + STORE_4X(0, 0, 0); + ptr_c += 16; + } + if (m_count > 0) { + __mmask16 mmask = (1UL << m_count) - 1; + ptr_b0 = ptr_b00; + DECLARE_A_PAIR(0); + DECLARE_B_PAIR(); + DECLARE_RESULT_4X(0, 0, 0); + for (k_count = k; k_count > 1; k_count -=2) { + MASK_LOAD_A_PAIR(0); + BROADCAST_B_PAIR(0, 0); MATMUL_4X(0, 0, 0); + ptr_b0 += 4 * 2; + ptr_a0 += m_count * 2; + } + if (k_count > 0) { + MASK_LOAD_A_PAIR_TAIL(0); + BROADCAST_B_PAIR_TAIL(0, 0); MATMUL_4X(0, 0, 0); + ptr_b0 += 4; + ptr_a0 += m_count; + } + ptr_c0 = ptr_c; + MASK_STORE_4X(0, 0, 0); + ptr_c += m_count; + } + ptr_b += 4 * k; + cn_offset += 4; + } + if (n_count > 0) { + __mmask8 nmask = (1UL << n_count) - 1; + IFLOAT *ptr_b00 = ptr_b; + ptr_a0 = ptr_a; + ptr_c = C + cn_offset * ldc; + m_count = m; + for (; m_count > 15; m_count -= 16) { + ptr_b0 = ptr_b00; + DECLARE_A_PAIR(0); + DECLARE_B_PAIR(); + DECLARE_RESULT_4X(0, 0, 0); + for (k_count = k; k_count > 1; k_count -=2) { + LOAD_A_PAIR(0); + MASK_BROADCAST_B_PAIR(0, 0); MATMUL_4X(0, 0, 0); + ptr_b0 += n_count * 2; + ptr_a0 += 16 * 2; + } + if (k_count > 0) { + LOAD_A_PAIR_TAIL(0); + MASK_BROADCAST_B_PAIR_TAIL(0, 0); MATMUL_4X(0, 0, 0); + ptr_b0 += n_count; + ptr_a0 += 16; + } + ptr_c0 = ptr_c; + N_STORE_4X(0, 0, 0); + ptr_c += 16; + } + if (m_count > 0) { + __mmask16 mmask = (1UL << m_count) - 1; + ptr_b0 = ptr_b00; + DECLARE_A_PAIR(0); + DECLARE_B_PAIR(); + DECLARE_RESULT_4X(0, 0, 0); + for (k_count = k; k_count > 1; k_count -=2) { + MASK_LOAD_A_PAIR(0); + MASK_BROADCAST_B_PAIR(0, 0); MATMUL_4X(0, 0, 0); + ptr_b0 += n_count * 2; + ptr_a0 += m_count * 2; + } + if (k_count > 0) { + MASK_LOAD_A_PAIR_TAIL(0); + MASK_BROADCAST_B_PAIR_TAIL(0, 0); MATMUL_4X(0, 0, 0); + ptr_b0 += n_count; + ptr_a0 += m_count; + } + ptr_c0 = ptr_c; + N_MASK_STORE_4X(0, 0, 0); + ptr_c += m_count; + } + } + return 0; +} diff --git a/kernel/x86_64/sbgemm_ncopy_16_cooperlake.c b/kernel/x86_64/sbgemm_ncopy_16_cooperlake.c new file mode 100644 index 000000000..95ed82d7c --- /dev/null +++ b/kernel/x86_64/sbgemm_ncopy_16_cooperlake.c @@ -0,0 +1,353 @@ +/*************************************************************************** +Copyright (c) 2021, The OpenBLAS Project +All rights reserved. +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions are +met: +1. Redistributions of source code must retain the above copyright +notice, this list of conditions and the following disclaimer. +2. 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. +3. Neither the name of the OpenBLAS project 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 OPENBLAS PROJECT 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 +#include +#include "common.h" + +#define _MM512_SHUFFLE_i32(result, in1, in2, imm8) \ + asm("vshufps %3, %2, %1, %0": "=v"(result): "v"(in1), "v"(in2), "N"(imm8)) + +#define REORDER_8x32(t0, t1, t2, t3, t4, t5, t6, t7) { \ + __m512i v; \ + t0 = _mm512_unpacklo_epi32(r0, r1); \ + t1 = _mm512_unpackhi_epi32(r0, r1); \ + t2 = _mm512_unpacklo_epi32(r2, r3); \ + t3 = _mm512_unpackhi_epi32(r2, r3); \ + t4 = _mm512_unpacklo_epi32(r4, r5); \ + t5 = _mm512_unpackhi_epi32(r4, r5); \ + t6 = _mm512_unpacklo_epi32(r6, r7); \ + t7 = _mm512_unpackhi_epi32(r6, r7); \ + _MM512_SHUFFLE_i32(v, t0, t2, 0x4E); \ + r0 = _mm512_mask_blend_epi32(kc, t0, v); \ + r1 = _mm512_mask_blend_epi32(k3, t2, v); \ + _MM512_SHUFFLE_i32(v, t1, t3, 0x4E); \ + r2 = _mm512_mask_blend_epi32(kc, t1, v); \ + r3 = _mm512_mask_blend_epi32(k3, t3, v); \ + _MM512_SHUFFLE_i32(v, t4, t6, 0x4E); \ + r4 = _mm512_mask_blend_epi32(kc, t4, v); \ + r5 = _mm512_mask_blend_epi32(k3, t6, v); \ + _MM512_SHUFFLE_i32(v, t5, t7, 0x4E); \ + r6 = _mm512_mask_blend_epi32(kc, t5, v); \ + r7 = _mm512_mask_blend_epi32(k3, t7, v); \ + t0 = _mm512_permutex2var_epi32(r0, idx_lo, r4); \ + t1 = _mm512_permutex2var_epi32(r1, idx_lo, r5); \ + t2 = _mm512_permutex2var_epi32(r2, idx_lo, r6); \ + t3 = _mm512_permutex2var_epi32(r3, idx_lo, r7); \ + t4 = _mm512_permutex2var_epi32(r0, idx_hi, r4); \ + t5 = _mm512_permutex2var_epi32(r1, idx_hi, r5); \ + t6 = _mm512_permutex2var_epi32(r2, idx_hi, r6); \ + t7 = _mm512_permutex2var_epi32(r3, idx_hi, r7); \ +} + +#define STORE_512_LO(x) \ + v = _mm512_permutex2var_epi64(t0##x, idx_lo2, t1##x); \ + _mm512_storeu_si512(boffset0 + x*32, v); + +#define STORE_512_HI(x) \ + v = _mm512_permutex2var_epi64(t0##x, idx_hi2, t1##x); \ + _mm512_storeu_si512(boffset0 + (x + 8)*32, v); + +#define MASK_STORE_512_LO(x) \ + v = _mm512_permutex2var_epi64(t0##x, idx_lo2, t1##x); \ + _mm512_mask_storeu_epi32(boffset0 + 2*x*remain_n, nmask, v); + +#define MASK_STORE_512_HI(x) \ + v = _mm512_permutex2var_epi64(t0##x, idx_hi2, t1##x); \ + _mm512_mask_storeu_epi32(boffset0 + 2*(x + 8)*remain_n, nmask, v); + +#define STORE_512(x, y) {\ + __m512i v; \ + if (x == 0) { STORE_512_LO(y); } \ + else { STORE_512_HI(y); } \ +} + +#define MASK_STORE_512(x, y) {\ + __m512i v; \ + if (x == 0) { MASK_STORE_512_LO(y); } \ + else { MASK_STORE_512_HI(y); } \ +} + +#define SET_TAIL(y, x) {\ + if (y == 0) tail = _mm512_permutex2var_epi64(t0##x, idx_lo2, t1##x); \ + else tail = _mm512_permutex2var_epi64(t0##x, idx_hi2, t1##x); \ +} + +#define GET_TAIL() \ + switch (n_store + 1) { \ + case 16: SET_TAIL(1, 7); break; \ + case 15: SET_TAIL(1, 6); break; \ + case 14: SET_TAIL(1, 5); break; \ + case 13: SET_TAIL(1, 4); break; \ + case 12: SET_TAIL(1, 3); break; \ + case 11: SET_TAIL(1, 2); break; \ + case 10: SET_TAIL(1, 1); break; \ + case 9: SET_TAIL(1, 0); break; \ + case 8: SET_TAIL(0, 7); break; \ + case 7: SET_TAIL(0, 6); break; \ + case 6: SET_TAIL(0, 5); break; \ + case 5: SET_TAIL(0, 4); break; \ + case 4: SET_TAIL(0, 3); break; \ + case 3: SET_TAIL(0, 2); break; \ + case 2: SET_TAIL(0, 1); break; \ + case 1: SET_TAIL(0, 0); break; \ + } + + +int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){ + BLASLONG i, j; + + IFLOAT *boffset0; + IFLOAT *aoffset; + IFLOAT *aoffset00, *aoffset01, *aoffset02, *aoffset03, *aoffset04, *aoffset05, *aoffset06, *aoffset07; + IFLOAT *aoffset10, *aoffset11, *aoffset12, *aoffset13, *aoffset14, *aoffset15, *aoffset16, *aoffset17; + aoffset = a; + boffset0 = b; + + BLASLONG n16 = n & ~15; + BLASLONG m32 = m & ~31; + + int permute_table[] = { + 0x0, 0x1, 0x2, 0x3, 0x10, 0x11, 0x12, 0x13, 0x8, 0x9, 0xa, 0xb, 0x18, 0x19, 0x1a, 0x1b, + 0x4, 0x5, 0x6, 0x7, 0x14, 0x15, 0x16, 0x17, 0xc, 0xd, 0xe, 0xf, 0x1c, 0x1d, 0x1e, 0x1f, + }; + u_int64_t permute_table2[] = { + 0x00, 0x01, 0x02, 0x03, 8|0x0, 8|0x1, 8|0x2, 8|0x3, + 0x04, 0x05, 0x06, 0x07, 8|0x4, 8|0x5, 8|0x6, 8|0x7, + }; + __m512i idx_lo = _mm512_loadu_si512(permute_table); + __m512i idx_hi = _mm512_loadu_si512(permute_table + 16); + __m512i idx_lo2 = _mm512_loadu_si512(permute_table2); + __m512i idx_hi2 = _mm512_loadu_si512(permute_table2 + 8); + __mmask16 kc = 0xcccc; + __mmask16 k3 = 0x3333; + __m512i r0, r1, r2, r3, r4, r5, r6, r7; + __m512i t00, t01, t02, t03, t04, t05, t06, t07; + __m512i t10, t11, t12, t13, t14, t15, t16, t17; + + for (j = 0; j < n16; j += 16) { + aoffset00 = aoffset; + aoffset01 = aoffset00 + lda; + aoffset02 = aoffset01 + lda; + aoffset03 = aoffset02 + lda; + aoffset04 = aoffset03 + lda; + aoffset05 = aoffset04 + lda; + aoffset06 = aoffset05 + lda; + aoffset07 = aoffset06 + lda; + aoffset10 = aoffset07 + lda; + aoffset11 = aoffset10 + lda; + aoffset12 = aoffset11 + lda; + aoffset13 = aoffset12 + lda; + aoffset14 = aoffset13 + lda; + aoffset15 = aoffset14 + lda; + aoffset16 = aoffset15 + lda; + aoffset17 = aoffset16 + lda; + aoffset += 16 * lda; + for (i = 0; i < m32; i += 32) { + r0 = _mm512_loadu_si512(aoffset00 + i); + r1 = _mm512_loadu_si512(aoffset01 + i); + r2 = _mm512_loadu_si512(aoffset02 + i); + r3 = _mm512_loadu_si512(aoffset03 + i); + r4 = _mm512_loadu_si512(aoffset04 + i); + r5 = _mm512_loadu_si512(aoffset05 + i); + r6 = _mm512_loadu_si512(aoffset06 + i); + r7 = _mm512_loadu_si512(aoffset07 + i); + REORDER_8x32(t00, t01, t02, t03, t04, t05, t06, t07); + r0 = _mm512_loadu_si512(aoffset10 + i); + r1 = _mm512_loadu_si512(aoffset11 + i); + r2 = _mm512_loadu_si512(aoffset12 + i); + r3 = _mm512_loadu_si512(aoffset13 + i); + r4 = _mm512_loadu_si512(aoffset14 + i); + r5 = _mm512_loadu_si512(aoffset15 + i); + r6 = _mm512_loadu_si512(aoffset16 + i); + r7 = _mm512_loadu_si512(aoffset17 + i); + REORDER_8x32(t10, t11, t12, t13, t14, t15, t16, t17); + STORE_512(0, 0); STORE_512(0, 1); STORE_512(0, 2); STORE_512(0, 3); + STORE_512(0, 4); STORE_512(0, 5); STORE_512(0, 6); STORE_512(0, 7); + STORE_512(1, 0); STORE_512(1, 1); STORE_512(1, 2); STORE_512(1, 3); + STORE_512(1, 4); STORE_512(1, 5); STORE_512(1, 6); STORE_512(1, 7); + boffset0 += 16 * 32; + } + if (i < m) { + int remain_m = m - i; + __mmask32 mmask = (1UL << remain_m) - 1; + r0 = _mm512_maskz_loadu_epi16(mmask, aoffset00 + i); + r1 = _mm512_maskz_loadu_epi16(mmask, aoffset01 + i); + r2 = _mm512_maskz_loadu_epi16(mmask, aoffset02 + i); + r3 = _mm512_maskz_loadu_epi16(mmask, aoffset03 + i); + r4 = _mm512_maskz_loadu_epi16(mmask, aoffset04 + i); + r5 = _mm512_maskz_loadu_epi16(mmask, aoffset05 + i); + r6 = _mm512_maskz_loadu_epi16(mmask, aoffset06 + i); + r7 = _mm512_maskz_loadu_epi16(mmask, aoffset07 + i); + REORDER_8x32(t00, t01, t02, t03, t04, t05, t06, t07); + r0 = _mm512_maskz_loadu_epi16(mmask, aoffset10 + i); + r1 = _mm512_maskz_loadu_epi16(mmask, aoffset11 + i); + r2 = _mm512_maskz_loadu_epi16(mmask, aoffset12 + i); + r3 = _mm512_maskz_loadu_epi16(mmask, aoffset13 + i); + r4 = _mm512_maskz_loadu_epi16(mmask, aoffset14 + i); + r5 = _mm512_maskz_loadu_epi16(mmask, aoffset15 + i); + r6 = _mm512_maskz_loadu_epi16(mmask, aoffset16 + i); + r7 = _mm512_maskz_loadu_epi16(mmask, aoffset17 + i); + REORDER_8x32(t10, t11, t12, t13, t14, t15, t16, t17); + int n_store = remain_m/2; + switch (n_store) { + case 15: STORE_512(1, 6); + case 14: STORE_512(1, 5); + case 13: STORE_512(1, 4); + case 12: STORE_512(1, 3); + case 11: STORE_512(1, 2); + case 10: STORE_512(1, 1); + case 9: STORE_512(1, 0); + case 8: STORE_512(0, 7); + case 7: STORE_512(0, 6); + case 6: STORE_512(0, 5); + case 5: STORE_512(0, 4); + case 4: STORE_512(0, 3); + case 3: STORE_512(0, 2); + case 2: STORE_512(0, 1); + case 1: STORE_512(0, 0); + } + boffset0 += n_store * 32; + if (m & 0x1) { + __m512i tail; + GET_TAIL(); + _mm256_storeu_si256((void *)boffset0, _mm512_cvtepi32_epi16(tail)); + boffset0 += 16; + } + } + + } + if (j < n) { + int remain_n = n - j; + __mmask16 nmask = (1UL << remain_n) - 1; + int load0, load1; + if (remain_n > 8) { + load0 = 8; + load1 = remain_n - 8; + } else { + load0 = remain_n; + load1 = 0; + } + aoffset00 = aoffset; + aoffset01 = aoffset00 + lda; + aoffset02 = aoffset01 + lda; + aoffset03 = aoffset02 + lda; + aoffset04 = aoffset03 + lda; + aoffset05 = aoffset04 + lda; + aoffset06 = aoffset05 + lda; + aoffset07 = aoffset06 + lda; + aoffset10 = aoffset07 + lda; + aoffset11 = aoffset10 + lda; + aoffset12 = aoffset11 + lda; + aoffset13 = aoffset12 + lda; + aoffset14 = aoffset13 + lda; + aoffset15 = aoffset14 + lda; + aoffset16 = aoffset15 + lda; + aoffset17 = aoffset16 + lda; + aoffset += 16 * lda; + for (i = 0; i < m32; i += 32) { + switch (load0) { + case 8: r7 = _mm512_loadu_si512(aoffset07 + i); + case 7: r6 = _mm512_loadu_si512(aoffset06 + i); + case 6: r5 = _mm512_loadu_si512(aoffset05 + i); + case 5: r4 = _mm512_loadu_si512(aoffset04 + i); + case 4: r3 = _mm512_loadu_si512(aoffset03 + i); + case 3: r2 = _mm512_loadu_si512(aoffset02 + i); + case 2: r1 = _mm512_loadu_si512(aoffset01 + i); + case 1: r0 = _mm512_loadu_si512(aoffset00 + i); + } + REORDER_8x32(t00, t01, t02, t03, t04, t05, t06, t07); + switch (load1) { + case 8: r7 = _mm512_loadu_si512(aoffset17 + i); + case 7: r6 = _mm512_loadu_si512(aoffset16 + i); + case 6: r5 = _mm512_loadu_si512(aoffset15 + i); + case 5: r4 = _mm512_loadu_si512(aoffset14 + i); + case 4: r3 = _mm512_loadu_si512(aoffset13 + i); + case 3: r2 = _mm512_loadu_si512(aoffset12 + i); + case 2: r1 = _mm512_loadu_si512(aoffset11 + i); + case 1: r0 = _mm512_loadu_si512(aoffset10 + i); + } + REORDER_8x32(t10, t11, t12, t13, t14, t15, t16, t17); + MASK_STORE_512(0, 0); MASK_STORE_512(0, 1); MASK_STORE_512(0, 2); MASK_STORE_512(0, 3); + MASK_STORE_512(0, 4); MASK_STORE_512(0, 5); MASK_STORE_512(0, 6); MASK_STORE_512(0, 7); + MASK_STORE_512(1, 0); MASK_STORE_512(1, 1); MASK_STORE_512(1, 2); MASK_STORE_512(1, 3); + MASK_STORE_512(1, 4); MASK_STORE_512(1, 5); MASK_STORE_512(1, 6); MASK_STORE_512(1, 7); + boffset0 += remain_n * 32; + } + if (i < m) { + int remain_m = m - i; + __mmask32 mmask = (1UL << remain_m) - 1; + switch (load0) { + case 8: r7 = _mm512_maskz_loadu_epi16(mmask, aoffset07 + i); + case 7: r6 = _mm512_maskz_loadu_epi16(mmask, aoffset06 + i); + case 6: r5 = _mm512_maskz_loadu_epi16(mmask, aoffset05 + i); + case 5: r4 = _mm512_maskz_loadu_epi16(mmask, aoffset04 + i); + case 4: r3 = _mm512_maskz_loadu_epi16(mmask, aoffset03 + i); + case 3: r2 = _mm512_maskz_loadu_epi16(mmask, aoffset02 + i); + case 2: r1 = _mm512_maskz_loadu_epi16(mmask, aoffset01 + i); + case 1: r0 = _mm512_maskz_loadu_epi16(mmask, aoffset00 + i); + } + REORDER_8x32(t00, t01, t02, t03, t04, t05, t06, t07); + switch (load1) { + case 8: r7 = _mm512_maskz_loadu_epi16(mmask, aoffset17 + i); + case 7: r6 = _mm512_maskz_loadu_epi16(mmask, aoffset16 + i); + case 6: r5 = _mm512_maskz_loadu_epi16(mmask, aoffset15 + i); + case 5: r4 = _mm512_maskz_loadu_epi16(mmask, aoffset14 + i); + case 4: r3 = _mm512_maskz_loadu_epi16(mmask, aoffset13 + i); + case 3: r2 = _mm512_maskz_loadu_epi16(mmask, aoffset12 + i); + case 2: r1 = _mm512_maskz_loadu_epi16(mmask, aoffset11 + i); + case 1: r0 = _mm512_maskz_loadu_epi16(mmask, aoffset10 + i); + } + REORDER_8x32(t10, t11, t12, t13, t14, t15, t16, t17); + int n_store = remain_m/2; + switch (n_store) { + case 15: MASK_STORE_512(1, 6); + case 14: MASK_STORE_512(1, 5); + case 13: MASK_STORE_512(1, 4); + case 12: MASK_STORE_512(1, 3); + case 11: MASK_STORE_512(1, 2); + case 10: MASK_STORE_512(1, 1); + case 9: MASK_STORE_512(1, 0); + case 8: MASK_STORE_512(0, 7); + case 7: MASK_STORE_512(0, 6); + case 6: MASK_STORE_512(0, 5); + case 5: MASK_STORE_512(0, 4); + case 4: MASK_STORE_512(0, 3); + case 3: MASK_STORE_512(0, 2); + case 2: MASK_STORE_512(0, 1); + case 1: MASK_STORE_512(0, 0); + } + boffset0 += n_store * remain_n * 2; + if (m & 0x1) { + __m512i tail; + GET_TAIL(); + _mm256_mask_storeu_epi16((void *)boffset0, nmask, _mm512_cvtepi32_epi16(tail)); + } + } + } + return 0; +} diff --git a/kernel/x86_64/sbgemm_ncopy_4_cooperlake.c b/kernel/x86_64/sbgemm_ncopy_4_cooperlake.c new file mode 100644 index 000000000..eefbd7355 --- /dev/null +++ b/kernel/x86_64/sbgemm_ncopy_4_cooperlake.c @@ -0,0 +1,208 @@ +/*************************************************************************** +Copyright (c) 2021, The OpenBLAS Project +All rights reserved. +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions are +met: +1. Redistributions of source code must retain the above copyright +notice, this list of conditions and the following disclaimer. +2. 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. +3. Neither the name of the OpenBLAS project 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 OPENBLAS PROJECT 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 +#include +#include "common.h" + +#define REORDER_4x32(r0, r1, r2, r3) {\ + __m512i t0, t1, t2, t3; \ + t0 = _mm512_unpacklo_epi32(r0, r1); \ + t1 = _mm512_unpackhi_epi32(r0, r1); \ + t2 = _mm512_unpacklo_epi32(r2, r3); \ + t3 = _mm512_unpackhi_epi32(r2, r3); \ + r0 = _mm512_unpacklo_epi64(t0, t2); \ + r1 = _mm512_unpackhi_epi64(t0, t2); \ + r2 = _mm512_unpacklo_epi64(t1, t3); \ + r3 = _mm512_unpackhi_epi64(t1, t3); \ + t0 = _mm512_permutex2var_epi32(r0, idx_lo_128, r1); \ + t1 = _mm512_permutex2var_epi32(r0, idx_hi_128, r1); \ + t2 = _mm512_permutex2var_epi32(r2, idx_lo_128, r3); \ + t3 = _mm512_permutex2var_epi32(r2, idx_hi_128, r3); \ + r0 = _mm512_permutex2var_epi32(t0, idx_lo_256, t2); \ + r1 = _mm512_permutex2var_epi32(t1, idx_lo_256, t3); \ + r2 = _mm512_permutex2var_epi32(t0, idx_hi_256, t2); \ + r3 = _mm512_permutex2var_epi32(t1, idx_hi_256, t3); \ +} + +#define REORDER_4x8(r0, r1, r2, r3) {\ + __m128i t0, t1, t2, t3; \ + t0 = _mm_unpacklo_epi32(r0, r1); \ + t1 = _mm_unpackhi_epi32(r0, r1); \ + t2 = _mm_unpacklo_epi32(r2, r3); \ + t3 = _mm_unpackhi_epi32(r2, r3); \ + r0 = _mm_unpacklo_epi64(t0, t2); \ + r1 = _mm_unpackhi_epi64(t0, t2); \ + r2 = _mm_unpacklo_epi64(t1, t3); \ + r3 = _mm_unpackhi_epi64(t1, t3); \ +} + +#define GET_TAIL(tail, remain_m) \ + switch((remain_m + 1)/2) { \ + case 1: tail = r0; break; \ + case 2: tail = r1; break; \ + case 3: tail = r2; break; \ + case 4: tail = r3; break; \ + } + +int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){ + BLASLONG i, j; + IFLOAT *aoffset; + IFLOAT *aoffset0, *aoffset1, *aoffset2, *aoffset3; + + IFLOAT *boffset; + + aoffset = a; + boffset = b; + + BLASLONG m32 = m & ~31; + BLASLONG m8 = m & ~7; + BLASLONG n4 = n & ~3; + + int permute_table[] = { + 0x0, 0x1, 0x2, 0x3, 0x10, 0x11, 0x12, 0x13, 0x8, 0x9, 0xa, 0xb, 0x18, 0x19, 0x1a, 0x1b, + 0x4, 0x5, 0x6, 0x7, 0x14, 0x15, 0x16, 0x17, 0xc, 0xd, 0xe, 0xf, 0x1c, 0x1d, 0x1e, 0x1f, + 0x0, 0x1, 0x2, 0x3, 0x4, 0x5, 0x6, 0x7, 0x10, 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17, + 0x8, 0x9, 0xa, 0xb, 0xc, 0xd, 0xe, 0xf, 0x18, 0x19, 0x1a, 0x1b, 0x1c, 0x1d, 0x1e, 0x1f, + }; + __m512i idx_lo_128 = _mm512_loadu_si512(permute_table); + __m512i idx_hi_128 = _mm512_loadu_si512(permute_table + 16); + __m512i idx_lo_256 = _mm512_loadu_si512(permute_table + 32); + __m512i idx_hi_256 = _mm512_loadu_si512(permute_table + 48); + + for (j = 0; j < n4; j += 4) { + aoffset0 = aoffset; + aoffset1 = aoffset0 + lda; + aoffset2 = aoffset1 + lda; + aoffset3 = aoffset2 + lda; + aoffset += 4 * lda; + + for (i = 0; i < m32; i += 32) { + __m512i r0, r1, r2, r3; + r0 = _mm512_loadu_si512(aoffset0 + i); + r1 = _mm512_loadu_si512(aoffset1 + i); + r2 = _mm512_loadu_si512(aoffset2 + i); + r3 = _mm512_loadu_si512(aoffset3 + i); + REORDER_4x32(r0, r1, r2, r3); + _mm512_storeu_si512(boffset + 32*0, r0); + _mm512_storeu_si512(boffset + 32*1, r1); + _mm512_storeu_si512(boffset + 32*2, r2); + _mm512_storeu_si512(boffset + 32*3, r3); + boffset += 32 * 4; + } + for (; i < m8; i += 8) { + __m128i r0 = _mm_loadu_si128((void *)(aoffset0 + i)); + __m128i r1 = _mm_loadu_si128((void *)(aoffset1 + i)); + __m128i r2 = _mm_loadu_si128((void *)(aoffset2 + i)); + __m128i r3 = _mm_loadu_si128((void *)(aoffset3 + i)); + REORDER_4x8(r0, r1, r2, r3); + _mm_storeu_si128((void *)(boffset + 8*0), r0); + _mm_storeu_si128((void *)(boffset + 8*1), r1); + _mm_storeu_si128((void *)(boffset + 8*2), r2); + _mm_storeu_si128((void *)(boffset + 8*3), r3); + boffset += 8 * 4; + } + if (i < m) { + int remain_m = m - i; + __mmask8 r_mask = (1UL << remain_m) - 1; + __m128i r0 = _mm_maskz_loadu_epi16(r_mask, aoffset0 + i); + __m128i r1 = _mm_maskz_loadu_epi16(r_mask, aoffset1 + i); + __m128i r2 = _mm_maskz_loadu_epi16(r_mask, aoffset2 + i); + __m128i r3 = _mm_maskz_loadu_epi16(r_mask, aoffset3 + i); + REORDER_4x8(r0, r1, r2, r3); + + // store should skip the tail odd line + int num_store = remain_m/2; + switch(num_store) { + case 3: _mm_storeu_si128((void *)(boffset + 8*2), r2); + case 2: _mm_storeu_si128((void *)(boffset + 8*1), r1); + case 1: _mm_storeu_si128((void *)(boffset + 8*0), r0); + } + boffset += 8 * num_store; + + if (m & 0x1) { // handling the tail + __m128i tail; + GET_TAIL(tail, remain_m); + /* tail vector is fill with zero like: + * a, 0, b, 0, c, 0, d, 0 + * need to extract lo words of data and store + */ + tail = _mm_cvtepi32_epi16(tail); + _mm_store_sd((double *)boffset, (__m128d) tail); // only lower 4 bfloat valid + boffset += 4; + } + } + } + if (j < n) { + int remain_n = n - j; + __mmask8 nmask = (1UL << remain_n) - 1; + aoffset0 = aoffset; + aoffset1 = aoffset0 + lda; + aoffset2 = aoffset1 + lda; + aoffset3 = aoffset2 + lda; + __m128i r0, r1, r2, r3; + for (i = 0; i < m8; i += 8) { + switch (remain_n) { + case 3: r2 = _mm_loadu_si128((void *)(aoffset2 + i)); + case 2: r1 = _mm_loadu_si128((void *)(aoffset1 + i)); + case 1: r0 = _mm_loadu_si128((void *)(aoffset0 + i)); + } + REORDER_4x8(r0, r1, r2, r3); + _mm_mask_storeu_epi32(boffset + remain_n * 0, nmask, r0); + _mm_mask_storeu_epi32(boffset + remain_n * 2, nmask, r1); + _mm_mask_storeu_epi32(boffset + remain_n * 4, nmask, r2); + _mm_mask_storeu_epi32(boffset + remain_n * 6, nmask, r3); + boffset += 8 * remain_n; + } + if (i < m) { + int remain_m = m - i; + __mmask8 mmask = (1UL << remain_m) - 1; + switch (remain_n) { + case 3: r2 = _mm_maskz_loadu_epi16(mmask, aoffset2 + i); + case 2: r1 = _mm_maskz_loadu_epi16(mmask, aoffset1 + i); + case 1: r0 = _mm_maskz_loadu_epi16(mmask, aoffset0 + i); + } + REORDER_4x8(r0, r1, r2, r3); + + int num_store = remain_m/2; + switch (num_store) { + case 3: _mm_mask_storeu_epi32(boffset + remain_n * 4, nmask, r2); + case 2: _mm_mask_storeu_epi32(boffset + remain_n * 2, nmask, r1); + case 1: _mm_mask_storeu_epi32(boffset + remain_n * 0, nmask, r0); + } + boffset += 2 * num_store * remain_n; + + if (m & 0x1) { + __m128i tail; + GET_TAIL(tail, remain_m); + tail = _mm_cvtepi32_epi16(tail); + _mm_mask_storeu_epi16(boffset, nmask, tail); + } + } + } + return 0; +} diff --git a/kernel/x86_64/sbgemm_small_kernel_permit_cooperlake.c b/kernel/x86_64/sbgemm_small_kernel_permit_cooperlake.c index 823aafbdd..70becd9fa 100644 --- a/kernel/x86_64/sbgemm_small_kernel_permit_cooperlake.c +++ b/kernel/x86_64/sbgemm_small_kernel_permit_cooperlake.c @@ -38,5 +38,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. int CNAME(int transa, int transb, BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, FLOAT beta) { - return 1; + double MNK = (double) M * (double) N * (double) K; + if (MNK > 256.0*256.0*256.0) // disable for big size matrix + return 0; + /* small matrix kernel works well for N = 8, 16, 32 */ + if (N == 8 || N == 16 || N == 32) + return 1; + return 0; } diff --git a/kernel/x86_64/sbgemm_tcopy_16_cooperlake.c b/kernel/x86_64/sbgemm_tcopy_16_cooperlake.c new file mode 100644 index 000000000..88725f343 --- /dev/null +++ b/kernel/x86_64/sbgemm_tcopy_16_cooperlake.c @@ -0,0 +1,164 @@ +/*************************************************************************** +Copyright (c) 2021, The OpenBLAS Project +All rights reserved. +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions are +met: +1. Redistributions of source code must retain the above copyright +notice, this list of conditions and the following disclaimer. +2. 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. +3. Neither the name of the OpenBLAS project 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 OPENBLAS PROJECT 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 +#include +#include "common.h" + + +int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){ + BLASLONG i, j; + + IFLOAT *boffset0, *boffset1; + + boffset0 = b; + + BLASLONG n32 = n & ~31; + BLASLONG m4 = m & ~3; + BLASLONG m2 = m & ~1; + + uint32_t permute_table[] = { + 0x00, 0x01, 0x02, 0x03, 0x10, 0x11, 0x12, 0x13, 0x04, 0x05, 0x06, 0x07, 0x14, 0x15, 0x16, 0x17, + 0x08, 0x09, 0x0a, 0x0b, 0x18, 0x19, 0x1a, 0x1b, 0x0c, 0x0d, 0x0e, 0x0f, 0x1c, 0x1d, 0x1e, 0x1f, + }; + + __m512i idx_lo = _mm512_loadu_si512(permute_table); + __m512i idx_hi = _mm512_loadu_si512(permute_table + 16); + + for (j = 0; j < n32; j += 32) { + /* process 2x16 n at the same time */ + boffset1 = boffset0 + m * 16; + for (i = 0; i < m4; i += 4) { + /* bf16 fma need special memory layout: + * for memory layout like below: + * a00, a01, a02, a03, a04, a05 .... + * a10, a11, a12, a13, a14, a15 .... + * need to copy as: + * a00, a10, a01, a11, a02, a12, a03, a13, ... + */ + __m512i a0 = _mm512_loadu_si512(&a[(i + 0)*lda + j]); + __m512i a1 = _mm512_loadu_si512(&a[(i + 1)*lda + j]); + __m512i a2 = _mm512_loadu_si512(&a[(i + 2)*lda + j]); + __m512i a3 = _mm512_loadu_si512(&a[(i + 3)*lda + j]); + + __m512i a00 = _mm512_unpacklo_epi16(a0, a1); + __m512i a01 = _mm512_unpackhi_epi16(a0, a1); + __m512i a10 = _mm512_unpacklo_epi16(a2, a3); + __m512i a11 = _mm512_unpackhi_epi16(a2, a3); + + a0 = _mm512_permutex2var_epi32(a00, idx_lo, a01); + a1 = _mm512_permutex2var_epi32(a00, idx_hi, a01); + a2 = _mm512_permutex2var_epi32(a10, idx_lo, a11); + a3 = _mm512_permutex2var_epi32(a10, idx_hi, a11); + + _mm512_storeu_si512(boffset0, a0); + _mm512_storeu_si512(boffset1, a1); + _mm512_storeu_si512(boffset0 + 32, a2); + _mm512_storeu_si512(boffset1 + 32, a3); + boffset0 += 64; + boffset1 += 64; + } + for (; i < m2; i += 2) { + __m512i a0 = _mm512_loadu_si512(&a[(i + 0)*lda + j]); + __m512i a1 = _mm512_loadu_si512(&a[(i + 1)*lda + j]); + + __m512i a00 = _mm512_unpacklo_epi16(a0, a1); + __m512i a01 = _mm512_unpackhi_epi16(a0, a1); + + a0 = _mm512_permutex2var_epi32(a00, idx_lo, a01); + a1 = _mm512_permutex2var_epi32(a00, idx_hi, a01); + + _mm512_storeu_si512(boffset0, a0); + _mm512_storeu_si512(boffset1, a1); + boffset0 += 32; + boffset1 += 32; + } + for (; i < m; i++) { + /* just copy the only remains row */ + __m256i a0 = _mm256_loadu_si256((void *)&a[(i + 0)*lda + j]); + __m256i a1 = _mm256_loadu_si256((void *)&a[(i + 0)*lda + j + 16]); + _mm256_storeu_si256((void *)boffset0, a0); + _mm256_storeu_si256((void *)boffset1, a1); + boffset0 += 16; + boffset1 += 16; + } + boffset0 = boffset1; + } + if (j < n) { + uint32_t remains = n - j; + __mmask32 r_mask = (1UL << remains) - 1; + if (remains > 16) { + boffset1 = boffset0 + m * 16; + uint32_t tail1 = remains - 16; + __mmask16 w_mask1 = (1UL << tail1) - 1; + for (i = 0; i < m2; i += 2) { + __m512i a0 = _mm512_maskz_loadu_epi16(r_mask, &a[(i + 0)*lda + j]); + __m512i a1 = _mm512_maskz_loadu_epi16(r_mask, &a[(i + 1)*lda + j]); + + __m512i a00 = _mm512_unpacklo_epi16(a0, a1); + __m512i a01 = _mm512_unpackhi_epi16(a0, a1); + + a0 = _mm512_permutex2var_epi32(a00, idx_lo, a01); + a1 = _mm512_permutex2var_epi32(a00, idx_hi, a01); + + _mm512_storeu_si512(boffset0, a0); + _mm512_mask_storeu_epi32(boffset1, w_mask1, a1); + + boffset0 += 32; + boffset1 += 2 * tail1; + } + for (; i < m; i++) { + __m256i a0 = _mm256_loadu_si256((void *)&a[(i + 0)*lda + j]); + __m256i a1 = _mm256_maskz_loadu_epi16(w_mask1, (void *)&a[(i + 0)*lda + j + 16]); + _mm256_storeu_si256((void *)boffset0, a0); + _mm256_mask_storeu_epi16((void *)boffset1, w_mask1, a1); + boffset0 += 16; + boffset1 += tail1; + } + } else { + __mmask16 w_mask = (1UL << remains ) - 1; + for (i = 0; i < m2; i += 2) { + __m512i a0 = _mm512_maskz_loadu_epi16(r_mask, &a[(i + 0)*lda + j]); + __m512i a1 = _mm512_maskz_loadu_epi16(r_mask, &a[(i + 1)*lda + j]); + + __m512i a00 = _mm512_unpacklo_epi16(a0, a1); + __m512i a01 = _mm512_unpackhi_epi16(a0, a1); + + a0 = _mm512_permutex2var_epi32(a00, idx_lo, a01); + + _mm512_mask_storeu_epi32(boffset0, w_mask, a0); + boffset0 += 2 * remains; + } + for (; i < m; i++) { + __m256i a0 = _mm256_maskz_loadu_epi16(w_mask, &a[(i + 0)*lda + j]); + _mm256_mask_storeu_epi16(boffset0, w_mask, a0); + boffset0 += remains; + } + } + } + return 0; +} diff --git a/kernel/x86_64/sbgemm_tcopy_4_cooperlake.c b/kernel/x86_64/sbgemm_tcopy_4_cooperlake.c new file mode 100644 index 000000000..e9edd4571 --- /dev/null +++ b/kernel/x86_64/sbgemm_tcopy_4_cooperlake.c @@ -0,0 +1,216 @@ +/*************************************************************************** +Copyright (c) 2021, The OpenBLAS Project +All rights reserved. +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions are +met: +1. Redistributions of source code must retain the above copyright +notice, this list of conditions and the following disclaimer. +2. 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. +3. Neither the name of the OpenBLAS project 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 OPENBLAS PROJECT 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 +#include +#include "common.h" + +#define STORE_VEC(Bx, By, vec) \ + if (By == 0) asm("vmovdqu16 %0, (%1)": : "v"(vec), "r"(boffset##Bx)); \ + else asm("vmovdqu16 %0, (%1, %2, %c3)": : "v"(vec), "r"(boffset##Bx), "r"(blk_size), "n"(By * 2)); + +int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){ + BLASLONG i, j; + + IFLOAT *boffset0, *boffset1; + + boffset0 = b; + + BLASLONG n24 = n - (n % 24); + BLASLONG n8 = n & ~7; + BLASLONG m8 = m & ~7; + BLASLONG m4 = m & ~3; + BLASLONG m2 = m & ~1; + + int permute_table[] = { + 0x0, 0x1, 0x2, 0x3, 0x10, 0x11, 0x12, 0x13, 0x8, 0x9, 0xa, 0xb, 0x18, 0x19, 0x1a, 0x1b, + 0x4, 0x5, 0x6, 0x7, 0x14, 0x15, 0x16, 0x17, 0xc, 0xd, 0xe, 0xf, 0x1c, 0x1d, 0x1e, 0x1f, + 0x0, 0x1, 0x2, 0x3, 0x4, 0x5, 0x6, 0x7, 0x10, 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17, + 0x8, 0x9, 0xa, 0xb, 0xc, 0xd, 0xe, 0xf, 0x18, 0x19, 0x1a, 0x1b, 0x1c, 0x1d, 0x1e, 0x1f, + }; + + j = 0; + if (n > 23) { + /* n = 24 is the max width in current blocking setting */ + __m512i idx_lo_128 = _mm512_loadu_si512(permute_table); + __m512i idx_hi_128 = _mm512_loadu_si512(permute_table + 16); + __m512i idx_lo_256 = _mm512_loadu_si512(permute_table + 32); + __m512i idx_hi_256 = _mm512_loadu_si512(permute_table + 48); + __mmask32 mask24 = (1UL << 24) - 1; + BLASLONG blk_size = m * 4; + BLASLONG stride = blk_size * 3; + + for (; j < n24; j += 24) { + boffset1 = boffset0 + stride; + for (i = 0; i < m8; i += 8) { + __m512i r0, r1, r2, r3, r4, r5, r6, r7; + __m512i t0, t1, t2, t3, t4, t5, t6, t7; + r0 = _mm512_maskz_loadu_epi16(mask24, &a[(i + 0)*lda + j]); + r1 = _mm512_maskz_loadu_epi16(mask24, &a[(i + 1)*lda + j]); + r2 = _mm512_maskz_loadu_epi16(mask24, &a[(i + 2)*lda + j]); + r3 = _mm512_maskz_loadu_epi16(mask24, &a[(i + 3)*lda + j]); + r4 = _mm512_maskz_loadu_epi16(mask24, &a[(i + 4)*lda + j]); + r5 = _mm512_maskz_loadu_epi16(mask24, &a[(i + 5)*lda + j]); + r6 = _mm512_maskz_loadu_epi16(mask24, &a[(i + 6)*lda + j]); + r7 = _mm512_maskz_loadu_epi16(mask24, &a[(i + 7)*lda + j]); + + t0 = _mm512_unpacklo_epi16(r0, r1); + t1 = _mm512_unpackhi_epi16(r0, r1); + t2 = _mm512_unpacklo_epi16(r2, r3); + t3 = _mm512_unpackhi_epi16(r2, r3); + t4 = _mm512_unpacklo_epi16(r4, r5); + t5 = _mm512_unpackhi_epi16(r4, r5); + t6 = _mm512_unpacklo_epi16(r6, r7); + t7 = _mm512_unpackhi_epi16(r6, r7); + + r0 = _mm512_permutex2var_epi32(t0, idx_lo_128, t2); + r1 = _mm512_permutex2var_epi32(t1, idx_lo_128, t3); + r2 = _mm512_permutex2var_epi32(t4, idx_lo_128, t6); + r3 = _mm512_permutex2var_epi32(t5, idx_lo_128, t7); + r4 = _mm512_permutex2var_epi32(t0, idx_hi_128, t2); + r5 = _mm512_permutex2var_epi32(t1, idx_hi_128, t3); + r6 = _mm512_permutex2var_epi32(t4, idx_hi_128, t6); + r7 = _mm512_permutex2var_epi32(t5, idx_hi_128, t7); + + t0 = _mm512_permutex2var_epi32(r0, idx_lo_256, r2); + t1 = _mm512_permutex2var_epi32(r1, idx_lo_256, r3); + t2 = _mm512_permutex2var_epi32(r4, idx_lo_256, r6); + t3 = _mm512_permutex2var_epi32(r5, idx_lo_256, r7); + t4 = _mm512_permutex2var_epi32(r0, idx_hi_256, r2); + t5 = _mm512_permutex2var_epi32(r1, idx_hi_256, r3); + + STORE_VEC(0, 0, t0); STORE_VEC(0, 1, t1); STORE_VEC(0, 2, t2); + STORE_VEC(1, 0, t3); STORE_VEC(1, 1, t4); STORE_VEC(1, 2, t5); + boffset0 += 32; + boffset1 += 32; + } + for (; i < m2; i += 2) { + __m512i r0, r1, t0, t1; + r0 = _mm512_maskz_loadu_epi16(mask24, &a[(i + 0)*lda + j]); + r1 = _mm512_maskz_loadu_epi16(mask24, &a[(i + 1)*lda + j]); + t0 = _mm512_unpacklo_epi16(r0, r1); + t1 = _mm512_unpackhi_epi16(r0, r1); + STORE_VEC(0, 0, _mm512_extracti32x4_epi32(t0, 0)); + STORE_VEC(0, 1, _mm512_extracti32x4_epi32(t1, 0)); + STORE_VEC(0, 2, _mm512_extracti32x4_epi32(t0, 1)); + STORE_VEC(1, 0, _mm512_extracti32x4_epi32(t1, 1)); + STORE_VEC(1, 1, _mm512_extracti32x4_epi32(t0, 2)); + STORE_VEC(1, 2, _mm512_extracti32x4_epi32(t1, 2)); + boffset0 += 8; + boffset1 += 8; + } + for (; i < m; i++) { + *(uint64_t *)(boffset0 + blk_size * 0) = *(uint64_t *)&a[i * lda + j + 0]; + *(uint64_t *)(boffset0 + blk_size * 1) = *(uint64_t *)&a[i * lda + j + 4]; + *(uint64_t *)(boffset0 + blk_size * 2) = *(uint64_t *)&a[i * lda + j + 8]; + *(uint64_t *)(boffset1 + blk_size * 0) = *(uint64_t *)&a[i * lda + j + 12]; + *(uint64_t *)(boffset1 + blk_size * 1) = *(uint64_t *)&a[i * lda + j + 16]; + *(uint64_t *)(boffset1 + blk_size * 2) = *(uint64_t *)&a[i * lda + j + 20]; + boffset0 += 4; + boffset1 += 4; + } + boffset0 += stride * 2; + } + } + + for (; j < n8; j += 8) { + boffset1 = boffset0 + m * 4; + for (i = 0; i < m4; i += 4) { + __m128i a0 = _mm_loadu_si128((void *)&a[(i + 0)*lda + j]); + __m128i a1 = _mm_loadu_si128((void *)&a[(i + 1)*lda + j]); + __m128i a2 = _mm_loadu_si128((void *)&a[(i + 2)*lda + j]); + __m128i a3 = _mm_loadu_si128((void *)&a[(i + 3)*lda + j]); + __m128i a00 = _mm_unpacklo_epi16(a0, a1); + __m128i a01 = _mm_unpackhi_epi16(a0, a1); + __m128i a10 = _mm_unpacklo_epi16(a2, a3); + __m128i a11 = _mm_unpackhi_epi16(a2, a3); + _mm_storeu_si128((void *)(boffset0 + 0), a00); + _mm_storeu_si128((void *)(boffset0 + 8), a10); + _mm_storeu_si128((void *)(boffset1 + 0), a01); + _mm_storeu_si128((void *)(boffset1 + 8), a11); + boffset0 += 16; + boffset1 += 16; + } + for (; i < m2; i+= 2) { + __m128i a0 = _mm_loadu_si128((void *)&a[(i + 0)*lda + j]); + __m128i a1 = _mm_loadu_si128((void *)&a[(i + 1)*lda + j]); + __m128i a00 = _mm_unpacklo_epi16(a0, a1); + __m128i a01 = _mm_unpackhi_epi16(a0, a1); + _mm_storeu_si128((void *)(boffset0 + 0), a00); + _mm_storeu_si128((void *)(boffset1 + 0), a01); + boffset0 += 8; + boffset1 += 8; + } + for (; i < m; i++) { + __m128d a0 = _mm_loadu_pd((void *)&a[(i + 0)*lda + j]); + _mm_store_sd((void *)boffset0, a0); + _mm_store_sd((void *)boffset1, _mm_permute_pd(a0, 0x1)); + boffset0 += 4; + boffset1 += 4; + } + boffset0 = boffset1; + } + if (j < n) { + uint32_t remains = n - j; + __mmask8 r_mask = (1UL << remains) - 1; + if (remains > 4) { + boffset1 = boffset0 + m * 4; + uint32_t tail1 = remains - 4; + __mmask8 w_mask1 = (1UL << tail1) - 1; + for (i = 0; i < m2; i += 2) { + __m128i a0 = _mm_maskz_loadu_epi16(r_mask, &a[(i + 0)*lda + j]); + __m128i a1 = _mm_maskz_loadu_epi16(r_mask, &a[(i + 1)*lda + j]); + __m128i a00 = _mm_unpacklo_epi16(a0, a1); + __m128i a01 = _mm_unpackhi_epi16(a0, a1); + _mm_storeu_si128((void *)boffset0, a00); + _mm_mask_storeu_epi32((void *)boffset1, w_mask1, a01); + boffset0 += 8; + boffset1 += 2 * tail1; + } + for (; i < m; i++) { + __m128i a0 = _mm_maskz_loadu_epi16(r_mask, &a[(i + 0)*lda + j]); + _mm_store_sd((void *)boffset0, (__m128d) a0); + _mm_mask_storeu_epi16((void *)boffset1, w_mask1, (__m128i) _mm_permute_pd((__m128d) a0, 0x1)); + boffset0 += 4; + boffset1 += tail1; + } + } else { + for (i = 0; i < m2; i += 2) { + __m128i a0 = _mm_maskz_loadu_epi16(r_mask, &a[(i + 0)*lda + j]); + __m128i a1 = _mm_maskz_loadu_epi16(r_mask, &a[(i + 1)*lda + j]); + __m128i a00 = _mm_unpacklo_epi16(a0, a1); + _mm_mask_storeu_epi32((void *)boffset0, r_mask, a00); + boffset0 += 2 * remains; + } + for (; i < m; i++) { + __m128i a0 = _mm_maskz_loadu_epi16(r_mask, &a[(i + 0)*lda + j]); + _mm_mask_storeu_epi16((void *)boffset0, r_mask, a0); + } + } + } + return 0; +} diff --git a/kernel/x86_64/sgemm_beta_skylakex.c b/kernel/x86_64/sgemm_beta_skylakex.c index 1c29c1168..6217acf48 100644 --- a/kernel/x86_64/sgemm_beta_skylakex.c +++ b/kernel/x86_64/sgemm_beta_skylakex.c @@ -41,7 +41,7 @@ #include int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT beta, - FLOAT *dummy2, BLASLONG dummy3, FLOAT *dummy4, BLASLONG dummy5, + IFLOAT *dummy2, BLASLONG dummy3, IFLOAT *dummy4, BLASLONG dummy5, FLOAT *c, BLASLONG ldc){ BLASLONG i, j; diff --git a/param.h b/param.h index 07397a66e..48770fa7a 100644 --- a/param.h +++ b/param.h @@ -1771,6 +1771,17 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #endif #define USE_SGEMM_KERNEL_DIRECT 1 +#undef SBGEMM_DEFAULT_UNROLL_N +#undef SBGEMM_DEFAULT_UNROLL_M +#undef SBGEMM_DEFAULT_P +#undef SBGEMM_DEFAULT_R +#undef SBGEMM_DEFAULT_Q +#define SBGEMM_DEFAULT_UNROLL_N 4 +#define SBGEMM_DEFAULT_UNROLL_M 16 +#define SBGEMM_DEFAULT_P 384 +#define SBGEMM_DEFAULT_Q 768 +#define SBGEMM_DEFAULT_R sbgemm_r + #ifdef ARCH_X86 #define SGEMM_DEFAULT_UNROLL_M 4