From 89a12fa08352da73b03225cf1743dead75669043 Mon Sep 17 00:00:00 2001 From: Chip Kerchner Date: Mon, 23 Sep 2024 06:32:14 -0500 Subject: [PATCH] MMA BF16 GEMV code. --- kernel/power/gemm_common.c | 2 + kernel/power/sbgemv_common.c | 2 +- kernel/power/sbgemv_common_power10.c | 265 +++++++++++++++++++++ kernel/power/sbgemv_n.c | 22 ++ kernel/power/sbgemv_n_power10.c | 306 ++++++++++++++++++++++++- kernel/power/sbgemv_n_vsx.c | 67 +++--- kernel/power/sbgemv_t.c | 15 ++ kernel/power/sbgemv_t_power10.c | 330 ++++++++++++++++++++++++++- kernel/power/sbgemv_t_vsx.c | 65 +++--- 9 files changed, 1008 insertions(+), 66 deletions(-) create mode 100644 kernel/power/sbgemv_common_power10.c diff --git a/kernel/power/gemm_common.c b/kernel/power/gemm_common.c index c33faffe0..0611ebc2a 100644 --- a/kernel/power/gemm_common.c +++ b/kernel/power/gemm_common.c @@ -4,6 +4,8 @@ #include +#define NBMAX 4096 + #define FORCEINLINE inline __attribute__((always_inline)) #ifdef __clang__ diff --git a/kernel/power/sbgemv_common.c b/kernel/power/sbgemv_common.c index 46dee74c3..ab50f430a 100644 --- a/kernel/power/sbgemv_common.c +++ b/kernel/power/sbgemv_common.c @@ -111,7 +111,7 @@ FORCEINLINE vec_f32 vec_loadNHi_mult(vec_bf16 *in, vec_f32 v_inp0, BLASLONG n, v return (v_inp0 * v_in00); } -FORCEINLINE vec_f32 vec_loadNHi_multi2(vec_f32 v_x0, vec_bf16 *in, BLASLONG n, vec_bf16 zero) +FORCEINLINE vec_f32 vec_loadNHi_mult2(vec_f32 v_x0, vec_bf16 *in, BLASLONG n, vec_bf16 zero) { vec_f32 v_in00 = vec_loadNHi(in, n, zero); diff --git a/kernel/power/sbgemv_common_power10.c b/kernel/power/sbgemv_common_power10.c new file mode 100644 index 000000000..da088014b --- /dev/null +++ b/kernel/power/sbgemv_common_power10.c @@ -0,0 +1,265 @@ +/*************************************************************************** +Copyright (c) 2024, 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. +*****************************************************************************/ + +#ifndef SBGEMV_COMMON_MMA_C +#define SBGEMV_COMMON_MMA_C +#include "sbgemv_common.c" + +FORCEINLINE void vec_load_mult_mma(__vector_quad *out, vec_bf16 *in, vec_bf16 inp) +{ + vec_bf16 in0 = (vec_bf16)vec_load_vec(in); + + __builtin_mma_xvbf16ger2pp(out, (vec_uc8)in0, (vec_uc8)inp); +} + +FORCEINLINE void vec_load_mult2_mma(__vector_quad *out, vec_bf16 *in, vec_bf16 *inp) +{ + vec_bf16 in0[2]; + + vec_load_pair((vec_f32 *)in0, (vec_f32 *)in); + + __builtin_mma_xvbf16ger2pp(out, (vec_uc8)in0[0], (vec_uc8)inp[0]); + __builtin_mma_xvbf16ger2pp(out, (vec_uc8)in0[1], (vec_uc8)inp[1]); +} + +FORCEINLINE void vec_loadN_mult_mma(__vector_quad *out, vec_bf16 *in, vec_bf16 inp, BLASLONG n) +{ + vec_bf16 in0 = vec_loadN(in, n); + + __builtin_mma_xvbf16ger2pp(out, (vec_uc8)in0, (vec_uc8)inp); +} + +FORCEINLINE void vec_mult1_mma(__vector_quad *out, vec_bf16 in0, vec_bf16 inp) +{ + vec_bf16 in00 = vec_mergeh(in0, in0); + + __builtin_mma_xvbf16ger2(out, (vec_uc8)inp, (vec_uc8)in00); +} + +FORCEINLINE void vec_mult2_mma(__vector_quad *out, vec_bf16 in0, vec_bf16 inp) +{ + vec_bf16 in01 = vec_mergel(in0, in0); + + vec_mult1_mma(&out[0], in0, inp); + + __builtin_mma_xvbf16ger2(&out[1], (vec_uc8)inp, (vec_uc8)in01); +} + +FORCEINLINE void vec_mult4_mma(__vector_quad *out, vec_bf16 *in0, vec_bf16 inp) +{ + vec_mult2_mma(out + 0, in0[0], inp); + vec_mult2_mma(out + 2, in0[1], inp); +} + +FORCEINLINE void vec_loadN_mult11_mma(__vector_quad *out, vec_bf16 *in, vec_bf16 inp, BLASLONG n) +{ + vec_bf16 in0 = vec_loadN(in, n); + + vec_mult1_mma(out, in0, inp); +} + +FORCEINLINE void vec_loadN_mult12_mma(__vector_quad *out, vec_bf16 *in, vec_bf16 inp, BLASLONG n) +{ + vec_bf16 in0 = vec_loadN(in, n); + + vec_mult2_mma(out, in0, inp); +} + +FORCEINLINE void vec_load_mult12_mma(__vector_quad *out, vec_bf16 *in, vec_bf16 inp) +{ + vec_bf16 in0 = (vec_bf16)vec_load_vec(in); + + vec_mult2_mma(out, in0, inp); +} + +FORCEINLINE void vec_load_mult18_mma(__vector_quad *out, vec_bf16 *in, vec_bf16 inp) +{ + vec_bf16 in0[4]; + + vec_load_pair((vec_f32 *)(in0 + 0), (vec_f32 *)(in + 0)); + vec_load_pair((vec_f32 *)(in0 + 2), (vec_f32 *)(in + 2)); + + vec_mult4_mma(&out[0], in0 + 0, inp); + vec_mult4_mma(&out[4], in0 + 2, inp); +} + +FORCEINLINE void vec_reduce1_mma(__vector_quad *out, vec_f32 *temp, vec_f32 v_alpha, vec_f32 *vy0) +{ + __builtin_mma_disassemble_acc((void*)temp, &out[0]); + + vy0[0] += (temp[0] * v_alpha); +} + +FORCEINLINE void vec_reduce2_mma(__vector_quad *out, vec_f32 *temp, vec_f32 v_alpha, vec_f32 *vy0) +{ + vec_reduce1_mma(&out[0], &temp[0], v_alpha, &vy0[0]); + vec_reduce1_mma(&out[1], &temp[4], v_alpha, &vy0[1]); +} + +FORCEINLINE void vec_reduce8_mma(__vector_quad *out, vec_f32 *temp, vec_f32 v_alpha, vec_f32 *vy0) +{ + vec_reduce2_mma(&out[0], &temp[0], v_alpha, vy0 + 0); + vec_reduce2_mma(&out[2], &temp[8], v_alpha, vy0 + 2); + vec_reduce2_mma(&out[4], &temp[16], v_alpha, vy0 + 4); + vec_reduce2_mma(&out[6], &temp[24], v_alpha, vy0 + 6); +} + +FORCEINLINE void vec_mult11a_mma(__vector_quad *out, vec_bf16 in0, vec_bf16 in1, vec_bf16 inp) +{ + vec_bf16 in00 = vec_mergeh(in0, in1); + + __builtin_mma_xvbf16ger2(out, (vec_uc8)inp, (vec_uc8)in00); +} + +FORCEINLINE void vec_mult2a_mma(__vector_quad *out, vec_bf16 in0, vec_bf16 in1, vec_bf16 inp) +{ + vec_bf16 in01 = vec_mergel(in0, in1); + + vec_mult11a_mma(&out[0], in0, in1, inp); + + __builtin_mma_xvbf16ger2(&out[1], (vec_uc8)inp, (vec_uc8)in01); +} + +FORCEINLINE void vec_mult4a_mma(__vector_quad *out, vec_bf16 *in0, vec_bf16 *in1, vec_bf16 inp) +{ + vec_mult2a_mma(out + 0, in0[0], in1[0], inp); + vec_mult2a_mma(out + 2, in0[1], in1[1], inp); +} + +FORCEINLINE void vec_loadN_mult11a_mma(__vector_quad *out, vec_bf16 *ina, vec_bf16 *inb, vec_bf16 inp, BLASLONG n) +{ + vec_bf16 in0 = vec_loadN(ina, n); + vec_bf16 in1 = vec_loadN(inb, n); + + vec_mult11a_mma(out, in0, in1, inp); +} + +FORCEINLINE void vec_load_mult22a_mma(__vector_quad *out, vec_bf16 *ina, vec_bf16 *inb, vec_bf16 inp) +{ + vec_bf16 in0 = (vec_bf16)vec_load_vec(ina); + vec_bf16 in1 = (vec_bf16)vec_load_vec(inb); + + vec_mult2a_mma(out, in0, in1, inp); +} + +FORCEINLINE void vec_load_mult28a_mma(__vector_quad *out, vec_bf16 *ina, vec_bf16 *inb, vec_bf16 inp) +{ + vec_bf16 in0[4], in1[4]; + + vec_load_pair((vec_f32 *)(in0 + 0), (vec_f32 *)(ina + 0)); + vec_load_pair((vec_f32 *)(in1 + 0), (vec_f32 *)(inb + 0)); + vec_load_pair((vec_f32 *)(in0 + 2), (vec_f32 *)(ina + 2)); + vec_load_pair((vec_f32 *)(in1 + 2), (vec_f32 *)(inb + 2)); + + vec_mult4a_mma(&out[0], in0 + 0, in1 + 0, inp); + vec_mult4a_mma(&out[4], in0 + 2, in1 + 2, inp); +} + +FORCEINLINE void vec_loadN_mult22a_mma(__vector_quad *out, vec_bf16 *ina, vec_bf16 *inb, vec_bf16 inp, BLASLONG n) +{ + vec_bf16 in0 = vec_loadN(ina, n); + vec_bf16 in1 = vec_loadN(inb, n); + + vec_mult2a_mma(out, in0, in1, inp); +} + +FORCEINLINE void vec_mult11b_mma(__vector_quad *out, vec_bf16 in0, vec_bf16 in1, vec_bf16 inp) +{ + vec_bf16 in00 = vec_mergeh(in0, in1); + + __builtin_mma_xvbf16ger2pp(out, (vec_uc8)inp, (vec_uc8)in00); +} + +FORCEINLINE void vec_mult2b_mma(__vector_quad *out, vec_bf16 in0, vec_bf16 in1, vec_bf16 inp) +{ + vec_bf16 in01 = vec_mergel(in0, in1); + + vec_mult11b_mma(&out[0], in0, in1, inp); + + __builtin_mma_xvbf16ger2pp(&out[1], (vec_uc8)inp, (vec_uc8)in01); +} + +FORCEINLINE void vec_mult4b_mma(__vector_quad *out, vec_bf16 *in0, vec_bf16 *in1, vec_bf16 inp) +{ + vec_mult2b_mma(out + 0, in0[0], in1[0], inp); + vec_mult2b_mma(out + 2, in0[1], in1[1], inp); +} + +FORCEINLINE void vec_loadN_mult11b_mma(__vector_quad *out, vec_bf16 *ina, vec_bf16 *inb, vec_bf16 inp, BLASLONG n) +{ + vec_bf16 in0 = vec_loadN(ina, n); + vec_bf16 in1 = vec_loadN(inb, n); + + vec_mult11b_mma(out, in0, in1, inp); +} + +FORCEINLINE void vec_load_mult22b_mma(__vector_quad *out, vec_bf16 *ina, vec_bf16 *inb, vec_bf16 inp) +{ + vec_bf16 in0 = (vec_bf16)vec_load_vec(ina); + vec_bf16 in1 = (vec_bf16)vec_load_vec(inb); + + vec_mult2b_mma(out, in0, in1, inp); +} + +FORCEINLINE void vec_load_mult28b_mma(__vector_quad *out, vec_bf16 *ina, vec_bf16 *inb, vec_bf16 inp) +{ + vec_bf16 in0[4], in1[4]; + + vec_load_pair((vec_f32 *)(in0 + 0), (vec_f32 *)(ina + 0)); + vec_load_pair((vec_f32 *)(in1 + 0), (vec_f32 *)(inb + 0)); + vec_load_pair((vec_f32 *)(in0 + 2), (vec_f32 *)(ina + 2)); + vec_load_pair((vec_f32 *)(in1 + 2), (vec_f32 *)(inb + 2)); + + vec_mult4b_mma(&out[0], in0 + 0, in1 + 0, inp); + vec_mult4b_mma(&out[4], in0 + 2, in1 + 2, inp); +} + +FORCEINLINE void vec_loadN_mult22b_mma(__vector_quad *out, vec_bf16 *ina, vec_bf16 *inb, vec_bf16 inp, BLASLONG n) +{ + vec_bf16 in0 = vec_loadN(ina, n); + vec_bf16 in1 = vec_loadN(inb, n); + + vec_mult2b_mma(out, in0, in1, inp); +} + +FORCEINLINE void vec_load4_pair(vec_f32 *vy0, vec_f32 *v_y) +{ + vec_load_pair(vy0 + 0, v_y + 0); + vec_load_pair(vy0 + 2, v_y + 2); + vec_load_pair(vy0 + 4, v_y + 4); + vec_load_pair(vy0 + 6, v_y + 6); +} + +FORCEINLINE void vec_store4_pair(vec_f32 *v_y, vec_f32 *vy0) +{ + vec_store_pair(v_y + 0, vy0 + 0); + vec_store_pair(v_y + 2, vy0 + 2); + vec_store_pair(v_y + 4, vy0 + 4); + vec_store_pair(v_y + 6, vy0 + 6); +} + +#endif diff --git a/kernel/power/sbgemv_n.c b/kernel/power/sbgemv_n.c index c7559a47c..4768be31f 100644 --- a/kernel/power/sbgemv_n.c +++ b/kernel/power/sbgemv_n.c @@ -87,6 +87,10 @@ static void BF16GEMV_N_beta(BLASLONG n, FLOAT *output_vector, FLOAT *input_vecto } } +#if (defined(_ARCH_PWR10) && (defined(USE_BFGEMV_8_N_MMA) || (!defined(USE_BFGEMV_N_MMA) && defined(USE_BFGEMV_8_N_VSX)))) || (!defined(_ARCH_PWR10) && defined(USE_BFGEMV_8_N_VSX)) +#define USE_N_8 +#endif + int CNAME(BLASLONG m, BLASLONG n, FLOAT alpha, IFLOAT *a, BLASLONG lda, IFLOAT *x, BLASLONG inc_x, FLOAT beta, FLOAT *y, BLASLONG inc_y) { IFLOAT *x_ptr, *ap[4]; @@ -100,7 +104,9 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT alpha, IFLOAT *a, BLASLONG lda, IFLOAT * y_ptr = y; BLASLONG lda4 = lda << 2; +#ifdef USE_N_8 BLASLONG lda8 = lda << 3; +#endif BLASLONG NB = NBMAX; BLASLONG m2 = (m & (NBMAX - 1)); @@ -126,6 +132,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT alpha, IFLOAT *a, BLASLONG lda, IFLOAT * ap[3] = ap[2] + lda; if (inc_x == 1) { +#ifdef USE_N_8 for (BLASLONG j = 0; j + 8 <= n; j += 8) { BF16GEMV_N_8(NB, ap, x_ptr, ybuffer, lda4, alpha); ap[0] += lda8; @@ -135,9 +142,16 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT alpha, IFLOAT *a, BLASLONG lda, IFLOAT * x_ptr += 8; } if (n & 4) { +#else + for (BLASLONG j = 0; j + 4 <= n; j += 4) { +#endif BF16GEMV_N_4(NB, ap, x_ptr, ybuffer, alpha); ap[0] += lda4; ap[1] += lda4; +#ifndef USE_N_8 + ap[2] += lda4; + ap[3] += lda4; +#endif x_ptr += 4; } if (n & 2) { @@ -149,6 +163,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT alpha, IFLOAT *a, BLASLONG lda, IFLOAT * BF16GEMV_N_1(NB, ap, x_ptr, ybuffer, alpha); } } else { +#ifdef USE_N_8 for (BLASLONG j = 0; j + 8 <= n; j += 8) { copy_x(8, x_ptr, xbuffer, inc_x); BF16GEMV_N_8(NB, ap, xbuffer, ybuffer, lda4, alpha); @@ -159,10 +174,17 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT alpha, IFLOAT *a, BLASLONG lda, IFLOAT * x_ptr += 8 * inc_x; } if (n & 4) { +#else + for (BLASLONG j = 0; j + 4 <= n; j += 4) { +#endif copy_x(4, x_ptr, xbuffer, inc_x); BF16GEMV_N_4(NB, ap, xbuffer, ybuffer, alpha); ap[0] += lda4; ap[1] += lda4; +#ifndef USE_N_8 + ap[2] += lda4; + ap[3] += lda4; +#endif x_ptr += 4 * inc_x; } if (n & 2) { diff --git a/kernel/power/sbgemv_n_power10.c b/kernel/power/sbgemv_n_power10.c index fc83b38c3..7b2beb0c7 100644 --- a/kernel/power/sbgemv_n_power10.c +++ b/kernel/power/sbgemv_n_power10.c @@ -25,9 +25,309 @@ 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 "sbgemv_common.c" +#ifndef SBGEMV_N_MMA_C +#define SBGEMV_N_MMA_C +#define USE_BFGEMV_N_MMA + +#ifdef USE_BFGEMV_N_MMA +#include "sbgemv_common_power10.c" + +#ifndef BF16GEMV_N_X +#define BF16GEMV_N_X +#define BF16GEMV_N_8 BF16GEMV_N_MMA_8 +#define BF16GEMV_N_4 BF16GEMV_N_MMA_4 +#define BF16GEMV_N_2 BF16GEMV_N_MMA_2 +#define BF16GEMV_N_1 BF16GEMV_N_MMA_1 +#endif + +#define USE_BFGEMV_8_N_MMA + +static void BF16GEMV_N_MMA_1(BLASLONG n, IFLOAT **ap, IFLOAT *xo, FLOAT *y, FLOAT alpha) +{ + IFLOAT *a0; + __vector_quad temp[2*4]; + vec_f32 temp0[8*4], vy0[2*4]; + vec_f32 v_alpha = { alpha, alpha, alpha, alpha }; + + a0 = ap[0]; + + vec_bf16 *va0 = (vec_bf16 *)a0; + + vec_bf16 *x_bf = (vec_bf16 *)(xo); + vec_bf16 v_x0 = vec_loadN(x_bf, 1); + + vec_f32 *v_y = (vec_f32 *)y; + BLASLONG n8 = n / 8; + BLASLONG i = 0; + + for (; i + 4 <= n8; i += 4) { + vec_load4_pair(vy0, &v_y[(i * 2) + 0]); + + vec_load_mult18_mma(&temp[0], &va0[i + 0], v_x0); + + vec_reduce8_mma(&temp[0], temp0, v_alpha, vy0); + + vec_store4_pair(&v_y[(i * 2) + 0], vy0); + } + + for (; i < n8; i++) { + vec_load_pair(vy0, &v_y[(i * 2) + 0]); + + vec_load_mult12_mma(&temp[0], &va0[i], v_x0); + + vec_reduce2_mma(&temp[0], temp0, v_alpha, vy0); + + vec_store_pair(&v_y[(i * 2) + 0], vy0); + } + + n &= 7; + if (n > 4) { + BLASLONG n3 = n & 3; + vec_loadN2_f32(vy0, &v_y[(i * 2) + 0], n3); + + vec_loadN_mult12_mma(&temp[0], &va0[i], v_x0, n); + + vec_reduce2_mma(&temp[0], temp0, v_alpha, vy0); + + vec_storeN2_f32(vy0, &v_y[(i * 2) + 0], n3); + } else if (n) { + vy0[0] = vec_loadN_f32(&v_y[(i * 2) + 0], n); + + vec_loadN_mult11_mma(&temp[0], &va0[i], v_x0, n); + + vec_reduce1_mma(&temp[0], temp0, v_alpha, vy0); + + vec_storeN_f32(vy0[0], &v_y[(i * 2) + 0], n); + } +} + +static void BF16GEMV_N_MMA_2(BLASLONG n, IFLOAT **ap, IFLOAT *xo, FLOAT *y, FLOAT alpha) +{ + IFLOAT *a0, *a1; + __vector_quad temp[2*4]; + vec_f32 temp0[8*4], vy0[2*4]; + vec_f32 v_alpha = { alpha, alpha, alpha, alpha }; + + a0 = ap[0]; + a1 = ap[1]; + + vec_bf16 *va0 = (vec_bf16 *)a0; + vec_bf16 *va1 = (vec_bf16 *)a1; + + vec_bf16 *x_bf = (vec_bf16 *)(xo); + vec_bf16 v_x0 = vec_loadN(x_bf, 2); + + vec_f32 *v_y = (vec_f32 *)y; + BLASLONG n8 = n / 8; + BLASLONG i = 0; + + for (; i + 4 <= n8; i += 4) { + vec_load4_pair(vy0, &v_y[(i * 2) + 0]); + + vec_load_mult28a_mma(&temp[0], &va0[i + 0], &va1[i + 0], v_x0); + + vec_reduce8_mma(&temp[0], temp0, v_alpha, vy0); + + vec_store4_pair(&v_y[(i * 2) + 0], vy0); + } + + for (; i < n8; i++) { + vec_load_pair(vy0, &v_y[(i * 2) + 0]); + + vec_load_mult22a_mma(&temp[0], &va0[i], &va1[i], v_x0); + + vec_reduce2_mma(&temp[0], temp0, v_alpha, vy0); + + vec_store_pair(&v_y[(i * 2) + 0], vy0); + } + + n &= 7; + if (n > 4) { + BLASLONG n3 = n & 3; + vec_loadN2_f32(vy0, &v_y[(i * 2) + 0], n3); + + vec_loadN_mult22a_mma(&temp[0], &va0[i], &va1[i], v_x0, n); + + vec_reduce2_mma(&temp[0], temp0, v_alpha, vy0); + + vec_storeN2_f32(vy0, &v_y[(i * 2) + 0], n3); + } else if (n) { + vy0[0] = vec_loadN_f32(&v_y[(i * 2) + 0], n); + + vec_loadN_mult11a_mma(&temp[0], &va0[i], &va1[i], v_x0, n); + + vec_reduce1_mma(&temp[0], temp0, v_alpha, vy0); + + vec_storeN_f32(vy0[0], &v_y[(i * 2) + 0], n); + } +} + +static void BF16GEMV_N_MMA_4(BLASLONG n, IFLOAT **ap, IFLOAT *xo, FLOAT *y, FLOAT alpha) +{ + IFLOAT *a0, *a1, *a2, *a3; + __vector_quad temp[2*4]; + vec_f32 temp0[8*4], vy0[2*4]; + vec_f32 v_alpha = { alpha, alpha, alpha, alpha }; + + a0 = ap[0]; + a1 = ap[1]; + a2 = ap[2]; + a3 = ap[3]; + + vec_bf16 *va0 = (vec_bf16 *)a0; + vec_bf16 *va1 = (vec_bf16 *)a1; + vec_bf16 *va2 = (vec_bf16 *)a2; + vec_bf16 *va3 = (vec_bf16 *)a3; + + vec_bf16 *x_bf = (vec_bf16 *)(xo); + vec_bf16 v_x00 = vec_loadN(x_bf, 4); + + vec_bf16 v_x01 = (vec_bf16)vec_splat((vec_f32)v_x00, 1); + + vec_f32 *v_y = (vec_f32 *)y; + BLASLONG n8 = n / 8; + BLASLONG i = 0; + + for (; i + 4 <= n8; i += 4) { + vec_load4_pair(vy0, &v_y[(i * 2) + 0]); + + vec_load_mult28a_mma(&temp[0], &va0[i + 0], &va1[i + 0], v_x00); + vec_load_mult28b_mma(&temp[0], &va2[i + 0], &va3[i + 0], v_x01); + + vec_reduce8_mma(&temp[0], temp0, v_alpha, vy0); + + vec_store4_pair(&v_y[(i * 2) + 0], vy0); + } + + for (; i < n8; i++) { + vec_load_pair(vy0, &v_y[(i * 2) + 0]); + + vec_load_mult22a_mma(&temp[0], &va0[i], &va1[i], v_x00); + vec_load_mult22b_mma(&temp[0], &va2[i], &va3[i], v_x01); + + vec_reduce2_mma(&temp[0], temp0, v_alpha, vy0); + + vec_store_pair(&v_y[(i * 2) + 0], vy0); + } + + n &= 7; + if (n > 4) { + BLASLONG n3 = n & 3; + vec_loadN2_f32(vy0, &v_y[(i * 2) + 0], n3); + + vec_loadN_mult22a_mma(&temp[0], &va0[i], &va1[i], v_x00, n); + vec_loadN_mult22b_mma(&temp[0], &va2[i], &va3[i], v_x01, n); + + vec_reduce2_mma(&temp[0], temp0, v_alpha, vy0); + + vec_storeN2_f32(vy0, &v_y[(i * 2) + 0], n3); + } else if (n) { + vy0[0] = vec_loadN_f32(&v_y[(i * 2) + 0], n); + + vec_loadN_mult11a_mma(&temp[0], &va0[i], &va1[i], v_x00, n); + vec_loadN_mult11b_mma(&temp[0], &va2[i], &va3[i], v_x01, n); + + vec_reduce1_mma(&temp[0], temp0, v_alpha, vy0); + + vec_storeN_f32(vy0[0], &v_y[(i * 2) + 0], n); + } +} + +#ifdef USE_BFGEMV_8_N_MMA +static void BF16GEMV_N_MMA_8(BLASLONG n, IFLOAT **ap, IFLOAT *xo, FLOAT *y, BLASLONG lda4, FLOAT alpha) +{ + IFLOAT *a0, *a1, *a2, *a3, *b0, *b1, *b2, *b3; + __vector_quad temp[2*4]; + vec_f32 temp0[8*4], vy0[2*4]; + vec_f32 v_alpha = { alpha, alpha, alpha, alpha }; + + a0 = ap[0]; + a1 = ap[1]; + a2 = ap[2]; + a3 = ap[3]; + b0 = a0 + lda4; + b1 = a1 + lda4; + b2 = a2 + lda4; + b3 = a3 + lda4; + + vec_bf16 *va0 = (vec_bf16 *)a0; + vec_bf16 *va1 = (vec_bf16 *)a1; + vec_bf16 *va2 = (vec_bf16 *)a2; + vec_bf16 *va3 = (vec_bf16 *)a3; + vec_bf16 *vb0 = (vec_bf16 *)b0; + vec_bf16 *vb1 = (vec_bf16 *)b1; + vec_bf16 *vb2 = (vec_bf16 *)b2; + vec_bf16 *vb3 = (vec_bf16 *)b3; + + vec_bf16 *x_bf = (vec_bf16 *)(xo); + vec_bf16 v_x00 = (vec_bf16)vec_load_vec(x_bf); + + vec_bf16 v_x01 = (vec_bf16)vec_splat((vec_f32)v_x00, 1); + vec_bf16 v_x02 = (vec_bf16)vec_splat((vec_f32)v_x00, 2); + vec_bf16 v_x03 = (vec_bf16)vec_splat((vec_f32)v_x00, 3); + + vec_f32 *v_y = (vec_f32 *)y; + BLASLONG n8 = n / 8; + BLASLONG i = 0; + + for (; i + 4 <= n8; i += 4) { + vec_load4_pair(vy0, &v_y[(i * 2) + 0]); + + vec_load_mult28a_mma(&temp[0], &va0[i + 0], &va1[i + 0], v_x00); + vec_load_mult28b_mma(&temp[0], &va2[i + 0], &va3[i + 0], v_x01); + vec_load_mult28b_mma(&temp[0], &vb0[i + 0], &vb1[i + 0], v_x02); + vec_load_mult28b_mma(&temp[0], &vb2[i + 0], &vb3[i + 0], v_x03); + + vec_reduce8_mma(&temp[0], temp0, v_alpha, vy0); + + vec_store4_pair(&v_y[(i * 2) + 0], vy0); + } + + for (; i < n8; i++) { + vec_load_pair(vy0, &v_y[(i * 2) + 0]); + + vec_load_mult22a_mma(&temp[0], &va0[i], &va1[i], v_x00); + vec_load_mult22b_mma(&temp[0], &va2[i], &va3[i], v_x01); + vec_load_mult22b_mma(&temp[0], &vb0[i], &vb1[i], v_x02); + vec_load_mult22b_mma(&temp[0], &vb2[i], &vb3[i], v_x03); + + vec_reduce2_mma(&temp[0], temp0, v_alpha, vy0); + + vec_store_pair(&v_y[(i * 2) + 0], vy0); + } + + n &= 7; + if (n > 4) { + BLASLONG n3 = n & 3; + vec_loadN2_f32(vy0, &v_y[(i * 2) + 0], n3); + + vec_loadN_mult22a_mma(&temp[0], &va0[i], &va1[i], v_x00, n); + vec_loadN_mult22b_mma(&temp[0], &va2[i], &va3[i], v_x01, n); + vec_loadN_mult22b_mma(&temp[0], &vb0[i], &vb1[i], v_x02, n); + vec_loadN_mult22b_mma(&temp[0], &vb2[i], &vb3[i], v_x03, n); + + vec_reduce2_mma(&temp[0], temp0, v_alpha, vy0); + + vec_storeN2_f32(vy0, &v_y[(i * 2) + 0], n3); + } else if (n) { + vy0[0] = vec_loadN_f32(&v_y[(i * 2) + 0], n); + + vec_loadN_mult11a_mma(&temp[0], &va0[i], &va1[i], v_x00, n); + vec_loadN_mult11b_mma(&temp[0], &va2[i], &va3[i], v_x01, n); + vec_loadN_mult11b_mma(&temp[0], &vb0[i], &vb1[i], v_x02, n); + vec_loadN_mult11b_mma(&temp[0], &vb2[i], &vb3[i], v_x03, n); + + vec_reduce1_mma(&temp[0], temp0, v_alpha, vy0); + + vec_storeN_f32(vy0[0], &v_y[(i * 2) + 0], n); + } +} +#endif + +#include "sbgemv_n.c" +#else #include "sbgemv_n_vsx.c" - -//#include "sbgemv_n.c" +#endif +#endif diff --git a/kernel/power/sbgemv_n_vsx.c b/kernel/power/sbgemv_n_vsx.c index cab2316d4..e8f6dca9f 100644 --- a/kernel/power/sbgemv_n_vsx.c +++ b/kernel/power/sbgemv_n_vsx.c @@ -25,12 +25,20 @@ 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. *****************************************************************************/ -#ifndef SBGEMV_N_VSX -#define SBGEMV_N_VSX +#ifndef SBGEMV_N_VSX_C +#define SBGEMV_N_VSX_C #include "sbgemv_common.c" -#define NBMAX 4096 +#ifndef BF16GEMV_N_X +#define BF16GEMV_N_X +#define BF16GEMV_N_8 BF16GEMV_N_VSX_8 +#define BF16GEMV_N_4 BF16GEMV_N_VSX_4 +#define BF16GEMV_N_2 BF16GEMV_N_VSX_2 +#define BF16GEMV_N_1 BF16GEMV_N_VSX_1 +#endif + +#define USE_BFGEMV_8_N_VSX static void BF16GEMV_N_VSX_1(BLASLONG n, IFLOAT **ap, IFLOAT *xo, FLOAT *y, FLOAT alpha) { @@ -70,11 +78,11 @@ static void BF16GEMV_N_VSX_1(BLASLONG n, IFLOAT **ap, IFLOAT *xo, FLOAT *y, FLOA vec_storeN2_f32(vy0, &v_y[(i * 2) + 0], n3); } else if (n) { - vec_f32 vy0 = vec_loadN_f32(&v_y[(i * 2) + 0], n); + vy0[0] = vec_loadN_f32(&v_y[(i * 2) + 0], n); - vy0 += vec_loadNHi_multi2(v_x0, &va0[i], n, zero); + vy0[0] += vec_loadNHi_mult2(v_x0, &va0[i], n, zero); - vec_storeN_f32(vy0, &v_y[(i * 2) + 0], n); + vec_storeN_f32(vy0[0], &v_y[(i * 2) + 0], n); } } @@ -121,12 +129,12 @@ static void BF16GEMV_N_VSX_2(BLASLONG n, IFLOAT **ap, IFLOAT *xo, FLOAT *y, FLOA vec_storeN2_f32(vy0, &v_y[(i * 2) + 0], n3); } else if (n) { - vec_f32 vy0 = vec_loadN_f32(&v_y[(i * 2) + 0], n); + vy0[0] = vec_loadN_f32(&v_y[(i * 2) + 0], n); - vy0 += vec_loadNHi_multi2(v_x0, &va0[i], n, zero); - vy0 += vec_loadNHi_multi2(v_x1, &va1[i], n, zero); + vy0[0] += vec_loadNHi_mult2(v_x0, &va0[i], n, zero); + vy0[0] += vec_loadNHi_mult2(v_x1, &va1[i], n, zero); - vec_storeN_f32(vy0, &v_y[(i * 2) + 0], n); + vec_storeN_f32(vy0[0], &v_y[(i * 2) + 0], n); } } @@ -183,17 +191,18 @@ static void BF16GEMV_N_VSX_4(BLASLONG n, IFLOAT **ap, IFLOAT *xo, FLOAT *y, FLOA vec_storeN2_f32(vy0, &v_y[(i * 2) + 0], n3); } else if (n) { - vec_f32 vy0 = vec_loadN_f32(&v_y[(i * 2) + 0], n); + vy0[0] = vec_loadN_f32(&v_y[(i * 2) + 0], n); - vy0 += vec_loadNHi_multi2(v_x0, &va0[i], n, zero); - vy0 += vec_loadNHi_multi2(v_x1, &va1[i], n, zero); - vy0 += vec_loadNHi_multi2(v_x2, &va2[i], n, zero); - vy0 += vec_loadNHi_multi2(v_x3, &va3[i], n, zero); + vy0[0] += vec_loadNHi_mult2(v_x0, &va0[i], n, zero); + vy0[0] += vec_loadNHi_mult2(v_x1, &va1[i], n, zero); + vy0[0] += vec_loadNHi_mult2(v_x2, &va2[i], n, zero); + vy0[0] += vec_loadNHi_mult2(v_x3, &va3[i], n, zero); - vec_storeN_f32(vy0, &v_y[(i * 2) + 0], n); + vec_storeN_f32(vy0[0], &v_y[(i * 2) + 0], n); } } +#ifdef USE_BFGEMV_8_N_VSX static void BF16GEMV_N_VSX_8(BLASLONG n, IFLOAT **ap, IFLOAT *xo, FLOAT *y, BLASLONG lda4, FLOAT alpha) { IFLOAT *a0, *a1, *a2, *a3, *b0, *b1, *b2, *b3; @@ -270,25 +279,21 @@ static void BF16GEMV_N_VSX_8(BLASLONG n, IFLOAT **ap, IFLOAT *xo, FLOAT *y, BLAS vec_storeN2_f32(vy0, &v_y[(i * 2) + 0], n3); } else if (n) { - vec_f32 vy0 = vec_loadN_f32(&v_y[(i * 2) + 0], n); + vy0[0] = vec_loadN_f32(&v_y[(i * 2) + 0], n); - vy0 += vec_loadNHi_multi2(v_x0, &va0[i], n, zero); - vy0 += vec_loadNHi_multi2(v_x1, &va1[i], n, zero); - vy0 += vec_loadNHi_multi2(v_x2, &va2[i], n, zero); - vy0 += vec_loadNHi_multi2(v_x3, &va3[i], n, zero); - vy0 += vec_loadNHi_multi2(v_x4, &vb0[i], n, zero); - vy0 += vec_loadNHi_multi2(v_x5, &vb1[i], n, zero); - vy0 += vec_loadNHi_multi2(v_x6, &vb2[i], n, zero); - vy0 += vec_loadNHi_multi2(v_x7, &vb3[i], n, zero); + vy0[0] += vec_loadNHi_mult2(v_x0, &va0[i], n, zero); + vy0[0] += vec_loadNHi_mult2(v_x1, &va1[i], n, zero); + vy0[0] += vec_loadNHi_mult2(v_x2, &va2[i], n, zero); + vy0[0] += vec_loadNHi_mult2(v_x3, &va3[i], n, zero); + vy0[0] += vec_loadNHi_mult2(v_x4, &vb0[i], n, zero); + vy0[0] += vec_loadNHi_mult2(v_x5, &vb1[i], n, zero); + vy0[0] += vec_loadNHi_mult2(v_x6, &vb2[i], n, zero); + vy0[0] += vec_loadNHi_mult2(v_x7, &vb3[i], n, zero); - vec_storeN_f32(vy0, &v_y[(i * 2) + 0], n); + vec_storeN_f32(vy0[0], &v_y[(i * 2) + 0], n); } } - -#define BF16GEMV_N_8 BF16GEMV_N_VSX_8 -#define BF16GEMV_N_4 BF16GEMV_N_VSX_4 -#define BF16GEMV_N_2 BF16GEMV_N_VSX_2 -#define BF16GEMV_N_1 BF16GEMV_N_VSX_1 +#endif #include "sbgemv_n.c" #endif diff --git a/kernel/power/sbgemv_t.c b/kernel/power/sbgemv_t.c index f0c79fe77..4cc8f060e 100644 --- a/kernel/power/sbgemv_t.c +++ b/kernel/power/sbgemv_t.c @@ -27,6 +27,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #ifndef SBGEMV_T_COMMON_C #define SBGEMV_T_COMMON_C + +#if (defined(_ARCH_PWR10) && (defined(USE_BFGEMV_8_T_MMA) || (!defined(USE_BFGEMV_N_MMA) && defined(USE_BFGEMV_8_T_VSX)))) || (!defined(_ARCH_PWR10) && defined(USE_BFGEMV_8_T_VSX)) +#define USE_T_8 +#endif + int CNAME(BLASLONG m, BLASLONG n, FLOAT alpha, IFLOAT *a, BLASLONG lda, IFLOAT *x, BLASLONG inc_x, FLOAT beta, FLOAT *y, BLASLONG inc_y) { IFLOAT *xbuffer, *a_ptr; @@ -39,7 +44,9 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT alpha, IFLOAT *a, BLASLONG lda, IFLOAT * xbuffer = buffer; BLASLONG lda4 = lda << 2; +#ifdef USE_T_8 BLASLONG lda8 = lda << 3; +#endif BLASLONG NB = NBMAX; BLASLONG m2 = (m & (NBMAX - 1)); @@ -60,12 +67,16 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT alpha, IFLOAT *a, BLASLONG lda, IFLOAT * } if (inc_y == 1) { +#ifdef USE_T_8 for (BLASLONG j = 0; j + 8 <= n; j += 8) { BF16GEMV_T_8(NB, lda, a_ptr, xbuffer, y_ptr, alpha, beta); y_ptr += 8; a_ptr += lda8; } if (n & 4) { +#else + for (BLASLONG j = 0; j + 4 <= n; j += 4) { +#endif BF16GEMV_T_4(NB, lda, a_ptr, xbuffer, y_ptr, alpha, beta); y_ptr += 4; a_ptr += lda4; @@ -79,6 +90,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT alpha, IFLOAT *a, BLASLONG lda, IFLOAT * BF16GEMV_T_1(NB, lda, a_ptr, xbuffer, y_ptr, alpha, beta); } } else { +#ifdef USE_T_8 for (BLASLONG j = 0; j + 8 <= n; j += 8) { memset(ybuffer, 0, sizeof(FLOAT) * 8); BF16GEMV_T_8(NB, lda, a_ptr, xbuffer, ybuffer, alpha, beta); @@ -87,6 +99,9 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT alpha, IFLOAT *a, BLASLONG lda, IFLOAT * a_ptr += lda8; } if (n & 4) { +#else + for (BLASLONG j = 0; j + 4 <= n; j += 4) { +#endif memset(ybuffer, 0, sizeof(FLOAT) * 4); BF16GEMV_T_4(NB, lda, a_ptr, xbuffer, ybuffer, alpha, beta); copy_y(4, ybuffer, y_ptr, inc_y, beta); diff --git a/kernel/power/sbgemv_t_power10.c b/kernel/power/sbgemv_t_power10.c index 08bc4237c..810287e89 100644 --- a/kernel/power/sbgemv_t_power10.c +++ b/kernel/power/sbgemv_t_power10.c @@ -25,8 +25,334 @@ 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 "sbgemv_common.c" +#ifndef SBGEMV_T_MMA_C +#define SBGEMV_T_MMA_C +#define USE_BFGEMV_T_MMA + +#ifdef USE_BFGEMV_T_MMA +#include "sbgemv_common_power10.c" + +#ifndef BF16GEMV_T_X +#define BF16GEMV_T_X +#define BF16GEMV_T_8 BF16GEMV_T_MMA_8 +#define BF16GEMV_T_4 BF16GEMV_T_MMA_4 +#define BF16GEMV_T_2 BF16GEMV_T_MMA_2 +#define BF16GEMV_T_1 BF16GEMV_T_MMA_1 +#endif + +#define USE_BFGEMV_8_T_MMA + +static void BF16GEMV_T_MMA_1(BLASLONG n, BLASLONG lda, IFLOAT *ap, IFLOAT *x, FLOAT *y, FLOAT alpha, FLOAT beta) +{ + IFLOAT *a0; + vec_bf16 *va0, *v_x; + __vector_quad temp0; + vec_f32 temp00[4]; + vec_bf16 inp[2]; + + __builtin_mma_xxsetaccz(&temp0); + + a0 = ap; + va0 = (vec_bf16 *)a0; + v_x = (vec_bf16 *)x; + BLASLONG n8 = n / 8; + BLASLONG i = 0; + + for (; i + 2 <= n8; i += 2) { + vec_load_pair((vec_f32 *)inp, (vec_f32 *)&v_x[i]); + + vec_load_mult2_mma(&temp0, &va0[i + 0], inp); + } + + if (n8 & 1) { + inp[0] = (vec_bf16)vec_load_vec(&v_x[i]); + + vec_load_mult_mma(&temp0, &va0[i], inp[0]); + + i++; + } + + n &= 7; + if (n) { + inp[0] = vec_loadN(&v_x[i], n); + + vec_loadN_mult_mma(&temp0, &va0[i], inp[0], n); + } + + __builtin_mma_disassemble_acc((void*)temp00, &temp0); + + y[0] = (alpha * (temp00[0][0] + temp00[1][1] + temp00[2][2] + temp00[3][3])) + (beta * y[0]); +} + +static void BF16GEMV_T_MMA_2(BLASLONG n, BLASLONG lda, IFLOAT *ap, IFLOAT *x, FLOAT *y, FLOAT alpha, FLOAT beta) +{ + IFLOAT *a0, *a1; + vec_bf16 *va0, *va1, *v_x; + __vector_quad temp0, temp1; + vec_f32 temp00[4], temp01[4]; + vec_bf16 inp[2]; + + __builtin_mma_xxsetaccz(&temp0); + __builtin_mma_xxsetaccz(&temp1); + + a0 = ap; + a1 = ap + lda; + va0 = (vec_bf16 *)a0; + va1 = (vec_bf16 *)a1; + v_x = (vec_bf16 *)x; + BLASLONG n8 = n / 8; + BLASLONG i = 0; + + for (; i + 2 <= n8; i += 2) { + vec_load_pair((vec_f32 *)inp, (vec_f32 *)&v_x[i]); + + vec_load_mult2_mma(&temp0, &va0[i + 0], inp); + vec_load_mult2_mma(&temp1, &va1[i + 0], inp); + } + + if (n8 & 1) { + inp[0] = (vec_bf16)vec_load_vec(&v_x[i]); + + vec_load_mult_mma(&temp0, &va0[i], inp[0]); + vec_load_mult_mma(&temp1, &va1[i], inp[0]); + + i++; + } + + n &= 7; + if (n) { + inp[0] = vec_loadN(&v_x[i], n); + + vec_loadN_mult_mma(&temp0, &va0[i], inp[0], n); + vec_loadN_mult_mma(&temp1, &va1[i], inp[0], n); + } + + __builtin_mma_disassemble_acc((void*)temp00, &temp0); + __builtin_mma_disassemble_acc((void*)temp01, &temp1); + + y[0] = (alpha * (temp00[0][0] + temp00[1][1] + temp00[2][2] + temp00[3][3])) + (beta * y[0]); + y[1] = (alpha * (temp01[0][0] + temp01[1][1] + temp01[2][2] + temp01[3][3])) + (beta * y[1]); +} + +static void BF16GEMV_T_MMA_4(BLASLONG n, BLASLONG lda, IFLOAT *ap, IFLOAT *x, FLOAT *y, FLOAT alpha, FLOAT beta) +{ + IFLOAT *a0, *a1, *a2, *a3; + vec_bf16 *va0, *va1, *va2, *va3, *v_x; + __vector_quad temp0, temp1, temp2, temp3; + vec_f32 temp00[4], temp01[4], temp02[4], temp03[4]; + vec_bf16 inp[2]; + + __builtin_mma_xxsetaccz(&temp0); + __builtin_mma_xxsetaccz(&temp1); + __builtin_mma_xxsetaccz(&temp2); + __builtin_mma_xxsetaccz(&temp3); + + a0 = ap; + a1 = ap + lda; + a2 = a1 + lda; + a3 = a2 + lda; + va0 = (vec_bf16 *)a0; + va1 = (vec_bf16 *)a1; + va2 = (vec_bf16 *)a2; + va3 = (vec_bf16 *)a3; + v_x = (vec_bf16 *)x; + BLASLONG n8 = n / 8; + BLASLONG i = 0; + + for (; i + 2 <= n8; i += 2) { + vec_load_pair((vec_f32 *)inp, (vec_f32 *)&v_x[i]); + + vec_load_mult2_mma(&temp0, &va0[i + 0], inp); + vec_load_mult2_mma(&temp1, &va1[i + 0], inp); + vec_load_mult2_mma(&temp2, &va2[i + 0], inp); + vec_load_mult2_mma(&temp3, &va3[i + 0], inp); + } + + if (n8 & 1) { + inp[0] = (vec_bf16)vec_load_vec(&v_x[i]); + + vec_load_mult_mma(&temp0, &va0[i], inp[0]); + vec_load_mult_mma(&temp1, &va1[i], inp[0]); + vec_load_mult_mma(&temp2, &va2[i], inp[0]); + vec_load_mult_mma(&temp3, &va3[i], inp[0]); + + i++; + } + + n &= 7; + if (n) { + inp[0] = vec_loadN(&v_x[i], n); + + vec_loadN_mult_mma(&temp0, &va0[i], inp[0], n); + vec_loadN_mult_mma(&temp1, &va1[i], inp[0], n); + vec_loadN_mult_mma(&temp2, &va2[i], inp[0], n); + vec_loadN_mult_mma(&temp3, &va3[i], inp[0], n); + } + + __builtin_mma_disassemble_acc((void*)temp00, &temp0); + __builtin_mma_disassemble_acc((void*)temp01, &temp1); + __builtin_mma_disassemble_acc((void*)temp02, &temp2); + __builtin_mma_disassemble_acc((void*)temp03, &temp3); + + vec_f32 t0, t1, t2, t3, t4, t5, t6, t7; + vec_f32 a = { alpha, alpha, alpha, alpha }; + vec_f32 b = { beta, beta, beta, beta }; + vec_f32 *v_y = (vec_f32 *) y; + + t0 = vec_mergeh(temp00[0], temp01[0]); + t1 = vec_mergeh(temp02[0], temp03[0]); + t2 = vec_mergeo(temp00[1], temp01[1]); + t3 = vec_mergeo(temp02[1], temp03[1]); + t4 = vec_mergel(temp00[2], temp01[2]); + t5 = vec_mergel(temp02[2], temp03[2]); + t6 = vec_mergeo(temp00[3], temp01[3]); + t7 = vec_mergeo(temp02[3], temp03[3]); + t0 = vec_xxpermdi(t0, t1, 0); + t2 = vec_xxpermdi(t2, t3, 0); + t4 = vec_xxpermdi(t4, t5, 0); + t6 = vec_xxpermdi(t6, t7, 3); + + t0 += t2 + t4 + t6; + + v_y[0] = (a * t0) + (b * v_y[0]); +} + +#ifdef USE_BFGEMV_8_T_MMA +static void BF16GEMV_T_MMA_8(BLASLONG n, BLASLONG lda, IFLOAT *ap, IFLOAT *x, FLOAT *y, FLOAT alpha, FLOAT beta) +{ + IFLOAT *a0, *a1, *a2, *a3, *a4, *a5, *a6, *a7; + vec_bf16 *va0, *va1, *va2, *va3, *va4, *va5, *va6, *va7, *v_x; + __vector_quad temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7; + vec_f32 temp00[4], temp01[4], temp02[4], temp03[4], temp04[4], temp05[4], temp06[4], temp07[4]; + vec_bf16 inp[2]; + + __builtin_mma_xxsetaccz(&temp0); + __builtin_mma_xxsetaccz(&temp1); + __builtin_mma_xxsetaccz(&temp2); + __builtin_mma_xxsetaccz(&temp3); + __builtin_mma_xxsetaccz(&temp4); + __builtin_mma_xxsetaccz(&temp5); + __builtin_mma_xxsetaccz(&temp6); + __builtin_mma_xxsetaccz(&temp7); + + a0 = ap; + a1 = ap + lda; + a2 = a1 + lda; + a3 = a2 + lda; + a4 = a3 + lda; + a5 = a4 + lda; + a6 = a5 + lda; + a7 = a6 + lda; + va0 = (vec_bf16 *)a0; + va1 = (vec_bf16 *)a1; + va2 = (vec_bf16 *)a2; + va3 = (vec_bf16 *)a3; + va4 = (vec_bf16 *)a4; + va5 = (vec_bf16 *)a5; + va6 = (vec_bf16 *)a6; + va7 = (vec_bf16 *)a7; + v_x = (vec_bf16 *)x; + BLASLONG n8 = n / 8; + BLASLONG i = 0; + + for (; i + 2 <= n8; i += 2) { + vec_load_pair((vec_f32 *)inp, (vec_f32 *)&v_x[i]); + + vec_load_mult2_mma(&temp0, &va0[i + 0], inp); + vec_load_mult2_mma(&temp1, &va1[i + 0], inp); + vec_load_mult2_mma(&temp2, &va2[i + 0], inp); + vec_load_mult2_mma(&temp3, &va3[i + 0], inp); + vec_load_mult2_mma(&temp4, &va4[i + 0], inp); + vec_load_mult2_mma(&temp5, &va5[i + 0], inp); + vec_load_mult2_mma(&temp6, &va6[i + 0], inp); + vec_load_mult2_mma(&temp7, &va7[i + 0], inp); + } + + if (n8 & 1) { + inp[0] = (vec_bf16)vec_load_vec(&v_x[i]); + + vec_load_mult_mma(&temp0, &va0[i], inp[0]); + vec_load_mult_mma(&temp1, &va1[i], inp[0]); + vec_load_mult_mma(&temp2, &va2[i], inp[0]); + vec_load_mult_mma(&temp3, &va3[i], inp[0]); + vec_load_mult_mma(&temp4, &va4[i], inp[0]); + vec_load_mult_mma(&temp5, &va5[i], inp[0]); + vec_load_mult_mma(&temp6, &va6[i], inp[0]); + vec_load_mult_mma(&temp7, &va7[i], inp[0]); + + i++; + } + + n &= 7; + if (n) { + inp[0] = vec_loadN(&v_x[i], n); + + vec_loadN_mult_mma(&temp0, &va0[i], inp[0], n); + vec_loadN_mult_mma(&temp1, &va1[i], inp[0], n); + vec_loadN_mult_mma(&temp2, &va2[i], inp[0], n); + vec_loadN_mult_mma(&temp3, &va3[i], inp[0], n); + vec_loadN_mult_mma(&temp4, &va4[i], inp[0], n); + vec_loadN_mult_mma(&temp5, &va5[i], inp[0], n); + vec_loadN_mult_mma(&temp6, &va6[i], inp[0], n); + vec_loadN_mult_mma(&temp7, &va7[i], inp[0], n); + } + + __builtin_mma_disassemble_acc((void*)temp00, &temp0); + __builtin_mma_disassemble_acc((void*)temp01, &temp1); + __builtin_mma_disassemble_acc((void*)temp02, &temp2); + __builtin_mma_disassemble_acc((void*)temp03, &temp3); + __builtin_mma_disassemble_acc((void*)temp04, &temp4); + __builtin_mma_disassemble_acc((void*)temp05, &temp5); + __builtin_mma_disassemble_acc((void*)temp06, &temp6); + __builtin_mma_disassemble_acc((void*)temp07, &temp7); + + vec_f32 t0, t1, t2, t3, t4, t5, t6, t7, t10, t11, t12, t13, t14, t15, t16, t17; + vec_f32 a = { alpha, alpha, alpha, alpha }; + vec_f32 b = { beta, beta, beta, beta }; + vec_f32 *v_y = (vec_f32 *) y; + + t0 = vec_mergeh(temp00[0], temp01[0]); + t1 = vec_mergeh(temp02[0], temp03[0]); + t2 = vec_mergeo(temp00[1], temp01[1]); + t3 = vec_mergeo(temp02[1], temp03[1]); + t4 = vec_mergel(temp00[2], temp01[2]); + t5 = vec_mergel(temp02[2], temp03[2]); + t6 = vec_mergeo(temp00[3], temp01[3]); + t7 = vec_mergeo(temp02[3], temp03[3]); + t0 = vec_xxpermdi(t0, t1, 0); + t2 = vec_xxpermdi(t2, t3, 0); + t4 = vec_xxpermdi(t4, t5, 0); + t6 = vec_xxpermdi(t6, t7, 3); + + t0 += t2 + t4 + t6; + + t10 = vec_mergeh(temp04[0], temp05[0]); + t11 = vec_mergeh(temp06[0], temp07[0]); + t12 = vec_mergeo(temp04[1], temp05[1]); + t13 = vec_mergeo(temp06[1], temp07[1]); + t14 = vec_mergel(temp04[2], temp05[2]); + t15 = vec_mergel(temp06[2], temp07[2]); + t16 = vec_mergeo(temp04[3], temp05[3]); + t17 = vec_mergeo(temp06[3], temp07[3]); + t10 = vec_xxpermdi(t10, t11, 0); + t12 = vec_xxpermdi(t12, t13, 0); + t14 = vec_xxpermdi(t14, t15, 0); + t16 = vec_xxpermdi(t16, t17, 3); + + t10 += t12 + t14 + t16; + + vec_f32 inp2[2]; + vec_load_pair(inp2, v_y); + inp2[0] = (a * t0) + (b * inp2[0]); + inp2[1] = (a * t10) + (b * inp2[1]); + vec_store_pair(v_y, inp2); +} +#endif + +#include "sbgemv_t.c" +#else #include "sbgemv_t_vsx.c" +#endif +#endif -//#include "sbgemv_t.c" diff --git a/kernel/power/sbgemv_t_vsx.c b/kernel/power/sbgemv_t_vsx.c index 7da894109..399989bb5 100644 --- a/kernel/power/sbgemv_t_vsx.c +++ b/kernel/power/sbgemv_t_vsx.c @@ -25,12 +25,20 @@ 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. *****************************************************************************/ -#ifndef SBGEMV_T_VSX -#define SBGEMV_T_VSX +#ifndef SBGEMV_T_VSX_C +#define SBGEMV_T_VSX_C #include "sbgemv_common.c" -#define NBMAX 4096 +#ifndef BF16GEMV_T_X +#define BF16GEMV_T_X +#define BF16GEMV_T_8 BF16GEMV_T_VSX_8 +#define BF16GEMV_T_4 BF16GEMV_T_VSX_4 +#define BF16GEMV_T_2 BF16GEMV_T_VSX_2 +#define BF16GEMV_T_1 BF16GEMV_T_VSX_1 +#endif + +#define USE_BFGEMV_8_T_VSX static void BF16GEMV_T_VSX_1(BLASLONG n, BLASLONG lda, IFLOAT *ap, IFLOAT *x, FLOAT *y, FLOAT alpha, FLOAT beta) { @@ -58,9 +66,9 @@ static void BF16GEMV_T_VSX_1(BLASLONG n, BLASLONG lda, IFLOAT *ap, IFLOAT *x, FL temp0 += vec_loadN_mult(&va0[i], inp, n, zero); } else if (n) { - vec_f32 v_inp0 = vec_loadNHi_vec(v_x, i, n, zero); + inp[0] = vec_loadNHi_vec(v_x, i, n, zero); - temp0 += vec_loadNHi_mult(&va0[i], v_inp0, n, zero); + temp0 += vec_loadNHi_mult(&va0[i], inp[0], n, zero); } y[0] = (alpha * (temp0[0] + temp0[1] + temp0[2] + temp0[3])) + (beta * y[0]); @@ -97,10 +105,10 @@ static void BF16GEMV_T_VSX_2(BLASLONG n, BLASLONG lda, IFLOAT *ap, IFLOAT *x, FL temp0 += vec_loadN_mult(&va0[i], inp, n, zero); temp1 += vec_loadN_mult(&va1[i], inp, n, zero); } else if (n) { - vec_f32 v_inp0 = vec_loadNHi_vec(v_x, i, n, zero); + inp[0] = vec_loadNHi_vec(v_x, i, n, zero); - temp0 += vec_loadNHi_mult(&va0[i], v_inp0, n, zero); - temp1 += vec_loadNHi_mult(&va1[i], v_inp0, n, zero); + temp0 += vec_loadNHi_mult(&va0[i], inp[0], n, zero); + temp1 += vec_loadNHi_mult(&va1[i], inp[0], n, zero); } y[0] = (alpha * (temp0[0] + temp0[1] + temp0[2] + temp0[3])) + (beta * y[0]); @@ -148,12 +156,12 @@ static void BF16GEMV_T_VSX_4(BLASLONG n, BLASLONG lda, IFLOAT *ap, IFLOAT *x, FL temp2 += vec_loadN_mult(&va2[i], inp, n, zero); temp3 += vec_loadN_mult(&va3[i], inp, n, zero); } else if (n) { - vec_f32 v_inp0 = vec_loadNHi_vec(v_x, i, n, zero); + inp[0] = vec_loadNHi_vec(v_x, i, n, zero); - temp0 += vec_loadNHi_mult(&va0[i], v_inp0, n, zero); - temp1 += vec_loadNHi_mult(&va1[i], v_inp0, n, zero); - temp2 += vec_loadNHi_mult(&va2[i], v_inp0, n, zero); - temp3 += vec_loadNHi_mult(&va3[i], v_inp0, n, zero); + temp0 += vec_loadNHi_mult(&va0[i], inp[0], n, zero); + temp1 += vec_loadNHi_mult(&va1[i], inp[0], n, zero); + temp2 += vec_loadNHi_mult(&va2[i], inp[0], n, zero); + temp3 += vec_loadNHi_mult(&va3[i], inp[0], n, zero); } vec_f32 t0, t1, t2, t3; @@ -174,6 +182,7 @@ static void BF16GEMV_T_VSX_4(BLASLONG n, BLASLONG lda, IFLOAT *ap, IFLOAT *x, FL v_y[0] = (a * temp0) + (b * v_y[0]); } +#ifdef USE_BFGEMV_8_T_VSX static void BF16GEMV_T_VSX_8(BLASLONG n, BLASLONG lda, IFLOAT *ap, IFLOAT *x, FLOAT *y, FLOAT alpha, FLOAT beta) { IFLOAT *a0, *a1, *a2, *a3, *a4, *a5, *a6, *a7; @@ -235,16 +244,16 @@ static void BF16GEMV_T_VSX_8(BLASLONG n, BLASLONG lda, IFLOAT *ap, IFLOAT *x, FL temp6 += vec_loadN_mult(&va6[i], inp, n, zero); temp7 += vec_loadN_mult(&va7[i], inp, n, zero); } else if (n) { - vec_f32 v_inp0 = vec_loadNHi_vec(v_x, i, n, zero); + inp[0] = vec_loadNHi_vec(v_x, i, n, zero); - temp0 += vec_loadNHi_mult(&va0[i], v_inp0, n, zero); - temp1 += vec_loadNHi_mult(&va1[i], v_inp0, n, zero); - temp2 += vec_loadNHi_mult(&va2[i], v_inp0, n, zero); - temp3 += vec_loadNHi_mult(&va3[i], v_inp0, n, zero); - temp4 += vec_loadNHi_mult(&va4[i], v_inp0, n, zero); - temp5 += vec_loadNHi_mult(&va5[i], v_inp0, n, zero); - temp6 += vec_loadNHi_mult(&va6[i], v_inp0, n, zero); - temp7 += vec_loadNHi_mult(&va7[i], v_inp0, n, zero); + temp0 += vec_loadNHi_mult(&va0[i], inp[0], n, zero); + temp1 += vec_loadNHi_mult(&va1[i], inp[0], n, zero); + temp2 += vec_loadNHi_mult(&va2[i], inp[0], n, zero); + temp3 += vec_loadNHi_mult(&va3[i], inp[0], n, zero); + temp4 += vec_loadNHi_mult(&va4[i], inp[0], n, zero); + temp5 += vec_loadNHi_mult(&va5[i], inp[0], n, zero); + temp6 += vec_loadNHi_mult(&va6[i], inp[0], n, zero); + temp7 += vec_loadNHi_mult(&va7[i], inp[0], n, zero); } vec_f32 t0, t1, t2, t3; @@ -272,14 +281,12 @@ static void BF16GEMV_T_VSX_8(BLASLONG n, BLASLONG lda, IFLOAT *ap, IFLOAT *x, FL temp7 = vec_mergel(t1, t3); temp4 += temp5 + temp6 + temp7; - v_y[0] = (a * temp0) + (b * v_y[0]); - v_y[1] = (a * temp4) + (b * v_y[1]); + vec_load_pair(inp, v_y); + inp[0] = (a * temp0) + (b * inp[0]); + inp[1] = (a * temp4) + (b * inp[1]); + vec_store_pair(v_y, inp); } - -#define BF16GEMV_T_8 BF16GEMV_T_VSX_8 -#define BF16GEMV_T_4 BF16GEMV_T_VSX_4 -#define BF16GEMV_T_2 BF16GEMV_T_VSX_2 -#define BF16GEMV_T_1 BF16GEMV_T_VSX_1 +#endif #include "sbgemv_t.c" #endif