mirror of
https://github.com/amd/blis.git
synced 2026-05-04 06:21:12 +00:00
c1e063e65c
Details: - For a variable x, Using address of x in an instruction throws exception if the difference between &x and access position is larger than 2 GiB. To solve this issue all variables are stored within the JIT code section and are accessed using relative addressing. - Fixed a bug in B matrix pack function for s8s8s32os32 API. - Fixed a bug in JIT code to apply bias on col-major matrices. AMD-Internal: [SWLCSG-2820] Change-Id: I82f117a0422c794cb9b1a4d65a89d60de4adfd96
199 lines
7.5 KiB
C++
199 lines
7.5 KiB
C++
/*
|
|
|
|
BLIS
|
|
An object-based framework for developing high-performance BLAS-like
|
|
libraries.
|
|
|
|
Copyright (C) 2024, Advanced Micro Devices, Inc. All rights reserved.
|
|
|
|
Redistribution and use in source and binary forms, with or without
|
|
modification, are permitted provided that the following conditions are
|
|
met:
|
|
- Redistributions of source code must retain the above copyright
|
|
notice, this list of conditions and the following disclaimer.
|
|
- Redistributions in binary form must reproduce the above copyright
|
|
notice, this list of conditions and the following disclaimer in the
|
|
documentation and/or other materials provided with the distribution.
|
|
- Neither the name(s) of the copyright holder(s) nor the names of its
|
|
contributors may be used to endorse or promote products derived
|
|
from this software without specific prior written permission.
|
|
|
|
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
|
|
"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
|
|
LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
|
|
A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
|
|
HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
|
|
SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
|
|
LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
|
|
DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
|
|
THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
|
(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
|
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
|
|
*/
|
|
|
|
#ifndef JIT_BF16_H
|
|
#define JIT_BF16_H
|
|
|
|
#include <stdio.h>
|
|
#include <stdlib.h>
|
|
#include <time.h>
|
|
#include <cstring>
|
|
#include <functional>
|
|
#include "blis.h"
|
|
#include <xbyak/xbyak.h>
|
|
|
|
using namespace Xbyak;
|
|
|
|
class bli_lpgemm_jit: public Xbyak::CodeGenerator
|
|
{
|
|
|
|
private :
|
|
void preamble();
|
|
void postamble();
|
|
void initialize_params( lpgemm_jit_inputs_t* params );
|
|
void reg_init(dim_t m_dim, dim_t n_dim );
|
|
void kernel_unroll( dim_t m_dim, dim_t n_dim );
|
|
void prefetchC( dim_t m_dim, dim_t n_dim );
|
|
void k_fringe_loop( dim_t m_dim, dim_t n_dim );
|
|
void scale_alpha( dim_t m_dim, dim_t n_dim );
|
|
// beta ops
|
|
void bf16_f32_beta_op( dim_t m_dim, dim_t n_dim );
|
|
void f32_f32_beta_op( dim_t m_dim, dim_t n_dim );
|
|
//postops
|
|
void clip_f32( dim_t m_dim, dim_t n_dim );
|
|
void f32_f32_matrix_add( dim_t m_dim, dim_t n_dim );
|
|
void bf16_f32_matrix_add( dim_t m_dim, dim_t n_dim );
|
|
void bias_row_major( dim_t m_dim, dim_t n_dim );
|
|
void bias_col_major( dim_t m_dim, dim_t n_dim );
|
|
void relu( dim_t m_dim, dim_t n_dim );
|
|
void relu_scale( dim_t m_dim, dim_t n_dim );
|
|
void gelu_tanh( dim_t m_dim, dim_t n_dim );
|
|
void POLY_EVAL_6_AVX512();
|
|
void EXPF_AVX512();
|
|
void TANHF_AVX512();
|
|
void GELU_TANH_F32_AVX512_DEF( dim_t reg );
|
|
void POLY_EVAL_HORNER_16_0_AVX512();
|
|
void ERF_AVX512();
|
|
void GELU_ERF_F32_AVX512_DEF( dim_t reg );
|
|
void gelu_erf( dim_t m_dim, dim_t n_dim );
|
|
void SWISH_F32_AVX512_DEF( dim_t reg );
|
|
void swish( dim_t m, dim_t n );
|
|
|
|
void apply_post_ops_in_high_reg_pressure
|
|
(
|
|
const dim_t num_post_op_regs,
|
|
std::function< void( dim_t ) > op_fn
|
|
);
|
|
// C store functions
|
|
void cvt_store_f32_bf16_mask( dim_t m_dim, dim_t n_dim );
|
|
void store_f32( dim_t m_dim, dim_t n_dim );
|
|
|
|
void post_op_label_lastk_safe_jump_with_next_ptr();
|
|
void post_op_label_lastk_safe_jump();
|
|
|
|
|
|
dim_t num_elems_per_reg = 64 / sizeof(float);
|
|
dim_t n_rem;
|
|
dim_t num_fma_regs;
|
|
dim_t fma_start_idx = 0;
|
|
dim_t load_start_idx = 0;
|
|
dim_t num_full_loads;
|
|
dim_t num_loads;
|
|
dim_t bcst_start_idx;
|
|
dim_t alpha_reg = fma_start_idx;
|
|
dim_t beta_reg;
|
|
|
|
// registers used for gelu_tanh
|
|
const dim_t num_gelu_regs = 9;
|
|
const dim_t const1 = load_start_idx;
|
|
const dim_t const2 = load_start_idx+1;
|
|
const dim_t x = load_start_idx+2;
|
|
const dim_t r = load_start_idx+3;
|
|
const dim_t r2 = load_start_idx+4;
|
|
const dim_t z = load_start_idx+5;
|
|
const dim_t dn = load_start_idx+6;
|
|
const dim_t x_tanh = load_start_idx+7;
|
|
const dim_t q = load_start_idx+8;
|
|
|
|
// registers for gelu_erf
|
|
const dim_t num_erf_regs = 5;
|
|
const dim_t x_erf = load_start_idx+4;
|
|
|
|
// registers used for swish. Reusing the gelu_tanh registers.
|
|
const dim_t num_swish_regs = 9;
|
|
|
|
const dim_t stack_off_ps_a = 8;
|
|
const dim_t stack_off_k_iter_before_prefetch = 16;
|
|
const dim_t stack_off_k_iter_after_prefetch = 24;
|
|
const dim_t stack_off_k_left = 32;
|
|
const dim_t stack_off_alpha = 40;
|
|
const dim_t stack_off_beta = 48;
|
|
const dim_t stack_off_b_ptr = 56;
|
|
const dim_t stack_off_postop = 64;
|
|
const dim_t stack_off_buf_downscale = stack_off_postop +
|
|
offsetof( lpgemm_post_op_attr,
|
|
buf_downscale );
|
|
const dim_t stack_off_temp_list = stack_off_postop +
|
|
sizeof( lpgemm_post_op );
|
|
|
|
|
|
const dim_t stack_off_zmm_stack = stack_off_temp_list + 8;
|
|
dim_t zmm_stack_top;
|
|
|
|
void store_zmms_in_stack( dim_t reg_start_idx,
|
|
dim_t num_regs,
|
|
dim_t stack_off
|
|
);
|
|
|
|
void get_zmms_from_stack( dim_t reg_start_idx,
|
|
dim_t num_regs,
|
|
dim_t stack_off
|
|
);
|
|
|
|
float gelu_consts[7] = { 0.044715, 0.797884, -2, 0.5, -1, 2, 1 };
|
|
float gelu_macros[6] = { 1.4426950408889634, 1.2582912E7,
|
|
-88.0f, 88.0f,
|
|
(float)(1.0/0.0), -2147483648 };
|
|
|
|
float lpgemm_exp[6] = { 1.0000000754895704, 0.6931472254087585,
|
|
0.2402210737432219, 0.05550297297702539,
|
|
0.009676036358193323, 0.001341000536524434 };
|
|
|
|
float erf_consts[4] = { 0.707107, 1.0, 0.5, 3.553f };
|
|
|
|
float lpgemm_erf[16] = { 1.1283793786592402, 2.5468861568875563E-5,
|
|
0.3756169877289898, 0.004025179163741976,
|
|
0.12947984300439994, 0.0412525204794885,
|
|
0.03918550001070417, 0.07104542913277255,
|
|
0.05717052146749476, 0.025310822854733135,
|
|
0.0067305713376882076, 0.0010410692067591445,
|
|
6.921588102382636E-5, 4.092409485758739E-6,
|
|
1.033131746125426E-6, 5.2927177513236435E-8 };
|
|
|
|
|
|
const dim_t gelu_consts_off = 0;
|
|
const dim_t gelu_macros_off = gelu_consts_off + sizeof(gelu_consts);
|
|
const dim_t lpgemm_exp_off = gelu_macros_off + sizeof(gelu_macros);
|
|
const dim_t erf_consts_off = lpgemm_exp_off + sizeof(lpgemm_exp);
|
|
const dim_t lpgemm_erf_off = erf_consts_off + sizeof(erf_consts);
|
|
|
|
Xbyak::Address get_constant( dim_t table_off, dim_t value_off )
|
|
{
|
|
return ptr[rip + tables + table_off + value_off * 4 ];
|
|
}
|
|
Xbyak::Label tables;
|
|
|
|
public:
|
|
bli_lpgemm_jit( void* buffer, size_t bufferSize );
|
|
void generate_kernel( lpgemm_jit_inputs_t* params );
|
|
const void (*get_function ()const)( lpgemm_jit_params_t*,
|
|
lpgemm_post_op_attr*,
|
|
lpgemm_post_op*
|
|
);
|
|
const void *get_code ()const;
|
|
dim_t get_size ();
|
|
|
|
};
|
|
#endif
|