mirror of
https://github.com/amd/blis.git
synced 2026-05-04 06:21:12 +00:00
fbb72d047f
Description: 1. Added group quantization and zero-point (zp) in aocl_gemm_bf16s4f32o<bf16|f32> API. 2. Group quantization is technique to improve accuracy where scale factors to dequantize weights varies at group level instead of per channel and per tensor level. 3. Added zp and scaling in woq packb kernels so that for large M values zp and scaling are performed at pack-b stage and bf16 kernels are called 4. Adding zp support and scaling to default path in WoQ kernels created some performance overhead when M value is very small. 5. Added string group_size to lpgemm bench to read group size from bench_input.txt and tested for various combinations of matrix dimensions. 6. The scalefactors could be of type float or bf16 and the zeropoint values are expected to be in int8 format. AMD-Internal: [SWLCSG-3168, SWLCSG-3172] Change-Id: Iff07b54d76edc7408eb2ea0b29ce8b4a04a38f57
157 lines
4.3 KiB
C
157 lines
4.3 KiB
C
/*
|
|
|
|
BLIS
|
|
An object-based framework for developing high-performance BLAS-like
|
|
libraries.
|
|
|
|
Copyright (C) 2022 - 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 LPGEMM_POST_OPS_H
|
|
#define LPGEMM_POST_OPS_H
|
|
|
|
typedef enum
|
|
{
|
|
POST_OPS_DISABLE = 0,
|
|
POST_OPS_BIAS = 1,
|
|
POST_OPS_RELU = 2,
|
|
POST_OPS_RELU_SCALE = 3,
|
|
POST_OPS_GELU_TANH = 4,
|
|
POST_OPS_GELU_ERF = 5,
|
|
POST_OPS_CLIP = 6,
|
|
POST_OPS_DOWNSCALE = 7,
|
|
POST_OPS_MATRIX_ADD = 8,
|
|
POST_OPS_SWISH = 9,
|
|
POST_OPS_MATRIX_MUL = 10,
|
|
POST_OPS_TANH = 11,
|
|
POST_OPS_SIGMOID = 12,
|
|
POST_OPS_SUM = 13
|
|
|
|
|
|
|
|
} LPGEMM_POST_OP_CODE;
|
|
|
|
// Used as an internal structure.
|
|
typedef struct lpgemm_post_op_t
|
|
{
|
|
uint64_t op_code;
|
|
void* op_args1; // zero_point, bias, sum_buff
|
|
void* op_args2; // alpha, storage order, sum_zero_point
|
|
void* op_args3; // beta, zero_point_len
|
|
void* scale_factor;
|
|
dim_t scale_factor_len;
|
|
bool is_power_of_2;
|
|
uint64_t bias_stor_type;
|
|
struct lpgemm_post_op_t* next;
|
|
} lpgemm_post_op;
|
|
|
|
// Used as an internal structure.
|
|
typedef struct lpgemm_pre_op_t
|
|
{
|
|
uint64_t op_code;
|
|
uint64_t group_size;
|
|
void *scale_factor;
|
|
uint64_t scale_factor_len;
|
|
uint64_t scale_factor_type;
|
|
void *zp;
|
|
uint64_t zp_len;
|
|
struct lpgemm_pre_op_t *next;
|
|
} lpgemm_pre_op;
|
|
|
|
// Used as an internal structure.
|
|
typedef struct lpgemm_post_op_attr_t
|
|
{
|
|
uint64_t post_op_c_i;
|
|
uint64_t post_op_c_j;
|
|
uint64_t rs_c_downscale;
|
|
uint64_t cs_c_downscale;
|
|
void* buf_downscale;
|
|
uint64_t is_first_k;
|
|
uint64_t is_last_k;
|
|
uint64_t c_stor_type;
|
|
uint64_t bias_stor_type;
|
|
uint64_t b_sum_offset;
|
|
int32_t* b_col_sum_vec;
|
|
int16_t* b_col_sum_vec_s16;
|
|
} lpgemm_post_op_attr;
|
|
|
|
typedef struct lpgemm_pre_op_attr_t
|
|
{
|
|
void* scale_factor;
|
|
uint64_t scale_factor_len;
|
|
uint64_t scale_factor_type;
|
|
void* zero_point;
|
|
uint64_t zero_point_len;
|
|
uint64_t pre_op_b_i;
|
|
uint64_t pre_op_b_j;
|
|
uint64_t group_size;
|
|
uint64_t pre_op_ld;
|
|
} lpgemm_pre_op_attr;
|
|
|
|
err_t lpgemm_translate_to_post_ops_list
|
|
(
|
|
aocl_post_op* post_op_unparsed,
|
|
lpgemm_post_op* post_op_list,
|
|
void* scale_buffer,
|
|
void* meta_arg,
|
|
dim_t m,
|
|
dim_t n
|
|
);
|
|
|
|
err_t lpgemm_translate_to_pre_ops_list
|
|
(
|
|
aocl_pre_op *pre_op_unparsed,
|
|
lpgemm_pre_op *pre_op_list,
|
|
dim_t m,
|
|
dim_t n,
|
|
dim_t k
|
|
);
|
|
|
|
#define POST_OP_LABEL_LASTK_SAFE_JUMP \
|
|
if ( ( post_ops_attr.is_last_k == TRUE ) && ( post_ops_list_temp != NULL ) ) \
|
|
{ \
|
|
goto *post_ops_labels[post_ops_list_temp->op_code]; \
|
|
} \
|
|
else \
|
|
{ \
|
|
goto *post_ops_labels[0]; \
|
|
}
|
|
|
|
#define POST_OP_LABEL_LASTK_SAFE_JUMP_WITH_NEXT_PTR \
|
|
post_ops_list_temp = post_ops_list_temp->next; \
|
|
if ( post_ops_list_temp != NULL ) \
|
|
{ \
|
|
goto *post_ops_labels[post_ops_list_temp->op_code]; \
|
|
} \
|
|
else \
|
|
{ \
|
|
goto *post_ops_labels[0]; \
|
|
}
|
|
|
|
#endif //LPGEMM_POST_OPS_H
|