Skip to content

Commit cb4e4de

Browse files
dcammarata01dcammarata01
committed
[SW] Restructured sw/quadrilateroBenchmarks repo
1 parent 0dd51fa commit cb4e4de

25 files changed

Lines changed: 631 additions & 1379 deletions

File tree

‎sw/quadrilateroBenchmarks/CMakeLists.txt‎

Lines changed: 12 additions & 16 deletions
Original file line numberDiff line numberDiff line change
@@ -38,33 +38,29 @@ macro(add_quadrilatero_test_threeParam name file param1 param2 param3)
3838
set(target_name ${name}_M${param1}_N${param2}_K${param3})
3939
add_snitch_test(${target_name} ${file})
4040
target_link_libraries(test-${SNITCH_TEST_PREFIX}${target_name} benchmark-quadrilatero ${SNITCH_RUNTIME})
41+
target_link_libraries(test-${SNITCH_TEST_PREFIX}${target_name} matmul ${SNITCH_RUNTIME})
4142
target_compile_definitions(test-${SNITCH_TEST_PREFIX}${target_name} PUBLIC DATAHEADER="data/data_${param1}_${param2}_${param3}.h" SNRT_NFPU_PER_CORE=16 QUAD_RLEN=${QUAD_RLEN})
4243
endmacro()
4344

4445
# Benchmark library
45-
add_library(benchmark-quadrilatero benchmark-quadrilatero/benchmark.c)
46+
add_library(benchmark-quadrilatero kernels/benchmark.c)
47+
add_library(matmul kernels/matmul.c)
4648

4749
# Kernels
48-
# add_library(sp-matmul-quadrilatero sp-matmul-quadrilatero/kernel/sp-matmul-quadrilatero.c)
49-
# add_library(widening2x-hp-matmul-quadrilatero widening2x-hp-matmul-quadrilatero/kernel/widening2x-hp-matmul-quadrilatero.c)
50-
# add_library(widening4x-bp-matmul-quadrilatero widening4x-bp-matmul-quadrilatero/kernel/widening4x-bp-matmul-quadrilatero.c)
51-
add_library(sp-fmatmul-quadrilatero sp-fmatmul-quadrilatero/kernel/sp-fmatmul-quadrilatero.c)
52-
# add_library(widening2x-hp-fmatmul-quadrilatero widening2x-hp-fmatmul-quadrilatero/kernel/widening2x-hp-fmatmul-quadrilatero.c)
53-
# add_library(widening4x-bp-fmatmul-quadrilatero widening4x-bp-fmatmul-quadrilatero/kernel/widening4x-bp-fmatmul-quadrilatero.c)
54-
add_library(verify-quadrilatero verify-quadrilatero/kernel/verify-quadrilatero.c)
50+
# add_library(verify-quadrilatero verify-quadrilatero/kernel/verify-quadrilatero.c)
5551

5652
# # Tests
5753
enable_testing()
5854
set(SNITCH_TEST_PREFIX quadrilateroBenchmarks-)
5955
# add_quadrilatero_test_noParam(quadrilatero-matmul quadrilatero_matmul/main.c)
6056

61-
# add_quadrilatero_test_threeParam(sp-matmul-quadrilatero sp-matmul-quadrilatero/main.c 64 64 64)
62-
# add_quadrilatero_test_threeParam(widening2x-hp-matmul-quadrilatero widening2x-hp-matmul-quadrilatero/main.c 64 64 64)
63-
# add_quadrilatero_test_threeParam(widening4x-bp-matmul-quadrilatero widening4x-bp-matmul-quadrilatero/main.c 64 64 64)
64-
# add_quadrilatero_test_threeParam(sp-fmatmul-quadrilatero sp-fmatmul-quadrilatero/main.c 64 64 64)
57+
add_quadrilatero_test_threeParam(sp-matmul-quadrilatero sp-matmul-quadrilatero/main.c 64 64 64)
58+
add_quadrilatero_test_threeParam(widening2x-hp-matmul-quadrilatero widening2x-hp-matmul-quadrilatero/main.c 64 64 64)
59+
add_quadrilatero_test_threeParam(widening4x-bp-matmul-quadrilatero widening4x-bp-matmul-quadrilatero/main.c 64 64 64)
60+
add_quadrilatero_test_threeParam(sp-fmatmul-quadrilatero sp-fmatmul-quadrilatero/main.c 64 64 64)
61+
add_quadrilatero_test_threeParam(widening2x-hp-fmatmul-quadrilatero widening2x-hp-fmatmul-quadrilatero/main.c 64 64 64)
62+
add_quadrilatero_test_threeParam(widening4x-bp-fmatmul-quadrilatero widening4x-bp-fmatmul-quadrilatero/main.c 64 64 64)
6563
# add_quadrilatero_test_threeParam(sp-fmatmul-quadrilatero sp-fmatmul-quadrilatero/main.c 64 64 8)
66-
add_quadrilatero_test_threeParam(sp-fmatmul-quadrilatero sp-fmatmul-quadrilatero/main.c 8 8 8)
64+
# add_quadrilatero_test_threeParam(sp-fmatmul-quadrilatero sp-fmatmul-quadrilatero/main.c 8 8 8)
6765
# add_quadrilatero_test_threeParam(sp-fmatmul-quadrilatero sp-fmatmul-quadrilatero/main.c 8 16 72)
68-
# add_quadrilatero_test_threeParam(widening2x-hp-fmatmul-quadrilatero widening2x-hp-fmatmul-quadrilatero/main.c 64 64 64)
69-
add_quadrilatero_test_threeParam(verify-quadrilatero verify-quadrilatero/main.c 64 64 64)
70-
# add_quadrilatero_test_threeParam(widening4x-bp-fmatmul-quadrilatero widening4x-bp-fmatmul-quadrilatero/main.c 64 64 64)
66+
# add_quadrilatero_test_threeParam(verify-quadrilatero verify-quadrilatero/main.c 64 64 64)

sw/quadrilateroBenchmarks/sp-fmatmul-quadrilatero/kernel/sp-fmatmul-quadrilatero.c renamed to sw/quadrilateroBenchmarks/include/matmul.h

Lines changed: 35 additions & 48 deletions
Original file line numberDiff line numberDiff line change
@@ -1,20 +1,9 @@
1-
// Copyright 2024 EPFL
2-
// Solderpad Hardware License, Version 2.1, see LICENSE.md for details.
3-
// SPDX-License-Identifier: Apache-2.0 WITH SHL-2.1
4-
//
5-
// Author: Danilo Cammarata
6-
7-
// Output tile size: 8x8
8-
#define FP32 6
9-
#define CMUL 2
10-
#define RMUL 2
11-
12-
void __attribute__ ((noinline)) matrixMul_8x8(float* addrA, float* addrB, float* addrC, int K, int N, int M, int shift)
13-
{
14-
int strideA = M * 4;
15-
int strideB = N * 4;
16-
int strideC = N * 4;
171

2+
void __attribute__ ((noinline)) FUNC_NAME(void* addrA,void* addrB, void* addrC, int K, int N, int M, int shift)
3+
{
4+
int strideA;
5+
int strideB;
6+
int strideC;
187
asm volatile(
198
// --- PROLOGUE: Save registers ---
209
"addi sp, sp, -0x30 \n\t"
@@ -32,45 +21,45 @@ void __attribute__ ((noinline)) matrixMul_8x8(float* addrA, float* addrB, float*
3221
"sw s11, 0x00(sp) \n\t"
3322

3423
// 1. Data Types Configuration
35-
"mmac.dt %[dt_type], %[dt_type], %[dt_type] \n\t"
24+
"mmac.dt %[dt_typeC], %[dt_typeA], %[dt_typeB] \n\t"
3625

3726
"add t0, x0, %[M] \n\t" // t0 = Remaining M
3827
"add s0, x0, %[addrC] \n\t" // s0 = Current row pointer for C
3928
"add s1, x0, %[addrA] \n\t" // s1 = Current row pointer for A (LHS)
4029

41-
"loopM_start_dyn: \n\t"
42-
"mcfgm t3, t0, 1 \n\t" // t3 = Processed M rows
30+
"sll %[strideA], %[M], %[shift] \n\t" // Compute strideA = M * 2^shift
31+
"sll %[strideB], %[N], %[shift] \n\t" // Compute strideB = M * 2^shift
32+
"sll %[strideC], %[N], %[shift] \n\t" // Compute strideC = M * 2^shift
33+
34+
"1: \n\t"
35+
"mcfgm t3, t0, %[rmul] \n\t" // t3 = Processed M rows
4336

4437
"add t1, x0, %[N] \n\t" // t1 = Remaining N
4538
"add s2, x0, s0 \n\t" // s2 = C tile pointer
46-
"add s3, x0, %[addrB] \n\t" // s3 = B (RHS) tile pointer
47-
48-
"loopN_start_dyn: \n\t"
49-
"mcfgn t4, t1, 1 \n\t" // t4 = Processed N columns
50-
51-
// Reset accumulator
52-
"mzero.a acc0 \n\t"
39+
"add s3, x0, %[addrB] \n\t" // s3 = B (RHS) tile pointer
5340

41+
"2: \n\t"
5442
"add t2, x0, %[K] \n\t" // t2 = Remaining K
43+
"mcfgk t5, t2 \n\t" // t5 = Processed K depth for a single block
44+
"mcfgn t4, t1, %[cmul] \n\t" // t4 = Processed N columns
45+
"mzero.a acc0 \n\t" // Reset accumulator
5546
"add s4, x0, s1 \n\t" // s4 = A tile pointer
5647
"add s5, x0, s3 \n\t" // s5 = B tile pointer
5748

58-
"loopK_start_dyn: \n\t"
59-
"mcfgk t5, t2 \n\t" // t5 = Processed K depth for a single block
49+
"3: \n\t"
6050

6151
// 2. Load First Tiles
6252
"mld.lhs m0, (s4), %[strideA] \n\t"
63-
"mld.rhs m4, (s5), %[strideB] \n\t"
64-
"mmacc acc0, m4, m0 \n\t"
65-
66-
"mul s8, t5, %[strideA] \n\t"
53+
"mld.rhs m4, (s5), %[strideB] \n\t"
54+
"mmacc acc0, m4, m0 \n\t"
55+
"mul s8, t5, %[strideA] \n\t"
56+
"mul s9, t5, %[strideB] \n\t"
57+
6758
"add s6, s4, s8 \n\t" // s6 = Pointer to 2nd tile of A
6859
"mld.lhs m2, (s6), %[strideA] \n\t"
69-
70-
"mul s9, t5, %[strideB] \n\t"
60+
7161
"add s7, s5, s9 \n\t" // s7 = Pointer to 2nd tile of B
72-
"mld.rhs m6, (s7), %[strideB] \n\t"
73-
"mmacc acc0, m6, m2 \n\t"
62+
"mld.rhs m6, (s7), %[strideB] \n\t"
7463

7564
// 5. Advance K pointers by TWO blocks
7665
"add t6, s8, s8 \n\t"
@@ -81,11 +70,13 @@ void __attribute__ ((noinline)) matrixMul_8x8(float* addrA, float* addrB, float*
8170

8271
// Decrease remaining K by two blocks
8372
"add t6, t5, t5 \n\t"
84-
"sub t2, t2, t6 \n\t"
85-
"bgtz t2, loopK_start_dyn \n\t"
73+
"sub t2, t2, t6 \n\t"
74+
"mmacc acc0, m6, m2 \n\t"
75+
"mcfgk t5, t2 \n\t" // t5 = Processed K depth for a single block
76+
"bgtz t2, 3b \n\t"
8677

8778
// 6. Transfer to MR and Store
88-
"mmov.am m8, acc0 \n\t"
79+
"mmov.am m8, acc0 \n\t"
8980
"mst m8, (s2), %[strideC] \n\t"
9081

9182
// 7. Advance along N
@@ -94,7 +85,7 @@ void __attribute__ ((noinline)) matrixMul_8x8(float* addrA, float* addrB, float*
9485
"add s3, s3, t6 \n\t"
9586

9687
"sub t1, t1, t4 \n\t"
97-
"bgtz t1, loopN_start_dyn \n\t"
88+
"bgtz t1, 2b \n\t"
9889

9990
// 8. Advance along M
10091
"mul t6, t3, %[strideC] \n\t"
@@ -104,7 +95,7 @@ void __attribute__ ((noinline)) matrixMul_8x8(float* addrA, float* addrB, float*
10495
"add s1, s1, t6 \n\t"
10596

10697
"sub t0, t0, t3 \n\t"
107-
"bgtz t0, loopM_start_dyn \n\t"
98+
"bgtz t0, 1b \n\t"
10899

109100
// --- EPILOGUE: Restore registers ---
110101
"lw s0 , 0x2c(sp) \n\t"
@@ -124,14 +115,10 @@ void __attribute__ ((noinline)) matrixMul_8x8(float* addrA, float* addrB, float*
124115

125116
:
126117
: [addrA] "r" (addrA), [addrB] "r" (addrB), [addrC] "r" (addrC),
127-
[M] "r" (M), [N] "r" (N), [K] "r" (K),
118+
[M] "r" (M), [N] "r" (N), [K] "r" (K), [shift] "r" (shift),
128119
[strideA] "r" (strideA), [strideB] "r" (strideB), [strideC] "r" (strideC),
129-
[dt_type] "i" (FP32)
120+
[dt_typeC] "i" (DTC), [dt_typeA] "i" (DTA), [dt_typeB] "i" (DTB),
121+
[rmul] "i" (RMUL_2), [cmul] "i" (CMUL_2)
130122
: "t0", "t1", "t2", "t3", "t4", "t5", "t6", "s0", "s1", "s2", "s3", "s4", "s5", "s6", "s7", "s8", "s9", "memory"
131123
);
132-
}
133-
134-
void __attribute__ ((noinline)) matrixMul_16x16(float* addrA, float* addrB, float* addrC, int K, int N, int M, int shift)
135-
{
136-
137124
}

‎sw/quadrilateroBenchmarks/include/quadrilatero.h‎

Lines changed: 29 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,29 @@
1+
#define __INT8__ 0
2+
#define __INT16__ 1
3+
#define __INT32__ 2
4+
#define __FP8__ 4
5+
#define __FP16__ 5
6+
#define __FP32__ 6
7+
#define __FP8ALT__ 12
8+
#define __FP16ALT__ 13
9+
#define __FP32ALT__ 14
10+
#define CMUL_1 0
11+
#define RMUL_1 0
12+
#define CMUL_2 1
13+
#define RMUL_2 1
14+
#define CMUL_4 3
15+
#define RMUL_4 3
16+
17+
#define MATMUL_NAME(DTA,DTB,DTC) \
18+
matmul_##DTA##_##DTB##_##DTC
19+
20+
#define matmul(DTA,DTB,DTC,A,B,C,K,N,M,shift) \
21+
MATMUL_NAME(DTA,DTB,DTC)(A,B,C,K,N,M,shift)
22+
23+
24+
void __attribute__ ((noinline)) matmul_INT8_INT8_INT32(void* addrA,void* addrB, void* addrC, int K, int N, int M, int shift);
25+
void __attribute__ ((noinline)) matmul_INT16_INT16_INT32(void* addrA,void* addrB, void* addrC, int K, int N, int M, int shift);
26+
void __attribute__ ((noinline)) matmul_INT32_INT32_INT32(void* addrA,void* addrB, void* addrC, int K, int N, int M, int shift);
27+
void __attribute__ ((noinline)) matmul_FP8_FP8_FP32(void* addrA,void* addrB, void* addrC, int K, int N, int M, int shift);
28+
void __attribute__ ((noinline)) matmul_FP16_FP16_FP32(void* addrA,void* addrB, void* addrC, int K, int N, int M, int shift);
29+
void __attribute__ ((noinline)) matmul_FP32_FP32_FP32(void* addrA,void* addrB, void* addrC, int K, int N, int M, int shift);

sw/quadrilateroBenchmarks/benchmark-quadrilatero/benchmark.c renamed to sw/quadrilateroBenchmarks/kernels/benchmark.c

File renamed without changes.

‎sw/quadrilateroBenchmarks/kernels/matmul.c‎

Lines changed: 67 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,67 @@
1+
#include <quadrilatero.h>
2+
3+
// INT8 x INT8 -> INT32
4+
#define DTA __INT8__
5+
#define DTB __INT8__
6+
#define DTC __INT32__
7+
#define FUNC_NAME MATMUL_NAME(INT8,INT8,INT32)
8+
#include <matmul.h>
9+
#undef FUNC_NAME
10+
#undef DTA
11+
#undef DTB
12+
#undef DTC
13+
14+
// INT16 x INT16 -> INT32
15+
#define DTA __INT16__
16+
#define DTB __INT16__
17+
#define DTC __INT32__
18+
#define FUNC_NAME MATMUL_NAME(INT16,INT16,INT32)
19+
#include <matmul.h>
20+
#undef FUNC_NAME
21+
#undef DTA
22+
#undef DTB
23+
#undef DTC
24+
25+
// INT32 x INT32 -> INT32
26+
#define DTA __INT32__
27+
#define DTB __INT32__
28+
#define DTC __INT32__
29+
#define FUNC_NAME MATMUL_NAME(INT32,INT32,INT32)
30+
#include <matmul.h>
31+
#undef FUNC_NAME
32+
#undef DTA
33+
#undef DTB
34+
#undef DTC
35+
36+
// FP8 x FP8 -> FP32
37+
#define DTA __FP8__
38+
#define DTB __FP8__
39+
#define DTC __FP32__
40+
#define FUNC_NAME MATMUL_NAME(FP8,FP8,FP32)
41+
#include <matmul.h>
42+
#undef FUNC_NAME
43+
#undef DTA
44+
#undef DTB
45+
#undef DTC
46+
47+
// FP16 x FP16 -> FP32
48+
#define DTA __FP16__
49+
#define DTB __FP16__
50+
#define DTC __FP32__
51+
#define FUNC_NAME MATMUL_NAME(FP16,FP16,FP32)
52+
#include <matmul.h>
53+
#undef FUNC_NAME
54+
#undef DTA
55+
#undef DTB
56+
#undef DTC
57+
58+
// FP32 x FP32 -> FP32
59+
#define DTA __FP32__
60+
#define DTB __FP32__
61+
#define DTC __FP32__
62+
#define FUNC_NAME MATMUL_NAME(FP32,FP32,FP32)
63+
#include <matmul.h>
64+
#undef FUNC_NAME
65+
#undef DTA
66+
#undef DTB
67+
#undef DTC

0 commit comments

Comments
 (0)