Vortex 2.0 changes:
+ Microarchitecture optimizations + 64-bit support + Xilinx FPGA support + LLVM-16 support + Refactoring and quality control fixes minor update minor update minor update minor update minor update minor update cleanup cleanup cache bindings and memory perf refactory minor update minor update hw unit tests fixes minor update minor update minor update minor update minor update minor udpate minor update minor update minor update minor update minor update minor update minor update minor updates minor updates minor update minor update minor update minor update minor update minor update minor updates minor updates minor updates minor updates minor update minor update
This commit is contained in:
Blaise Tine
@@ -13,8 +13,8 @@ ARG LLVM_HOME=/opt/llvm-10
|
||||
ARG POCL_CC_PATH=/opt/pocl_cc
|
||||
ARG POCL_RT_PATH=/opt/pocl_rt
|
||||
ARG VORTEX_HOME=/home/vortex
|
||||
ARG VORTEX_DRIVER_INC=$VORTEX_HOME/driver/sw/include
|
||||
ARG VORTEX_DRIVER_LIB=$VORTEX_HOME/driver/sw/stub/libvortex.so
|
||||
ARG VORTEX_RUNTIME_INC=$VORTEX_HOME/runtime/sw/include
|
||||
ARG VORTEX_RUNTIME_LIB=$VORTEX_HOME/runtime/sw/stub/libvortex.so
|
||||
ARG VORTEX_RUNTIME_PATH=$VORTEX_HOME/runtime
|
||||
|
||||
# system update
|
||||
@@ -65,13 +65,13 @@ RUN git clone https://$LOGIN@github.gatech.edu/casl/pocl.git /tmp/pocl
|
||||
RUN cd /tmp/pocl; \
|
||||
mkdir build_cc
|
||||
RUN cd /tmp/pocl/build_cc; \
|
||||
cmake -G Ninja -DCMAKE_INSTALL_PREFIX=$POCL_CC_PATH -DOCS_AVAILABLE=ON -DWITH_LLVM_CONFIG=$LLVM_HOME/bin/llvm-config -DENABLE_VORTEX=ON -DVORTEX_RUNTIME_PATH=$VORTEX_RUNTIME_PATH -DVORTEX_DRIVER_INC=$VORTEX_DRIVER_INC -DVORTEX_DRIVER_LIB=$VORTEX_DRIVER_LIB -DBUILD_TESTS=OFF -DPOCL_DEBUG_MESSAGES=ON ..
|
||||
cmake -G Ninja -DCMAKE_INSTALL_PREFIX=$POCL_CC_PATH -DOCS_AVAILABLE=ON -DWITH_LLVM_CONFIG=$LLVM_HOME/bin/llvm-config -DENABLE_VORTEX=ON -DVORTEX_RUNTIME_PATH=$VORTEX_RUNTIME_PATH -DVORTEX_RUNTIME_INC=$VORTEX_RUNTIME_INC -DVORTEX_RUNTIME_LIB=$VORTEX_RUNTIME_LIB -DBUILD_TESTS=OFF -DPOCL_DEBUG_MESSAGES=ON ..
|
||||
RUN cd /tmp/pocl/build_cc; \
|
||||
cmake --build . --target install
|
||||
RUN cd /tmp/pocl; \
|
||||
mkdir build_rt
|
||||
RUN cd /tmp/pocl/build_rt; \
|
||||
cmake -G Ninja -DCMAKE_INSTALL_PREFIX=$POCL_RT_PATH -DOCS_AVAILABLE=OFF -DHOST_DEVICE_BUILD_HASH=riscv32-unknown-unknown-elf -DENABLE_VORTEX=ON -DVORTEX_RUNTIME_PATH=$VORTEX_RUNTIME_PATH -DVORTEX_DRIVER_INC=$VORTEX_DRIVER_INC -DVORTEX_DRIVER_LIB=$VORTEX_DRIVER_LIB -DBUILD_TESTS=OFF -DPOCL_DEBUG_MESSAGES=ON ..
|
||||
cmake -G Ninja -DCMAKE_INSTALL_PREFIX=$POCL_RT_PATH -DOCS_AVAILABLE=OFF -DHOST_DEVICE_BUILD_HASH=riscv32-unknown-unknown-elf -DENABLE_VORTEX=ON -DVORTEX_RUNTIME_PATH=$VORTEX_RUNTIME_PATH -DVORTEX_RUNTIME_INC=$VORTEX_RUNTIME_INC -DVORTEX_RUNTIME_LIB=$VORTEX_RUNTIME_LIB -DBUILD_TESTS=OFF -DPOCL_DEBUG_MESSAGES=ON ..
|
||||
RUN cd /tmp/pocl/build_rt; \
|
||||
cmake --build . --target install
|
||||
RUN rm -rf /tmp/pocl
|
||||
|
||||
@@ -1,41 +0,0 @@
|
||||
|
||||
LIB_PATH = ../../runtime
|
||||
|
||||
|
||||
COMP = /home/fares/dev/riscv-gnu-toolchain-vector/drops/bin/riscv32-unknown-elf-gcc
|
||||
|
||||
CC_FLAGS = -ffreestanding -O0 -Wl,--gc-sections -nostartfiles -nostdlib -nostartfiles -nodefaultlibs -Wl,-Bstatic,-T,$(LIB_PATH)/startup/vx_link.ld -march=rv32imv -mabi=ilp32
|
||||
|
||||
DMP = /home/fares/dev/riscv-gnu-toolchain-vector/drops/bin/riscv32-unknown-elf-objdump
|
||||
CPY = /home/fares/dev/riscv-gnu-toolchain-vector/drops/bin/riscv32-unknown-elf-objcopy
|
||||
|
||||
# VX_STR = ../../startup/vx_start.S
|
||||
|
||||
|
||||
|
||||
NEWLIB = $(LIB_PATH)/newlib/newlib.c
|
||||
VX_STR = $(LIB_PATH)/startup/vx_start.S
|
||||
VX_INT = $(LIB_PATH)/intrinsics/vx_intrinsics.S
|
||||
VX_IO = $(LIB_PATH)/io/vx_io.S $(LIB_PATH)/io/vx_io.c
|
||||
VX_API = $(LIB_PATH)/vx_api/vx_api.c
|
||||
VX_FIO = $(LIB_PATH)/fileio/fileio.S
|
||||
VX_VEC = vx_vec.s
|
||||
LIBS = /home/fares/dev/riscv-gnu-toolchain-vector/drops/riscv32-unknown-elf/lib/libc.a /home/fares/dev/riscv-gnu-toolchain-vector/drops/riscv32-unknown-elf/lib/libstdc++.a -static-libgcc -lgcc
|
||||
|
||||
VX_MAIN = vx_vector_main
|
||||
|
||||
all: HEX DUMP ELF
|
||||
|
||||
DUMP: ELF
|
||||
$(DMP) -D $(VX_MAIN).elf > $(VX_MAIN).dump
|
||||
|
||||
HEX: ELF
|
||||
$(CPY) -O ihex $(VX_MAIN).elf $(VX_MAIN).hex
|
||||
|
||||
ELF:
|
||||
$(COMP) $(CC_FLAGS) $(VX_STR) $(VX_VEC) $(VX_FIO) $(NEWLIB) $(VX_INT) $(VX_IO) $(VX_API) $(VX_MAIN).c $(LIBS) -Iinclude -o $(VX_MAIN).elf
|
||||
|
||||
run:
|
||||
../../simx/obj_dir/Vcache_simX -E -a rv32i --core vx_vector_main.hex -s -b 1> emulator.debug
|
||||
|
||||
|
||||
@@ -1,30 +0,0 @@
|
||||
|
||||
|
||||
|
||||
.type vx_vec_test, @function
|
||||
.global vx_vec_test
|
||||
vx_vec_test:
|
||||
li a1, 7
|
||||
sw a1, 0(a0)
|
||||
ret
|
||||
|
||||
|
||||
|
||||
|
||||
# slli a0, a0, 2
|
||||
# add a0, a0, a3
|
||||
# vmv.v.x vv0, a2
|
||||
# # vsplat4 vv0, a2
|
||||
# stripmine_loop:
|
||||
# vlb4 vv1, (a1)
|
||||
# vcmpez4 vp0, vv1
|
||||
# !vp0 vlw4 vv1, (a3)
|
||||
# !vp0 vlw4 vv2, (a4)
|
||||
# !vp0 vfma4 vv1, vv0, vv1, vv2
|
||||
# !vp0 vsw4 vv1, (a4)
|
||||
# addi a1, a1, 4
|
||||
# addi a3, a3, 16
|
||||
# addi a4, a4, 16
|
||||
# bleu a3, a0, stripmine_loop
|
||||
# handle edge cases
|
||||
# when (n % 4) != 0 ...
|
||||
@@ -1,32 +0,0 @@
|
||||
|
||||
#include "../../runtime/intrinsics/vx_intrinsics.h"
|
||||
#include "vx_vec.h"
|
||||
|
||||
int main()
|
||||
{
|
||||
vx_tmc(1);
|
||||
// int * a = malloc(4);
|
||||
// int * b = malloc(4);
|
||||
// int * c = malloc(4);
|
||||
|
||||
|
||||
int * a = malloc(4);
|
||||
*a = 5;
|
||||
printf("Value of a: %d\n", *a);
|
||||
|
||||
vx_vec_test(a);
|
||||
|
||||
printf("Value of a: %d\n", *a);
|
||||
|
||||
|
||||
// for (int i = 0; i < 4; i++)
|
||||
// {
|
||||
// if (c[i] != (a[i] + b[i]))
|
||||
// {
|
||||
// printf("Fail\n");
|
||||
// break;
|
||||
// }
|
||||
// }
|
||||
|
||||
vx_tmc(0);
|
||||
}
|
||||
@@ -1,91 +0,0 @@
|
||||
#include <stdio.h>
|
||||
#include <stdlib.h>
|
||||
#include "../../runtime/intrinsics/vx_intrinsics.h"
|
||||
#include "vx_vec.h"
|
||||
|
||||
int main()
|
||||
{
|
||||
vx_tmc(1);
|
||||
#if 0
|
||||
# vector-vector add routine of 32-bit integers
|
||||
# void vvaddint32(size_t n, const int*x, const int*y, int*z)
|
||||
# { for (size_t i=0; i<n; i++) { z[i]=x[i]+y[i]; } }
|
||||
#
|
||||
# a0 = n, a1 = x, a2 = y, a3 = z
|
||||
# Non-vector instructions are indented
|
||||
#endif
|
||||
#if 1
|
||||
int n = 5;
|
||||
int *a = (int*)malloc(sizeof(int) * n); //{1, 1, 1, 1, 1};
|
||||
int *b = (int*)malloc(sizeof(int) * n); //{1, 1, 1, 1, 1};
|
||||
int *c = (int*)malloc(sizeof(int) * n); //{1, 1, 1, 1, 1};
|
||||
|
||||
for(int i = 0; i < n; ++i)
|
||||
{
|
||||
a[i] = b[i] = c[i] = 1;
|
||||
}
|
||||
|
||||
for(int i = 0; i < n; ++i) printf("%d, ", a[i]);
|
||||
printf("\n");
|
||||
// for(int i = 0; i < n; ++i) printf("%d, ", b[i]);
|
||||
// printf("\n");
|
||||
// for(int i = 0; i < n; ++i) printf("%d, ", c[i]);
|
||||
|
||||
int *d;
|
||||
*d = 1;
|
||||
vx_vec_test(n, d, b, c);
|
||||
|
||||
|
||||
printf("(after: n = %d, %d)\n", n, *d);
|
||||
for(int i = 0; i < n; ++i) printf("%d, ", a[i]);
|
||||
// printf("\n");
|
||||
// for(int i = 0; i < n; ++i) printf("%d, ", b[i]);
|
||||
// printf("\n");
|
||||
// for(int i = 0; i < n; ++i) printf("%d, ", c[i]);
|
||||
|
||||
#endif
|
||||
#if 0
|
||||
int * a = malloc(sizeof(int) * 10);
|
||||
for(int i = 0; i < 10; ++i) a[i] = 5;
|
||||
|
||||
|
||||
for(int i = 0; i < 10; ++i)
|
||||
printf("%d, ", a[i]);
|
||||
|
||||
vx_vec_test(a);
|
||||
//vx_vec_test(2, a, a, a);
|
||||
|
||||
printf("after--------\n");
|
||||
for(int i = 0; i < 10; ++i)
|
||||
printf("%d, ", a[i]);
|
||||
#endif
|
||||
#if 0
|
||||
int n = 5;
|
||||
int *a = (int*)malloc(sizeof(int) * 5); //{1, 1, 1, 1, 1};
|
||||
int *b = (int*)malloc(sizeof(int) * 5); //{1, 1, 1, 1, 1};
|
||||
int *c = (int*)malloc(sizeof(int) * 5); //{1, 1, 1, 1, 1};
|
||||
|
||||
for(int i = 0; i < n; ++i)
|
||||
{
|
||||
a[i] = 1;
|
||||
b[i] = 1;
|
||||
c[i] = 0;
|
||||
}
|
||||
|
||||
printf("Value of a: %d, b: %d, c: %d, n: %d\n", a[0], b[0], c[0], n);
|
||||
vx_vec_test(n, a, b, c);
|
||||
printf("Value of a: %d, b: %d, c: %d, n: %d\n", a[0], b[0], c[0], n);
|
||||
|
||||
#endif
|
||||
|
||||
// for (int i = 0; i < 4; i++)
|
||||
// {
|
||||
// if (c[i] != (a[i] + b[i]))
|
||||
// {
|
||||
// printf("Fail\n");
|
||||
// break;
|
||||
// }
|
||||
// }
|
||||
|
||||
vx_tmc(0);
|
||||
}
|
||||
@@ -1,15 +0,0 @@
|
||||
|
||||
|
||||
#pragma once
|
||||
|
||||
|
||||
#ifdef __cplusplus
|
||||
extern "C" {
|
||||
#endif
|
||||
|
||||
void vx_vec_test(int n, int* a, int* b, int* c); //vvaddint32
|
||||
|
||||
|
||||
#ifdef __cplusplus
|
||||
}
|
||||
#endif
|
||||
@@ -1,23 +0,0 @@
|
||||
|
||||
|
||||
.type vx_vec_test, @function
|
||||
.global vx_vec_test
|
||||
vx_vec_test:
|
||||
# vector-vector add routine of 32-bit integers
|
||||
# void vvaddint32(size_t n, const int*x, const int*y, int*z)
|
||||
# { for (size_t i=0; i<n; i++) { z[i]=x[i]+y[i]; } }
|
||||
#
|
||||
# a0 = n, a1 = x, a2 = y, a3 = z
|
||||
# Non-vector instructions are indented
|
||||
vsetvli t0, a0, e32 # Set vector length based on 32-bit vectors
|
||||
vlw.v v0, (a1) # Get first vector
|
||||
sub a0, a0, t0 # Decrement number done
|
||||
slli t0, t0, 2 # Multiply number done by 4 bytes
|
||||
add a1, a1, t0 # Bump pointer
|
||||
vlw.v v1, (a2) # Get second vector
|
||||
add a2, a2, t0 # Bump pointer
|
||||
vadd.vv v2, v0, v1 # Sum vectors
|
||||
vsw.v v2, (a3) # Store result
|
||||
add a3, a3, t0 # Bump pointer
|
||||
bnez a0, vx_vec_test # Loop back
|
||||
ret # Finished
|
||||
@@ -1,27 +0,0 @@
|
||||
#include "../../runtime/intrinsics/vx_intrinsics.h"
|
||||
#include "vx_vec.h"
|
||||
|
||||
int main()
|
||||
{
|
||||
vx_tmc(1);
|
||||
printf("----------------hello!!! \n");
|
||||
|
||||
int n = 8;
|
||||
int *a = (int*)malloc(sizeof(int) * n); //{1, 1, 1, 1, 1};
|
||||
int *b = (int*)malloc(sizeof(int) * n); //{1, 1, 1, 1, 1};
|
||||
int *c = (int*)malloc(sizeof(int) * n); //{1, 1, 1, 1, 1};
|
||||
|
||||
printf("hello!!! \n");
|
||||
|
||||
for(int i = 0; i < n; ++i)
|
||||
{
|
||||
a[i] = b[i] = c[i] = 1;
|
||||
}
|
||||
|
||||
vx_vec_test(n, a, b, c);
|
||||
|
||||
for(int i = 0; i < n; ++i)
|
||||
printf("%d ", c[i]);
|
||||
|
||||
vx_tmc(0);
|
||||
}
|
||||
@@ -1,29 +0,0 @@
|
||||
#include "../../runtime/intrinsics/vx_intrinsics.h"
|
||||
#include "vx_vec.h"
|
||||
|
||||
int main()
|
||||
{
|
||||
vx_tmc(1);
|
||||
|
||||
printf("Hello\n");
|
||||
|
||||
int n = 64;
|
||||
int *a = (int*)malloc(sizeof(int) * n); //{1, 1, 1, 1, 1};
|
||||
int *b = (int*)malloc(sizeof(int) * n); //{1, 1, 1, 1, 1};
|
||||
int *c = (int*)malloc(sizeof(int) * n); //{1, 1, 1, 1, 1};
|
||||
|
||||
for(int i = 0; i < n; ++i)
|
||||
{
|
||||
a[i] = b[i] = c[i] = 1;
|
||||
}
|
||||
|
||||
vx_vec_test(n, a, b, c);
|
||||
|
||||
for (int i = 0; i < n; ++i)
|
||||
{
|
||||
printf("a[%d]=%d, b[%d]=%d, c[%d]=%d\n", i, a[i], i, b[i], i, c[i]);
|
||||
}
|
||||
|
||||
|
||||
vx_tmc(0);
|
||||
}
|
||||
@@ -1,39 +0,0 @@
|
||||
LIB_PATH = ../../runtime
|
||||
|
||||
COMP = /home/priya/dev/riscv_vec/riscv-gnu/bin/riscv32-unknown-elf-gcc
|
||||
|
||||
CC_FLAGS = -ffreestanding -O0 -Wl,--gc-sections -nostartfiles -nostdlib -nostartfiles -nodefaultlibs -Wl,-Bstatic,-T,$(LIB_PATH)/startup/vx_link.ld -march=rv32imv -mabi=ilp32
|
||||
|
||||
DMP = /home/priya/dev/riscv_vec/riscv-gnu/bin/riscv32-unknown-elf-objdump
|
||||
CPY = /home/priya/dev/riscv_vec/riscv-gnu/bin/riscv32-unknown-elf-objcopy
|
||||
|
||||
# VX_STR = ../../startup/vx_start.S
|
||||
|
||||
NEWLIB = $(LIB_PATH)/newlib/newlib.c
|
||||
VX_STR = $(LIB_PATH)/startup/vx_start.S
|
||||
VX_INT = $(LIB_PATH)/intrinsics/vx_intrinsics.S
|
||||
VX_IO = $(LIB_PATH)/io/vx_io.S $(LIB_PATH)/io/vx_io.c
|
||||
VX_API = $(LIB_PATH)/vx_api/vx_api.c
|
||||
VX_FIO = $(LIB_PATH)/fileio/fileio.S
|
||||
VX_VEC1 = vx_vec_vvaddint32.s
|
||||
VX_VEC2 = vx_vec_saxpy.s #float --> int
|
||||
VX_VEC3 = vx_vec_sgemm_float.s #float --> int
|
||||
VX_VEC4 = vx_vec_vsadd.s
|
||||
VX_VEC5 = vx_vec_memcpy.s
|
||||
LIBS = /home/priya/dev/riscv_vec/riscv-gnu/riscv32-unknown-elf/lib/libc.a /home/priya/dev/riscv_vec/riscv-gnu/riscv32-unknown-elf/lib/libstdc++.a -static-libgcc -lgcc
|
||||
|
||||
VX_MAIN = vx_vec_benchmark
|
||||
|
||||
all: HEX DUMP ELF
|
||||
|
||||
DUMP: ELF
|
||||
$(DMP) -D $(VX_MAIN).elf > $(VX_MAIN).dump
|
||||
|
||||
HEX: ELF
|
||||
$(CPY) -O ihex $(VX_MAIN).elf $(VX_MAIN).hex
|
||||
|
||||
ELF:
|
||||
$(COMP) $(CC_FLAGS) $(VX_STR) $(VX_VEC2) $(VX_FIO) $(NEWLIB) $(VX_INT) $(VX_IO) $(VX_API) $(VX_MAIN).c $(LIBS) -Iinclude -o $(VX_MAIN).elf
|
||||
# $(COMP) $(CC_FLAGS) $(VX_STR) $(VX_VEC3) $(VX_FIO) $(NEWLIB) $(VX_INT) $(VX_IO) $(VX_API) $(VX_MAIN).c $(LIBS) -Iinclude -o $(VX_MAIN).elf
|
||||
# $(COMP) $(CC_FLAGS) $(VX_STR) $(VX_VEC4) $(VX_FIO) $(NEWLIB) $(VX_INT) $(VX_IO) $(VX_API) $(VX_MAIN).c $(LIBS) -Iinclude -o $(VX_MAIN).elf
|
||||
# $(COMP) $(CC_FLAGS) $(VX_STR) $(VX_VEC5) $(VX_FIO) $(NEWLIB) $(VX_INT) $(VX_IO) $(VX_API) $(VX_MAIN).c $(LIBS) -Iinclude -o $(VX_MAIN).elf~
|
||||
@@ -1,9 +0,0 @@
|
||||
1. add benchmarks under (Vortex/benchmarks/..)
|
||||
1.1 bfs --> blas spmv approach
|
||||
1.2 kmeans // stage 2
|
||||
1.3 saxpy --> sample
|
||||
1.4 sfilter // stage 2
|
||||
1.5 sgemm --> sample modify (float --> int)
|
||||
1.6 vecadd --> sample
|
||||
|
||||
|
||||
@@ -1,177 +0,0 @@
|
||||
#include <stdio.h>
|
||||
#include <stdlib.h>
|
||||
#include "../../runtime/intrinsics/vx_intrinsics.h"
|
||||
#include "vx_vec_benchmark.h"
|
||||
|
||||
int main()
|
||||
{
|
||||
vx_tmc(1);
|
||||
|
||||
int n = 5;
|
||||
int scalar = 10;
|
||||
|
||||
int *a = (int*)malloc(sizeof(int) * n); //{1, 1, 1, 1, 1};
|
||||
int *b = (int*)malloc(sizeof(int) * n); //{1, 1, 1, 1, 1};
|
||||
int *c = (int*)malloc(sizeof(int) * n); //{1, 1, 1, 1, 1};
|
||||
|
||||
for (int i = 0; i < n; ++i) { a[i] = 1; b[i] = 2; c[i] = 5; }
|
||||
|
||||
#if 0
|
||||
//---------------------------------------------------------------
|
||||
/* vvaddint32
|
||||
* # vector-vector add routine of 32-bit integers
|
||||
* # void vvaddint32(size_t n, const int*x, const int*y, int*z)
|
||||
* # { for (size_t i=0; i<n; i++) { z[i]=x[i]+y[i]; } } */
|
||||
printf("vvaddint...\na[%d]: ", n);
|
||||
for(int i = 0; i < n; ++i) printf("%d ", a[i]);
|
||||
printf("\nb[%d]: ", n);
|
||||
for(int i = 0; i < n; ++i) printf("%d ", b[i]);
|
||||
printf("\nc[%d] = a[%d] + b[%d]: ", n, n, n);
|
||||
for(int i = 0; i < n; ++i) printf("%d ", c[i]);
|
||||
|
||||
vx_vec_vvaddint32(n, a, b, c);
|
||||
|
||||
for(int i = 0; i < n; ++i)
|
||||
{
|
||||
if(c[i] != (a[i]+b[i]))
|
||||
{
|
||||
printf("\n<vddint32> failed at <index: %d>! \n", i);
|
||||
return 1;
|
||||
}
|
||||
}
|
||||
printf("\nPASSED.......................... <vddint32> \n");
|
||||
#endif
|
||||
#if 0
|
||||
//---------------------------------------------------------------
|
||||
/* # vector-scalar add
|
||||
# for (i=0; i<N; i++) { C[i] = A[i] + B; } // 32-bit ints */
|
||||
for (int i = 0; i < n; ++i) { a[i] = 1; b[i] = 1;}
|
||||
printf("vsadd...scalar:%d\na[%d]: ", scalar, n);
|
||||
for(int i = 0; i < n; ++i) printf("%d \n", a[i]);
|
||||
printf("\nb: %d", scalar);
|
||||
|
||||
vx_vec_vsadd(n, a, scalar);
|
||||
|
||||
for(int i = 0; i < n; ++i)
|
||||
{
|
||||
if(a[i] != (b[i] * scalar))
|
||||
{
|
||||
printf("\n<vsadd> failed at <index: %d>! \n", i);
|
||||
return 1;
|
||||
}
|
||||
}
|
||||
printf("\nPASSED.......................... <vsadd> \n");
|
||||
|
||||
#endif
|
||||
#if 0
|
||||
//---------------------------------------------------------------
|
||||
/* # memory copy
|
||||
# void *memcpy(void* dest, const void* src, size_t n) */
|
||||
for (int i = 0; i < n; ++i) { a[i] = 1; b[i] = 2;}
|
||||
printf("memcpy\na[%d]: ", n);
|
||||
for(int i = 0; i < n; ++i) printf("%d \n", a[i]);
|
||||
printf("\nb[%d]: ", n);
|
||||
for(int i = 0; i < n; ++i) printf("%d \n", b[i]);
|
||||
|
||||
vx_vec_memcpy(a, b, n);
|
||||
|
||||
for(int i = 0; i < n; ++i)
|
||||
{
|
||||
if(a[i] != b[i])
|
||||
{
|
||||
printf("\n<memcpy> failed at <index: %d>! \n", i);
|
||||
<<<<<<< HEAD
|
||||
return;
|
||||
}
|
||||
}
|
||||
printf("\nPASSED.......................... <memcpy> \n");
|
||||
=======
|
||||
return 1;
|
||||
}
|
||||
}
|
||||
printf("\nPASSED.......................... <memcpy> \n");
|
||||
#endif
|
||||
#if 1
|
||||
//---------------------------------------------------------------
|
||||
/* # void saxpy(size_t n, const float a, const float *x, float *y)
|
||||
# ==> convert to int!!
|
||||
# void saxpy(size_t n, const int a, const int *x, int *y)
|
||||
# {
|
||||
# size_t i;
|
||||
# for (i=0; i<n; i++) y[i] = a * x[i] + y[i];
|
||||
# } */
|
||||
for (int i = 0; i < n; ++i) { a[i] = 1; b[i] = 2; c[i] = 2;}
|
||||
|
||||
printf("saxpy\na[%d]: ", n);
|
||||
for(int i = 0; i < n; ++i) printf("%d \n", a[i]);
|
||||
printf("\nb[%d]: ", n);
|
||||
for(int i = 0; i < n; ++i) printf("%d \n", b[i]);
|
||||
|
||||
vx_vec_saxpy(n, scalar, a, b);
|
||||
|
||||
printf("saxpy\na[%d]: ", n);
|
||||
for(int i = 0; i < n; ++i) printf("%d \n", a[i]);
|
||||
printf("\nb[%d]: ", n);
|
||||
for(int i = 0; i < n; ++i) printf("%d \n", b[i]);
|
||||
|
||||
for(int i = 0; i < n; ++i)
|
||||
{
|
||||
if(b[i] != ((a[i] * scalar) + c[i]))
|
||||
{
|
||||
printf("\n<saxpy> failed at <index: %d>! \n", i);
|
||||
return;
|
||||
}
|
||||
}
|
||||
printf("\nPASSED.......................... <saxpy> \n");
|
||||
|
||||
return 1;
|
||||
}
|
||||
}
|
||||
printf("\nPASSED.......................... <saxpy> \n");
|
||||
#endif
|
||||
#if 0
|
||||
//---------------------------------------------------------------
|
||||
/* # void sgemm_nn(size_t n, size_t m, size_t k, const float*a, // m * k matrix
|
||||
# size_t lda, const float*b, // k * n matrix
|
||||
# size_t ldb, float*c, // m * n matrix
|
||||
# size_t ldc)
|
||||
# c += a*b (alpha=1, no transpose on input matrices)
|
||||
# matrices stored in C row-major order */
|
||||
|
||||
int m = 8;
|
||||
int k = 8;
|
||||
int n = 8
|
||||
int lda = 4;
|
||||
int ldb = 4;
|
||||
int ldc = 4;
|
||||
|
||||
int* a1 = (int*)malloc(sizeof(m * k));
|
||||
int* b1 = (int*)malloc(sizeof(k * n));
|
||||
int* c1 = (int*)malloc(sizeof(m * n));
|
||||
|
||||
for(int i = 0; i < (m * k); ++i) a1[i] = 1;
|
||||
for(int i = 0; i < (k * n); ++i) b1[i] = 1;
|
||||
for(int i = 0; i < (m * n); ++i) c1[i] = 1;
|
||||
|
||||
printf("sgemm_nn\na[%d]: ", n);
|
||||
for(int i = 0; i < n; ++i) printf("%d \n", a1[i]);
|
||||
printf("\nb[%d]: ", n);
|
||||
for(int i = 0; i < n; ++i) printf("%d \n", b1[i]);
|
||||
|
||||
vx_vec_sgemm_nn(n, m, k, a1, lda, b1, ldb, c1, ldc);
|
||||
|
||||
//for(int i = 0; i < n; ++i)
|
||||
//{
|
||||
// if(b[i] != ((a[i] * scalar) + c[i]))
|
||||
// {
|
||||
// printf("\n<sgemm_nn> failed at <index: %d>! \n", i);
|
||||
// return;
|
||||
// }
|
||||
//}
|
||||
printf("\nNOT TESTED.......................... <sgemm_nn> \n");
|
||||
//---------------------------------------------------------------
|
||||
#endif
|
||||
|
||||
vx_tmc(0);
|
||||
return 0;
|
||||
}
|
||||
File diff suppressed because it is too large
Load Diff
Binary file not shown.
@@ -1,16 +0,0 @@
|
||||
#pragma once
|
||||
|
||||
|
||||
#ifdef __cplusplus
|
||||
extern "C" {
|
||||
#endif
|
||||
|
||||
//void vx_vec_vvaddint32(int n, int* a, int* b, int *c);
|
||||
//void vx_vec_vsadd(int n, int* a, int scalar);
|
||||
//void vx_vec_memcpy(int* a, int* b, int n);
|
||||
void vx_vec_saxpy(int n, int scalar, int* a, int* b);
|
||||
//void vx_vec_sgemm_nn(int n, int m, int k, int* a1, int lda, int* b1, int ldb, int* c1, int ldc);
|
||||
|
||||
#ifdef __cplusplus
|
||||
}
|
||||
#endif
|
||||
File diff suppressed because it is too large
Load Diff
@@ -1,17 +0,0 @@
|
||||
.type vx_vec_memcpy, @function
|
||||
.global vx_vec_memcpy
|
||||
# void *memcpy(void* dest, const void* src, size_t n)
|
||||
# a0=dest, a1=src, a2=n
|
||||
#
|
||||
vx_vec_memcpy:
|
||||
# memcpy
|
||||
mv a3, a0 # Copy destination
|
||||
vsetvli t0, a2, e8,m8 # Vectors of 8b
|
||||
loop:
|
||||
vlb.v v0, (a1) # Load bytes
|
||||
add a1, a1, t0 # Bump pointer
|
||||
sub a2, a2, t0 # Decrement count
|
||||
vsb.v v0, (a3) # Store bytes
|
||||
add a3, a3, t0 # Bump pointer
|
||||
bnez a2, loop # Any more?
|
||||
ret # Return
|
||||
@@ -1,62 +0,0 @@
|
||||
.type vx_vec_saxpy, @function
|
||||
.global vx_vec_saxpy
|
||||
# void
|
||||
# saxpy(size_t n, const float a, const float *x, float *y)
|
||||
# {
|
||||
# size_t i;
|
||||
# for (i=0; i<n; i++)
|
||||
# y[i] = a * x[i] + y[i];
|
||||
# }
|
||||
#
|
||||
# register arguments:
|
||||
# a0 n
|
||||
# fa0 a
|
||||
# a1 x
|
||||
# a2 y
|
||||
vx_vec_saxpy:
|
||||
vsetvli a4, a0, e32, m8
|
||||
saxpy:
|
||||
vlw.v v0, (a1)
|
||||
sub a0, a0, a4
|
||||
slli a4, a4, 2
|
||||
add a1, a1, a4
|
||||
vlw.v v8, (a2)
|
||||
vfmacc.vf v8, fa0, v0
|
||||
vsw.v v8, (a2)
|
||||
add a2, a2, a4
|
||||
bnez a0, saxpy
|
||||
ret
|
||||
#vx_vec_saxpy:
|
||||
# vsetvli a4, a0, e32, m8
|
||||
#saxpy:
|
||||
# vlw.v v0, (a1)
|
||||
# sub a0, a0, a4
|
||||
# slli a4, a4, 2
|
||||
# add a1, a1, a4
|
||||
# vlw.v v8, (a2)
|
||||
# vfmacc.vf v8, fa0, v0
|
||||
# vsw.v v8, (a2)
|
||||
# add a2, a2, a4
|
||||
# bnez a0, saxpy
|
||||
# ret
|
||||
|
||||
# a0 n, rs1 a, a2 x, a3 y
|
||||
|
||||
# a0 n, a1 a, a2 x, a3 y
|
||||
vx_vec_saxpy:
|
||||
vsetvli a4, a0, e32, m1
|
||||
saxpy:
|
||||
vlw.v v0, (a2)
|
||||
sub a0, a0, a4
|
||||
slli a4, a4, 2
|
||||
add a2, a2, a4
|
||||
vlw.v v1, (a3)
|
||||
vmul.vx v0, v0, a1
|
||||
# vmul.vv v0, v0, v1
|
||||
# li x1, 10
|
||||
# vmul.vx v0, v0, x1
|
||||
vadd.vv v1, v0, v1
|
||||
vsw.v v1, (a3)
|
||||
add a3, a3, a4
|
||||
bnez a0, saxpy
|
||||
ret
|
||||
@@ -1,28 +0,0 @@
|
||||
.type vx_vec_saxpy, @function
|
||||
.global vx_vec_saxpy_float
|
||||
# void
|
||||
# saxpy(size_t n, const float a, const float *x, float *y)
|
||||
# {
|
||||
# size_t i;
|
||||
# for (i=0; i<n; i++)
|
||||
# y[i] = a * x[i] + y[i];
|
||||
# }
|
||||
#
|
||||
# register arguments:
|
||||
# a0 n
|
||||
# fa0 a
|
||||
# a1 x
|
||||
# a2 y
|
||||
vx_vec_saxpy_float:
|
||||
vsetvli a4, a0, e32, m8
|
||||
saxpy:
|
||||
vlw.v v0, (a1)
|
||||
sub a0, a0, a4
|
||||
slli a4, a4, 2
|
||||
add a1, a1, a4
|
||||
vlw.v v8, (a2)
|
||||
vfmacc.vf v8, fa0, v0
|
||||
vsw.v v8, (a2)
|
||||
add a2, a2, a4
|
||||
bnez a0, saxpy
|
||||
ret
|
||||
@@ -1,220 +0,0 @@
|
||||
.type vx_vec_sgemm_nn, @function
|
||||
.global vx_vec_sgemm_nn
|
||||
# RV64IDV system
|
||||
#
|
||||
# void
|
||||
# sgemm_nn(size_t n,
|
||||
# size_t m,
|
||||
# size_t k,
|
||||
# const float*a, // m * k matrix
|
||||
# size_t lda,
|
||||
# const float*b, // k * n matrix
|
||||
# size_t ldb,
|
||||
# float*c, // m * n matrix
|
||||
# size_t ldc)
|
||||
#
|
||||
# c += a*b (alpha=1, no transpose on input matrices)
|
||||
# matrices stored in C row-major order
|
||||
|
||||
#define n a0
|
||||
#define m a1
|
||||
#define k a2
|
||||
#define ap a3
|
||||
#define astride a4
|
||||
#define bp a5
|
||||
#define bstride a6
|
||||
#define cp a7
|
||||
#define cstride t0
|
||||
#define kt t1
|
||||
#define nt t2
|
||||
#define bnp t3
|
||||
#define cnp t4
|
||||
#define akp t5
|
||||
#define bkp s0
|
||||
#define nvl s1
|
||||
#define ccp s2
|
||||
#define amp s3
|
||||
|
||||
# Use args as additional temporaries
|
||||
#define ft12 fa0
|
||||
#define ft13 fa1
|
||||
#define ft14 fa2
|
||||
#define ft15 fa3
|
||||
|
||||
# This version holds a 16*VLMAX block of C matrix in vector registers
|
||||
# in inner loop, but otherwise does not cache or TLB tiling.
|
||||
vx_vec_sgemm_nn:
|
||||
#sgemm_nn:
|
||||
addi sp, sp, -FRAMESIZE
|
||||
sd s0, OFFSET(sp)
|
||||
sd s1, OFFSET(sp)
|
||||
sd s2, OFFSET(sp)
|
||||
|
||||
# Check for zero size matrices
|
||||
beqz n, exit
|
||||
beqz m, exit
|
||||
beqz k, exit
|
||||
|
||||
# Convert elements strides to byte strides.
|
||||
ld cstride, OFFSET(sp) # Get arg from stack frame
|
||||
slli astride, astride, 2
|
||||
slli bstride, bstride, 2
|
||||
slli cstride, cstride, 2
|
||||
|
||||
slti t6, m, 16
|
||||
bnez t6, end_rows
|
||||
|
||||
c_row_loop: # Loop across rows of C blocks
|
||||
|
||||
mv nt, n # Initialize n counter for next row of C blocks
|
||||
|
||||
mv bnp, bp # Initialize B n-loop pointer to start
|
||||
mv cnp, cp # Initialize C n-loop pointer
|
||||
|
||||
c_col_loop: # Loop across one row of C blocks
|
||||
vsetvli nvl, nt, e32 # 32-bit vectors, LMUL=1
|
||||
|
||||
mv akp, ap # reset pointer into A to beginning
|
||||
mv bkp, bnp # step to next column in B matrix
|
||||
|
||||
# Initalize current C submatrix block from memory.
|
||||
vlw.v v0, (cnp); add ccp, cnp, cstride;
|
||||
vlw.v v1, (ccp); add ccp, ccp, cstride;
|
||||
vlw.v v2, (ccp); add ccp, ccp, cstride;
|
||||
vlw.v v3, (ccp); add ccp, ccp, cstride;
|
||||
vlw.v v4, (ccp); add ccp, ccp, cstride;
|
||||
vlw.v v5, (ccp); add ccp, ccp, cstride;
|
||||
vlw.v v6, (ccp); add ccp, ccp, cstride;
|
||||
vlw.v v7, (ccp); add ccp, ccp, cstride;
|
||||
vlw.v v8, (ccp); add ccp, ccp, cstride;
|
||||
vlw.v v9, (ccp); add ccp, ccp, cstride;
|
||||
vlw.v v10, (ccp); add ccp, ccp, cstride;
|
||||
vlw.v v11, (ccp); add ccp, ccp, cstride;
|
||||
vlw.v v12, (ccp); add ccp, ccp, cstride;
|
||||
vlw.v v13, (ccp); add ccp, ccp, cstride;
|
||||
vlw.v v14, (ccp); add ccp, ccp, cstride;
|
||||
vlw.v v15, (ccp)
|
||||
|
||||
|
||||
mv kt, k # Initialize inner loop counter
|
||||
|
||||
# Inner loop scheduled assuming 4-clock occupancy of vfmacc instruction and single-issue pipeline
|
||||
# Software pipeline loads
|
||||
flw ft0, (akp); add amp, akp, astride;
|
||||
flw ft1, (amp); add amp, amp, astride;
|
||||
flw ft2, (amp); add amp, amp, astride;
|
||||
flw ft3, (amp); add amp, amp, astride;
|
||||
# Get vector from B matrix
|
||||
vlw.v v16, (bkp)
|
||||
|
||||
# Loop on inner dimension for current C block
|
||||
k_loop:
|
||||
vfmacc.vf v0, ft0, v16
|
||||
add bkp, bkp, bstride
|
||||
flw ft4, (amp)
|
||||
add amp, amp, astride
|
||||
vfmacc.vf v1, ft1, v16
|
||||
addi kt, kt, -1 # Decrement k counter
|
||||
flw ft5, (amp)
|
||||
add amp, amp, astride
|
||||
vfmacc.vf v2, ft2, v16
|
||||
flw ft6, (amp)
|
||||
add amp, amp, astride
|
||||
flw ft7, (amp)
|
||||
vfmacc.vf v3, ft3, v16
|
||||
add amp, amp, astride
|
||||
flw ft8, (amp)
|
||||
add amp, amp, astride
|
||||
vfmacc.vf v4, ft4, v16
|
||||
flw ft9, (amp)
|
||||
add amp, amp, astride
|
||||
vfmacc.vf v5, ft5, v16
|
||||
flw ft10, (amp)
|
||||
add amp, amp, astride
|
||||
vfmacc.vf v6, ft6, v16
|
||||
flw ft11, (amp)
|
||||
add amp, amp, astride
|
||||
vfmacc.vf v7, ft7, v16
|
||||
flw ft12, (amp)
|
||||
add amp, amp, astride
|
||||
vfmacc.vf v8, ft8, v16
|
||||
flw ft13, (amp)
|
||||
add amp, amp, astride
|
||||
vfmacc.vf v9, ft9, v16
|
||||
flw ft14, (amp)
|
||||
add amp, amp, astride
|
||||
vfmacc.vf v10, ft10, v16
|
||||
flw ft15, (amp)
|
||||
add amp, amp, astride
|
||||
addi akp, akp, 4 # Move to next column of a
|
||||
vfmacc.vf v11, ft11, v16
|
||||
beqz kt, 1f # Don't load past end of matrix
|
||||
flw ft0, (akp)
|
||||
add amp, akp, astride
|
||||
1: vfmacc.vf v12, ft12, v16
|
||||
beqz kt, 1f
|
||||
flw ft1, (amp)
|
||||
add amp, amp, astride
|
||||
1: vfmacc.vf v13, ft13, v16
|
||||
beqz kt, 1f
|
||||
flw ft2, (amp)
|
||||
add amp, amp, astride
|
||||
1: vfmacc.vf v14, ft14, v16
|
||||
beqz kt, 1f # Exit out of loop
|
||||
flw ft3, (amp)
|
||||
add amp, amp, astride
|
||||
vfmacc.vf v15, ft15, v16
|
||||
vlw.v v16, (bkp) # Get next vector from B matrix, overlap loads with jump stalls
|
||||
j k_loop
|
||||
|
||||
1: vfmacc.vf v15, ft15, v16
|
||||
|
||||
# Save C matrix block back to memory
|
||||
vsw.v v0, (cnp); add ccp, cnp, cstride;
|
||||
vsw.v v1, (ccp); add ccp, ccp, cstride;
|
||||
vsw.v v2, (ccp); add ccp, ccp, cstride;
|
||||
vsw.v v3, (ccp); add ccp, ccp, cstride;
|
||||
vsw.v v4, (ccp); add ccp, ccp, cstride;
|
||||
vsw.v v5, (ccp); add ccp, ccp, cstride;
|
||||
vsw.v v6, (ccp); add ccp, ccp, cstride;
|
||||
vsw.v v7, (ccp); add ccp, ccp, cstride;
|
||||
vsw.v v8, (ccp); add ccp, ccp, cstride;
|
||||
vsw.v v9, (ccp); add ccp, ccp, cstride;
|
||||
vsw.v v10, (ccp); add ccp, ccp, cstride;
|
||||
vsw.v v11, (ccp); add ccp, ccp, cstride;
|
||||
vsw.v v12, (ccp); add ccp, ccp, cstride;
|
||||
vsw.v v13, (ccp); add ccp, ccp, cstride;
|
||||
vsw.v v14, (ccp); add ccp, ccp, cstride;
|
||||
vsw.v v15, (ccp)
|
||||
|
||||
# Following tail instructions should be scheduled earlier in free slots during C block save.
|
||||
# Leaving here for clarity.
|
||||
|
||||
# Bump pointers for loop across blocks in one row
|
||||
slli t6, nvl, 2
|
||||
add cnp, cnp, t6 # Move C block pointer over
|
||||
add bnp, bnp, t6 # Move B block pointer over
|
||||
sub nt, nt, nvl # Decrement element count in n dimension
|
||||
bnez nt, c_col_loop # Any more to do?
|
||||
|
||||
# Move to next set of rows
|
||||
addi m, m, -16 # Did 16 rows above
|
||||
slli t6, astride, 4 # Multiply astride by 16
|
||||
add ap, ap, t6 # Move A matrix pointer down 16 rows
|
||||
slli t6, cstride, 4 # Multiply cstride by 16
|
||||
add cp, cp, t6 # Move C matrix pointer down 16 rows
|
||||
|
||||
slti t6, m, 16
|
||||
beqz t6, c_row_loop
|
||||
|
||||
# Handle end of matrix with fewer than 16 rows.
|
||||
# Can use smaller versions of above decreasing in powers-of-2 depending on code-size concerns.
|
||||
end_rows:
|
||||
# Not done.
|
||||
|
||||
exit:
|
||||
ld s0, OFFSET(sp)
|
||||
ld s1, OFFSET(sp)
|
||||
ld s2, OFFSET(sp)
|
||||
addi sp, sp, FRAMESIZE
|
||||
ret
|
||||
@@ -1,220 +0,0 @@
|
||||
.type vx_vec_sgemm_nn_float, @function
|
||||
.global vx_vec_sgemm_nn_float
|
||||
# RV64IDV system
|
||||
#
|
||||
# void
|
||||
# sgemm_nn(size_t n,
|
||||
# size_t m,
|
||||
# size_t k,
|
||||
# const float*a, // m * k matrix
|
||||
# size_t lda,
|
||||
# const float*b, // k * n matrix
|
||||
# size_t ldb,
|
||||
# float*c, // m * n matrix
|
||||
# size_t ldc)
|
||||
#
|
||||
# c += a*b (alpha=1, no transpose on input matrices)
|
||||
# matrices stored in C row-major order
|
||||
|
||||
#define n a0
|
||||
#define m a1
|
||||
#define k a2
|
||||
#define ap a3
|
||||
#define astride a4
|
||||
#define bp a5
|
||||
#define bstride a6
|
||||
#define cp a7
|
||||
#define cstride t0
|
||||
#define kt t1
|
||||
#define nt t2
|
||||
#define bnp t3
|
||||
#define cnp t4
|
||||
#define akp t5
|
||||
#define bkp s0
|
||||
#define nvl s1
|
||||
#define ccp s2
|
||||
#define amp s3
|
||||
|
||||
# Use args as additional temporaries
|
||||
#define ft12 fa0
|
||||
#define ft13 fa1
|
||||
#define ft14 fa2
|
||||
#define ft15 fa3
|
||||
|
||||
# This version holds a 16*VLMAX block of C matrix in vector registers
|
||||
# in inner loop, but otherwise does not cache or TLB tiling.
|
||||
vx_vec_sgemm_nn_float:
|
||||
#sgemm_nn:
|
||||
addi sp, sp, -FRAMESIZE
|
||||
sd s0, OFFSET(sp)
|
||||
sd s1, OFFSET(sp)
|
||||
sd s2, OFFSET(sp)
|
||||
|
||||
# Check for zero size matrices
|
||||
beqz n, exit
|
||||
beqz m, exit
|
||||
beqz k, exit
|
||||
|
||||
# Convert elements strides to byte strides.
|
||||
ld cstride, OFFSET(sp) # Get arg from stack frame
|
||||
slli astride, astride, 2
|
||||
slli bstride, bstride, 2
|
||||
slli cstride, cstride, 2
|
||||
|
||||
slti t6, m, 16
|
||||
bnez t6, end_rows
|
||||
|
||||
c_row_loop: # Loop across rows of C blocks
|
||||
|
||||
mv nt, n # Initialize n counter for next row of C blocks
|
||||
|
||||
mv bnp, bp # Initialize B n-loop pointer to start
|
||||
mv cnp, cp # Initialize C n-loop pointer
|
||||
|
||||
c_col_loop: # Loop across one row of C blocks
|
||||
vsetvli nvl, nt, e32 # 32-bit vectors, LMUL=1
|
||||
|
||||
mv akp, ap # reset pointer into A to beginning
|
||||
mv bkp, bnp # step to next column in B matrix
|
||||
|
||||
# Initalize current C submatrix block from memory.
|
||||
vlw.v v0, (cnp); add ccp, cnp, cstride;
|
||||
vlw.v v1, (ccp); add ccp, ccp, cstride;
|
||||
vlw.v v2, (ccp); add ccp, ccp, cstride;
|
||||
vlw.v v3, (ccp); add ccp, ccp, cstride;
|
||||
vlw.v v4, (ccp); add ccp, ccp, cstride;
|
||||
vlw.v v5, (ccp); add ccp, ccp, cstride;
|
||||
vlw.v v6, (ccp); add ccp, ccp, cstride;
|
||||
vlw.v v7, (ccp); add ccp, ccp, cstride;
|
||||
vlw.v v8, (ccp); add ccp, ccp, cstride;
|
||||
vlw.v v9, (ccp); add ccp, ccp, cstride;
|
||||
vlw.v v10, (ccp); add ccp, ccp, cstride;
|
||||
vlw.v v11, (ccp); add ccp, ccp, cstride;
|
||||
vlw.v v12, (ccp); add ccp, ccp, cstride;
|
||||
vlw.v v13, (ccp); add ccp, ccp, cstride;
|
||||
vlw.v v14, (ccp); add ccp, ccp, cstride;
|
||||
vlw.v v15, (ccp)
|
||||
|
||||
|
||||
mv kt, k # Initialize inner loop counter
|
||||
|
||||
# Inner loop scheduled assuming 4-clock occupancy of vfmacc instruction and single-issue pipeline
|
||||
# Software pipeline loads
|
||||
flw ft0, (akp); add amp, akp, astride;
|
||||
flw ft1, (amp); add amp, amp, astride;
|
||||
flw ft2, (amp); add amp, amp, astride;
|
||||
flw ft3, (amp); add amp, amp, astride;
|
||||
# Get vector from B matrix
|
||||
vlw.v v16, (bkp)
|
||||
|
||||
# Loop on inner dimension for current C block
|
||||
k_loop:
|
||||
vfmacc.vf v0, ft0, v16
|
||||
add bkp, bkp, bstride
|
||||
flw ft4, (amp)
|
||||
add amp, amp, astride
|
||||
vfmacc.vf v1, ft1, v16
|
||||
addi kt, kt, -1 # Decrement k counter
|
||||
flw ft5, (amp)
|
||||
add amp, amp, astride
|
||||
vfmacc.vf v2, ft2, v16
|
||||
flw ft6, (amp)
|
||||
add amp, amp, astride
|
||||
flw ft7, (amp)
|
||||
vfmacc.vf v3, ft3, v16
|
||||
add amp, amp, astride
|
||||
flw ft8, (amp)
|
||||
add amp, amp, astride
|
||||
vfmacc.vf v4, ft4, v16
|
||||
flw ft9, (amp)
|
||||
add amp, amp, astride
|
||||
vfmacc.vf v5, ft5, v16
|
||||
flw ft10, (amp)
|
||||
add amp, amp, astride
|
||||
vfmacc.vf v6, ft6, v16
|
||||
flw ft11, (amp)
|
||||
add amp, amp, astride
|
||||
vfmacc.vf v7, ft7, v16
|
||||
flw ft12, (amp)
|
||||
add amp, amp, astride
|
||||
vfmacc.vf v8, ft8, v16
|
||||
flw ft13, (amp)
|
||||
add amp, amp, astride
|
||||
vfmacc.vf v9, ft9, v16
|
||||
flw ft14, (amp)
|
||||
add amp, amp, astride
|
||||
vfmacc.vf v10, ft10, v16
|
||||
flw ft15, (amp)
|
||||
add amp, amp, astride
|
||||
addi akp, akp, 4 # Move to next column of a
|
||||
vfmacc.vf v11, ft11, v16
|
||||
beqz kt, 1f # Don't load past end of matrix
|
||||
flw ft0, (akp)
|
||||
add amp, akp, astride
|
||||
1: vfmacc.vf v12, ft12, v16
|
||||
beqz kt, 1f
|
||||
flw ft1, (amp)
|
||||
add amp, amp, astride
|
||||
1: vfmacc.vf v13, ft13, v16
|
||||
beqz kt, 1f
|
||||
flw ft2, (amp)
|
||||
add amp, amp, astride
|
||||
1: vfmacc.vf v14, ft14, v16
|
||||
beqz kt, 1f # Exit out of loop
|
||||
flw ft3, (amp)
|
||||
add amp, amp, astride
|
||||
vfmacc.vf v15, ft15, v16
|
||||
vlw.v v16, (bkp) # Get next vector from B matrix, overlap loads with jump stalls
|
||||
j k_loop
|
||||
|
||||
1: vfmacc.vf v15, ft15, v16
|
||||
|
||||
# Save C matrix block back to memory
|
||||
vsw.v v0, (cnp); add ccp, cnp, cstride;
|
||||
vsw.v v1, (ccp); add ccp, ccp, cstride;
|
||||
vsw.v v2, (ccp); add ccp, ccp, cstride;
|
||||
vsw.v v3, (ccp); add ccp, ccp, cstride;
|
||||
vsw.v v4, (ccp); add ccp, ccp, cstride;
|
||||
vsw.v v5, (ccp); add ccp, ccp, cstride;
|
||||
vsw.v v6, (ccp); add ccp, ccp, cstride;
|
||||
vsw.v v7, (ccp); add ccp, ccp, cstride;
|
||||
vsw.v v8, (ccp); add ccp, ccp, cstride;
|
||||
vsw.v v9, (ccp); add ccp, ccp, cstride;
|
||||
vsw.v v10, (ccp); add ccp, ccp, cstride;
|
||||
vsw.v v11, (ccp); add ccp, ccp, cstride;
|
||||
vsw.v v12, (ccp); add ccp, ccp, cstride;
|
||||
vsw.v v13, (ccp); add ccp, ccp, cstride;
|
||||
vsw.v v14, (ccp); add ccp, ccp, cstride;
|
||||
vsw.v v15, (ccp)
|
||||
|
||||
# Following tail instructions should be scheduled earlier in free slots during C block save.
|
||||
# Leaving here for clarity.
|
||||
|
||||
# Bump pointers for loop across blocks in one row
|
||||
slli t6, nvl, 2
|
||||
add cnp, cnp, t6 # Move C block pointer over
|
||||
add bnp, bnp, t6 # Move B block pointer over
|
||||
sub nt, nt, nvl # Decrement element count in n dimension
|
||||
bnez nt, c_col_loop # Any more to do?
|
||||
|
||||
# Move to next set of rows
|
||||
addi m, m, -16 # Did 16 rows above
|
||||
slli t6, astride, 4 # Multiply astride by 16
|
||||
add ap, ap, t6 # Move A matrix pointer down 16 rows
|
||||
slli t6, cstride, 4 # Multiply cstride by 16
|
||||
add cp, cp, t6 # Move C matrix pointer down 16 rows
|
||||
|
||||
slti t6, m, 16
|
||||
beqz t6, c_row_loop
|
||||
|
||||
# Handle end of matrix with fewer than 16 rows.
|
||||
# Can use smaller versions of above decreasing in powers-of-2 depending on code-size concerns.
|
||||
end_rows:
|
||||
# Not done.
|
||||
|
||||
exit:
|
||||
ld s0, OFFSET(sp)
|
||||
ld s1, OFFSET(sp)
|
||||
ld s2, OFFSET(sp)
|
||||
addi sp, sp, FRAMESIZE
|
||||
ret
|
||||
@@ -1,20 +0,0 @@
|
||||
.type vx_vec_vsadd, @function
|
||||
.global vx_vec_vsadd
|
||||
# vector-scalar add
|
||||
# N = a0, a[] = a1, B = a2
|
||||
# for (i=0; i<N; i++)
|
||||
# { C[i] = A[i] + B; } // 32-bit ints
|
||||
#
|
||||
vx_vec_vsadd:
|
||||
# vcfg2*V32bINT, 1*S32bINT #
|
||||
# vmv v2, a2 # Copy B to vector unit scalar
|
||||
loop:
|
||||
# setvl t0, a0 # a0 holds N, t0 holds amount done
|
||||
ld v0, a1 # load strip of A vector
|
||||
vadd v1, v0, v2 # add vectors
|
||||
st v1, a3 # store strip of C vector
|
||||
slli t1, t0, 2 # multiply by 4 to get bytes
|
||||
add a1, a1, t1 # bump pointers
|
||||
add a3, a3, t1
|
||||
sub a0, a0, t0 # Subtract amount done
|
||||
bnez a0, loop
|
||||
@@ -1,22 +0,0 @@
|
||||
.type vx_vec_vvaddi32, @function
|
||||
.global vx_vec_vvaddint32
|
||||
# vector-vector add routine of 32-bit integers
|
||||
# void vvaddint32(size_t n, const int*x, const int*y, int*z)
|
||||
# { for (size_t i=0; i<n; i++) { z[i]=x[i]+y[i]; } }
|
||||
#
|
||||
# a0 = n, a1 = x, a2 = y, a3 = z
|
||||
# Non-vector instructions are indented
|
||||
vx_vec_vvaddint32:
|
||||
vsetvli t0, a0, e32 # Set vector length based on 32-bit vectors
|
||||
loop:
|
||||
vlw.v v0, (a1) # Get first vector
|
||||
sub a0, a0, t0 # Decrement number done
|
||||
slli t0, t0, 2 # Multiply number done by 4 bytes
|
||||
add a1, a1, t0 # Bump pointer
|
||||
vlw.v v1, (a2) # Get second vector
|
||||
add a2, a2, t0 # Bump pointer
|
||||
vadd.vv v2, v0, v1 # Sum vectors
|
||||
vsw.v v2, (a3) # Store result
|
||||
add a3, a3, t0 # Bump pointer
|
||||
bnez a0, loop # Loop back
|
||||
ret # Finished
|
||||
@@ -1,34 +0,0 @@
|
||||
RISCV_TOOLCHAIN_PATH ?= /opt/riscv-gnu-toolchain
|
||||
VORTEX_RT_PATH ?= $(wildcard ../..)
|
||||
|
||||
CC = $(RISCV_TOOLCHAIN_PATH)/bin/riscv32-unknown-elf-gcc
|
||||
AR = $(RISCV_TOOLCHAIN_PATH)/bin/riscv32-unknown-elf-gcc-ar
|
||||
DP = $(RISCV_TOOLCHAIN_PATH)/bin/riscv32-unknown-elf-objdump
|
||||
CP = $(RISCV_TOOLCHAIN_PATH)/bin/riscv32-unknown-elf-objcopy
|
||||
|
||||
CFLAGS += -march=rv32imf -mabi=ilp32f -O3 -Wl,-Bstatic,-T,$(VORTEX_RT_PATH)/linker/vx_link.ld
|
||||
CFLAGS += -nostartfiles -ffreestanding -fno-exceptions -Wl,--gc-sections
|
||||
CFLAGS += -I$(VORTEX_RT_PATH)/include
|
||||
|
||||
LDFLAGS += $(VORTEX_RT_PATH)/libvortexrt.a
|
||||
|
||||
PROJECT = vx_vector_main
|
||||
|
||||
SRCS = vx_vector_main.c vx_vec.s
|
||||
|
||||
all: $(PROJECT).elf $(PROJECT).hex $(PROJECT).dump
|
||||
|
||||
$(PROJECT).dump: $(PROJECT).elf
|
||||
$(DP) -D $(PROJECT).elf > $(PROJECT).dump
|
||||
|
||||
$(PROJECT).hex: $(PROJECT).elf
|
||||
$(CP) -O ihex $(PROJECT).elf $(PROJECT).hex
|
||||
|
||||
$(PROJECT).elf: $(SRCS)
|
||||
$(CC) $(CFLAGS) $(SRCS) $(LDFLAGS) -o $(PROJECT).elf
|
||||
|
||||
.depend: $(SRCS)
|
||||
$(CC) $(CFLAGS) -MM $^ > .depend;
|
||||
|
||||
clean:
|
||||
rm -rf *.elf *.hex *.dump .depend
|
||||
@@ -1,13 +0,0 @@
|
||||
#pragma once
|
||||
|
||||
|
||||
#ifdef __cplusplus
|
||||
extern "C" {
|
||||
#endif
|
||||
|
||||
void vx_vec_test(int n, int* a, int* b, int* c); //vvaddint32
|
||||
//void vx_vec_test(int*);
|
||||
|
||||
#ifdef __cplusplus
|
||||
}
|
||||
#endif
|
||||
@@ -1,23 +0,0 @@
|
||||
.type vx_vec_test, @function
|
||||
.global vx_vec_test
|
||||
vx_vec_test:
|
||||
# vector-vector add routine of 32-bit integers
|
||||
# void vvaddint32(size_t n, const int*x, const int*y, int*z)
|
||||
# { for (size_t i=0; i<n; i++) { z[i]=x[i]+y[i]; } }
|
||||
#
|
||||
# a0 = n, a1 = x, a2 = y, a3 = z
|
||||
# Non-vector instructions are indented
|
||||
vsetvli t0, a0, e32 # Set vector length based on 32-bit vectors
|
||||
loop:
|
||||
vlw.v v0, (a1) # Get first vector
|
||||
sub a0, a0, t0 # Decrement number done
|
||||
slli t0, t0, 2 # Multiply number done by 4 bytes
|
||||
add a1, a1, t0 # Bump pointer
|
||||
vlw.v v1, (a2) # Get second vector
|
||||
add a2, a2, t0 # Bump pointer
|
||||
vadd.vv v2, v0, v1 # Sum vectors
|
||||
vsw.v v2, (a3) # Store result
|
||||
add a3, a3, t0 # Bump pointer
|
||||
bnez a0, loop # Loop back
|
||||
vmacc.vv v1, v2, v2
|
||||
ret # Finished
|
||||
@@ -1,22 +0,0 @@
|
||||
.type vx_vec_test, @function
|
||||
.global vx_vec_test
|
||||
vx_vec_test:
|
||||
# vector-vector add routine of 32-bit integers
|
||||
# void vvaddint32(size_t n, const int*x, const int*y, int*z)
|
||||
# { for (size_t i=0; i<n; i++) { z[i]=x[i]+y[i]; } }
|
||||
#
|
||||
# a0 = n, a1 = x, a2 = y, a3 = z
|
||||
# Non-vector instructions are indented
|
||||
vsetvli t0, a0, e32 # Set vector length based on 32-bit vectors
|
||||
loop:
|
||||
vlw.v v0, (a1) # Get first vector
|
||||
sub a0, a0, t0 # Decrement number done
|
||||
slli t0, t0, 2 # Multiply number done by 4 bytes
|
||||
add a1, a1, t0 # Bump pointer
|
||||
vlw.v v1, (a2) # Get second vector
|
||||
add a2, a2, t0 # Bump pointer
|
||||
vadd.vv v2, v0, v1 # Sum vectors
|
||||
vsw.v v2, (a3) # Store result
|
||||
add a3, a3, t0 # Bump pointer
|
||||
bnez a0, loop # Loop back
|
||||
ret # Finished
|
||||
@@ -1,24 +0,0 @@
|
||||
#include <vx_intrinsics.h>
|
||||
#include "vx_vec.h"
|
||||
|
||||
int main() {
|
||||
vx_tmc(1);
|
||||
|
||||
int n = 32;
|
||||
int *a = (int*)malloc(sizeof(int) * n); //{1, 1, 1, 1, 1};
|
||||
int *b = (int*)malloc(sizeof(int) * n); //{1, 1, 1, 1, 1};
|
||||
int *c = (int*)malloc(sizeof(int) * n); //{1, 1, 1, 1, 1};
|
||||
|
||||
for(int i = 0; i < n; ++i) {
|
||||
a[i] = 1;
|
||||
b[i] = 1;
|
||||
c[i] = 1;
|
||||
}
|
||||
|
||||
vx_vec_test(n, a, b, c);
|
||||
for (int i = 0; i < n; ++i) {
|
||||
vx_printf("%d", c[i]);
|
||||
}
|
||||
|
||||
vx_tmc(0);
|
||||
}
|
||||
Reference in New Issue
Block a user