代码拉取完成，页面将自动刷新
// Author: Xiaohui Duan
// A set of Macros with DMA and athread_get/put implementation.
// #define DMA_FAST to use DMA implementation, by default, athread
// implementation is used.
// PLEASE DO NOT USE DMA_FAST at begining.
// DMA_FAST can be tried when athread_get/put implementation can
// produce excepted results.
#ifndef DMA_MACROS_H
#define DMA_MACROS_H
#include <slave.h>
#ifdef DMA_FAST
#include <simd.h>
#include <dma.h>
// A magic macro that pretends that reply depends on dma instruction.
// Thus compiler will do less evil.
#ifndef DMA_COUNT_BYTES
#define dma_rpl(desc, mem, ldm, reply) \
asm("dma %0, %1, %2\n\t" \
: : "r"(desc), "r"(mem), "r"(ldm), "r"(&reply) \
: "memory")
#else
#define dma_rpl(desc, mem, ldm, reply) { \
long t0, t1, t2; \
asm("dma %[DESC], %[MEM], %[LDM]\n\t" \
"zap %[DESC], 0xf8, %[DSIZE]\n\t" \
"addl %[MEM], %[DSIZE], %[CNT]\n\t" \
"addl %[CNT], 0x7f, %[CNT]\n\t" \
"bic %[CNT], 0x7f, %[CNT]\n\t" \
"bic %[MEM], 0x7f, %[MPTR]\n\t" \
"subl %[CNT], %[MPTR], %[DSIZE]\n\t" \
"srl %[DESC], 56, %[MPTR]\n\t" \
"and %[MPTR], 0x7, %[MPTR]\n\t" \
"s8addl %[MPTR], %[CNTS], %[MPTR]\n\t" \
"ldl %[CNT], 64(%[MPTR])\n\t" \
"addl %[CNT], %[DSIZE], %[CNT]\n\t" \
"stl %[CNT], 64(%[MPTR])\n\t" \
"zap %[DESC], 0xf8, %[DSIZE]\n\t" \
"ldl %[CNT], 0(%[MPTR])\n\t" \
"addl %[DSIZE], %[CNT], %[CNT]\n\t" \
"stl %[CNT], 0(%[MPTR])\n\t" \
: [DSIZE]"=&r"(t0), [MPTR]"=&r"(t1), [CNT]"=&r"(t2) \
: [CNTS]"r"(bytecount_shadow), [DESC]"r"(desc), [MEM]"r"(mem), [LDM]"r"(ldm), [REPLY]"r"(&reply) \
: "memory"); \
}
#endif
#define dma_bytecount(op) bytecount_shadow[op]
#define dma_bytecount_aligned(op) bytecount_shadow[op + 8]
// Pretending to be initializing, actually, it is defining descriptors,
// reply and count variable.
#define dma_init() \
volatile int reply_shadow = 0; \
int count_shadow = 0; \
long bytecount_shadow[16]; \
bytecount_shadow[0] = bytecount_shadow[1] = 0; \
bytecount_shadow[8] = bytecount_shadow[9] = 0; \
dma_desc pe_get_desc = 0, pe_put_desc = 0; \
dma_desc row_get_desc = 0, row_put_desc = 0; \
dma_desc rank_get_desc = 0, rank_put_desc = 0; \
dma_desc bcast_get_desc = 0, brow_get_desc = 0; \
dma_set_op(&pe_get_desc, DMA_GET); \
dma_set_op(&pe_put_desc, DMA_PUT); \
dma_set_op(&row_get_desc, DMA_GET); \
dma_set_op(&row_put_desc, DMA_PUT); \
dma_set_op(&rank_get_desc, DMA_GET); \
dma_set_op(&rank_put_desc, DMA_PUT); \
dma_set_op(&bcast_get_desc, DMA_GET); \
dma_set_op(&brow_get_desc, DMA_GET); \
\
dma_set_mode(&pe_get_desc, PE_MODE); \
dma_set_mode(&pe_put_desc, PE_MODE); \
dma_set_mode(&row_get_desc, ROW_MODE); \
dma_set_mode(&row_put_desc, ROW_MODE); \
dma_set_mode(&rank_get_desc, RANK_MODE); \
dma_set_mode(&rank_put_desc, RANK_MODE); \
dma_set_mode(&bcast_get_desc, BCAST_MODE); \
dma_set_mode(&brow_get_desc, BROW_MODE); \
\
dma_set_reply(&pe_get_desc, &reply_shadow); \
dma_set_reply(&pe_put_desc, &reply_shadow); \
dma_set_reply(&row_get_desc, &reply_shadow); \
dma_set_reply(&row_put_desc, &reply_shadow); \
dma_set_reply(&rank_get_desc, &reply_shadow); \
dma_set_reply(&rank_put_desc, &reply_shadow); \
dma_set_reply(&bcast_get_desc, &reply_shadow); \
dma_set_reply(&brow_get_desc, &reply_shadow); \
\
dma_set_mask(&bcast_get_desc, 0xff); \
dma_set_mask(&brow_get_desc, 0xff); \
// Set stride and bsize for strided get and rank mode.
#define dma_set_stride(mode, stride, bsize) { \
dma_set_stepsize(&(mode ## _desc), (long)(stride)); \
dma_set_bsize(&(mode ## _desc), (long)(bsize)); \
} \
// Set mask for brow and bcast mode
#define bcast_set_mask(mode, mask){ \
dma_set_mask(&(mode ## _desc), (long)(mask)); \
}
// PE_MODE, DMA_GET, stride and bsize can be set by dma_set_stride
#define pe_get(mem, ldm, size) { \
dma_set_size(&pe_get_desc, (long)(size)); \
dma_rpl(pe_get_desc, (mem), (ldm), reply_shadow); \
count_shadow ++; \
}
// PE_MODE, DMA_PUT, stride and bsize can be set by dma_set_stride
#define pe_put(mem, ldm, size) { \
dma_set_size(&pe_put_desc, (long)(size)); \
dma_rpl(pe_put_desc, (mem), (ldm), reply_shadow); \
count_shadow ++; \
}
// ROW_MODE, DMA_GET, stride and bsize can be set by dma_set_stride
#define row_get(mem, ldm, size) { \
dma_set_size(&row_get_desc, (long)(size)); \
dma_rpl(row_get_desc, (mem), (ldm), reply_shadow); \
count_shadow ++; \
}
// ROW_MODE, DMA_PUT, stride and bsize can be set by dma_set_stride
#define row_put(mem, ldm, size) { \
dma_set_size(&row_put_desc, (long)(size)); \
dma_rpl(row_put_desc, (mem), (ldm), reply_shadow); \
count_shadow ++; \
}
// RANK_MODE, DMA_GET, bsize can be set by dma_set_stride
#define rank_get(mem, ldm, size) { \
dma_set_size(&rank_get_desc, (long)(size)); \
dma_rpl(rank_get_desc, (mem), (ldm), reply_shadow); \
count_shadow ++; \
}
// RANK_MODE, DMA_PUT, stride and bsize can be set by dma_set_stride
#define rank_put(mem, ldm, size) { \
dma_set_size(&rank_put_desc, (long)(size)); \
dma_rpl(rank_put_desc, (mem), (ldm), reply_shadow); \
count_shadow ++; \
}
// BCAST_MODE, DMA_GET, stride and bsize can be set by dma_set_stride,
// broadcasting mask can be set by bcast_set_mask
#define bcast_get(mem, ldm, size) { \
dma_set_size(&bcast_get_desc, (long)(size)); \
dma_rpl(bcast_get_desc, (mem), (ldm), reply_shadow); \
count_shadow ++; \
}
// BROW_MODE, DMA_GET, stride and bsize can be set by dma_set_stride,
// broadcasting mask can be set by bcast_set_mask
#define brow_get(mem, ldm, size) { \
dma_set_size(&brow_get_desc, (long)(size)); \
dma_rpl(brow_get_desc, (mem), (ldm), reply_shadow); \
count_shadow ++; \
}
#else
#ifndef DMA_COUNT_BYTES
#define __athread_get athread_get
#define __athread_put athread_put
#else
#define __athread_get(mode, src, dst, size, reply, mask, stride, bsize) { \
bytecount_shadow[1] += size; \
athread_get(mode, src, dst, size, reply, mask, stride, bsize); \
}
#define __athread_put(mode, src, dst, size, reply, stride, bsize) { \
bytecount_shadow[0] += size; \
athread_put(mode, src, dst, size, reply, stride, bsize); \
}
#endif
// Pretending to be initializing, actually, it is defining reply, count.
// No descriptor required, setting stride, bsize and mask are done by
// setting variables.
// REMINDER: please clear masks, strides and bsizes after using strided or
// rank mode, and, currently it is not so compatible with DMA_FAST since
// all modes share a same set of masks.
#define dma_init() \
volatile int reply_shadow = 0; \
int count_shadow = 0; \
long bytecount_shadow[16]; \
bytecount_shadow[0] = bytecount_shadow[1] = 0; \
bytecount_shadow[8] = bytecount_shadow[9] = 0; \
long pe_get_stride_shadow = 0; \
long pe_put_stride_shadow = 0; \
long row_get_stride_shadow = 0; \
long row_put_stride_shadow = 0; \
long rank_get_stride_shadow = 0; \
long rank_put_stride_shadow = 0; \
long brow_get_stride_shadow = 0; \
long bcast_get_stride_shadow = 0; \
\
long pe_get_bsize_shadow = 0; \
long pe_put_bsize_shadow = 0; \
long row_get_bsize_shadow = 0; \
long row_put_bsize_shadow = 0; \
long rank_get_bsize_shadow = 0; \
long rank_put_bsize_shadow = 0; \
long brow_get_bsize_shadow = 0; \
long bcast_get_bsize_shadow = 0; \
\
int bcast_get_mask_shadow = 0xff; \
int brow_get_mask_shadow = 0xff; \
// The comments for the following macros can be found in DMA_FAST mode.
#define dma_set_stride(mode, stride, bsize) { \
mode ## _stride_shadow = (long)(stride); \
mode ## _bsize_shadow = (long)(bsize); \
}
#define bcast_set_mask(mode, mask) { \
mode ## _mask_shadow = (int)(mask); \
}
#define pe_get(mem, ldm, size) { \
__athread_get(PE_MODE, (mem), (ldm), (size), (void*)&reply_shadow \
, 0, pe_get_stride_shadow, pe_get_bsize_shadow); \
count_shadow ++; \
}
#define row_get(mem, ldm, size) { \
__athread_get(ROW_MODE, (mem), (ldm), (size), (void *)&reply_shadow \
, 0, row_get_stride_shadow, row_get_bsize_shadow); \
count_shadow ++; \
}
#define rank_get(mem, ldm, size) { \
__athread_get(RANK_MODE, (mem), (ldm), (size), (void *)&reply_shadow \
, 0, rank_get_stride_shadow, rank_get_bsize_shadow); \
count_shadow ++; \
}
#define bcast_get(mem, ldm, size) { \
__athread_get(BCAST_MODE, (mem), (ldm), (size), (void*)&reply_shadow \
, bcast_get_mask_shadow, bcast_get_stride_shadow \
, bcast_get_bsize_shadow); \
count_shadow ++; \
}
#define brow_get(mem, ldm, size) { \
__athread_get(BROW_MODE, (mem), (ldm), (size), (void *)&reply_shadow \
, brow_get_mask_shadow, brow_get_stride_shadow \
, brow_get_bsize_shadow); \
count_shadow ++; \
}
#define pe_put(mem, ldm, size) { \
__athread_put(PE_MODE, (ldm), (mem), (size), (void*)&reply_shadow \
, pe_put_stride_shadow, pe_put_bsize_shadow); \
count_shadow ++; \
}
#define row_put(mem, ldm, size) { \
__athread_put(ROW_MODE, (ldm), (mem), (size), (void *)&reply_shadow \
, row_put_stride_shadow, row_put_bsize_shadow); \
count_shadow ++; \
}
#define rank_put(mem, ldm, size) { \
__athread_put(RANK_MODE, (ldm), (mem), (size), (void *)&reply_shadow \
, rank_put_stride_shadow, rank_put_bsize_shadow); \
count_shadow ++; \
}
#endif
#define bcast_reset_mask(mode) bcast_set_mask(mode, 0xff)
#define dma_reset_stride(mode) dma_set_stride(mode, 0, 0)
// Sync the reply and count. "memb" is always a good choice to avoid
// god's punishment.
#define dma_syn() { \
while (reply_shadow != count_shadow) { \
}; \
asm volatile("memb\n\t"); \
}
#endif
此处可能存在不合适展示的内容，页面不予展示。您可通过相关编辑功能自查并修改。
如您确认内容无涉及 不当用语 / 纯广告导流 / 暴力 / 低俗色情 / 侵权 / 盗版 / 虚假 / 无价值内容或违法国家有关法律法规的内容，可点击提交进行申诉，我们将尽快为您处理。
马建仓 AI 助手
