static gpudata *cuda_transfer(gpudata *src, size_t offset, size_t sz,
void *dst_c, int may_share) {
cuda_context *ctx = src->ctx;
cuda_context *dst_ctx = (cuda_context *)dst_c;
gpudata *dst;
ASSERT_BUF(src);
ASSERT_CTX(ctx);
ASSERT_CTX(dst_ctx);
if (ctx == dst_ctx) {
if (may_share && offset == 0) {
cuda_retain(src);
return src;
}
dst = cuda_alloc(ctx, sz, NULL, 0, NULL);
if (dst == NULL) return NULL;
cuda_enter(ctx);
cuda_wait(src, CUDA_WAIT_READ);
cuda_wait(dst, CUDA_WAIT_WRITE);
ctx->err = cuMemcpyDtoDAsync(dst->ptr, src->ptr+offset, sz, ctx->s);
if (ctx->err != CUDA_SUCCESS) {
cuda_exit(ctx);
cuda_free(dst);
return NULL;
}
cuda_record(src, CUDA_WAIT_READ);
cuda_record(dst, CUDA_WAIT_WRITE);
cuda_exit(ctx);
return dst;
}
dst = cuda_alloc(dst_ctx, sz, NULL, 0, NULL);
if (dst == NULL)
return NULL;
cuda_enter(ctx);
cuda_waits(src, CUDA_WAIT_READ, dst_ctx->mem_s);
cuda_waits(dst, CUDA_WAIT_WRITE, dst_ctx->mem_s);
ctx->err = cuMemcpyPeerAsync(dst->ptr, dst->ctx->ctx, src->ptr+offset,
src->ctx->ctx, sz, dst_ctx->mem_s);
if (ctx->err != CUDA_SUCCESS) {
cuda_free(dst);
cuda_exit(ctx);
return NULL;
}
cuda_records(dst, CUDA_WAIT_WRITE, dst_ctx->mem_s);
cuda_records(src, CUDA_WAIT_READ, dst_ctx->mem_s);
cuda_exit(ctx);
return dst;
}
SEXP R_auto_cuMemcpyPeerAsync(SEXP r_dstDevice, SEXP r_dstContext, SEXP r_srcDevice, SEXP r_srcContext, SEXP r_ByteCount, SEXP r_hStream)
{
SEXP r_ans = R_NilValue;
CUdeviceptr dstDevice = REAL(r_dstDevice)[0];
CUcontext dstContext = (CUcontext) getRReference(r_dstContext);
CUdeviceptr srcDevice = REAL(r_srcDevice)[0];
CUcontext srcContext = (CUcontext) getRReference(r_srcContext);
size_t ByteCount = REAL(r_ByteCount)[0];
CUstream hStream = (CUstream) getRReference(r_hStream);
CUresult ans;
ans = cuMemcpyPeerAsync(dstDevice, dstContext, srcDevice, srcContext, ByteCount, hStream);
r_ans = Renum_convert_CUresult(ans) ;
return(r_ans);
}
