WEAK void halide_dev_malloc(void *user_context, buffer_t* buf) {
if (buf->dev) {
// This buffer already has a device allocation
return;
}
size_t size = __buf_size(user_context, buf);
#ifdef DEBUG
halide_printf(user_context, "dev_malloc allocating buffer of %zd bytes, "
"extents: %zdx%zdx%zdx%zd strides: %zdx%zdx%zdx%zd (%d bytes per element)\n",
size, buf->extent[0], buf->extent[1], buf->extent[2], buf->extent[3],
buf->stride[0], buf->stride[1], buf->stride[2], buf->stride[3],
buf->elem_size);
#endif
CUdeviceptr p;
TIME_CALL( cuMemAlloc(&p, size), "dev_malloc");
buf->dev = (uint64_t)p;
halide_assert(user_context, buf->dev);
#ifdef DEBUG
halide_assert(user_context, halide_validate_dev_pointer(user_context, buf));
#endif
}
WEAK int halide_dev_free(void *user_context, buffer_t* buf) {
DEBUG_PRINTF( user_context, "CL: halide_dev_free (user_context: %p, buf: %p)\n", user_context, buf );
ClContext ctx(user_context);
// halide_dev_free, at present, can be exposed to clients and they
// should be allowed to call halide_dev_free on any buffer_t
// including ones that have never been used with a GPU.
if (buf->dev == 0) {
return 0;
}
#ifdef DEBUG
uint64_t t_before = halide_current_time_ns(user_context);
#endif
halide_assert(user_context, halide_validate_dev_pointer(user_context, buf));
DEBUG_PRINTF(user_context, " clReleaseMemObject %p\n", (cl_mem)buf->dev );
cl_int result = clReleaseMemObject((cl_mem)buf->dev);
// If clReleaseMemObject fails, it is unlikely to succeed in a later call, so
// we just end our reference to it regardless.
buf->dev = 0;
if (result != CL_SUCCESS) {
halide_error_varargs(user_context, "CL: clReleaseMemObject failed (%d)", result);
return result;
}
#ifdef DEBUG
uint64_t t_after = halide_current_time_ns(user_context);
halide_printf(user_context, " Time: %f ms\n", (t_after - t_before) / 1.0e6);
#endif
return 0;
}
WEAK int halide_copy_to_host(void *user_context, buffer_t* buf) {
if (!buf->dev_dirty) {
return 0;
}
DEBUG_PRINTF( user_context, "CUDA: halide_copy_to_host (user_context: %p, buf: %p)\n", user_context, buf );
CudaContext ctx(user_context);
if (ctx.error != CUDA_SUCCESS) {
return ctx.error;
}
// Need to check dev_dirty again, in case another thread did the
// copy_to_host before the serialization point above.
if (buf->dev_dirty) {
#ifdef DEBUG
uint64_t t_before = halide_current_time_ns(user_context);
#endif
halide_assert(user_context, buf->dev && buf->dev);
halide_assert(user_context, halide_validate_dev_pointer(user_context, buf));
_dev_copy c = _make_dev_to_host_copy(buf);
for (int w = 0; w < c.extent[3]; w++) {
for (int z = 0; z < c.extent[2]; z++) {
for (int y = 0; y < c.extent[1]; y++) {
for (int x = 0; x < c.extent[0]; x++) {
uint64_t off = (x * c.stride_bytes[0] +
y * c.stride_bytes[1] +
z * c.stride_bytes[2] +
w * c.stride_bytes[3]);
CUdeviceptr src = (CUdeviceptr)(c.src + off);
void *dst = (void *)(c.dst + off);
uint64_t size = c.chunk_size;
DEBUG_PRINTF( user_context, " cuMemcpyDtoH (%d, %d, %d, %d), %p -> %p, %lld bytes\n",
x, y, z, w,
(void *)src, dst, (long long)size );
CUresult err = cuMemcpyDtoH(dst, src, size);
if (err != CUDA_SUCCESS) {
halide_error_varargs(user_context, "CUDA: cuMemcpyDtoH failed (%s)",
_get_error_name(err));
return err;
}
}
}
}
}
#ifdef DEBUG
uint64_t t_after = halide_current_time_ns(user_context);
halide_printf(user_context, " Time: %f ms\n", (t_after - t_before) / 1.0e6);
#endif
}
buf->dev_dirty = false;
return 0;
}
WEAK void halide_dev_free(buffer_t* buf) {
#ifndef NDEBUG
fprintf(stderr, "In dev_free of %p - dev: 0x%zx\n", buf, buf->dev);
#endif
assert(halide_validate_dev_pointer(buf));
CHECK_CALL( cuMemFree(buf->dev), "cuMemFree" );
buf->dev = 0;
}
WEAK void halide_dev_free(buffer_t* buf) {
#ifdef DEBUG
halide_printf("In dev_free of %p - dev: 0x%p\n", buf, (void*)buf->dev);
halide_assert(halide_validate_dev_pointer(buf));
#endif
CHECK_CALL( cuMemFree(buf->dev), "cuMemFree" );
buf->dev = 0;
}
WEAK int halide_dev_malloc(void *user_context, buffer_t *buf) {
DEBUG_PRINTF( user_context, "CUDA: halide_dev_malloc (user_context: %p, buf: %p)\n", user_context, buf );
CudaContext ctx(user_context);
if (ctx.error != CUDA_SUCCESS) {
return ctx.error;
}
size_t size = _buf_size(user_context, buf);
if (buf->dev) {
// This buffer already has a device allocation
halide_assert(user_context, halide_validate_dev_pointer(user_context, buf, size));
return 0;
}
halide_assert(user_context, buf->stride[0] >= 0 && buf->stride[1] >= 0 &&
buf->stride[2] >= 0 && buf->stride[3] >= 0);
DEBUG_PRINTF(user_context, " allocating buffer of %lld bytes, "
"extents: %lldx%lldx%lldx%lld strides: %lldx%lldx%lldx%lld (%d bytes per element)\n",
(long long)size,
(long long)buf->extent[0], (long long)buf->extent[1],
(long long)buf->extent[2], (long long)buf->extent[3],
(long long)buf->stride[0], (long long)buf->stride[1],
(long long)buf->stride[2], (long long)buf->stride[3],
buf->elem_size);
#ifdef DEBUG
uint64_t t_before = halide_current_time_ns(user_context);
#endif
CUdeviceptr p;
DEBUG_PRINTF( user_context, " cuMemAlloc %lld -> ", size );
CUresult err = cuMemAlloc(&p, size);
if (err != CUDA_SUCCESS) {
DEBUG_PRINTF( user_context, "%s\n", _get_error_name(err));
halide_error_varargs(user_context, "CUDA: cuMemAlloc failed (%s)",
_get_error_name(err));
return err;
} else {
DEBUG_PRINTF( user_context, "%p\n", p );
}
halide_assert(user_context, p);
buf->dev = (uint64_t)p;
#ifdef DEBUG
uint64_t t_after = halide_current_time_ns(user_context);
halide_printf(user_context, " Time: %f ms\n", (t_after - t_before) / 1.0e6);
#endif
return 0;
}
WEAK void halide_copy_to_dev(void *user_context, buffer_t* buf) {
if (buf->host_dirty) {
halide_assert(user_context, buf->host && buf->dev);
size_t size = __buf_size(user_context, buf);
#ifdef DEBUG
halide_printf(user_context, "copy_to_dev (%lld bytes) %p -> %p\n", (long long)size, buf->host, (void*)buf->dev);
#endif
halide_assert(user_context, halide_validate_dev_pointer(user_context, buf));
int err = clEnqueueWriteBuffer( *cl_q, (cl_mem)((void*)buf->dev), CL_TRUE, 0, size, buf->host, 0, NULL, NULL );
CHECK_ERR( err, "clEnqueueWriteBuffer" );
}
buf->host_dirty = false;
}
WEAK int halide_dev_malloc(void *user_context, buffer_t* buf) {
DEBUG_PRINTF( user_context, "CL: halide_dev_malloc (user_context: %p, buf: %p)\n", user_context, buf );
ClContext ctx(user_context);
if (ctx.error != CL_SUCCESS) {
return ctx.error;
}
size_t size = _buf_size(user_context, buf);
if (buf->dev) {
halide_assert(user_context, halide_validate_dev_pointer(user_context, buf, size));
return 0;
}
halide_assert(user_context, buf->stride[0] >= 0 && buf->stride[1] >= 0 &&
buf->stride[2] >= 0 && buf->stride[3] >= 0);
DEBUG_PRINTF(user_context, " Allocating buffer of %lld bytes, "
"extents: %lldx%lldx%lldx%lld strides: %lldx%lldx%lldx%lld (%d bytes per element)\n",
(long long)size,
(long long)buf->extent[0], (long long)buf->extent[1],
(long long)buf->extent[2], (long long)buf->extent[3],
(long long)buf->stride[0], (long long)buf->stride[1],
(long long)buf->stride[2], (long long)buf->stride[3],
buf->elem_size);
#ifdef DEBUG
uint64_t t_before = halide_current_time_ns(user_context);
#endif
cl_int err;
DEBUG_PRINTF( user_context, " clCreateBuffer -> ", size );
buf->dev = (uint64_t)clCreateBuffer(ctx.context, CL_MEM_READ_WRITE, size, NULL, &err);
if (err != CL_SUCCESS || buf->dev == 0) {
DEBUG_PRINTF( user_context, "%d\n", err);
halide_error_varargs(user_context, "CL: clCreateBuffer failed (%d)\n", err);
return err;
} else {
DEBUG_PRINTF( user_context, "%p\n", (cl_mem)buf->dev );
}
DEBUG_PRINTF(user_context, " Allocated device buffer %p for buffer %p\n",
(void *)buf->dev, buf);
#ifdef DEBUG
uint64_t t_after = halide_current_time_ns(user_context);
halide_printf(user_context, " Time: %f ms\n", (t_after - t_before) / 1.0e6);
#endif
return CL_SUCCESS;
}
WEAK void halide_copy_to_host(buffer_t* buf) {
if (buf->dev_dirty) {
halide_assert(buf->dev);
halide_assert(buf->host);
size_t size = buf_size(buf);
#ifdef DEBUG
char msg[256];
snprintf(msg, 256, "copy_to_host (%zu bytes) %p -> %p", size, (void*)buf->dev, buf->host );
halide_assert(halide_validate_dev_pointer(buf));
#endif
TIME_CALL( cuMemcpyDtoH(buf->host, buf->dev, size), msg );
}
buf->dev_dirty = false;
}
WEAK void halide_copy_to_dev(buffer_t* buf) {
if (buf->host_dirty) {
halide_assert(buf->host && buf->dev);
size_t size = buf_size(buf);
#ifdef DEBUG
char msg[256];
snprintf(msg, 256, "copy_to_dev (%zu bytes) %p -> %p (t=%lld)",
size, buf->host, (void*)buf->dev, (long long)halide_current_time_ns() );
halide_assert(halide_validate_dev_pointer(buf));
#endif
TIME_CALL( cuMemcpyHtoD(buf->dev, buf->host, size), msg );
}
buf->host_dirty = false;
}
WEAK void halide_dev_free(buffer_t* buf) {
// halide_dev_free, at present, can be exposed to clients and they
// should be allowed to call halide_dev_free on any buffer_t
// including ones that have never been used with a GPU.
if (buf->dev == 0)
return;
#ifdef DEBUG
halide_printf("In dev_free of %p - dev: 0x%p\n", buf, (void*)buf->dev);
#endif
halide_assert(halide_validate_dev_pointer(buf));
CHECK_CALL( clReleaseMemObject((cl_mem)buf->dev), "clReleaseMemObject" );
buf->dev = 0;
}
WEAK void halide_copy_to_host(void *user_context, buffer_t* buf) {
if (buf->dev_dirty) {
clFinish(*cl_q); // block on completion before read back
halide_assert(user_context, buf->host && buf->dev);
size_t size = __buf_size(user_context, buf);
#ifdef DEBUG
halide_printf(user_context, "copy_to_host buf %p (%lld bytes) %p -> %p\n", buf, (long long)size,
(void*)buf->dev, buf->host );
#endif
halide_assert(user_context, halide_validate_dev_pointer(user_context, buf, size));
int err = clEnqueueReadBuffer( *cl_q, (cl_mem)((void*)buf->dev), CL_TRUE, 0, size, buf->host, 0, NULL, NULL );
CHECK_ERR( err, "clEnqueueReadBuffer" );
}
buf->dev_dirty = false;
}
WEAK void halide_dev_free(void *user_context, buffer_t* buf) {
// halide_dev_free, at present, can be exposed to clients and they
// should be allowed to call halide_dev_free on any buffer_t
// including ones that have never been used with a GPU.
if (buf->dev == 0)
return;
#ifdef DEBUG
halide_printf(user_context, "In dev_free of %p - dev: 0x%p\n", buf, (void*)buf->dev);
halide_assert(user_context, halide_validate_dev_pointer(user_context, buf));
#endif
CHECK_CALL( cuMemFree(buf->dev), "cuMemFree" );
buf->dev = 0;
}
WEAK void halide_copy_to_dev(buffer_t* buf) {
if (buf->host_dirty) {
halide_assert(buf->host && buf->dev);
size_t size = __buf_size(buf);
#ifdef DEBUG
char msg[256];
snprintf(msg, 256, "copy_to_dev (%lld bytes) %p -> %p (t=%lld)",
(long long)size, buf->host, (void*)buf->dev, (long long)halide_current_time_ns() );
#endif
halide_assert(halide_validate_dev_pointer(buf));
TIME_START();
int err = clEnqueueWriteBuffer( cl_q, (cl_mem)((void*)buf->dev), CL_TRUE, 0, size, buf->host, 0, NULL, NULL );
CHECK_ERR( err, msg );
TIME_CHECK(msg);
}
buf->host_dirty = false;
}
WEAK void halide_dev_malloc(buffer_t* buf) {
if (buf->dev) {
halide_assert(halide_validate_dev_pointer(buf));
return;
}
size_t size = __buf_size(buf);
#ifdef DEBUG
halide_printf("dev_malloc allocating buffer of %zd bytes, extents: %zdx%zdx%zdx%zd strides: %zdx%zdx%zdx%zd (%d bytes per element)\n",
size, buf->extent[0], buf->extent[1], buf->extent[2], buf->extent[3],
buf->stride[0], buf->stride[1], buf->stride[2], buf->stride[3],
buf->elem_size);
#endif
buf->dev = (uint64_t)__dev_malloc(size);
halide_assert(buf->dev);
}
WEAK void halide_dev_malloc(buffer_t* buf) {
if (buf->dev) {
// This buffer already has a device allocation
return;
}
#ifndef NDEBUG
fprintf(stderr, "dev_malloc of %zdx%zdx%zdx%zd (%zd bytes per element) (buf->dev = %p) buffer\n",
buf->extent[0], buf->extent[1], buf->extent[2], buf->extent[3], buf->elem_size, (void*)buf->dev);
#endif
CUdeviceptr p;
TIME_CALL( cuMemAlloc(&p, buf_size(buf)), "dev_malloc");
buf->dev = (uint64_t)p;
assert(buf->dev);
#ifndef NDEBUG
assert(halide_validate_dev_pointer(buf));
#endif
}
WEAK void halide_copy_to_host(buffer_t* buf) {
if (buf->dev_dirty) {
clFinish(cl_q); // block on completion before read back
halide_assert(buf->host && buf->dev);
size_t size = __buf_size(buf);
#ifndef DEBUG
char msg[1] = { 0 };
#else
char msg[256];
snprintf(msg, 256, "copy_to_host (%lld bytes) %p -> %p", (long long)size, (void*)buf->dev, buf->host );
#endif
halide_assert(halide_validate_dev_pointer(buf, size));
TIME_START();
#ifdef DEBUG
halide_printf("%s\n", msg);
#endif
int err = clEnqueueReadBuffer( cl_q, (cl_mem)((void*)buf->dev), CL_TRUE, 0, size, buf->host, 0, NULL, NULL );
CHECK_ERR( err, msg );
TIME_CHECK(msg);
}
buf->dev_dirty = false;
}
WEAK int halide_dev_free(void *user_context, buffer_t* buf) {
// halide_dev_free, at present, can be exposed to clients and they
// should be allowed to call halide_dev_free on any buffer_t
// including ones that have never been used with a GPU.
if (buf->dev == 0) {
return 0;
}
DEBUG_PRINTF( user_context, "CUDA: halide_dev_free (user_context: %p, buf: %p)\n", user_context, buf );
CudaContext ctx(user_context);
if (ctx.error != CUDA_SUCCESS)
return ctx.error;
#ifdef DEBUG
uint64_t t_before = halide_current_time_ns(user_context);
#endif
halide_assert(user_context, halide_validate_dev_pointer(user_context, buf));
DEBUG_PRINTF( user_context, " cuMemFree %p\n", buf->dev );
CUresult err = cuMemFree(buf->dev);
// If cuMemFree fails, it isn't likely to succeed later, so just drop
// the reference.
buf->dev = 0;
if (err != CUDA_SUCCESS) {
halide_error_varargs(user_context, "CUDA: cuMemFree failed (%s)",
_get_error_name(err));
return err;
}
#ifdef DEBUG
uint64_t t_after = halide_current_time_ns(user_context);
halide_printf(user_context, " Time: %f ms\n", (t_after - t_before) / 1.0e6);
#endif
return 0;
}
WEAK void halide_dev_malloc(void *user_context, buffer_t* buf) {
if (buf->dev) {
halide_assert(user_context, halide_validate_dev_pointer(user_context, buf));
return;
}
size_t size = __buf_size(user_context, buf);
#ifdef DEBUG
halide_printf(user_context, "dev_malloc allocating buffer of %lld bytes, "
"extents: %lldx%lldx%lldx%lld strides: %lldx%lldx%lldx%lld (%d bytes per element)\n",
(long long)size, (long long)buf->extent[0], (long long)buf->extent[1],
(long long)buf->extent[2], (long long)buf->extent[3],
(long long)buf->stride[0], (long long)buf->stride[1],
(long long)buf->stride[2], (long long)buf->stride[3],
buf->elem_size);
#endif
buf->dev = (uint64_t)__dev_malloc(user_context, size);
#ifdef DEBUG
halide_printf(user_context, "dev_malloc allocated buffer %p of with buf->dev of %p\n",
buf, (void *)buf->dev);
#endif
halide_assert(user_context, buf->dev);
}
WEAK int halide_copy_to_host(void *user_context, buffer_t* buf) {
DEBUG_PRINTF(user_context, "CL: halide_copy_to_host (user_context: %p, buf: %p)\n", user_context, buf );
// Acquire the context so we can use the command queue. This also avoids multiple
// redundant calls to clEnqueueReadBuffer when multiple threads are trying to copy
// the same buffer.
ClContext ctx(user_context);
if (ctx.error != CL_SUCCESS) {
return ctx.error;
}
if (buf->dev_dirty) {
#ifdef DEBUG
uint64_t t_before = halide_current_time_ns(user_context);
#endif
halide_assert(user_context, buf->dev && buf->dev);
halide_assert(user_context, halide_validate_dev_pointer(user_context, buf));
_dev_copy c = _make_dev_to_host_copy(buf);
for (int w = 0; w < c.extent[3]; w++) {
for (int z = 0; z < c.extent[2]; z++) {
#ifdef ENABLE_OPENCL_11
// OpenCL 1.1 supports stride-aware memory transfers up to 3D, so we
// can deal with the 2 innermost strides with OpenCL.
uint64_t off = z * c.stride_bytes[2] + w * c.stride_bytes[3];
size_t offset[3] = { off, 0, 0 };
size_t region[3] = { c.chunk_size, c.extent[0], c.extent[1] };
DEBUG_PRINTF( user_context, " clEnqueueReadBufferRect ((%d, %d), (%p -> %p) + %d, %dx%dx%d bytes, %dx%d)\n",
z, w,
(void *)c.src, c.dst, (int)off,
(int)region[0], (int)region[1], (int)region[2],
(int)c.stride_bytes[0], (int)c.stride_bytes[1]);
cl_int err = clEnqueueReadBufferRect(ctx.cmd_queue, (cl_mem)c.src, CL_FALSE,
offset, offset, region,
c.stride_bytes[0], c.stride_bytes[1],
c.stride_bytes[0], c.stride_bytes[1],
(void *)c.dst,
0, NULL, NULL);
if (err != CL_SUCCESS) {
halide_error_varargs(user_context, "CL: clEnqueueReadBufferRect failed (%d)\n", err);
return err;
}
#else
for (int y = 0; y < c.extent[1]; y++) {
for (int x = 0; x < c.extent[0]; x++) {
uint64_t off = (x * c.stride_bytes[0] +
y * c.stride_bytes[1] +
z * c.stride_bytes[2] +
w * c.stride_bytes[3]);
void *src = (void *)(c.src + off);
void *dst = (void *)(c.dst + off);
uint64_t size = c.chunk_size;
DEBUG_PRINTF( user_context, " clEnqueueReadBuffer ((%d, %d, %d, %d), %lld bytes, %p -> %p)\n",
x, y, z, w,
(long long)size, (void *)src, dst );
cl_int err = clEnqueueReadBuffer(ctx.cmd_queue, (cl_mem)c.src,
CL_FALSE, off, size, dst, 0, NULL, NULL);
if (err != CL_SUCCESS) {
halide_error_varargs(user_context, "CL: clEnqueueReadBuffer failed (%d)\n", err);
return err;
}
}
}
#endif
}
}
// The writes above are all non-blocking, so empty the command
// queue before we proceed so that other host code won't read
// bad data.
clFinish(ctx.cmd_queue);
#ifdef DEBUG
uint64_t t_after = halide_current_time_ns(user_context);
halide_printf(user_context, " Time: %f ms\n", (t_after - t_before) / 1.0e6);
#endif
}
buf->dev_dirty = false;
return 0;
}
