[v4l-utils PATCH v2 0/4] Add fwht stateless decoder

[v4l-utils PATCH v2 0/4] Add fwht stateless decoder

[Dafna Hirschfeld]

Changes from v1:

- patches 1,2 did not change

- patches 3,4 changed, main change was removing the header
from the compressed frame buffer.

Dafna Hirschfeld (4):
  v4l2-ctl: move stateful m2m decode code to a separate function
  (c)v4l-helpers.h: Add support for the request api
  v4l2-ctl: Add functions and variables to support fwht stateless
    decoder
  v4l2-ctl: Add implementation for the stateless fwht decoder.

 utils/common/cv4l-helpers.h           |   5 +
 utils/common/v4l-helpers.h            |  22 ++
 utils/v4l2-ctl/v4l2-ctl-streaming.cpp | 449 +++++++++++++++++++++++---
 3 files changed, 429 insertions(+), 47 deletions(-)

-- 
2.17.1

[v4l-utils PATCH v2 1/4] v4l2-ctl: move stateful m2m decode code to a separate function

[Dafna Hirschfeld]

Add the function stateful_m2m that implements the stateful
codec api.
This is a preparation for having both stateful and stateless
implementations

Signed-off-by: Dafna Hirschfeld <dafna3@gmail.com>
---
 utils/v4l2-ctl/v4l2-ctl-streaming.cpp | 95 ++++++++++++++-------------
 1 file changed, 50 insertions(+), 45 deletions(-)

diff --git a/utils/v4l2-ctl/v4l2-ctl-streaming.cpp b/utils/v4l2-ctl/v4l2-ctl-streaming.cpp
index 766872b5..40ddc1c3 100644
--- a/utils/v4l2-ctl/v4l2-ctl-streaming.cpp
+++ b/utils/v4l2-ctl/v4l2-ctl-streaming.cpp
@@ -1933,16 +1933,12 @@ static int capture_setup(cv4l_fd &fd, cv4l_queue &in, cv4l_fd *exp_fd)
 	return 0;
 }
 
-static void streaming_set_m2m(cv4l_fd &fd, cv4l_fd &exp_fd)
+static void stateful_m2m(cv4l_fd &fd, cv4l_queue &in, cv4l_queue &out,
+			 FILE *fin, FILE *fout, cv4l_fd *exp_fd_p)
 {
 	int fd_flags = fcntl(fd.g_fd(), F_GETFL);
-	cv4l_queue in(fd.g_type(), memory);
-	cv4l_queue out(v4l_type_invert(fd.g_type()), out_memory);
-	cv4l_queue exp_q(exp_fd.g_type(), V4L2_MEMORY_MMAP);
-	cv4l_fd *exp_fd_p = NULL;
 	fps_timestamps fps_ts[2];
 	unsigned count[2] = { 0, 0 };
-	FILE *file[2] = {NULL, NULL};
 	fd_set fds[3];
 	fd_set *rd_fds = &fds[0]; /* for capture */
 	fd_set *ex_fds = &fds[1]; /* for capture */
@@ -1953,19 +1949,6 @@ static void streaming_set_m2m(cv4l_fd &fd, cv4l_fd &exp_fd)
 	fd.g_fmt(fmt[OUT], out.g_type());
 	fd.g_fmt(fmt[CAP], in.g_type());
 
-	if (!fd.has_vid_m2m()) {
-		fprintf(stderr, "unsupported m2m stream type\n");
-		return;
-	}
-	if (options[OptStreamDmaBuf] && options[OptStreamOutDmaBuf]) {
-		fprintf(stderr, "--stream-dmabuf and --stream-out-dmabuf not supported for m2m devices\n");
-		return;
-	}
-	if ((options[OptStreamDmaBuf] || options[OptStreamOutDmaBuf]) && exp_fd.g_fd() < 0) {
-		fprintf(stderr, "--stream-dmabuf or --stream-out-dmabuf can only work in combination with --export-device\n");
-		return;
-	}
-
 	struct v4l2_event_subscription sub;
 
 	memset(&sub, 0, sizeof(sub));
@@ -1986,34 +1969,18 @@ static void streaming_set_m2m(cv4l_fd &fd, cv4l_fd &exp_fd)
 	sub.type = V4L2_EVENT_SOURCE_CHANGE;
 	bool have_source_change = !fd.subscribe_event(sub);
 
-	file[CAP] = open_output_file(fd);
-	file[OUT] = open_input_file(fd, out.g_type());
-
 	if (out.reqbufs(&fd, reqbufs_count_out))
-		goto done;
-
-	if (options[OptStreamDmaBuf]) {
-		if (exp_q.reqbufs(&exp_fd, reqbufs_count_cap))
-			goto done;
-		exp_fd_p = &exp_fd;
-	}
-
-	if (options[OptStreamOutDmaBuf]) {
-		if (exp_q.reqbufs(&exp_fd, reqbufs_count_out))
-			goto done;
-		if (out.export_bufs(&exp_fd, exp_fd.g_type()))
-			goto done;
-	}
+		return;
 
-	if (do_setup_out_buffers(fd, out, file[OUT], true))
-		goto done;
+	if (do_setup_out_buffers(fd, out, fout, true))
+		return;
 
 	if (fd.streamon(out.g_type()))
-		goto done;
+		return;
 
 	if (codec_type != DECODER || !have_source_change)
 		if (capture_setup(fd, in, exp_fd_p))
-			goto done;
+			return;
 
 	fps_ts[CAP].determine_field(fd.g_fd(), in.g_type());
 	fps_ts[OUT].determine_field(fd.g_fd(), out.g_type());
@@ -2049,15 +2016,15 @@ static void streaming_set_m2m(cv4l_fd &fd, cv4l_fd &exp_fd)
 				continue;
 			fprintf(stderr, "select error: %s\n",
 					strerror(errno));
-			goto done;
+			return;
 		}
 		if (r == 0) {
 			fprintf(stderr, "select timeout\n");
-			goto done;
+			return;
 		}
 
 		if (rd_fds && FD_ISSET(fd.g_fd(), rd_fds)) {
-			r = do_handle_cap(fd, in, file[CAP], NULL,
+			r = do_handle_cap(fd, in, fin, NULL,
 					  count[CAP], fps_ts[CAP], fmt[CAP]);
 			if (r < 0) {
 				rd_fds = NULL;
@@ -2069,7 +2036,7 @@ static void streaming_set_m2m(cv4l_fd &fd, cv4l_fd &exp_fd)
 		}
 
 		if (wr_fds && FD_ISSET(fd.g_fd(), wr_fds)) {
-			r = do_handle_out(fd, out, file[OUT], NULL,
+			r = do_handle_out(fd, out, fout, NULL,
 					  count[OUT], fps_ts[OUT], fmt[OUT]);
 			if (r < 0)  {
 				wr_fds = NULL;
@@ -2120,7 +2087,7 @@ static void streaming_set_m2m(cv4l_fd &fd, cv4l_fd &exp_fd)
 				in_source_change_event = false;
 				last_buffer = false;
 				if (capture_setup(fd, in, exp_fd_p))
-					goto done;
+					return;
 				fd.g_fmt(fmt[OUT], out.g_type());
 				fd.g_fmt(fmt[CAP], in.g_type());
 				cap_streaming = true;
@@ -2138,7 +2105,45 @@ static void streaming_set_m2m(cv4l_fd &fd, cv4l_fd &exp_fd)
 	in.free(&fd);
 	out.free(&fd);
 	tpg_free(&tpg);
+}
+
+static void streaming_set_m2m(cv4l_fd &fd, cv4l_fd &exp_fd)
+{
+	cv4l_queue in(fd.g_type(), memory);
+	cv4l_queue out(v4l_type_invert(fd.g_type()), out_memory);
+	cv4l_queue exp_q(exp_fd.g_type(), V4L2_MEMORY_MMAP);
+	cv4l_fd *exp_fd_p = NULL;
+	FILE *file[2] = {NULL, NULL};
+
+	if (!fd.has_vid_m2m()) {
+		fprintf(stderr, "unsupported m2m stream type\n");
+		return;
+	}
+	if (options[OptStreamDmaBuf] && options[OptStreamOutDmaBuf]) {
+		fprintf(stderr, "--stream-dmabuf and --stream-out-dmabuf not supported for m2m devices\n");
+		return;
+	}
+	if ((options[OptStreamDmaBuf] || options[OptStreamOutDmaBuf]) && exp_fd.g_fd() < 0) {
+		fprintf(stderr, "--stream-dmabuf or --stream-out-dmabuf can only work in combination with --export-device\n");
+		return;
+	}
 
+	file[CAP] = open_output_file(fd);
+	file[OUT] = open_input_file(fd, out.g_type());
+
+	if (options[OptStreamDmaBuf]) {
+		if (exp_q.reqbufs(&exp_fd, reqbufs_count_cap))
+			return;
+		exp_fd_p = &exp_fd;
+	}
+
+	if (options[OptStreamOutDmaBuf]) {
+		if (exp_q.reqbufs(&exp_fd, reqbufs_count_out))
+			return;
+		if (out.export_bufs(&exp_fd, exp_fd.g_type()))
+			return;
+	}
+	stateful_m2m(fd, in, out, file[CAP], file[OUT], exp_fd_p);
 done:
 	if (options[OptStreamDmaBuf] || options[OptStreamOutDmaBuf])
 		exp_q.close_exported_fds();
-- 
2.17.1

[v4l-utils PATCH v2 2/4] (c)v4l-helpers.h: Add support for the request api

[Dafna Hirschfeld]

Add an array of request file descriptors to v4l_queue
and add methods to allocate and get them.

Signed-off-by: Dafna Hirschfeld <dafna3@gmail.com>
---
 utils/common/cv4l-helpers.h |  5 +++++
 utils/common/v4l-helpers.h  | 22 ++++++++++++++++++++++
 2 files changed, 27 insertions(+)

diff --git a/utils/common/cv4l-helpers.h b/utils/common/cv4l-helpers.h
index 1cd2b6b2..551d4673 100644
--- a/utils/common/cv4l-helpers.h
+++ b/utils/common/cv4l-helpers.h
@@ -745,6 +745,7 @@ public:
 	unsigned g_capabilities() const { return v4l_queue_g_capabilities(this); }
 	unsigned g_length(unsigned plane) const { return v4l_queue_g_length(this, plane); }
 	unsigned g_mem_offset(unsigned index, unsigned plane) const { return v4l_queue_g_mem_offset(this, index, plane); }
+	unsigned g_req_fd(unsigned index) const { return v4l_queue_g_req_fd(this, index); }
 	void *g_mmapping(unsigned index, unsigned plane) const { return v4l_queue_g_mmapping(this, index, plane); }
 	void s_mmapping(unsigned index, unsigned plane, void *m) { v4l_queue_s_mmapping(this, index, plane, m); }
 	void *g_userptr(unsigned index, unsigned plane) const { return v4l_queue_g_userptr(this, index, plane); }
@@ -797,6 +798,10 @@ public:
 	{
 		return v4l_queue_export_bufs(fd->g_v4l_fd(), this, exp_type);
 	}
+	int alloc_req(int media_fd, unsigned index)
+	{
+		return v4l_queue_alloc_req(this, media_fd, index);
+	}
 	void close_exported_fds()
 	{
 		v4l_queue_close_exported_fds(this);
diff --git a/utils/common/v4l-helpers.h b/utils/common/v4l-helpers.h
index 59d8566a..daa49a1f 100644
--- a/utils/common/v4l-helpers.h
+++ b/utils/common/v4l-helpers.h
@@ -10,6 +10,7 @@
 #define _V4L_HELPERS_H_
 
 #include <linux/videodev2.h>
+#include <linux/media.h>
 #include <string.h>
 #include <stdlib.h>
 #include <stdio.h>
@@ -1414,6 +1415,7 @@ struct v4l_queue {
 	void *mmappings[VIDEO_MAX_FRAME][VIDEO_MAX_PLANES];
 	unsigned long userptrs[VIDEO_MAX_FRAME][VIDEO_MAX_PLANES];
 	int fds[VIDEO_MAX_FRAME][VIDEO_MAX_PLANES];
+	int req_fds[VIDEO_MAX_FRAME];
 };
 
 static inline void v4l_queue_init(struct v4l_queue *q,
@@ -1445,6 +1447,11 @@ static inline __u32 v4l_queue_g_mem_offset(const struct v4l_queue *q, unsigned i
 	return q->mem_offsets[index][plane];
 }
 
+static inline unsigned v4l_queue_g_req_fd(const struct v4l_queue *q, unsigned index)
+{
+	return q->req_fds[index];
+}
+
 static inline void v4l_queue_s_mmapping(struct v4l_queue *q, unsigned index, unsigned plane, void *m)
 {
 	q->mmappings[index][plane] = m;
@@ -1701,6 +1708,21 @@ static inline int v4l_queue_export_bufs(struct v4l_fd *f, struct v4l_queue *q,
 	return 0;
 }
 
+static inline int v4l_queue_alloc_req(struct v4l_queue *q, int media_fd, unsigned index)
+{
+	int rc = 0;
+
+	rc = ioctl(media_fd, MEDIA_IOC_REQUEST_ALLOC, &q->req_fds[index]);
+	if (rc < 0) {
+		fprintf(stderr, "Unable to allocate media request: %s\n",
+			strerror(errno));
+		return rc;
+	}
+
+	return 0;
+}
+
+
 static inline void v4l_queue_close_exported_fds(struct v4l_queue *q)
 {
 	unsigned b, p;
-- 
2.17.1

[v4l-utils PATCH v2 3/4] v4l2-ctl: Add functions and variables to support fwht stateless decoder

[Dafna Hirschfeld]

Add the variable 'last_fwht_bf_ts' and the array 'fwht_reqs' to
allow the fwht stateless decoder to maintain the requests.

Signed-off-by: Dafna Hirschfeld <dafna3@gmail.com>
---
 utils/v4l2-ctl/v4l2-ctl-streaming.cpp | 120 ++++++++++++++++++++++++++
 1 file changed, 120 insertions(+)

diff --git a/utils/v4l2-ctl/v4l2-ctl-streaming.cpp b/utils/v4l2-ctl/v4l2-ctl-streaming.cpp
index 40ddc1c3..ec43bd03 100644
--- a/utils/v4l2-ctl/v4l2-ctl-streaming.cpp
+++ b/utils/v4l2-ctl/v4l2-ctl-streaming.cpp
@@ -20,11 +20,17 @@
 
 #include "v4l2-ctl.h"
 #include "v4l-stream.h"
+#include <media-info.h>
 
 extern "C" {
 #include "v4l2-tpg.h"
 }
 
+#define VICODEC_CID_CUSTOM_BASE        (V4L2_CID_MPEG_BASE | 0xf000)
+#define VICODEC_CID_I_FRAME_QP         (VICODEC_CID_CUSTOM_BASE + 0)
+#define VICODEC_CID_P_FRAME_QP         (VICODEC_CID_CUSTOM_BASE + 1)
+#define VICODEC_CID_STATELESS_FWHT     (VICODEC_CID_CUSTOM_BASE + 2)
+
 static unsigned stream_count;
 static unsigned stream_skip;
 static __u32 memory = V4L2_MEMORY_MMAP;
@@ -80,6 +86,16 @@ static bool support_cap_compose;
 static bool support_out_crop;
 static bool in_source_change_event;
 
+static __u64 last_fwht_bf_ts;
+
+struct request_fwht {
+	int fd;
+	__u64 ts;
+	struct v4l2_ctrl_fwht_params params;
+};
+
+static request_fwht fwht_reqs[VIDEO_MAX_FRAME];
+
 #define TS_WINDOW 241
 #define FILE_HDR_ID			v4l2_fourcc('V', 'h', 'd', 'r')
 
@@ -420,6 +436,12 @@ static int get_out_crop_rect(cv4l_fd &fd)
 	return 0;
 }
 
+static __u64 get_ns_timestamp(cv4l_buffer &buf)
+{
+	const struct timeval tv = buf.g_timestamp();
+	return v4l2_timeval_to_ns(&tv);
+}
+
 static void set_time_stamp(cv4l_buffer &buf)
 {
 	if ((buf.g_flags() & V4L2_BUF_FLAG_TIMESTAMP_MASK) != V4L2_BUF_FLAG_TIMESTAMP_COPY)
@@ -749,6 +771,104 @@ void streaming_cmd(int ch, char *optarg)
 	}
 }
 
+/*
+ * Assume that the fwht stream is valid and that each
+ * frame starts right after the previous one.
+ */
+static void read_fwht_frame(cv4l_fmt &fmt, unsigned char *buf,
+			    FILE *fpointer, unsigned &sz,
+			    unsigned &len)
+{
+	struct fwht_cframe_hdr *h = (struct fwht_cframe_hdr *)buf;
+
+	len = sizeof(struct fwht_cframe_hdr);
+	sz = fread(buf, 1, sizeof(struct fwht_cframe_hdr), fpointer);
+	if (sz < sizeof(struct fwht_cframe_hdr))
+		return;
+
+	len += ntohl(h->size);
+	sz += fread(buf + sz, 1, ntohl(h->size), fpointer);
+}
+
+static void set_fwht_stateless_params(struct v4l2_ctrl_fwht_params &fwht_params,
+				      const struct fwht_cframe_hdr *hdr,
+				      __u64 last_bf_ts)
+{
+	fwht_params.backward_ref_ts = last_bf_ts;
+	fwht_params.version = ntohl(hdr->version);
+	fwht_params.width = ntohl(hdr->width);
+	fwht_params.height = ntohl(hdr->height);
+	fwht_params.flags = ntohl(hdr->flags);
+	fwht_params.colorspace = ntohl(hdr->colorspace);
+	fwht_params.xfer_func = ntohl(hdr->xfer_func);
+	fwht_params.ycbcr_enc = ntohl(hdr->ycbcr_enc);
+	fwht_params.quantization = ntohl(hdr->quantization);
+	fwht_params.comp_frame_size = ntohl(hdr->size);
+
+	if (!last_bf_ts)
+		fwht_params.flags |= FWHT_FL_I_FRAME;
+}
+
+static void set_fwht_req_by_idx(unsigned idx, struct fwht_cframe_hdr *hdr,
+				int req_fd, __u64 last_bf_ts, __u64 ts)
+{
+	struct v4l2_ctrl_fwht_params fwht_params;
+
+	set_fwht_stateless_params(fwht_params, hdr, last_bf_ts);
+
+	fwht_reqs[idx].fd = req_fd;
+	fwht_reqs[idx].ts = ts;
+	fwht_reqs[idx].params = fwht_params;
+}
+
+static int get_fwht_req_by_ts(__u64 ts)
+{
+	for (int idx = 0; idx < VIDEO_MAX_FRAME; idx++) {
+		if (fwht_reqs[idx].ts == ts)
+			return idx;
+	}
+	return -1;
+}
+
+static bool set_fwht_req_by_fd(struct fwht_cframe_hdr *hdr,
+			       int req_fd, __u64 last_bf_ts, __u64 ts)
+{
+	struct v4l2_ctrl_fwht_params fwht_params;
+
+	set_fwht_stateless_params(fwht_params, hdr, last_bf_ts);
+
+	for (int idx = 0; idx < VIDEO_MAX_FRAME; idx++) {
+		if (fwht_reqs[idx].fd == req_fd) {
+			fwht_reqs[idx].ts = ts;
+			fwht_reqs[idx].params = fwht_params;
+			return true;
+		}
+	}
+	return false;
+}
+
+static int set_fwht_ext_ctrl(cv4l_fd &fd, struct fwht_cframe_hdr *hdr,
+			     __u64 last_bf_ts, int req_fd)
+{
+	v4l2_ext_controls controls;
+	struct v4l2_ext_control control;
+	struct v4l2_ctrl_fwht_params fwht_params;
+
+	memset(&control, 0, sizeof(control));
+	memset(&controls, 0, sizeof(controls));
+
+	set_fwht_stateless_params(fwht_params, hdr, last_bf_ts);
+
+	control.id = VICODEC_CID_STATELESS_FWHT;
+	control.ptr = &fwht_params;
+	control.size = sizeof(fwht_params);
+	controls.which = V4L2_CTRL_WHICH_REQUEST_VAL;
+	controls.request_fd = req_fd;
+	controls.controls = &control;
+	controls.count = 1;
+	return fd.s_ext_ctrls(controls);
+}
+
 static void read_write_padded_frame(cv4l_fmt &fmt, unsigned char *buf,
 				    FILE *fpointer, unsigned &sz,
 				    unsigned &len, bool is_read)
-- 
2.17.1

[v4l-utils PATCH v2 4/4] v4l2-ctl: Add implementation for the stateless fwht decoder.

[Dafna Hirschfeld]

Add code to support the stateless decoder
and the function 'stateless_m2m' that implements it.

Signed-off-by: Dafna Hirschfeld <dafna3@gmail.com>
---
 utils/v4l2-ctl/v4l2-ctl-streaming.cpp | 236 +++++++++++++++++++++++++-
 1 file changed, 233 insertions(+), 3 deletions(-)

diff --git a/utils/v4l2-ctl/v4l2-ctl-streaming.cpp b/utils/v4l2-ctl/v4l2-ctl-streaming.cpp
index ec43bd03..2bdee280 100644
--- a/utils/v4l2-ctl/v4l2-ctl-streaming.cpp
+++ b/utils/v4l2-ctl/v4l2-ctl-streaming.cpp
@@ -1054,8 +1054,9 @@ restart:
 				return false;
 			}
 		}
-
-		if (support_out_crop && v4l2_fwht_find_pixfmt(fmt.g_pixelformat()))
+		if (q.g_capabilities() & V4L2_BUF_CAP_SUPPORTS_REQUESTS)
+			read_fwht_frame(fmt, (unsigned char *)buf, fin, sz, len);
+		else if (support_out_crop && v4l2_fwht_find_pixfmt(fmt.g_pixelformat()))
 			read_write_padded_frame(fmt, (unsigned char *)buf, fin, sz, len, true);
 		else
 			sz = fread(buf, 1, len, fin);
@@ -1082,6 +1083,21 @@ restart:
 	return true;
 }
 
+static bool split_fwht_frame(u8 *frame, struct fwht_cframe_hdr *hdr, unsigned max_len)
+{
+	unsigned int len;
+	bool ret = true;
+
+	memcpy(hdr, frame, sizeof(struct fwht_cframe_hdr));
+	len = ntohl(hdr->size);
+	if (len > max_len) {
+		len = max_len;
+		ret = false;
+	}
+	memmove(frame, frame + sizeof(struct fwht_cframe_hdr), len);
+	return ret;
+}
+
 static int do_setup_out_buffers(cv4l_fd &fd, cv4l_queue &q, FILE *fin, bool qbuf)
 {
 	tpg_pixel_aspect aspect = TPG_PIXEL_ASPECT_SQUARE;
@@ -1186,6 +1202,35 @@ static int do_setup_out_buffers(cv4l_fd &fd, cv4l_queue &q, FILE *fin, bool qbuf
 		if (fin && !fill_buffer_from_file(fd, q, buf, fmt, fin))
 			return -2;
 
+		struct fwht_cframe_hdr hdr;
+
+		if (q.g_capabilities() & V4L2_BUF_CAP_SUPPORTS_REQUESTS) {
+			int media_fd = mi_get_media_fd(fd.g_fd());
+
+			if (!split_fwht_frame((u8 *)q.g_dataptr(buf.g_index(), 0),
+					      &hdr, buf.g_length(0)))
+				fprintf(stderr, "%s: warning: size field in fwht header is larger than buf size\n",
+					__func__);
+
+			if (media_fd < 0) {
+				fprintf(stderr, "%s: mi_get_media_fd failed\n", __func__);
+				return media_fd;
+			}
+
+			if (q.alloc_req(media_fd, i)) {
+				fprintf(stderr, "%s: q.alloc_req failed\n", __func__);
+				return -1;
+			}
+			buf.s_request_fd(q.g_req_fd(i));
+			buf.or_flags(V4L2_BUF_FLAG_REQUEST_FD);
+
+			if (set_fwht_ext_ctrl(fd, &hdr, last_fwht_bf_ts,
+					      buf.g_request_fd())) {
+				fprintf(stderr, "%s: set_fwht_ext_ctrls failed on %dth buf: %s\n",
+					__func__, i, strerror(errno));
+				return -1;
+			}
+		}
 		if (qbuf) {
 			set_time_stamp(buf);
 			if (fd.qbuf(buf))
@@ -1195,6 +1240,16 @@ static int do_setup_out_buffers(cv4l_fd &fd, cv4l_queue &q, FILE *fin, bool qbuf
 				fprintf(stderr, ">");
 			fflush(stderr);
 		}
+		if (q.g_capabilities() & V4L2_BUF_CAP_SUPPORTS_REQUESTS) {
+			set_fwht_req_by_idx(i, &hdr, q.g_req_fd(i),
+					    last_fwht_bf_ts, get_ns_timestamp(buf));
+			last_fwht_bf_ts = get_ns_timestamp(buf);
+			if (ioctl(q.g_req_fd(i), MEDIA_REQUEST_IOC_QUEUE) < 0) {
+				fprintf(stderr, "Unable to queue media request: %s\n",
+					strerror(errno));
+				return -1;
+			}
+		}
 	}
 	if (qbuf)
 		output_field = field;
@@ -1433,12 +1488,48 @@ static int do_handle_out(cv4l_fd &fd, cv4l_queue &q, FILE *fin, cv4l_buffer *cap
 				       (u8 *)q.g_dataptr(buf.g_index(), j));
 	}
 
+	struct fwht_cframe_hdr hdr;
+
+	if (q.g_capabilities() & V4L2_BUF_CAP_SUPPORTS_REQUESTS) {
+		if (ioctl(buf.g_request_fd(), MEDIA_REQUEST_IOC_REINIT, NULL)) {
+			fprintf(stderr, "Unable to reinit media request: %s\n",
+				strerror(errno));
+			return -1;
+		}
+		if (!split_fwht_frame((u8 *)q.g_dataptr(buf.g_index(), 0), &hdr, buf.g_length(0)))
+			fprintf(stderr, "%s: warning: size field in fwht header is larger than buf size\n",
+					__func__);
+
+		if (set_fwht_ext_ctrl(fd, &hdr, last_fwht_bf_ts,
+				      buf.g_request_fd())) {
+			fprintf(stderr, "%s: set_fwht_ext_ctrls failed: %s\n",
+				__func__, strerror(errno));
+			return -1;
+		}
+	}
+
 	set_time_stamp(buf);
 
 	if (fd.qbuf(buf)) {
 		fprintf(stderr, "%s: failed: %s\n", "VIDIOC_QBUF", strerror(errno));
 		return -1;
 	}
+	if (q.g_capabilities() & V4L2_BUF_CAP_SUPPORTS_REQUESTS) {
+		if (!set_fwht_req_by_fd(&hdr, buf.g_request_fd(), last_fwht_bf_ts,
+					get_ns_timestamp(buf))) {
+			fprintf(stderr, "%s: request for fd %d does not exist\n",
+				__func__, buf.g_request_fd());
+			return -1;
+		}
+
+		last_fwht_bf_ts = get_ns_timestamp(buf);
+		if (ioctl(buf.g_request_fd(), MEDIA_REQUEST_IOC_QUEUE) < 0) {
+			fprintf(stderr, "Unable to queue media request: %s\n",
+				strerror(errno));
+			return -1;
+		}
+	}
+
 	tpg_update_mv_count(&tpg, V4L2_FIELD_HAS_T_OR_B(output_field));
 
 	if (!verbose)
@@ -2227,6 +2318,140 @@ static void stateful_m2m(cv4l_fd &fd, cv4l_queue &in, cv4l_queue &out,
 	tpg_free(&tpg);
 }
 
+static void stateless_m2m(cv4l_fd &fd, cv4l_queue &in, cv4l_queue &out,
+			  FILE *fin, FILE *fout, cv4l_fd *exp_fd_p)
+{
+	fps_timestamps fps_ts[2];
+	unsigned count[2] = { 0, 0 };
+	cv4l_fmt fmt[2];
+	int fd_flags = fcntl(fd.g_fd(), F_GETFL);
+
+	fd.g_fmt(fmt[OUT], out.g_type());
+	fd.g_fmt(fmt[CAP], in.g_type());
+
+	if (out.reqbufs(&fd, reqbufs_count_out)) {
+		fprintf(stderr, "%s: out.reqbufs failed\n", __func__);
+		return;
+	}
+
+	if (in.reqbufs(&fd, reqbufs_count_cap)) {
+		fprintf(stderr, "%s: in.reqbufs failed\n", __func__);
+		return;
+	}
+
+	if (exp_fd_p && in.export_bufs(exp_fd_p, exp_fd_p->g_type()))
+		return;
+
+	if (in.obtain_bufs(&fd)) {
+		fprintf(stderr, "%s: in.obtain_bufs error\n", __func__);
+		return;
+	}
+
+	if (do_setup_out_buffers(fd, out, fout, true)) {
+		fprintf(stderr, "%s: do_setup_out_buffers failed\n", __func__);
+		return;
+	}
+
+	if (in.queue_all(&fd)) {
+		fprintf(stderr, "%s: in.queue_all failed\n", __func__);
+		return;
+	}
+
+	if (fd.streamon(out.g_type())) {
+		fprintf(stderr, "%s: streamon for out failed\n", __func__);
+		return;
+	}
+
+	if (fd.streamon(in.g_type())) {
+		fprintf(stderr, "%s: streamon for in failed\n", __func__);
+		return;
+	}
+	int index = 0;
+	bool queue_lst_buf = false;
+	cv4l_buffer last_in_buf;
+
+	fcntl(fd.g_fd(), F_SETFL, fd_flags | O_NONBLOCK);
+
+	while (true) {
+		fd_set except_fds;
+		int req_fd = out.g_req_fd(index);
+		struct timeval tv = { 2, 0 };
+
+		FD_ZERO(&except_fds);
+		FD_SET(req_fd, &except_fds);
+
+		int rc = select(req_fd + 1, NULL, NULL, &except_fds, &tv);
+
+		if (rc == 0) {
+			fprintf(stderr, "Timeout when waiting for media request\n");
+			return;
+		} else if (rc < 0) {
+			fprintf(stderr, "Unable to select media request: %s\n",
+				strerror(errno));
+			return;
+		}
+		/*
+		 * it is safe to queue back last cap buffer only after
+		 * the following request is done so that the buffer
+		 * is not needed anymore as a reference frame
+		 */
+		if (queue_lst_buf) {
+			if (fd.qbuf(last_in_buf) < 0) {
+				fprintf(stderr, "%s: qbuf failed\n", __func__);
+				return;
+			}
+		}
+		int buf_idx = -1;
+		  /*
+		   * fin is not sent to do_handle_cap since the capture buf is
+		   * written to the file in current function
+		   */
+		rc = do_handle_cap(fd, in, NULL, &buf_idx, count[CAP],
+				   fps_ts[CAP], fmt[CAP]);
+		if (rc) {
+			fprintf(stderr, "%s: do_handle_cap err\n", __func__);
+			return;
+		}
+		/*
+		 * in case of an error in the frame, set last ts to 0 as a mean
+		 * to recover so that next request will not use a
+		 * reference buffer. Otherwise the error flag will be set to
+		 * all the future capture buffers.
+		 */
+		if (buf_idx == -1) {
+			fprintf(stderr, "%s: frame returned with error\n", __func__);
+			last_fwht_bf_ts	= 0;
+		} else {
+			cv4l_buffer cap_buf(in, index);
+			if (fd.querybuf(cap_buf))
+				return;
+			last_in_buf = cap_buf;
+			queue_lst_buf = true;
+			if (fin && cap_buf.g_bytesused(0) &&
+			    !(cap_buf.g_flags() & V4L2_BUF_FLAG_ERROR)) {
+				int idx = get_fwht_req_by_ts(get_ns_timestamp(cap_buf));
+
+				if (idx < 0) {
+					fprintf(stderr, "%s: could not find request from buffer\n", __func__);
+					fprintf(stderr, "%s: ts = %llu\n", __func__, get_ns_timestamp(cap_buf));
+					return;
+				}
+				composed_width = fwht_reqs[idx].params.width;
+				composed_height = fwht_reqs[idx].params.height;
+				write_buffer_to_file(fd, in, cap_buf,
+						     fmt[CAP], fin);
+			}
+		}
+		rc = do_handle_out(fd, out, fout, NULL, count[OUT],
+				   fps_ts[OUT], fmt[OUT]);
+		if (rc) {
+			fprintf(stderr, "%s: output stream ended\n", __func__);
+			close(req_fd);
+		}
+		index = (index + 1) % out.g_buffers();
+	}
+}
+
 static void streaming_set_m2m(cv4l_fd &fd, cv4l_fd &exp_fd)
 {
 	cv4l_queue in(fd.g_type(), memory);
@@ -2263,7 +2488,12 @@ static void streaming_set_m2m(cv4l_fd &fd, cv4l_fd &exp_fd)
 		if (out.export_bufs(&exp_fd, exp_fd.g_type()))
 			return;
 	}
-	stateful_m2m(fd, in, out, file[CAP], file[OUT], exp_fd_p);
+	if (out.reqbufs(&fd, 0))
+		goto done;
+	if (out.g_capabilities() & V4L2_BUF_CAP_SUPPORTS_REQUESTS)
+		stateless_m2m(fd, in, out, file[CAP], file[OUT], exp_fd_p);
+	else
+		stateful_m2m(fd, in, out, file[CAP], file[OUT], exp_fd_p);
 done:
 	if (options[OptStreamDmaBuf] || options[OptStreamOutDmaBuf])
 		exp_q.close_exported_fds();
-- 
2.17.1