X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=nageru%2Fpbo_frame_allocator.cpp;h=709a2bf359e505ab7e089f567fce33b502427834;hb=HEAD;hp=6211937b701f21dc2bdfbe916781f71f4629fb6b;hpb=b44bf7cfce6a5aaffbcd1e37df39068a163438ad;p=nageru

diff --git a/nageru/pbo_frame_allocator.cpp b/nageru/pbo_frame_allocator.cpp
index 6211937..6ebe13a 100644
--- a/nageru/pbo_frame_allocator.cpp
+++ b/nageru/pbo_frame_allocator.cpp
@@ -1,14 +1,22 @@
 #include "pbo_frame_allocator.h"
 
 #include <bmusb/bmusb.h>
+#include <assert.h>
+#include <epoxy/gl.h>
 #include <movit/util.h>
+#include <mutex>
 #include <stdbool.h>
 #include <stdint.h>
 #include <stdio.h>
 #include <cstddef>
+#include <utility>
+#include <va/va.h>
 
-#include "flags.h"
+#include "mjpeg_encoder.h"
+#include "defs.h"
+#include "shared/va_resource_pool.h"
 #include "v210_converter.h"
+#include "shared/va_display.h"
 
 using namespace std;
 
@@ -26,12 +34,23 @@ void set_clamp_to_edge()
 
 }  // namespace
 
-PBOFrameAllocator::PBOFrameAllocator(bmusb::PixelFormat pixel_format, size_t frame_size, GLuint width, GLuint height, size_t num_queued_frames, GLenum buffer, GLenum permissions, GLenum map_bits)
-        : pixel_format(pixel_format), buffer(buffer)
+PBOFrameAllocator::PBOFrameAllocator(bmusb::PixelFormat pixel_format, size_t frame_size, GLuint width, GLuint height, unsigned card_index, MJPEGEncoder *mjpeg_encoder, size_t num_queued_frames, GLenum buffer, GLenum permissions, GLenum map_bits)
+	: card_index(card_index),
+	  mjpeg_encoder(mjpeg_encoder),
+	  pixel_format(pixel_format),
+	  buffer(buffer),
+	  frame_size(frame_size),
+	  num_queued_frames(num_queued_frames),
+	  width(width),
+	  height(height),
+	  permissions(permissions),
+	  map_bits(map_bits)
 {
 	userdata.reset(new Userdata[num_queued_frames]);
 	for (size_t i = 0; i < num_queued_frames; ++i) {
-		init_frame(i, frame_size, width, height, permissions, map_bits);
+		Frame frame;
+		init_frame(frame, &userdata[i], this, pixel_format, frame_size, width, height, permissions, map_bits, buffer, generation);
+		freelist.push(frame);
 	}
 	glBindBuffer(buffer, 0);
 	check_error();
@@ -39,7 +58,7 @@ PBOFrameAllocator::PBOFrameAllocator(bmusb::PixelFormat pixel_format, size_t fra
 	check_error();
 }
 
-void PBOFrameAllocator::init_frame(size_t frame_idx, size_t frame_size, GLuint width, GLuint height, GLenum permissions, GLenum map_bits)
+void PBOFrameAllocator::init_frame(Frame &frame, Userdata *ud, PBOFrameAllocator *owner, bmusb::PixelFormat pixel_format, size_t frame_size, GLuint width, GLuint height, GLenum permissions, GLenum map_bits, GLenum buffer, int generation)
 {
 	GLuint pbo;
 	glGenBuffers(1, &pbo);
@@ -49,16 +68,16 @@ void PBOFrameAllocator::init_frame(size_t frame_idx, size_t frame_size, GLuint w
 	glBufferStorage(buffer, frame_size, nullptr, permissions | GL_MAP_PERSISTENT_BIT);
 	check_error();
 
-	Frame frame;
 	frame.data = (uint8_t *)glMapBufferRange(buffer, 0, frame_size, permissions | map_bits | GL_MAP_PERSISTENT_BIT);
 	frame.data2 = frame.data + frame_size / 2;
-	frame.data_copy = new uint8_t[frame_size];
 	check_error();
 	frame.size = frame_size;
-	frame.userdata = &userdata[frame_idx];
-	userdata[frame_idx].pbo = pbo;
-	userdata[frame_idx].pixel_format = pixel_format;
-	frame.owner = this;
+	frame.userdata = ud;
+	ud->generation = generation;
+	ud->pbo = pbo;
+	ud->pixel_format = pixel_format;
+	ud->data_copy_malloc = new uint8_t[frame_size];
+	frame.owner = owner;
 
 	// For 8-bit non-planar Y'CbCr, we ask the driver to split Y' and Cb/Cr
 	// into separate textures. For 10-bit, the input format (v210)
@@ -72,48 +91,48 @@ void PBOFrameAllocator::init_frame(size_t frame_idx, size_t frame_size, GLuint w
 	// resolution is progressive.
 	switch (pixel_format) {
 	case bmusb::PixelFormat_8BitYCbCr:
-		glGenTextures(2, userdata[frame_idx].tex_y);
+		glGenTextures(2, ud->tex_y);
 		check_error();
-		glGenTextures(2, userdata[frame_idx].tex_cbcr);
+		glGenTextures(2, ud->tex_cbcr);
 		check_error();
 		break;
 	case bmusb::PixelFormat_10BitYCbCr:
-		glGenTextures(2, userdata[frame_idx].tex_v210);
+		glGenTextures(2, ud->tex_v210);
 		check_error();
-		glGenTextures(2, userdata[frame_idx].tex_444);
+		glGenTextures(2, ud->tex_444);
 		check_error();
 		break;
 	case bmusb::PixelFormat_8BitBGRA:
-		glGenTextures(2, userdata[frame_idx].tex_rgba);
+		glGenTextures(2, ud->tex_rgba);
 		check_error();
 		break;
 	case bmusb::PixelFormat_8BitYCbCrPlanar:
-		glGenTextures(2, userdata[frame_idx].tex_y);
+		glGenTextures(2, ud->tex_y);
 		check_error();
-		glGenTextures(2, userdata[frame_idx].tex_cb);
+		glGenTextures(2, ud->tex_cb);
 		check_error();
-		glGenTextures(2, userdata[frame_idx].tex_cr);
+		glGenTextures(2, ud->tex_cr);
 		check_error();
 		break;
 	default:
 		assert(false);
 	}
 
-	userdata[frame_idx].last_width[0] = width;
-	userdata[frame_idx].last_height[0] = height;
-	userdata[frame_idx].last_cbcr_width[0] = width / 2;
-	userdata[frame_idx].last_cbcr_height[0] = height;
-	userdata[frame_idx].last_v210_width[0] = 0;
-
-	userdata[frame_idx].last_width[1] = 0;
-	userdata[frame_idx].last_height[1] = 0;
-	userdata[frame_idx].last_cbcr_width[1] = 0;
-	userdata[frame_idx].last_cbcr_height[1] = 0;
-	userdata[frame_idx].last_v210_width[1] = 0;
-
-	userdata[frame_idx].last_interlaced = false;
-	userdata[frame_idx].last_has_signal = false;
-	userdata[frame_idx].last_is_connected = false;
+	ud->last_width[0] = width;
+	ud->last_height[0] = height;
+	ud->last_cbcr_width[0] = width / 2;
+	ud->last_cbcr_height[0] = height;
+	ud->last_v210_width[0] = 0;
+
+	ud->last_width[1] = 0;
+	ud->last_height[1] = 0;
+	ud->last_cbcr_width[1] = 0;
+	ud->last_cbcr_height[1] = 0;
+	ud->last_v210_width[1] = 0;
+
+	ud->last_interlaced = false;
+	ud->last_has_signal = false;
+	ud->last_is_connected = false;
 	for (unsigned field = 0; field < 2; ++field) {
 		switch (pixel_format) {
 		case bmusb::PixelFormat_10BitYCbCr: {
@@ -122,17 +141,17 @@ void PBOFrameAllocator::init_frame(size_t frame_idx, size_t frame_size, GLuint w
 			// Seemingly we need to set the minification filter even though
 			// shader image loads don't use them, or NVIDIA will just give us
 			// zero back.
-			glBindTexture(GL_TEXTURE_2D, userdata[frame_idx].tex_v210[field]);
+			glBindTexture(GL_TEXTURE_2D, ud->tex_v210[field]);
 			check_error();
 			glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 			check_error();
 			if (field == 0) {
-				userdata[frame_idx].last_v210_width[0] = v210_width;
+				ud->last_v210_width[0] = v210_width;
 				glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, v210_width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
 				check_error();
 			}
 
-			glBindTexture(GL_TEXTURE_2D, userdata[frame_idx].tex_444[field]);
+			glBindTexture(GL_TEXTURE_2D, ud->tex_444[field]);
 			check_error();
 			set_clamp_to_edge();
 			if (field == 0) {
@@ -142,7 +161,7 @@ void PBOFrameAllocator::init_frame(size_t frame_idx, size_t frame_size, GLuint w
 			break;
 		}
 		case bmusb::PixelFormat_8BitYCbCr:
-			glBindTexture(GL_TEXTURE_2D, userdata[frame_idx].tex_y[field]);
+			glBindTexture(GL_TEXTURE_2D, ud->tex_y[field]);
 			check_error();
 			set_clamp_to_edge();
 			if (field == 0) {
@@ -150,7 +169,7 @@ void PBOFrameAllocator::init_frame(size_t frame_idx, size_t frame_size, GLuint w
 				check_error();
 			}
 
-			glBindTexture(GL_TEXTURE_2D, userdata[frame_idx].tex_cbcr[field]);
+			glBindTexture(GL_TEXTURE_2D, ud->tex_cbcr[field]);
 			check_error();
 			set_clamp_to_edge();
 			if (field == 0) {
@@ -159,20 +178,16 @@ void PBOFrameAllocator::init_frame(size_t frame_idx, size_t frame_size, GLuint w
 			}
 			break;
 		case bmusb::PixelFormat_8BitBGRA:
-			glBindTexture(GL_TEXTURE_2D, userdata[frame_idx].tex_rgba[field]);
+			glBindTexture(GL_TEXTURE_2D, ud->tex_rgba[field]);
 			check_error();
 			set_clamp_to_edge();
 			if (field == 0) {
-				if (global_flags.can_disable_srgb_decoder) {  // See the comments in tweaked_inputs.h.
-					glTexImage2D(GL_TEXTURE_2D, 0, GL_SRGB8_ALPHA8, width, height, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, nullptr);
-				} else {
-					glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, width, height, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, nullptr);
-				}
+				glTexImage2D(GL_TEXTURE_2D, 0, GL_SRGB8_ALPHA8, width, height, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, nullptr);
 				check_error();
 			}
 			break;
 		case bmusb::PixelFormat_8BitYCbCrPlanar:
-			glBindTexture(GL_TEXTURE_2D, userdata[frame_idx].tex_y[field]);
+			glBindTexture(GL_TEXTURE_2D, ud->tex_y[field]);
 			check_error();
 			set_clamp_to_edge();
 			if (field == 0) {
@@ -180,7 +195,7 @@ void PBOFrameAllocator::init_frame(size_t frame_idx, size_t frame_size, GLuint w
 				check_error();
 			}
 
-			glBindTexture(GL_TEXTURE_2D, userdata[frame_idx].tex_cb[field]);
+			glBindTexture(GL_TEXTURE_2D, ud->tex_cb[field]);
 			check_error();
 			set_clamp_to_edge();
 			if (field == 0) {
@@ -188,7 +203,7 @@ void PBOFrameAllocator::init_frame(size_t frame_idx, size_t frame_size, GLuint w
 				check_error();
 			}
 
-			glBindTexture(GL_TEXTURE_2D, userdata[frame_idx].tex_cr[field]);
+			glBindTexture(GL_TEXTURE_2D, ud->tex_cr[field]);
 			check_error();
 			set_clamp_to_edge();
 			if (field == 0) {
@@ -200,8 +215,6 @@ void PBOFrameAllocator::init_frame(size_t frame_idx, size_t frame_size, GLuint w
 			assert(false);
 		}
 	}
-
-	freelist.push(frame);
 }
 
 PBOFrameAllocator::~PBOFrameAllocator()
@@ -215,9 +228,10 @@ PBOFrameAllocator::~PBOFrameAllocator()
 
 void PBOFrameAllocator::destroy_frame(Frame *frame)
 {
-	delete[] frame->data_copy;
+	Userdata *ud = (Userdata *)frame->userdata;
+	delete[] ud->data_copy_malloc;
 
-	GLuint pbo = ((Userdata *)frame->userdata)->pbo;
+	GLuint pbo = ud->pbo;
 	glBindBuffer(buffer, pbo);
 	check_error();
 	glUnmapBuffer(buffer);
@@ -226,40 +240,48 @@ void PBOFrameAllocator::destroy_frame(Frame *frame)
 	check_error();
 	glDeleteBuffers(1, &pbo);
 	check_error();
-	switch (pixel_format) {
+	switch (ud->pixel_format) {
 	case bmusb::PixelFormat_10BitYCbCr:
-		glDeleteTextures(2, ((Userdata *)frame->userdata)->tex_v210);
+		glDeleteTextures(2, ud->tex_v210);
 		check_error();
-		glDeleteTextures(2, ((Userdata *)frame->userdata)->tex_444);
+		glDeleteTextures(2, ud->tex_444);
 		check_error();
 		break;
 	case bmusb::PixelFormat_8BitYCbCr:
-		glDeleteTextures(2, ((Userdata *)frame->userdata)->tex_y);
+		glDeleteTextures(2, ud->tex_y);
 		check_error();
-		glDeleteTextures(2, ((Userdata *)frame->userdata)->tex_cbcr);
+		glDeleteTextures(2, ud->tex_cbcr);
 		check_error();
 		break;
 	case bmusb::PixelFormat_8BitBGRA:
-		glDeleteTextures(2, ((Userdata *)frame->userdata)->tex_rgba);
+		glDeleteTextures(2, ud->tex_rgba);
 		check_error();
 		break;
 	case bmusb::PixelFormat_8BitYCbCrPlanar:
-		glDeleteTextures(2, ((Userdata *)frame->userdata)->tex_y);
+		glDeleteTextures(2, ud->tex_y);
 		check_error();
-		glDeleteTextures(2, ((Userdata *)frame->userdata)->tex_cb);
+		glDeleteTextures(2, ud->tex_cb);
 		check_error();
-		glDeleteTextures(2, ((Userdata *)frame->userdata)->tex_cr);
+		glDeleteTextures(2, ud->tex_cr);
 		check_error();
 		break;
 	default:
 		assert(false);
 	}
+
+	if (ud->generation != generation) {
+		auto it = lingering_generations.find(ud->generation);
+		assert(it != lingering_generations.end());
+		if (--it->second.num_frames_left == 0) {
+			lingering_generations.erase(it);  // Deallocates the userdata block.
+		}
+	}
 }
 //static int sumsum = 0;
 
 bmusb::FrameAllocator::Frame PBOFrameAllocator::alloc_frame()
 {
-        Frame vf;
+	Frame vf;
 
 	lock_guard<mutex> lock(freelist_mutex);  // Meh.
 	if (freelist.empty()) {
@@ -271,6 +293,86 @@ bmusb::FrameAllocator::Frame PBOFrameAllocator::alloc_frame()
 	}
 	vf.len = 0;
 	vf.overflow = 0;
+
+	if (mjpeg_encoder != nullptr &&
+	    mjpeg_encoder->should_encode_mjpeg_for_card(card_index) &&
+	    vf.userdata != nullptr) {
+		Userdata *ud = (Userdata *)vf.userdata;
+		vf.data_copy = ud->data_copy_malloc;
+		ud->data_copy_current_src = Userdata::FROM_MALLOC;
+	} else {
+		vf.data_copy = nullptr;
+	}
+
+	return vf;
+}
+
+bmusb::FrameAllocator::Frame PBOFrameAllocator::create_frame(size_t width, size_t height, size_t stride)
+{
+	Frame vf;
+
+	size_t desired_frame_bytes = width * stride;
+	if (stride > 8192 * 4 || height > 8192 || desired_frame_bytes > MAX_FRAME_SIZE) {
+		return vf;
+	}
+
+	{
+		lock_guard<mutex> lock(freelist_mutex);
+		if (freelist.empty()) {
+			printf("Frame overrun (no more spare PBO frames), dropping frame!\n");
+			vf.len = 0;
+			vf.overflow = 0;
+			return vf;
+		} else {
+			vf = freelist.front();
+			freelist.pop();
+		}
+	}
+
+	Userdata *userdata = (Userdata *)vf.userdata;
+	assert(generation == userdata->generation);
+	if (vf.size < desired_frame_bytes || (vf.size > FRAME_SIZE && vf.size > desired_frame_bytes * 2)) {
+		// Frame is either too small or way too large, so reallocate it.
+		// Note that width and height now automatically becomes the right size
+		// (the one we just asked for, instead of the default for the allocator,
+		// which is generally the global resolution); it doesn't matter
+		// for correctness, since we'll recreate the texture on upload if needed,
+		// but it is nice to save that step.
+		destroy_frame(&vf);
+		init_frame(vf, userdata, this, pixel_format, std::max<size_t>(desired_frame_bytes, FRAME_SIZE), width, height, permissions, map_bits, buffer, generation);
+	};
+
+	vf.len = 0;
+	vf.overflow = 0;
+
+	if (mjpeg_encoder != nullptr &&
+	    mjpeg_encoder->should_encode_mjpeg_for_card(card_index)) {
+		if (mjpeg_encoder->using_vaapi()) {
+			VADisplay va_dpy = mjpeg_encoder->va_dpy->va_dpy;
+			VAResourcePool::VAResources resources = mjpeg_encoder->get_va_pool()->get_va_resources(width, height, VA_FOURCC_UYVY);  // Only used by DeckLinkCapture, so always 4:2:2.
+			ReleaseVAResources release(mjpeg_encoder->get_va_pool(), resources);
+
+			if (resources.image.pitches[0] == stride) {
+				userdata->va_resources = move(resources);
+				userdata->va_resources_release = move(release);
+
+				VAStatus va_status = vaMapBuffer(va_dpy, resources.image.buf, (void **)&vf.data_copy);
+				CHECK_VASTATUS(va_status, "vaMapBuffer");
+				vf.data_copy += resources.image.offsets[0];
+				userdata->data_copy_current_src = Userdata::FROM_VA_API;
+			} else {
+				printf("WARNING: Could not copy directly into VA-API MJPEG buffer for %zu x %zu, since producer and consumer disagreed on stride (%zu != %d).\n", width, height, stride, resources.image.pitches[0]);
+				vf.data_copy = userdata->data_copy_malloc;
+				userdata->data_copy_current_src = Userdata::FROM_MALLOC;
+			}
+		} else {
+			vf.data_copy = userdata->data_copy_malloc;
+			userdata->data_copy_current_src = Userdata::FROM_MALLOC;
+		}
+	} else {
+		vf.data_copy = nullptr;
+	}
+
 	return vf;
 }
 
@@ -309,7 +411,82 @@ void PBOFrameAllocator::release_frame(Frame frame)
 	}
 #endif
 
+	{
+		// In case we never got to upload the frame to MJPEGEncoder.
+		Userdata *userdata = (Userdata *)frame.userdata;
+		VAResourcePool::VAResources resources __attribute__((unused)) = move(userdata->va_resources);
+		ReleaseVAResources release = move(userdata->va_resources_release);
+
+		if (frame.data_copy != nullptr && userdata->data_copy_current_src == Userdata::FROM_VA_API) {
+			VADisplay va_dpy = mjpeg_encoder->va_dpy->va_dpy;
+			VAStatus va_status = vaUnmapBuffer(va_dpy, resources.image.buf);
+			CHECK_VASTATUS(va_status, "vaUnmapBuffer");
+
+			frame.data_copy = nullptr;
+		}
+	}
+
 	lock_guard<mutex> lock(freelist_mutex);
-	freelist.push(frame);
+	Userdata *userdata = (Userdata *)frame.userdata;
+	if (userdata->generation == generation) {
+		freelist.push(frame);
+	} else {
+		destroy_frame(&frame);
+	}
 	//--sumsum;
 }
+
+void PBOFrameAllocator::reconfigure(bmusb::PixelFormat pixel_format,
+	                 size_t frame_size,
+	                 GLuint width, GLuint height,
+	                 unsigned card_index,
+	                 MJPEGEncoder *mjpeg_encoder,
+	                 size_t num_queued_frames,
+	                 GLenum buffer,
+	                 GLenum permissions,
+	                 GLenum map_bits)
+{
+	if (pixel_format == this->pixel_format &&
+	    frame_size == this->frame_size &&
+	    width == this->width && height == this->height &&
+	    card_index == this->card_index &&
+	    mjpeg_encoder == this->mjpeg_encoder &&
+	    num_queued_frames == this->num_queued_frames &&
+	    buffer == this->buffer &&
+	    permissions == this->permissions &&
+	    map_bits == this->map_bits) {
+		return;
+	}
+
+	lock_guard<mutex> lock(freelist_mutex);
+	lingering_generations[generation] = LingeringGeneration{ move(userdata), this->num_queued_frames };
+	++generation;
+
+	while (!freelist.empty()) {
+		Frame frame = freelist.front();
+		freelist.pop();
+		destroy_frame(&frame);
+	}
+
+	this->pixel_format = pixel_format;
+	this->frame_size = frame_size;
+	this->width = width;
+	this->height = height;
+	this->card_index = card_index;
+	this->mjpeg_encoder = mjpeg_encoder;
+	this->num_queued_frames = num_queued_frames;
+	this->buffer = buffer;
+	this->permissions = permissions;
+	this->map_bits = map_bits;
+
+	userdata.reset(new Userdata[num_queued_frames]);
+	for (size_t i = 0; i < num_queued_frames; ++i) {
+		Frame frame;
+		init_frame(frame, &userdata[i], this, pixel_format, frame_size, width, height, permissions, map_bits, buffer, generation);
+		freelist.push(frame);
+	}
+
+	// There may still be frames out with the old configuration
+	// (for instance, living in GLWidget); they will be destroyed
+	// when they come back in release_frame().
+}