ngscopeclient · Feb 23, 2020
Showing with 53 additions and 38 deletions.

+0 −9 glscopeclient/WaveformArea_events.cpp

+14 −8 glscopeclient/WaveformArea_rendering.cpp

+39 −21 glscopeclient/shaders/waveform-compute.glsl
diff --git a/glscopeclient/WaveformArea_events.cpp b/glscopeclient/WaveformArea_events.cpp
@@ -566,15 +566,6 @@ void WaveformArea::OnWaveformDataReady()
 		m_group->m_xAxisOffset = -eye->GetUIWidth();
 	}
 
-	//Download the waveform to the GPU and kick off the compute shader for rendering it
-	if(!IsEye() && !IsWaterfall())
-	{
-		make_current();
-		m_geometryOK = PrepareGeometry();
-		if(m_geometryOK)
-			RenderTrace();
-	}
-
 	//Update our measurements and redraw the waveform
 	SetGeometryDirty();
 	queue_draw();

diff --git a/glscopeclient/WaveformArea_rendering.cpp b/glscopeclient/WaveformArea_rendering.cpp
@@ -211,7 +211,13 @@ bool WaveformArea::on_render(const Glib::RefPtr<Gdk::GLContext>& /*context*/)
 		RenderPersistenceOverlay();
 	*/
 
-	//PrepareGeometry() / RenderTrace() are now launched early on, by OnWaveformDataReady()
+	//Download the waveform to the GPU and kick off the compute shader for rendering it
+	if(!IsEye() && !IsWaterfall())
+	{
+		m_geometryOK = PrepareGeometry();
+		if(m_geometryOK)
+			RenderTrace();
+	}
 
 	//Launch software rendering passes and push these to the GPU
 	ComputeAndDownloadCairoUnderlays();
@@ -350,15 +356,15 @@ void WaveformArea::RenderPersistenceOverlay()
 
 void WaveformArea::RenderTrace()
 {
-	//Round thread count up to next multiple of the local size (must be power of two)
-	int localSize = 128;
-	int numThreads = m_plotRight;
-	if(0 != (numThreads % localSize) )
+	//Round thread block size up to next multiple of the local size (must be power of two)
+	int localSize = 2;
+	int numCols = m_plotRight;
+	if(0 != (numCols % localSize) )
 	{
-		numThreads |= (localSize-1);
-		numThreads ++;
+		numCols |= (localSize-1);
+		numCols ++;
 	}
-	int numGroups = numThreads / localSize;
+	int numGroups = numCols / localSize;
 
 	m_waveformComputeProgram.Bind();
 	m_waveformComputeProgram.SetImageUniform(m_waveformTextureResolved, "outputTex");

diff --git a/glscopeclient/shaders/waveform-compute.glsl b/glscopeclient/shaders/waveform-compute.glsl
@@ -28,35 +28,53 @@ layout(std430, binding=3) buffer index
 	uint xind[];
 };
 
-layout(local_size_x=128, local_size_y=1, local_size_z=1) in;
+//Maximum height of a single waveform, in pixels.
+//This is enough for a nearly fullscreen 4K window so should be plenty.
+#define MAX_HEIGHT		2048
+
+//Number of columns of pixels per thread block
+#define COLS_PER_BLOCK	2
+
+layout(local_size_x=COLS_PER_BLOCK, local_size_y=1, local_size_z=1) in;
 
 //Interpolate a Y coordinate
 float InterpolateY(vec2 left, vec2 right, float slope, float x)
 {
 	return left.y + ( (x - left.x) * slope );
 }
 
-//Maximum height of a single waveform, in pixels.
-//This is enough for a nearly fullscreen 4K window so should be plenty.
-#define MAX_HEIGHT 2048
+/*
+NEW IDEA
+Multiple threads per X coordinate (say, 32 - 1 warp)
+Parallel fetch base[i+z] and atomically increment local memory
+
+Each local has a 2D shared array
+Assuming 96 KB shared memory, we can fit a total of 24K float32 temp pixels
+Assuming 2K max line height, that's up to 12 pixels of width per local
+*/
+
+//Shared buffer for the local working buffer
+shared float g_workingBuffer[COLS_PER_BLOCK][MAX_HEIGHT];
 
 void main()
 {
-	//Make sure image isn't too big for our hard coded max
-	//TODO: truncate in this case??
-	float g_workingBuffer[MAX_HEIGHT];
+	//Abort if window height is too big, or if we're off the end of the window
 	if(windowHeight > MAX_HEIGHT)
 		return;
+	if(gl_GlobalInvocationID.x > windowWidth)
+		return;
 
 	//Save some constants
-	float x = gl_GlobalInvocationID.x;
-	if(x > windowWidth)
-		return;
 	float alpha = float(alpha_scaled) / 256;
 
-	//Clear column to blank
-	for(uint y=0; y<windowHeight; y++)
-		g_workingBuffer[y] = 0;
+	//Clear column to blank in the first thread of the block
+	if(gl_LocalInvocationID.y == 0)
+	{
+		for(uint y=0; y<windowHeight; y++)
+			g_workingBuffer[gl_LocalInvocationID.x][y] = 0;
+	}
+	barrier();
+	memoryBarrierShared();
 
 	//Loop over the waveform, starting at the leftmost point that overlaps this column
 	uint istart = xind[gl_GlobalInvocationID.x];
@@ -68,11 +86,11 @@ void main()
 		right = vec2(data[i+1].x, data[i+1].voltage);
 
 		//If the current point is right of us, stop
-		if(left.x > x+1)
+		if(left.x > gl_GlobalInvocationID.x + 1)
 			break;
 
 		//If the upcoming point is still left of us, we're not there yet
-		if(right.x < x)
+		if(right.x < gl_GlobalInvocationID.x)
 		{
 			left = right;
 			continue;
@@ -84,10 +102,10 @@ void main()
 
 		//Interpolate if either end is outside our column
 		float slope = (right.y - left.y) / (right.x - left.x);
-		if(left.x < x)
-			starty = InterpolateY(left, right, slope, x);
-		if(right.x > x+1)
-			endy = InterpolateY(left, right, slope, x+1);
+		if(left.x < gl_GlobalInvocationID.x)
+			starty = InterpolateY(left, right, slope, gl_GlobalInvocationID.x);
+		if(right.x > gl_GlobalInvocationID.x + 1)
+			endy = InterpolateY(left, right, slope, gl_GlobalInvocationID.x + 1);
 
 		//Sort Y coordinates from min to max
 		int ymin = int(min(starty, endy));
@@ -99,7 +117,7 @@ void main()
 		//Fill in the space between min and max for this segment
 		for(int y=ymin; y <= ymax; y++)
 		{
-			g_workingBuffer[y] += alpha;
+			g_workingBuffer[gl_LocalInvocationID.x][y] += alpha;
 		}
 
 		//TODO: antialiasing
@@ -108,5 +126,5 @@ void main()
 
 	//Copy working buffer to RGB output
 	for(uint y=0; y<windowHeight; y++)
-		imageStore(outputTex, ivec2(gl_GlobalInvocationID.x, y), vec4(0, 0, 0, g_workingBuffer[y]));
+		imageStore(outputTex, ivec2(gl_GlobalInvocationID.x, y), vec4(0, 0, 0, g_workingBuffer[gl_LocalInvocationID.x][y]));
 }