ogl_beamforming

Commit: c7a78f342b205e06b80aada2871b81dce7b6f6f9
Parent: a425672a4da9bce718f4469b058fa25acdd0a440
Author: Randy Palamar
Date:   Fri,  5 Jul 2024 15:44:33 -0600

change decoding to work on i16 data directly

Diffstat:
M
beamformer.c
 | 
2
+-
M
main.c
 | 
2
+-
M
shaders/hadamard.glsl
 | 
22
+++++++++++++++++++---

3 files changed, 21 insertions(+), 5 deletions(-)
diff --git a/beamformer.c b/beamformer.c
@@ -124,7 +124,7 @@ do_beamformer(BeamformerCtx *ctx, Arena arena)
 		void *rf_data_buf = glMapBuffer(GL_SHADER_STORAGE_BUFFER, GL_WRITE_ONLY);
 		ASSERT(rf_data_buf);
 		uv3  rf_data_dim  = ctx->csctx.rf_data_dim;
-		size rf_raw_size  = rf_data_dim.w * rf_data_dim.h * rf_data_dim.d * sizeof(i32);
+		size rf_raw_size  = rf_data_dim.w * rf_data_dim.h * rf_data_dim.d * sizeof(i16);
 		size rlen         = os_read_pipe_data(ctx->data_pipe, rf_data_buf, rf_raw_size);
 		glUnmapBuffer(GL_SHADER_STORAGE_BUFFER);
 		if (rlen == rf_raw_size) {
diff --git a/main.c b/main.c
@@ -121,7 +121,7 @@ init_compute_shader_ctx(ComputeShaderCtx *ctx, Arena a, uv3 rf_data_dim)
 {
 	ctx->rf_data_dim  = rf_data_dim;
 	/* TODO: send i16 data and convert to i32 on GPU */
-	size rf_raw_size     = rf_data_dim.w * rf_data_dim.h * rf_data_dim.d * sizeof(i32);
+	size rf_raw_size     = rf_data_dim.w * rf_data_dim.h * rf_data_dim.d * sizeof(i16);
 	size rf_decoded_size = rf_data_dim.w * rf_data_dim.h * rf_data_dim.d * sizeof(i32);
 
 	glGenBuffers(ARRAY_COUNT(ctx->rf_data_ssbos), ctx->rf_data_ssbos);
diff --git a/shaders/hadamard.glsl b/shaders/hadamard.glsl
@@ -38,10 +38,26 @@ void main()
 	uint rstride = u_rf_data_dim.x * u_rf_data_dim.y;
 	uint rfoff   = u_rf_data_dim.x * channel + time_sample;
 
-	/* N-D dot product */
+	uint ridx       = rfoff / 2;
+	uint ridx_delta = rstride / 2;
+
+	/* NOTE: Compute N-D dot product */
 	int sum = 0;
-	for (int i = 0; i < u_rf_data_dim.z; i++)
-		sum += hadamard[hoff + i] * rf_data[rfoff + rstride * i];
+	for (int i = 0; i < u_rf_data_dim.z; i++) {
+		int data = rf_data[ridx];
+
+		/* NOTE: rf_data is i16 so each access grabs two time samples at time.
+		 * We need to shift and mask to get the desired element. If the time sample
+		 * is even we take the upper half and if its odd we take the lower half.
+		 * Hopefully shader compiler is smart enough to lift this out of the loop. */
+		if ((int(time_sample) & 1) == 1)
+			data = data >> 16;
+		else
+			data = (data << 16) >> 16;
+
+		sum += hadamard[hoff + i] * data;
+		ridx += ridx_delta;
+	}
 
 	out_data[rfoff + rstride * acq] = float(sum);
 }