Commit: d0bb8be1de86b0e323254d85ea7b87e2ce1e8d84
Parent: d23026486d5af255051d3d4c61d157845931917a
Author: Randy Palamar
Date: Fri, 18 Jul 2025 12:29:50 -0600
shaders/das: early out when channel apodization is 0
This is most obviously beneficial for FORCES in the DASFast path
where it is clear that it skips the entire computation for that
thread. When you consider how the voxels are clustered it is
obvious that a large portion of warps can be immediately retired
when there is a receive f-number applied (think on individual
channel level where a large portion of the image is completely
black).
This also gives a performance boost to HERCULES and the RCA
methods but to a lesser extent since it only avoids rf data
sampling in those threads at the expense thread divergence.
On its own this change seems trivial but it brings to light that
the correct way to do DAS is via indirect dispatch from another
compute shader. That way we won't even need to issue warps for
threads which would early out due to apodization. As I implement
it I'm sure there will be other conditions for skipping full DAS
on a voxel which will be highly beneficial.
Diffstat:
| M | shaders/das.glsl | | | 78 | +++++++++++++++++++++++++++++++++++++++++++++--------------------------------- |
1 file changed, 45 insertions(+), 33 deletions(-)
diff --git a/shaders/das.glsl b/shaders/das.glsl
@@ -119,9 +119,11 @@ vec3 RCA(vec3 world_point)
float receive_distance = length(receive_vector);
float apodization = apodize(f_number * radians(180) / abs(xdc_world_point.y) * receive_vector.x);
- float sidx = sample_index(transmit_distance + receive_distance);
- vec2 valid = vec2(sidx >= 0) * vec2(sidx < dec_data_dim.x);
- result += valid * apodization * sample_rf(channel, u_channel, sidx);
+ if (apodization > 0) {
+ float sidx = sample_index(transmit_distance + receive_distance);
+ vec2 valid = vec2(sidx >= 0) * vec2(sidx < dec_data_dim.x);
+ result += valid * apodization * sample_rf(channel, u_channel, sidx);
+ }
}
return vec3(result, 0);
}
@@ -151,10 +153,12 @@ vec3 RCA(vec3 world_point)
vec2 receive_vector = xdc_world_point - rca_plane_projection(rx_center, rx_rows);
float apodization = apodize(f_number * radians(180) / abs(xdc_world_point.y) * receive_vector.x);
- float sidx = sample_index(transmit_distance + length(receive_vector));
- vec2 valid = vec2(sidx >= 0) * vec2(sidx < dec_data_dim.x);
- vec2 value = valid * apodization * sample_rf(rx_channel, transmit, sidx);
- sum += vec3(value, length(value));
+ if (apodization > 0) {
+ float sidx = sample_index(transmit_distance + length(receive_vector));
+ vec2 valid = vec2(sidx >= 0) * vec2(sidx < dec_data_dim.x);
+ vec2 value = valid * apodization * sample_rf(rx_channel, transmit, sidx);
+ sum += vec3(value, length(value));
+ }
}
}
return sum;
@@ -189,12 +193,14 @@ vec3 HERCULES(vec3 world_point)
float apodization = apodize(f_number * radians(180) / abs(xdc_world_point.z) *
distance(xdc_world_point.xy, element_position.xy));
- /* NOTE: tribal knowledge */
- if (transmit == 0) apodization *= inversesqrt(dec_data_dim.z);
+ if (apodization > 0) {
+ /* NOTE: tribal knowledge */
+ if (transmit == 0) apodization *= inversesqrt(dec_data_dim.z);
- float sidx = sample_index(transmit_distance + distance(xdc_world_point, element_position));
- vec2 valid = vec2(sidx >= 0) * vec2(sidx < dec_data_dim.x);
- result += valid * apodization * sample_rf(u_channel, transmit, sidx);
+ float sidx = sample_index(transmit_distance + distance(xdc_world_point, element_position));
+ vec2 valid = vec2(sidx >= 0) * vec2(sidx < dec_data_dim.x);
+ result += valid * apodization * sample_rf(u_channel, transmit, sidx);
+ }
}
return vec3(result, 0);
}
@@ -227,13 +233,15 @@ vec3 HERCULES(vec3 world_point)
float apodization = apodize(f_number * radians(180) / abs(xdc_world_point.z) *
distance(xdc_world_point.xy, element_position.xy));
- /* NOTE: tribal knowledge */
- if (transmit == 0) apodization *= inversesqrt(dec_data_dim.z);
-
- float sidx = sample_index(transmit_distance + distance(xdc_world_point, element_position));
- vec2 valid = vec2(sidx >= 0) * vec2(sidx < dec_data_dim.x);
- vec2 value = valid * apodization * sample_rf(rx_channel, transmit, sidx);
- result += vec3(value, length(value));
+ if (apodization > 0) {
+ /* NOTE: tribal knowledge */
+ if (transmit == 0) apodization *= inversesqrt(dec_data_dim.z);
+
+ float sidx = sample_index(transmit_distance + distance(xdc_world_point, element_position));
+ vec2 valid = vec2(sidx >= 0) * vec2(sidx < dec_data_dim.x);
+ vec2 value = valid * apodization * sample_rf(rx_channel, transmit, sidx);
+ result += vec3(value, length(value));
+ }
}
}
return result;
@@ -250,13 +258,15 @@ vec3 FORCES(vec3 world_point)
(xdc_world_point.x - u_channel * xdc_element_pitch.x));
vec2 result = vec2(0);
- for (int transmit = int(uforces); transmit < dec_data_dim.z; transmit++) {
- int tx_channel = uforces ? imageLoad(sparse_elements, transmit - int(uforces)).x : transmit;
- vec3 transmit_center = vec3(xdc_element_pitch * vec2(tx_channel, floor(dec_data_dim.y / 2)), 0);
+ if (apodization > 0) {
+ for (int transmit = int(uforces); transmit < dec_data_dim.z; transmit++) {
+ int tx_channel = uforces ? imageLoad(sparse_elements, transmit - int(uforces)).x : transmit;
+ vec3 transmit_center = vec3(xdc_element_pitch * vec2(tx_channel, floor(dec_data_dim.y / 2)), 0);
- float sidx = sample_index(distance(xdc_world_point, transmit_center) + receive_distance);
- vec2 valid = vec2(sidx >= 0) * vec2(sidx < dec_data_dim.x);
- result += valid * apodization * sample_rf(u_channel, transmit, sidx);
+ float sidx = sample_index(distance(xdc_world_point, transmit_center) + receive_distance);
+ vec2 valid = vec2(sidx >= 0) * vec2(sidx < dec_data_dim.x);
+ result += valid * apodization * sample_rf(u_channel, transmit, sidx);
+ }
}
return vec3(result, 0);
}
@@ -271,14 +281,16 @@ vec3 FORCES(vec3 world_point)
float receive_distance = distance(xdc_world_point.xz, vec2(rx_channel * xdc_element_pitch.x, 0));
float apodization = apodize(f_number * radians(180) / abs(xdc_world_point.z) *
(xdc_world_point.x - rx_channel * xdc_element_pitch.x));
- for (int transmit = int(uforces); transmit < dec_data_dim.z; transmit++) {
- int tx_channel = uforces ? imageLoad(sparse_elements, transmit - int(uforces)).x : transmit;
- vec3 transmit_center = vec3(xdc_element_pitch * vec2(tx_channel, floor(dec_data_dim.y / 2)), 0);
-
- float sidx = sample_index(distance(xdc_world_point, transmit_center) + receive_distance);
- vec2 valid = vec2(sidx >= 0) * vec2(sidx < dec_data_dim.x);
- vec2 value = valid * apodization * sample_rf(rx_channel, tx_channel, sidx);
- result += vec3(value, length(value));
+ if (apodization > 0) {
+ for (int transmit = int(uforces); transmit < dec_data_dim.z; transmit++) {
+ int tx_channel = uforces ? imageLoad(sparse_elements, transmit - int(uforces)).x : transmit;
+ vec3 transmit_center = vec3(xdc_element_pitch * vec2(tx_channel, floor(dec_data_dim.y / 2)), 0);
+
+ float sidx = sample_index(distance(xdc_world_point, transmit_center) + receive_distance);
+ vec2 valid = vec2(sidx >= 0) * vec2(sidx < dec_data_dim.x);
+ vec2 value = valid * apodization * sample_rf(rx_channel, tx_channel, sidx);
+ result += vec3(value, length(value));
+ }
}
}
return result;