aco: skip unused channels at the start when fetching vertices

pipeline-db (Vega):
Totals from affected shaders:
SGPRS: 161320 -> 161224 (-0.06 %)
VGPRS: 153968 -> 149408 (-2.96 %)
Spilled SGPRs: 0 -> 0 (0.00 %)
Spilled VGPRs: 0 -> 0 (0.00 %)
Private memory VGPRs: 0 -> 0 (0.00 %)
Scratch size: 0 -> 0 (0.00 %) dwords per thread
Code Size: 4331496 -> 4331308 (-0.00 %) bytes
LDS: 0 -> 0 (0.00 %) blocks
Max Waves: 27814 -> 28594 (2.80 %)

pipeline-db (Navi):
Totals from affected shaders:
SGPRS: 161504 -> 161408 (-0.06 %)
VGPRS: 153836 -> 149440 (-2.86 %)
Spilled SGPRs: 0 -> 0 (0.00 %)
Spilled VGPRs: 0 -> 0 (0.00 %)
Private memory VGPRs: 0 -> 0 (0.00 %)
Scratch size: 0 -> 0 (0.00 %) dwords per thread
Code Size: 4327572 -> 4327604 (0.00 %) bytes
LDS: 0 -> 0 (0.00 %) blocks
Max Waves: 27837 -> 28618 (2.81 %)

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3086>

diff --git a/src/amd/compiler/aco_instruction_selection.cpp b/src/amd/compiler/aco_instruction_selection.cpp
index 11f7805b56d21b54ed205d86b182c31258b02df4..0847d5b6ca8cc61e948ab677532003adcec8b86e 100644 (file)
--- a/src/amd/compiler/aco_instruction_selection.cpp
+++ b/src/amd/compiler/aco_instruction_selection.cpp
@@ -3214,6 +3214,15 @@ void visit_load_input(isel_context *ctx, nir_intrinsic_instr *instr)
       unsigned channel_start = 0;
       bool direct_fetch = false;
 
+      /* skip unused channels at the start */
+      if (vtx_info->chan_byte_size && !post_shuffle) {
+         channel_start = ffs(mask) - 1;
+         for (unsigned i = 0; i < channel_start; i++)
+            channels[i] = Temp(0, s1);
+      } else if (vtx_info->chan_byte_size && post_shuffle && !(mask & 0x8)) {
+         num_channels = 3 - (ffs(mask) - 1);
+      }
+
       /* load channels */
       while (channel_start < num_channels) {
          unsigned fetch_size = num_channels - channel_start;
@@ -3290,7 +3299,7 @@ void visit_load_input(isel_context *ctx, nir_intrinsic_instr *instr)
          unsigned num_temp = 0;
          for (unsigned i = 0; i < dst.size(); i++) {
             unsigned idx = i + component;
-            if (idx < num_channels && channels[swizzle[idx]].id()) {
+            if (swizzle[idx] < num_channels && channels[swizzle[idx]].id()) {
                Temp channel = channels[swizzle[idx]];
                if (idx == 3 && alpha_adjust != RADV_ALPHA_ADJUST_NONE)
                   channel = adjust_vertex_fetch_alpha(ctx, alpha_adjust, channel);