radeonsi/nir: add support for packed inputs

Because NIR can create non vec4 variables when implementing component
packing we need to make sure not to reprocess the same slot again.

Also we can drop the fs_attr_idx counter and just use driver_location.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

diff --git a/src/gallium/drivers/radeonsi/si_shader_nir.c b/src/gallium/drivers/radeonsi/si_shader_nir.c
index 5d82715f7a7c8aea226165992729de1d2965f873..12eb8763edb2cbd4bf0075e8e1fc914e5b448914 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_shader_nir.c
+++ b/src/gallium/drivers/radeonsi/si_shader_nir.c
@@ -444,21 +444,18 @@ si_lower_nir(struct si_shader_selector* sel)
 }
 
 static void declare_nir_input_vs(struct si_shader_context *ctx,
-                                struct nir_variable *variable, unsigned rel,
+                                struct nir_variable *variable,
                                 LLVMValueRef out[4])
 {
-       si_llvm_load_input_vs(ctx, variable->data.driver_location / 4 + rel, out);
+       si_llvm_load_input_vs(ctx, variable->data.driver_location / 4, out);
 }
 
 static void declare_nir_input_fs(struct si_shader_context *ctx,
-                                struct nir_variable *variable, unsigned rel,
-                                unsigned *fs_attr_idx,
+                                struct nir_variable *variable,
+                                unsigned input_index,
                                 LLVMValueRef out[4])
 {
-       unsigned slot = variable->data.location + rel;
-
-       assert(variable->data.location >= VARYING_SLOT_VAR0 || rel == 0);
-
+       unsigned slot = variable->data.location;
        if (slot == VARYING_SLOT_POS) {
                out[0] = LLVMGetParam(ctx->main_fn, SI_PARAM_POS_X_FLOAT);
                out[1] = LLVMGetParam(ctx->main_fn, SI_PARAM_POS_Y_FLOAT);
@@ -468,8 +465,7 @@ static void declare_nir_input_fs(struct si_shader_context *ctx,
                return;
        }
 
-       si_llvm_load_input_fs(ctx, *fs_attr_idx, out);
-       (*fs_attr_idx)++;
+       si_llvm_load_input_fs(ctx, input_index, out);
 }
 
 static LLVMValueRef
@@ -523,25 +519,33 @@ bool si_nir_build_llvm(struct si_shader_context *ctx, struct nir_shader *nir)
 {
        struct tgsi_shader_info *info = &ctx->shader->selector->info;
 
-       unsigned fs_attr_idx = 0;
+       uint64_t processed_inputs = 0;
        nir_foreach_variable(variable, &nir->inputs) {
                unsigned attrib_count = glsl_count_attribute_slots(variable->type,
                                                                   nir->info.stage == MESA_SHADER_VERTEX);
                unsigned input_idx = variable->data.driver_location;
 
-               for (unsigned i = 0; i < attrib_count; ++i) {
-                       LLVMValueRef data[4];
+               assert(attrib_count == 1);
 
-                       if (nir->info.stage == MESA_SHADER_VERTEX)
-                               declare_nir_input_vs(ctx, variable, i, data);
-                       else if (nir->info.stage == MESA_SHADER_FRAGMENT)
-                               declare_nir_input_fs(ctx, variable, i, &fs_attr_idx, data);
+               LLVMValueRef data[4];
+               unsigned loc = variable->data.location;
 
-                       for (unsigned chan = 0; chan < 4; chan++) {
-                               ctx->inputs[input_idx + chan] =
-                                       LLVMBuildBitCast(ctx->ac.builder, data[chan], ctx->ac.i32, "");
-                       }
+               /* Packed components share the same location so skip
+                * them if we have already processed the location.
+                */
+               if (processed_inputs & ((uint64_t)1 << loc))
+                       continue;
+
+               if (nir->info.stage == MESA_SHADER_VERTEX)
+                       declare_nir_input_vs(ctx, variable, data);
+               else if (nir->info.stage == MESA_SHADER_FRAGMENT)
+                       declare_nir_input_fs(ctx, variable, input_idx / 4, data);
+
+               for (unsigned chan = 0; chan < 4; chan++) {
+                       ctx->inputs[input_idx + chan] =
+                               LLVMBuildBitCast(ctx->ac.builder, data[chan], ctx->ac.i32, "");
                }
+               processed_inputs |= ((uint64_t)1 << loc);
        }
 
        ctx->abi.inputs = &ctx->inputs[0];