[Mesa-dev] [PATCH 12/14] radeonsi/nir: add support for packed inputs

[Timothy Arceri]

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.
---
 src/gallium/drivers/radeonsi/si_shader_nir.c | 46 +++++++++++++++-------------
 1 file changed, 25 insertions(+), 21 deletions(-)

diff --git a/src/gallium/drivers/radeonsi/si_shader_nir.c b/src/gallium/drivers/radeonsi/si_shader_nir.c
index 5d82715f7a..ec748c9679 100644
--- a/src/gallium/drivers/radeonsi/si_shader_nir.c
+++ b/src/gallium/drivers/radeonsi/si_shader_nir.c
@@ -437,46 +437,42 @@ si_lower_nir(struct si_shader_selector* sel)
 
 		NIR_PASS(progress, sel->nir, nir_opt_undef);
 		NIR_PASS(progress, sel->nir, nir_opt_conditional_discard);
 		if (sel->nir->options->max_unroll_iterations) {
 			NIR_PASS(progress, sel->nir, nir_opt_loop_unroll, 0);
 		}
 	} while (progress);
 }
 
 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);
 		out[2] = LLVMGetParam(ctx->main_fn, SI_PARAM_POS_Z_FLOAT);
 		out[3] = ac_build_fdiv(&ctx->ac, ctx->ac.f32_1,
 				LLVMGetParam(ctx->main_fn, SI_PARAM_POS_W_FLOAT));
 		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
 si_nir_load_sampler_desc(struct ac_shader_abi *abi,
 		         unsigned descriptor_set, unsigned base_index,
 		         unsigned constant_index, LLVMValueRef dynamic_index,
 		         enum ac_descriptor_type desc_type, bool image,
 			 bool write)
 {
 	struct si_shader_context *ctx = si_shader_context_from_abi(abi);
@@ -516,39 +512,47 @@ si_nir_load_sampler_desc(struct ac_shader_abi *abi,
 	index = LLVMBuildAdd(ctx->gallivm.builder, index,
 			     LLVMConstInt(ctx->i32, SI_NUM_IMAGES / 2, 0), "");
 
 	return si_load_sampler_desc(ctx, list, index, desc_type);
 }
 
 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_outputs = 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_outputs & ((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_outputs |= ((uint64_t)1 << loc);
 	}
 
 	ctx->abi.inputs = &ctx->inputs[0];
 	ctx->abi.load_sampler_desc = si_nir_load_sampler_desc;
 	ctx->abi.clamp_shadow_reference = true;
 
 	ctx->num_samplers = util_last_bit(info->samplers_declared);
 	ctx->num_images = util_last_bit(info->images_declared);
 
 	ac_nir_translate(&ctx->ac, &ctx->abi, nir, NULL);
-- 
2.14.3