radeonsi: pack GS output components for each vertex stream contiguously

Note that the memory layout of one vertex stream inside one "item" (= memory
written by one GS wave) on the GSVS ring is:

  t0v0c0 ... t15v0c0 t0v1c0 ... t15v1c0 ... t0vLc0 ... t15vLc0
  t0v0c1 ... t15v0c1 t0v1c1 ... t15v1c1 ... t0vLc1 ... t15vLc1
                        ...
  t0v0cL ... t15v0cL t0v1cL ... t15v1cL ... t0vLcL ... t15vLcL
  t16v0c0 ... t31v0c0 t16v1c0 ... t31v1c0 ... t16vLc0 ... t31vLc0
  t16v0c1 ... t31v0c1 t16v1c1 ... t31v1c1 ... t16vLc1 ... t31vLc1
                        ...
  t16v0cL ... t31v0cL t16v1cL ... t31v1cL ... t16vLcL ... t31vLcL

                        ...

  t48v0c0 ... t63v0c0 t48v1c0 ... t63v1c0 ... t48vLc0 ... t63vLc0
  t48v0c1 ... t63v0c1 t48v1c1 ... t63v1c1 ... t48vLc1 ... t63vLc1
                        ...
  t48v0cL ... t63v0cL t48v1cL ... t63v1cL ... t48vLcL ... t63vLcL

where tNN indicates the thread number, vNN the vertex number (in the order of
EMIT_VERTEX), and cNN the output component (vL and cL are the last vertex and
component, respectively).

The vertex streams are laid out sequentially.

The swizzling by 16 threads is hard-coded in the way the VGT generates the
offset passed into the GS copy shader, and the jump every 16 threads is
calculated from VGT_GSVS_RING_OFFSET_n and VGT_GSVS_RING_ITEMSIZE in a way
that makes it difficult to deviate from this layout (at least that's what
I've experimentally confirmed on VI after first trying to go the simpler
route of just interleaving the vertex streams).

Reviewed-by: Marek Olšák <[email protected]>

1 files changed, 8 insertions, 3 deletions

diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index 7324eb38a1c..48ccd83b396 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -5277,7 +5277,7 @@ static void si_llvm_emit_vertex(
 	LLVMValueRef gs_next_vertex;
 	LLVMValueRef can_emit, kill;
 	LLVMValueRef args[2];
-	unsigned chan;
+	unsigned chan, offset;
 	int i;
 	unsigned stream;
 
@@ -5312,6 +5312,7 @@ static void si_llvm_emit_vertex(
 		lp_build_if(&if_state, gallivm, can_emit);
 	}
 
+	offset = 0;
 	for (i = 0; i < info->num_outputs; i++) {
 		LLVMValueRef *out_ptr =
 			ctx->soa.outputs[i];
@@ -5323,8 +5324,9 @@ static void si_llvm_emit_vertex(
 
 			LLVMValueRef out_val = LLVMBuildLoad(gallivm->builder, out_ptr[chan], "");
 			LLVMValueRef voffset =
-				lp_build_const_int32(gallivm, (i * 4 + chan) *
+				lp_build_const_int32(gallivm, offset *
 						     shader->selector->gs_max_out_vertices);
+			offset++;
 
 			voffset = lp_build_add(uint, voffset, gs_next_vertex);
 			voffset = lp_build_mul_imm(uint, voffset, 4);
@@ -6419,6 +6421,7 @@ si_generate_gs_copy_shader(struct si_screen *sscreen,
 
 	for (int stream = 0; stream < 4; stream++) {
 		LLVMBasicBlockRef bb;
+		unsigned offset;
 
 		if (!gsinfo->num_stream_output_components[stream])
 			continue;
@@ -6431,6 +6434,7 @@ si_generate_gs_copy_shader(struct si_screen *sscreen,
 		LLVMPositionBuilderAtEnd(builder, bb);
 
 		/* Fetch vertex data from GSVS ring */
+		offset = 0;
 		for (i = 0; i < gsinfo->num_outputs; ++i) {
 			for (unsigned chan = 0; chan < 4; chan++) {
 				if (!(gsinfo->output_usagemask[i] & (1 << chan)) ||
@@ -6441,7 +6445,8 @@ si_generate_gs_copy_shader(struct si_screen *sscreen,
 
 				args[2] = lp_build_const_int32(
 					gallivm,
-					(i * 4 + chan) * gs_selector->gs_max_out_vertices * 16 * 4);
+					offset * gs_selector->gs_max_out_vertices * 16 * 4);
+				offset++;
 
 				outputs[i].values[chan] =
 					LLVMBuildBitCast(gallivm->builder,