[Mesa-dev] [PATCH 3/3] radeonsi: document why we can't use SMEM for shader buffer loads yet

[Marek Olšák]

From: Marek Olšák <marek.olsak at amd.com>

---
 src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c | 22 ++++++++++++++++++++--
 1 file changed, 20 insertions(+), 2 deletions(-)

diff --git a/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c b/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c
index c5b94b9..9519e0d 100644
--- a/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c
+++ b/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c
@@ -410,27 +410,45 @@ static unsigned get_store_intr_attribs(bool writeonly_memory)
 }
 
 static void load_emit_buffer(struct si_shader_context *ctx,
 			     struct lp_build_emit_data *emit_data,
 			     bool readonly_memory)
 {
 	const struct tgsi_full_instruction *inst = emit_data->inst;
 	uint writemask = inst->Dst[0].Register.WriteMask;
 	uint count = util_last_bit(writemask);
 	LLVMValueRef *args = emit_data->args;
+	bool coherent = (inst->Memory.Qualifier &
+			 (TGSI_MEMORY_COHERENT | TGSI_MEMORY_VOLATILE)) != 0;
 
+	/* Force the use of VMEM for shader buffer loads, because LLVM doesn't
+	 * select SMEM for SI.load.const with a non-constant offset, and
+	 * constant offsets practically don't exist with shader buffers.
+	 *
+	 * Also, SI.load.const doesn't use inst_offset when it's lowered
+	 * to VMEM, so we just end up with more VALU instructions in the end
+	 * and no benefit.
+	 *
+	 * TODO: Remove this line once LLVM can select SMEM with a non-constant
+	 *       offset, and can derive inst_offset when VMEM is selected.
+	 *       After that, si_memory_barrier should invalidate sL1 for shader
+	 *       buffers.
+	 */
+	coherent = true;
+
+	assert(LLVMConstIntGetZExtValue(args[1]) == 0); /* vindex */
 	emit_data->output[emit_data->chan] =
-		ac_build_buffer_load(&ctx->ac, args[0], count, args[1],
+		ac_build_buffer_load(&ctx->ac, args[0], count, NULL,
 				     args[2], NULL, 0,
 				     LLVMConstIntGetZExtValue(args[3]),
 				     LLVMConstIntGetZExtValue(args[4]),
-				     readonly_memory, true);
+				     readonly_memory, coherent);
 }
 
 static LLVMValueRef get_memory_ptr(struct si_shader_context *ctx,
                                    const struct tgsi_full_instruction *inst,
                                    LLVMTypeRef type, int arg)
 {
 	struct gallivm_state *gallivm = &ctx->gallivm;
 	LLVMBuilderRef builder = gallivm->builder;
 	LLVMValueRef offset, ptr;
 	int addr_space;
-- 
2.7.4