Mesa (main): radeonsi: try to keep all VS input loads together for better perf

[GitLab Mirror]

Module: Mesa
Branch: main
Commit: b35b5926c9b077087bcfc614ecd4203b710a57aa
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=b35b5926c9b077087bcfc614ecd4203b710a57aa

Author: Marek Olšák <marek.olsak at amd.com>
Date:   Mon May 24 19:42:17 2021 -0400

radeonsi: try to keep all VS input loads together for better perf

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer at amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11102>

---

 src/gallium/drivers/radeonsi/si_shader_llvm_vs.c | 16 ++++++++++++++--
 1 file changed, 14 insertions(+), 2 deletions(-)

diff --git a/src/gallium/drivers/radeonsi/si_shader_llvm_vs.c b/src/gallium/drivers/radeonsi/si_shader_llvm_vs.c
index f49e0de6563..0d07aeb7a7c 100644
--- a/src/gallium/drivers/radeonsi/si_shader_llvm_vs.c
+++ b/src/gallium/drivers/radeonsi/si_shader_llvm_vs.c
@@ -95,6 +95,18 @@ static void load_input_vs(struct si_shader_context *ctx, unsigned input_index, L
       return;
    }
 
+   /* Set can_speculate=false to help keep all loads grouped together
+    * for better latency hiding. If it was true, LLVM could move the loads forward
+    * and accidentally double memory latency by doing:
+    *
+    *    buffer_load_dword_xyzw
+    *    s_waitcnt vmcnt(0)
+    *    buffer_load_dword_xyzw
+    *    s_waitcnt vmcnt(0)
+    *
+    * ... which is what we must prevent at all cost.
+    */
+   const bool can_speculate = false;
    unsigned bit_size = info->input_fp16_lo_hi_valid[input_index] & 0x1 ? 16 : 32;
    LLVMTypeRef int_type = bit_size == 16 ? ctx->ac.i16 : ctx->ac.i32;
    LLVMTypeRef float_type = bit_size == 16 ? ctx->ac.f16 : ctx->ac.f32;
@@ -125,7 +137,7 @@ static void load_input_vs(struct si_shader_context *ctx, unsigned input_index, L
       tmp = ac_build_opencoded_load_format(&ctx->ac, fix_fetch.u.log_size,
                                            fix_fetch.u.num_channels_m1 + 1, fix_fetch.u.format,
                                            fix_fetch.u.reverse, !opencode, vb_desc, vertex_index,
-                                           ctx->ac.i32_0, ctx->ac.i32_0, 0, true);
+                                           ctx->ac.i32_0, ctx->ac.i32_0, 0, can_speculate);
       for (unsigned i = 0; i < 4; ++i)
          out[i] =
             LLVMBuildExtractElement(ctx->ac.builder, tmp, LLVMConstInt(ctx->ac.i32, i, false), "");
@@ -171,7 +183,7 @@ static void load_input_vs(struct si_shader_context *ctx, unsigned input_index, L
    for (unsigned i = 0; i < num_fetches; ++i) {
       LLVMValueRef voffset = LLVMConstInt(ctx->ac.i32, fetch_stride * i, 0);
       fetches[i] = ac_build_buffer_load_format(&ctx->ac, vb_desc, vertex_index, voffset,
-                                               channels_per_fetch, 0, true,
+                                               channels_per_fetch, 0, can_speculate,
                                                bit_size == 16, false);
    }