Mesa (master): radeonsi/gfx10: use fma for TGSI_OPCODE_FMA

[GitLab Mirror]

Module: Mesa
Branch: master
Commit: 2f42d4cacc9a12bc210c279d86833bf210c9c806
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=2f42d4cacc9a12bc210c279d86833bf210c9c806

Author: Marek Olšák <marek.olsak at amd.com>
Date:   Mon Aug 26 16:20:44 2019 -0400

radeonsi/gfx10: use fma for TGSI_OPCODE_FMA

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer at amd.com>

---

 src/gallium/drivers/radeonsi/si_shader_internal.h   |  2 +-
 src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c   | 17 ++++++++++++++---
 src/gallium/drivers/radeonsi/si_shader_tgsi_setup.c |  2 +-
 3 files changed, 16 insertions(+), 5 deletions(-)

diff --git a/src/gallium/drivers/radeonsi/si_shader_internal.h b/src/gallium/drivers/radeonsi/si_shader_internal.h
index b576d94a63f..98abbdfc693 100644
--- a/src/gallium/drivers/radeonsi/si_shader_internal.h
+++ b/src/gallium/drivers/radeonsi/si_shader_internal.h
@@ -342,7 +342,7 @@ LLVMValueRef si_get_bounded_indirect_index(struct si_shader_context *ctx,
 					   int rel_index, unsigned num);
 LLVMValueRef si_get_sample_id(struct si_shader_context *ctx);
 
-void si_shader_context_init_alu(struct lp_build_tgsi_context *bld_base);
+void si_shader_context_init_alu(struct si_shader_context *ctx);
 void si_shader_context_init_mem(struct si_shader_context *ctx);
 
 LLVMValueRef si_load_sampler_desc(struct si_shader_context *ctx,
diff --git a/src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c b/src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c
index adad3223d99..4be410ec331 100644
--- a/src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c
+++ b/src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c
@@ -23,6 +23,7 @@
  */
 
 #include "si_shader_internal.h"
+#include "si_pipe.h"
 #include "ac_llvm_util.h"
 
 void si_llvm_emit_kill(struct ac_shader_abi *abi, LLVMValueRef visible)
@@ -677,8 +678,10 @@ static void dfracexp_emit(const struct lp_build_tgsi_action *action,
 				   ctx->ac.i32, &in, 1, 0);
 }
 
-void si_shader_context_init_alu(struct lp_build_tgsi_context *bld_base)
+void si_shader_context_init_alu(struct si_shader_context *ctx)
 {
+	struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
+
 	lp_set_default_actions(bld_base);
 
 	bld_base->op_actions[TGSI_OPCODE_AND].emit = emit_and;
@@ -722,8 +725,16 @@ void si_shader_context_init_alu(struct lp_build_tgsi_context *bld_base)
 	bld_base->op_actions[TGSI_OPCODE_EX2].intr_name = "llvm.exp2.f32";
 	bld_base->op_actions[TGSI_OPCODE_FLR].emit = build_tgsi_intrinsic_nomem;
 	bld_base->op_actions[TGSI_OPCODE_FLR].intr_name = "llvm.floor.f32";
-	bld_base->op_actions[TGSI_OPCODE_FMA].emit =
-		bld_base->op_actions[TGSI_OPCODE_MAD].emit;
+
+	/* FMA is better on GFX10, because it has FMA units instead of MUL-ADD units. */
+	if (ctx->screen->info.chip_class >= GFX10) {
+		bld_base->op_actions[TGSI_OPCODE_FMA].emit = build_tgsi_intrinsic_nomem;
+		bld_base->op_actions[TGSI_OPCODE_FMA].intr_name = "llvm.fma.f32";
+	} else {
+		bld_base->op_actions[TGSI_OPCODE_FMA].emit =
+			bld_base->op_actions[TGSI_OPCODE_MAD].emit;
+	}
+
 	bld_base->op_actions[TGSI_OPCODE_FRC].emit = emit_frac;
 	bld_base->op_actions[TGSI_OPCODE_F2I].emit = emit_f2i;
 	bld_base->op_actions[TGSI_OPCODE_F2U].emit = emit_f2u;
diff --git a/src/gallium/drivers/radeonsi/si_shader_tgsi_setup.c b/src/gallium/drivers/radeonsi/si_shader_tgsi_setup.c
index 39abd4b18f6..1c6522322ff 100644
--- a/src/gallium/drivers/radeonsi/si_shader_tgsi_setup.c
+++ b/src/gallium/drivers/radeonsi/si_shader_tgsi_setup.c
@@ -1008,7 +1008,7 @@ void si_llvm_context_init(struct si_shader_context *ctx,
 	bld_base->op_actions[TGSI_OPCODE_ENDIF].emit = endif_emit;
 	bld_base->op_actions[TGSI_OPCODE_ENDLOOP].emit = endloop_emit;
 
-	si_shader_context_init_alu(&ctx->bld_base);
+	si_shader_context_init_alu(ctx);
 	si_shader_context_init_mem(ctx);
 
 	ctx->voidt = LLVMVoidTypeInContext(ctx->ac.context);