radeonsi: emit DLDEXP and DFRACEXP TGSI opcodes

Note: this causes spurious regressions in some current piglit tests,
because the tests incorrectly assume that there is no denorm support for
doubles. I'm going to send out a fix for those tests as well.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>

diff --git a/src/gallium/drivers/radeonsi/si_pipe.c b/src/gallium/drivers/radeonsi/si_pipe.c
index 9a5b789a2c290d7b9d9b45d02d443ce067030151..954f9ff063fc6d8b36a3c711cd52a2f7a5965a83 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_pipe.c
+++ b/src/gallium/drivers/radeonsi/si_pipe.c
@@ -749,6 +749,7 @@ static int si_get_shader_param(struct pipe_screen* pscreen,
        case PIPE_SHADER_CAP_TGSI_SKIP_MERGE_REGISTERS:
        case PIPE_SHADER_CAP_TGSI_DROUND_SUPPORTED:
        case PIPE_SHADER_CAP_TGSI_LDEXP_SUPPORTED:
+       case PIPE_SHADER_CAP_TGSI_DFRACEXP_DLDEXP_SUPPORTED:
                return 1;
 
        case PIPE_SHADER_CAP_INDIRECT_INPUT_ADDR:
@@ -768,7 +769,6 @@ static int si_get_shader_param(struct pipe_screen* pscreen,
        /* Unsupported boolean features. */
        case PIPE_SHADER_CAP_SUBROUTINES:
        case PIPE_SHADER_CAP_SUPPORTED_IRS:
-       case PIPE_SHADER_CAP_TGSI_DFRACEXP_DLDEXP_SUPPORTED:
                return 0;
        }
        return 0;

diff --git a/src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c b/src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c
index ba7ec4f4107fe5e502dbdfa71b8632117bac6f50..818ca499d9086b0a7fff8d879d4b4e5f534b8ae2 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c
+++ b/src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c
@@ -734,6 +734,27 @@ static void emit_rsq(const struct lp_build_tgsi_action *action,
                                          bld_base->base.one, sqrt);
 }
 
+static void dfracexp_fetch_args(struct lp_build_tgsi_context *bld_base,
+                               struct lp_build_emit_data *emit_data)
+{
+       emit_data->args[0] = lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_X);
+       emit_data->arg_count = 1;
+}
+
+static void dfracexp_emit(const struct lp_build_tgsi_action *action,
+                         struct lp_build_tgsi_context *bld_base,
+                         struct lp_build_emit_data *emit_data)
+{
+       struct si_shader_context *ctx = si_shader_context(bld_base);
+
+       emit_data->output[emit_data->chan] =
+               lp_build_intrinsic(ctx->ac.builder, "llvm.amdgcn.frexp.mant.f64",
+                                  ctx->ac.f64, &emit_data->args[0], 1, 0);
+       emit_data->output1[emit_data->chan] =
+               lp_build_intrinsic(ctx->ac.builder, "llvm.amdgcn.frexp.exp.i32.f64",
+                                  ctx->ac.i32, &emit_data->args[0], 1, 0);
+}
+
 void si_shader_context_init_alu(struct lp_build_tgsi_context *bld_base)
 {
        lp_set_default_actions(bld_base);
@@ -772,6 +793,10 @@ void si_shader_context_init_alu(struct lp_build_tgsi_context *bld_base)
        bld_base->op_actions[TGSI_OPCODE_DSQRT].intr_name = "llvm.sqrt.f64";
        bld_base->op_actions[TGSI_OPCODE_DTRUNC].emit = build_tgsi_intrinsic_nomem;
        bld_base->op_actions[TGSI_OPCODE_DTRUNC].intr_name = "llvm.trunc.f64";
+       bld_base->op_actions[TGSI_OPCODE_DFRACEXP].fetch_args = dfracexp_fetch_args;
+       bld_base->op_actions[TGSI_OPCODE_DFRACEXP].emit = dfracexp_emit;
+       bld_base->op_actions[TGSI_OPCODE_DLDEXP].emit = build_tgsi_intrinsic_nomem;
+       bld_base->op_actions[TGSI_OPCODE_DLDEXP].intr_name = "llvm.amdgcn.ldexp.f64";
        bld_base->op_actions[TGSI_OPCODE_EX2].emit = build_tgsi_intrinsic_nomem;
        bld_base->op_actions[TGSI_OPCODE_EX2].intr_name = "llvm.exp2.f32";
        bld_base->op_actions[TGSI_OPCODE_FLR].emit = build_tgsi_intrinsic_nomem;