nir: Add nir_lower_dsign as 64-bit fsign lowering.

Right now some drivers are doing dsign lowering in GLSL and haven't had to
have a NIR path due to there not being a corresponding vulkan driver.  We
want this in NIR now so that we can retire that batch of GLSL lowering
code.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Acked-by: Faith Ekstrand <faith.ekstrand@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25777>

diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
index 5c4c4ad..db1e264 100644 (file)
--- a/src/compiler/nir/nir.h
+++ b/src/compiler/nir/nir.h
@@ -3471,7 +3471,8 @@ typedef enum {
    nir_lower_dmod = (1 << 8),
    nir_lower_dsub = (1 << 9),
    nir_lower_ddiv = (1 << 10),
-   nir_lower_fp64_full_software = (1 << 11),
+   nir_lower_dsign = (1 << 11),
+   nir_lower_fp64_full_software = (1 << 12),
 } nir_lower_doubles_options;
 
 typedef enum {

diff --git a/src/compiler/nir/nir_opt_algebraic.py b/src/compiler/nir/nir_opt_algebraic.py
index da422a7..08f4099 100644 (file)
--- a/src/compiler/nir/nir_opt_algebraic.py
+++ b/src/compiler/nir/nir_opt_algebraic.py
@@ -921,6 +921,10 @@ optimizations.extend([
 
 # Float sizes
 for s in [16, 32, 64]:
+    if s == 64:
+        match_fsign_cond = "!options->lower_fsign & !(options->lower_doubles_options & nir_lower_dsign)"
+    else:
+        match_fsign_cond = "!options->lower_fsign"
     optimizations.extend([
        # These derive from the previous patterns with the application of b < 0 <=>
        # 0 < -b.  The transformation should be applied if either comparison is
@@ -979,8 +983,8 @@ for s in [16, 32, 64]:
        (('~f2u{}'.format(s), ('i2f', 'a@{}'.format(s))), a),
        (('~f2u{}'.format(s), ('u2f', 'a@{}'.format(s))), a),
 
-       (('fadd', ('b2f{}'.format(s), ('flt', 0.0, 'a@{}'.format(s))), ('fneg', ('b2f{}'.format(s), ('flt', 'a@{}'.format(s), 0.0)))), ('fsign', a), '!options->lower_fsign'),
-       (('iadd', ('b2i{}'.format(s), ('flt', 0, 'a@{}'.format(s))), ('ineg', ('b2i{}'.format(s), ('flt', 'a@{}'.format(s), 0)))), ('f2i{}'.format(s), ('fsign', a)), '!options->lower_fsign'),
+       (('fadd', ('b2f{}'.format(s), ('flt', 0.0, 'a@{}'.format(s))), ('fneg', ('b2f{}'.format(s), ('flt', 'a@{}'.format(s), 0.0)))), ('fsign', a), match_fsign_cond),
+       (('iadd', ('b2i{}'.format(s), ('flt', 0, 'a@{}'.format(s))), ('ineg', ('b2i{}'.format(s), ('flt', 'a@{}'.format(s), 0)))), ('f2i{}'.format(s), ('fsign', a)), match_fsign_cond),
 
        # float? -> float? -> floatS ==> float? -> floatS
        (('~f2f{}'.format(s), ('f2f', a)), ('f2f{}'.format(s), a)),
@@ -2217,6 +2221,7 @@ optimizations.extend([
    # Mark the new comparisons precise to prevent them being changed to 'a !=
    # 0' or 'a == 0'.
    (('fsign', a), ('fsub', ('b2f', ('!flt', 0.0, a)), ('b2f', ('!flt', a, 0.0))), 'options->lower_fsign'),
+   (('fsign', 'a@64'), ('fsub', ('b2f', ('!flt', 0.0, a)), ('b2f', ('!flt', a, 0.0))), 'options->lower_doubles_options & nir_lower_dsign'),
 
    # Address/offset calculations:
    # Drivers supporting imul24 should use the nir_lower_amul() pass, this

diff --git a/src/gallium/drivers/svga/svga_screen.c b/src/gallium/drivers/svga/svga_screen.c
index 10b82fe..26e26bc 100644 (file)
--- a/src/gallium/drivers/svga/svga_screen.c
+++ b/src/gallium/drivers/svga/svga_screen.c
@@ -752,7 +752,7 @@ vgpu10_get_shader_param(struct pipe_screen *screen,
    .use_interpolated_input_intrinsics = true
 
 #define VGPU10_OPTIONS                                                        \
-   .lower_doubles_options = nir_lower_dfloor,                                 \
+   .lower_doubles_options = nir_lower_dfloor | nir_lower_dsign,               \
    .lower_fmod = true,                                                        \
    .lower_fpow = true