nir: split fuse_ffma into fuse_ffma16/32/64

AMD wants different behavior for each bit size

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6756>

10 files changed, 39 insertions, 13 deletions

diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
index d3ba1bf0f00..d048cb35b5d 100644
--- a/src/compiler/nir/nir.h
+++ b/src/compiler/nir/nir.h
@@ -3056,7 +3056,9 @@ typedef enum {
 typedef struct nir_shader_compiler_options {
    bool lower_fdiv;
    bool lower_ffma;
-   bool fuse_ffma;
+   bool fuse_ffma16;
+   bool fuse_ffma32;
+   bool fuse_ffma64;
    bool lower_flrp16;
    bool lower_flrp32;
    /** Lowers flrp when it does not support doubles */
diff --git a/src/compiler/nir/nir_opt_algebraic.py b/src/compiler/nir/nir_opt_algebraic.py
index 39c07ce0b7f..a03cc549e5f 100644
--- a/src/compiler/nir/nir_opt_algebraic.py
+++ b/src/compiler/nir/nir_opt_algebraic.py
@@ -195,7 +195,9 @@ optimizations.extend([
    (('fceil', a), ('fneg', ('ffloor', ('fneg', a))), 'options->lower_fceil'),
    (('ffma', a, b, c), ('fadd', ('fmul', a, b), c), 'options->lower_ffma'),
    # Always lower inexact ffma, because it will be fused back by late optimizations (nir_opt_algebraic_late).
-   (('~ffma', a, b, c), ('fadd', ('fmul', a, b), c), 'options->fuse_ffma'),
+   (('~ffma@16', a, b, c), ('fadd', ('fmul', a, b), c), 'options->fuse_ffma16'),
+   (('~ffma@32', a, b, c), ('fadd', ('fmul', a, b), c), 'options->fuse_ffma32'),
+   (('~ffma@64', a, b, c), ('fadd', ('fmul', a, b), c), 'options->fuse_ffma64'),
 
    (('~fmul', ('fadd', ('iand', ('ineg', ('b2i', 'a@bool')), ('fmul', b, c)), '#d'), '#e'),
     ('bcsel', a, ('fmul', ('fadd', ('fmul', b, c), d), e), ('fmul', d, e))),
@@ -2028,7 +2030,9 @@ late_optimizations = [
    (('fneg', a), ('fsub', 0.0, a), 'options->lower_negate'),
    (('ineg', a), ('isub', 0, a), 'options->lower_negate'),
    (('iabs', a), ('imax', a, ('ineg', a)), 'options->lower_iabs'),
-   (('~fadd', ('fmul', a, b), c), ('ffma', a, b, c), 'options->fuse_ffma'),
+   (('~fadd@16', ('fmul', a, b), c), ('ffma', a, b, c), 'options->fuse_ffma16'),
+   (('~fadd@32', ('fmul', a, b), c), ('ffma', a, b, c), 'options->fuse_ffma32'),
+   (('~fadd@64', ('fmul', a, b), c), ('ffma', a, b, c), 'options->fuse_ffma64'),
 
    # These are duplicated from the main optimizations table.  The late
    # patterns that rearrange expressions like x - .5 < 0 to x < .5 can create
diff --git a/src/freedreno/ir3/ir3_nir.c b/src/freedreno/ir3/ir3_nir.c
index 308d42542f2..f33f048a505 100644
--- a/src/freedreno/ir3/ir3_nir.c
+++ b/src/freedreno/ir3/ir3_nir.c
@@ -47,7 +47,9 @@ static const nir_shader_compiler_options options = {
 		.lower_usub_borrow = true,
 		.lower_mul_high = true,
 		.lower_mul_2x32_64 = true,
-		.fuse_ffma = true,
+		.fuse_ffma16 = true,
+		.fuse_ffma32 = true,
+		.fuse_ffma64 = true,
 		.vertex_id_zero_based = true,
 		.lower_extract_byte = true,
 		.lower_extract_word = true,
@@ -97,7 +99,9 @@ static const nir_shader_compiler_options options_a6xx = {
 		.lower_usub_borrow = true,
 		.lower_mul_high = true,
 		.lower_mul_2x32_64 = true,
-		.fuse_ffma = true,
+		.fuse_ffma16 = true,
+		.fuse_ffma32 = true,
+		.fuse_ffma64 = true,
 		.vertex_id_zero_based = false,
 		.lower_extract_byte = true,
 		.lower_extract_word = true,
diff --git a/src/gallium/drivers/etnaviv/etnaviv_screen.c b/src/gallium/drivers/etnaviv/etnaviv_screen.c
index 1637eaa3f5f..3cc7dcec148 100644
--- a/src/gallium/drivers/etnaviv/etnaviv_screen.c
+++ b/src/gallium/drivers/etnaviv/etnaviv_screen.c
@@ -1004,7 +1004,9 @@ etna_screen_create(struct etna_device *dev, struct etna_gpu *gpu,
       .lower_fpow = true,
       .lower_sub = true,
       .lower_ftrunc = true,
-      .fuse_ffma = true,
+      .fuse_ffma16 = true,
+      .fuse_ffma32 = true,
+      .fuse_ffma64 = true,
       .lower_bitops = true,
       .lower_all_io_to_temps = true,
       .vertex_id_zero_based = true,
diff --git a/src/gallium/drivers/freedreno/a2xx/ir2_nir.c b/src/gallium/drivers/freedreno/a2xx/ir2_nir.c
index 58e99f0943b..6cf95d5f4e3 100644
--- a/src/gallium/drivers/freedreno/a2xx/ir2_nir.c
+++ b/src/gallium/drivers/freedreno/a2xx/ir2_nir.c
@@ -35,7 +35,9 @@ static const nir_shader_compiler_options options = {
 	.lower_fmod = true,
 	.lower_fdiv = true,
 	.lower_fceil = true,
-	.fuse_ffma = true,
+	.fuse_ffma16 = true,
+	.fuse_ffma32 = true,
+	.fuse_ffma64 = true,
 	/* .fdot_replicates = true, it is replicated, but it makes things worse */
 	.lower_all_io_to_temps = true,
 	.vertex_id_zero_based = true, /* its not implemented anyway */
diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_from_nir.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_from_nir.cpp
index c466d249e58..c5e54779ad7 100644
--- a/src/gallium/drivers/nouveau/codegen/nv50_ir_from_nir.cpp
+++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_from_nir.cpp
@@ -3207,7 +3207,9 @@ nvir_nir_shader_compiler_options(int chipset)
    nir_shader_compiler_options op = {};
    op.lower_fdiv = (chipset >= NVISA_GV100_CHIPSET);
    op.lower_ffma = false;
-   op.fuse_ffma = false; /* nir doesn't track mad vs fma */
+   op.fuse_ffma16 = false; /* nir doesn't track mad vs fma */
+   op.fuse_ffma32 = false; /* nir doesn't track mad vs fma */
+   op.fuse_ffma64 = false; /* nir doesn't track mad vs fma */
    op.lower_flrp16 = (chipset >= NVISA_GV100_CHIPSET);
    op.lower_flrp32 = true;
    op.lower_flrp64 = true;
diff --git a/src/gallium/drivers/nouveau/nv50/nv50_screen.c b/src/gallium/drivers/nouveau/nv50/nv50_screen.c
index 06f73ab6be2..29201eee649 100644
--- a/src/gallium/drivers/nouveau/nv50/nv50_screen.c
+++ b/src/gallium/drivers/nouveau/nv50/nv50_screen.c
@@ -923,7 +923,9 @@ int nv50_tls_realloc(struct nv50_screen *screen, unsigned tls_space)
 }
 
 static const nir_shader_compiler_options nir_options = {
-   .fuse_ffma = false, /* nir doesn't track mad vs fma */
+   .fuse_ffma16 = false, /* nir doesn't track mad vs fma */
+   .fuse_ffma32 = false, /* nir doesn't track mad vs fma */
+   .fuse_ffma64 = false, /* nir doesn't track mad vs fma */
    .lower_flrp32 = true,
    .lower_flrp64 = true,
    .lower_fpow = false,
diff --git a/src/gallium/drivers/r600/r600_pipe_common.c b/src/gallium/drivers/r600/r600_pipe_common.c
index b4dffa76a11..eab7ce91dd7 100644
--- a/src/gallium/drivers/r600/r600_pipe_common.c
+++ b/src/gallium/drivers/r600/r600_pipe_common.c
@@ -1179,7 +1179,9 @@ struct pipe_resource *r600_resource_create_common(struct pipe_screen *screen,
 }
 
 const struct nir_shader_compiler_options r600_nir_fs_options = {
-	.fuse_ffma = true,
+	.fuse_ffma16 = true,
+	.fuse_ffma32 = true,
+	.fuse_ffma64 = true,
 	.lower_scmp = true,
 	.lower_flrp32 = true,
 	.lower_flrp64 = true,
@@ -1203,7 +1205,9 @@ const struct nir_shader_compiler_options r600_nir_fs_options = {
 };
 
 const struct nir_shader_compiler_options r600_nir_options = {
-	.fuse_ffma = true,
+	.fuse_ffma16 = true,
+	.fuse_ffma32 = true,
+	.fuse_ffma64 = true,
 	.lower_scmp = true,
 	.lower_flrp32 = true,
 	.lower_flrp64 = true,
diff --git a/src/gallium/drivers/radeonsi/si_get.c b/src/gallium/drivers/radeonsi/si_get.c
index a511bab4f0d..1f19fdd8817 100644
--- a/src/gallium/drivers/radeonsi/si_get.c
+++ b/src/gallium/drivers/radeonsi/si_get.c
@@ -946,7 +946,9 @@ void si_init_screen_get_functions(struct si_screen *sscreen)
        * for gfx10.3 on gfx10.
        */
       .lower_ffma = sscreen->info.chip_class <= GFX9,
-      .fuse_ffma = sscreen->info.chip_class >= GFX10,
+      .fuse_ffma16 = sscreen->info.chip_class >= GFX10,
+      .fuse_ffma32 = sscreen->info.chip_class >= GFX10,
+      .fuse_ffma64 = sscreen->info.chip_class >= GFX10,
       .lower_fmod = true,
       .lower_pack_snorm_4x8 = true,
       .lower_pack_unorm_4x8 = true,
diff --git a/src/panfrost/bifrost/bifrost_compile.h b/src/panfrost/bifrost/bifrost_compile.h
index cc28b42c3af..15b90788133 100644
--- a/src/panfrost/bifrost/bifrost_compile.h
+++ b/src/panfrost/bifrost/bifrost_compile.h
@@ -69,7 +69,9 @@ static const nir_shader_compiler_options bifrost_nir_options = {
 
         .lower_bitfield_extract_to_shifts = true,
         .vectorize_io = true,
-        .fuse_ffma = true,
+	.fuse_ffma16 = true,
+	.fuse_ffma32 = true,
+	.fuse_ffma64 = true,
         .use_interpolated_input_intrinsics = true
 };