panfrost: LogicOp fixes and non 8-bit format support

With the previous LogicOp commit almost half of the blend modes were
broken because the surplus bits were not cleared after an inot.

v2:
 - Remove u8 "fast path" as 8-bit is not well optimised yet
 - Don't mask for 32-bit formats as that triggers an assert

Fixes: 068806c9f6b ("panfrost: LogicOp support")
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Tested-by: Marge Bot <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/3943>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/3943>

diff --git a/src/gallium/drivers/panfrost/nir/nir_lower_blend.c b/src/gallium/drivers/panfrost/nir/nir_lower_blend.c
index a2ff37922d9c942a04d97192925837a4fdda5792..99e34e48724aa5d5b4912414f71528affa74c254 100644 (file)
--- a/src/gallium/drivers/panfrost/nir/nir_lower_blend.c
+++ b/src/gallium/drivers/panfrost/nir/nir_lower_blend.c
@@ -32,6 +32,7 @@
 
 #include "compiler/nir/nir.h"
 #include "compiler/nir/nir_builder.h"
+#include "compiler/nir/nir_format_convert.h"
 #include "nir_lower_blend.h"
 
 /* Given processed factors, combine them per a blend function */
@@ -206,18 +207,30 @@ nir_blend_logicop(
    nir_lower_blend_options options,
    nir_ssa_def *src, nir_ssa_def *dst)
 {
-   /* TODO: Support other sizes */
-   nir_ssa_def *factor = nir_imm_float(b, 255);
+   const struct util_format_description *format_desc =
+      util_format_description(options.format);
 
-   src = nir_fmin(b, nir_fmax(b, src, nir_imm_float(b, -1)), nir_imm_float(b, 1));
-   src = nir_f2u32(b, nir_fround_even(b, nir_fmul(b, src, factor)));
+   assert(src->num_components <= 4);
+   assert(dst->num_components <= 4);
 
-   dst = nir_fmin(b, nir_fmax(b, dst, nir_imm_float(b, -1)), nir_imm_float(b, 1));
-   dst = nir_f2u32(b, nir_fround_even(b, nir_fmul(b, dst, factor)));
+   unsigned bits[4];
+   for (int i = 0; i < 4; ++i)
+       bits[i] = format_desc->channel[i].size;
+
+   src = nir_format_float_to_unorm(b, src, bits);
+   dst = nir_format_float_to_unorm(b, dst, bits);
 
    nir_ssa_def *out = nir_logicop_func(b, options.logicop_func, src, dst);
 
-   return nir_fdiv(b, nir_u2f32(b, out), factor);
+   if (bits[0] < 32) {
+       nir_const_value mask[4];
+       for (int i = 0; i < 4; ++i)
+           mask[i] = nir_const_value_for_int((1u << bits[i]) - 1, 32);
+
+       out = nir_iand(b, out, nir_build_imm(b, 4, 32, mask));
+   }
+
+   return nir_format_unorm_to_float(b, out, bits);
 }
 
 /* Given a blend state, the source color, and the destination color,

diff --git a/src/gallium/drivers/panfrost/nir/nir_lower_blend.h b/src/gallium/drivers/panfrost/nir/nir_lower_blend.h
index 99530c3e973b639a970afddea9921dbbff52a80e..b2885d5b5c2c5e5f0e3c87a45b584b277dc50830 100644 (file)
--- a/src/gallium/drivers/panfrost/nir/nir_lower_blend.h
+++ b/src/gallium/drivers/panfrost/nir/nir_lower_blend.h
@@ -51,6 +51,7 @@ typedef struct {
 
    bool logicop_enable;
    unsigned logicop_func;
+   enum pipe_format format;
 } nir_lower_blend_options;
 
 void nir_lower_blend(nir_shader *shader, nir_lower_blend_options options);

diff --git a/src/gallium/drivers/panfrost/pan_blend_shaders.c b/src/gallium/drivers/panfrost/pan_blend_shaders.c
index 0f0e1b5ffd91b91b2f3ae357d6fbaad052a02f4e..a293c6bc3a84fbc117c89ff243d282535ca5d433 100644 (file)
--- a/src/gallium/drivers/panfrost/pan_blend_shaders.c
+++ b/src/gallium/drivers/panfrost/pan_blend_shaders.c
@@ -173,6 +173,8 @@ panfrost_compile_blend_shader(
 
         nir_lower_blend_options options =
                 nir_make_options(cso, rt);
+        options.format = format;
+
         NIR_PASS_V(shader, nir_lower_blend, options);
 
         NIR_PASS_V(shader, nir_lower_framebuffer, format, screen->gpu_id);