util: treat denorm'ed floats like zero

The D3D10 spec is very explicit about treatment of denorm floats and
the behavior is exactly the same for them as it would be for -0 or
+0. This makes our shading code match that behavior, since OpenGL
doesn't care and on a few cpu's it's faster (worst case the same).
Float16 conversions will likely break but we'll fix them in a follow
up commit.

Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>

6 files changed, 83 insertions, 0 deletions

diff --git a/src/gallium/auxiliary/draw/draw_pt.c b/src/gallium/auxiliary/draw/draw_pt.c
index d2fe0025bfc..ccde371ffcd 100644
--- a/src/gallium/auxiliary/draw/draw_pt.c
+++ b/src/gallium/auxiliary/draw/draw_pt.c
@@ -459,8 +459,14 @@ draw_vbo(struct draw_context *draw,
    unsigned instance;
    unsigned index_limit;
    unsigned count;
+   unsigned fpstate = util_fpstate_get();
    struct pipe_draw_info resolved_info;
 
+   /* Make sure that denorms are treated like zeros. This is 
+    * the behavior required by D3D10. OpenGL doesn't care.
+    */
+   util_fpstate_set_denorms_to_zero(fpstate);
+
    resolve_draw_info(info, &resolved_info);
    info = &resolved_info;
 
@@ -518,6 +524,7 @@ draw_vbo(struct draw_context *draw,
       if (index_limit == 0) {
       /* one of the buffers is too small to do any valid drawing */
          debug_warning("draw: VBO too small to draw anything\n");
+         util_fpstate_set(fpstate);
          return;
       }
    }
@@ -558,4 +565,5 @@ draw_vbo(struct draw_context *draw,
    if (draw->collect_statistics) {
       draw->render->pipeline_statistics(draw->render, &draw->statistics);
    }
+   util_fpstate_set(fpstate);
 }
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_arit.c b/src/gallium/auxiliary/gallivm/lp_bld_arit.c
index 08aec799a33..c006ac537c1 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_arit.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_arit.c
@@ -62,6 +62,7 @@
 #include "lp_bld_debug.h"
 #include "lp_bld_bitarit.h"
 #include "lp_bld_arit.h"
+#include "lp_bld_flow.h"
 
 
 #define EXP_POLY_DEGREE 5
diff --git a/src/gallium/auxiliary/util/u_math.c b/src/gallium/auxiliary/util/u_math.c
index 2811475fa00..2487bc74b25 100644
--- a/src/gallium/auxiliary/util/u_math.c
+++ b/src/gallium/auxiliary/util/u_math.c
@@ -28,6 +28,7 @@
 
 
 #include "util/u_math.h"
+#include "util/u_cpu_detect.h"
 
 
 /** 2^x, for x in [-1.0, 1.0) */
@@ -70,4 +71,59 @@ util_init_math(void)
    }
 }
 
+/**
+ * Fetches the contents of the fpstate (mxcsr on x86) register.
+ *
+ * On platforms without support for it just returns 0.
+ */
+unsigned
+util_fpstate_get(void)
+{
+   unsigned mxcsr = 0;
+
+#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
+   if (util_cpu_caps.has_sse) {
+      mxcsr = __builtin_ia32_stmxcsr();
+   }
+#endif
 
+   return mxcsr;
+}
+
+/**
+ * Make sure that the fp treats the denormalized floating
+ * point numbers as zero.
+ *
+ * This is the behavior required by D3D10. OpenGL doesn't care.
+ */
+unsigned
+util_fpstate_set_denorms_to_zero(unsigned current_mxcsr)
+{
+#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
+#define MXCSR_DAZ (1 << 6)	/* Enable denormals are zero mode */
+#define MXCSR_FTZ (1 << 15)	/* Enable flush to zero mode */
+   if (util_cpu_caps.has_sse) {
+      current_mxcsr |= MXCSR_FTZ;
+      if (util_cpu_caps.has_sse3) {
+         current_mxcsr |= MXCSR_DAZ;
+      }
+      util_fpstate_set(current_mxcsr);
+   }
+#endif
+   return current_mxcsr;
+}
+
+/**
+ * Set the state of the fpstate (mxcsr on x86) register.
+ *
+ * On platforms without support for it's a noop.
+ */
+void
+util_fpstate_set(unsigned mxcsr)
+{
+#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
+   if (util_cpu_caps.has_sse) {
+      __builtin_ia32_ldmxcsr(mxcsr);
+   }
+#endif
+}
diff --git a/src/gallium/auxiliary/util/u_math.h b/src/gallium/auxiliary/util/u_math.h
index 64d16cbe715..bc3948875a9 100644
--- a/src/gallium/auxiliary/util/u_math.h
+++ b/src/gallium/auxiliary/util/u_math.h
@@ -763,6 +763,13 @@ static INLINE int32_t util_signed_fixed(float value, unsigned frac_bits)
    return (int32_t)(value * (1<<frac_bits));
 }
 
+unsigned
+util_fpstate_get(void);
+unsigned
+util_fpstate_set_denorms_to_zero(unsigned current_fpstate);
+void
+util_fpstate_set(unsigned fpstate);
+
 
 
 #ifdef __cplusplus
diff --git a/src/gallium/drivers/llvmpipe/lp_rast.c b/src/gallium/drivers/llvmpipe/lp_rast.c
index 871cc50fb4c..e172dac40cb 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast.c
+++ b/src/gallium/drivers/llvmpipe/lp_rast.c
@@ -751,6 +751,12 @@ static PIPE_THREAD_ROUTINE( thread_function, init_data )
    struct lp_rasterizer_task *task = (struct lp_rasterizer_task *) init_data;
    struct lp_rasterizer *rast = task->rast;
    boolean debug = false;
+   unsigned fpstate = util_fpstate_get();
+
+   /* Make sure that denorms are treated like zeros. This is 
+    * the behavior required by D3D10. OpenGL doesn't care.
+    */
+   util_fpstate_set_denorms_to_zero(fpstate);
 
    while (1) {
       /* wait for work */
diff --git a/src/gallium/drivers/llvmpipe/lp_test_main.c b/src/gallium/drivers/llvmpipe/lp_test_main.c
index 4c610923146..8a896becf53 100644
--- a/src/gallium/drivers/llvmpipe/lp_test_main.c
+++ b/src/gallium/drivers/llvmpipe/lp_test_main.c
@@ -370,6 +370,11 @@ int main(int argc, char **argv)
    unsigned i;
    boolean success;
    boolean single = FALSE;
+   unsigned fpstate;
+
+   util_cpu_detect();
+   fpstate = util_fpstate_get();
+   util_fpstate_set_denorms_to_zero(fpstate);
 
    for(i = 1; i < argc; ++i) {
       if(strcmp(argv[i], "-v") == 0)