i965: Drop support for the legacy SNORM -> Float equation.
authorKenneth Graunke <kenneth@whitecape.org>
Tue, 26 Dec 2017 03:10:22 +0000 (19:10 -0800)
committerKenneth Graunke <kenneth@whitecape.org>
Wed, 3 Jan 2018 00:51:42 +0000 (16:51 -0800)
Older OpenGL defines two equations for converting from signed-normalized
to floating point data.  These are:

    f = (2c + 1)/(2^b - 1)                (equation 2.2)
    f = max{c/2^(b-1) - 1), -1.0}         (equation 2.3)

Both OpenGL 4.2+ and OpenGL ES 3.0+ mandate that equation 2.3 is to be
used in all scenarios, and remove equation 2.2.  DirectX uses equation
2.3 as well.  Intel hardware only supports equation 2.3, so Gen7.5+
systems that use the vertex fetcher hardware to do the conversions
always get formula 2.3.

This can make a big difference for 10-10-10-2 formats - the 2-bit value
can represent 0 with equation 2.3, and cannot with equation 2.2.

Ivybridge and older were using equation 2.2 for OpenGL, and 2.3 for ES.
Now that Ivybridge supports OpenGL 4.2, this is wrong - we need to use
the new rules, at least in core profile.  That would leave Gen4-6 doing
something different than all other hardware, which seems...lame.

With context version promotion, applications that requested a pre-4.2
context may get promoted to 4.2, and thus get the new rules.  Zero cases
have been reported of this being a problem.  However, we've received a
report that following the old rules breaks expectations.  SuperTuxKart
apparently renders the cars red when following equation 2.2, and works
correctly when following equation 2.3:

https://github.com/supertuxkart/stk-code/issues/2885#issuecomment-353858405

So, this patch deletes the legacy equation 2.2 support entirely, making
all hardware and APIs consistently use the new equation 2.3 rules.

If we ever find an application that truly requires the old formula, then
we'd likely want that application to work on modern hardware, too.  We'd
likely restore this support as a driconf option.  Until then, drop it.

This commit will regress Piglit's draw-vertices-2101010 test on
pre-Haswell without the corresponding Piglit patch to accept either
formula (commit 35daaa1695ea01eb85bc02f9be9b6ebd1a7113a1):

    draw-vertices-2101010: Accept either SNORM conversion formula.

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Chris Forbes <chrisforbes@google.com>
src/intel/blorp/blorp.c
src/intel/compiler/brw_compiler.h
src/intel/compiler/brw_nir.c
src/intel/compiler/brw_nir.h
src/intel/compiler/brw_nir_attribute_workarounds.c
src/intel/compiler/brw_vec4.cpp
src/intel/compiler/brw_vec4_vs.h
src/intel/compiler/brw_vec4_vs_visitor.cpp
src/intel/vulkan/anv_pipeline.c
src/mesa/drivers/dri/i965/brw_vs.c

index 8a9d2fd3b97e1e6fa12652d4e0fdfffec5ab8547..e8a2c6135f54b25933675889d9e17162914f913e 100644 (file)
@@ -223,8 +223,7 @@ blorp_compile_vs(struct blorp_context *blorp, void *mem_ctx,
 
    const unsigned *program =
       brw_compile_vs(compiler, blorp->driver_ctx, mem_ctx,
-                     &vs_key, vs_prog_data, nir,
-                     false, -1, NULL);
+                     &vs_key, vs_prog_data, nir, -1, NULL);
 
    return program;
 }
index 28aed833245aa6feb0176d03804efe5d4b18332e..0060c381c0d4a12cab32d81aca13114abd029726 100644 (file)
@@ -1123,7 +1123,6 @@ brw_compile_vs(const struct brw_compiler *compiler, void *log_data,
                const struct brw_vs_prog_key *key,
                struct brw_vs_prog_data *prog_data,
                const struct nir_shader *shader,
-               bool use_legacy_snorm_formula,
                int shader_time_index,
                char **error_str);
 
index 265c63efddaa10ce700ea9d5cfeb628827ddc2ff..dbddef0d04d93ef3f5d2a5e9372b82aa8606110c 100644 (file)
@@ -211,7 +211,6 @@ remap_patch_urb_offsets(nir_block *block, nir_builder *b,
 
 void
 brw_nir_lower_vs_inputs(nir_shader *nir,
-                        bool use_legacy_snorm_formula,
                         const uint8_t *vs_attrib_wa_flags)
 {
    /* Start with the location of the variable's base. */
@@ -230,8 +229,7 @@ brw_nir_lower_vs_inputs(nir_shader *nir,
 
    add_const_offset_to_base(nir, nir_var_shader_in);
 
-   brw_nir_apply_attribute_workarounds(nir, use_legacy_snorm_formula,
-                                       vs_attrib_wa_flags);
+   brw_nir_apply_attribute_workarounds(nir, vs_attrib_wa_flags);
 
    /* The last step is to remap VERT_ATTRIB_* to actual registers */
 
index e9cb6f89948186050f81a529f0c84d1e4cd5724c..809d4c338dc4f0afdffe72243fd7b9f7d2dbdb62 100644 (file)
@@ -102,7 +102,6 @@ brw_nir_link_shaders(const struct brw_compiler *compiler,
 bool brw_nir_lower_cs_intrinsics(nir_shader *nir,
                                  unsigned dispatch_width);
 void brw_nir_lower_vs_inputs(nir_shader *nir,
-                             bool use_legacy_snorm_formula,
                              const uint8_t *vs_attrib_wa_flags);
 void brw_nir_lower_vue_inputs(nir_shader *nir,
                               const struct brw_vue_map *vue_map);
@@ -121,7 +120,6 @@ nir_shader *brw_postprocess_nir(nir_shader *nir,
                                 bool is_scalar);
 
 bool brw_nir_apply_attribute_workarounds(nir_shader *nir,
-                                         bool use_legacy_snorm_formula,
                                          const uint8_t *attrib_wa_flags);
 
 bool brw_nir_apply_trig_workarounds(nir_shader *nir);
index c719371ddf199ff661ef620b1a25381f58a24bcb..55a3b3e454696154973485a65fe90f18c557f793 100644 (file)
@@ -33,7 +33,6 @@
 struct attr_wa_state {
    nir_builder builder;
    bool impl_progress;
-   bool use_legacy_snorm_formula;
    const uint8_t *wa_flags;
 };
 
@@ -88,12 +87,15 @@ apply_attr_wa_block(nir_block *block, struct attr_wa_state *state)
          /* ES 3.0 has different rules for converting signed normalized
           * fixed-point numbers than desktop GL.
           */
-         if ((wa_flags & BRW_ATTRIB_WA_SIGN) &&
-             !state->use_legacy_snorm_formula) {
+         if (wa_flags & BRW_ATTRIB_WA_SIGN) {
             /* According to equation 2.2 of the ES 3.0 specification,
              * signed normalization conversion is done by:
              *
              * f = c / (2^(b-1)-1)
+             *
+             * OpenGL 4.2+ uses this equation as well.  Since most contexts
+             * promote to the new higher version, and this is what Haswell+
+             * hardware does anyway, we just always use this formula.
              */
             nir_ssa_def *es3_normalize_factor =
                nir_imm_vec4(b, 1.0f / ((1 << 9) - 1), 1.0f / ((1 << 9) - 1),
@@ -102,31 +104,16 @@ apply_attr_wa_block(nir_block *block, struct attr_wa_state *state)
                            nir_fmul(b, nir_i2f32(b, val), es3_normalize_factor),
                            nir_imm_float(b, -1.0f));
          } else {
-            /* The following equations are from the OpenGL 3.2 specification:
+            /* The following equation is from the OpenGL 3.2 specification:
              *
              * 2.1 unsigned normalization
              * f = c/(2^n-1)
-             *
-             * 2.2 signed normalization
-             * f = (2c+1)/(2^n-1)
-             *
-             * Both of these share a common divisor, which we handle by
-             * multiplying by 1 / (2^b - 1) for b = <10, 10, 10, 2>.
              */
             nir_ssa_def *normalize_factor =
                nir_imm_vec4(b, 1.0f / ((1 << 10) - 1), 1.0f / ((1 << 10) - 1),
                                1.0f / ((1 << 10) - 1), 1.0f / ((1 << 2)  - 1));
 
-            if (wa_flags & BRW_ATTRIB_WA_SIGN) {
-               /* For signed normalization, the numerator is 2c+1. */
-               nir_ssa_def *two = nir_imm_float(b, 2.0f);
-               nir_ssa_def *one = nir_imm_float(b, 1.0f);
-               val = nir_fadd(b, nir_fmul(b, nir_i2f32(b, val), two), one);
-            } else {
-               /* For unsigned normalization, the numerator is just c. */
-               val = nir_u2f32(b, val);
-            }
-            val = nir_fmul(b, val, normalize_factor);
+            val = nir_fmul(b, nir_u2f32(b, val), normalize_factor);
          }
       }
 
@@ -145,12 +132,10 @@ apply_attr_wa_block(nir_block *block, struct attr_wa_state *state)
 
 bool
 brw_nir_apply_attribute_workarounds(nir_shader *shader,
-                                    bool use_legacy_snorm_formula,
                                     const uint8_t *attrib_wa_flags)
 {
    bool progress = false;
    struct attr_wa_state state = {
-      .use_legacy_snorm_formula = use_legacy_snorm_formula,
       .wa_flags = attrib_wa_flags,
    };
 
index 3ddbe6c57fcc181720dc504fdaa2e48f3d7cea2e..ad6d8f9d6bc10670ac01a00460e278664d1505ba 100644 (file)
@@ -2744,7 +2744,6 @@ brw_compile_vs(const struct brw_compiler *compiler, void *log_data,
                const struct brw_vs_prog_key *key,
                struct brw_vs_prog_data *prog_data,
                const nir_shader *src_shader,
-               bool use_legacy_snorm_formula,
                int shader_time_index,
                char **error_str)
 {
@@ -2772,8 +2771,7 @@ brw_compile_vs(const struct brw_compiler *compiler, void *log_data,
    prog_data->inputs_read = shader->info.inputs_read;
    prog_data->double_inputs_read = shader->info.double_inputs_read;
 
-   brw_nir_lower_vs_inputs(shader, use_legacy_snorm_formula,
-                           key->gl_attrib_wa_flags);
+   brw_nir_lower_vs_inputs(shader, key->gl_attrib_wa_flags);
    brw_nir_lower_vue_outputs(shader, is_scalar);
    shader = brw_postprocess_nir(shader, compiler, is_scalar);
 
@@ -2891,8 +2889,7 @@ brw_compile_vs(const struct brw_compiler *compiler, void *log_data,
       prog_data->base.dispatch_mode = DISPATCH_MODE_4X2_DUAL_OBJECT;
 
       vec4_vs_visitor v(compiler, log_data, key, prog_data,
-                        shader, mem_ctx,
-                        shader_time_index, use_legacy_snorm_formula);
+                        shader, mem_ctx, shader_time_index);
       if (!v.run()) {
          if (error_str)
             *error_str = ralloc_strdup(mem_ctx, v.fail_msg);
index b2a862fdbde3f90e064dd6f32e67be7f24a72a37..b62e03aa8d9fc0acba428453dc201c3aef47d691 100644 (file)
@@ -37,8 +37,7 @@ public:
                    struct brw_vs_prog_data *vs_prog_data,
                    const nir_shader *shader,
                    void *mem_ctx,
-                   int shader_time_index,
-                   bool use_legacy_snorm_formula);
+                   int shader_time_index);
 
 protected:
    virtual void setup_payload();
@@ -55,8 +54,6 @@ private:
 
    const struct brw_vs_prog_key *const key;
    struct brw_vs_prog_data * const vs_prog_data;
-
-   bool use_legacy_snorm_formula;
 };
 
 } /* namespace brw */
index 4d8ae23b0c73d25e9fdd76a3bb636dc0afaeba71..8f15bc30a7c5a91f4cf89a3b662bb1c56414ba97 100644 (file)
@@ -172,13 +172,11 @@ vec4_vs_visitor::vec4_vs_visitor(const struct brw_compiler *compiler,
                                  struct brw_vs_prog_data *vs_prog_data,
                                  const nir_shader *shader,
                                  void *mem_ctx,
-                                 int shader_time_index,
-                                 bool use_legacy_snorm_formula)
+                                 int shader_time_index)
    : vec4_visitor(compiler, log_data, &key->tex, &vs_prog_data->base, shader,
                   mem_ctx, false /* no_spills */, shader_time_index),
      key(key),
-     vs_prog_data(vs_prog_data),
-     use_legacy_snorm_formula(use_legacy_snorm_formula)
+     vs_prog_data(vs_prog_data)
 {
 }
 
index 407f29701991ab997e7c86690a014bdcc6d0c33d..4e66f8665fa8824b450ae9a26fb27ae15a88f37a 100644 (file)
@@ -545,7 +545,7 @@ anv_pipeline_compile_vs(struct anv_pipeline *pipeline,
 
       const unsigned *shader_code =
          brw_compile_vs(compiler, NULL, mem_ctx, &key, &prog_data, nir,
-                        false, -1, NULL);
+                        -1, NULL);
       if (shader_code == NULL) {
          ralloc_free(mem_ctx);
          return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
index a56b256bc3887a64a33cde4a401485b23e694872..985d0bf04f05fe3964ec2bebea8c072b8aa2be15 100644 (file)
@@ -221,7 +221,6 @@ brw_codegen_vs_prog(struct brw_context *brw,
    char *error_str;
    program = brw_compile_vs(compiler, brw, mem_ctx, key, &prog_data,
                             vp->program.nir,
-                            !_mesa_is_gles3(&brw->ctx),
                             st_index, &error_str);
    if (program == NULL) {
       if (!vp->program.is_arb_asm) {