glsl/tests/test_gl_lower_mediump.cpp

/*
 * Copyright © 2023 Google LLC
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice (including the next
 * paragraph) shall be included in all copies or substantial portions of the
 * Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
 * IN THE SOFTWARE.
 */

#include <gtest/gtest.h>
#include <gtest/gtest-spi.h>

#include "main/mtypes.h"
#include "standalone_scaffolding.h"
#include "ir.h"
#include "ir_optimization.h"
#include "nir.h"
#include "builtin_functions.h"
#include "nir.h"
#include "gl_nir.h"
#include "gl_nir_linker.h"
#include "glsl_parser_extras.h"
#include "glsl_to_nir.h"
#include "linker_util.h"
#include "nir_builder.h"

/* The printed-GLSL-IR tests use fmemopen so we can do stdio to memory (or you'd
 * need equivalent tempfiles that you manage).  Just disable this test on those
 * platforms (aka Windows).
 */
#ifdef HAVE_FMEMOPEN

namespace
{
   class gl_nir_lower_mediump_test : public ::testing::Test
   {
   protected:
      gl_nir_lower_mediump_test();
      ~gl_nir_lower_mediump_test();

      struct gl_shader *compile_shader(GLenum type, const char *source);
      void compile(const char *source);

      struct gl_context local_ctx;
      struct gl_context *ctx;

      nir_alu_instr *find_op(nir_op op)
      {
         if (!nir)
            return NULL;

         nir_foreach_function_impl(impl, nir)
         {
            nir_foreach_block(block, impl)
            {
               nir_foreach_instr(instr, block)
               {
                  if (instr->type == nir_instr_type_alu)
                  {
                     nir_alu_instr *alu = nir_instr_as_alu(instr);
                     if (alu->op == op)
                        return alu;
                  }
               }
            }
         }
         return NULL;
      }

      uint32_t op_dest_bits(nir_op op)
      {
         nir_alu_instr *alu = find_op(op);
         EXPECT_TRUE(alu != NULL);
         return alu->def.bit_size;
      }

      /* Returns the common bit size of all src operands (failing if not matching). */
      uint32_t op_src_bits(nir_op op)
      {
         nir_alu_instr *alu = find_op(op);
         EXPECT_TRUE(alu != NULL);

         for (int i = 0; i < nir_op_infos[op].num_inputs; i++) {
            EXPECT_EQ(alu->src[i].src.ssa->bit_size, alu->src[0].src.ssa->bit_size);
         }
         return alu->src[0].src.ssa->bit_size;
      }

      nir_shader *nir;
      struct gl_shader_program *whole_program;
      const char *source;
      char *fs_ir;
   };

   gl_nir_lower_mediump_test::gl_nir_lower_mediump_test()
       : nir(NULL), source(NULL), fs_ir(NULL)
   {
      glsl_type_singleton_init_or_ref();
   }

   gl_nir_lower_mediump_test::~gl_nir_lower_mediump_test()
   {
      if (HasFailure())
      {
         if (source)
            printf("\nSource for the failed test:\n%s\n", source);
         if (fs_ir) {
            printf("\nGLSL IR from the failed test:\n\n");
            printf("%s", fs_ir);

         }
         if (nir) {
            printf("\nNIR from the failed test:\n\n");
            nir_print_shader(nir, stdout);
         }
      }

      standalone_destroy_shader_program(whole_program);

      free(fs_ir);

      glsl_type_singleton_decref();
   }

   struct gl_shader *
   gl_nir_lower_mediump_test::compile_shader(GLenum type, const char *source)
   {
      struct gl_shader *shader = standalone_add_shader_source(ctx, whole_program, type, source);

      /* Save off the GLSL IR, since the compile frees it. */
      char temp[4096];
      FILE *ftemp = NULL;
      if (type == GL_FRAGMENT_SHADER)
         ftemp = fmemopen(temp, sizeof(temp), "w");

      _mesa_glsl_compile_shader(ctx, shader, ftemp, false, false, true);

      if (type == GL_FRAGMENT_SHADER) {
         fclose(ftemp);
         fs_ir = strdup(temp);
      }

      return shader;
   }

   void
   gl_nir_lower_mediump_test::compile(const char *source)
   {
      ctx = &local_ctx;

      static const struct nir_shader_compiler_options compiler_options = {
          .support_16bit_alu = true,
      };

      /* Get better variable names from GLSL IR for debugging. */
      ir_variable::temporaries_allocate_names = true;

      initialize_context_to_defaults(ctx, API_OPENGLES2);
      ctx->Version = 31;
      for (int i = 0; i < MESA_SHADER_STAGES; i++) {
         ctx->Const.ShaderCompilerOptions[i].LowerPrecisionFloat16 = true;
         ctx->Const.ShaderCompilerOptions[i].LowerPrecisionInt16 = true;
         ctx->Const.ShaderCompilerOptions[i].NirOptions = &compiler_options;
      }

      /* GL_ARB_explicit_uniform_location, GL_MAX_UNIFORM_LOCATIONS */
      ctx->Const.MaxUserAssignableUniformLocations =
         4 * MESA_SHADER_STAGES * MAX_UNIFORMS;

      ctx->Const.Program[MESA_SHADER_VERTEX].MaxCombinedUniformComponents = 128 * 4;
      ctx->Const.Program[MESA_SHADER_FRAGMENT].MaxCombinedUniformComponents = 16 * 4;

      _mesa_glsl_builtin_functions_init_or_ref();

      whole_program = standalone_create_shader_program();
      whole_program->IsES = true;

      const char *vs_source = R"(#version 310 es
      void main() {
         gl_Position = vec4(0.0);
      })";
      compile_shader(GL_VERTEX_SHADER, vs_source);

      compile_shader(GL_FRAGMENT_SHADER, source);

      for (unsigned i = 0; i < whole_program->NumShaders; i++)
      {
         struct gl_shader *shader = whole_program->Shaders[i];
         if (shader->CompileStatus != COMPILE_SUCCESS)
            fprintf(stderr, "Compiler error: %s", shader->InfoLog);
         ASSERT_EQ(shader->CompileStatus, COMPILE_SUCCESS);
      }

      link_shaders_init(ctx, whole_program);
      gl_nir_link_glsl(ctx, whole_program);
      if (whole_program->data->LinkStatus != LINKING_SUCCESS)
         fprintf(stderr, "Linker error: %s", whole_program->data->InfoLog);
      EXPECT_EQ(whole_program->data->LinkStatus, LINKING_SUCCESS);

      nir = whole_program->_LinkedShaders[MESA_SHADER_FRAGMENT]->Program->nir;

      /* Store the source for printing from later assertions. */
      this->source = source;
   }

   // A predicate-formatter for asserting that two integers are mutually prime.
   testing::AssertionResult glsl_ir_contains(const char *glsl_ir_expr,
                                             const char *needle_expr,
                                             const char *glsl_ir,
                                             const char *needle)
   {
      /* If we didn't HAVE_FMEMOPEN, we won't have GLSL IR to look at.  Just
       * skip those parts of the tests on such platforms.
       */
      if (!glsl_ir)
         return testing::AssertionSuccess();

      if (strstr(glsl_ir, needle))
         return testing::AssertionSuccess();

      return testing::AssertionFailure() << " " << needle_expr << " not found in GLSL IR";
   }
} // namespace

TEST_F(gl_nir_lower_mediump_test, float_simple_mul)
{
   ASSERT_NO_FATAL_FAILURE(compile(
       R"(#version 310 es
         uniform mediump float a, b;
         out mediump float result;

         void main()
         {
            result = a * b;
         }
    )"));

   EXPECT_EQ(op_dest_bits(nir_op_fmul), 16);
}

TEST_F(gl_nir_lower_mediump_test, int_simple_mul)
{
   ASSERT_NO_FATAL_FAILURE(compile(
       R"(#version 310 es
         precision highp float;
         precision mediump int;
         uniform mediump int a, b;
         out mediump int result;

         void main()
         {
            result = a * b;
         }
    )"));

   EXPECT_EQ(op_dest_bits(nir_op_imul), 16);
}

TEST_F(gl_nir_lower_mediump_test, int_default_precision_med)
{
   ASSERT_NO_FATAL_FAILURE(compile(
       R"(#version 310 es
         precision highp float;
         precision mediump int;
         uniform int a, b;
         out int result;

         void main()
         {
            result = a * b;
         }
    )"));

   EXPECT_EQ(op_dest_bits(nir_op_imul), 16);
}

TEST_F(gl_nir_lower_mediump_test, int_default_precision_high)
{
   ASSERT_NO_FATAL_FAILURE(compile(
       R"(#version 310 es
         precision mediump float;
         precision highp int;
         uniform int a, b;
         out int result;

         void main()
         {
            result = a * b;
         }
    )"));

   EXPECT_EQ(op_dest_bits(nir_op_imul), 32);
}

/* Test that a builtin with mediump args does mediump computation. */
TEST_F(gl_nir_lower_mediump_test, dot_builtin)
{
   ASSERT_NO_FATAL_FAILURE(compile(
       R"(#version 310 es
         precision highp float;
         precision highp int;
         uniform mediump vec4 a, b;
         out float result;

         void main()
         {
            result = dot(a, b);
         }
    )"));

   EXPECT_EQ(op_dest_bits(nir_op_fdot4), 16);
}

/* Test that a constant-index array deref is mediump */
TEST_F(gl_nir_lower_mediump_test, array_const_index)
{
   ASSERT_NO_FATAL_FAILURE(compile(
       R"(#version 310 es
         precision highp float;
         precision highp int;
         uniform mediump float a, b[2];
         out float result;

         void main()
         {
            result = a * b[1];
         }
    )"));

   EXPECT_EQ(op_dest_bits(nir_op_fmul), 16);
}

/* Test that a variable-index array deref is mediump, even if the array index is highp */
TEST_F(gl_nir_lower_mediump_test, array_uniform_index)
{
   ASSERT_NO_FATAL_FAILURE(compile(
       R"(#version 310 es
         precision highp float;
         uniform mediump float a, b[2];
         uniform highp int i;
         out float result;

         void main()
         {
            result = a * b[i];
         }
    )"));

   EXPECT_EQ(op_dest_bits(nir_op_fmul), 16);
}

/* Test that a variable-index array deref is highp, even if the array index is mediump */
TEST_F(gl_nir_lower_mediump_test, array_mediump_index)
{
   ASSERT_NO_FATAL_FAILURE(compile(
       R"(#version 310 es
         precision highp float;
         uniform highp int b[2];
         uniform mediump int a, i;
         out highp int result;

         void main()
         {
            result = a * b[i];
         }
    )"));

   EXPECT_PRED_FORMAT2(glsl_ir_contains, fs_ir, "expression int * ");

   EXPECT_EQ(op_dest_bits(nir_op_imul), 32);
}

TEST_F(gl_nir_lower_mediump_test, func_return)
{
   ASSERT_NO_FATAL_FAILURE(compile(
       R"(#version 310 es
         precision highp float; /* Make sure that default highp temps in function handling don't break our mediump return. */
         uniform mediump float a;
         uniform highp float b;
         out float result;

         mediump float func()
         {
            return b; /* Returning highp b here, but it should be the mediump return value qualifier that matters */
         }

         void main()
         {
            /* "If a function returns a value, then a call to that function may
             *  be used as an expression, whose type will be the type that was
             *  used to declare or define the function."
             */
            result = a * func();
         }
    )"));

   EXPECT_PRED_FORMAT2(glsl_ir_contains, fs_ir, "expression float16_t * ");

   EXPECT_EQ(op_dest_bits(nir_op_fmul), 16);
}

TEST_F(gl_nir_lower_mediump_test, func_args_in_mediump)
{
   ASSERT_NO_FATAL_FAILURE(compile(
       R"(#version 310 es
         precision highp float; /* Make sure that default highp temps in function handling don't break our mediump return. */
         uniform highp float a, b;
         out highp float result;

         highp float func(mediump float x, mediump float y)
         {
            return x * y; /* should be mediump due to x and y, but propagating qualifiers from a,b by inlining could trick it. */
         }

         void main()
         {
            result = func(a, b);
         }
    )"));

   EXPECT_PRED_FORMAT2(glsl_ir_contains, fs_ir, "expression float f162f (expression float16_t * (expression float16_t f2fmp (var_ref x) ) (expression float16_t f2fmp (var_ref y) ) )");

   EXPECT_EQ(op_dest_bits(nir_op_fmul), 16);
}

TEST_F(gl_nir_lower_mediump_test, func_args_inout_mediump)
{
   ASSERT_NO_FATAL_FAILURE(compile(
       R"(#version 310 es
         precision highp float; /* Make sure that default highp temps in function handling don't break our mediump inout. */
         uniform highp float a, b;
         out float result;

         void func(inout mediump float x, mediump float y)
         {
            x = x * y; /* should be mediump due to x and y, but propagating qualifiers from a,b by inlining could trick it. */
         }

         void main()
         {
            /* The spec says "function input and output is done through copies,
             * and therefore qualifiers do not have to match."  So we use a
             * highp here for our mediump inout.
             */
            highp float x = a;
            func(x, b);
            result = x;
         }
    )"));

   EXPECT_PRED_FORMAT2(glsl_ir_contains, fs_ir, "expression float16_t * ");

   EXPECT_EQ(op_dest_bits(nir_op_fmul), 16);
}

TEST_F(gl_nir_lower_mediump_test, func_args_in_out_mediump)
{
   ASSERT_NO_FATAL_FAILURE(compile(
       R"(#version 310 es
         precision highp float; /* Make sure that default highp temps in function handling don't break our mediump inout. */
         uniform highp float a, b;
         out float result;

         void func(mediump float x, mediump float y, out mediump float w)
         {
            w = x * y; /* should be mediump due to x and y, but propagating qualifiers from a,b by inlining could trick it. */
         }

         void main()
         {
            /* The spec says "function input and output is done through copies,
             * and therefore qualifiers do not have to match."  So we use a
             * highp here for our mediump out.
             */
            highp float x;
            func(a, b, x);
            result = x;
         }
    )"));

   EXPECT_PRED_FORMAT2(glsl_ir_contains, fs_ir, "expression float16_t * ");

   EXPECT_EQ(op_dest_bits(nir_op_fmul), 16);
}

TEST_F(gl_nir_lower_mediump_test, func_args_inout_highp)
{
   ASSERT_NO_FATAL_FAILURE(compile(
       R"(#version 310 es
         precision mediump float; /* Make sure that default mediump temps in function handling don't break our highp inout. */
         uniform mediump float a, b;
         out float result;

         void func(inout highp float x, highp float y)
         {
            x = x * y; /* should be highp due to x and y, but propagating qualifiers from a,b by inlining could trick it. */
         }

         void main()
         {
            mediump float x = a;
            func(x, b);
            result = x;
         }
    )"));

   EXPECT_PRED_FORMAT2(glsl_ir_contains, fs_ir, "expression float * ");

   EXPECT_EQ(op_dest_bits(nir_op_fmul), 32);
}

TEST_F(gl_nir_lower_mediump_test, if_mediump)
{
   ASSERT_NO_FATAL_FAILURE(compile(
       R"(#version 310 es
         precision highp float;
         uniform mediump float a, b, c;
         out float result;

         void main()
         {
            if (a * b < c)
               result = 1.0;
            else
               result = 0.0;
         }
    )"));

   EXPECT_EQ(op_dest_bits(nir_op_fmul), 16);
   EXPECT_EQ(op_src_bits(nir_op_flt), 16);
}

TEST_F(gl_nir_lower_mediump_test, mat_mul_mediump)
{
   ASSERT_NO_FATAL_FAILURE(compile(
       R"(#version 310 es
         precision highp float;
         uniform mediump mat2 a;
         uniform mediump vec2 b;
         out highp vec2 result;

         void main()
         {
            result = a * b;
         }
    )"));

   EXPECT_EQ(op_dest_bits(nir_op_fmul), 16);
}

TEST_F(gl_nir_lower_mediump_test, struct_default_precision_lvalue)
{
   ASSERT_NO_FATAL_FAILURE(compile(
       R"(#version 310 es
         precision highp float;
         precision mediump int;
         struct S {
            float x, y;
            int z, w;
         };
         uniform S a;
         out mediump vec2 result;

         void main()
         {
            /* I believe that structure members don't have a precision
             * qualifier, so we expect the precision of these operations to come
             * from the lvalue (which is higher precedence than the default
             * precision).
             */
            mediump float resultf = a.x * a.y;
            highp int resulti = a.z * a.w;
            result = vec2(resultf, float(resulti));
         }
    )"));

   /* GLSL fails to implement this correctly. */
   EXPECT_NONFATAL_FAILURE(
       EXPECT_PRED_FORMAT2(glsl_ir_contains, fs_ir,
                           "expression float16_t * (record_ref (var_ref a)  x) (record_ref (var_ref a)  y) "),
       "not found in GLSL IR");
   EXPECT_NONFATAL_FAILURE(
       EXPECT_PRED_FORMAT2(glsl_ir_contains, fs_ir,
                           "expression int * (record_ref (var_ref a)  z) (record_ref (var_ref a)  w) "),
       "not found in GLSL IR");

   // Enable these checks once we fix the GLSL.
   //EXPECT_EQ(op_dest_bits(nir_op_fmul), 16);
   //EXPECT_EQ(op_dest_bits(nir_op_imul), 32);
}

TEST_F(gl_nir_lower_mediump_test, float_constructor)
{
   ASSERT_NO_FATAL_FAILURE(compile(
       R"(#version 310 es
         precision mediump float;
         uniform highp uint a;
         uniform mediump float b;
         out mediump float result;

         void main()
         {
            /* It's tricky to reconcile these two bits of spec: "Literal
             * constants do not have precision qualifiers. Neither do Boolean
             * variables. Neither do constructors."
             *
             * and
             *
             * "For this paragraph, “operation” includes operators, built-in
             * functions, and constructors, and “operand” includes function
             * arguments and constructor arguments."
             *
             * I take this to mean that the language doesn't let you put a
             * precision qualifier on a constructor (or literal), but the
             * constructor operation gets precision qualification inference
             * based on its args like normal.
             */
            result = float(a) * b;
         }
    )"));

   EXPECT_EQ(op_dest_bits(nir_op_fmul), 32);
}

TEST_F(gl_nir_lower_mediump_test, vec2_constructor)
{
   ASSERT_NO_FATAL_FAILURE(compile(
       R"(#version 310 es
         precision mediump float;
         uniform highp float a, b;
         uniform mediump float c;
         out mediump vec2 result;

         void main()
         {
            result = c * vec2(a, b);
         }
    )"));

   EXPECT_EQ(op_dest_bits(nir_op_fmul), 32);
}
TEST_F(gl_nir_lower_mediump_test, vec4_of_float_constructor)
{
   ASSERT_NO_FATAL_FAILURE(compile(
       R"(#version 310 es
         precision mediump float;
         uniform highp float a;
         uniform mediump float b;
         out mediump vec4 result;

         void main()
         {
            result = b * vec4(a);
         }
    )"));

   EXPECT_EQ(op_dest_bits(nir_op_fmul), 32);
}

TEST_F(gl_nir_lower_mediump_test, vec4_of_vec2_constructor)
{
   ASSERT_NO_FATAL_FAILURE(compile(
       R"(#version 310 es
         precision mediump float;
         uniform highp vec2 a, b;
         uniform mediump vec4 c;
         out mediump vec4 result;

         void main()
         {
            /* GLSL IR has to either have a temp for a*b, or clone the
             * expression and let it get CSEed later.  If it chooses temp, that
             * may confuse us.
             */
            result = c + vec4(a * b, 0.0, 0.0);
         }
    )"));

   EXPECT_EQ(op_dest_bits(nir_op_fmul), 32);
   EXPECT_EQ(op_dest_bits(nir_op_fadd), 32);
}

TEST_F(gl_nir_lower_mediump_test, float_literal_mediump)
{
   ASSERT_NO_FATAL_FAILURE(compile(
       R"(#version 310 es
         precision highp float;
         uniform mediump float a;
         out highp float result;

         void main()
         {
            /* The literal is unqualified, so it shouldn't promote the expression to highp. */
            result = a * 2.0;
         }
    )"));

   EXPECT_EQ(op_dest_bits(nir_op_fmul), 16);
}

TEST_F(gl_nir_lower_mediump_test, float_const_highp)
{
   ASSERT_NO_FATAL_FAILURE(compile(
       R"(#version 310 es
         precision highp float;
         uniform mediump float a;
         out highp float result;

         void main()
         {
            highp float two = 2.0;
            /* The constant is highp, so even with constant propagation the expression should be highp. */
            result = a * two;
         }
    )"));

   EXPECT_EQ(op_dest_bits(nir_op_fmul), 32);
}

TEST_F(gl_nir_lower_mediump_test, float_const_expr_mediump)
{
   ASSERT_NO_FATAL_FAILURE(compile(
       R"(#version 310 es
         precision highp float;
         uniform mediump float a;
         out highp float result;

         void main()
         {
            /* "Where the precision of a constant integral or constant floating
             * point expression is not specified, evaluation is performed at
             * highp. This rule does not affect the precision qualification of the
             * expression."
             * So the 5.0 is calculated at highp, but a * 5.0 is calculated at mediump.
             */
            result = a * (2.0 + 3.0);
         }
    )"));

   EXPECT_EQ(op_dest_bits(nir_op_fmul), 16);
}

TEST_F(gl_nir_lower_mediump_test, unpackUnorm4x8)
{
   ASSERT_NO_FATAL_FAILURE(compile(
       R"(#version 310 es
         precision highp float;
         uniform highp uint a;
         uniform mediump float b;
         out highp float result;

         void main()
         {
            result = unpackUnorm4x8(a).x * b;
         }
    )"));

   /* XXX: GLSL doesn't lower this one correctly, currently.  It returns highp despite the prototype being mediump. */
   EXPECT_NONFATAL_FAILURE(
      EXPECT_PRED_FORMAT2(glsl_ir_contains, fs_ir, "expression f16vec4 unpackUnorm4x8 (var_ref a"),
      "not found in GLSL IR");
   EXPECT_PRED_FORMAT2(glsl_ir_contains, fs_ir, "expression float16_t *");

   /* XXX: NIR insists that nir_op_unpack_unorm_4x8 returns 32 bits per channel, too. */
   EXPECT_NONFATAL_FAILURE(
      EXPECT_EQ(op_dest_bits(nir_op_unpack_unorm_4x8), 16),
      "op_dest_bits");
   EXPECT_EQ(op_dest_bits(nir_op_fmul), 16);
}

TEST_F(gl_nir_lower_mediump_test, packUnorm4x8)
{
   ASSERT_NO_FATAL_FAILURE(compile(
       R"(#version 310 es
         precision highp float;
         uniform mediump vec4 a;
         uniform mediump uint b;
         out highp uint result;

         void main()
         {
            result = packUnorm4x8(a) & b;
         }
    )"));

   /* Test both the GLSL IR return value and an op using it with a mediump
    * value, so we can be sure it's not just that we're assigning to highp.
    */
   EXPECT_PRED_FORMAT2(glsl_ir_contains, fs_ir, "expression uint packUnorm4x8 (var_ref a)");
   EXPECT_PRED_FORMAT2(glsl_ir_contains, fs_ir, "expression uint &");

   EXPECT_EQ(op_dest_bits(nir_op_pack_unorm_4x8), 32);
}

/* XXX: Add unit tests getting at precision of temporaries inside builtin function impls. */
/* XXX: Add unit tests getting at precision of any other temps internally generated by the compiler */
/* XXX: Add unit tests checking for default precision on user-declared function temps*/

#endif /* HAVE_FMEMOPEN */