Because samplers can be stored within structures, register
allocation of entire structures within the samplers registers
was pushing sampler indices outside of the allowed limit.
In order to solve this, sampler registers now exclusively
contain samplers, and utility functions to compute sampler
only type size were added to make this possible.

Bug chromium:797264

Change-Id: Ic5a6f09665c39661944444cd736547bce4dff2ab
Reviewed-on: https://swiftshader-review.googlesource.com/15728Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

This also improves GetProcAddress performance by using binary search
instead of linear search.

Fixes bug b/20110899

Change-Id: I6c58e17f0580904338e4d806e310cccbec398f28
Reviewed-on: https://swiftshader-review.googlesource.com/15748Tested-by: Krzysztof Kosiński <krzysio@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

Change-Id: I25567211a4603713ad89a1eb58aa3306b11be0ab
Reviewed-on: https://swiftshader-review.googlesource.com/15689Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

-march=core2 implies support for both the SSE3 and SSSE3 instruction
set extensions. Chrome should run on CPUs with only SSE2 as well.

Also, make use of SSE on x86-32 for floating-point operations, to be
consistent with x86-64, and don't favor any specific architecture for
tuning.

Bug chromium:797763

Change-Id: Ia17428734460ed855f3aa0f83c6d2f6b3775094c
Reviewed-on: https://swiftshader-review.googlesource.com/15688Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

OpenGL ES 3.0 specifies that "For fixed-point depth buffers, the final
fragment depth written by a fragment shader is first clamped to [0, 1]
and then converted to fixed-point as if it were a window z value (see
section 2.13.1). For floating-point depth buffers, conversion is not
performed but clamping is."

Change-Id: Ic7f22f7e47106aaa86c4916fbbbe23ad9c52e186
Reviewed-on: https://swiftshader-review.googlesource.com/15628Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

When the pixel shader writes the depth value, we can't use the
interpolated implicit depth to perform a depth buffer test before the
pixel shader executes.

Change-Id: Icf80de50f4aa79f73b43e341c18478e7fd928b8d
Reviewed-on: https://swiftshader-review.googlesource.com/15608Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

The LOD value was not being computed for 3D texture sampling when no
mipmapping is done. We still need the LOD in case the minification and
magnification filters differ.

Change-Id: If32d7aaf8a92ace33bcc6ca6f0b60fe33c7e771b
Reviewed-on: https://swiftshader-review.googlesource.com/15588Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

The array folding logic generated an error when a constant array didn't
have an array of literal constants.

Change-Id: I48f58936cfa7e19e0f4026fb940654f20d9a9d5e
Reviewed-on: https://swiftshader-review.googlesource.com/15551Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

GLSL ES 3.00 spec says that length is only allowed on "array names",
but according to
https://bugs.chromium.org/p/angleproject/issues/detail?id=2142 that
was not intended to be interpreted strictly and it has been clarified
in GLSL ES 3.20.

Change-Id: Id8944346908e1c122230fd2f94c9a26cf389498b
Reviewed-on: https://swiftshader-review.googlesource.com/15550Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

Arrays of constants can be dynamically indexed, so they can't be
propagated by value and handled as literals like other constants.
Instead we store them in the constant register file.

Change-Id: Idcee033d694b20d047c70e2d13974e5a6dce127b
Reviewed-on: https://swiftshader-review.googlesource.com/15488Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

We were using a quad-layout depth format as the destination format for
glReadPixels calls with GL_DEPTH_COMPONENT / GL_FLOAT format. Upload to
a texture was also affected.

Bug b/64542192

Change-Id: I8ad700bf86737debe3268ec6a3dbd237cd85ea1a
Reviewed-on: https://swiftshader-review.googlesource.com/15548Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

This reverts commit a141a075.

Reason for revert: We need a different solution to fix this issue which doesn't
                   require having a different sampler index and register index.

Change-Id: If08a0fc4713e2bd8277630ca760bb8686a2effd2
Reviewed-on: https://swiftshader-review.googlesource.com/15649Reviewed-by: Alexis Hétu <sugoi@google.com>
Tested-by: Alexis Hétu <sugoi@google.com>

On low-resolution Android Emulator AVDs run with -gpu
swiftshader_indirect, it's possible to run fast enough to see count ==
-1 on Resource acquisition. Check for that.

bug: 70950184

Test: API 27 Chrome is usable on a 480x800 or smaller device
Change-Id: I4f4cf9cd0ab1bc2bce846f0af6ffbbe243bc1a96
Reviewed-on: https://swiftshader-review.googlesource.com/15508Reviewed-by: Lingfeng Yang <lfy@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Lingfeng Yang <lfy@google.com>

Unlike OpenGL ES 2.0, 3.0 allows framebuffer attachments to be of
different sizes: "If the attachment sizes are not all identical, the
results of rendering are defined only within the largest area that
can fit in all of the attachments. This area is defined as the
intersection of rectangles having a lower left of (0, 0) and an upper
right of (width, height) for each attachment. Contents of attachments
outside this area are undefined after execution of a rendering command
(as defined in section 2.6)."

Change-Id: Ib75be93650edd7d6a7f3d08d9206830094e3fbd8
Reviewed-on: https://swiftshader-review.googlesource.com/15469Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

Since OpenGL uses the last vertex, not the leading
vertex, as the vertex used to select the triangle
color, then the last vertex, instead of the first
vertex, has to be consistently incremented for each
triangle in the strip or fan. In order to make this
work, when leadingVertexFirst is true, the first
vertex is consistently the lowest (consistently
incremented) vertex index and when leadingVertexFirst
is false, the last vertex is consistently the highest
vertex index.

Fixes the 2 failures in:
dEQP-GLES3.functional.rasterization.flatshading

Change-Id: Ib2247a219fe8c6725dc141e6860ba2ff521dde8c
Reviewed-on: https://swiftshader-review.googlesource.com/15468Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

GL_R11F_G11F_B10F is a renderable format (in the EXT_color_buffer_float
extension) with unsigned floating-point values, which we implement using
(signed) half- or single-precision floating-point types. So we need to
clamp values to a positive range before writing to them.

Change-Id: Ic21a5b0b33905c0aeab35299fc268158f8c679f9
Reviewed-on: https://swiftshader-review.googlesource.com/15448Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

This change uses half-precision floating-point implementation formats
for all OpenGL half-precision floating-point formats, as well as the
R11F_G11F_B10F and RGB9_E5 special internal formats.

sw::FORMAT_X16B16G16R16F was implemented for the formats without alpha.

RGB9E5 conversion was optimized to not require powf(2.0, x), and 11-
and 10-bit floating-point formats were optimized to map directly to
16-bit half-precision floating-point.

Change-Id: Ic33f903d01f37394244aec9f53b0e67d1c978764
Reviewed-on: https://swiftshader-review.googlesource.com/15410Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

GL_UNSIGNED_INT_10F_11F_11F_REV and GL_UNSIGNED_INT_5_9_9_9_REV were
using sw::FORMAT_A32B32G32R32F as the implementation format, which
resulted in using bad alpha values when it was rendered into.

Change-Id: If29da49a9dc7f7a52d8846417da66f7af929f276
Reviewed-on: https://swiftshader-review.googlesource.com/15408Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

Instead of testing for the number of samplers used so far
in the shader, the Program object was checking for the
register index, which could have been larger or equal to
MAX_VERTEX_TEXTURE_IMAGE_UNITS or MAX_TEXTURE_IMAGE_UNITS,
making the shader compilation fail erroneously.

Changed sampler arrays to maps in order to get the sampler
data from the register index.

Change-Id: Iba6cbf0dbbcdbd926f2670af4413710550e341a9
Reviewed-on: https://swiftshader-review.googlesource.com/15428Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

Change-Id: I68d7dbb445eee85e8a2ff4ac3a6fc6b31c6ea305
Reviewed-on: https://swiftshader-review.googlesource.com/15369Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

Change-Id: I9c01d1a20920ab7e33d979fc69711c7e69113912
Reviewed-on: https://swiftshader-review.googlesource.com/15368Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

We can avoid adding the base level to the computed lod by putting the
base level at index 0 of the mipmap array.

Change-Id: I0bdebec3e4d0a3617dea9e3d28fdeb5dae40444b
Reviewed-on: https://swiftshader-review.googlesource.com/15329Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

Change-Id: I611815fb0dcbba97e67f3c146dffb463f133447e
Reviewed-on: https://swiftshader-review.googlesource.com/15248Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

The 'top' level is the last valid mipmap level nearest to the top of
the pyramid. Start searching for it from the 'base' of the pyramid.

Change-Id: I2d2ea3a1834f6cff2e2d342876d0649205004581
Reviewed-on: https://swiftshader-review.googlesource.com/15348Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

Change-Id: Ibc557e4843d8106c07ae2f92c9628e10f228b1ae
Reviewed-on: https://swiftshader-review.googlesource.com/15208Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

The OpenGL ES 3.0.5 spec section 3.8.10.5 states: "For cube map
textures an INVALID_OPERATION error is generated if the texture
bound to target is not cube complete".

Change-Id: I4304ecf125528b9406a011a95d7822b1f868ee14
Reviewed-on: https://swiftshader-review.googlesource.com/15328Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

Updated glGetProgramBinary and glProgramBinary in order to properly
reflect that SwiftShader doesn't support any binary format.

Fixes dEQP-GLES3.functional.negative_api.shader.get_program_binary

Change-Id: I068ce7eb7ce824719cecca2fdc651a9e6c14190c
Reviewed-on: https://swiftshader-review.googlesource.com/15330Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

Change-Id: If37158efec1555a3b4e14774dff081f284292245
Reviewed-on: https://swiftshader-review.googlesource.com/15308Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

Change-Id: Ifa2239e897cfa8878a43ce6e51ab731f2b67b7b1
Reviewed-on: https://swiftshader-review.googlesource.com/15288Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

Change-Id: I93171764df5e13a7a5be6a9116e1a5db1ed59ad5
Reviewed-on: https://swiftshader-review.googlesource.com/15268Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

<unordered_map> is not available on Android J-L.

Bug b/28006371

Change-Id: I7e06ebef092a7d3eef217f88e6fb6347c8d40ac1
Reviewed-on: https://swiftshader-review.googlesource.com/15149Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

Since EAC is no longer decoded to an 8 bit format, but ETC2 alpha
still is, we need to distinguish between the 2. Added an argument
to DecodeBlock for that purpose.

Change-Id: Ied4263cecc43f88609a357437c11f4e43af229b8
Reviewed-on: https://swiftshader-review.googlesource.com/15228Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

The convert_float_int was not scaling the [0,1] value to the full
1.31 fixed-point range before clamping the value.

Fixes all failures in:
dEQP-GLES3.functional.state_query.floats*

Change-Id: Ic558eb45f1f351d3c30a0a5047a1b252edb56bd0
Reviewed-on: https://swiftshader-review.googlesource.com/14968Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Alexis Hétu <sugoi@google.com>

Note that glReadPixels() does not perform sRGB conversion.

Change-Id: I3f9089b79652ce42cb5695d5b6a8ce92d15c27a8
Reviewed-on: https://swiftshader-review.googlesource.com/14492Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

Change-Id: Ife3342c64dfc846f8b7722f2e80612fc71f93688
Reviewed-on: https://swiftshader-review.googlesource.com/15128Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

Change-Id: Idba9017de6bf92142e2281bdca3cb6c73ea2451d
Reviewed-on: https://swiftshader-review.googlesource.com/14491Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

Previously sRGB data was converted to linear space on upload. This
caused a loss of precision. This change performs the conversion after
texel lookup. Note that we had a code path for performing the
conversion after filtering, but that leads to failures in dEQP and
unacceptable darkening between texels.

Also, glTexSubImage calls can update sRGB textures using a format/type
combination with no indication of the color space, which caused an
unintentional conversion on upload. Likewise we were missing support
for an A2B10G10R10UI implementation format.

Change-Id: Ib10845f628fb2d1849e88d7a9350868cdec32fa2
Reviewed-on: https://swiftshader-review.googlesource.com/15068Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

- Updated preprocessor code from Angle revision
  9fc8733187c02470a9eadb6d295348b6d37a2004
- Reran generate_parser.sh (flex 2.6.4, bison (GNU Bison) 3.0.4)
- Made a few trivial changes in src/OpenGL/compiler in order to
  adapt to the new preprocessor code.

Fixes all 24 failures in:
dEQP-GLES3.functional.shaders.preprocessor.*

Change-Id: I00d0b511d617ab81a0f57310174e1ba8bf7c22e5
Reviewed-on: https://swiftshader-review.googlesource.com/15109Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

The ETC2 decoder spec has a slight difference with the EAC decoder
spec when it comes to handling the 0 multiplier corner case.

For ETC2, we have (OpenGL ES 3.0 spec, section C.1.3):
"An encoder is not allowed to produce a multiplier of zero, but
 the decoder should still be able to handle also this case (and
 produce 0 x modifier = 0 in that case)."

For EAC, we have (OpenGL ES 3.0 spec, section C.1.5):
"If the multiplier value is zero, we should set the multiplier
 to 1.0/8.0"

In order to take this into account, the EAC decoded output can no
longer be represented by an 8 bit value, but must be represented by
a minimum of 11 bits, as the spec requires. For now, the EAC decoder
decodes EAC into a 32 bit integer format, which then gets converted
to a 32 bit float format internally.

Eventually, it would be possible for the EAC decoder to decode the
image to a signed 16 bit integer internal format, if it was supported.

Fixes all failures in:
dEQP-GLES3.functional.texture.wrap*

Change-Id: I32106383ade56e375229231ff230a2574791caa6
Reviewed-on: https://swiftshader-review.googlesource.com/15188Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

Change-Id: I879eb3c3d9093f769f88bd3a2dcc67f91bd6eb47
Reviewed-on: https://swiftshader-review.googlesource.com/15049Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>