[Mesa-dev] [PATCH 4/6] nir/lower_tex: support for lowering RECT textures

[Rob Clark]

On Fri, Sep 18, 2015 at 11:22 AM, Kenneth Graunke <kenneth at whitecape.org> wrote:
> On Friday, September 18, 2015 10:55:07 AM Rob Clark wrote:
>> From: Rob Clark <robclark at freedesktop.org>
>>
>> v2: comments/suggestions from Ilia and Eric, split out get_texture_size()
>> helper so we can use it in the next commit for clamping RECT textures.
>>
>> Signed-off-by: Rob Clark <robclark at freedesktop.org>
>> ---
>>  src/glsl/nir/nir.h           |  7 +++++
>>  src/glsl/nir/nir_lower_tex.c | 62 +++++++++++++++++++++++++++++++++++++++++---
>>  2 files changed, 66 insertions(+), 3 deletions(-)
>>
>> diff --git a/src/glsl/nir/nir.h b/src/glsl/nir/nir.h
>> index 4600fb0..3c908b9 100644
>> --- a/src/glsl/nir/nir.h
>> +++ b/src/glsl/nir/nir.h
>> @@ -1843,6 +1843,13 @@ typedef struct nir_lower_tex_options {
>>      * sampler types a texture projector is lowered.
>>      */
>>     unsigned lower_txp;
>> +
>> +   /**
>> +    * If true, lower rect textures to 2D, using txs to fetch the
>> +    * texture dimensions and dividing the texture coords by the
>> +    * texture dims to normalize.
>> +    */
>> +   bool lower_rect;
>>  } nir_lower_tex_options;
>>
>>  void nir_lower_tex(nir_shader *shader,
>> diff --git a/src/glsl/nir/nir_lower_tex.c b/src/glsl/nir/nir_lower_tex.c
>> index 281fc9f..a71a5c5 100644
>> --- a/src/glsl/nir/nir_lower_tex.c
>> +++ b/src/glsl/nir/nir_lower_tex.c
>> @@ -22,9 +22,13 @@
>>   */
>>
>>  /*
>> - * This lowering pass converts the coordinate division for texture projection
>> - * to be done in ALU instructions instead of asking the texture operation to
>> - * do so.
>> + * This lowering pass supports (as configured via nir_lower_tex_options)
>> + * various texture related conversions:
>> + *   + texture projector lowering: converts the coordinate division for
>> + *     texture projection to be done in ALU instructions instead of
>> + *     asking the texture operation to do so.
>> + *   + lowering RECT: converts the un-normalized RECT texture coordinates
>> + *     to normalized coordinates with txs plus ALU instructions
>>   */
>>
>>  #include "nir.h"
>> @@ -111,6 +115,55 @@ project_src(nir_builder *b, nir_tex_instr *tex)
>>     tex->num_srcs--;
>>  }
>>
>> +static nir_ssa_def *
>> +get_texture_size(nir_builder *b, nir_tex_instr *tex)
>> +{
>> +   b->cursor = nir_before_instr(&tex->instr);
>> +
>> +   /* RECT textures should not be array: */
>> +   assert(!tex->is_array);
>> +
>> +   nir_tex_instr *txs;
>> +
>> +   txs = nir_tex_instr_create(b->shader, 1);
>> +   txs->op = nir_texop_txs;
>> +   txs->sampler_dim = GLSL_SAMPLER_DIM_RECT;
>> +   txs->sampler_index = tex->sampler_index;
>> +
>> +   /* only single src, the lod: */
>> +   txs->src[0].src = nir_src_for_ssa(nir_imm_int(b, 0));
>> +   txs->src[0].src_type = nir_tex_src_lod;
>> +
>> +   nir_ssa_dest_init(&txs->instr, &txs->dest, 3, NULL);
>
> Shouldn't this be 2? textureSize of 2DRect returns an ivec2...

yup

>> +   nir_builder_instr_insert(b, &txs->instr);
>> +
>> +   return nir_i2f(b, &txs->dest.ssa);
>> +}
>> +
>> +static void
>> +lower_rect(nir_builder *b, nir_tex_instr *tex)
>> +{
>> +   nir_ssa_def *txs = get_texture_size(b, tex);
>> +
>> +   nir_ssa_def *scalex = nir_frcp(b, nir_channel(b, txs, 0));
>> +   nir_ssa_def *scaley = nir_frcp(b, nir_channel(b, txs, 1));
>> +   nir_ssa_def *scale  = nir_vec2(b, scalex, scaley);
>
> Why not just do 1 / txs.xy directly?

I had assumed that frcp just operates on src.x (at least that it is
how it is in tgsi and hw that I have seen).. if it operates on each
component, then yes I could do it in one shot.  And I guess since txs
is already a vec2, I could just:

  nir_ssa_def *scale = nir_frcp(b, txs);

> unsigned swizzle[4] = { SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Y, SWIZZLE_Y };
> nir_ssa_def *scale = nir_frcp(b, nir_swizzle(b, txs, swizzle, 2, false);
>
>> +
>> +   /* Walk through the sources normalizing the requested arguments. */
>> +   for (unsigned i = 0; i < tex->num_srcs; i++) {
>> +      if (tex->src[i].src_type != nir_tex_src_coord)
>> +         continue;
>> +
>> +      nir_ssa_def *coords =
>> +         nir_ssa_for_src(b, tex->src[i].src, tex->coord_components);
>> +      nir_instr_rewrite_src(&tex->instr,
>> +                            &tex->src[i].src,
>> +                            nir_src_for_ssa(nir_fmul(b, coords, scale)));
>> +   }
>> +
>> +   tex->sampler_dim = GLSL_SAMPLER_DIM_2D;
>> +}
>> +
>>  static bool
>>  nir_lower_tex_block(nir_block *block, void *void_state)
>>  {
>> @@ -127,6 +180,9 @@ nir_lower_tex_block(nir_block *block, void *void_state)
>>        if (lower_txp)
>>           project_src(b, tex);
>>
>> +      if ((tex->sampler_dim == GLSL_SAMPLER_DIM_RECT) &&
>
> Please remove the bonus parens.

hmm, I'm the only one who *likes* bonus parens?  (Then again, I'd
always liked lisp/scheme so maybe I have a paren fetish..)

BR,
-R

>> +          state->options->lower_rect)
>> +         lower_rect(b, tex);
>>     }
>>
>>     return true;
>>
>
> With the ivec2/ivec3 thing fixed, this would be:
> Reviewed-by: Kenneth Graunke <kenneth at whitecape.org>