[Mesa-dev] [PATCH 1/2] gallium: allow 512-bit vectors

[Jose Fonseca]

On 11/10/17 16:50, Tim Rowley wrote:
> Increase the max allowed vector size from 256 to 512.
> 
> No piglit llvmpipe regressions running on avx2.
> 
> Cc: Dave Airlie <airlied at redhat.com>
> Cc: Jose Fonseca <jfonseca at vmware.com>
> ---
>   src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c | 14 +++++++-------
>   src/gallium/auxiliary/gallivm/lp_bld_type.h     |  4 ++--
>   2 files changed, 9 insertions(+), 9 deletions(-)
> 
> diff --git a/src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c b/src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c
> index de18f629cd..97efc3a399 100644
> --- a/src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c
> +++ b/src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c
> @@ -1272,9 +1272,9 @@ emit_fetch_constant(
>   /**
>    * Fetch 64-bit values from two separate channels.
>    * 64-bit values are stored split across two channels, like xy and zw.
> - * This function creates a set of 16 floats,
> + * This function creates a set of vec_length*2 floats,
>    * extracts the values from the two channels,
> - * puts them in the correct place, then casts to 8 64-bits.
> + * puts them in the correct place, then casts to vec_length 64-bits.
>    */
>   static LLVMValueRef
>   emit_fetch_64bit(
> @@ -1289,9 +1289,9 @@ emit_fetch_64bit(
>      LLVMValueRef res;
>      struct lp_build_context *bld_fetch = stype_to_fetch(bld_base, stype);
>      int i;
> -   LLVMValueRef shuffles[16];
> +   LLVMValueRef shuffles[2 * (LP_MAX_VECTOR_WIDTH/32)];
>      int len = bld_base->base.type.length * 2;
> -   assert(len <= 16);
> +   assert(len <= (2 * (LP_MAX_VECTOR_WIDTH/32)));
>   
>      for (i = 0; i < bld_base->base.type.length * 2; i+=2) {
>         shuffles[i] = lp_build_const_int32(gallivm, i / 2);
> @@ -1691,7 +1691,7 @@ emit_fetch_deriv(
>   }
>   
>   /**
> - * store an array of 8 64-bit into two arrays of 8 floats
> + * store an array of vec-length 64-bit into two arrays of vec_length floats
>    * i.e.
>    * value is d0, d1, d2, d3 etc.
>    * each 64-bit has high and low pieces x, y
> @@ -1710,8 +1710,8 @@ emit_store_64bit_chan(struct lp_build_tgsi_context *bld_base,
>      struct lp_build_context *float_bld = &bld_base->base;
>      unsigned i;
>      LLVMValueRef temp, temp2;
> -   LLVMValueRef shuffles[8];
> -   LLVMValueRef shuffles2[8];
> +   LLVMValueRef shuffles[LP_MAX_VECTOR_WIDTH/32];
> +   LLVMValueRef shuffles2[LP_MAX_VECTOR_WIDTH/32];
>   
>      for (i = 0; i < bld_base->base.type.length; i++) {
>         shuffles[i] = lp_build_const_int32(gallivm, i * 2);
> diff --git a/src/gallium/auxiliary/gallivm/lp_bld_type.h b/src/gallium/auxiliary/gallivm/lp_bld_type.h
> index afe8722b05..62f1f85461 100644
> --- a/src/gallium/auxiliary/gallivm/lp_bld_type.h
> +++ b/src/gallium/auxiliary/gallivm/lp_bld_type.h
> @@ -59,7 +59,7 @@ extern unsigned lp_native_vector_width;
>    * Should only be used when lp_native_vector_width isn't available,
>    * i.e. sizing/alignment of non-malloced variables.
>    */
> -#define LP_MAX_VECTOR_WIDTH 256
> +#define LP_MAX_VECTOR_WIDTH 512
>   
>   /**
>    * Minimum vector alignment for static variable alignment
> @@ -67,7 +67,7 @@ extern unsigned lp_native_vector_width;
>    * It should always be a constant equal to LP_MAX_VECTOR_WIDTH/8.  An
>    * expression is non-portable.
>    */
> -#define LP_MIN_VECTOR_ALIGN 32
> +#define LP_MIN_VECTOR_ALIGN 64
>   
>   /**
>    * Several functions can only cope with vectors of length up to this value.
> 

LGTM.  But better let Roland give a look too, as he's maintaining 
llvmpipe nowadays.

Reviewed-by: Jose Fonseca <jfonseca at vmware.com>