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

[Roland Scheidegger]

This looks alright to me. There's quite a few places where this will
increase the memory requirements a bit but there's no way around it
(well unless deciding at runtime alignments/allocations, but it's
probably not worth the trouble).

Reviewed-by: Roland Scheidegger <sroland at vmware.com>

Am 11.10.2017 um 17:50 schrieb Tim Rowley:
> 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.
>