From 728741a3bd27a3aaab50393665d25776ad06fcb3 Mon Sep 17 00:00:00 2001 From: =?utf8?q?Jos=C3=A9=20Fonseca?= Date: Tue, 6 Apr 2010 22:49:33 +0100 Subject: [PATCH] gallivm: Fallback to calling util_format_description::fetch_float for any format we can't code LLVM IR directly. --- src/gallium/auxiliary/gallivm/lp_bld_format.h | 11 ++ src/gallium/auxiliary/gallivm/lp_bld_format_soa.c | 140 ++++++++++++++++++++++ src/gallium/auxiliary/gallivm/lp_bld_sample.c | 5 +- src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c | 63 ++++++---- 4 files changed, 192 insertions(+), 27 deletions(-) diff --git a/src/gallium/auxiliary/gallivm/lp_bld_format.h b/src/gallium/auxiliary/gallivm/lp_bld_format.h index 73ab6de3f21..bb1298ed3f4 100644 --- a/src/gallium/auxiliary/gallivm/lp_bld_format.h +++ b/src/gallium/auxiliary/gallivm/lp_bld_format.h @@ -80,4 +80,15 @@ lp_build_unpack_rgba_soa(LLVMBuilderRef builder, LLVMValueRef *rgba); +void +lp_build_fetch_rgba_soa(LLVMBuilderRef builder, + const struct util_format_description *format_desc, + struct lp_type type, + LLVMValueRef base_ptr, + LLVMValueRef offsets, + LLVMValueRef i, + LLVMValueRef j, + LLVMValueRef *rgba); + + #endif /* !LP_BLD_FORMAT_H */ diff --git a/src/gallium/auxiliary/gallivm/lp_bld_format_soa.c b/src/gallium/auxiliary/gallivm/lp_bld_format_soa.c index 9f242844e5a..433134b8090 100644 --- a/src/gallium/auxiliary/gallivm/lp_bld_format_soa.c +++ b/src/gallium/auxiliary/gallivm/lp_bld_format_soa.c @@ -27,10 +27,14 @@ #include "util/u_format.h" +#include "util/u_memory.h" +#include "util/u_string.h" #include "lp_bld_type.h" #include "lp_bld_const.h" #include "lp_bld_conv.h" +#include "lp_bld_sample.h" /* for lp_build_gather */ +#include "lp_bld_init.h" #include "lp_bld_format.h" @@ -240,3 +244,139 @@ lp_build_unpack_rgba_soa(LLVMBuilderRef builder, lp_build_format_swizzle_soa(format_desc, type, inputs, rgba); } + + +/** + * Fetch a pixel into a SoA. + * + * i and j are the sub-block pixel coordinates. + */ +void +lp_build_fetch_rgba_soa(LLVMBuilderRef builder, + const struct util_format_description *format_desc, + struct lp_type type, + LLVMValueRef base_ptr, + LLVMValueRef offset, + LLVMValueRef i, + LLVMValueRef j, + LLVMValueRef *rgba) +{ + + if (format_desc->block.width == 1 && + format_desc->block.height == 1 && + format_desc->block.bits <= type.width) + { + /* + * The packed pixel fits into an element of the destination format. Put + * the packed pixels into a vector and estract each component for all + * vector elements in parallel. + */ + + LLVMValueRef packed; + + /* + * gather the texels from the texture + */ + packed = lp_build_gather(builder, + type.length, + format_desc->block.bits, + type.width, + base_ptr, offset); + + /* + * convert texels to float rgba + */ + lp_build_unpack_rgba_soa(builder, + format_desc, + type, + packed, rgba); + } + else { + /* + * Fallback to calling util_format_description::fetch_float for each + * pixel. + * + * This is definitely not the most efficient way of fetching pixels, as + * we miss the opportunity to do vectorization, but this it is a + * convenient for formats or scenarios for which there was no opportunity + * or incentive to optimize. + */ + + LLVMModuleRef module = LLVMGetGlobalParent(LLVMGetBasicBlockParent(LLVMGetInsertBlock(builder))); + char name[256]; + LLVMValueRef function; + LLVMValueRef tmp; + unsigned k, chan; + + assert(type.floating); + + util_snprintf(name, sizeof name, "util_format_%s_fetch_float", format_desc->short_name); + + /* + * Declare and bind format_desc->fetch_float(). + */ + + function = LLVMGetNamedFunction(module, name); + if (!function) { + LLVMTypeRef ret_type; + LLVMTypeRef arg_types[4]; + LLVMTypeRef function_type; + + ret_type = LLVMVoidType(); + arg_types[0] = LLVMPointerType(LLVMFloatType(), 0); + arg_types[1] = LLVMPointerType(LLVMInt8Type(), 0); + arg_types[3] = arg_types[2] = LLVMIntType(sizeof(unsigned) * 8); + function_type = LLVMFunctionType(ret_type, arg_types, Elements(arg_types), 0); + function = LLVMAddFunction(module, name, function_type); + + LLVMSetFunctionCallConv(function, LLVMCCallConv); + LLVMSetLinkage(function, LLVMExternalLinkage); + + assert(LLVMIsDeclaration(function)); + + LLVMAddGlobalMapping(lp_build_engine, function, format_desc->fetch_float); + } + + for (chan = 0; chan < 4; ++chan) { + rgba[chan] = lp_build_undef(type); + } + + tmp = LLVMBuildArrayAlloca(builder, + LLVMFloatType(), + LLVMConstInt(LLVMInt32Type(), 4, 0), + ""); + + /* + * Invoke format_desc->fetch_float() for each pixel and insert the result + * in the SoA vectors. + */ + + for(k = 0; k < type.length; ++k) { + LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), k, 0); + LLVMValueRef offset_elem; + LLVMValueRef ptr; + LLVMValueRef i_elem, j_elem; + LLVMValueRef args[4]; + + offset_elem = LLVMBuildExtractElement(builder, offset, index, ""); + ptr = LLVMBuildGEP(builder, base_ptr, &offset_elem, 1, ""); + + i_elem = LLVMBuildExtractElement(builder, i, index, ""); + j_elem = LLVMBuildExtractElement(builder, j, index, ""); + + args[0] = tmp; + args[1] = ptr; + args[2] = i_elem; + args[3] = j_elem; + + LLVMBuildCall(builder, function, args, 4, ""); + + for (chan = 0; chan < 4; ++chan) { + LLVMValueRef chan_val = LLVMConstInt(LLVMInt32Type(), chan, 0), + tmp_chan = LLVMBuildGEP(builder, tmp, &chan_val, 1, ""); + tmp_chan = LLVMBuildLoad(builder, tmp_chan, ""); + rgba[chan] = LLVMBuildInsertElement(builder, rgba[chan], tmp_chan, index, ""); + } + } + } +} diff --git a/src/gallium/auxiliary/gallivm/lp_bld_sample.c b/src/gallium/auxiliary/gallivm/lp_bld_sample.c index bb76ad4c6bf..0744be6b41f 100644 --- a/src/gallium/auxiliary/gallivm/lp_bld_sample.c +++ b/src/gallium/auxiliary/gallivm/lp_bld_sample.c @@ -157,9 +157,10 @@ lp_build_gather(LLVMBuilderRef builder, /** - * Compute the offset of a pixel. + * Compute the offset of a pixel block. * - * x, y, z, y_stride, z_stride are vectors + * x, y, z, y_stride, z_stride are vectors, and they refer to pixel blocks, as + * per format description, and not individual pixels. */ LLVMValueRef lp_build_sample_offset(struct lp_build_context *bld, diff --git a/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c b/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c index 5b56f2cf3b7..395eaaba269 100644 --- a/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c +++ b/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c @@ -211,7 +211,7 @@ lp_build_sample_texel_soa(struct lp_build_sample_context *bld, const int dims = texture_dims(bld->static_state->target); struct lp_build_context *int_coord_bld = &bld->int_coord_bld; LLVMValueRef offset; - LLVMValueRef packed; + LLVMValueRef i, j; LLVMValueRef use_border = NULL; /* use_border = x < 0 || x >= width || y < 0 || y >= height */ @@ -249,6 +249,43 @@ lp_build_sample_texel_soa(struct lp_build_sample_context *bld, } /* + * Describe the coordinates in terms of pixel blocks. + * + * TODO: pixel blocks are power of two. LLVM should convert rem/div to + * bit arithmetic. Verify this. + */ + + if (bld->format_desc->block.width == 1) { + i = bld->uint_coord_bld.zero; + } + else { + LLVMValueRef block_width = lp_build_const_int_vec(bld->uint_coord_bld.type, bld->format_desc->block.width); + i = LLVMBuildURem(bld->builder, x, block_width, ""); + x = LLVMBuildUDiv(bld->builder, x, block_width, ""); + } + + if (bld->format_desc->block.height == 1) { + j = bld->uint_coord_bld.zero; + } + else { + LLVMValueRef block_height = lp_build_const_int_vec(bld->uint_coord_bld.type, bld->format_desc->block.height); + j = LLVMBuildURem(bld->builder, y, block_height, ""); + y = LLVMBuildUDiv(bld->builder, y, block_height, ""); + } + + /* convert x,y,z coords to linear offset from start of texture, in bytes */ + offset = lp_build_sample_offset(&bld->uint_coord_bld, + bld->format_desc, + x, y, z, y_stride, z_stride); + + lp_build_fetch_rgba_soa(bld->builder, + bld->format_desc, + bld->texel_type, + data_ptr, offset, + i, j, + texel); + + /* * Note: if we find an app which frequently samples the texture border * we might want to implement a true conditional here to avoid sampling * the texture whenever possible (since that's quite a bit of code). @@ -263,30 +300,6 @@ lp_build_sample_texel_soa(struct lp_build_sample_context *bld, * the texel color results with the border color. */ - /* convert x,y,z coords to linear offset from start of texture, in bytes */ - offset = lp_build_sample_offset(&bld->uint_coord_bld, - bld->format_desc, - x, y, z, y_stride, z_stride); - - assert(bld->format_desc->block.width == 1); - assert(bld->format_desc->block.height == 1); - assert(bld->format_desc->block.bits <= bld->texel_type.width); - - /* gather the texels from the texture */ - packed = lp_build_gather(bld->builder, - bld->texel_type.length, - bld->format_desc->block.bits, - bld->texel_type.width, - data_ptr, offset); - - texel[0] = texel[1] = texel[2] = texel[3] = NULL; - - /* convert texels to float rgba */ - lp_build_unpack_rgba_soa(bld->builder, - bld->format_desc, - bld->texel_type, - packed, texel); - if (use_border) { /* select texel color or border color depending on use_border */ int chan; -- 2.11.0