From 741c40a232637c933c9273bbdef905397e54bc94 Mon Sep 17 00:00:00 2001 From: José Fonseca Date: Tue, 29 Sep 2009 16:59:13 +0100 Subject: llvmpipe: Implement non SSE4.1 versions of floor and round. --- src/gallium/drivers/llvmpipe/lp_bld_arit.c | 177 +++++++++++++++++++---- src/gallium/drivers/llvmpipe/lp_bld_arit.h | 13 +- src/gallium/drivers/llvmpipe/lp_bld_sample_soa.c | 4 +- 3 files changed, 159 insertions(+), 35 deletions(-) (limited to 'src/gallium/drivers') diff --git a/src/gallium/drivers/llvmpipe/lp_bld_arit.c b/src/gallium/drivers/llvmpipe/lp_bld_arit.c index e8c5fa3c2a..f878706ad1 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_arit.c +++ b/src/gallium/drivers/llvmpipe/lp_bld_arit.c @@ -675,6 +675,8 @@ lp_build_round_sse41(struct lp_build_context *bld, assert(type.floating); assert(type.width*type.length == 128); + assert(lp_check_value(type, a)); + assert(util_cpu_caps.has_sse4_1); switch(type.width) { case 32: @@ -693,6 +695,28 @@ lp_build_round_sse41(struct lp_build_context *bld, } +LLVMValueRef +lp_build_trunc(struct lp_build_context *bld, + LLVMValueRef a) +{ + const struct lp_type type = bld->type; + + assert(type.floating); + assert(lp_check_value(type, a)); + + if(util_cpu_caps.has_sse4_1) + return lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_TRUNCATE); + else { + LLVMTypeRef vec_type = lp_build_vec_type(type); + LLVMTypeRef int_vec_type = lp_build_int_vec_type(type); + LLVMValueRef res; + res = LLVMBuildFPToSI(bld->builder, a, int_vec_type, ""); + res = LLVMBuildSIToFP(bld->builder, res, vec_type, ""); + return res; + } +} + + LLVMValueRef lp_build_round(struct lp_build_context *bld, LLVMValueRef a) @@ -700,13 +724,17 @@ lp_build_round(struct lp_build_context *bld, const struct lp_type type = bld->type; assert(type.floating); + assert(lp_check_value(type, a)); if(util_cpu_caps.has_sse4_1) return lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_NEAREST); - - /* FIXME */ - assert(0); - return bld->undef; + else { + LLVMTypeRef vec_type = lp_build_vec_type(type); + LLVMValueRef res; + res = lp_build_iround(bld, a); + res = LLVMBuildSIToFP(bld->builder, res, vec_type, ""); + return res; + } } @@ -720,10 +748,13 @@ lp_build_floor(struct lp_build_context *bld, if(util_cpu_caps.has_sse4_1) return lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_FLOOR); - - /* FIXME */ - assert(0); - return bld->undef; + else { + LLVMTypeRef vec_type = lp_build_vec_type(type); + LLVMValueRef res; + res = lp_build_ifloor(bld, a); + res = LLVMBuildSIToFP(bld->builder, res, vec_type, ""); + return res; + } } @@ -734,47 +765,74 @@ lp_build_ceil(struct lp_build_context *bld, const struct lp_type type = bld->type; assert(type.floating); + assert(lp_check_value(type, a)); if(util_cpu_caps.has_sse4_1) return lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_CEIL); - - /* FIXME */ - assert(0); - return bld->undef; + else { + LLVMTypeRef vec_type = lp_build_vec_type(type); + LLVMValueRef res; + res = lp_build_iceil(bld, a); + res = LLVMBuildSIToFP(bld->builder, res, vec_type, ""); + return res; + } } +/** + * Convert to integer, through whichever rounding method that's fastest, + * typically truncating to zero. + */ LLVMValueRef -lp_build_trunc(struct lp_build_context *bld, - LLVMValueRef a) +lp_build_itrunc(struct lp_build_context *bld, + LLVMValueRef a) { const struct lp_type type = bld->type; + LLVMTypeRef int_vec_type = lp_build_int_vec_type(type); assert(type.floating); + assert(lp_check_value(type, a)); - if(util_cpu_caps.has_sse4_1) - return lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_TRUNCATE); - - /* FIXME */ - assert(0); - return bld->undef; + return LLVMBuildFPToSI(bld->builder, a, int_vec_type, ""); } -/** - * Convert to integer, through whichever rounding method that's fastest, - * typically truncating to zero. - */ LLVMValueRef -lp_build_int(struct lp_build_context *bld, - LLVMValueRef a) +lp_build_iround(struct lp_build_context *bld, + LLVMValueRef a) { const struct lp_type type = bld->type; LLVMTypeRef int_vec_type = lp_build_int_vec_type(type); + LLVMValueRef res; assert(type.floating); + assert(lp_check_value(type, a)); - return LLVMBuildFPToSI(bld->builder, a, int_vec_type, ""); + if(util_cpu_caps.has_sse4_1) { + res = lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_NEAREST); + } + else { + LLVMTypeRef vec_type = lp_build_vec_type(type); + LLVMValueRef mask = lp_build_int_const_scalar(type, (unsigned long long)1 << (type.width - 1)); + LLVMValueRef sign; + LLVMValueRef half; + + /* get sign bit */ + sign = LLVMBuildBitCast(bld->builder, a, int_vec_type, ""); + sign = LLVMBuildAnd(bld->builder, sign, mask, ""); + + /* sign * 0.5 */ + half = lp_build_const_scalar(type, 0.5); + half = LLVMBuildBitCast(bld->builder, half, int_vec_type, ""); + half = LLVMBuildOr(bld->builder, sign, half, ""); + half = LLVMBuildBitCast(bld->builder, half, vec_type, ""); + + res = LLVMBuildAdd(bld->builder, a, half, ""); + } + + res = LLVMBuildFPToSI(bld->builder, res, int_vec_type, ""); + + return res; } @@ -782,9 +840,68 @@ LLVMValueRef lp_build_ifloor(struct lp_build_context *bld, LLVMValueRef a) { - a = lp_build_floor(bld, a); - a = lp_build_int(bld, a); - return a; + const struct lp_type type = bld->type; + LLVMTypeRef int_vec_type = lp_build_int_vec_type(type); + LLVMValueRef res; + + assert(type.floating); + assert(lp_check_value(type, a)); + + if(util_cpu_caps.has_sse4_1) { + res = lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_FLOOR); + } + else { + /* Take the sign bit and add it to 1 constant */ + LLVMTypeRef vec_type = lp_build_vec_type(type); + unsigned mantissa = lp_mantissa(type); + LLVMValueRef mask = lp_build_int_const_scalar(type, (unsigned long long)1 << (type.width - 1)); + LLVMValueRef sign; + LLVMValueRef offset; + + /* sign = a < 0 ? ~0 : 0 */ + sign = LLVMBuildBitCast(bld->builder, a, int_vec_type, ""); + sign = LLVMBuildAnd(bld->builder, sign, mask, ""); + sign = LLVMBuildAShr(bld->builder, sign, lp_build_int_const_scalar(type, type.width - 1), ""); + + /* offset = -0.99999(9)f */ + offset = lp_build_const_scalar(type, -(double)(((unsigned long long)1 << mantissa) - 1)/((unsigned long long)1 << mantissa)); + offset = LLVMConstBitCast(offset, int_vec_type); + + /* offset = a < 0 ? -0.99999(9)f : 0.0f */ + offset = LLVMBuildAnd(bld->builder, offset, sign, ""); + offset = LLVMBuildBitCast(bld->builder, offset, vec_type, ""); + + res = LLVMBuildAdd(bld->builder, a, offset, ""); + } + + res = LLVMBuildFPToSI(bld->builder, res, int_vec_type, ""); + + return res; +} + + +LLVMValueRef +lp_build_iceil(struct lp_build_context *bld, + LLVMValueRef a) +{ + const struct lp_type type = bld->type; + LLVMTypeRef int_vec_type = lp_build_int_vec_type(type); + LLVMValueRef res; + + assert(type.floating); + assert(lp_check_value(type, a)); + + if(util_cpu_caps.has_sse4_1) { + res = lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_CEIL); + } + else { + assert(0); + res = bld->undef; + } + + res = LLVMBuildFPToSI(bld->builder, res, int_vec_type, ""); + + return res; } diff --git a/src/gallium/drivers/llvmpipe/lp_bld_arit.h b/src/gallium/drivers/llvmpipe/lp_bld_arit.h index d68a97c4b8..095a8e1cab 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_arit.h +++ b/src/gallium/drivers/llvmpipe/lp_bld_arit.h @@ -126,11 +126,18 @@ lp_build_trunc(struct lp_build_context *bld, LLVMValueRef a); LLVMValueRef -lp_build_int(struct lp_build_context *bld, - LLVMValueRef a); +lp_build_ifloor(struct lp_build_context *bld, + LLVMValueRef a); +LLVMValueRef +lp_build_iceil(struct lp_build_context *bld, + LLVMValueRef a); LLVMValueRef -lp_build_ifloor(struct lp_build_context *bld, +lp_build_iround(struct lp_build_context *bld, + LLVMValueRef a); + +LLVMValueRef +lp_build_itrunc(struct lp_build_context *bld, LLVMValueRef a); LLVMValueRef diff --git a/src/gallium/drivers/llvmpipe/lp_bld_sample_soa.c b/src/gallium/drivers/llvmpipe/lp_bld_sample_soa.c index 8ca1be6f1b..1a47ca32d2 100644 --- a/src/gallium/drivers/llvmpipe/lp_bld_sample_soa.c +++ b/src/gallium/drivers/llvmpipe/lp_bld_sample_soa.c @@ -274,8 +274,8 @@ lp_build_sample_2d_linear_soa(struct lp_build_sample_context *bld, s_fpart = lp_build_sub(&bld->coord_bld, s, s_ipart); t_fpart = lp_build_sub(&bld->coord_bld, t, t_ipart); - x0 = lp_build_int(&bld->coord_bld, s_ipart); - y0 = lp_build_int(&bld->coord_bld, t_ipart); + x0 = lp_build_itrunc(&bld->coord_bld, s_ipart); + y0 = lp_build_itrunc(&bld->coord_bld, t_ipart); x0 = lp_build_sample_wrap(bld, x0, width, bld->static_state->pot_width, bld->static_state->wrap_s); y0 = lp_build_sample_wrap(bld, y0, height, bld->static_state->pot_height, bld->static_state->wrap_t); -- cgit v1.2.3