*/
-#include "util/u_format.h"
+#include "util/format/u_format.h"
#include "util/u_memory.h"
#include "util/u_math.h"
#include "util/u_pointer.h"
#include "util/u_string.h"
+#include "util/u_cpu_detect.h"
#include "lp_bld_arit.h"
#include "lp_bld_init.h"
#include "lp_bld_gather.h"
#include "lp_bld_debug.h"
#include "lp_bld_format.h"
+#include "lp_bld_pack.h"
#include "lp_bld_intr.h"
-
+#include "lp_bld_logic.h"
+#include "lp_bld_bitarit.h"
+#include "lp_bld_misc.h"
/**
* Basic swizzling. Rearrange the order of the unswizzled array elements
- * according to the format description. PIPE_SWIZZLE_ZERO/ONE are supported
+ * according to the format description. PIPE_SWIZZLE_0/ONE are supported
* too.
* Ex: if unswizzled[4] = {B, G, R, x}, then swizzled_out[4] = {R, G, B, 1}.
*/
assert(bld->type.length % 4 == 0);
for (chan = 0; chan < 4; ++chan) {
- enum util_format_swizzle swizzle;
+ enum pipe_swizzle swizzle;
if (desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) {
/*
* For ZS formats do RGBA = ZZZ1
*/
if (chan == 3) {
- swizzle = UTIL_FORMAT_SWIZZLE_1;
- } else if (desc->swizzle[0] == UTIL_FORMAT_SWIZZLE_NONE) {
- swizzle = UTIL_FORMAT_SWIZZLE_0;
+ swizzle = PIPE_SWIZZLE_1;
+ } else if (desc->swizzle[0] == PIPE_SWIZZLE_NONE) {
+ swizzle = PIPE_SWIZZLE_0;
} else {
swizzle = desc->swizzle[0];
}
return TRUE;
}
+/*
+ * Do rounding when converting small unorm values to larger ones.
+ * Not quite 100% accurate, as it's done by appending MSBs, but
+ * should be good enough.
+ */
+
+static inline LLVMValueRef
+scale_bits_up(struct gallivm_state *gallivm,
+ int src_bits,
+ int dst_bits,
+ LLVMValueRef src,
+ struct lp_type src_type)
+{
+ LLVMBuilderRef builder = gallivm->builder;
+ LLVMValueRef result = src;
+
+ if (src_bits == 1 && dst_bits > 1) {
+ /*
+ * Useful for a1 - we'd need quite some repeated copies otherwise.
+ */
+ struct lp_build_context bld;
+ LLVMValueRef dst_mask;
+ lp_build_context_init(&bld, gallivm, src_type);
+ dst_mask = lp_build_const_int_vec(gallivm, src_type,
+ (1 << dst_bits) - 1),
+ result = lp_build_cmp(&bld, PIPE_FUNC_EQUAL, src,
+ lp_build_const_int_vec(gallivm, src_type, 0));
+ result = lp_build_andnot(&bld, dst_mask, result);
+ }
+ else if (dst_bits > src_bits) {
+ /* Scale up bits */
+ int db = dst_bits - src_bits;
+
+ /* Shift left by difference in bits */
+ result = LLVMBuildShl(builder,
+ src,
+ lp_build_const_int_vec(gallivm, src_type, db),
+ "");
+
+ if (db <= src_bits) {
+ /* Enough bits in src to fill the remainder */
+ LLVMValueRef lower = LLVMBuildLShr(builder,
+ src,
+ lp_build_const_int_vec(gallivm, src_type,
+ src_bits - db),
+ "");
+
+ result = LLVMBuildOr(builder, result, lower, "");
+ } else if (db > src_bits) {
+ /* Need to repeatedly copy src bits to fill remainder in dst */
+ unsigned n;
+
+ for (n = src_bits; n < dst_bits; n *= 2) {
+ LLVMValueRef shuv = lp_build_const_int_vec(gallivm, src_type, n);
+
+ result = LLVMBuildOr(builder,
+ result,
+ LLVMBuildLShr(builder, result, shuv, ""),
+ "");
+ }
+ }
+ } else {
+ assert (dst_bits == src_bits);
+ }
+
+ return result;
+}
/**
* Unpack a single pixel into its XYZW components.
LLVMValueRef shifts[4];
LLVMValueRef masks[4];
LLVMValueRef scales[4];
+ LLVMTypeRef vec32_type;
boolean normalized;
boolean needs_uitofp;
* matches floating point size */
assert (LLVMTypeOf(packed) == LLVMInt32TypeInContext(gallivm->context));
+ vec32_type = LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4);
+
/* Broadcast the packed value to all four channels
* before: packed = BGRA
* after: packed = {BGRA, BGRA, BGRA, BGRA}
*/
- packed = LLVMBuildInsertElement(builder,
- LLVMGetUndef(LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4)),
- packed,
+ packed = LLVMBuildInsertElement(builder, LLVMGetUndef(vec32_type), packed,
LLVMConstNull(LLVMInt32TypeInContext(gallivm->context)),
"");
- packed = LLVMBuildShuffleVector(builder,
- packed,
- LLVMGetUndef(LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4)),
- LLVMConstNull(LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4)),
+ packed = LLVMBuildShuffleVector(builder, packed, LLVMGetUndef(vec32_type),
+ LLVMConstNull(vec32_type),
"");
/* Initialize vector constants */
/* Ex: convert packed = {XYZW, XYZW, XYZW, XYZW}
* into masked = {X, Y, Z, W}
*/
- shifted = LLVMBuildLShr(builder, packed, LLVMConstVector(shifts, 4), "");
- masked = LLVMBuildAnd(builder, shifted, LLVMConstVector(masks, 4), "");
+ if (desc->block.bits < 32 && normalized) {
+ /*
+ * Note: we cannot do the shift below on x86 natively until AVX2.
+ *
+ * Old llvm versions will resort to scalar extract/shift insert,
+ * which is definitely terrible, new versions will just do
+ * several vector shifts and shuffle/blend results together.
+ * We could turn this into a variable left shift plus a constant
+ * right shift, and llvm would then turn the variable left shift
+ * into a mul for us (albeit without sse41 the mul needs emulation
+ * too...). However, since we're going to do a float mul
+ * anyway, we just adjust that mul instead (plus the mask), skipping
+ * the shift completely.
+ * We could also use a extra mul when the format isn't normalized and
+ * we don't have AVX2 support, but don't bother for now. Unfortunately,
+ * this strategy doesn't work for 32bit formats (such as rgb10a2 or even
+ * rgba8 if it ends up here), as that would require UIToFP, albeit that
+ * would be fixable with easy 16bit shuffle (unless there's channels
+ * crossing 16bit boundaries).
+ */
+ for (i = 0; i < 4; ++i) {
+ if (desc->channel[i].type != UTIL_FORMAT_TYPE_VOID) {
+ unsigned bits = desc->channel[i].size;
+ unsigned shift = desc->channel[i].shift;
+ unsigned long long mask = ((1ULL << bits) - 1) << shift;
+ scales[i] = lp_build_const_float(gallivm, 1.0 / mask);
+ masks[i] = lp_build_const_int32(gallivm, mask);
+ }
+ }
+ masked = LLVMBuildAnd(builder, packed, LLVMConstVector(masks, 4), "");
+ } else {
+ shifted = LLVMBuildLShr(builder, packed, LLVMConstVector(shifts, 4), "");
+ masked = LLVMBuildAnd(builder, shifted, LLVMConstVector(masks, 4), "");
+ }
if (!needs_uitofp) {
/* UIToFP can't be expressed in SSE2 */
casted = LLVMBuildUIToFP(builder, masked, LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), 4), "");
}
- /* At this point 'casted' may be a vector of floats such as
- * {255.0, 255.0, 255.0, 255.0}. Next, if the pixel values are normalized
+ /*
+ * At this point 'casted' may be a vector of floats such as
+ * {255.0, 255.0, 255.0, 255.0}. (Normalized values may be multiplied
+ * by powers of two). Next, if the pixel values are normalized
* we'll scale this to {1.0, 1.0, 1.0, 1.0}.
*/
* \param ptr address of the pixel block (or the texel if uncompressed)
* \param i, j the sub-block pixel coordinates. For non-compressed formats
* these will always be (0, 0).
+ * \param cache optional value pointing to a lp_build_format_cache structure
* \return a 4 element vector with the pixel's RGBA values.
*/
LLVMValueRef
LLVMValueRef base_ptr,
LLVMValueRef offset,
LLVMValueRef i,
- LLVMValueRef j)
+ LLVMValueRef j,
+ LLVMValueRef cache)
{
+ const struct util_format_unpack_description *unpack =
+ util_format_unpack_description(format_desc->format);
LLVMBuilderRef builder = gallivm->builder;
unsigned num_pixels = type.length / 4;
struct lp_build_context bld;
if (format_matches_type(format_desc, type) &&
format_desc->block.bits <= type.width * 4 &&
- util_is_power_of_two(format_desc->block.bits)) {
+ /* XXX this shouldn't be needed */
+ util_is_power_of_two_or_zero(format_desc->block.bits)) {
LLVMValueRef packed;
LLVMTypeRef dst_vec_type = lp_build_vec_type(gallivm, type);
+ struct lp_type fetch_type;
unsigned vec_len = type.width * type.length;
/*
* scaling or converting.
*/
+ fetch_type = lp_type_uint(type.width*4);
packed = lp_build_gather(gallivm, type.length/4,
- format_desc->block.bits, type.width*4,
+ format_desc->block.bits, fetch_type,
aligned, base_ptr, offset, TRUE);
assert(format_desc->block.bits <= vec_len);
return lp_build_format_swizzle_aos(format_desc, &bld, packed);
}
+ /*
+ * Bit arithmetic for converting small_unorm to unorm8.
+ *
+ * This misses some opportunities for optimizations (like skipping mask
+ * for the highest channel for instance, or doing bit scaling in parallel
+ * for channels with the same bit width) but it should be passable for
+ * all arithmetic formats.
+ */
+ if (format_desc->layout == UTIL_FORMAT_LAYOUT_PLAIN &&
+ format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB &&
+ util_format_fits_8unorm(format_desc) &&
+ type.width == 8 && type.norm == 1 && type.sign == 0 &&
+ type.fixed == 0 && type.floating == 0) {
+ LLVMValueRef packed, res = NULL, chans[4], rgba[4];
+ LLVMTypeRef dst_vec_type, conv_vec_type;
+ struct lp_type fetch_type, conv_type;
+ struct lp_build_context bld_conv;
+ unsigned j;
+
+ fetch_type = lp_type_uint(type.width*4);
+ conv_type = lp_type_int_vec(type.width*4, type.width * type.length);
+ dst_vec_type = lp_build_vec_type(gallivm, type);
+ conv_vec_type = lp_build_vec_type(gallivm, conv_type);
+ lp_build_context_init(&bld_conv, gallivm, conv_type);
+
+ packed = lp_build_gather(gallivm, type.length/4,
+ format_desc->block.bits, fetch_type,
+ aligned, base_ptr, offset, TRUE);
+
+ assert(format_desc->block.bits * type.length / 4 <=
+ type.width * type.length);
+
+ packed = LLVMBuildBitCast(gallivm->builder, packed, conv_vec_type, "");
+
+ for (j = 0; j < format_desc->nr_channels; ++j) {
+ unsigned mask = 0;
+ unsigned sa = format_desc->channel[j].shift;
+
+ mask = (1 << format_desc->channel[j].size) - 1;
+
+ /* Extract bits from source */
+ chans[j] = LLVMBuildLShr(builder, packed,
+ lp_build_const_int_vec(gallivm, conv_type, sa),
+ "");
+
+ chans[j] = LLVMBuildAnd(builder, chans[j],
+ lp_build_const_int_vec(gallivm, conv_type, mask),
+ "");
+
+ /* Scale bits */
+ if (type.norm) {
+ chans[j] = scale_bits_up(gallivm, format_desc->channel[j].size,
+ type.width, chans[j], conv_type);
+ }
+ }
+ /*
+ * This is a hacked lp_build_format_swizzle_soa() since we need a
+ * normalized 1 but only 8 bits in a 32bit vector...
+ */
+ for (j = 0; j < 4; ++j) {
+ enum pipe_swizzle swizzle = format_desc->swizzle[j];
+ if (swizzle == PIPE_SWIZZLE_1) {
+ rgba[j] = lp_build_const_int_vec(gallivm, conv_type, (1 << type.width) - 1);
+ } else {
+ rgba[j] = lp_build_swizzle_soa_channel(&bld_conv, chans, swizzle);
+ }
+ if (j == 0) {
+ res = rgba[j];
+ } else {
+ rgba[j] = LLVMBuildShl(builder, rgba[j],
+ lp_build_const_int_vec(gallivm, conv_type,
+ j * type.width), "");
+ res = LLVMBuildOr(builder, res, rgba[j], "");
+ }
+ }
+ res = LLVMBuildBitCast(gallivm->builder, res, dst_vec_type, "");
+
+ return res;
+ }
+
/*
* Bit arithmetic
*/
format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) &&
format_desc->block.width == 1 &&
format_desc->block.height == 1 &&
- util_is_power_of_two(format_desc->block.bits) &&
+ /* XXX this shouldn't be needed */
+ util_is_power_of_two_or_zero(format_desc->block.bits) &&
format_desc->block.bits <= 32 &&
format_desc->is_bitmask &&
!format_desc->is_mixed &&
!format_desc->channel[0].pure_integer) {
LLVMValueRef tmps[LP_MAX_VECTOR_LENGTH/4];
- LLVMValueRef res;
- unsigned k;
+ LLVMValueRef res[LP_MAX_VECTOR_WIDTH / 128];
+ struct lp_type conv_type;
+ unsigned k, num_conv_src, num_conv_dst;
+
+ /*
+ * Note this path is generally terrible for fetching multiple pixels.
+ * We should make sure we cannot hit this code path for anything but
+ * single pixels.
+ */
/*
* Unpack a pixel at a time into a <4 x float> RGBA vector
__FUNCTION__, format_desc->short_name);
}
- lp_build_conv(gallivm,
- lp_float32_vec4_type(),
- type,
- tmps, num_pixels, &res, 1);
+ conv_type = lp_float32_vec4_type();
+ num_conv_src = num_pixels;
+ num_conv_dst = 1;
+
+ if (num_pixels % 8 == 0) {
+ lp_build_concat_n(gallivm, lp_float32_vec4_type(),
+ tmps, num_pixels, tmps, num_pixels / 2);
+ conv_type.length *= num_pixels / 4;
+ num_conv_src = 4 * num_pixels / 8;
+ if (type.width == 8 && type.floating == 0 && type.fixed == 0) {
+ /*
+ * FIXME: The fast float->unorm path (which is basically
+ * skipping the MIN/MAX which are extremely pointless in any
+ * case) requires that there's 2 destinations...
+ * In any case, we really should make sure we don't hit this
+ * code with multiple pixels for unorm8 dst types, it's
+ * completely hopeless even if we do hit the right conversion.
+ */
+ type.length /= num_pixels / 4;
+ num_conv_dst = num_pixels / 4;
+ }
+ }
- return lp_build_format_swizzle_aos(format_desc, &bld, res);
+ lp_build_conv(gallivm, conv_type, type,
+ tmps, num_conv_src, res, num_conv_dst);
+
+ if (num_pixels % 8 == 0 &&
+ (type.width == 8 && type.floating == 0 && type.fixed == 0)) {
+ lp_build_concat_n(gallivm, type, res, num_conv_dst, res, 1);
+ }
+
+ return lp_build_format_swizzle_aos(format_desc, &bld, res[0]);
}
/* If all channels are of same type and we are not using half-floats */
return tmp;
}
+ /*
+ * s3tc rgb formats
+ */
+
+ if (format_desc->layout == UTIL_FORMAT_LAYOUT_S3TC) {
+ struct lp_type tmp_type;
+ LLVMValueRef tmp;
+
+ memset(&tmp_type, 0, sizeof tmp_type);
+ tmp_type.width = 8;
+ tmp_type.length = num_pixels * 4;
+ tmp_type.norm = TRUE;
+
+ tmp = lp_build_fetch_s3tc_rgba_aos(gallivm,
+ format_desc,
+ num_pixels,
+ base_ptr,
+ offset,
+ i, j,
+ cache);
+
+ lp_build_conv(gallivm,
+ tmp_type, type,
+ &tmp, 1, &tmp, 1);
+
+ return tmp;
+ }
+
+ /*
+ * rgtc rgb formats
+ */
+
+ if (format_desc->layout == UTIL_FORMAT_LAYOUT_RGTC) {
+ struct lp_type tmp_type;
+ LLVMValueRef tmp;
+
+ memset(&tmp_type, 0, sizeof tmp_type);
+ tmp_type.width = 8;
+ tmp_type.length = num_pixels * 4;
+ tmp_type.norm = TRUE;
+ tmp_type.sign = (format_desc->format == PIPE_FORMAT_RGTC1_SNORM ||
+ format_desc->format == PIPE_FORMAT_RGTC2_SNORM ||
+ format_desc->format == PIPE_FORMAT_LATC1_SNORM ||
+ format_desc->format == PIPE_FORMAT_LATC2_SNORM);
+
+ tmp = lp_build_fetch_rgtc_rgba_aos(gallivm,
+ format_desc,
+ num_pixels,
+ base_ptr,
+ offset,
+ i, j,
+ cache);
+
+ lp_build_conv(gallivm,
+ tmp_type, type,
+ &tmp, 1, &tmp, 1);
+
+ return tmp;
+ }
+
/*
* Fallback to util_format_description::fetch_rgba_8unorm().
*/
- if (format_desc->fetch_rgba_8unorm &&
+ if (unpack->fetch_rgba_8unorm &&
!type.floating && type.width == 8 && !type.sign && type.norm) {
/*
* Fallback to calling util_format_description::fetch_rgba_8unorm.
arg_types[2] = i32t;
arg_types[3] = i32t;
function_type = LLVMFunctionType(ret_type, arg_types,
- Elements(arg_types), 0);
+ ARRAY_SIZE(arg_types), 0);
+ if (gallivm->cache)
+ gallivm->cache->dont_cache = true;
/* make const pointer for the C fetch_rgba_8unorm function */
function = lp_build_const_int_pointer(gallivm,
- func_to_pointer((func_pointer) format_desc->fetch_rgba_8unorm));
+ func_to_pointer((func_pointer) unpack->fetch_rgba_8unorm));
/* cast the callee pointer to the function's type */
function = LLVMBuildBitCast(builder, function,
args[3] = LLVMBuildExtractElement(builder, j, index, "");
}
- LLVMBuildCall(builder, function, args, Elements(args), "");
+ LLVMBuildCall(builder, function, args, ARRAY_SIZE(args), "");
tmp = LLVMBuildLoad(builder, tmp_ptr, "");
}
/*
- * Fallback to util_format_description::fetch_rgba_float().
+ * Fallback to fetch_rgba().
*/
- if (format_desc->fetch_rgba_float) {
+ util_format_fetch_rgba_func_ptr fetch_rgba =
+ util_format_fetch_rgba_func(format_desc->format);
+ if (fetch_rgba) {
/*
* Fallback to calling util_format_description::fetch_rgba_float.
*
}
/*
- * Declare and bind format_desc->fetch_rgba_float().
+ * Declare and bind unpack->fetch_rgba_float().
*/
{
arg_types[2] = i32t;
arg_types[3] = i32t;
+ if (gallivm->cache)
+ gallivm->cache->dont_cache = true;
function = lp_build_const_func_pointer(gallivm,
- func_to_pointer((func_pointer) format_desc->fetch_rgba_float),
+ func_to_pointer((func_pointer) fetch_rgba),
ret_type,
- arg_types, Elements(arg_types),
+ arg_types, ARRAY_SIZE(arg_types),
format_desc->short_name);
}
args[3] = LLVMBuildExtractElement(builder, j, index, "");
}
- LLVMBuildCall(builder, function, args, Elements(args), "");
+ LLVMBuildCall(builder, function, args, ARRAY_SIZE(args), "");
tmps[k] = LLVMBuildLoad(builder, tmp_ptr, "");
}
projects / mesa.git / blobdiff