#include "lp_tex_sample.h"
#include "lp_flush.h"
#include "lp_state_fs.h"
+#include "lp_rast.h"
/** Fragment shader number (for debugging) */
if (depth_mode & EARLY_DEPTH_TEST) {
lp_build_depth_stencil_load_swizzled(gallivm, type,
- zs_format_desc,
+ zs_format_desc, key->resource_1d,
depth_ptr, depth_stride,
&z_fb, &s_fb, loop_state.counter);
lp_build_depth_stencil_test(gallivm,
!simple_shader);
if (depth_mode & EARLY_DEPTH_WRITE) {
- lp_build_depth_stencil_write_swizzled(gallivm, type, zs_format_desc,
+ lp_build_depth_stencil_write_swizzled(gallivm, type,
+ zs_format_desc, key->resource_1d,
NULL, NULL, NULL, loop_state.counter,
depth_ptr, depth_stride,
z_value, s_value);
}
lp_build_depth_stencil_load_swizzled(gallivm, type,
- zs_format_desc,
+ zs_format_desc, key->resource_1d,
depth_ptr, depth_stride,
&z_fb, &s_fb, loop_state.counter);
!simple_shader);
/* Late Z write */
if (depth_mode & LATE_DEPTH_WRITE) {
- lp_build_depth_stencil_write_swizzled(gallivm, type, zs_format_desc,
+ lp_build_depth_stencil_write_swizzled(gallivm, type,
+ zs_format_desc, key->resource_1d,
NULL, NULL, NULL, loop_state.counter,
depth_ptr, depth_stride,
z_value, s_value);
* depth value, update from zs_value with the new mask value and
* write that out.
*/
- lp_build_depth_stencil_write_swizzled(gallivm, type, zs_format_desc,
+ lp_build_depth_stencil_write_swizzled(gallivm, type,
+ zs_format_desc, key->resource_1d,
&mask, z_fb, s_fb, loop_state.counter,
depth_ptr, depth_stride,
z_value, s_value);
*
* @param type fragment shader type (4x or 8x float)
* @param num_fs number of fs_src
+ * @param is_1d whether we're outputting to a 1d resource
* @param dst_channels number of output channels
* @param fs_src output from fragment shader
* @param dst pointer to store result
bool twiddle;
bool split;
- unsigned pixels = num_fs == 4 ? 1 : 2;
+ unsigned pixels = type.length / 4;
unsigned reorder_group;
unsigned src_channels;
unsigned src_count;
src_channels = dst_channels < 3 ? dst_channels : 4;
src_count = num_fs * src_channels;
- assert(pixels == 2 || num_fs == 4);
+ assert(pixels == 2 || pixels == 1);
assert(num_fs * src_channels <= Elements(src));
/*
}
+/**
+ * Checks if this format requires special handling due to required expansion
+ * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
+ * SoA conversion.
+ */
+static INLINE boolean
+format_expands_to_float_soa(const struct util_format_description *format_desc)
+{
+ if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT ||
+ format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
+ return true;
+ }
+ return false;
+}
+
+
/**
* Retrieves the type representing the memory layout for a format
*
unsigned i;
unsigned chan;
- if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT) {
+ if (format_expands_to_float_soa(format_desc)) {
/* just make this a 32bit uint */
type->floating = false;
type->fixed = false;
unsigned i;
unsigned chan;
- if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT) {
+ if (format_expands_to_float_soa(format_desc)) {
/* always use ordinary floats for blending */
type->floating = true;
type->fixed = false;
/**
- * Scale a normalized value from src_bits to dst_bits
+ * Scale a normalized value from src_bits to dst_bits.
+ *
+ * The exact calculation is
+ *
+ * dst = iround(src * dst_mask / src_mask)
+ *
+ * or with integer rounding
+ *
+ * dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
+ *
+ * where
+ *
+ * src_mask = (1 << src_bits) - 1
+ * dst_mask = (1 << dst_bits) - 1
+ *
+ * but we try to avoid division and multiplication through shifts.
*/
static INLINE LLVMValueRef
scale_bits(struct gallivm_state *gallivm,
LLVMValueRef result = src;
if (dst_bits < src_bits) {
- /* Scale down by LShr */
- result = LLVMBuildLShr(builder,
- src,
- lp_build_const_int_vec(gallivm, src_type, src_bits - dst_bits),
- "");
+ int delta_bits = src_bits - dst_bits;
+
+ if (delta_bits <= dst_bits) {
+ /*
+ * Approximate the rescaling with a single shift.
+ *
+ * This gives the wrong rounding.
+ */
+
+ result = LLVMBuildLShr(builder,
+ src,
+ lp_build_const_int_vec(gallivm, src_type, delta_bits),
+ "");
+
+ } else {
+ /*
+ * Try more accurate rescaling.
+ */
+
+ /*
+ * Drop the least significant bits to make space for the multiplication.
+ *
+ * XXX: A better approach would be to use a wider integer type as intermediate. But
+ * this is enough to convert alpha from 16bits -> 2 when rendering to
+ * PIPE_FORMAT_R10G10B10A2_UNORM.
+ */
+ result = LLVMBuildLShr(builder,
+ src,
+ lp_build_const_int_vec(gallivm, src_type, dst_bits),
+ "");
+
+
+ result = LLVMBuildMul(builder,
+ result,
+ lp_build_const_int_vec(gallivm, src_type, (1LL << dst_bits) - 1),
+ "");
+
+ /*
+ * Add a rounding term before the division.
+ *
+ * TODO: Handle signed integers too.
+ */
+ if (!src_type.sign) {
+ result = LLVMBuildAdd(builder,
+ result,
+ lp_build_const_int_vec(gallivm, src_type, (1LL << (delta_bits - 1))),
+ "");
+ }
+
+ /*
+ * Approximate the division by src_mask with a src_bits shift.
+ *
+ * Given the src has already been shifted by dst_bits, all we need
+ * to do is to shift by the difference.
+ */
+
+ result = LLVMBuildLShr(builder,
+ result,
+ lp_build_const_int_vec(gallivm, src_type, delta_bits),
+ "");
+ }
+
} else if (dst_bits > src_bits) {
/* Scale up bits */
int db = dst_bits - src_bits;
*/
static void
convert_to_blend_type(struct gallivm_state *gallivm,
+ unsigned block_size,
const struct util_format_description *src_fmt,
struct lp_type src_type,
struct lp_type dst_type,
struct lp_type blend_type;
struct lp_type mem_type;
unsigned i, j, k;
- unsigned pixels = 16 / num_srcs;
+ unsigned pixels = block_size / num_srcs;
bool is_arith;
/*
- * full custom path for packed floats - none of the later functions would do
- * anything useful, and given the lp_type representation they can't be fixed.
+ * full custom path for packed floats and srgb formats - none of the later
+ * functions would do anything useful, and given the lp_type representation they
+ * can't be fixed. Should really have some SoA blend path for these kind of
+ * formats rather than hacking them in here.
*/
- if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
+ if (format_expands_to_float_soa(src_fmt)) {
LLVMValueRef tmpsrc[4];
/*
* This is pretty suboptimal for this case blending in SoA would be much
assert(dst_type.floating);
assert(dst_type.width == 32);
assert(dst_type.length % 4 == 0);
+ assert(num_srcs % 4 == 0);
+
for (i = 0; i < 4; i++) {
tmpsrc[i] = src[i];
}
for (i = 0; i < num_srcs / 4; i++) {
LLVMValueRef tmpsoa[4];
LLVMValueRef tmps = tmpsrc[i];
- if (num_srcs == 8) {
+ if (dst_type.length == 8) {
LLVMValueRef shuffles[8];
unsigned j;
/* fetch was 4 values but need 8-wide output values */
tmps = LLVMBuildShuffleVector(builder, tmps, tmps,
LLVMConstVector(shuffles, 8), "");
}
- lp_build_r11g11b10_to_float(gallivm, tmps, tmpsoa);
+ if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
+ lp_build_r11g11b10_to_float(gallivm, tmps, tmpsoa);
+ }
+ else {
+ lp_build_unpack_rgba_soa(gallivm, src_fmt, dst_type, tmps, tmpsoa);
+ }
lp_build_transpose_aos(gallivm, dst_type, tmpsoa, &src[i * 4]);
}
return;
for (i = 0; i < num_srcs; ++i) {
LLVMValueRef chans[4];
LLVMValueRef res = NULL;
- unsigned sa = 0;
dst[i] = LLVMBuildZExt(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
for (j = 0; j < src_fmt->nr_channels; ++j) {
unsigned mask = 0;
+ unsigned sa = src_fmt->channel[j].shift;
+#ifdef PIPE_ARCH_LITTLE_ENDIAN
+ unsigned from_lsb = j;
+#else
+ unsigned from_lsb = src_fmt->nr_channels - j - 1;
+#endif
for (k = 0; k < src_fmt->channel[j].size; ++k) {
mask |= 1 << k;
/* Insert bits into correct position */
chans[j] = LLVMBuildShl(builder,
chans[j],
- lp_build_const_int_vec(gallivm, src_type, j * blend_type.width),
+ lp_build_const_int_vec(gallivm, src_type, from_lsb * blend_type.width),
"");
- sa += src_fmt->channel[j].size;
-
if (j == 0) {
res = chans[j];
} else {
*/
static void
convert_from_blend_type(struct gallivm_state *gallivm,
+ unsigned block_size,
const struct util_format_description *src_fmt,
struct lp_type src_type,
struct lp_type dst_type,
struct lp_type mem_type;
struct lp_type blend_type;
LLVMBuilderRef builder = gallivm->builder;
- unsigned pixels = 16 / num_srcs;
+ unsigned pixels = block_size / num_srcs;
bool is_arith;
/*
- * full custom path for packed floats - none of the later functions would do
- * anything useful, and given the lp_type representation they can't be fixed.
+ * full custom path for packed floats and srgb formats - none of the later
+ * functions would do anything useful, and given the lp_type representation they
+ * can't be fixed. Should really have some SoA blend path for these kind of
+ * formats rather than hacking them in here.
*/
- if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
+ if (format_expands_to_float_soa(src_fmt)) {
/*
* This is pretty suboptimal for this case blending in SoA would be much
* better - we need to transpose the AoS values back to SoA values for
assert(src_type.width == 32);
assert(src_type.length % 4 == 0);
assert(dst_type.width == 32);
+
for (i = 0; i < num_srcs / 4; i++) {
LLVMValueRef tmpsoa[4], tmpdst;
lp_build_transpose_aos(gallivm, src_type, &src[i * 4], tmpsoa);
- tmpdst = lp_build_float_to_r11g11b10(gallivm, tmpsoa);
- if (num_srcs == 8) {
+ /* really really need SoA here */
+
+ if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
+ tmpdst = lp_build_float_to_r11g11b10(gallivm, tmpsoa);
+ }
+ else {
+ tmpdst = lp_build_float_to_srgb_packed(gallivm, src_fmt,
+ src_type, tmpsoa);
+ }
+
+ if (src_type.length == 8) {
LLVMValueRef tmpaos, shuffles[8];
unsigned j;
/*
for (i = 0; i < num_srcs; ++i) {
LLVMValueRef chans[4];
LLVMValueRef res = NULL;
- unsigned sa = 0;
dst[i] = LLVMBuildBitCast(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
for (j = 0; j < src_fmt->nr_channels; ++j) {
unsigned mask = 0;
+ unsigned sa = src_fmt->channel[j].shift;
+#ifdef PIPE_ARCH_LITTLE_ENDIAN
+ unsigned from_lsb = j;
+#else
+ unsigned from_lsb = src_fmt->nr_channels - j - 1;
+#endif
assert(blend_type.width > src_fmt->channel[j].size);
/* Extract bits */
chans[j] = LLVMBuildLShr(builder,
dst[i],
- lp_build_const_int_vec(gallivm, src_type, j * blend_type.width),
+ lp_build_const_int_vec(gallivm, src_type, from_lsb * blend_type.width),
"");
chans[j] = LLVMBuildAnd(builder,
row_type.length = alpha_type.length;
/* Twiddle the alpha to match pixels */
- lp_bld_quad_twiddle(gallivm, alpha_type, src_alpha, 4, src_alpha);
+ lp_bld_quad_twiddle(gallivm, alpha_type, src_alpha, block_height, src_alpha);
- for (i = 0; i < 4; ++i) {
+ /*
+ * TODO this should use single lp_build_conv call for
+ * src_count == 1 && dst_channels == 1 case (dropping the concat below)
+ */
+ for (i = 0; i < block_height; ++i) {
lp_build_conv(gallivm, alpha_type, row_type, &src_alpha[i], 1, &src_alpha[i], 1);
}
row_type.length = length;
/* If only one channel we can only need the single alpha value per pixel */
- if (src_count == 1) {
- assert(dst_channels == 1);
+ if (src_count == 1 && dst_channels == 1) {
- lp_build_concat_n(gallivm, alpha_type, src_alpha, 4, src_alpha, src_count);
+ lp_build_concat_n(gallivm, alpha_type, src_alpha, block_height, src_alpha, src_count);
} else {
/* If there are more srcs than rows then we need to split alpha up */
if (src_count > block_height) {
unsigned pixels = block_size / src_count;
unsigned idx = i - 1;
- src_alpha[idx] = lp_build_extract_range(gallivm, src_alpha[(idx * pixels) / 4], (idx * pixels) % 4, pixels);
+ src_alpha[idx] = lp_build_extract_range(gallivm, src_alpha[(idx * pixels) / 4],
+ (idx * pixels) % 4, pixels);
}
}
boolean do_branch)
{
const unsigned alpha_channel = 3;
- const unsigned block_width = 4;
- const unsigned block_height = 4;
+ const unsigned block_width = LP_RASTER_BLOCK_SIZE;
+ const unsigned block_height = LP_RASTER_BLOCK_SIZE;
const unsigned block_size = block_width * block_height;
const unsigned lp_integer_vector_width = 128;
LLVMValueRef blend_alpha;
LLVMValueRef i32_zero;
LLVMValueRef check_mask;
+ LLVMValueRef undef_src_val;
struct lp_build_mask_context mask_ctx;
struct lp_type mask_type;
const boolean dual_source_blend = variant->key.blend.rt[0].blend_enable &&
util_blend_state_is_dual(&variant->key.blend, 0);
+ const boolean is_1d = variant->key.resource_1d;
+ unsigned num_fullblock_fs = is_1d ? 2 * num_fs : num_fs;
+
mask_type = lp_int32_vec4_type();
mask_type.length = fs_type.length;
- /* Compute the alignment of the destination pointer in bytes */
-#if 0
- dst_alignment = (block_width * out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
-#else
- /* FIXME -- currently we're fetching pixels one by one, instead of row by row */
- dst_alignment = (1 * out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
-#endif
- /* Force power-of-two alignment by extracting only the least-significant-bit */
- dst_alignment = 1 << (ffs(dst_alignment) - 1);
- /* Resource base and stride pointers are aligned to 16 bytes, so that's the maximum alignment we can guarantee */
- dst_alignment = MIN2(dst_alignment, 16);
+ for (i = num_fs; i < num_fullblock_fs; i++) {
+ fs_mask[i] = lp_build_zero(gallivm, mask_type);
+ }
/* Do not bother executing code when mask is empty.. */
if (do_branch) {
check_mask = LLVMConstNull(lp_build_int_vec_type(gallivm, mask_type));
- for (i = 0; i < num_fs; ++i) {
+ for (i = 0; i < num_fullblock_fs; ++i) {
check_mask = LLVMBuildOr(builder, check_mask, fs_mask[i], "");
}
partial_mask |= !variant->opaque;
i32_zero = lp_build_const_int32(gallivm, 0);
+#if HAVE_LLVM < 0x0302
+ /*
+ * undef triggers a crash in LLVMBuildTrunc in convert_from_blend_type in some
+ * cases (seen with r10g10b10a2, 128bit wide vectors) (only used for 1d case).
+ */
+ undef_src_val = lp_build_zero(gallivm, fs_type);
+#else
+ undef_src_val = lp_build_undef(gallivm, fs_type);
+#endif
+
+
/* Get type from output format */
lp_blend_type_from_format_desc(out_format_desc, &row_type);
lp_mem_type_from_format_desc(out_format_desc, &dst_type);
}
}
- if (out_format == PIPE_FORMAT_R11G11B10_FLOAT) {
- /* the code above can't work for layout_other */
+ if (format_expands_to_float_soa(out_format_desc)) {
+ /*
+ * the code above can't work for layout_other
+ * for srgb it would sort of work but we short-circuit swizzles, etc.
+ * as that is done as part of unpack / pack.
+ */
dst_channels = 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
has_alpha = true;
swizzle[0] = 0;
/*
* Load shader output
*/
- for (i = 0; i < num_fs; ++i) {
+ for (i = 0; i < num_fullblock_fs; ++i) {
/* Always load alpha for use in blending */
- LLVMValueRef alpha = LLVMBuildLoad(builder, fs_out_color[rt][alpha_channel][i], "");
+ LLVMValueRef alpha;
+ if (i < num_fs) {
+ alpha = LLVMBuildLoad(builder, fs_out_color[rt][alpha_channel][i], "");
+ }
+ else {
+ alpha = undef_src_val;
+ }
/* Load each channel */
for (j = 0; j < dst_channels; ++j) {
assert(swizzle[j] < 4);
- fs_src[i][j] = LLVMBuildLoad(builder, fs_out_color[rt][swizzle[j]][i], "");
+ if (i < num_fs) {
+ fs_src[i][j] = LLVMBuildLoad(builder, fs_out_color[rt][swizzle[j]][i], "");
+ }
+ else {
+ fs_src[i][j] = undef_src_val;
+ }
}
/* If 3 channels then pad to include alpha for 4 element transpose */
}
if (dual_source_blend) {
/* same as above except different src/dst, skip masks and comments... */
- for (i = 0; i < num_fs; ++i) {
- LLVMValueRef alpha = LLVMBuildLoad(builder, fs_out_color[1][alpha_channel][i], "");
+ for (i = 0; i < num_fullblock_fs; ++i) {
+ LLVMValueRef alpha;
+ if (i < num_fs) {
+ alpha = LLVMBuildLoad(builder, fs_out_color[1][alpha_channel][i], "");
+ }
+ else {
+ alpha = undef_src_val;
+ }
for (j = 0; j < dst_channels; ++j) {
assert(swizzle[j] < 4);
- fs_src1[i][j] = LLVMBuildLoad(builder, fs_out_color[1][swizzle[j]][i], "");
+ if (i < num_fs) {
+ fs_src1[i][j] = LLVMBuildLoad(builder, fs_out_color[1][swizzle[j]][i], "");
+ }
+ else {
+ fs_src1[i][j] = undef_src_val;
+ }
}
if (dst_channels == 3 && !has_alpha) {
fs_src1[i][3] = alpha;
*/
fs_type.floating = 0;
fs_type.sign = dst_type.sign;
- for (i = 0; i < num_fs; ++i) {
+ for (i = 0; i < num_fullblock_fs; ++i) {
for (j = 0; j < dst_channels; ++j) {
fs_src[i][j] = LLVMBuildBitCast(builder, fs_src[i][j],
lp_build_vec_type(gallivm, fs_type), "");
/*
* Pixel twiddle from fragment shader order to memory order
*/
- src_count = generate_fs_twiddle(gallivm, fs_type, num_fs, dst_channels, fs_src, src, pad_inline);
+ src_count = generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs,
+ dst_channels, fs_src, src, pad_inline);
if (dual_source_blend) {
- generate_fs_twiddle(gallivm, fs_type, num_fs, dst_channels, fs_src1, src1, pad_inline);
+ generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs, dst_channels,
+ fs_src1, src1, pad_inline);
}
src_channels = dst_channels < 3 ? dst_channels : 4;
- if (src_count != num_fs * src_channels) {
- unsigned ds = src_count / (num_fs * src_channels);
+ if (src_count != num_fullblock_fs * src_channels) {
+ unsigned ds = src_count / (num_fullblock_fs * src_channels);
row_type.length /= ds;
fs_type.length = row_type.length;
}
unsigned bits = row_type.width * row_type.length;
unsigned combined;
+ assert(src_count >= (vector_width / bits));
+
dst_count = src_count / (vector_width / bits);
+
combined = lp_build_concat_n(gallivm, row_type, src, src_count, src, dst_count);
if (dual_source_blend) {
lp_build_concat_n(gallivm, row_type, src1, src_count, src1, dst_count);
/* Convert */
lp_build_conv(gallivm, fs_type, blend_type, &blend_color, 1, &blend_color, 1);
+ if (out_format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
+ /*
+ * since blending is done with floats, there was no conversion.
+ * However, the rules according to fixed point renderbuffers still
+ * apply, that is we must clamp inputs to 0.0/1.0.
+ * (This would apply to separate alpha conversion too but we currently
+ * force has_alpha to be true.)
+ * TODO: should skip this with "fake" blend, since post-blend conversion
+ * will clamp anyway.
+ * TODO: could also skip this if fragment color clamping is enabled. We
+ * don't support it natively so it gets baked into the shader however, so
+ * can't really tell here.
+ */
+ struct lp_build_context f32_bld;
+ assert(row_type.floating);
+ lp_build_context_init(&f32_bld, gallivm, row_type);
+ for (i = 0; i < src_count; i++) {
+ src[i] = lp_build_clamp(&f32_bld, src[i], f32_bld.zero, f32_bld.one);
+ }
+ if (dual_source_blend) {
+ for (i = 0; i < src_count; i++) {
+ src1[i] = lp_build_clamp(&f32_bld, src1[i], f32_bld.zero, f32_bld.one);
+ }
+ }
+ /* probably can't be different than row_type but better safe than sorry... */
+ lp_build_context_init(&f32_bld, gallivm, blend_type);
+ blend_color = lp_build_clamp(&f32_bld, blend_color, f32_bld.zero, f32_bld.one);
+ }
+
/* Extract alpha */
blend_alpha = lp_build_extract_broadcast(gallivm, blend_type, row_type, blend_color, lp_build_const_int32(gallivm, 3));
/*
* Mask conversion
*/
- lp_bld_quad_twiddle(gallivm, mask_type, &src_mask[0], 4, &src_mask[0]);
+ lp_bld_quad_twiddle(gallivm, mask_type, &src_mask[0], block_height, &src_mask[0]);
if (src_count < block_height) {
lp_build_concat_n(gallivm, mask_type, src_mask, 4, src_mask, src_count);
unsigned pixels = block_size / src_count;
unsigned idx = i - 1;
- src_mask[idx] = lp_build_extract_range(gallivm, src_mask[(idx * pixels) / 4], (idx * pixels) % 4, pixels);
+ src_mask[idx] = lp_build_extract_range(gallivm, src_mask[(idx * pixels) / 4],
+ (idx * pixels) % 4, pixels);
}
}
dst_count = src_count;
}
- dst_type.length *= 16 / dst_count;
+ dst_type.length *= block_size / dst_count;
- if (out_format == PIPE_FORMAT_R11G11B10_FLOAT) {
+ if (format_expands_to_float_soa(out_format_desc)) {
/*
* we need multiple values at once for the conversion, so can as well
* load them vectorized here too instead of concatenating later.
dst_type.length = block_width;
}
- load_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
- dst, dst_type, dst_count, dst_alignment);
+ /*
+ * Compute the alignment of the destination pointer in bytes
+ * We fetch 1-4 pixels, if the format has pot alignment then those fetches
+ * are always aligned by MIN2(16, fetch_width) except for buffers (not
+ * 1d tex but can't distinguish here) so need to stick with per-pixel
+ * alignment in this case.
+ */
+ if (is_1d) {
+ dst_alignment = (out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
+ }
+ else {
+ dst_alignment = dst_type.length * dst_type.width / 8;
+ }
+ /* Force power-of-two alignment by extracting only the least-significant-bit */
+ dst_alignment = 1 << (ffs(dst_alignment) - 1);
+ /*
+ * Resource base and stride pointers are aligned to 16 bytes, so that's
+ * the maximum alignment we can guarantee
+ */
+ dst_alignment = MIN2(16, dst_alignment);
+
+ if (is_1d) {
+ load_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
+ dst, dst_type, dst_count / 4, dst_alignment);
+ for (i = dst_count / 4; i < dst_count; i++) {
+ dst[i] = lp_build_undef(gallivm, dst_type);
+ }
+
+ }
+ else {
+ load_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
+ dst, dst_type, dst_count, dst_alignment);
+ }
/*
* It seems some cleanup could be done here (like skipping conversion/blend
* when not needed).
*/
- convert_to_blend_type(gallivm, out_format_desc, dst_type, row_type, dst, src_count);
+ convert_to_blend_type(gallivm, block_size, out_format_desc, dst_type, row_type, dst, src_count);
+ /*
+ * FIXME: Really should get logic ops / masks out of generic blend / row
+ * format. Logic ops will definitely not work on the blend float format
+ * used for SRGB here and I think OpenGL expects this to work as expected
+ * (that is incoming values converted to srgb then logic op applied).
+ */
for (i = 0; i < src_count; ++i) {
dst[i] = lp_build_blend_aos(gallivm,
&variant->key.blend,
pad_inline ? 4 : dst_channels);
}
- convert_from_blend_type(gallivm, out_format_desc, row_type, dst_type, dst, src_count);
+ convert_from_blend_type(gallivm, block_size, out_format_desc, row_type, dst_type, dst, src_count);
/* Split the blend rows back to memory rows */
if (dst_count > src_count) {
src_count *= 2;
}
-
/*
* Store blend result to memory
*/
- store_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
- dst, dst_type, dst_count, dst_alignment);
+ if (is_1d) {
+ store_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
+ dst, dst_type, dst_count / 4, dst_alignment);
+ }
+ else {
+ store_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
+ dst, dst_type, dst_count, dst_alignment);
+ }
if (do_branch) {
lp_build_mask_end(&mask_ctx);
fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
fs_type.width = 32; /* 32-bit float */
fs_type.length = MIN2(lp_native_vector_width / 32, 16); /* n*4 elements per vector */
- num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
memset(&blend_type, 0, sizeof blend_type);
blend_type.floating = FALSE; /* values are integers */
/* code generated texture sampling */
sampler = lp_llvm_sampler_soa_create(key->state, context_ptr);
+ num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
+ /* for 1d resources only run "upper half" of stamp */
+ if (key->resource_1d)
+ num_fs /= 2;
+
{
LLVMValueRef num_loop = lp_build_const_int32(gallivm, num_fs);
LLVMTypeRef mask_type = lp_build_int_vec_type(gallivm, fs_type);
!key->alpha.enabled &&
!key->depth.enabled &&
!shader->info.base.uses_kill
- ? TRUE : FALSE;
+ ? TRUE : FALSE;
+
+ if ((!shader || shader->info.base.num_tokens <= 1) &&
+ !key->depth.enabled && !key->stencil[0].enabled) {
+ variant->ps_inv_multiplier = 0;
+ } else {
+ variant->ps_inv_multiplier = 1;
+ }
if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
lp_debug_fs_variant(variant);
* Return the blend factor equivalent to a destination alpha of one.
*/
static INLINE unsigned
-force_dst_alpha_one(unsigned factor)
+force_dst_alpha_one(unsigned factor, boolean clamped_zero)
{
switch(factor) {
case PIPE_BLENDFACTOR_DST_ALPHA:
case PIPE_BLENDFACTOR_INV_DST_ALPHA:
return PIPE_BLENDFACTOR_ZERO;
case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
- return PIPE_BLENDFACTOR_ZERO;
+ if (clamped_zero)
+ return PIPE_BLENDFACTOR_ZERO;
+ else
+ return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE;
}
return factor;
key->zsbuf_format = zsbuf_format;
memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
}
+ if (llvmpipe_resource_is_1d(lp->framebuffer.zsbuf->texture)) {
+ key->resource_1d = TRUE;
+ }
}
/* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
/* alpha.ref_value is passed in jit_context */
key->flatshade = lp->rasterizer->flatshade;
- if (lp->active_occlusion_query) {
+ if (lp->active_occlusion_queries) {
key->occlusion_count = TRUE;
}
key->cbuf_format[i] = format;
+ /*
+ * Figure out if this is a 1d resource. Note that OpenGL allows crazy
+ * mixing of 2d textures with height 1 and 1d textures, so make sure
+ * we pick 1d if any cbuf or zsbuf is 1d.
+ */
+ if (llvmpipe_resource_is_1d(lp->framebuffer.cbufs[0]->texture)) {
+ key->resource_1d = TRUE;
+ }
+
format_desc = util_format_description(format);
assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
*/
if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W ||
format_desc->swizzle[3] == format_desc->swizzle[0]) {
- blend_rt->rgb_src_factor = force_dst_alpha_one(blend_rt->rgb_src_factor);
- blend_rt->rgb_dst_factor = force_dst_alpha_one(blend_rt->rgb_dst_factor);
+ /* Doesn't cover mixed snorm/unorm but can't render to them anyway */
+ boolean clamped_zero = !util_format_is_float(format) &&
+ !util_format_is_snorm(format);
+ blend_rt->rgb_src_factor = force_dst_alpha_one(blend_rt->rgb_src_factor,
+ clamped_zero);
+ blend_rt->rgb_dst_factor = force_dst_alpha_one(blend_rt->rgb_dst_factor,
+ clamped_zero);
blend_rt->alpha_func = blend_rt->rgb_func;
blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
}
+
+/*
+ * Rasterization is disabled if there is no pixel shader and
+ * both depth and stencil testing are disabled:
+ * http://msdn.microsoft.com/en-us/library/windows/desktop/bb205125
+ */
+boolean
+llvmpipe_rasterization_disabled(struct llvmpipe_context *lp)
+{
+ boolean null_fs = !lp->fs || lp->fs->info.base.num_tokens <= 1;
+
+ return (null_fs &&
+ !lp->depth_stencil->depth.enabled &&
+ !lp->depth_stencil->stencil[0].enabled);
+}