Big opimizations, almost 2x performance increase!
This commit is contained in:
Krzosa Karol
@@ -1,3 +1,6 @@
|
||||
#define I(x,i) (((F32 *)&x)[i])
|
||||
#define Is(x,i) (((S32 *)&x)[i])
|
||||
|
||||
function
|
||||
void draw_triangle_nearest_a(Bitmap* dst, F32 *depth_buffer, Bitmap *src, Vec3 light_direction,
|
||||
Vec4 p0, Vec4 p1, Vec4 p2,
|
||||
@@ -1313,295 +1316,3 @@ void draw_triangle_nearest_g(Bitmap* dst, F32 *depth_buffer, Bitmap *src, Vec3 l
|
||||
filled_pixel_total_time += end_time - fill_pixels_begin;
|
||||
filled_pixel_count += (max_x - min_x)*(max_y - min_y);
|
||||
}
|
||||
|
||||
function
|
||||
void draw_triangle_nearest_iter(Bitmap* dst, F32 *depth_buffer, Bitmap *src, Vec3 light_direction,
|
||||
Vec4 p0, Vec4 p1, Vec4 p2,
|
||||
Vec2 tex0, Vec2 tex1, Vec2 tex2,
|
||||
Vec3 norm0, Vec3 norm1, Vec3 norm2) {
|
||||
if(src->pixels == 0) return;
|
||||
|
||||
PROFILE_SCOPE(draw_triangle);
|
||||
|
||||
F32 min_x1 = (F32)(min(p0.x, min(p1.x, p2.x)));
|
||||
F32 min_y1 = (F32)(min(p0.y, min(p1.y, p2.y)));
|
||||
F32 max_x1 = (F32)(max(p0.x, max(p1.x, p2.x)));
|
||||
F32 max_y1 = (F32)(max(p0.y, max(p1.y, p2.y)));
|
||||
|
||||
S64 min_x = (S64)max(0.f, floor(min_x1));
|
||||
S64 min_y = (S64)max(0.f, floor(min_y1));
|
||||
S64 max_x = (S64)min((F32)dst->x, ceil(max_x1));
|
||||
S64 max_y = (S64)min((F32)dst->y, ceil(max_y1));
|
||||
|
||||
if (min_y >= max_y) return;
|
||||
if (min_x >= max_x) return;
|
||||
|
||||
F32 dy10 = (p1.y - p0.y);
|
||||
F32 dy21 = (p2.y - p1.y);
|
||||
F32 dy02 = (p0.y - p2.y);
|
||||
|
||||
F32 dx10 = (p1.x - p0.x);
|
||||
F32 dx21 = (p2.x - p1.x);
|
||||
F32 dx02 = (p0.x - p2.x);
|
||||
|
||||
F32 C0 = dy10 * (p0.x) - dx10 * (p0.y);
|
||||
F32 C1 = dy21 * (p1.x) - dx21 * (p1.y);
|
||||
F32 C2 = dy02 * (p2.x) - dx02 * (p2.y);
|
||||
|
||||
F32 Cy0 = dy10 * min_x - dx10 * min_y - C0;
|
||||
F32 Cy1 = dy21 * min_x - dx21 * min_y - C1;
|
||||
F32 Cy2 = dy02 * min_x - dx02 * min_y - C2;
|
||||
|
||||
F32x8 var255 = _mm256_set1_ps(255);
|
||||
F32x8 var0 = _mm256_set1_ps(0);
|
||||
F32x8 var1 = _mm256_set1_ps(1);
|
||||
F32x8 inv255 = _mm256_div_ps(var1, var255);
|
||||
F32x8 var_max_x = _mm256_set1_ps(max_x);
|
||||
F32x8 var07 = _mm256_set_ps(7,6,5,4,3,2,1,0);
|
||||
|
||||
F32x8 var_1_8 = _mm256_set_ps(8,7,6,5,4,3,2,1);
|
||||
F32x8 Dy10 = _mm256_mul_ps(_mm256_set1_ps(dy10), var_1_8);
|
||||
F32x8 Dy21 = _mm256_mul_ps(_mm256_set1_ps(dy21), var_1_8);
|
||||
F32x8 Dy02 = _mm256_mul_ps(_mm256_set1_ps(dy02), var_1_8);
|
||||
|
||||
F32x8 var_src_x_minus_one = _mm256_set1_ps(src->x-1);
|
||||
F32x8 var_src_y_minus_one = _mm256_set1_ps(src->y-1);
|
||||
S32x8 var_src_y_minus_one_int = _mm256_set1_epi32(src->y-1);
|
||||
S32x8 var_src_x_int = _mm256_set1_epi32(src->x);
|
||||
|
||||
S32x8 var_0xff000000 = _mm256_set1_epi32(0xff000000);
|
||||
S32x8 var_0x00ff0000 = _mm256_set1_epi32(0x00ff0000);
|
||||
S32x8 var_0x0000ff00 = _mm256_set1_epi32(0x0000ff00);
|
||||
S32x8 var_0x000000ff = _mm256_set1_epi32(0x000000ff);
|
||||
|
||||
F32x8 var_tex0x = _mm256_set1_ps(tex0.x);
|
||||
F32x8 var_tex1x = _mm256_set1_ps(tex1.x);
|
||||
F32x8 var_tex2x = _mm256_set1_ps(tex2.x);
|
||||
F32x8 var_tex0y = _mm256_set1_ps(tex0.y);
|
||||
F32x8 var_tex1y = _mm256_set1_ps(tex1.y);
|
||||
F32x8 var_tex2y = _mm256_set1_ps(tex2.y);
|
||||
|
||||
F32x8 var_p0w = _mm256_set1_ps(p0.w);
|
||||
F32x8 var_p1w = _mm256_set1_ps(p1.w);
|
||||
F32x8 var_p2w = _mm256_set1_ps(p2.w);
|
||||
F32x8 one_over_p0w = _mm256_set1_ps(1.f / p0.w);
|
||||
F32x8 one_over_p1w = _mm256_set1_ps(1.f / p1.w);
|
||||
F32x8 one_over_p2w = _mm256_set1_ps(1.f / p2.w);
|
||||
|
||||
U32 *destination = dst->pixels + dst->x*min_y;
|
||||
F32 area = (p1.y - p0.y) * (p2.x - p0.x) - (p1.x - p0.x) * (p2.y - p0.y);
|
||||
F32x8 area8 = _mm256_set1_ps(area);
|
||||
U64 fill_pixels_begin = __rdtsc();
|
||||
for (S64 y = min_y; y < max_y; y++) {
|
||||
F32x8 Cx0 = _mm256_set1_ps(Cy0);
|
||||
F32x8 Cx1 = _mm256_set1_ps(Cy1);
|
||||
F32x8 Cx2 = _mm256_set1_ps(Cy2);
|
||||
|
||||
for (S64 x8 = min_x; x8 < max_x; x8+=8) {
|
||||
{
|
||||
F32x8 i0 = _mm256_set1_ps(I(Cx0, 7));
|
||||
Cx0 = _mm256_add_ps(i0, Dy10);
|
||||
|
||||
F32x8 i2 = _mm256_set1_ps(I(Cx1, 7));
|
||||
Cx1 = _mm256_add_ps(i2, Dy21);
|
||||
|
||||
F32x8 i4 = _mm256_set1_ps(I(Cx2, 7));
|
||||
Cx2 = _mm256_add_ps(i4, Dy02);
|
||||
}
|
||||
|
||||
|
||||
F32x8 should_fill;
|
||||
F32x8 i11 = _mm256_set1_ps(x8);
|
||||
F32x8 i12 = _mm256_add_ps(i11, var07);
|
||||
F32x8 i13 = _mm256_cmp_ps(i12, var_max_x, _CMP_LT_OQ);
|
||||
|
||||
F32x8 i6 = _mm256_cmp_ps(Cx0, var0, _CMP_GE_OQ);
|
||||
F32x8 i7 = _mm256_cmp_ps(Cx1, var0, _CMP_GE_OQ);
|
||||
F32x8 i8 = _mm256_cmp_ps(Cx2, var0, _CMP_GE_OQ);
|
||||
F32x8 i9 = _mm256_and_ps(i6, i7);
|
||||
F32x8 i10 = _mm256_and_ps(i9, i8);
|
||||
should_fill = _mm256_and_ps(i13, i10);
|
||||
|
||||
F32x8 w0 = _mm256_div_ps(Cx1, area8);
|
||||
F32x8 w1 = _mm256_div_ps(Cx2, area8);
|
||||
F32x8 w2 = _mm256_div_ps(Cx0, area8);
|
||||
|
||||
// @Note: We could do: interpolated_w = 1.f / interpolated_w to get proper depth
|
||||
// but why waste an instruction, the smaller the depth value the farther the object
|
||||
F32x8 interpolated_w;
|
||||
F32x8 i14 = _mm256_mul_ps(one_over_p0w, w0); //
|
||||
F32x8 i15 = _mm256_mul_ps(one_over_p1w, w1);
|
||||
F32x8 i16 = _mm256_mul_ps(one_over_p2w, w2);
|
||||
F32x8 i17 = _mm256_add_ps(i14, i15);
|
||||
F32x8 i18 = _mm256_add_ps(i16, i17);
|
||||
interpolated_w = {i18};
|
||||
|
||||
F32 *depth_pointer = (depth_buffer + (x8 + y * dst->x));
|
||||
F32x8 depth = _mm256_loadu_ps((float *)depth_pointer);
|
||||
|
||||
F32x8 should_fill_term = _mm256_cmp_ps(depth, interpolated_w, _CMP_LT_OQ);
|
||||
should_fill = _mm256_and_ps(should_fill, should_fill_term);
|
||||
|
||||
#if 0
|
||||
// If all pixels are not going to get drawn then opt out
|
||||
// Seems to decrease perf
|
||||
F32x8 compare_with_zero = _mm256_cmpeq_epi32(should_fill, var0);
|
||||
int mask = _mm256_movemask_epi8(compare_with_zero);
|
||||
if(mask == 1) continue;
|
||||
#endif
|
||||
|
||||
F32x8 invw0 = _mm256_div_ps(w0, var_p0w);
|
||||
F32x8 invw1 = _mm256_div_ps(w1, var_p1w);
|
||||
F32x8 invw2 = _mm256_div_ps(w2, var_p2w);
|
||||
|
||||
F32x8 u_term0 = _mm256_mul_ps(var_tex0x, invw0);
|
||||
F32x8 u_term1 = _mm256_mul_ps(var_tex1x, invw1);
|
||||
F32x8 u_term2 = _mm256_mul_ps(var_tex2x, invw2);
|
||||
F32x8 u_term3 = _mm256_add_ps(u_term0, u_term1);
|
||||
F32x8 u0 = _mm256_add_ps(u_term2, u_term3);
|
||||
|
||||
F32x8 v_term0 = _mm256_mul_ps(var_tex0y, invw0);
|
||||
F32x8 v_term1 = _mm256_mul_ps(var_tex1y, invw1);
|
||||
F32x8 v_term2 = _mm256_mul_ps(var_tex2y, invw2);
|
||||
F32x8 v_term3 = _mm256_add_ps(v_term0, v_term1);
|
||||
F32x8 v0 = _mm256_add_ps(v_term2, v_term3);
|
||||
|
||||
F32x8 u1 = _mm256_div_ps(u0, interpolated_w);
|
||||
F32x8 v1 = _mm256_div_ps(v0, interpolated_w);
|
||||
|
||||
F32x8 u_floored = _mm256_floor_ps(u1);
|
||||
F32x8 v_floored = _mm256_floor_ps(v1);
|
||||
F32x8 u2 = _mm256_sub_ps(u1, u_floored);
|
||||
F32x8 v2 = _mm256_sub_ps(v1, v_floored);
|
||||
F32x8 u3 = _mm256_mul_ps(u2, var_src_x_minus_one);
|
||||
F32x8 v3 = _mm256_mul_ps(v2, var_src_y_minus_one);
|
||||
|
||||
F32x8 ui = _mm256_cvtps_epi32(u3);
|
||||
F32x8 vi = _mm256_cvtps_epi32(v3);
|
||||
|
||||
// Origin UV (0,0) is in bottom left
|
||||
_mm256_maskstore_epi32((int *)depth_pointer, should_fill, interpolated_w);
|
||||
|
||||
S32x8 indices1 = _mm256_sub_epi32(var_src_y_minus_one_int, vi);
|
||||
S32x8 indices3 = _mm256_mullo_epi32(var_src_x_int, indices1);
|
||||
S32x8 indices = _mm256_add_epi32(indices3, ui);
|
||||
|
||||
//
|
||||
// Fetch and calculate texel values
|
||||
//
|
||||
S32x8 pixel;
|
||||
if(I(should_fill, 0)) Is(pixel, 0) = src->pixels[Is(indices, 0)];
|
||||
if(I(should_fill, 1)) Is(pixel, 1) = src->pixels[Is(indices, 1)];
|
||||
if(I(should_fill, 2)) Is(pixel, 2) = src->pixels[Is(indices, 2)];
|
||||
if(I(should_fill, 3)) Is(pixel, 3) = src->pixels[Is(indices, 3)];
|
||||
if(I(should_fill, 4)) Is(pixel, 4) = src->pixels[Is(indices, 4)];
|
||||
if(I(should_fill, 5)) Is(pixel, 5) = src->pixels[Is(indices, 5)];
|
||||
if(I(should_fill, 6)) Is(pixel, 6) = src->pixels[Is(indices, 6)];
|
||||
if(I(should_fill, 7)) Is(pixel, 7) = src->pixels[Is(indices, 7)];
|
||||
|
||||
S32x8 texel_i_a = _mm256_and_si256(pixel, var_0xff000000);
|
||||
S32x8 texel_i_b = _mm256_and_si256(pixel, var_0x00ff0000);
|
||||
S32x8 texel_i_g = _mm256_and_si256(pixel, var_0x0000ff00);
|
||||
S32x8 texel_i_r = _mm256_and_si256(pixel, var_0x000000ff);
|
||||
|
||||
// Alpha is done this way because signed integer shift is weird
|
||||
// When sign bit is set it sets all bits that we shift the sign through
|
||||
// So first we shift
|
||||
texel_i_a = _mm256_srai_epi32(texel_i_a, 24);
|
||||
texel_i_a = _mm256_and_si256(texel_i_a, var_0x000000ff);
|
||||
texel_i_b = _mm256_srai_epi32(texel_i_b, 16);
|
||||
texel_i_g = _mm256_srai_epi32(texel_i_g, 8 );
|
||||
texel_i_r = _mm256_srai_epi32(texel_i_r, 0 );
|
||||
|
||||
F32x8 texel_a0 = _mm256_cvtepi32_ps(texel_i_a);
|
||||
F32x8 texel_b0 = _mm256_cvtepi32_ps(texel_i_b);
|
||||
F32x8 texel_g0 = _mm256_cvtepi32_ps(texel_i_g);
|
||||
F32x8 texel_r0 = _mm256_cvtepi32_ps(texel_i_r);
|
||||
|
||||
F32x8 texel_a1 = _mm256_mul_ps(texel_a0, inv255);
|
||||
F32x8 texel_b1 = _mm256_mul_ps(texel_b0, inv255);
|
||||
F32x8 texel_g1 = _mm256_mul_ps(texel_g0, inv255);
|
||||
F32x8 texel_r1 = _mm256_mul_ps(texel_r0, inv255);
|
||||
|
||||
texel_r1 = _mm256_mul_ps(texel_r1, texel_r1);
|
||||
texel_g1 = _mm256_mul_ps(texel_g1, texel_g1);
|
||||
texel_b1 = _mm256_mul_ps(texel_b1, texel_b1);
|
||||
|
||||
//
|
||||
// Fetch and calculate dst pixels
|
||||
//
|
||||
U32 *dst_memory = destination + x8;
|
||||
S32x8 dst_pixel = _mm256_maskload_epi32((const int *)dst_memory, should_fill);
|
||||
|
||||
S32x8 dst_i_a0 = _mm256_and_si256(dst_pixel, var_0xff000000);
|
||||
S32x8 dst_i_b0 = _mm256_and_si256(dst_pixel, var_0x00ff0000);
|
||||
S32x8 dst_i_g0 = _mm256_and_si256(dst_pixel, var_0x0000ff00);
|
||||
S32x8 dst_i_r0 = _mm256_and_si256(dst_pixel, var_0x000000ff);
|
||||
|
||||
S32x8 dst_i_a1 = _mm256_srai_epi32(dst_i_a0, 24);
|
||||
dst_i_a1 = _mm256_and_si256(dst_i_a1, var_0x000000ff);
|
||||
S32x8 dst_i_b1 = _mm256_srai_epi32(dst_i_b0, 16);
|
||||
S32x8 dst_i_g1 = _mm256_srai_epi32(dst_i_g0, 8);
|
||||
S32x8 dst_i_r1 = dst_i_r0;
|
||||
|
||||
F32x8 dst_a = _mm256_cvtepi32_ps(dst_i_a1);
|
||||
F32x8 dst_b = _mm256_cvtepi32_ps(dst_i_b1);
|
||||
F32x8 dst_g = _mm256_cvtepi32_ps(dst_i_g1);
|
||||
F32x8 dst_r = _mm256_cvtepi32_ps(dst_i_r1);
|
||||
|
||||
dst_a = _mm256_mul_ps(dst_a, inv255);
|
||||
dst_b = _mm256_mul_ps(dst_b, inv255);
|
||||
dst_g = _mm256_mul_ps(dst_g, inv255);
|
||||
dst_r = _mm256_mul_ps(dst_r, inv255);
|
||||
|
||||
dst_r = _mm256_mul_ps(dst_r, dst_r);
|
||||
dst_g = _mm256_mul_ps(dst_g, dst_g);
|
||||
dst_b = _mm256_mul_ps(dst_b, dst_b);
|
||||
|
||||
// Premultiplied alpha
|
||||
{
|
||||
dst_r = _mm256_add_ps(texel_r1, _mm256_mul_ps(_mm256_sub_ps(var1,texel_a1), dst_r));
|
||||
dst_g = _mm256_add_ps(texel_g1, _mm256_mul_ps(_mm256_sub_ps(var1,texel_a1), dst_g));
|
||||
dst_b = _mm256_add_ps(texel_b1, _mm256_mul_ps(_mm256_sub_ps(var1,texel_a1), dst_b));
|
||||
dst_a = _mm256_sub_ps(_mm256_add_ps(texel_a1, dst_a), _mm256_mul_ps(texel_a1,dst_a));
|
||||
}
|
||||
|
||||
// Almost linear to srgb
|
||||
{
|
||||
dst_r = _mm256_sqrt_ps(dst_r);
|
||||
dst_g = _mm256_sqrt_ps(dst_g);
|
||||
dst_b = _mm256_sqrt_ps(dst_b);
|
||||
}
|
||||
|
||||
// Convert to integer format
|
||||
dst_r = _mm256_mul_ps(dst_r, var255);
|
||||
dst_g = _mm256_mul_ps(dst_g, var255);
|
||||
dst_b = _mm256_mul_ps(dst_b, var255);
|
||||
dst_a = _mm256_mul_ps(dst_a, var255);
|
||||
|
||||
S32x8 dst_r_int = _mm256_cvtps_epi32(dst_r);
|
||||
S32x8 dst_g_int = _mm256_cvtps_epi32(dst_g);
|
||||
S32x8 dst_b_int = _mm256_cvtps_epi32(dst_b);
|
||||
S32x8 dst_a_int = _mm256_cvtps_epi32(dst_a);
|
||||
|
||||
S32x8 dst_int_a_shifted = _mm256_slli_epi32(dst_a_int, 24);
|
||||
S32x8 dst_int_b_shifted = _mm256_slli_epi32(dst_b_int, 16);
|
||||
S32x8 dst_int_g_shifted = _mm256_slli_epi32(dst_g_int, 8);
|
||||
S32x8 dst_int_r_shifted = dst_r_int;
|
||||
|
||||
S32x8 packed_abgr0 = _mm256_or_si256(dst_int_a_shifted, dst_int_b_shifted);
|
||||
S32x8 packed_abgr1 = _mm256_or_si256(dst_int_r_shifted, dst_int_g_shifted);
|
||||
S32x8 packed_abgr2 = _mm256_or_si256(packed_abgr1, packed_abgr0);
|
||||
|
||||
_mm256_maskstore_epi32((int *)dst_memory, should_fill, packed_abgr2);
|
||||
}
|
||||
Cy0 -= dx10;
|
||||
Cy1 -= dx21;
|
||||
Cy2 -= dx02;
|
||||
destination += dst->x;
|
||||
}
|
||||
U64 end_time = __rdtsc();
|
||||
|
||||
filled_pixel_total_time += end_time - fill_pixels_begin;
|
||||
filled_pixel_count += (max_x - min_x)*(max_y - min_y);
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user