diff options
Diffstat (limited to 'plugins/gpu_neon/psx_gpu/psx_gpu_arm_neon.S')
| -rw-r--r-- | plugins/gpu_neon/psx_gpu/psx_gpu_arm_neon.S | 11 |
1 files changed, 8 insertions, 3 deletions
diff --git a/plugins/gpu_neon/psx_gpu/psx_gpu_arm_neon.S b/plugins/gpu_neon/psx_gpu/psx_gpu_arm_neon.S index 6393e15..3239412 100644 --- a/plugins/gpu_neon/psx_gpu/psx_gpu_arm_neon.S +++ b/plugins/gpu_neon/psx_gpu/psx_gpu_arm_neon.S @@ -182,6 +182,7 @@ #define uvrg_dx3l d6 #define uvrg_dx3h d7 +#define uvrgb_phase q13 .align 4 @@ -313,11 +314,16 @@ function(compute_all_gradients) vmull.s16 ga_uvrg_y, d0_b, d1_b rsbmi ga_bx, ga_bx, #0 + @ r12 = psx_gpu->uvrgb_phase + ldr r12, [ psx_gpu, #psx_gpu_uvrgb_phase_offset ] + vmlsl.s16 ga_uvrg_y, d2_b, d3_b movs gs_by, ga_by, asr #31 vshr.u64 d0, d30, #22 - mov b_base, b0, lsl #16 + add b_base, r12, b0, lsl #16 + + vdup.u32 uvrgb_phase, r12 rsbmi ga_by, ga_by, #0 vclt.s32 gs_uvrg_x, ga_uvrg_x, #0 @ gs_uvrg_x = ga_uvrg_x < 0 @@ -326,7 +332,6 @@ function(compute_all_gradients) ldrb r12, [ psx_gpu, #psx_gpu_triangle_winding_offset ] vclt.s32 gs_uvrg_y, ga_uvrg_y, #0 @ gs_uvrg_y = ga_uvrg_y < 0 - add b_base, b_base, #0x8000 rsb r12, r12, #0 @ r12 = -(triangle->winding) vdup.u32 w_mask, r12 @ w_mask = { -w, -w, -w, -w } @@ -335,7 +340,7 @@ function(compute_all_gradients) vshll.u16 uvrg_base, uvrg0, #16 @ uvrg_base = uvrg0 << 16 vdup.u32 r_shift, r14 @ r_shift = { shift, shift, shift, shift } - vorr.u32 uvrg_base, #0x8000 + vadd.u32 uvrg_base, uvrgb_phase vabs.s32 ga_uvrg_x, ga_uvrg_x @ ga_uvrg_x = abs(ga_uvrg_x) vmov area_r_s, s0 @ area_r_s = triangle_reciprocal |